An ai-based power spectrum-based power supply device state monitoring system and method

Abstract

The application discloses a power supply equipment state monitoring system and method based on AI power spectrum, and relates to the technical field of electrical intelligent sensing; the method comprises the following steps: obtaining original analog signals of the power supply equipment through real-time synchronous acquisition, and preprocessing the original analog signals to obtain digital waveform data frames; performing time domain feature calculation on the digital waveform data frames to obtain a time domain feature set; performing frequency spectrum feature extraction on the digital waveform data frames to obtain a frequency domain feature set; performing high-order statistical analysis and nonlinear signal processing on the digital waveform data frames to obtain a nonlinear feature vector; based on the frequency domain feature set, the nonlinear feature vector and the time domain feature set, combining a pre-trained artificial intelligence model, outputting a fault type identifier and a corresponding fault probability, and calculating an abnormal score based on the fault type identifier and the corresponding fault probability; the application can effectively identify the camouflage fault of the power supply equipment and reduce the cumulative risk of potential faults.

Description

Technical Field

[0001] This invention relates to the field of electrical intelligent sensing technology, and more specifically, to a power equipment status monitoring system and method based on AI power spectrum. Background Technology

[0002] During the operation of power systems, various power supply equipment, such as grid-connected inverters, switchgear, and industrial motors, need to use electrical sensing technology to collect key signals such as voltage and current, and use condition monitoring technology to assess their operating status in real time to ensure system stability. In existing technologies, equipment condition monitoring largely relies on voltage and current waveform data acquired by electrical sensors. It extracts explicit features such as signal frequency amplitude and harmonic distortion rate, and combines these with preset thresholds or AI models to determine the condition. The underlying principle is that normally operating equipment exhibits a stable frequency distribution and amplitude range, while fault conditions are usually accompanied by abnormal deviations in these explicit features. Therefore, traditional monitoring methods and AI models mostly use these explicit features as the core basis for judgment.

[0003] However, when some equipment malfunctions, its built-in control system activates protection mechanisms, such as entering standby mode or bypass operation. It actively adjusts output parameters to control explicit characteristics like frequency amplitude and harmonic distortion rate within the acceptable range specified by industry standards, thus meeting basic compliance requirements. But at this point, a fault already exists within the equipment, causing changes in its deeper operational characteristics. For example, the randomness of frequency fluctuations, the evolution of phase over time, and higher-order statistical properties all show significant differences from normal conditions. Because existing monitoring technologies and AI models do not effectively mine and analyze these deeper characteristics in the signals captured by electrical sensors, relying solely on explicit characteristics for status assessment makes it difficult to identify this seemingly compliant but actually abnormal state of the equipment.

[0004] The aforementioned problems can cause monitoring systems to misjudge equipment as being in normal operating condition, preventing maintenance personnel from promptly identifying potential faults and taking corrective action. As faults accumulate, equipment may experience sudden failures without warning, leading to unplanned power outages, power system fluctuations, and other issues that affect the safe and stable operation of the power system.

[0005] In view of this, the present invention proposes a power equipment condition monitoring system and method based on AI power spectrum to solve the above problems. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art and achieve the above objectives, the present invention provides the following technical solution: a power equipment status monitoring method based on AI power spectrum, comprising:

[0007] The original analog signals of the power supply equipment are acquired in real time and synchronously, and the original analog signals are preprocessed to obtain digital waveform data frames;

[0008] Time-domain features are calculated on digital waveform data frames to obtain a set of time-domain features;

[0009] Spectral features are extracted from digital waveform data frames to obtain a set of frequency domain features;

[0010] High-order statistical analysis and nonlinear signal processing are performed on digital waveform data frames to obtain nonlinear eigenvectors.

[0011] Based on the frequency domain feature set, nonlinear feature vector and time domain feature set, combined with the pre-trained artificial intelligence model, the fault type identifier and the corresponding fault probability are output, and the anomaly score is calculated based on the fault type identifier and the corresponding fault probability.

[0012] Based on the fault type identifier, the corresponding fault probability, and the anomaly score, a Bayesian confidence assessment is performed to obtain the final equipment status and confidence score.

[0013] Furthermore, the nonlinear eigenvectors include noise features, bispectral features, and envelope demodulation features;

[0014] Noise characteristics include the kurtosis values ​​and center frequencies of the characteristic frequency bands of the three-phase voltage and current; bispectral characteristics include the total bispectral energy and peak percentage of the three-phase voltage and current; envelope demodulation characteristics include the peak values ​​of the characteristic frequencies of the envelope spectrum of the three-phase voltage and current, the peak signal-to-noise ratio, and the number of peaks.

[0015] Furthermore, methods for obtaining noise features include:

[0016] The digital waveform data frame is filtered based on the set frequency band to obtain high-frequency signals of three-phase voltage and current; the kurtosis values ​​of three-phase voltage and current are calculated using a preset kurtosis calculation formula.

[0017] Multi-resolution filtering is performed on the digital waveform data frame to obtain multiple equal-width sub-bands. Each sub-band corresponds to a center frequency. The kurtosis of each sub-band signal is calculated to obtain multiple kurtosis values, forming a spectral kurtosis curve. The center frequency and kurtosis peak value of the sub-band corresponding to the peak point of the spectral kurtosis curve are recorded, and the center frequency of the sub-band corresponding to the peak point is taken as the center frequency of the characteristic frequency band.

[0018] Furthermore, methods for obtaining bispectral features include:

[0019] Multiple sampling points of three-phase voltage and current digital waveform data frames are pre-selected, and the mean value of each sampling point is subtracted to obtain a zero-mean sequence.

[0020] The third-order cumulant of a zero-mean sequence is calculated using a sample estimation method.

[0021] Performing a fast Fourier transform on the third-order cumulants yields a bispectral output.

[0022] Sum the amplitudes of all bispectral values ​​to obtain the total bispectral energy of the three-phase voltage and current.

[0023] The mean amplitude of the bispectral spectrum is calculated, and a screening threshold is set based on the mean amplitude of the bispectral spectrum. The frequency points corresponding to the bispectral amplitudes that are greater than the screening threshold are recorded as peak points.

[0024] The ratio of the total energy at the peak point to the total energy of the bispectral spectrum is denoted as the peak percentage.

[0025] Furthermore, methods for obtaining envelope demodulation features include:

[0026] Filtering digital waveform data frames based on the center frequency of the characteristic frequency band yields narrowband signals within the characteristic frequency bands of three-phase voltage and current.

[0027] Perform a Hilbert transform on a narrowband signal to obtain an analytic signal;

[0028] Calculate the amplitude of the analytic signal to obtain the envelope signal;

[0029] Perform a fast Fourier transform on the envelope signal to obtain the envelope spectrum;

[0030] In the envelope spectrum, the maximum amplitude within a set range near the pre-calculated fault characteristic frequency is searched, and the maximum amplitude is recorded as the characteristic frequency peak value of the three-phase voltage and current; the ratio of the characteristic frequency peak value to the average amplitude of the envelope spectrum is recorded as the peak signal-to-noise ratio; the number of peak values ​​in the envelope spectrum with a peak signal-to-noise ratio greater than or equal to the set peak value threshold is counted and recorded as the peak value.

[0031] Furthermore, the structure of the artificial intelligence model includes convolutional neural network subnetworks and fully connected subnetworks, wherein:

[0032] The convolutional neural network subnetwork consists of an input layer, a first convolutional layer, a second convolutional layer, a first max pooling layer, a second max pooling layer, a global average pooling layer, and a residual connection layer. The connection relationship is as follows: the input layer outputs to the first convolutional layer, the first convolutional layer outputs to the first max pooling layer, the first max pooling layer outputs to the second convolutional layer, and the residual connection layer adds the outputs of the first and second convolutional layers and inputs them to the second max pooling layer.

[0033] The fully connected subnetwork consists of an input layer, a first fully connected layer, a second fully connected layer, and a batch normalization layer. The connection relationship is as follows: the input layer outputs to the first fully connected layer, the first fully connected layer outputs to the batch normalization layer, and the batch normalization layer outputs to the second fully connected layer.

[0034] Furthermore, it also includes fusing the output features of the convolutional neural network subnetwork and the fully connected subnetwork, specifically including a fusion layer, an attention layer, and an output layer.

[0035] Furthermore, the input to the convolutional neural network subnetwork is a set of time-frequency spectrogram matrices, and the output is a time-frequency feature vector of a set dimension, which serves as a compressed representation of the time-frequency feature set;

[0036] The input of the fully connected subnetwork is a total feature vector formed by concatenating all one-dimensional features of the frequency domain feature set, nonlinear feature vector, and time domain feature set. The output is a numerical feature vector of a set dimension, which serves as a compressed representation of the numerical features.

[0037] Furthermore, the fault type identifiers are 0, 1, 2, and 3, where 1 corresponds to an explicit fault, 2 corresponds to a camouflaged fault, and 3 corresponds to uncertainty; the fault probability corresponding to the fault type identifier is the confidence level of the fault type identifier output by the artificial intelligence model; the anomaly score is obtained by adding the fault probability of the camouflaged fault to the fault probability of the explicit fault.

[0038] A power equipment condition monitoring system based on AI power spectrum includes:

[0039] The data acquisition module is used to acquire the raw analog signals of the power supply equipment in real time and synchronously, and to preprocess the raw analog signals to obtain digital waveform data frames.

[0040] The time-domain feature module is used to perform time-domain feature calculations on digital waveform data frames to obtain a time-domain feature set.

[0041] The frequency domain feature module is used to extract spectral features from digital waveform data frames to obtain a set of frequency domain features.

[0042] The nonlinear module performs high-order statistical analysis and nonlinear signal processing on digital waveform data frames to obtain nonlinear eigenvectors.

[0043] The fault identification module, based on frequency domain feature set, nonlinear feature vector and time domain feature set, combined with a pre-trained artificial intelligence model, outputs fault type identifier and corresponding fault probability, and calculates an anomaly score based on fault type identifier and corresponding fault probability;

[0044] The confidence assessment module performs Bayesian confidence assessment based on the fault type identifier, the corresponding fault probability, and the anomaly score to obtain the final equipment status and confidence score.

[0045] This invention utilizes a data acquisition module to synchronously acquire raw analog signals from power supply equipment in real time based on electrical sensing technology and preprocess them to obtain digital waveform data frames. A time-domain feature module calculates a time-domain feature set from the digital waveform data frames, a frequency-domain feature module extracts a frequency-domain feature set, and a nonlinear module obtains nonlinear feature vectors through high-order statistical analysis and nonlinear signal processing. A fault identification module combines these three types of features with a pre-trained artificial intelligence model to output fault type identification, fault probability, and anomaly score. A confidence assessment module obtains the final equipment status and confidence score through Bayesian confidence assessment. The artificial intelligence model includes a convolutional neural network subnetwork and a fully connected subnetwork, and also integrates the outputs of both and focuses on key features through an attention layer. All modules work together to achieve a complete monitoring process.

[0046] This invention addresses the problem in existing technologies where equipment status monitoring relies on explicit features such as frequency amplitude and harmonic distortion rate obtained through electrical sensing. This fails to identify the seemingly compliant but actually abnormal states presented by the built-in control system after equipment failure. Its advantages lie in its comprehensive extraction of equipment operating characteristics from multiple dimensions—time domain, frequency domain, and nonlinearity—based on comprehensive signals captured by electrical sensing technology. This covers characteristic patterns from normal equipment operation to various faults, avoiding omissions of information from a single dimension. The AI ​​model specifically processes and effectively integrates different morphological features, particularly improving the detection capability of disguised faults. Combined with Bayesian confidence assessment, historical data is integrated with real-time features acquired through electrical sensing to correct model bias and improve decision reliability. Each module has clearly defined functions and works collaboratively and efficiently, enabling maintenance personnel to grasp equipment status promptly and accurately, reducing the risk of potential fault accumulation, and ensuring the safe and stable operation of the power system and the continuity of industrial production. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of a power equipment status monitoring system based on AI power spectrum according to an embodiment of the present invention;

[0048] Figure 2 This is a flowchart of a power equipment status monitoring method based on AI power spectrum according to an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the structure of the artificial intelligence model in an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be described in detail, clearly, and completely below with reference to the accompanying drawings. It should be particularly noted that the specific embodiments described below are only for better illustrating and explaining the technical solutions of the present invention, and are intended to enable those skilled in the art to better understand and implement the present invention, and should not be construed as limiting the scope of protection of the present invention. Without departing from the spirit and substance of the present invention, those skilled in the art can modify, adjust, or make equivalent substitutions based on the content disclosed in the present invention, and these should all be considered within the scope of protection of the present invention.

[0051] Example 1:

[0052] Please see Figure 1 As shown, this embodiment discloses a power equipment status monitoring system based on AI power spectrum for use in a three-phase system. It includes a data acquisition module, a time-domain feature module, a frequency-domain feature module, a nonlinear module, a fault identification module, and a confidence assessment module. Each module is connected via wired and / or wireless connections to achieve data transmission.

[0053] The data acquisition module is used to acquire the raw analog signals of the power supply equipment in real time and to preprocess the raw analog signals to obtain digital waveform data frames.

[0054] The original analog signals include phase A voltage, phase B voltage, phase C voltage, phase A current, phase B current, and phase C current.

[0055] Phase A voltage is acquired using voltage sensors such as PT and resistor dividers, and the instantaneous voltage waveform is obtained through synchronous sampling by an ADC. A sampling rate of 10kHz is recommended, covering the 0–5kHz frequency band to capture the 50 / 60Hz fundamental frequency and its higher harmonics. Phase B and C voltages are acquired using the same method as phase A voltage, with synchronous sampling ensuring alignment of all phases. Phase A current is acquired using current sensors such as Hall effect sensors, sampling resistors, and current transformers, with synchronous sampling by an ADC. A sampling rate of 10kHz is recommended, synchronized with voltage sampling, covering the 0–5kHz frequency band. Phase B and C currents are acquired using the same method as phase A current.

[0056] The raw analog signal is acquired through timed sampling using the ARM SoC's built-in ADC or an external DSP and stored in a buffer for subsequent processing. An anti-aliasing filter is added to the analog front end to prevent high-frequency noise from entering; the bandwidth of the anti-aliasing filter must be less than or equal to half the sampling frequency. Acquisition of each channel must be synchronized, especially in a three-phase system to ensure correct phase relationships between phases.

[0057] A 16-bit resolution multi-channel ADC (Analog-to-Digital Converter) is used to synchronously acquire the original analog signals. Specifically, in this embodiment, the sampling rate for both voltage and current signals is set to 10kHz, meaning one sample is acquired every 0.1 milliseconds, ensuring coverage of the 0–5kHz frequency band. Hardware trigger signals, such as synchronization pulses generated by the ADC's built-in timer, control the sampling timing of all channels, ensuring that phase A voltage, phase B voltage, phase C voltage, phase A current, phase B current, and phase C current are sampled simultaneously, maintaining the phase relationship of the three-phase system. The ADC converts the analog signals into 16-bit binary two's complement digital quantities. The digital quantity range for voltage signals is 0–65535, corresponding to 0–5V, and the digital quantity range for current signals is 0–65535, corresponding to 0–3V.

[0058] An 8th-order low-pass FIR filter is applied to the digital sequences of voltage and current signals to eliminate high-frequency noise and power frequency interference. Specifically, a Hanning window is used to design the FIR filter. In this embodiment, the cutoff frequency is set to 5kHz, the passband ripple is set to less than or equal to 0.1dB, and the stopband attenuation is set to greater than or equal to 60dB. The filter coefficients are calculated using the window function method; the 8th-order filter corresponds to 9 coefficients. Real-time filtering is performed on the digital sequence of each channel using a sliding window convolution. That is, the filtering result of each new sampling point is equal to the sum of the product of that point and the previous 8 sampling points multiplied by their respective filter coefficients. For example, the filtered value of phase A voltage (a1th) = (the original value of phase A voltage (a1th) × the first coefficient) + (the original value of phase A voltage (a1-1th) × the second coefficient) + ... + (the original value of phase A voltage (a1-8th) × the ninth coefficient). The filtered signal retains only the 0-5kHz frequency band components, removing high-frequency interference.

[0059] Time alignment calibration is performed on the filtered phase A, phase B, and phase C voltages and currents, specifically: The zero-crossing times of the three-phase voltage signals are compared. The zero-crossing interval for the 50Hz fundamental wave is 10ms. The time deviations of phase B relative to phase A and phase C relative to phase A are calculated, in units of sampling points. For example, if phase A's zero-crossing point is at the 100th sampling point and phase B's is at the 102nd sampling point, the time deviation between phase B and phase A is 2 sampling points, corresponding to 0.2 milliseconds. For phases with positive deviations, such as phase B with a deviation of 2 sampling points, the signal sequence of phase B is shifted forward by 2 sampling points, i.e., the last 2 sampling points are discarded, and the copied values ​​of the first 2 sampling points are added. For phases with a deviation of -N, the signal sequence is shifted backward by the corresponding sampling points, i.e., the first N sampling points are discarded, and the copied values ​​of the last N sampling points are added.

[0060] To ensure numerical accuracy, voltage and current signals undergo sensor calibration and DC offset removal. The specific method is as follows: Calculate the correction coefficient based on the sensor's turns ratio. For example, if the turns ratio of the phase A voltage sensor is 1000:1, 1000V on the primary side corresponds to 1V on the secondary side. The ADC's 1V corresponds to a digital value of 13107, and 5V corresponds to 65535. Therefore, the phase A voltage correction coefficient = 1000 / 13107. Multiply the filtered phase A voltage digital value by this coefficient to obtain the actual voltage value in volts. The current signal is processed using the same method, calculating the correction coefficient based on the current sensor's turns ratio and converting it to amperes. Under no-load conditions, the filtered digital value of the phase A current signal is collected, and its average value is calculated as the zero-bias value. During normal operation, this zero-bias value is subtracted from the real-time collected phase A current digital value to remove the sensor's inherent DC offset. The voltage signal is processed similarly, calculating and removing the zero-bias value from the average value of samples taken under no-input conditions.

[0061] The corrected voltage and current signals are divided into time frames of fixed duration for batch processing analysis. Specifically, in this embodiment, each frame has a duration of 0.1 seconds, corresponding to 1000 sampling points (10kHz sampling rate × 0.1 seconds = 1000 points). Starting from the signal's initial point, 1000 sampling points are continuously extracted as the first frame. Then, starting from the 1001st sampling point, 1000 sampling points are extracted as the second frame, and so on, with no overlap between frames. The sampled values ​​of each frame are stored in chronological order as an array of length 1000. Each element represents the actual physical quantity after filtering, calibration, and offset removal; voltage is expressed in volts, and current in amperes.

[0062] The final result is a digital waveform data frame in the time domain, including A-phase voltage data frame, B-phase voltage data frame, C-phase voltage data frame, A-phase current data frame, B-phase current data frame, and C-phase current data frame. Each data frame is an array containing 1000 elements, and the element values ​​are the sampled values ​​of the corresponding signal within 0.1 seconds, in units of V or A.

[0063] The time-domain feature module is used to perform time-domain feature calculations on digital waveform data frames to obtain a time-domain feature set.

[0064] The time-domain feature set includes the effective values ​​of three-phase voltage and current, power and power factor, harmonic indices, and three-phase imbalance characteristics; the effective values ​​of three-phase voltage and current include the effective values ​​of phase A voltage, phase B voltage, phase C voltage, phase A current, phase B current, and phase C current; power and power factor include the total active power of phase A, phase B, phase C, total active power of all three phases, apparent power of phase A, apparent power of phase B, apparent power of phase C, total reactive power of all three phases, and power of phase A. Power factors, B-phase power factor, and C-phase power factor; harmonic indicators include the total harmonic distortion rate of A-phase voltage, the total harmonic distortion rate of A-phase current, the total harmonic distortion rate of B-phase voltage, the total harmonic distortion rate of B-phase current, the total harmonic distortion rate of C-phase voltage, the total harmonic distortion rate of C-phase current, and the amplitudes of the 2nd, 3rd, 5th, 7th, and 11th harmonics of each phase; three-phase imbalance characteristics include the amplitudes of the positive-sequence current component, the negative-sequence current component, and the zero-sequence current component, as well as potential fault signs of three-phase imbalance and potential fault signs of grounding.

[0065] The calculation method for the effective values ​​of three-phase voltage and current includes: For a 50Hz fundamental signal with a period of 20ms, a 0.1-second time frame contains 5 complete fundamental cycles, corresponding to 200 sampling points. 20ms / 0.1ms = 200 sampling points, with 0.1ms as the sampling interval, calculated using a 10kHz sampling rate. From 1000 sampling points per frame, 5 consecutive complete cycle sampling points are selected, totaling 1000 sampling points, which exactly cover 5 cycles, without truncation. The effective value is then calculated for the selected sampling points according to the definition formula. Taking phase A voltage as an example, let the value of the nth sampling point of phase A voltage be... (n) (n=1,2,...,1000), the calculation process for the effective value of phase A voltage is as follows: First, calculate the square of the value at each sampling point, then sum all the squared values, divide the sum by the total number of sampling points, and finally take the square root of the result to obtain the effective value of phase A voltage. For three consecutive frames of data of the same phase, calculate their effective values ​​and take the arithmetic mean as the final effective value at the previous moment of that phase. By averaging across multiple frames, the interference of instantaneous noise on the effective value calculation results is suppressed, ensuring characteristic stability. Using the same process, calculate the effective values ​​of phase B voltage, phase C voltage, phase A current, phase B current, and phase C current respectively.

[0066] Calculate power and power factor:

[0067] Taking phase A as an example, let the value of the nth sampling point of the phase A voltage be... (n), the value of the nth sampling point of the A-phase current is (n), then the instantaneous power of the nth sampling point of phase A. (n)= (n)× (n). For each of the 1000 sampling points in each frame, calculate the instantaneous power one by one to obtain the instantaneous power sequence of phase A. Summate the instantaneous power sequences of phase A and divide the sum by the total number of sampling points to obtain the total active power of phase A. Use the same process to calculate the total active power of phase B and phase C respectively. Add the total active power of phase A, phase B, and phase C to obtain the total active power of the three phases. Taking phase A as an example, the apparent power of phase A... =Equation: Phase A voltage RMS value × Phase A current RMS value. Using the same procedure, calculate the apparent power of Phase B and Phase C respectively. Taking Phase A as an example, first extract the fundamental phase of the Phase A voltage using Fourier transform. Phase A of the fundamental frequency Calculate the phase difference Δφ=| - |;A-phase power factor cos =cos(Δφ), or it can be calculated by the ratio of the total active power of phase A to the apparent power of phase A, i.e., cos =Total active power of phase A / The results from the two methods are cross-checked, and the average is taken as the final value. The reactive power of each phase is calculated, for example, the reactive power of phase A. = ×sin(Δφ), then add the three-phase reactive power together to get the total three-phase reactive power.

[0068] Calculate harmonic distortion:

[0069] In this embodiment, a 1024-point Discrete Fourier Transform (DFT) is used for the 1000 sampling points of each phase voltage and current. This involves padding with 24 zeros up to 1024 points to ensure the result is a power of 2, improving computational efficiency. With a sampling rate of 10kHz, the calculated frequency resolution is 10kHz / 1024 ≈ 9.77Hz. The fundamental frequency is located using the correspondence between frequency and DFT points. In this embodiment, 50Hz corresponds to the 5th DFT point, and 50Hz / 9.77Hz ≈ 5.12. Linear interpolation is used to correct the amplitude, improving the accuracy of the fundamental amplitude. The fundamental amplitude is then extracted. For example, the fundamental amplitude of phase A voltage is... Simultaneously, frequency points corresponding to the 2nd to 30th harmonics are extracted, which in this embodiment is 100Hz to 1500Hz, covering the main harmonic range. The amplitude of each harmonic is obtained, for example, the amplitude of the 2nd harmonic of phase A voltage. 3rd harmonic amplitude up to the amplitude of the 30th harmonic. Taking phase A voltage as an example, the total harmonic distortion (THD) is calculated as follows: First, calculate the sum of squares of the amplitudes of the 2nd to 30th harmonics. Take the square root of the sum to obtain the total harmonic amplitude. Then, divide the total harmonic amplitude by the fundamental frequency amplitude to obtain the THD of phase A voltage, i.e., THD = (200 - 300) / (300) = 300. Using the same procedure, the total harmonic distortion (THD) rates of phase A current, phase B voltage, phase B current, phase C voltage, and phase C current were calculated respectively. For each phase voltage and current, the amplitudes of the 2nd, 3rd, 5th, 7th, and 11th harmonics were recorded as supplementary features for subsequent diagnosis.

[0070] Take 1000 sampling points of the A-phase current, B-phase current, and C-phase current within the current frame to form a sequence of instantaneous three-phase current values. (n) (n) (n) (n=1,2,...,1000). Introduce the complex operator a, a= =-0.5+j a² = -0.5 - j , where j is the imaginary unit, and the positive-sequence, negative-sequence, and zero-sequence current components are calculated using a transformation formula. Among these, the positive-sequence current component... =[ (n)+a× (n)+a²× [(n)] / 3; Negative sequence component of current =[ (n)+a²× (n)+a× [(n)] / 3; Zero-sequence component of current =[ (n)+ (n)+ [(n)] / 3. Calculated from 1000 sampling points within each frame. , , The effective values ​​of the current components are calculated to obtain the amplitudes of the positive-sequence, negative-sequence, and zero-sequence components. Thresholds are set: if the amplitude of the negative-sequence component exceeds 5% of the amplitude of the positive-sequence component, it is marked as a potential sign of three-phase imbalance; if the amplitude of the zero-sequence component exceeds 2% of the amplitude of the positive-sequence component, it is marked as a potential sign of grounding fault. These markings will serve as auxiliary criteria for subsequent diagnosis.

[0071] The resulting time-domain feature set is used for monitoring equipment operation trends and serves as auxiliary input for the artificial intelligence model in subsequent steps, providing a basic electrical condition basis for fault identification. The time-domain feature set can reflect obvious faults; for example, severe overload will cause an increase in current RMS and THD. However, at risk points of disguised faults, such as when the equipment's control system maintains normal apparent values ​​of voltage and current through adjustments, keeping indicators like RMS and power within normal ranges, hidden fault signs may be masked by the protection controller and cannot be detected solely by the time-domain feature set. To address this risk, subsequent steps will introduce more in-depth spectrum analysis and higher-order statistical features to uncover masked anomalies.

[0072] The frequency domain feature module is used to extract spectral features from digital waveform data frames to obtain a set of frequency domain features.

[0073] The frequency domain feature set includes a set of time-spectrum graph matrices, a set of multi-channel time-frequency matrices, a set of wavelet energy spectrum matrices, and a set of quantitative frequency domain feature vectors. The time-spectrum graph matrix set includes single-channel time-spectrum graph matrices for all phase voltages and currents. The multi-channel time-frequency matrices include three-phase voltage multi-channel time-frequency matrices and three-phase current multi-channel time-frequency matrices, used for multi-channel parallel feature learning in artificial intelligence models. The wavelet energy spectrum matrix set includes wavelet energy spectrum matrices for all phase voltages and currents, specifically: phase B voltage wavelet energy spectrum matrix, phase C voltage wavelet energy spectrum matrix, phase A current wavelet energy spectrum matrix, phase B current wavelet energy spectrum matrix, and phase C current wavelet energy spectrum matrix. The quantitative frequency domain feature vector set includes frequency band energy feature vectors for all phase voltages and currents, specifically: frequency band energy feature vector A, frequency band energy feature vector B, frequency band energy feature vector C, frequency band energy feature vector A (current), frequency band energy feature vector B (current), and frequency band energy feature vector C (current).

[0074] By employing two time-frequency analysis methods—short-time Fourier transform and wavelet packet transform—one-dimensional digital waveform data frames are converted into two-dimensional time-frequency feature representations, comprehensively capturing the frequency components and dynamic variation patterns related to faults. Specific methods include:

[0075] In this embodiment, a window length of 256 points is selected for a data frame with a sampling rate of 10kHz. Based on the sampling interval, the time length corresponding to 256 points is 256 × 0.1ms = 25.6ms, which covers 1.28 cycles of the 50Hz fundamental frequency, ensuring both the integrity of the fundamental frequency information and maintaining sufficient time resolution. To improve the smoothness and continuity of the time-spectrum, the overlap rate between adjacent windows is set to 50%, meaning the movement step of each window is 128 points, corresponding to a time step of 128 × 0.1ms = 12.8ms. A Hanning window is used to weight the time-domain data within each window. The time-domain expression of the Hanning window is w(l) = 0.5 - 0.5cos(2πl / (L-1)), where L is the window length of 256, and l is the sampling point number within the window, ranging from 0 to 255. The Hanning window effectively suppresses spectral leakage, reduces frequency component diffusion caused by signal truncation of non-integer periods, and ensures the accuracy of the subsequent Fast Fourier Transform results. To improve computational efficiency and meet the requirement of being a power of 2, each windowed 256-point data point is padded with zeros to 512 points before a 512-point Fast Fourier Transform is performed. According to the frequency resolution formula, frequency resolution = sampling rate / number of Fast Fourier Transform points, the frequency resolution can be calculated as 10kHz / 512≈19.53Hz. This frequency resolution can clearly distinguish between adjacent harmonic and interharmonic components in the power system.

[0076] Taking the A-phase voltage data frame as an example, starting from the first sampling point, a 256-point window is extracted and a Hanning window is applied. A 512-point Fast Fourier Transform is then performed to obtain the complex spectrum of this window. According to the Nyquist criterion, only the spectral data in the 0–5kHz frequency band is retained, corresponding to the first 256 points of the Fast Fourier Transform result, while the mirror frequency band data in the 5–10kHz range is discarded. For the retained 0–5kHz complex spectrum, the power density at each frequency point is calculated. The power density is calculated by squared the amplitude at each point in the complex spectrum and then dividing by the window length and sampling rate, i.e., power density S(f) = |X(f)|² / (L× ), where X(f) is the complex spectral amplitude at frequency f, With a sampling rate of 10kHz, this calculation reflects the signal power per unit frequency and unit time, better reflecting the energy distribution analysis of fault characteristics. All 1000 sampling points of the A-phase voltage data frame are traversed sequentially according to the window movement step size, generating 8 windows: Window 1: points 1-256, Window 2: points 129-384, ..., Window 8: points 897-1000. If the last window has fewer than 256 points, it is padded with zeros to reach 256 points. The power density data of each window are arranged according to the dimension of the frequency point multiplied by the window number, forming the A-phase voltage time-spectrum matrix, denoted as time-spectrum matrix A. The height of time-spectrum matrix A is 256, corresponding to 256 frequency points in the 0-5kHz band, with a frequency interval of 19.53Hz. The width is 8, corresponding to the 8 windows, with a time interval of 12.8ms. The value of each element in the matrix is ​​the power density of the corresponding frequency and time window. Using the same process, window traversal, fast Fourier transform, and power density calculation are performed on the B-phase voltage data frame, C-phase voltage data frame, A-phase current data frame, B-phase current data frame, and C-phase current data frame, respectively, to generate the corresponding time-frequency spectrum matrix B, time-frequency spectrum matrix C, A-phase current time-frequency spectrum matrix, B-phase current time-frequency spectrum matrix, and C-phase current time-frequency spectrum matrix.

[0077] For time-frequency spectrum matrices A, B, and C, or the time-frequency spectrum matrices for phase A current, phase B current, and phase C current, double alignment of the frequency and time axes is first performed. Frequency axis alignment is achieved by confirming that the frequency points of each matrix are in the range of 0–5kHz and the intervals are all 19.53Hz. Time axis alignment is achieved by the correspondence between window numbers and actual time, i.e., all 8 windows correspond to the same time interval of 0–0.1 seconds, ensuring that the positions of the three-phase data in the time-frequency dimension are one-to-one. The aligned time-frequency spectrum matrices A, B, and C are stacked according to the channel dimension to form a three-phase voltage multi-channel time-frequency matrix. The dimension of the three-phase voltage multi-channel time-frequency matrix is ​​the number of frequency points × the number of windows × the number of channels, i.e., 256 × 8 × 3, where the first channel corresponds to the phase A voltage, the second channel corresponds to the phase B voltage, and the third channel corresponds to the phase C voltage. Using the same method, the time-frequency spectrum matrices of phase A current, phase B current, and phase C current are stacked into a three-phase current multi-channel time-frequency matrix. The structure of the multi-channel matrix is ​​consistent with the multi-channel format of image data, which can be directly input into artificial intelligence models that support multi-channel input to achieve parallel extraction and correlation analysis of three-phase features.

[0078] To address the insufficient resolution of the short-time Fourier transform in the low-frequency band, the db4 wavelet is used as the basis function for the wavelet packet transform. This is because the db4 wavelet possesses good temporal locality and frequency domain compactness, making it suitable for extracting abrupt changes in power fault signals. Based on a 10kHz sampling rate and a target frequency band of 0–5kHz, the decomposition layer is set to 5 layers: layer 1 yields 2 frequency bands (0–5kHz and 5–10kHz), layer 2 yields 4 frequency bands (0–2.5kHz, 2.5–5kHz, 5–7.5kHz, and 7.5–10kHz), and so on, until layer 5 yields 32 frequency bands. Each band has a bandwidth of 5kHz / 32 ≈ 156.25Hz, which is sufficient to finely divide the low-order harmonics, high-order harmonics, and interharmonics in the power system. Taking the A-phase voltage data frame as an example, performing 5-layer wavelet packet decomposition yields 32 wavelet packet subbands. The wavelet coefficients of each sub-band are reconstructed to recover the corresponding time-domain signal, ensuring that each sub-band signal contains only information of its corresponding frequency band. For example, sub-band 1 corresponds to 0–156.25 Hz, sub-band 2 corresponds to 156.25–312.5 Hz, ..., sub-band 32 corresponds to 4843.75–5000 Hz. The energy of each reconstructed sub-band time-domain signal is calculated. The calculation method is as follows: =Σ| (n)|², where k is the sub-band index, ranging from 1 to 32. (n) represents the time-domain sampled value of the k-th sub-band, where n is the sampling point number. The energy value reflects the signal strength of the corresponding frequency band. When a fault occurs, the energy of a specific frequency band will change significantly; for example, partial discharge will cause the energy in the 1–3 kHz frequency band to increase. The energy values ​​of the 32 sub-bands are arranged according to the dimension of sub-band number × window number to form the A-phase voltage wavelet energy spectrum matrix. The A-phase voltage wavelet energy spectrum matrix has a height of 32, corresponding to 32 frequency bands, a bandwidth of 156.25 Hz, and a width of 8, consistent with the number of windows in the short-time Fourier transform, corresponding to the same time interval. The value of each element in the A-phase voltage wavelet energy spectrum matrix is ​​the sub-band energy of the corresponding frequency band and the corresponding time window. Using the same process, the B-phase voltage wavelet energy spectrum matrix, C-phase voltage wavelet energy spectrum matrix, A-phase current wavelet energy spectrum matrix, B-phase current wavelet energy spectrum matrix, and C-phase current wavelet energy spectrum matrix are generated respectively.

[0079] The 0–5 kHz frequency band is divided into five characteristic frequency bands: low frequency, second-low frequency, mid frequency, second-high frequency, and high frequency. For example, the low frequency band is 0–500 Hz, corresponding to the fundamental frequency and the second harmonic; the second-low frequency band is 500–1000 Hz, corresponding to the third–5th harmonics; the mid frequency band is 1000–2000 Hz, corresponding to the sixth–10th harmonics; the second-high frequency band is 2000–3000 Hz, corresponding to the 11th–15th harmonics; and the high frequency band is 3000–5000 Hz, corresponding to the 16th–25th harmonics and interharmonics. Taking the time-spectrum matrix A as an example, for each time window, the sum of the power densities of all frequency points within the five characteristic frequency bands is calculated to obtain the energy values ​​of the five frequency bands for that window, such as the low frequency band energy E1, the second-low frequency band energy E2, etc. The average of the frequency band energy values ​​from the eight windows is used to obtain the frequency band energy feature vector of phase A voltage, denoted as frequency band energy feature vector A. Frequency band energy feature vector A has a dimension of 5, with each element corresponding to the average energy of a characteristic frequency band. In the time-spectrum matrix A, the maximum power density within each characteristic frequency band is located, i.e., the characteristic peak value. The frequency and power density value corresponding to the characteristic peak value are recorded to form a peak feature array, which serves as a supplement to frequency band energy feature vector A. Using the same process, quantitative frequency domain feature vectors are extracted from other time-spectrum matrices, such as frequency band energy feature vector B and frequency band energy feature vector C.

[0080] The nonlinear module performs high-order statistical analysis and nonlinear signal processing on digital waveform data frames to obtain nonlinear feature vectors.

[0081] Nonlinear eigenvectors include noise features, bispectral features, and envelope demodulation features.

[0082] By employing advanced statistical analysis and nonlinear signal processing, hidden fault information that cannot be captured by conventional features is uncovered. All calculations are performed based on single-frame data of digital waveform data, and features are extracted for each phase voltage and current separately.

[0083] Methods for obtaining noise characteristics include:

[0084] Noise characteristics include the kurtosis values ​​of the three-phase voltage and current, and the center frequency of the characteristic frequency band.

[0085] In this embodiment, a 10th-order Butterworth bandpass filter is used to filter the digital waveform data frame. The filter's passband is set to 500Hz–5kHz, the stopband attenuation is set to greater than or equal to 40dB, and the passband ripple is set to less than or equal to 1dB. Taking phase A voltage as an example, the phase A voltage data frame is input into the Butterworth bandpass filter, and the output is a high-frequency residual signal containing only components in the 500Hz–5kHz frequency band, denoted as the phase A voltage high-frequency signal. The same process is used to filter other phase voltage and current data frames to obtain the corresponding high-frequency residual signals, such as the phase A current high-frequency signal and the phase B voltage high-frequency signal. The 500Hz–5kHz frequency band avoids the power frequency (50Hz) and the 2nd–10th harmonics (100Hz–500Hz), and can focus on the high-frequency impact components caused by faults, such as the 1–3kHz pulses generated by partial discharge and the 2–5kHz vibration coupling signals caused by mechanical friction.

[0086] Kurtosis is a high-order statistic describing the steepness of a signal's probability distribution. The formula is: Kurtosis = (Fourth-order central moment of the signal) / (Square of the second-order central moment of the signal). During normal operation, high-frequency signals are mostly Gaussian noise, with a kurtosis close to 3. When a fault occurs, the non-Gaussian pulse component increases, and the kurtosis rises significantly; for example, during partial discharge, the kurtosis can rise to over 5.

[0087] Spectral kurtosis is used to automatically locate the optimal frequency band containing fault characteristics. The calculation steps are as follows:

[0088] Multi-resolution filtering is applied to the digital waveform data frames. For example, the 0–5 kHz frequency band is divided into 32 equal-width sub-bands with a bandwidth of 156.25 Hz. Each sub-band corresponds to a center frequency of 78.125 Hz, 234.375 Hz, ..., 4921.875 Hz. The kurtosis of each sub-band signal is calculated, resulting in 32 kurtosis values, forming a spectral kurtosis curve. The peak point of the spectral kurtosis curve is located, and the sub-band corresponding to the peak point is the frequency band with the most significant fault characteristics. The center frequency and kurtosis peak value of the sub-band corresponding to the peak point are recorded, and the center frequency of the sub-band corresponding to the peak point is used as the center frequency of the characteristic frequency band. The center frequency of the characteristic frequency band determined by spectral kurtosis can eliminate interference from irrelevant frequency bands, focusing on the high-frequency components related to the fault and improving the targeting of subsequent analysis.

[0089] Methods for obtaining bispectral features include:

[0090] The bispectral characteristics include the total bispectral energy and peak percentage of the three-phase voltage and current. The nonlinear phase coupling relationship of the frequency components in the signal is revealed through a two-dimensional Fourier transform of the third-order cumulants. The calculation steps are as follows:

[0091] A stable segment is selected from the digital waveform data frame, for example, 500 sampling points in the middle of the digital waveform data frame. To avoid transitional interference at the beginning and end of the frame, the mean of each sampling point is subtracted to obtain a zero-mean sequence, denoted as y(n), where n=1,2,...,500. The third-order cumulant C(τ1,τ2) is defined as C(τ1,τ2)=E[y(n)×y(n+τ1)×y(n+τ2)], where E[·] is the expectation operator, and τ1 and τ2 are time delays. In actual calculations, sample estimation is used: C(τ1,τ2)=Σy(n)×y(n+τ1)×y(n+τ2) / (N-τ1-τ2), where N=500, and the values ​​of τ1 and τ2 range from 0 to 50, covering a 5ms time delay to ensure the capture of the coupling relationship between the fundamental wave and harmonics.

[0092] Performing a 256×256-point two-dimensional fast Fourier transform on C(τ1,τ2) yields a bispectral B(ω1,ω2), where ω1 and ω2 are frequency variables in rad / s, corresponding to the 0–5 kHz frequency band. The amplitude |B(ω1,ω2)| of the bispectral reflects the coupling strength of the frequency pair (ω1,ω2). When there is nonlinearity introduced by a fault, such as frequency modulation caused by a broken rotor bar in a motor, significant peaks will appear at ω1=f1 and ω2=f2, where f1 and f2 are the fundamental or harmonic frequencies, and f1+f2 is the modulation product frequency.

[0093] Taking phase A voltage as an example, sum |B(ω1,ω2)| for all frequency pairs (ω1,ω2) to obtain the total bispectral energy, denoted as the total bispectral energy of phase A. Count the total number of frequency pairs participating in the bispectral calculation. The total number of frequency pairs is determined by the number of points in the two-dimensional fast Fourier transform, such as 256×256=65536 frequency pairs. Calculate the average value of the bispectral amplitude, i.e., the total bispectral energy of phase A divided by the total number of frequency pairs. Use 10 times the average value of the bispectral amplitude as the screening threshold to screen out peak points where |B(ω1,ω2)| is greater than the screening threshold. Record their corresponding frequency values ​​(ω1,ω2) and amplitudes to form a bispectral peak characteristic array, denoted as the bispectral peak characteristic array of phase A. The ratio of the total energy of the peak points to the total bispectral energy is denoted as the peak percentage of phase A. The higher the peak percentage, the more significant the nonlinear coupling. The peak percentage of normal equipment is usually less than 5%, but can rise to more than 15% during faults. Using the same process, the bispectral characteristics of other phase voltages and currents are calculated separately to obtain the total bispectral energy, peak percentage, and bispectral peak characteristic arrays of the three-phase voltage and current.

[0094] Methods for obtaining envelope demodulation features include:

[0095] Envelope demodulation features include the peak frequency, peak signal-to-noise ratio, and number of peaks of the envelope spectrum of three-phase voltage and current. Envelope demodulation is used to extract low-frequency fault features hidden in high-frequency carrier waves, such as the modulation signal generated by bearing faults. The calculation steps for envelope demodulation features are as follows:

[0096] The digital waveform data frame is filtered based on the center frequency of the characteristic frequency band. For example, the characteristic frequency band of phase A current is 2~2.156kHz. An 8th-order Chebyshev type I bandpass filter is used to filter the phase A current to obtain the narrowband signal z(n) within the characteristic frequency band of phase A current. A Hilbert transform is performed on z(n) to obtain the analytic signal s(n)=z(n)+j×H[z(n)], where H[·] is the Hilbert transform operator and j is the imaginary unit. The amplitude of the analytic signal is calculated, i.e., the envelope signal e(n)=|s(n)|= e(n) reflects the amplitude change of the narrowband signal over time and contains low-frequency fault modulation information. A 1024-point Fast Fourier Transform is performed on the envelope signal e(n) to obtain the envelope spectrum, with a frequency resolution of 10kHz / 1024≈9.77Hz, covering the 0–5kHz frequency band. Significant peaks appear in the envelope spectrum for fault characteristic frequencies related to the equipment's mechanical structure, such as the bearing inner ring fault frequency and rolling element fault frequency. These fault characteristic frequencies related to the equipment's mechanical structure can be pre-calculated based on the equipment parameters. In the envelope spectrum, the maximum amplitude within ±5Hz of the pre-calculated fault characteristic frequency is searched. For example, the fault frequency of the inner ring of a motor bearing is 120Hz. The maximum amplitude is recorded as the characteristic frequency peak value of the three-phase voltage and current. The ratio of the characteristic frequency peak value to the average amplitude of the envelope spectrum is recorded as the peak signal-to-noise ratio. When the peak signal-to-noise ratio is greater than or equal to 3, it indicates that the peak value is significant and can effectively distinguish between faults and noise. The number of peak values ​​with a peak signal-to-noise ratio greater than or equal to 3 in the envelope spectrum is counted and recorded as the number of peak values. The number of peak values ​​under fault conditions is usually more than that under normal conditions.

[0097] It should be noted that frequency domain feature sets may be insufficient in hidden fault scenarios. For example, weak fault signals may result in indistinct features, or multiple faults may overlap, leading to complex spectral patterns that are difficult to distinguish using only linear features. The higher-order and nonlinear analysis used in this embodiment can serve as a supplement. Even if the device controller adjusts the effective values ​​of voltage and current, harmonics, and other linear features to approach the normal range, the high-frequency impacts, nonlinear coupling, or modulation characteristics caused by hidden faults can still be reflected through indicators such as kurtosis, bispectrum, and envelope spectrum, thereby improving the confidence level of fault identification.

[0098] High-order feature computation places high demands on the computing power of embedded devices. This can be mitigated by reducing the computation frequency, such as analyzing once every second instead of once every 0.1 seconds per frame, to reduce real-time pressure. At the same time, fixed-point algorithm optimization can be adopted, such as converting floating-point operations to 16-bit fixed-point operations, to reduce computational complexity and ensure stable operation on embedded platforms such as ARM Cortex-M or DSP.

[0099] The fault identification module, based on frequency domain feature set, nonlinear feature vector and time domain feature set, combined with a pre-trained artificial intelligence model, outputs fault type identifier and corresponding fault probability, and calculates an anomaly score based on fault type identifier and corresponding fault probability.

[0100] This embodiment employs a multi-input fusion artificial intelligence model to jointly analyze frequency domain feature sets, nonlinear feature vectors, and time domain feature sets, thereby achieving accurate identification of equipment status and significantly improving the ability to detect disguised faults.

[0101] The choice of frequency domain feature sets, nonlinear feature vectors, and time domain feature sets as inputs to the artificial intelligence model stems from their comprehensive characterization of equipment operation across different dimensions, forming a complementary feature system: the time domain feature set directly reflects the statistical characteristics and basic electrical indicators of the signal in the time dimension, such as the effective values ​​of three-phase voltage and current, and power, serving as a fundamental representation of normal equipment operation; the frequency domain feature set transforms the time-domain signal into a two-dimensional frequency-time distribution through time-frequency transformation, revealing fault-related frequency components and their dynamic changes, such as power density fluctuations at specific frequencies; and the nonlinear feature vectors focus on higher-order statistical laws and nonlinear coupling relationships, such as the impact components reflected by kurtosis and the frequency coupling revealed by bispectral density, specifically capturing disguised faults that are difficult to detect using conventional features. The combination of frequency domain feature sets, nonlinear feature vectors, and time domain feature sets can cover all feature patterns of equipment from normal operation to various faults, avoiding information omissions from single feature dimensions.

[0102] Please see Figure 3 As shown, the structure of the artificial intelligence model includes a convolutional neural network subnetwork and a fully connected subnetwork, specifically constructed as follows:

[0103] A convolutional neural network subnetwork is designed for the time-spectrum matrix set to extract local correlation features in the spatial and frequency dimensions. Specific methods include:

[0104] The convolutional neural network subnetwork consists of an input layer, a first convolutional layer, a second convolutional layer, a first max pooling layer, a second max pooling layer, a global average pooling layer, and a residual connection layer.

[0105] For example, the input layer adjusts the input dimension based on the time-frequency spectrum matrix set. The first convolutional layer uses 16 3×3 convolutional kernels with a stride of 1 and the same edge padding to maintain the output size consistent with the input. Combined with the ReLU activation function, it extracts local frequency abrupt changes in the time-frequency spectrum, such as power density jumps at specific frequency points caused by faults. The first max-pooling layer uses 2×2 pooling kernels with a stride of 2, compressing the output size to 128×4×16, reducing the total number of parameters while retaining key features. The second convolutional layer uses 8 3×3 convolutional kernels with a stride of 1 and the same edge padding, combined with the ReLU activation function, to capture dynamic changes across windows in the time-frequency spectrum, such as abnormal frequency fluctuations in camouflaged faults. The residual connection layer adds the output of the first convolutional layer to the output of the second convolutional layer, alleviating the gradient vanishing problem in deep networks and enhancing the propagation of weak features such as subtle spectral changes in camouflaged faults. The second max-pooling layer uses 2×2 pooling kernels with a stride of 2, compressing the output size to 64×2×8. The global average pooling layer is used to perform global average pooling on the output of the second max pooling layer, converting the 64×2×8 three-dimensional features into an 8-dimensional time-frequency feature vector as a compressed representation of the time-frequency features.

[0106] Numerical design of fully connected subnetworks for frequency domain feature sets, nonlinear feature vectors, and time domain feature sets to extract global correlations between features. Specific methods include:

[0107] The fully connected subnetwork consists of an input layer, a first fully connected layer, a second fully connected layer, and a batch normalization layer.

[0108] For example, the input layer concatenates all one-dimensional features from the frequency domain feature set, nonlinear feature vector, and time domain feature set (excluding the time-spectrum matrix) into a total feature vector. The total feature vector has a dimension of 128, including 5×6=30 dimensions of energy feature vectors for each phase band, 36 dimensions of nonlinear feature vectors, and 62 dimensions of the time domain feature set. The first fully connected layer contains 32 neurons and, with the ReLU activation function, performs a nonlinear transformation on the input vector to extract low-order correlation features, such as the correlation between harmonic distortion rate and effective current value. The batch normalization layer batch normalizes the output of the first fully connected layer, accelerating model convergence and enhancing generalization ability. The second fully connected layer contains 16 neurons and, with the ReLU activation function, extracts high-order correlation features, such as the co-occurrence anomaly pattern of kurtosis and bispectral peaks, which is a typical feature of masquerading faults. The output of the fully connected sub-network is the 16-dimensional feature vector generated by the second fully connected layer, serving as a compressed representation of the numerical features.

[0109] The separate input of frequency domain feature sets, nonlinear feature vectors, and time domain feature sets into convolutional neural network subnetworks and fully connected subnetworks is determined by the morphology of the features and the strengths of the networks: the time-spectrum matrix in the frequency domain feature set is a two-dimensional image-like feature containing spatial local correlation patterns. The convolutional neural network subnetwork, through convolution and pooling operations, excels at extracting these local correlation features, such as the power density variation trend at different frequency points. Conversely, the quantitative frequency domain feature vectors, nonlinear feature vectors, and time domain feature sets in the frequency domain feature set are all one-dimensional numerical features, focusing on global correlations between features, such as the linkage between harmonic indicators and power. The fully connected subnetwork, through nonlinear transformations of multiple layers of neurons, is more suitable for uncovering these global correlation patterns. Separate processing fully leverages the advantages of both types of networks, improving the targeting and efficiency of feature extraction.

[0110] The output features of the convolutional neural network (CNN) subnetwork and the fully connected subnetwork are fused to achieve joint decision-making based on global and local features. This process includes a fusion layer, an attention layer, and an output layer. The fusion layer concatenates the 8-dimensional feature vector from the CNN subnetwork with the 16-dimensional feature vector from the fully connected branch, forming a 24-dimensional fused feature vector. The attention layer applies a channel attention mechanism to the fused feature vector, assigning different weights to each dimension of the feature vector by learning a weight matrix (24×1). For example, in cases of camouflaged faults, the dimension corresponding to the nonlinear feature vector is given a higher weight, enhancing the model's attention to hidden features. The output layer contains four neurons, corresponding to four fault type labels: 0 for normal, 1 for overt fault, 2 for camouflaged fault, and 3 for uncertain. Using a softmax activation function, it outputs the probability values ​​for each device state, with a sum of probabilities equal to 1.

[0111] The intrinsic relationship between frequency domain feature sets, nonlinear feature vectors, and time domain feature sets is a progression and complement from basic to advanced, and from linear to nonlinear: time domain feature sets are the most direct representation of signals, providing raw data for frequency domain analysis; frequency domain feature sets are an extension of time domain features in the frequency dimension, transforming imperceptible dynamic changes in the time domain into identifiable frequency features; nonlinear feature vectors are a deeper exploration of the former two, capturing the implicit higher-order statistical properties and nonlinear interactions. Together, the three constitute a complete feature chain, ensuring comprehensive coverage of device status information.

[0112] In this embodiment, the total number of parameters of the artificial intelligence model is controlled to within 80,000 through lightweight design: about 50,000 parameters for convolutional layers and about 30,000 parameters for fully connected layers, adapting to the memory limitations of embedded platforms.

[0113] Training methods for artificial intelligence models include:

[0114] Collect feature data of the equipment under different states, including normal operation samples, explicit fault samples, camouflaged fault samples, and uncertain state samples. Each state should have at least 5000 samples to ensure the model fully learns various patterns. Manually label each sample with a state label. For camouflaged fault samples, internal faults must be verified by disassembling the equipment, such as winding aging or bearing wear, to ensure label accuracy. Divide the data into training, validation, and test sets in a 7:2:1 ratio, ensuring consistent proportions of samples for each state across all three sets to avoid class imbalance.

[0115] In this embodiment, the cross-entropy loss function is used, and the calculation formula is Loss=-Σ( ×log( ), where Loss is the loss value, One-hot encoding of the true labels of the samples. This represents the probability value output by the model. In this embodiment, the Adam optimizer is used, with an initial learning rate of 0.001, which decays to 0.9 times every 10 training epochs. Momentum parameters β1=0.9, β2=0.999, and weight decay coefficient 0.0001 to suppress overfitting. The training epochs are set to 50, with a batch size of 32. An early stopping mechanism is used: if the validation set loss does not decrease for 5 consecutive epochs, training stops, and the optimal model parameters are saved.

[0116] In this embodiment, accuracy and spoofing fault recall are used as the core evaluation metrics. Accuracy is required to be greater than or equal to 92%, and spoofing fault recall greater than or equal to 90%. Accuracy represents the proportion of correctly classified samples, and spoofing fault recall represents the proportion of correctly identified spoofing fault samples out of actual spoofing fault samples. If the spoofing fault recall is insufficient, the weight of spoofing fault samples in the training set is increased, and the corresponding class weight in the loss function is set to 2. If overfitting exists, a dropout layer is added to the fully connected layer, with a dropout rate of 0.2.

[0117] The diagnostic results of the artificial intelligence model include fault type identification, corresponding fault probability, and anomaly score. The fault probability comprises four probability values, each corresponding to the confidence level of the fault type identification, with the sum of the probability values ​​being 1. The anomaly score is obtained by adding the probability of a camouflaged fault to the probability of a visible fault, with a value ranging from 0 to 1, where 0 indicates no anomaly and 1 indicates an extremely high risk of anomaly, used to quickly determine the overall degree of anomaly of the equipment.

[0118] The confidence assessment module performs Bayesian confidence assessment based on the fault type identifier, the corresponding fault probability, and the anomaly score to obtain the final equipment status and confidence score.

[0119] Based on historical fault data and real-time features, confidence correction of the fault probabilities output by the artificial intelligence model can improve decision reliability. In this embodiment, the prior probabilities are set according to the historical occurrence probabilities of each state based on the equipment's operating records over the past three years. The prior probability of normal state P(normal) = 0.85, the prior probability of overt fault P(overt fault) = 0.10, the prior probability of camouflaged fault P(camouflaged fault) = 0.04, and the prior probability of uncertain state P(uncertain) = 0.01. The likelihood function is defined using the fault probabilities output by the AI ​​model as the likelihood function. That is, given real-time feature X, the conditional probabilities of each state are P(X|normal) = normal probability, P(X|overt fault) = overt fault probability, P(X|camouflaged fault) = camouflaged fault probability, and P(X|uncertain) = uncertain probability. The posterior probability is calculated based on Bayes' theorem, where the posterior probabilities of each state are P(normal|X)=[P(X|normal)×P(normal)] / P(X), P(dominant fault|X)=[P(X|dominant fault)×P(dominant fault)] / P(X), and P(masked fault|X)=[P(X|masked fault)×P(masked fault)] / P(X), where P(X) is the evidence factor. Then the posterior probability is equal to P(X|normal)×P(normal)+P(X|dominant fault)×P(dominant fault)+P(X|masked fault)×P(masked fault)+P(X|uncertain)×P(uncertain). The confidence score generation process involves normalizing the posterior probability (ensuring the sum is 1) to obtain the Bayesian confidence score for each state, denoted as the normal Bayesian confidence score, the explicit fault Bayesian confidence score, and the masquerade fault Bayesian confidence score. The Bayesian confidence score integrates historical data and real-time features, and can correct model bias when the samples are imbalanced.

[0120] For example, based on Bayesian confidence scores, nonlinear feature vectors, and time-domain feature sets, setting three-level decision rules can clarify the criteria for determining the equipment status.

[0121] A fault state is determined when any of the following conditions are met: a manifest fault has a Bayesian confidence score greater than or equal to 0.8; a camouflaged fault has a Bayesian confidence score greater than or equal to 0.7; a camouflaged fault has a Bayesian confidence score between 0.5 and 0.7, and at least two abnormal items in the nonlinear feature vector; or an abnormal score greater than or equal to 0.9, and one or more items in the time-domain feature set exceed the limit, such as the effective value of the phase A current exceeding 1.1 times the rated value, or the total harmonic distortion rate of the phase A voltage greater than or equal to 5%.

[0122] A warning state is determined when the fault condition is not met, but any of the following conditions are met: the Bayesian confidence score for a manifest fault is between 0.3 and 0.8; the Bayesian confidence score for a camouflaged fault is between 0.3 and 0.7, and there is one anomalous term in the nonlinear feature vector; the anomalous score is between 0.6 and 0.9, and the time-domain feature set does not exceed the limit but there is an abnormal trend, such as the effective value of the A-phase current increasing by more than or equal to 2% for three consecutive frames.

[0123] A state is considered normal when all of the following conditions are met: a normal Bayesian confidence score greater than or equal to 0.7; an anomaly score less than or equal to 0.3; no anomalies in the nonlinear feature vectors; and all key temporal feature sets are within the normal range. If none of the above rules are met, the state is temporarily suspended, triggering a secondary analysis of three consecutive frames of data. The majority value of the three analysis results is taken as the final state; for example, two warnings result in a warning.

[0124] Based on the final determined equipment status, the corresponding alarm actions are executed. In a fault state, a Level 1 alarm is triggered, including a local audible and visual alarm, a remote communication alarm, and a protection action command. The local audible and visual alarm is indicated by a constantly lit red indicator light and an intermittent buzzer sound every 2 seconds. The remote communication alarm sends a fault message to the monitoring center via the Modbus protocol, including the fault type, confidence score, and occurrence time. If it is a masquerade fault, only an alarm is triggered without shutting down the equipment, awaiting manual confirmation. If it is a manifest fault with a Bayesian confidence score greater than or equal to 0.9, the equipment is shut down for protection. In a warning state, a Level 2 alarm is triggered, including a flashing local yellow indicator light and a remote communication warning message. No protection action is triggered, but the next monitoring cycle is shortened, for example, from 0.1 seconds / frame to 0.05 seconds / frame. In a normal state, there is no alarm output; the normal monitoring cycle is maintained, and only the normal status is displayed on the monitoring interface.

[0125] The final equipment status assessment results and control commands include equipment status, fault category, confidence score, and control commands. Equipment status is categorized as normal, warning, and fault, clearly indicating the current operating status of the equipment; fault category is categorized as explicit fault or camouflaged fault, corresponding to fault type identifier 1 or 2; confidence score is the score for the corresponding status in the Bayesian confidence score, ranging from 0 to 1; control commands include alarm commands, protection commands, and monitoring cycle adjustment commands.

[0126] The above output is executed through the embedded system's I / O interface and communication module, ensuring that maintenance personnel can monitor the equipment status in real time, especially providing early warnings of camouflaged faults, and avoiding serious consequences caused by the accumulation of faults.

[0127] Example 2:

[0128] Please see Figure 2As shown, this embodiment provides a power equipment status monitoring method based on AI power spectrum, including:

[0129] The original analog signals of the power supply equipment are acquired in real time and synchronously, and the original analog signals are preprocessed to obtain digital waveform data frames;

[0130] Time-domain features are calculated on digital waveform data frames to obtain a set of time-domain features;

[0131] Spectral features are extracted from digital waveform data frames to obtain a set of frequency domain features;

[0132] High-order statistical analysis and nonlinear signal processing are performed on digital waveform data frames to obtain nonlinear eigenvectors.

[0133] Based on the frequency domain feature set, nonlinear feature vector and time domain feature set, combined with the pre-trained artificial intelligence model, the fault type identifier and the corresponding fault probability are output, and the anomaly score is calculated based on the fault type identifier and the corresponding fault probability.

[0134] Based on the fault type identifier, the corresponding fault probability, and the anomaly score, a Bayesian confidence assessment is performed to obtain the final equipment status and confidence score.

[0135] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0136] In conclusion, 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 spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring the condition of power supply equipment based on AI power spectrum, characterized in that, include: The original analog signals of the power supply equipment are acquired in real time and synchronously, and the original analog signals are preprocessed to obtain digital waveform data frames; Time-domain features are calculated on digital waveform data frames to obtain a set of time-domain features; Spectral features are extracted from digital waveform data frames to obtain a set of frequency domain features; High-order statistical analysis and nonlinear signal processing are performed on digital waveform data frames to obtain nonlinear eigenvectors. Based on the frequency domain feature set, nonlinear feature vector and time domain feature set, combined with the pre-trained artificial intelligence model, the fault type identifier and the corresponding fault probability are output, and the anomaly score is calculated based on the fault type identifier and the corresponding fault probability. Based on the fault type identifier, the corresponding fault probability, and the anomaly score, a Bayesian confidence assessment is performed to obtain the final equipment status and confidence score.

2. The power equipment status monitoring method based on AI power spectrum according to claim 1, characterized in that, Nonlinear eigenvectors include noise features, bispectral features, and envelope demodulation features; Noise characteristics include the kurtosis values ​​and center frequencies of the characteristic frequency bands of the three-phase voltage and current; bispectral characteristics include the total bispectral energy and peak percentage of the three-phase voltage and current; envelope demodulation characteristics include the peak values ​​of the characteristic frequencies of the envelope spectrum of the three-phase voltage and current, the peak signal-to-noise ratio, and the number of peaks.

3. The power equipment status monitoring method based on AI power spectrum according to claim 2, characterized in that, Methods for obtaining noise characteristics include: The digital waveform data frame is filtered based on the set frequency band to obtain high-frequency signals of three-phase voltage and current; the kurtosis values ​​of three-phase voltage and current are calculated using a preset kurtosis calculation formula. Multi-resolution filtering is performed on the digital waveform data frame to obtain multiple equal-width sub-bands. Each sub-band corresponds to a center frequency. The kurtosis of each sub-band signal is calculated to obtain multiple kurtosis values, forming a spectral kurtosis curve. The center frequency and kurtosis peak value of the sub-band corresponding to the peak point of the spectral kurtosis curve are recorded, and the center frequency of the sub-band corresponding to the peak point is taken as the center frequency of the characteristic frequency band.

4. The power equipment status monitoring method based on AI power spectrum according to claim 2, characterized in that, Methods for obtaining bispectral features include: Multiple sampling points of three-phase voltage and current digital waveform data frames are pre-selected, and the mean value of each sampling point is subtracted to obtain a zero-mean sequence. The third-order cumulant of a zero-mean sequence is calculated using a sample estimation method. Performing a fast Fourier transform on the third-order cumulants yields a bispectral output. Sum the amplitudes of all bispectral values ​​to obtain the total bispectral energy of the three-phase voltage and current. The mean amplitude of the bispectral spectrum is calculated, and a screening threshold is set based on the mean amplitude of the bispectral spectrum. The frequency points corresponding to the bispectral amplitudes that are greater than the screening threshold are recorded as peak points. The ratio of the total energy at the peak point to the total energy of the bispectral spectrum is denoted as the peak percentage.

5. The power equipment status monitoring method based on AI power spectrum according to claim 2, characterized in that, Methods for obtaining envelope demodulation features include: Filtering digital waveform data frames based on the center frequency of the characteristic frequency band yields narrowband signals within the characteristic frequency bands of three-phase voltage and current. Perform a Hilbert transform on a narrowband signal to obtain an analytic signal; Calculate the amplitude of the analytic signal to obtain the envelope signal; Perform a fast Fourier transform on the envelope signal to obtain the envelope spectrum; In the envelope spectrum, the maximum amplitude within a set range near the pre-calculated fault characteristic frequency is searched, and the maximum amplitude is recorded as the characteristic frequency peak value of the three-phase voltage and current; the ratio of the characteristic frequency peak value to the average amplitude of the envelope spectrum is recorded as the peak signal-to-noise ratio; the number of peak values ​​in the envelope spectrum with a peak signal-to-noise ratio greater than or equal to the set peak value threshold is counted and recorded as the peak value.

6. The power equipment status monitoring method based on AI power spectrum according to claim 1, characterized in that, The structure of an artificial intelligence model includes a convolutional neural network subnetwork and a fully connected subnetwork, wherein: The convolutional neural network subnetwork consists of an input layer, a first convolutional layer, a second convolutional layer, a first max pooling layer, a second max pooling layer, a global average pooling layer, and a residual connection layer. The connection relationship is as follows: the input layer outputs to the first convolutional layer, the first convolutional layer outputs to the first max pooling layer, the first max pooling layer outputs to the second convolutional layer, and the residual connection layer adds the outputs of the first and second convolutional layers and inputs them to the second max pooling layer. The fully connected subnetwork consists of an input layer, a first fully connected layer, a second fully connected layer, and a batch normalization layer. The connection relationship is as follows: the input layer outputs to the first fully connected layer, the first fully connected layer outputs to the batch normalization layer, and the batch normalization layer outputs to the second fully connected layer.

7. The power equipment status monitoring method based on AI power spectrum according to claim 6, characterized in that, It also includes fusing the output features of convolutional neural network subnetworks and fully connected subnetworks, specifically including a fusion layer, an attention layer, and an output layer.

8. The power equipment status monitoring method based on AI power spectrum according to claim 6, characterized in that, The input to the convolutional neural network subnetwork is a set of time-frequency spectrogram matrices, and the output is a time-frequency feature vector of a set dimension, which serves as a compressed representation of the time-frequency feature set. The input of the fully connected subnetwork is a total feature vector formed by concatenating all one-dimensional features of the frequency domain feature set, nonlinear feature vector, and time domain feature set. The output is a numerical feature vector of a set dimension, which serves as a compressed representation of the numerical features.

9. The power equipment condition monitoring method based on AI power spectrum according to claim 1, characterized in that, Fault type identifiers are 0, 1, 2, and 3, where 1 corresponds to explicit fault, 2 corresponds to camouflaged fault, and 3 corresponds to uncertainty; the fault probability corresponding to the fault type identifier is the confidence level of the fault type identifier output by the artificial intelligence model; the anomaly score is obtained by adding the fault probability of camouflaged fault to the fault probability of explicit fault.

10. A power supply equipment condition monitoring system based on AI power spectrum, used to implement the power supply equipment condition monitoring method based on AI power spectrum as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire the raw analog signals of the power supply equipment in real time and synchronously, and to preprocess the raw analog signals to obtain digital waveform data frames. The time-domain feature module is used to perform time-domain feature calculations on digital waveform data frames to obtain a time-domain feature set. The frequency domain feature module is used to extract spectral features from digital waveform data frames to obtain a set of frequency domain features. The nonlinear module performs high-order statistical analysis and nonlinear signal processing on digital waveform data frames to obtain nonlinear eigenvectors. The fault identification module, based on frequency domain feature set, nonlinear feature vector and time domain feature set, combined with a pre-trained artificial intelligence model, outputs fault type identifier and corresponding fault probability, and calculates an anomaly score based on fault type identifier and corresponding fault probability; The confidence assessment module performs Bayesian confidence assessment based on the fault type identifier, the corresponding fault probability, and the anomaly score to obtain the final equipment status and confidence score.

Images