A power distribution terminal fault monitoring system and method

Abstract

The application relates to the technical field of power system monitoring, in particular to a power distribution terminal fault monitoring system and method. The system comprises a double-channel acquisition module which acquires two analog quantities of a power distribution room and maps the two analog quantities into an environment characteristic matrix of a power grid; a cyclic mapping module which performs decoupling operation to generate a noise separation characteristic vector; a background noise dynamic benchmark is constructed based on the noise separation characteristic vector to generate a decision boundary threshold value; a terminal state data stream is formed by cyclically increasing and resetting a write-in pointer in a fault memory storage unit; a fault determination module which compares the amplitude of the noise separation characteristic vector with the decision boundary threshold value to generate a fault locking instruction; based on the fault locking instruction, a termination signal is generated to freeze the write-in pointer, and a partial discharge data file is extracted from the terminal state data stream. The application constructs a closed-loop monitoring system in which an environment adaptive benchmark and signal characteristic decoupling support each other, and provides a fault backtracking sample which is time-space aligned and logically complete.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system monitoring, in particular to a power distribution terminal fault monitoring system and method. BACKGROUND

[0002] In the field of distributed electrical parameter measurement and fault diagnosis, the multi-channel signal acquisition unit deployed at each key node of the measured line is the core instrument for obtaining the operating characteristics of electrical facilities, and the confidence of its sampling data directly determines the accuracy of subsequent fault location. At present, the embedded electrical monitoring device widely used in such test scenarios is the basic equipment for realizing online signal capture and analysis. Its main function is to record the voltage waveform, current amplitude and contact on-off state of the measured circuit in real time through high-precision analog-to-digital conversion and signal conditioning technology. In order to realize the automatic abnormal screening of a large number of measurement nodes, the existing technology generally adopts a relative measurement strategy based on multi-source feature comparison, that is, selecting "redundant monitoring channels" in the same electrical topology or physical area as a reference, and calculating the Euclidean distance between the target channel and the reference channel in the signal feature space to measure the measurement consistency. When the calculated difference measure exceeds the preset tolerance threshold, it is determined that the target node has an electrical fault or measurement error, which is the mainstream technical means for edge-side abnormal diagnosis in large-scale electrical test systems.

[0003] However, the above-mentioned existing technology faces serious feature decoupling challenges in actual application. First, its associated terminal model has the idealized assumption of environmental homogeneity, ignoring the objective "inherent dispersion" between individual terminal devices, such as inconsistent aging of components, dynamic differences in load characteristics, and asynchronization of operating conditions. This endogenous non-uniformity can cause significant non-fault deviations in the state vectors of associated nodes, causing the linear criterion based on a single Euclidean distance to fail, resulting in high-frequency false positives. Second, the existing technology lacks immunity to complex electromagnetic environments. Frequent switching operations in the power distribution room can produce broadband electromagnetic interference, making the system extremely sensitive to interference signals; in addition, the existing relative diagnosis technology has insufficient robustness in the closed environment of the power distribution room, and cannot accurately define the physical properties of abnormal monitoring data in the presence of multi-source noise, making it difficult for the system to trace faults under complex operating conditions.

[0004] Therefore, a power distribution terminal fault monitoring system and method are proposed. SUMMARY

[0005] The purpose of the present application is to provide a power distribution terminal fault monitoring system and method to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides the following technical solution: a power distribution terminal fault monitoring system, comprising:

[0007] Dual-channel acquisition module: acquire the insulation state analog quantity of power distribution equipment and the space environment analog quantity of power distribution room, and map them into a power grid environment feature matrix containing an induced voltage sequence and a radiation field strength digital sequence through an analog-to-digital converter;

[0008] Cyclic mapping module: decouples the power grid environment feature matrix to generate a noise separation feature vector; based on the sliding average amplitude of the noise separation feature vector within a sliding time window, constructs a background noise dynamic reference, and superimposes a bias voltage to generate a decision boundary threshold; writes the power grid environment feature matrix into a fault memory storage unit, and performs cyclic increment and reset in the fault memory storage unit through a write pointer to form a terminal state data stream;

[0009] Fault determination module: compares the amplitude of the noise separation feature vector with the decision boundary threshold, and generates a fault locking instruction when the amplitude is greater than the decision boundary threshold; based on the fault locking instruction, monitors the energy envelope value of the noise separation feature vector, and generates a termination signal when the energy envelope value continuously falls within the range of the background noise dynamic reference; based on the termination signal, freezes the write pointer, and extracts a partial discharge data file containing the entire process of the fault precursor and evolution from the terminal state data stream.

[0010] Preferably, the specific generation process of the power grid environment feature matrix includes: using a contact sensor to couple the transient pulse on the surface of the metal cabinet of the power distribution terminal to generate a power distribution equipment partial discharge analog quantity; using a non-contact antenna to capture an electromagnetic wave signal to generate a power distribution room space environment analog quantity; inputting the power distribution equipment insulation state analog quantity and the power distribution room space environment analog quantity into a signal conditioning circuit in parallel to output a standardized analog voltage signal, and outputting a digital sampling point through a synchronous analog-to-digital converter; arranging in sequence according to time sequence logic to generate an induced voltage sequence and a radiation field strength digital sequence, respectively; aligning the time stamps and compensating for the time delay between channels to generate time sequence synchronized induced voltage sequence and radiation field strength digital sequence, and combining them according to vector splicing rules to generate the power grid environment feature matrix.

[0011] Preferably, the specific generation process of the noise separation feature vector comprises: receiving a power grid environment feature matrix, calculating the arithmetic mean of each row vector and performing a mean centering process to generate a zero-mean observation matrix; constructing a covariance matrix of the zero-mean observation matrix and performing eigenvalue decomposition to generate a whitening transformation matrix through eigenvalues and eigenvectors; projecting the zero-mean observation matrix to a whitening space based on the whitening transformation matrix to output a whitened data stream; initializing a separation matrix and constructing a fixed point iteration rule using an independent component analysis algorithm; in the calibration stage, performing cyclic iteration on the whitened data stream based on the fixed point iteration rule until the separation matrix converges, and updating the separation matrix coefficients; in the monitoring stage, performing linear projection operation on the whitened data stream using the updated separation matrix to extract the source signal component as the noise separation feature vector.

[0012] Preferably, the specific generation process of the decision boundary threshold comprises: writing the noise separation feature vector into a sliding buffer area of a preset length point by point; during the sliding buffer area updating process, performing an absolute value operation on the resident discrete sampling points to generate a unipolar modulus sequence; performing mean value calculation on the unipolar modulus sequence to obtain an average background noise level; inputting the average background noise level into a digital smoothing filter to suppress random fluctuations through weighted iteration, outputting a trend curve following the change of background noise, which is defined as a background noise dynamic reference; calling a direct current bias voltage constant, superimposing the background noise dynamic reference and the direct current bias voltage constant to generate a level value as the decision boundary threshold.

[0013] Preferably, the specific generation process of the terminal state data stream comprises: taking the original sampling points of the power grid environment feature matrix as the data frames to be stored, activating the direct memory access channel of the fault memory storage unit, positioning to the physical storage address currently pointed by the write pointer, and performing write operation; performing address arithmetic operation to linearly increment the write pointer to point to the next adjacent storage space; performing boundary detection logic in parallel to compare the incremented write pointer with the physical end address of the fault memory storage unit, and if it exceeds the physical end address, triggering a pointer unwinding mechanism to reset the write pointer to the physical first address; when the storage space is saturated, using the new data frame to overwrite the earliest stored historical data frame; through continuous address writing and unwinding operations, a terminal state data stream containing historical state information is constructed.

[0014] Preferably, the specific generation process of the fault locking instruction comprises: taking absolute value operation on the noise separation feature vector by the digital signal processing unit to generate a non-negative instantaneous amplitude sequence; synchronously calling the decision boundary threshold as a dynamic comparison reference; inputting the instantaneous amplitude sequence point by point into the amplitude comparison logic circuit; in the amplitude comparison logic circuit, comparing the instantaneous amplitude at the current time with the decision boundary threshold; when the comparison result shows that the instantaneous amplitude exceeds the decision boundary threshold, triggering the signal debouncing and width checking logic; if the signal continuously exceeds the threshold and meets the preset timing constraint, confirming that an abnormal transient event is captured; driving the fault state flip-flop to flip, generating the fault locking instruction.

[0015] Preferably, the specific generation process of the termination signal comprises: based on the fault locking instruction, activating the envelope monitoring logic channel; performing square operation on the noise separation feature vector to generate a power sequence; performing sliding window peak holding operation on the power sequence to extract an energy envelope value; reading the background noise dynamic reference as a reference bottom line for signal falling back; continuously comparing the energy envelope value with the reference bottom line; when the comparison result shows that the energy envelope value is lower than the reference bottom line, starting the continuous falling back counter; during the counting process, when the energy envelope value is higher than the reference bottom line again, resetting the continuous falling back counter; when the accumulated value of the continuous falling back counter reaches the preset silence confirmation period, determining that the fault transient process ends; generating a high-level pulse as the termination signal.

[0016] Preferably, the specific generation process of the partial discharge data file comprises: receiving the termination signal to stop the increment operation of the write pointer; obtaining the memory address of the current write pointer as the fault end point; according to the preset pre-trigger length, reversely tracing back along the fault memory storage unit to calculate the memory address of the fault start point; continuously reading data from the fault start point to the fault end point; packaging the read data sequence and adding a channel identifier to generate the partial discharge data file.

[0017] A power distribution terminal fault monitoring method, comprising:

[0018] Collecting power distribution equipment insulation state analog quantity and power distribution room space environment analog quantity, mapping into a power grid environment feature matrix containing induced voltage sequence and radiation field strength digital sequence through an analog-to-digital converter;

[0019] Performing decoupling operation on the power grid environment feature matrix to generate a noise separation feature vector; constructing a background noise dynamic reference based on the sliding average amplitude of the noise separation feature vector within a sliding time window, and superimposing a bias voltage to generate a decision boundary threshold; writing the power grid environment feature matrix into a fault memory storage unit, and performing cyclic increment and reset in the fault memory storage unit through a write pointer to form a terminal state data stream;

[0020] The amplitude of the noise separation feature vector is compared with the decision boundary threshold. When the amplitude is greater than the decision boundary threshold, a fault lockout command is generated. Based on the fault lockout command, the energy envelope value of the noise separation feature vector is monitored. When the energy envelope value continues to fall back to the background noise dynamic reference range, a termination signal is generated. Based on the termination signal, the write pointer is frozen, and a partial discharge data file containing fault precursors and the entire evolution process is extracted from the terminal status data stream.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] 1. A dual-channel acquisition module collects simulated quantities of the insulation status of the power distribution equipment and the simulated quantities of the power distribution room's spatial environment in real time. A cyclic mapping module generates a noise separation feature vector, and calculates the moving average amplitude of this vector within a sliding time window. This is used to construct a dynamic background noise benchmark and generate a decision boundary threshold. This allows the fault determination module to establish a trigger threshold that fluctuates with the environmental electromagnetic background, ensuring that the fault locking command is only generated when the amplitude of the noise separation feature vector exceeds the current environmental noise floor due to physical anomalies. This environment-aware triggering mechanism locks in the true moment of the fault occurrence, providing a more accurate time anchor point for locating the write pointer in the fault recall storage unit. This ensures that when the system performs fault backtracking, the extracted partial discharge data file strictly corresponds to the actual insulation fault event, establishing an effective starting point for data backtracking.

[0023] 2. By performing decoupling operations on the power grid environment feature matrix, which includes simulated quantities of the insulation state of power distribution equipment and simulated quantities of the spatial environment of the power distribution room, a noise separation feature vector free from environmental coupling interference is generated through the cyclic mapping module. The system uses this feature vector for fault logic judgment and writes the power grid environment feature matrix into the fault recall storage unit to form a terminal status data stream. This means that the fault locking criterion comes from data that has already undergone physical attribute definition and purification. When the system performs fault backtracking, although the partial discharge data file extracted from the data stream contains the original environmental feature matrix information, its truncated time period is locked based on independent source signals with high signal-to-noise ratio. This mechanism ensures that the backtracked data can completely reproduce the physical scenario of the insulation state of the power distribution terminal, enabling maintenance personnel to identify detailed features based on the fault time window during the backtracking process, without causing missed or misjudgments due to environmental noise interference.

[0024] 3. Upon detecting a fault, the energy envelope value of the noise separation feature vector is compared with the background noise dynamic benchmark. Only when the energy envelope value completely falls back to the benchmark range does the system determine the end of the physical process, generate a termination signal, and freeze the write pointer. Utilizing the energy dissipation characteristics of the decoupled signal itself as the recording cutoff condition ensures that the partial discharge data file extracted from the terminal state data stream completely encompasses the entire evolution process from fault precursors, through fault outbreak, to complete energy dissipation. This guarantees the continuity and integrity of the retrospective data in the time domain, providing a more complete chain of evidence for analyzing the fault's occurrence mechanism and evolution trend.

[0025] 4. By deeply coupling the dual-channel acquisition module, the cyclic mapping module, and the fault determination module, a closed-loop monitoring system is constructed, in which environmental adaptive benchmarks and signal feature decoupling support each other, ensuring the reliability of fault backtracking from the overall system perspective. The background noise dynamic benchmark plays a dual constraint role, serving both as the decision boundary threshold determining the triggering point of the fault locking command and as the reference bottom line for the fallback of the energy envelope value of the noise separation feature vector, determining the generation of the termination signal and the freezing point of the write pointer. At the same time, the power grid environmental feature matrix is ​​the stored subject, and the noise separation feature vector is the monitored subject, ensuring that the signal determination reflects the insulation status of the equipment, rather than environmental noise. This logical coupling allows the recording window of the terminal status data stream to adaptively expand and contract with the fluctuations of the environmental electromagnetic background, and the triggering basis is high-fidelity physical characteristics. The final generated partial discharge data file achieves complete coverage of the entire process of fault precursors and evolution in the time domain and eliminates environmental interference in the logical domain, thus providing maintenance personnel with a spatiotemporally aligned and logically complete fault backtracking sample. Attached Figure Description

[0026] Figure 1 This is a structural diagram of a power distribution terminal fault monitoring system proposed in an embodiment of this invention application;

[0027] Figure 2 This is a flowchart of the digital signal processing logic proposed in an embodiment of this invention application;

[0028] Figure 3 This is a flowchart of a power distribution terminal fault monitoring method proposed in an embodiment of this invention. Detailed Implementation

[0029] 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, and 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] Please see Figures 1 to 3 The present invention provides a power distribution terminal fault monitoring system, the specific modules of which are as follows:

[0031] Dual-channel acquisition module: Acquires analog quantities of insulation status of power distribution equipment and analog quantities of spatial environment of power distribution room, and maps them into a power grid environment feature matrix containing digital sequences of induced voltage and radiation field strength through analog-to-digital converter;

[0032] The cyclic mapping module performs decoupling operations on the power grid environment feature matrix to generate a noise separation feature vector; based on the sliding average amplitude of the noise separation feature vector within the sliding time window, it constructs a dynamic benchmark for background noise and superimposes a bias voltage to generate a decision boundary threshold; it writes the power grid environment feature matrix into the fault recall storage unit, and uses a write pointer to cyclically increment and reset within the fault recall storage unit to form a terminal status data stream.

[0033] Fault determination module: Compares the amplitude of the noise separation feature vector with the decision boundary threshold. When the amplitude is greater than the decision boundary threshold, a fault locking command is generated. Based on the fault locking command, the energy envelope value of the noise separation feature vector is monitored. When the energy envelope value continues to fall back to the background noise dynamic reference range, a termination signal is generated. Based on the termination signal, the write pointer is frozen, and a partial discharge data file containing fault precursors and the entire evolution process is extracted from the terminal status data stream.

[0034] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0035] Example 1

[0036] This application discloses a power distribution terminal fault monitoring system, see below. Figure 1The specific modules proposed in this application include: a dual-channel acquisition module: acquiring analog quantities of the insulation status of power distribution equipment and analog quantities of the spatial environment of the power distribution room, and mapping them into a power grid environment feature matrix containing induced voltage sequences and radiated field strength digital sequences through an analog-to-digital converter; a cyclic mapping module: performing decoupling operations on the power grid environment feature matrix to generate a noise separation feature vector; constructing a background noise dynamic benchmark based on the sliding average amplitude of the noise separation feature vector within a sliding time window, and superimposing a bias voltage to generate a decision boundary threshold; writing the power grid environment feature matrix into a fault recall storage unit, and cyclically incrementing and resetting the fault recall storage unit through a write pointer to form a terminal status data stream; a fault determination module: comparing the amplitude of the noise separation feature vector with the decision boundary threshold, and generating a fault locking command when the amplitude is greater than the decision boundary threshold; monitoring the energy envelope value of the noise separation feature vector based on the fault locking command, and generating a termination signal when the energy envelope value continuously falls back to the range of the background noise dynamic benchmark; and freezing the write pointer based on the termination signal to extract a partial discharge data file containing fault precursors and the entire evolution process from the terminal status data stream.

[0037] Furthermore, the analog quantities of the insulation status of the power distribution equipment and the analog quantities of the spatial environment of the power distribution room are collected and mapped into a power grid environment feature matrix containing induced voltage sequences and radiated field strength digital sequences through an analog-to-digital converter; this corresponds to a dual-channel acquisition module; the specific implementation process includes:

[0038] A contact sensor is used to couple transient pulses onto the surface of the metal cabinet of the power distribution terminal to generate a simulated partial discharge quantity of the power distribution equipment; a non-contact antenna is used to capture electromagnetic wave signals to generate a simulated quantity of the power distribution room's spatial environment; the simulated quantity of the power distribution equipment's insulation state and the simulated quantity of the power distribution room's spatial environment are input in parallel to a signal conditioning circuit, which outputs a standardized analog voltage signal, which is then converted into digital sampling points through a synchronous analog-to-digital converter; these are arranged sequentially according to timing logic to generate induced voltage sequences and radiated field strength digital sequences, respectively; through timestamp alignment and inter-channel delay compensation, time-synchronized induced voltage sequences and radiated field strength digital sequences are generated, and then combined according to vector splicing rules to generate a power grid environment feature matrix.

[0039] Specifically, the method involves using a contact sensor to couple transient pulses onto the surface of the metal cabinet of the power distribution terminal to generate a simulated partial discharge quantity for the power distribution equipment; and using a non-contact antenna to capture electromagnetic wave signals to generate a simulated quantity for the spatial environment of the power distribution room. In this embodiment, the acquisition of the simulated partial discharge quantity for the power distribution equipment relies on a transient ground voltage sensor installed on the surface of the metal casing of the switchgear. This sensor adopts the principle of capacitive coupling, with an equivalent coupling capacitance value designed to be 50pF to 100pF, and its detection frequency band is locked to 3MHz to 100MHz by a hardware filter. Simultaneously, the acquisition of the simulated spatial environment of the power distribution room uses a wideband ultra-high frequency antenna, with an operating frequency band covering 300MHz to 1.5GHz, and an average effective height of 8.5mm. This parameter design ensures that the antenna can sensitively capture mobile communication signals and background white noise in the space, thereby generating a simulated quantity of electromagnetic radiation in the spatial environment of the power distribution room.

[0040] Specifically, the analog quantities of the insulation state of the power distribution equipment and the electromagnetic radiation of the power distribution room are input in parallel to the signal conditioning circuit, which outputs a standardized analog voltage signal. This signal is then converted into digital sampling points by a synchronous analog-to-digital converter (ADC). These digital sampling points are arranged sequentially according to timing logic to generate induced voltage and radiated field strength digital sequences, respectively. In this embodiment, the two analog signals undergo parallel signal conditioning before digital processing. This circuit first eliminates signal reflections through a 50Ω impedance matching network, then passes the signal into a programmable gain amplifier (PGA), which linearly maps the original transient ground voltage and electromagnetic radiation of the space environment to a standard analog voltage range of ±2V. Next, the analog signal is input to a broadband logarithmic detector circuit. This circuit extracts the voltage envelope characteristics of high-frequency transient pulses, converting nanosecond-level oscillation signals into microsecond-level unipolar envelope signals, limiting the output bandwidth to within 15MHz. The system uses a dual-channel synchronous sampling pipelined ADC to perform quantization, with a sampling rate set to 40MS / s and a quantization precision of 14 bits. Driven by the same crystal oscillator clock source, the two signals are kept in strict phase synchronization, and finally output digital sequences of induced voltage and radiation field strength containing continuous discrete values ​​respectively.

[0041] In this embodiment, an automatic gain control mechanism can be introduced after the output digital sampling points. The system runs a first-order saturation detection logic in parallel to monitor whether the digital sampling points output by the analog-to-digital converter are close to the clipping distortion region and the quantization noise region. When the sampling points touch the clipping distortion region for three consecutive clock cycles, a gain reduction command is sent to the programmable gain amplifier; conversely, if the sampling points are below the preset quantization lower limit threshold for two consecutive power frequency cycles, they are in the quantization noise region, and a gain command is sent. The quantization noise region refers to the area where the signal amplitude is close to 1-2 least significant bits, and the quantization lower limit threshold can be set to 100mV. This closed-loop adaptive gain adjustment mechanism ensures that the analog-to-digital converter always operates in the optimal linear region, avoiding signal truncation due to strong interference or weak signals being overwhelmed by quantization noise, thereby improving the dynamic range of the subsequent feature matrix.

[0042] Specifically, the system generates a time-synchronized induced voltage sequence and a digital sequence of radiated field strength through timestamp alignment and inter-channel delay compensation, and combines them according to vector splicing rules to generate a power grid environment feature matrix. In this embodiment, the system needs to correct for time errors introduced by the physical transmission path during the construction of the final feature matrix. In this implementation, the transient ground voltage sensor and the UHF antenna are connected via coaxial cable, resulting in a transmission delay deviation of approximately 15 ns between the two signals. At a sampling rate of 40 MS / s, the time interval between each sampling point is 25 ns. Since the physical delay (15 ns) is less than one sampling period, and the rising edge of the envelope signal after detection is relatively slow, the system performs delay calibration during the initialization phase, measuring the fixed physical delay (e.g., 15 ns) introduced by the difference in coaxial cable length and converting it into a fractional delay parameter. During the continuous monitoring process of generating the power grid environment feature matrix, the system employs a fractional time delay filter algorithm (e.g., using Lagrange interpolation with a Farrow structure) to continuously compensate for the time shift of the digital sequence in the lagging channel based on the preset fixed fractional time delay parameters, thereby achieving physical alignment of timestamps. Finally, the logic unit uses the aligned induced voltage digital sequence as the first row vector and the radiation field strength digital sequence as the second row vector, constructing a power grid environment feature matrix with a dimension of 2×1024 (assuming a sliding window length N=1024) according to the matrix dimension splicing rules.

[0043] By integrating contact sensors and non-contact antennas, the system can simultaneously capture transient pulses from the surface of the metal cabinet and electromagnetic wave signals from the surrounding space, reducing the information blind spots that may exist with single monitoring methods. In the signal processing link, the system uses timestamp alignment and inter-channel delay compensation techniques to rigorously calibrate the timing of the parallel-input induced voltage sequence and the digital sequence of radiated field strength. This calibration mechanism effectively reduces phase deviations caused by differences in sensor response and transmission paths, improving the synchronization accuracy of each element in the power grid environment characteristic matrix in the time dimension. The resulting characteristic matrix can realistically reproduce the transient electromagnetic field distribution at the time of a fault.

[0044] Furthermore, decoupling operations are performed on the power grid environment feature matrix to generate noise separation feature vectors; corresponding to the cyclic mapping module; see [link / reference]. Figure 2 The specific implementation process includes:

[0045] The system receives the power grid environment feature matrix, calculates the arithmetic mean of each row vector, and performs mean-neutralization to generate a zero-mean observation matrix. It then constructs the covariance matrix of the zero-mean observation matrix and performs eigenvalue decomposition, generating a whitening transformation matrix using eigenvalues ​​and eigenvectors. Based on the whitening transformation matrix, it projects the zero-mean observation matrix onto the whitening space, outputting a whitened data stream. The system initializes the separation matrix and constructs a fixed-point iteration rule using independent component analysis. During the calibration phase, it iterates on the whitened data stream based on the fixed-point iteration rule until the separation matrix converges, updating the separation matrix coefficients. Finally, during the monitoring phase, it performs a linear projection operation on the whitened data stream using the updated separation matrix, extracting the source signal components as noise separation feature vectors.

[0046] Specifically, for the received power grid environment feature matrix, the arithmetic mean of each row vector is calculated and mean-neutralization is performed to generate a zero-mean observation matrix; the covariance matrix of the zero-mean observation matrix is ​​constructed and eigenvalue decomposition is performed, generating a whitening transformation matrix through eigenvalues ​​and eigenvectors; based on the whitening transformation matrix, the zero-mean observation matrix is ​​projected onto the whitening space, and a whitened data stream is output. In this embodiment, the system first receives a power grid environment feature matrix with a dimension of 2 rows and 1024 columns. To eliminate the convolution mixing effect introduced by the reflection of the metal cabinet of the substation and the difference in frequency response of different sensor transmission channels, the system sets the delay step size d (e.g., 1 sampling point) and the embedding dimension m (e.g., 5), and stacks the original induced voltage sequence and the radiation field strength sequence and their respective time delay sequences as row vectors to form a new 2 (m+1) row (e.g., 12 row) multidimensional augmented observation matrix. First, the arithmetic mean of each row of the matrix is ​​calculated. For example, if the DC component offset of the transient ground voltage channel is 20mV, the system will perform a subtraction operation to remove the DC offset, ensuring that the statistical center of the data is at zero, thus generating a zero-mean observation matrix. Subsequently, to eliminate the second-order statistical correlation between the two signals, fuse multipath energy, and normalize the variance, the system constructs a multidimensional (e.g., 12×12) covariance matrix. For example, due to external interference sources and multipath effects, the covariance matrix reflects not only the correlation between the two physical channels (e.g., 0.6) but also the autocorrelation between each time delay component. The system performs eigenvalue decomposition on this covariance matrix to obtain multiple sets of eigenvectors and their corresponding eigenvalues, and calculates the kurtosis of each principal component. Components with larger absolute kurtosis values ​​(representing signals with significant pulse characteristics) and components with larger eigenvalues ​​(representing major interference sources) are preferentially retained, while small eigenvalue components corresponding to secondary multipath reflections and thermal noise are discarded. Based on this, a whitening transformation matrix is ​​constructed, which is equal to the reciprocal of the square root of the diagonal matrix retaining eigenvalues ​​multiplied by the transpose of the matrix retaining eigenvectors. This whitening transformation matrix is ​​then multiplied left by the zero-mean observation matrix to project the original data onto the whitening space, outputting a whitened data stream. At this point, the covariance matrix of the whitened data stream becomes the identity matrix I, meaning that the components are uncorrelated and have a variance of 1.

[0047] Specifically, for initializing the separation matrix, a fixed-point iteration rule is constructed using the Independent Component Analysis (ICA) algorithm. During the calibration phase, based on the fixed-point iteration rule, iterative iterations are performed on the whitened data stream until the separation matrix converges, and the separation matrix coefficients are updated. In this embodiment, both the induced voltage sequence and the radiated field strength digital sequence are voltage envelope signals processed by a broadband logarithmic detector circuit. At the time-domain scale of the detected envelope, the difference in response of different sensors to the same discharge source is mainly reflected in the linear attenuation of amplitude, and the phase distortion caused by the transmission delay is negligible relative to the envelope wavelength. This embodiment utilizes the sparsity of the envelope signal in the time domain to satisfy the mathematical premise of using the ICA algorithm to separate the main fault features under underdetermined conditions.

[0048] The system employs the FastICA algorithm. First, during the calibration phase (e.g., the first 500ms after system startup or when a drastic change in the environmental background is detected), the data buffer logic is activated. A static data frame of a preset length (e.g., 1024 to 2048 sampling points) is extracted and locked from the whitened data stream as a training sample set. Based on this static data frame, the system first calculates the arithmetic mean of each channel to obtain a global mean vector, and then constructs a covariance matrix for eigenvalue decomposition to generate a global whitening transformation matrix. Using this global mean vector and global whitening transformation matrix, the static data frame undergoes mean removal and whitening preprocessing to generate a whitened data stream. At this point, the system stores the global mean vector and global whitening transformation matrix as a preprocessing benchmark for subsequent monitoring phases.

[0049] Subsequently, a random 2×2 initial separation matrix is ​​generated, and the tanh function (hyperbolic tangent function) is selected as the nonlinear comparison function. The fixed-point iteration rule is set to Newton's iteration method. The specific calculation steps are as follows: calculate the inner product of the row vectors of the separation matrix and the whitened data stream; apply the tanh function to the inner product result and calculate the mean again with the whitened data stream; subtract the product of the mean of the inner product result with the tanh derivative function and the current row vector, and use this as the new weight vector. To prevent different weight vectors from converging to the same source signal, the system performs a symmetric orthogonalization operation after each iteration, that is, normalizes the separation matrix using the inverse operation of the square root of the matrix. Based on the fixed-point iteration rule, the above iteration is repeatedly performed on the whitened data stream, and the absolute value of the dot product of the row vectors of the separation matrix before and after is calculated. When the difference between this absolute value and 1 is less than the preset convergence threshold (e.g., 10), the weight vector is selected. -5 When the algorithm converges, the coefficients of the finally converged separation matrix are locked together with the previously fixed global mean vector and global whitening transformation matrix, and all transformation parameters are updated.

[0050] Specifically, during the monitoring phase, the updated separation matrix is ​​used to perform a linear projection operation on the whitened data stream to extract the source signal components as noise separation feature vectors. In this embodiment, once the separation matrix is ​​calibrated and locked, the system enters the monitoring phase. At this time, the system receives the real-time input power grid environment feature matrix and directly calls the global mean vector and global whitening transformation matrix fixed during the calibration phase to perform a linear transformation, generating a whitened data stream. Subsequently, the updated 2×2 separation matrix is ​​used to perform matrix multiplication (i.e., linear projection operation) on the real-time input whitened data stream. This operation decouples the mixed observation signal into two statistically independent source signal components. To automatically identify which component is the desired partial discharge signal, the system calculates the kurtosis (i.e., fourth-order central moment) of the two source signal components. Because the partial discharge signal has extremely strong pulse characteristics, its probability density distribution exhibits a long-tailed super-Gaussian distribution, with a kurtosis value usually much greater than 3; while background noise (such as white noise or stationary interference) usually follows a Gaussian or sub-Gaussian distribution, with a kurtosis value close to 3 or smaller. The system compares the kurtosis of two components and automatically selects the component with the larger kurtosis value as the noise separation feature vector. For example, in cable terminal defect monitoring, the signal-to-noise ratio (SNR) of the original input signal is only -5dB. After the above linear projection and feature extraction, the SNR of the partial discharge pulse in the output noise separation feature vector is improved to more than 15dB because the component related to spatial environmental noise has been removed.

[0051] By performing mean-reduction and whitening transformations on the power grid environmental feature matrix, second-order correlations among observed data are eliminated, reducing data redundancy. Based on this, a fixed-point iterative rule is constructed using independent component analysis (ICM) algorithms, enabling the system to dynamically update the separation matrix through iterative calculations even with unknown mixing parameters. This adaptive linear projection operation can extract independent source signal components from the mixed whitened data stream, reducing the masking effect of random background noise and periodic interference on target fault signals. By generating high-purity noise separation feature vectors, the signal-to-noise ratio is improved, thereby enhancing the ability to capture weak signals of latent insulation defects.

[0052] Furthermore, based on the moving average amplitude of the noise separation feature vector within the sliding time window, a dynamic benchmark for background noise is constructed, and a bias voltage is superimposed to generate a decision boundary threshold; this corresponds to the cyclic mapping module; see [link to relevant documentation]. Figure 2 The specific implementation process includes:

[0053] The noise separation feature vector is written point by point into a sliding buffer of a preset length. During the sliding buffer update process, the absolute value operation is performed on the stationed discrete sampling points to generate a unipolar modulus sequence. The mean value is calculated on the unipolar modulus sequence to obtain the average background noise level. The average background noise level is input into a digital smoothing filter, and random fluctuations are suppressed through weighted iteration. The output is a trend curve that smoothly follows the changes in background noise, which is defined as the background noise dynamic reference. The DC bias voltage constant is retrieved, and the background noise dynamic reference is superimposed with the DC bias voltage constant to generate a level value, which is used as the decision boundary threshold.

[0054] Specifically, the noise separation feature vector is written point by point into a sliding buffer of a preset length. During the sliding buffer update process, absolute value operations are performed on the resident discrete sampling points to generate a unipolar modulus sequence. The mean of the unipolar modulus sequence is calculated to obtain the average background noise level. In this embodiment, the system uses a first-in-first-out (FIFO) ring-shaped storage space as the sliding buffer, with a preset length of 1024 to 2048 sampling points, corresponding to a duration of 25.6 μs to 51.2 μs at a sampling rate of 40 MS / s. As new feature vector data s(n) is written point by point in a pipeline manner, the oldest data is removed, and the buffer always maintains the latest environmental state snapshot. In each clock cycle, the digital logic unit performs absolute value operations on all resident discrete sampling points in the buffer in parallel or pipelined manner, converting the bipolar signal that originally fluctuated in the positive and negative voltage range into a non-negative unipolar modulus sequence |s(n)|. Subsequently, the arithmetic logic unit calculates the arithmetic mean of the sequence. For example, if the noise amplitude in the current buffer fluctuates randomly between -5mV and +5mV, after absolute value processing and mean calculation, the system will output an average background noise level of approximately 3.2mV, which quantifies the "silent" background noise level of the electromagnetic environment at the current moment.

[0055] Specifically, for the digital smoothing filter that inputs the average background noise level, random fluctuations are suppressed through weighted iteration, and the output curve that smoothly follows the trend of background noise changes is defined as the background noise dynamic reference. In this embodiment, to prevent the noise reference from drastically changing due to individual sudden non-fault spikes (such as electromagnetic pulses during switching operations), which could lead to erroneous increases or decreases in the threshold, the system introduces an exponentially weighted moving average (EWMA) filter. This filter adopts a first-order recursive structure, and its core parameter, the smoothing factor α, is set between 0.01 and 0.05. The specific operation logic is as follows: the current background noise dynamic reference is equal to the reference value of the previous moment multiplied by (1-α), plus the average background noise level calculated at the current moment multiplied by α. For example, setting α=0.02, when the input average noise level suddenly changes from 3mV to 10mV, the filter output will not immediately follow the change, but will rise slowly over hundreds of clock cycles. In actual operation, this curve represents the "zero energy" reference line of the equipment when no partial discharge occurs. For example, when the load is low and there is less interference at night, the reference is maintained at about 2mV, while during the peak industrial load period during the day, the reference automatically drifts to 5mV.

[0056] Specifically, for retrieving the DC bias voltage constant, the background noise dynamic reference is superimposed on the DC bias voltage constant to generate a level value, which serves as the decision boundary threshold. In this embodiment, the system retrieves a pre-set DC bias voltage constant. The value of this constant depends on the inherent thermal noise level of the signal conditioning circuit and the sum of quantization errors, and is typically set to 1.2 to 1.5 times the measured peak-to-peak noise floor of the system under no-input conditions (e.g., 15mV to 50mV). This voltage value acts as a "safety margin" to cover quantization noise, thermal noise peaks, and minor fluctuations not completely filtered out by the filter in the signal link. The logic unit performs algebraic addition on the real-time calculated background noise dynamic reference (e.g., 5mV) and the DC bias voltage constant (e.g., 20mV) to generate the final decision boundary threshold of 25mV. This threshold is not fixed but "floats" up and down with the fluctuation of the background noise reference. If increased ambient noise causes the reference voltage to rise to 8mV, the threshold will automatically adjust to 28mV to prevent noise from drowning out the signal and triggering false alarms. Conversely, if the environment becomes quiet and the reference voltage drops, the threshold will also decrease accordingly, ensuring that the system can detect weak partial discharge signals. For example, this mechanism successfully maintained stable detection of a 50pC small signal without false triggering even when the noise reference fluctuated by ±30%.

[0057] After generating the decision boundary threshold, this embodiment introduces a temperature and humidity multi-factor correction factor. The system reads the real-time ambient temperature T and relative humidity H in the power distribution room through an auxiliary interface and corrects the air ionization threshold model based on Paschen's law. When the relative humidity H exceeds 85%, the system recognizes that the air insulation strength is reduced, making non-faulty corona discharge more likely. Therefore, it automatically adds a humidity compensation coefficient to the decision boundary threshold, thereby dynamically raising the decision boundary threshold. For example, the humidity compensation coefficient is increased by 15%. Partial discharge in the power distribution room is easily affected by humidity. By introducing a physical model to correct the threshold, the false alarm rate in humid weather can be reduced, increasing environmental robustness.

[0058] By writing the noise separation feature vector point by point into a sliding buffer, calculating the average background noise level using a unipolar modulus sequence, and suppressing random fluctuations with a digital smoothing filter, a smooth dynamic background noise benchmark is generated. This allows the decision benchmark to follow the fluctuations of the ambient noise level in real time, reducing the possibility of the fixed threshold failing when the ambient noise changes abruptly. Simultaneously, by superimposing a DC bias voltage constant to generate the final decision boundary threshold, the system establishes a defined safety margin on top of the dynamic benchmark, effectively reducing false triggering caused by signal glitches or transient disturbances. This threshold generation mechanism based on real-time data statistics improves the adaptability and stability of the monitoring system under different operating conditions.

[0059] Furthermore, the power grid environment feature matrix is ​​written into the fault recall storage unit, and the write pointer is used to cyclically increment and reset within the fault recall storage unit to form a terminal status data stream; corresponding to the cyclic mapping module; the specific implementation process includes:

[0060] The original sampling points of the power grid environment feature matrix are used as the data frames to be stored. The direct memory access channel of the fault memory storage unit is activated, and the physical storage address currently pointed to by the write pointer is located to perform a write operation. Address arithmetic operations are performed to linearly increment the write pointer, pointing to the next adjacent storage space. Boundary detection logic is executed in parallel to compare the incremented write pointer with the physical end address of the fault memory storage unit. If it exceeds the physical end address, the pointer rollback mechanism is triggered to reset the write pointer to the physical start address. When the storage space is saturated, the earliest stored historical data frame is overwritten with a new data frame. Through continuous address writing and rollback operations, a terminal status data stream containing historical status information is constructed.

[0061] Specifically, for the original sampling points of the power grid environment feature matrix as the data frame to be stored, the direct memory access channel of the fault recall storage unit is activated, the physical storage address currently pointed to by the write pointer is located, and a write operation is performed. In this embodiment, after the power grid environment feature matrix is ​​spliced ​​and encapsulated in the logic unit, the 1024 original sampling points it contains are packaged into a group of data frames to be stored. The direct memory access controller reads the 30-bit wide physical storage address (e.g., hexadecimal address 0x1000_0000) stored in the register where the current write pointer is located, occupies the bus through burst transfer mode, and writes the data frame into the physical sector corresponding to the fault recall storage unit. This ensures that the data acquisition and storage process maintains strict pipeline synchronization on a nanosecond-level time scale, without frame loss or blocking.

[0062] Specifically, the write pointer is linearly incremented using address arithmetic to point to the next adjacent memory space. Concurrently, boundary detection logic compares the incremented write pointer with the physical end address of the fault recall memory unit. If the physical end address is exceeded, a pointer rollback mechanism is triggered, resetting the write pointer to the physical start address. In this embodiment, to achieve seamless recording of fault precursor information, after each data frame write operation, the storage control logic unit immediately performs address arithmetic, adding the current write pointer value to the byte length of the data frame (e.g., for a dual-channel, two-byte-per-point, 1024-point matrix, the step size is 4096 bytes) to generate a new write pointer that precisely points to the next adjacent and empty memory space. Simultaneously, the hardware comparator executes boundary detection logic in parallel, comparing the updated write pointer with the preset physical end address of the fault recall memory unit. When the incremented write pointer value is detected to be greater than or equal to the physical end address, it is determined that the memory space has reached its physical boundary, and the system triggers the pointer rollback mechanism. By forcibly assigning the write pointer to the physical starting address (0x1000_0000), the linearly arranged physical memory is logically connected end to end, forming a closed-loop circular buffer.

[0063] After pointing to the next adjacent storage space, this embodiment can introduce lightweight indexing and pre-marking techniques. While the write pointer increments, the system maintains an independent time-sector index table. Whenever the write pointer crosses a physical sector of a storage unit, the starting physical address and corresponding timestamp of the current sector are recorded in the time-sector index table, and a suspected pulse exceeding a threshold is marked within that sector. The suspected pulse refers to data whose amplitude does not trigger fault locking but is higher than 50% of the background baseline. Through lightweight indexing and pre-marking techniques, during subsequent file extraction, it is not necessary to traverse the entire memory; key data segments can be quickly located simply by consulting the time-sector index table. Furthermore, it can retain suspected data that "does not trigger alarms but has research value," enhancing the system's backtracking retrieval efficiency.

[0064] Specifically, when the storage space is saturated, the earliest stored historical data frame is overwritten using a new data frame. Through continuous address write and rollback operations, a terminal status data stream containing historical state information is constructed. In this embodiment, as the monitoring process continues, the fault recall storage unit enters a saturated state after a complete pointer traversal. At this time, the system does not stop recording, but strictly follows the "new in, old out" time sequence principle, using the newly generated power grid environment feature matrix data frame to directly overwrite the earliest timestamp historical data frame stored at the physical starting address. This cyclic overwrite operation continues in the storage unit at millisecond intervals, ensuring that the storage unit always retains a complete waveform data of a fixed duration closest to the current moment. Through this continuous address write, increment, and rollback operation, the static data block in physical memory is dynamically transformed into a terminal status data stream that flows and updates over time. This data stream not only contains the current power grid environment characteristics, but also completely preserves the voltage transient evolution process within several power frequency cycles before the fault occurred.

[0065] By utilizing direct memory access channels and address arithmetic operations, sampled data frames can be written to physical storage space, reducing CPU resource consumption during data transfer. Combined with a pointer rollback mechanism, when storage space is saturated, new data automatically overwrites old data, eliminating the need for frequent memory allocation operations and reducing system overhead and the risk of memory fragmentation. This circular storage method ensures that the fault recall storage unit always retains a complete historical state data stream immediately adjacent to the current moment. At the instant a fault is triggered, the system can directly use the stored data stream for tracing, guaranteeing real-time data availability.

[0066] Furthermore, the amplitude of the noise separation feature vector is compared with the decision boundary threshold. When the amplitude is greater than the decision boundary threshold, a fault locking command is generated; this corresponds to the fault determination module. The specific implementation process includes:

[0067] The digital signal processing unit performs absolute value operations on the noise separation feature vector to generate a non-negative instantaneous amplitude sequence; a decision boundary threshold is synchronously retrieved as a dynamic comparison benchmark; the instantaneous amplitude sequence is input point by point to the amplitude comparison logic circuit; in the amplitude comparison logic circuit, the instantaneous amplitude at the current moment is logically compared with the decision boundary threshold; when the comparison result shows that the instantaneous amplitude exceeds the decision boundary threshold, signal dejittering and width verification logic is triggered; if the signal continues to exceed the threshold and meets the preset timing constraints, an abnormal transient event is confirmed to have been captured; the fault state trigger is driven to flip, generating a fault lockout command.

[0068] Specifically, the system uses a digital signal processing unit to perform absolute value operations on the noise separation feature vector to generate a non-negative instantaneous amplitude sequence; it synchronously retrieves a decision boundary threshold as a dynamic comparison benchmark; and inputs the instantaneous amplitude sequence point by point to the amplitude comparison logic circuit. In the amplitude comparison logic circuit, the instantaneous amplitude at the current moment is logically compared with the decision boundary threshold. In this embodiment, the system performs real-time absolute value transformation operations on the noise separation feature vector and maps all negative polarity sampling points to positive polarity intervals using a two's complement inversion algorithm to generate a non-negative instantaneous amplitude sequence. Simultaneously, the system retrieves the decision boundary threshold, which is a dynamic value that fluctuates with the ambient noise floor, for example, 35.5mV at a certain moment. Driven by each clock cycle, the amplitude comparison logic circuit receives the current instantaneous amplitude sampling point and the real-time decision boundary threshold in parallel. The comparator performs a numerical magnitude judgment. If the amplitude of the current sampling point (e.g., 120mV) is greater than the decision boundary threshold, the comparator outputs a logic high level "1"; otherwise, if the amplitude (e.g., 10mV) is less than or equal to the threshold, it outputs a logic low level "0".

[0069] Specifically, when the comparison result indicates that the instantaneous amplitude exceeds the decision boundary threshold, signal dejittering and width verification logic is triggered. If the signal continues to exceed the threshold and meets the preset timing constraints, an abnormal transient event is confirmed to have been captured. In this embodiment, the system cascades a signal dejittering and width verification logic module after the amplitude comparator. The core of this module is a high-speed counter based on clock counting. Its preset timing constraint parameters are set to 5 to 10 sampling points, corresponding to a duration of 125 ns to 250 ns, based on the physical characteristics of the partial discharge envelope signal after detection (usually lasting more than 1 μs). When the amplitude comparator outputs a logic high level, the counter starts and begins to accumulate; once the comparator output jumps to a low level, the counter is immediately reset to zero. Only when the accumulated value of the counter continuously reaches the preset minimum effective pulse width threshold (e.g., 5) does the logic circuit determine that the currently captured signal is a valid abnormal transient event. Taking a certain test as an example, there was a fast transient interference with an amplitude of up to 80mV but a duration of only 8ns (2 sampling points). Although its amplitude far exceeded the threshold of 35.5mV, it was judged as interference and filtered out by the verification logic because it did not meet the 20ns width constraint. The following pulse with an amplitude of 60mV and a duration of 150ns was successfully identified as a partial discharge signal because the duration of its continuous exceedance of the threshold exceeded the preset constraint.

[0070] Specifically, regarding the toggling of the fault state trigger to generate a fault lockout command, in this embodiment, once the width verification logic confirms the validity of the abnormal transient event, the system outputs an acknowledgment pulse signal to the fault state trigger (typically a state machine constructed using D flip-flops). The trigger toggles its state instantly upon receiving the rising edge of the acknowledgment pulse, locking the system's internal state flag from "monitoring scan state" to "fault lockout state." This state toggle is irreversible (until the system is reset or data retrieval is complete), directly generating a high-priority fault lockout command. This command is instantly broadcast to all submodules of the system via the internal bus, and its logic level transition (e.g., from 0V to 3.3V) serves as a global trigger signal for subsequent freezing of the storage pointer, initiation of energy envelope analysis, and triggering of alarm indicators. For example, after detecting the aforementioned 150ns valid pulse, the trigger toggles within 4ns, generating the fault lockout command.

[0071] The generation of fault lockout commands not only relies on the instantaneous amplitude of the noise separation feature vector exceeding the decision boundary threshold, but also introduces signal dejittering and width verification logic. This logic requires that the over-limit signal must continuously meet preset timing constraints, thereby effectively filtering out nanosecond-level narrow pulse interference generated by lightning induction or switching operations. This design reduces the probability of non-faulty transient spikes causing system malfunctions. By driving the fault state trigger to flip only after meeting strict timing conditions, the system improves the accuracy of judging real power distribution fault events, ensuring that each fault lockout corresponds to an abnormal physical process with substantial energy characteristics.

[0072] Furthermore, based on the fault locking command, the energy envelope value of the noise separation feature vector is monitored. When the energy envelope value continuously falls back to the background noise dynamic reference range, a termination signal is generated; this corresponds to the fault determination module. The specific implementation process includes:

[0073] Based on the fault lockout command, the envelope monitoring logic channel is activated; the noise separation feature vector is squared to generate a power sequence; a sliding window peak hold operation is performed on the power sequence to extract the energy envelope value; the background noise dynamic benchmark is read as the reference baseline for signal fallback; the energy envelope value is continuously compared with the reference baseline; when the comparison result shows that the energy envelope value is lower than the reference baseline, a continuous fallback counter is started; during the counting process, when the energy envelope value is higher than the reference baseline again, the continuous fallback counter is reset; when the accumulated value of the continuous fallback counter reaches the preset silent confirmation period, the fault transient process is determined to have ended; a high-level pulse is generated as a termination signal.

[0074] Specifically, based on the fault lockout command, the envelope monitoring logic channel is activated; the noise separation feature vector is squared to generate a power sequence; a sliding window peak hold operation is performed on the power sequence to extract the energy envelope value. In this embodiment, when the system receives the fault lockout command, the parallel envelope monitoring logic channel is activated, and a point-by-point square operation is performed on the input noise separation feature vector (i.e., the source signal component after blind source separation) to convert the voltage amplitude sequence containing positive and negative polarity oscillations into a non-negative instantaneous power sequence. The sliding window peak hold algorithm is then applied to this power sequence. The preset sliding time window length parameter is set to 50 sampling points. The selection of this window length is based on the typical oscillation decay time constant of the partial discharge pulse, smoothing out the high-frequency zero-crossing fluctuations inside the pulse, while retaining the overall rising and falling edge contours of the pulse, outputting an energy envelope value curve that can closely enclose the pulse energy change.

[0075] Specifically, the system reads the background noise dynamic reference as the reference baseline for signal fallback; continuously compares the energy envelope value with the reference baseline; when the comparison result shows that the energy envelope value is lower than the reference baseline, a continuous fallback counter is started; during the counting process, when the energy envelope value is higher than the reference baseline again, the continuous fallback counter is reset. In this embodiment, the system adopts a silent counting mechanism, reads the background noise dynamic reference (e.g., 5mV) updated in real time by the preceding loop mapping module from the register, and retrieves the preset fallback decision ratio coefficient k (recommended value range is 1.2 to 1.5) to compensate for the statistical deviation between the peak and mean values ​​of the background noise. The background noise dynamic reference is multiplied by the ratio coefficient k to generate the signal fallback reference baseline. The comparison logic unit compares the current energy envelope value with the reference baseline every clock cycle (5ns). If the current energy envelope value is less than the reference baseline, it indicates that the device is in an electrically silent state at this moment, and the system starts a 32-bit wide continuous fallback counter for accumulation. However, partial discharge is often accompanied by multiple secondary discharges or aftershocks caused by voltage oscillations. If the energy envelope value changes abruptly at any moment during the counting process and exceeds the reference baseline again (for example, if a subsequent reignition pulse is detected), the comparator outputs a flip signal and immediately sends a clear command to the reset terminal of the counter, forcing the counting process to restart.

[0076] Specifically, when the accumulated value of the continuous fallback counter reaches a preset silent acknowledgment period, the fault transient process is determined to have ended; a high-level pulse is generated as a termination signal. In this embodiment, the preset silent acknowledgment period parameter is set to 800,000 clock cycles. Once the accumulated value of the continuous fallback counter reaches this preset threshold, the counter outputs a carry flag, driving the final stage output circuit to generate a logic high-level pulse with a duration of 1μs as a termination signal. This termination signal acts directly on the direct memory access controller through the interrupt control line, stopping the write operation to the terminal status data stream.

[0077] By extracting the energy envelope value of the signal using square operations and a sliding window peak hold technique, and comparing it with a dynamic background noise benchmark, misjudgments caused by signal oscillations crossing zero points are reduced. In particular, the introduction of a continuous fall-off counter requires the energy envelope to remain below the reference baseline for a preset silent confirmation period before triggering a termination signal. This mechanism effectively addresses the intermittent nature of partial discharge, reducing the risk of prematurely truncating waveform recordings before the fault energy is fully released. By defining the endpoint of the fault transient process, the system improves the integrity of the waveform data, ensuring that the data file covers the entire process of fault arc extinguishing and recovery to stability.

[0078] Furthermore, based on the termination signal, the write pointer is frozen, and a partial discharge data file containing fault precursors and the entire evolution process is extracted from the terminal status data stream; this corresponds to the fault determination module; the specific implementation process includes:

[0079] Upon receiving the termination signal, the increment operation of the write pointer is stopped; the memory address of the current write pointer is obtained as the fault end point; based on the preset pre-trigger length, the fault recall storage unit is traced backward to calculate the memory address of the fault start point; starting from the fault start point, data is continuously read until the fault end point; the read data sequence is packaged and a channel identifier is added to generate a partial discharge data file.

[0080] Specifically, upon receiving a termination signal, the incrementing operation of the write pointer is stopped; the memory address of the current write pointer is obtained as the fault termination point; based on the preset pre-trigger length, the system traces back along the fault recall memory unit to calculate the memory address of the fault start point. In this embodiment, when the system receives the termination signal (logic 1), it sends a mask instruction to the write enable port of the Direct Memory Access (DMA) controller. This operation forcibly stops the linear incrementing of the write pointer within 1 to 2 clock cycles (i.e., 4ns to 8ns), thereby instantly "freezing" the terminal state data stream in the fault recall memory unit. At this time, the write pointer value residing in the register (e.g., the hexadecimal physical address 0x3000_ABCD) is immediately latched and defined as the fault termination point address. To fully reproduce the entire process of the fault from latency to outbreak, the system needs to trace back along the fault based on the preset pre-trigger length parameter L_pre. This parameter is usually set according to the power frequency cycle of the distribution network, for example, set to a historical duration of 1,000,000 points. The address calculation logic unit performs a subtraction operation with modulo to handle boundary crossings in the circular buffer: if the byte offset corresponding to the pre-trigger length minus the fault end address is greater than the physical starting address of the storage unit (e.g., 0x2000_0000), the fault start address is directly the difference between the two; if the calculation result is less than the physical starting address, it indicates that the data segment has crossed the tail and head of the buffer, and the system maps it back to the high-order address range at the end of the buffer using two's complement arithmetic. For example, if the total buffer capacity is 256MB and the current pointer is at 0x2000_0100 (near the head), when backtracking 10MB of data, the calculated fault start address will automatically "roll back" to around 0x2F60_0100. This precise address arithmetic operation ensures that the system can logically and continuously extract the complete data segment containing fault precursor information, regardless of when the fault occurs.

[0081] Specifically, starting from the fault initiation point, data is continuously read until the fault termination point. The read data sequence is packaged and channel identifiers are added to generate a partial discharge data file. In this embodiment, the start and end boundaries in the physical storage space are determined, and the system initiates the data transfer and packaging process. Since the fault data may exhibit a "segmented" distribution characteristic in physical memory (i.e., crossing the loop point of the circular buffer), the reading logic is designed as an adaptive two-stage direct memory access transmission mode. If the fault initiation point address is less than the fault termination point address, the DMA controller performs a single burst read; if the initiation point address is greater than the termination point address, the controller automatically splits the task into two transmissions: the first reads from the fault initiation point to the physical termination address, and the second reads from the physical termination address to the fault termination point, and automatically splices them in the internal first-in-first-out (FIFO) queue to restore the continuous induced voltage sequence and radiation field strength digital sequence on the time axis. Subsequently, the original binary data stream is standardized and packaged. The system constructs a combined document containing configuration files and data files based on the general format standard for power system transient data exchange. During the encapsulation process, the system automatically adds key channel identification metadata to the header, including but not limited to: sampling rate (40MS / s), ADC quantization resolution (14bit), detection circuit type (envelope detection), and sensor transformation coefficient (e.g., 100mV / V for transient ground voltage channels). Furthermore, the system writes the background noise dynamic reference value calculated in the preceding steps as a static parameter of the analog channel into the configuration file, enabling subsequent analysis software to reconstruct the signal-to-noise ratio environment at that time. The final generated partial discharge data file is compressed and stored on a non-volatile solid-state drive, awaiting upload to the distribution automation master station via Ethernet or wireless network, providing maintenance personnel with detailed fault waveform fingerprints and evolution evidence.

[0082] After receiving the termination signal, the system uses reverse backtracking to calculate the fault initiation point, automatically locating and retrieving historical data prior to the fault trigger from the circular buffer, reducing the complexity of manually retrieving precursor information. By continuously reading data from the fault initiation point to the end point and packaging the data sequence with channel identifiers, the system generates a partial discharge data file containing information on the entire process of precursors, outbreaks, and recovery. This improves the universality and readability of monitoring data across different analysis systems. Furthermore, the complete time-series recording reduces the difficulty for maintenance personnel in performing offline waveform reproduction and feature analysis.

[0083] This invention proposes a power distribution terminal fault monitoring system. It utilizes a dual-channel acquisition module to simultaneously acquire analog quantities of insulation status and spatial environment. By mapping these to a unified power grid environment feature matrix, the system improves the completeness of monitoring data coverage of both the equipment itself and the external field. A cyclic mapping module calculates the moving average amplitude based on noise separation feature vectors, constructs a dynamic background noise benchmark that changes with the environment, and generates decision boundary thresholds. This dynamic following mechanism reduces the probability of false alarms caused by environmental noise fluctuations and also improves the system's ability to identify weak early fault signals. Furthermore, through the collaboration of the fault determination module and the fault recall storage unit, a cyclic write pointer refreshes the terminal status data stream in real time. Upon detecting a fault locking command and a termination signal, the system freezes and extracts a data file containing the entire process of fault precursors, evolution, and extinction. This fault recall logic ensures the continuity of transient waveforms on the time axis, reduces the risk of fault information loss, and improves the intelligence level of power distribution network fault diagnosis.

[0084] Example 2

[0085] This embodiment demonstrates the implementation process of the present invention applied to the monitoring of medium and low voltage distribution networks. (See attached document.) Figure 3 The specific implementation method is as follows:

[0086] The system first receives the preprocessed digital sequence of induced voltage and digital sequence of radiation field strength. The time resolution of these two sequences is 20 ns. The logic control unit concatenates these two sequences into vectors according to each processing frame of 1024 sampling points to construct a power grid environment feature matrix with a dimension of 2×1024. This matrix covers a physical time window of approximately 20.48 μs.

[0087] Subsequently, the feature matrix is ​​fed into the cyclic mapping module to perform blind source separation operations. The processor first calculates the covariance matrix of the data within the window. If the calculated off-diagonal element value is 0.75, it indicates significant electromagnetic coupling interference between the two channels. The system then uses eigenvalue decomposition to generate a whitening matrix and orthogonally projects the data. During the iterative process of independent component analysis, the system sets the convergence threshold to 10. -4 Typically, statistically independent source signals can be separated after 15 to 20 fixed-point iterations. The separated noise feature vector is fed into the dynamic threshold generation logic. The system allocates a sliding buffer with a length of 500 sampling points (corresponding to a duration of 10μs) in the field-programmable gate array (FPGA), and calculates the arithmetic mean of the signal magnitudes within the buffer to obtain the real-time background noise level. The digital smoothing filter uses a weighting factor of 0.05, meaning the new reference value is a weighted synthesis of 5% of the current average noise value and 95% of the historical reference value. Assuming the current ambient noise reference value is calculated to be 8mV, the system superimposes a set DC bias constant of 12mV to finally generate a dynamic decision boundary threshold of 20mV.

[0088] In the fault monitoring and data tracing phase, when the instantaneous amplitude of the noise separation feature vector exceeds the 20mV decision boundary threshold, the amplitude comparison logic circuit activates a width verification mechanism. Only when the signal amplitude is above the threshold for four consecutive sampling points (i.e., 80ns) is a valid discharge pulse captured and a fault lockout command generated. Once in the fault lockout state, the system immediately initiates energy envelope monitoring, continuously comparing the signal power with the background reference. To ensure a complete record of the fault, the system sets a silent confirmation period of up to 1,000,000 sampling points. Only when the energy envelope value is below the background noise reference for 1,000,000 consecutive clock cycles, the system generates a termination signal and freezes the write pointer. At this time, the memory controller immediately performs a reverse backtracking operation based on the current pointer address, with the backtracking length also set to 1,000,000 sampling points, thereby accurately extracting the partial discharge data file containing historical data from 20ms before the fault trigger moment from the circular buffer. The file was eventually packaged into a binary waveform document containing 2,000,000 data points, which not only fully reproduced the voltage transient at the time of the fault, but also preserved the electrical environment characteristics within the previous cycle.

[0089] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A power distribution terminal fault monitoring system, characterized in that, include: Dual-channel acquisition module: Acquires analog quantities of insulation status of power distribution equipment and analog quantities of spatial environment of power distribution room, and maps them into a power grid environment feature matrix containing digital sequences of induced voltage and radiation field strength through analog-to-digital converter; The cyclic mapping module performs decoupling operations on the power grid environment feature matrix, calculates the arithmetic mean of each row vector and performs mean-neutralization to generate a zero-mean observation matrix; it constructs the covariance matrix of the zero-mean observation matrix and performs eigenvalue decomposition, generating a whitening transformation matrix through eigenvalues ​​and eigenvectors. Based on the whitening transformation matrix, the zero-mean observation matrix is ​​projected onto the whitening space to output the whitening data stream; Initialize the separation matrix and construct fixed-point iteration rules using the independent component analysis algorithm; during the calibration phase, perform iterative iterations on the whitened data stream based on the fixed-point iteration rules until the separation matrix converges, and then update the separation matrix coefficients. During the monitoring phase, the updated separation matrix is ​​used to perform a linear projection operation on the whitened data stream to extract the source signal components and generate a noise separation feature vector. Based on the moving average amplitude of the noise separation feature vector within the sliding time window, a dynamic baseline for background noise is constructed, and a decision boundary threshold is generated by superimposing a bias voltage. The original sampling points of the power grid environment feature matrix are used as the data frame to be stored. The direct memory access channel of the fault memory storage unit is activated, and the physical storage address currently pointed to by the write pointer of the fault memory storage unit is written. Address arithmetic is performed to linearly increment the write pointer, pointing to the next adjacent storage space. The boundary detection logic is executed in parallel. The incremented write pointer is compared with the physical end address of the fault memory storage unit. If it exceeds the physical end address, the pointer rollback mechanism is triggered to reset the write pointer to the physical start address. When the storage space is saturated, the earliest stored historical data frame is overwritten with a new data frame; the terminal status data stream is formed by cyclically incrementing and resetting the write pointer in the fault memory storage unit. Fault determination module: compares the amplitude of the noise separation feature vector with the decision boundary threshold. When the amplitude is greater than the decision boundary threshold, a fault locking command is generated. Based on the fault locking command, the energy envelope value of the noise separation feature vector is monitored. When the energy envelope value continues to fall back to the background noise dynamic reference range, a termination signal is generated. Based on the termination signal, the write pointer is frozen, and a partial discharge data file containing fault precursors and the entire evolution process is extracted from the terminal status data stream.

2. The power distribution terminal fault monitoring system according to claim 1, characterized in that, The specific generation process of the power grid environment feature matrix includes: using a contact sensor to couple transient pulses from the surface of the metal cabinet of the power distribution terminal to generate a partial discharge simulation quantity of the power distribution equipment; using a non-contact antenna to capture electromagnetic wave signals to generate a spatial environment simulation quantity of the power distribution room; inputting the insulation state simulation quantity of the power distribution equipment and the spatial environment simulation quantity of the power distribution room in parallel to the signal conditioning circuit, outputting a standardized analog voltage signal, and outputting digital sampling points through a synchronous analog-to-digital converter; arranging them sequentially according to time-series logic to generate an induced voltage sequence and a radiation field strength digital sequence respectively; generating a time-synchronized induced voltage sequence and radiation field strength digital sequence through timestamp alignment and inter-channel delay compensation, and combining them according to vector splicing rules to generate the power grid environment feature matrix.

3. The power distribution terminal fault monitoring system according to claim 1, characterized in that, The specific generation process of the decision boundary threshold includes: writing the noise separation feature vector point by point into a sliding buffer of a preset length; performing absolute value operation on the stationed discrete sampling points during the sliding buffer update process to generate a unipolar modulus sequence; performing mean calculation on the unipolar modulus sequence to obtain the average background noise level; inputting the average background noise level into a digital smoothing filter, suppressing random fluctuations through weighted iteration, and outputting a trend curve that smoothly follows the changes in background noise, which is defined as the background noise dynamic reference; retrieving the DC bias voltage constant, superimposing the background noise dynamic reference with the DC bias voltage constant to generate a level value, which is used as the decision boundary threshold.

4. The power distribution terminal fault monitoring system according to claim 1, characterized in that, The specific generation process of the fault lockout command includes: using a digital signal processing unit to perform absolute value calculation on the noise separation feature vector to generate a non-negative instantaneous amplitude sequence; synchronously retrieving the decision boundary threshold as a dynamic comparison benchmark; inputting the instantaneous amplitude sequence point by point into the amplitude comparison logic circuit; in the amplitude comparison logic circuit, logically comparing the instantaneous amplitude at the current moment with the decision boundary threshold; when the comparison result shows that the instantaneous amplitude exceeds the decision boundary threshold, triggering signal dejittering and width verification logic; if the signal continues to exceed the threshold and meets the preset timing constraints, confirming the capture of an abnormal transient event; driving the fault state trigger to flip, generating a fault lockout command.

5. A power distribution terminal fault monitoring system according to claim 1, characterized in that, The specific generation process of the termination signal includes: activating the envelope monitoring logic channel based on the fault lockout command; squaring the noise separation feature vector to generate a power sequence; performing a sliding window peak hold operation on the power sequence to extract the energy envelope value; reading the background noise dynamic benchmark as the reference baseline for signal fallback; continuously comparing the energy envelope value with the reference baseline; starting a continuous fallback counter when the comparison result shows that the energy envelope value is lower than the reference baseline; resetting the continuous fallback counter when the energy envelope value is higher than the reference baseline again during the counting process; determining that the fault transient process ends when the accumulated value of the continuous fallback counter reaches the preset silent confirmation period; and generating a high-level pulse as a termination signal.

6. The power distribution terminal fault monitoring system according to claim 1, characterized in that, The specific process of generating the partial discharge data file includes: receiving a termination signal to stop the incrementing operation of the write pointer; obtaining the memory address of the current write pointer as the fault end point; calculating the memory address of the fault start point by tracing back along the fault recall storage unit according to the preset pre-trigger length; continuously reading data from the fault start point until the fault end point; packaging the read data sequence and adding channel identifiers to generate the partial discharge data file.

7. A method for monitoring faults in a power distribution terminal, characterized in that, include: The analog quantities of insulation status of power distribution equipment and spatial environment of power distribution room are collected and mapped into a power grid environment feature matrix containing induced voltage sequence and radiated field strength digital sequence through analog-to-digital converter; The power grid environment feature matrix is ​​decoupled, the arithmetic mean of each row vector is calculated and mean-neutralization is performed to generate a zero-mean observation matrix; the covariance matrix of the zero-mean observation matrix is ​​constructed and eigenvalue decomposition is performed, and a whitening transformation matrix is ​​generated through eigenvalues ​​and eigenvectors; Based on the whitening transformation matrix, the zero-mean observation matrix is ​​projected onto the whitening space to output the whitening data stream; Initialize the separation matrix and construct fixed-point iteration rules using the independent component analysis algorithm; during the calibration phase, perform iterative iterations on the whitened data stream based on the fixed-point iteration rules until the separation matrix converges, and then update the separation matrix coefficients. During the monitoring phase, the updated separation matrix is ​​used to perform a linear projection operation on the whitened data stream to extract the source signal components and generate a noise separation feature vector. Based on the moving average amplitude of the noise separation feature vector within the sliding time window, a dynamic baseline for background noise is constructed, and a decision boundary threshold is generated by superimposing a bias voltage. The original sampling points of the power grid environment feature matrix are used as the data frame to be stored. The direct memory access channel of the fault memory storage unit is activated, and the physical storage address currently pointed to by the write pointer of the fault memory storage unit is written. Address arithmetic is performed to linearly increment the write pointer, pointing to the next adjacent storage space. The boundary detection logic is executed in parallel. The incremented write pointer is compared with the physical end address of the fault memory storage unit. If it exceeds the physical end address, the pointer rollback mechanism is triggered to reset the write pointer to the physical start address. When the storage space is saturated, the earliest stored historical data frame is overwritten with a new data frame; the terminal status data stream is formed by cyclically incrementing and resetting the write pointer in the fault memory storage unit. The amplitude of the noise separation feature vector is compared with the decision boundary threshold. When the amplitude is greater than the decision boundary threshold, a fault locking command is generated. Based on the fault locking command, the energy envelope value of the noise separation feature vector is monitored. When the energy envelope value continues to fall back to the background noise dynamic reference range, a termination signal is generated. Based on the termination signal, the write pointer is frozen, and a partial discharge data file containing fault precursors and the entire evolution process is extracted from the terminal status data stream.

Images