A method and system for identifying partial discharge audio inside a distribution box and a medium

Abstract

The present application relates to the technical field of discharge identification, and particularly relates to a power distribution box internal partial discharge audio identification method and system and a medium. The method comprises the following steps: obtaining physical structure parameters and electrical connection relationship of the power distribution box, and constructing an acoustic-electric topology matrix; forming a periodic phase window sequence based on the internal acoustic signal at the alternating current zero-crossing point; extracting short-time energy envelope and mel frequency band features for each window, and combining the topology matrix to perform cavity acoustic reflection correction to generate a cavity corrected acoustic feature tensor; then, phase information and electrical topology are fused through a phase gating convolution network to realize discharge probability estimation and node-level positioning; multi-cycle propagation consistency and physical attenuation law are combined for verification to construct a stable abnormal discharge event chain. The present application solves the problems of severe acoustic signal distortion in the sealed power distribution box, non-use of phase information and inaccurate identification caused by multi-source noise interference, and realizes high-precision identification of partial discharge and reliable positioning of the discharge node.

Description

Technical Field

[0001] This invention relates to the field of discharge identification technology, and in particular to a method, system and medium for partial discharge audio identification inside a distribution box. Background Technology

[0002] As the core terminal equipment of the power distribution system, the internal insulation condition of the distribution box directly affects the reliability of power supply. Partial discharge is an early sign of insulation degradation. By detecting and identifying the acoustic signals generated by the discharge, early warning intervention can be achieved before a fault occurs, which has significant operational and maintenance value.

[0003] However, existing technologies face multiple challenges in audio recognition of partial discharge in enclosed metal cavities such as distribution boxes. First, the cavity structure of the distribution box leads to severe acoustic signal distortion. Sound waves undergo multiple reflections and interferences between the metal inner walls and partitions, resulting in significant differences in the attenuation of different frequency components—high frequencies attenuate rapidly due to their short wavelengths and large reflection losses, while low frequencies may be abnormally amplified due to cavity resonance. The signal ultimately received by the sensor is an aliased signal modulated by the cavity, fundamentally deviating from the original sound wave from the discharge source. Traditional methods ignore the propagation characteristics of the cavity's physical structure, directly using the acquired signal for feature extraction, severely interfering with recognition accuracy. Second, partial discharge has an inherent coupling relationship with the AC voltage phase. In the voltage peak region, the electric field strength is highest, and breakdown discharge easily occurs at weak insulation points, concentrating sound wave energy in the high-frequency range; while in the voltage zero-crossing region, discharge activity weakens, and the acoustic characteristics are dominated by low-frequency reverberation. Existing technologies typically process acoustic signals from different phases together, forcibly merging different discharge modes, causing feature space chaos and degrading recognition performance. Furthermore, the distribution box site is subject to instantaneous interference such as mechanical vibration and switch operation, and its short-term energy change characteristics are highly similar to those of partial discharge. Relying solely on the classification and discrimination of single-frame data makes it difficult to distinguish between real discharge and occasional noise, resulting in a persistently high false alarm rate.

[0004] In summary, how to correct signal distortion in a complex acoustic field within a sealed cavity, establish a correlation model between acoustic propagation and electrical topology, dynamically adjust the identification strategy based on phase evolution, and filter out interference through multi-cycle physical consistency verification are key issues that urgently need to be addressed in this field. Summary of the Invention

[0005] This invention provides a method, system, and medium for audio recognition of partial discharge inside a distribution box, which is used to accurately detect, identify the type, and locate nodes of internal partial discharge in the closed cavity environment of the distribution box, providing a reliable basis for condition-based maintenance and fault early warning of power distribution equipment.

[0006] In a first aspect, the present invention provides a method for partial discharge audio recognition inside a distribution box, the method comprising: S1. Obtain the physical structure parameters and internal electrical connection relationships of the distribution box, determine the spatial location of the internal electrical nodes, calculate the acoustic attenuation coefficient, electrical coupling weight and node distance between nodes, and generate an acoustic-electric topology matrix. S2. Acquire the AC periodic synchronization signal, extract its zero crossover point as the starting reference, divide the acquired internal acoustic signal into half-cycle length acoustic data segments, add phase angle labels to each acoustic data segment, and form a periodic phase window sequence. S3. Calculate the short-time energy envelope and locate the energy mutation point for the periodic phase window sequence, extract the frequency band features of the Mel filter bank, use the acoustic reflection side weights of the acoustic-electric topology matrix to perform reflection gain correction on each frequency band feature, and generate the cavity correction acoustic feature tensor by combining the energy mutation point and the phase angle label. S4. Input the cavity correction acoustic feature tensor into the phase-gated convolutional unit, and modulate the convolutional kernel weights by using the phase angle label and the node distance and electrical adjacency edge weights in the acoustic-electric topology matrix, and output the abnormal discharge candidate probability and the distribution of suspected discharge nodes. S5. Based on the distribution of the suspected discharge nodes, and combined with the node distance and sound reflection weight in the acoustic-electric topology matrix, perform acoustic propagation consistency verification on adjacent periodic phase windows, screen out discharge events that conform to the physical laws of cavity sound propagation, and determine the abnormal discharge event chain.

[0007] In a second aspect, the present invention provides a partial discharge audio recognition system for use in the above-described method, the system comprising: The acoustic-electric topology construction unit is used to obtain the physical structure parameters and internal electrical connection relationships of the distribution box, determine the spatial position of the internal electrical nodes, calculate the acoustic attenuation coefficient, electrical coupling weight and node distance between nodes, and generate the acoustic-electric topology matrix. The phase window division unit is used to acquire the AC periodic synchronization signal, extract its zero crossover point as the starting reference, divide the acquired internal acoustic signal into acoustic data segments of half-cycle length, attach phase angle labels to each acoustic data segment, and form a periodic phase window sequence. The cavity feature correction unit is used to calculate the short-time energy envelope and locate the energy mutation point of the periodic phase window sequence, extract the frequency band features of the Mel filter bank, use the acoustic reflection side weight of the acoustic-electric topology matrix to perform reflection gain correction on each frequency band feature, and generate the cavity correction acoustic feature tensor by combining the energy mutation point and the phase angle label. The phase-gated convolutional unit is used to input the cavity correction acoustic feature tensor into the phase-gated convolutional unit, and modulate the convolutional kernel weights by using the phase angle label and the node distance and electrical adjacency edge weight in the acoustic-electric topology matrix, and output the abnormal discharge candidate probability and the distribution of suspected discharge nodes. The propagation consistency verification unit is used to perform acoustic propagation consistency verification on adjacent periodic phase windows based on the distribution of suspected discharge nodes and the node distance and acoustic reflection weight in the acoustic-electric topology matrix, to screen discharge events that conform to the physical laws of acoustic propagation in the cavity, and to determine the abnormal discharge event chain.

[0008] Thirdly, the present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the above-described method.

[0009] The beneficial effects of this invention are as follows: This invention constructs an acoustic-electric topology matrix to uniformly model the spatial location of electrical nodes inside the distribution box, the attenuation characteristics of sound propagation paths, and the coupling relationship of electrical connections. This provides a clear physical constraint basis for subsequent acoustic feature extraction and structural priors for discharge source location. Based on the zero-crossing point of AC current, the acoustic signal is divided into half-cycle phases, and load current and switch state information are introduced to form a periodic phase window sequence. This enables phase alignment and decoupling of the discharge signal in the time dimension, reducing feature aliasing caused by different operating states from the source. Furthermore, short-time energy analysis and frequency domain Mel feature extraction are performed on the signal within each phase window. Cavity propagation correction of the frequency band energy is performed using the acoustic reflection side weights in the acoustic-electric topology matrix. Simultaneously, an adaptive noise compensation mechanism based on load current is introduced, effectively eliminating the interference of multi-reflection effects in the sealed cavity and low-frequency mechanical noise on the features, resulting in a physically consistent cavity-corrected acoustic feature tensor. In the identification phase, a phase-gating mechanism is used to embed AC phase information into the convolution kernel generation process, and dynamic convolution weight modulation is achieved by fusing node distance and electrical adjacency relationships. This enables the model to adaptively extract discharge features for different electrical nodes and phase states, realizing the transformation from "signal classification" to "node-level localization." Finally, dual physical constraints of cross-cycle arrival time consistency and frequency band attenuation consistency are introduced to verify and filter the identification results over multiple cycles, retaining only stable discharge events that conform to the acoustic propagation laws, thereby constructing a continuous and reliable chain of abnormal discharge events. Through the synergy of the above technical solutions, high-precision identification, reliable localization, and continuous evolution analysis of partial discharge signals inside the distribution box are achieved under conditions of complex cavities, multiple noises, and multi-source interference. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1This is a flowchart of a partial discharge audio recognition method inside a distribution box, as described in the embodiment. Figure 2 This is a schematic diagram comparing the spectrum before and after cavity reflection gain correction in the embodiment; Figure 3 This is a structural diagram of a partial discharge audio recognition system inside a distribution box, as shown in the embodiment. Detailed Implementation

[0012] This invention provides a method, system, and medium for partial discharge audio recognition inside a distribution box. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0013] This solution is applied to online monitoring and early warning scenarios of partial discharge inside distribution boxes, and is particularly suitable for enclosed distribution boxes containing multiple electrical nodes such as circuit breakers, contactors, busbars, and terminals, with internal spaces separated by partitions. The distribution box has a physical cavity structure and electrical connection topology that can be calibrated using three-dimensional coordinates. It houses acoustic acquisition equipment and externally connects to AC periodic synchronization signals, such as square wave signals provided by voltage transformers, and is equipped with sensing units to acquire real-time load current and switch status. The acoustic acquisition equipment receives discharge sound waves within the cavity after multiple reflections and attenuation through the metal walls and partitions. By fusing the cavity's physical structural parameters and electrical connection relationships into an acoustic-electric topology matrix, and combining this with the AC phase synchronization signal to perform half-cycle segmentation and phase labeling of the continuous acoustic signal, after cavity reflection gain correction, a multi-dimensional acoustic feature tensor is extracted. Phase-gated convolution and multi-cycle propagation consistency verification are then used to ultimately achieve accurate location of the discharge source node and reliable construction of the abnormal discharge event chain.

[0014] For ease of understanding, the specific process of the embodiments of the present invention will be described below, such as... Figure 1 As shown in the figure, an embodiment of the present invention provides a method for partial discharge audio recognition inside a distribution box, the method comprising: S1. Obtain the physical structure parameters and internal electrical connections of the distribution box, determine the spatial location of the internal electrical nodes, calculate the acoustic attenuation coefficient, electrical coupling weight, and node distance between nodes, and generate an acoustic-electric topology matrix; including: Read the spatial coordinates and voltage levels of each electrical node inside the distribution box, and combine the box cavity dimensions and the position and material of the internal partitions to generate multiple physical spatial sound propagation paths between the nodes. Based on the length and medium properties of each physical space acoustic propagation path, the acoustic attenuation coefficient of the corresponding path is calculated. Based on the charge level and electrical connection relationship of the internal electrical nodes, the electrical coupling weight between the nodes is calculated. The acoustic propagation path with the smallest acoustic attenuation coefficient is selected as the optimal path, and the length of the optimal path is extracted as the node distance. The acoustic attenuation coefficient corresponding to the optimal path is used as the acoustic reflection edge weight, and the electrical coupling weight is used as the electrical adjacency edge weight. Combined with the node distance, an acoustic-electric topology matrix is ​​constructed.

[0015] Specifically, a three-dimensional Cartesian coordinate system is established with the lower left corner of the distribution box as the origin. The spatial coordinates of each electrical node, including the main circuit breaker, branch circuit breakers, and terminal blocks, are then determined within this system. Each node is also labeled with its voltage level, which has a pre-defined mapping relationship with a specific operating voltage value. For example, incoming line nodes are labeled as high-voltage and correspond to the first voltage level, outgoing line nodes as low-voltage and correspond to the second voltage level, and neutral line nodes as zero-potential and correspond to the zero voltage level. Simultaneously, the length, width, and height dimensions of the distribution box cavity, as well as the position, material, and thickness parameters of each internal partition, are obtained. The partition material is categorized as either metal sheet or insulating sheet. Based on this data, for any two electrical nodes, the mirror method is used to determine all possible sound propagation paths between them. The length of the direct path is directly taken as the Euclidean distance between the two nodes in three-dimensional space. For first-order and higher-order reflection paths, the source node is mirrored relative to the enclosure wall or partition surface to obtain the coordinates of the mirrored source node. The intersection of the line connecting the mirrored source node and the target node with the reflecting surface is the reflection point. The total length of this reflection path is equal to the sum of the distance from the source node to the reflection point and the distance from the reflection point to the target node. By enumerating the combinations of each wall and partition as reflecting surfaces, first-order, second-order, and third-order reflection paths are calculated. Since the sound energy attenuation increases with the number of reflections, usually only reflection paths of the third order or lower are considered. If the mirror method cannot generate a valid intersection point due to discontinuous partitions or obstructed nodes, the path is deemed non-existent and discarded. Thus, a set of physical spatial sound propagation paths, including one direct path and multiple reflection paths, is established for each pair of nodes.

[0016] Then, based on the length and medium properties of each physical sound propagation path, the acoustic attenuation coefficient of the corresponding path is calculated. Based on the voltage level and electrical connection relationship of the internal electrical nodes, the electrical coupling weight between nodes is calculated. The acoustic attenuation coefficient quantifies the energy loss of sound waves along a specific propagation path. Sound waves attenuate due to medium absorption when propagating in air; the degree of attenuation is related to the propagation distance and frequency. Specifically, for each path, the peak frequency of a typical partial discharge signal inside the distribution box, such as 40 kHz, is used as the reference frequency for calculation. The air attenuation coefficient used is the value at this reference frequency, and its value is selected based on the temperature and humidity conditions of the air medium inside the distribution box. The aforementioned typical partial discharge signal is a frequency signal that represents the main energy distribution of the partial discharge acoustic signal inside the distribution box and is sensitive to the attenuation response of cavity sound reflection. The acoustic attenuation coefficient of this path is obtained by summing the following components: air absorption attenuation, which is equal to the selected air attenuation coefficient multiplied by the total propagation length of the path in the air; if the path penetrates the partition, the transmission loss determined by the partition material and thickness is added; if the path includes wall reflection, the reflection loss corresponding to each reflection is added. The transmission loss and reflection loss mentioned above can be obtained through experimental simulation measurement. Electrical coupling weight is used to quantify the electrical association strength between nodes. Electrical nodes inside the distribution box are connected by conductors such as busbars and wires to form electrical loops. When a node experiences partial discharge, the high-frequency electromagnetic pulse generated by the discharge propagates through conductors to adjacent nodes, affecting the electric field distribution of adjacent nodes. The calculation of electrical coupling weight is based on the electrical topology relationship between nodes. First, an electrical connection topology graph is constructed with electrical nodes as vertices and busbar or wire connections as edges. The shortest path algorithm is used to calculate the electrical distance between two nodes. This electrical distance is defined as the number of conductor segments traversed by the shortest connection path. When two nodes are directly connected, the electrical distance is 1. The electrical coupling weight between two nodes is then defined as the reciprocal of this electrical distance. If there is no electrical connection between two nodes, the electrical distance is considered infinite, and the electrical coupling weight is reset to 0. The sound propagation path with the smallest acoustic attenuation coefficient is selected as the optimal path, and the length of this optimal path is extracted as the node distance.

[0017] Finally, the acoustic attenuation coefficient corresponding to the optimal path is used as the acoustic reflection edge weight, and the electrical coupling weight is used as the electrical adjacency edge weight. Combined with the node distance, an acoustic-electric topology matrix is ​​constructed. The rows and columns of this matrix correspond to the electrical nodes inside the distribution box, and each element in the matrix is ​​a data structure containing three components. The first component is the acoustic reflection edge weight, which is the total attenuation of the optimal acoustic propagation path from the row node to the column node at the reference frequency. The second component is the electrical adjacency edge weight, which is the electrical coupling weight between the two nodes. The third component is the node distance. If there is no effective acoustic propagation path between the row node and the column node, both the acoustic reflection edge weight and the node distance are set to positive infinity; if there is no electrical connection between the two nodes, the electrical adjacency edge weight is set to 0. Through the above technical solution, the constructed acoustic-electric topology matrix encodes the physical attenuation characteristics of acoustic propagation and the coupling relationship of electrical circuits within the distribution box into unified structured parameters, providing a complete and self-consistent quantitative basis for subsequent steps such as modulating convolution kernel weights based on phase angle and node distance, and performing physical consistency verification based on topological distance.

[0018] S2. Acquire the AC periodic synchronization signal, extract its zero-crossing point as the starting reference, divide the acquired internal acoustic signal into half-cycle length acoustic data segments, attach phase angle labels to each acoustic data segment, and form a periodic phase window sequence; including: Acquire the AC periodic synchronization signal from the incoming side of the distribution box and the continuously acquired internal acoustic signal, and extract the zero crossover point of the AC periodic synchronization signal as the segmentation starting reference. Using the zero-crossing point as the time node, the continuous internal acoustic signal is divided into multiple half-cycle length acoustic data segments, and a phase angle label is added according to the phase interval corresponding to each acoustic data segment in the AC cycle. The real-time load current data and switch status data of the distribution box are acquired, and the real-time load current data, switch status data and the phase angle label of the corresponding acoustic data segment are bound together to the acoustic data segment to form a periodic phase window sequence.

[0019] Specifically, since the generation of partial discharge acoustic signals is not an isolated event, its probability of occurrence, intensity, and spectral characteristics are closely coupled with the voltage phase of the AC power grid to which the distribution box is connected. Discharge modes in different phase intervals are fundamentally different; if all acoustic signals are processed in the same way, it will cause confusion in the feature space, severely affecting the accuracy of subsequent identification.

[0020] Therefore, the AC periodic synchronization signal from the incoming side of the distribution box and the continuously acquired internal acoustic signal are first obtained, and the zero-crossing point of the AC periodic synchronization signal is extracted as the starting reference for segmentation. The AC periodic synchronization signal is a reference signal used to accurately identify the phase information of the AC voltage. It is usually extracted from the incoming voltage of the distribution box through a voltage transformer, and its form is a square wave signal or pulse signal with the same frequency and phase as the AC voltage of the power grid. The rising edge of this synchronization signal corresponds to the positive zero-crossing point where the AC voltage transitions from the negative half-cycle to the positive half-cycle, and the falling edge corresponds to the negative zero-crossing point where the AC voltage transitions from the positive half-cycle to the negative half-cycle. The zero-crossing point is the moment when the instantaneous value of the AC voltage crosses zero, at which point the voltage amplitude is zero. The continuously acquired internal acoustic signal is time-domain waveform data continuously acquired at a fixed sampling rate by acoustic acquisition equipment installed inside or on the outer wall of the distribution box.

[0021] Then, using the zero-crossing point as the time node, the continuous internal acoustic signal is divided into multiple half-cycle length acoustic data segments, and a phase angle label is attached to each acoustic data segment according to its corresponding phase interval in the AC cycle. Specifically, for a 50Hz AC power grid, the duration of a complete AC cycle is 20ms, and the half-cycle duration is 10ms. The acoustic signal is sequentially segmented based on each adjacent rising and falling edge of the synchronization signal. The data segment starting from the positive zero-crossing point and ending at the next negative zero-crossing point corresponds to the positive half-cycle of the AC cycle, with a phase interval of 0 to 180 degrees; the data segment starting from the negative zero-crossing point and ending at the next positive zero-crossing point corresponds to the negative half-cycle of the AC cycle, with a phase interval of 180 to 360 degrees. Each data segment is assigned a phase angle label, which represents the position of that signal segment within the complete AC cycle. Within the data segment, the precise phase angle of each sampling point can be calculated based on the sampling period and time offset. The phase angle increases at a rate of 18 degrees / ms under 50Hz AC power.

[0022] Next, real-time load current data and switch status data of the distribution box are acquired. The real-time load current data, switch status data, and the phase angle tags of the corresponding acoustic data segments are bound together to the acoustic data segments to form a periodic phase window sequence. Specific real-time load current data is collected through current transformers or Hall current sensors installed on the main circuit and each branch circuit. This data reflects the load level of the distribution box within the corresponding half-cycle. The magnitude of the load current directly affects the intensity of conductor electromagnetic vibration noise and may mask weak discharge acoustic signals. For each half-cycle acoustic data segment, a sequence of current sampling points corresponding to the timestamp of that acoustic data segment is extracted from the sampling data stream of the current sensor. The arithmetic mean of this sequence is calculated as the real-time load current data. Switch status data is obtained by reading the auxiliary contact signals of each circuit breaker or the communication interface of the intelligent circuit breaker. The on / off state of each circuit is represented in binary form, with 1 for closed and 0 for open, thus forming a switch status vector. The correspondence between each element in this vector and the electrical nodes defined in step S1 is arranged in a preset order. The switch state vector defines the range of potentially energized nodes within the distribution box at the current moment. When a circuit breaker in a branch circuit is open, its downstream nodes cannot become discharge sources. The aforementioned real-time load current data, along with the switch state vector and the phase angle label of that data segment, are encapsulated into the corresponding acoustic data segment, forming a multi-dimensional data window containing an acoustic sampling point sequence, a phase angle label sequence, an average load current value, and a switch state vector. All consecutive data windows are arranged in chronological order, forming a periodic phase window sequence.

[0023] The obtained periodic phase window sequence through the above technical solution enables the subsequent feature extraction and pattern recognition processes to be processed in phase-by-phase and under different operating conditions. This not only effectively utilizes the phase-dependent characteristics of partial discharge activity, but also suppresses noise interference introduced by changes in operating conditions and narrows the search range of discharge nodes by introducing load current and switch status information, thereby significantly improving the accuracy and robustness of partial discharge audio recognition inside the distribution box.

[0024] S3. Calculate the short-time energy envelope and locate energy abrupt change points for the periodic phase window sequence, extract the frequency band features of the Mel filter bank, perform reflection gain correction on each frequency band feature using the acoustic reflection side weights of the acoustic-electric topology matrix, and generate a cavity correction acoustic feature tensor by combining the energy abrupt change points and phase angle labels; including: Each window in the periodic phase window sequence is traversed, and the energy mutation threshold is dynamically set in combination with the real-time load current data bound to the window. The frame energy of the acoustic signal in the window is calculated frame by frame to generate a short-time energy envelope. The time point when the energy change rate exceeds the energy mutation threshold is extracted as the short-time energy mutation point. The acoustic signal in each periodic phase window is frequency domain transformed to extract the frequency band features of the Mel filter bank. Based on the acoustic reflection side weights in the acoustic-electric topology matrix, and combined with real-time load current data, attenuation compensation is performed on low-frequency noise. The reflection gain compensation amount corresponding to each frequency band is calculated, and the energy of each frequency band of the Mel filter bank frequency band characteristics is weighted and compensated to obtain the corrected frequency band characteristics. The corrected frequency band characteristics, the feature parameters of short-time energy mutation points, and the phase angle labels of the corresponding windows are concatenated to generate the cavity corrected acoustic feature tensor.

[0025] Specifically, after constructing the aforementioned periodic phase window sequence, the continuous acoustic signal has been divided into half-cycle data windows carrying phase tags and operating status information, i.e., real-time load current data. However, the original acoustic signal within each window is still an aliased signal after cavity reflection distortion and contains a large amount of steady-state background noise. In the time domain, the energy released during partial discharge is much higher than the background noise, exhibiting a short-term energy mutation; in the frequency domain, the discharge sound wave has a wide spectral distribution, but the energy in each frequency band is attenuated or enhanced to varying degrees due to differences in the cavity propagation path.

[0026] Therefore, it is necessary to first traverse each window in the periodic phase window sequence, dynamically set the energy mutation threshold based on the real-time load current data bound to the window, calculate the frame energy of the acoustic signal within the window, generate a short-time energy envelope, and extract the time points where the energy change rate exceeds the energy mutation threshold as short-time energy mutation points. The short-time energy envelope is a curve describing the trend of signal energy change along the time axis. Specifically, the acoustic signal within the window is divided into frames according to a preset frame length and frame shift. The frame length is 1 / 10 to 1 / 20 of the number of sampling points in half a cycle, and the frame shift is 1 / 4 to 1 / 2 of the frame length to ensure overlap between adjacent frames and to avoid missing transient events. For each frame signal, the sum of the squares of the amplitudes of all sampling points within the frame is calculated as the frame energy. Arranging the frame energies of all frames in chronological order yields the short-time energy envelope. The energy mutation threshold is used to distinguish between energy mutations caused by partial discharge and energy fluctuations under normal operating conditions. A threshold that is too low may misjudge background noise as discharge, while a threshold that is too high may miss weak discharge signals. The aforementioned energy mutation threshold is dynamically set based on the real-time load current data bound to the window. Because a larger load current results in stronger electromagnetic vibration noise in the conductor and a higher background energy level, the threshold should be adjusted accordingly. The specific setting method is as follows: Under normal operating conditions with no discharge in the distribution box, acoustic signal samples are collected at different load current levels. The mean and standard deviation of the short-time energy envelope energy change rate at each load current level are calculated, and a mapping table between the load current and the mean and standard deviation is established. For any half-cycle window, the corresponding mean and standard deviation are obtained by looking up the table based on the bound real-time load current value. The energy mutation threshold is set to the mean plus three times the standard deviation. The short-time energy envelope is numerically differentiated along the time axis, and the difference in frame energy between adjacent frames is calculated and divided by the frame time interval to obtain the energy change rate sequence. This energy change rate sequence is traversed, and the moments when the change rate exceeds the energy mutation threshold are marked as short-time energy mutation points. The relative time position and corresponding instantaneous phase angle of each mutation point within the window are recorded.

[0027] Then, a frequency domain transformation is performed on the acoustic signal within each periodic phase window to extract the frequency band features of the Mel filter bank. The Mel filter bank is a spectral analysis tool that simulates the characteristics of human auditory perception. The human ear has a higher resolution for low-frequency sounds than high-frequency sounds, and the Mel scale converts physical frequencies into psychoacoustic frequencies that conform to human auditory perception through nonlinear mapping. Specifically, a Fourier transform is first performed on the acoustic signal within the window, converting the time-domain signal into a frequency-domain spectrum covering the frequency range from 0 Hz to half the sampling rate of the acoustic signal. A set of triangular bandpass filters, uniformly distributed on the Mel scale, is then designed, with 20 to 40 filters. The center frequencies of each filter are arranged at equal intervals on the Mel scale, resulting in a dense low-frequency and sparse high-frequency distribution after mapping back to the physical frequencies. Each frequency component of the signal spectrum is multiplied by the triangular weight function of each filter, and the product is summed within the filter's bandwidth to obtain the output energy of each filter. The logarithm of the output energy of all filters is then taken to form the frequency band feature vector of the Mel filter bank.

[0028] Next, based on the acoustic reflection weights in the acoustic-electric topology matrix and combined with real-time load current data, attenuation compensation is performed on low-frequency noise. The reflection gain compensation for each frequency band is calculated, and the energy of each frequency band in the Mel filter group's frequency band characteristics is weighted and compensated to obtain the corrected frequency band characteristics. Because the distribution box cavity has different reflection and attenuation characteristics for sound waves of different frequencies, high-frequency components are significantly attenuated in multiple reflections, while low-frequency components may be enhanced due to cavity resonance. Therefore, the directly extracted Mel filter group frequency band characteristics deviate from the true spectrum of the sound source. The acoustic reflection weights in the acoustic-electric topology matrix record the total attenuation of the optimal sound propagation path from each node to the acquisition device at the reference frequency. This attenuation implicitly contains the modulation information of the cavity for sound wave propagation. To decompose the acoustic reflection weights at a single reference frequency into reflection gain compensation for each frequency band, a frequency attenuation allocation function is established based on the acoustic theory that air absorption attenuation is proportional to the square of the frequency and wall reflection loss increases with frequency. For the center frequency of the frequency band corresponding to the k-th Mel filter... The reflection gain compensation amount in this frequency band Calculate using the following formula: Where A is the acoustic reflection edge weight of the optimal path corresponding to this node in the acoustic-electric topology matrix. The peak frequency of the typical partial discharge signal determined in step S1 is α, which is the attenuation frequency dependence index, selected within the range of 1 to 2 based on the material and structure of the distribution box cavity. Specifically, for propagation paths dominated by metal wall reflection, wall reflection loss increases with frequency but the growth trend is relatively gentle, and the value of α approaches 1; for propagation paths dominated by partition transmission, transmission loss increases significantly with frequency and the growth trend is steep, and the value of α approaches 2. Meanwhile, the electromagnetic vibration noise energy generated by the load current is mainly concentrated in the low-frequency band. When performing cavity correction, it is necessary to combine load current data to compensate for noise attenuation in the low-frequency band to prevent mechanical noise from being mistakenly amplified into a discharge signal. For low-frequency filters with a center frequency lower than the preset cutoff frequency, the noise attenuation factor is calculated. In the middle, I represents the real-time load current value for the current window. γ is the rated current of the distribution box, and γ is the noise coupling coefficient. A preset upper limit for noise attenuation is set to avoid excessive attenuation. This noise attenuation factor is subtracted from the reflection gain compensation amount in the corresponding low-frequency band to obtain the net compensation amount in the low-frequency band. The noise coupling coefficient described above characterizes the energy transfer efficiency of electromagnetic vibration noise in the distribution box through the cavity structure to the acoustic acquisition equipment. Its value is related to the specific mechanical structure of the distribution box, the sensor installation position, and the partition layout. Specifically, under the condition that the distribution box has no partial discharge and is under known load current conditions, the output energy of the low-frequency band Mel filter bank is collected, and the linear regression slope between this energy and the change in load current is calculated. This slope is used as the calibration value of γ. The final compensation amount for each frequency band is added to the logarithmic energy value of the corresponding frequency band of the Mel filter bank's frequency band characteristics, thus completing the weighted compensation. The compensated frequency band characteristics are closer to the true spectrum of the discharge source in a free field, and are called the corrected frequency band characteristics, such as... Figure 2 As shown, (a) is the spectrum before correction, i.e., the frequency band energy distribution of the original acquired signal after extraction by the Mel filter bank without cavity correction. Due to the acoustic reflection and interference effects of the sealed metal cavity of the distribution box, the low-frequency band is abnormally enhanced due to cavity resonance, while the high-frequency band attenuates rapidly with increasing frequency due to wall reflection loss and air absorption; (b) is the spectrum after correction, i.e., the corrected frequency band characteristics after acoustic reflection side-weight compensation. At the same time, the true original spectrum is shown in dotted line form as a reference. After acoustic reflection side-weight compensation, the low-frequency resonance enhancement is effectively suppressed, the high-frequency energy loss is recovered, and the corrected spectrum is highly consistent with the true original spectrum.

[0029] Finally, the correction band features, the feature parameters of short-time energy mutation points, and the phase angle labels of the corresponding windows are concatenated to generate the cavity correction acoustic feature tensor. The feature parameters of the short-time energy mutation points include the relative temporal position of the mutation point within the window and the normalized energy amplitude. The relative temporal position is represented by the proportion of the mutation point's time to the half-cycle duration, ranging from 0 to 1. The normalized energy amplitude is obtained by dividing the frame energy of the mutation point by the maximum value of the short-time energy envelope of the window. If multiple mutation points exist within a window, the three with the largest energy amplitudes are selected, with each mutation point contributing two dimensions: relative temporal position and normalized energy amplitude, for a total of six dimensions. The phase angle label is the normalized value of the initial phase angle of the window, i.e., the phase angle divided by 360 degrees. The correction band features, mutation point feature parameters, and phase angle labels are concatenated in a preset order to form a multi-dimensional vector, which is the cavity correction acoustic feature tensor for that window. This tensor integrates three complementary pieces of information: cavity-corrected frequency domain energy distribution, temporal transient change location, and phase context, providing structured input for the phase-gated convolutional units in subsequent steps.

[0030] Through the above technical solution, the transient change point of discharge is accurately located in the time domain using the load adaptive threshold. In the frequency domain, the spectral distortion caused by cavity propagation is corrected by the acoustic reflection side weights and frequency attenuation allocation function of the acoustic-electric topology matrix. In addition, the low-frequency noise related to the working condition is suppressed by combining the load current. The cavity-corrected acoustic feature tensor generated at the end carries the time-frequency features and phase information of the discharge event. It effectively overcomes the adverse effects of sound field distortion in the sealed cavity and interference from working condition noise on feature extraction, laying a solid foundation for subsequent high-precision recognition.

[0031] S4. Input the cavity correction acoustic feature tensor into a phase-gated convolutional unit, and modulate the convolutional kernel weights by using the phase angle label, the node distance in the acoustic-electric topology matrix, and the electrical adjacency edge weights, to output the abnormal discharge candidate probability and the distribution of suspected discharge nodes; including: Phase angle labels are extracted from the cavity correction acoustic feature tensor, and the phase angle labels are processed by a sine mapping function to generate a phase gate vector. The phase gate vector is then used to perform tensor operations with the node distances and electrical adjacency weights in the acoustic-electric topology matrix to generate dynamic convolution kernel weights. The dynamic convolution kernels are then used to perform convolution feature extraction on the cavity correction acoustic feature tensor to obtain a fused feature map. The fused feature map is input into a fully connected layer for classification and regression processing, and the abnormal discharge candidate probability and initial node distribution are output. The initial node distribution is then constrained by a charged node mask based on the switch state data bound to the window to obtain the distribution of suspected discharge nodes.

[0032] Specifically, after processing in step S3, each periodic phase window has been transformed into a cavity-corrected acoustic feature tensor that integrates cavity-corrected frequency domain features, time-domain energy abrupt change features, and phase information. However, this feature tensor is still a static signal representation, failing to utilize the strong coupling relationship between partial discharge and AC voltage phase, nor combining it with the electrical topology between nodes inside the distribution box for targeted feature extraction. Traditional fixed-weight convolutional neural networks apply the same feature transformation to all input samples, failing to distinguish the essential differences in discharge modes under different phases, and struggling to differentiate the perception of electrical relationships between different nodes. To address this, this step introduces a phase-gated convolutional unit, directly injecting phase information and electrical topology information into the feature extraction process. This allows the weights of the convolutional kernel to dynamically change with the phase context of the input sample and the electrical coupling relationship between nodes, thereby achieving phase-adaptive recognition and node localization of partial discharge modes.

[0033] The aforementioned phase-gated convolutional unit is a neural network module with dynamically generated convolutional kernel weights, and its training process is jointly performed with a classification and regression network. The training data consists of cavity-corrected acoustic feature tensor samples collected and labeled under different operating conditions of the distribution box. The operating conditions cover the normal state of the distribution box under different load current levels and different combinations of switching states, as well as the abnormal state where partial discharge occurs at each node. The labeling information for each sample includes a binary label indicating whether the window contains a discharge event, and an index label for the node where the discharge source is located when a discharge event is present. During the training phase, for each input sample, the phase-gated convolutional unit calculates a dynamic convolutional kernel in real time based on its phase angle label and the node distance and electrical adjacency weights in the acoustic-electric topology matrix. It performs convolution on the input feature tensor to extract and fuse the feature map, and then the fully connected layer branches output the discharge probability and node distribution respectively. The classification branch uses a binary cross-entropy loss function, while the regression branch uses a classification cross-entropy loss function. A mask is applied to the regression error of nodes downstream of the disconnected loop to eliminate its impact on parameter updates. The total loss function is a weighted sum of the classification and regression losses, with weighting coefficients set according to the importance of the two tasks. All trainable parameters of the network are updated using the backpropagation algorithm until convergence. After training, this unit can perform forward computation on new input feature tensors during the inference phase, directly outputting the probability of abnormal discharge candidates and the distribution of suspected discharge nodes.

[0034] During inference, the phase angle label is first extracted from the cavity correction acoustic feature tensor. This label is then processed using a sine mapping function to generate a phase-gated vector. The phase angle label is a scalar appended to the last dimension of the cavity correction acoustic feature tensor; its value is a normalized value obtained by dividing the initial phase angle of the window by 360 degrees, ranging from 0 to 1. Before generating the phase-gated vector, this normalized value is multiplied by 360 degrees to restore it to an angle value, then multiplied by π and divided by 180 degrees to convert it to a radian value φ. The sine mapping function is used to convert φ into a multi-dimensional vector that comprehensively reflects the periodic characteristics of the AC voltage, including not only the fundamental frequency component but also the harmonic components, to capture phase patterns at different time scales. The phase-gated vector is designed as a four-dimensional vector, and the calculation formulas for each dimension are as follows: , , , Where g1 and g2 represent the sine and cosine components of the fundamental frequency, respectively, reflecting the voltage amplitude and polarity; g3 and g4 represent the sine and cosine components of the second harmonic, respectively, reflecting the periodicity of the voltage change rate.

[0035] Then, the phase-gated vector is used to perform tensor operations with the node distances and electrical adjacency weights in the acoustic-electric topology matrix to generate dynamic convolution kernel weights. Specifically, let the total number of electrical nodes inside the distribution box be N, and the basic convolution kernel be a kernel with a size of N. The weight matrix is ​​initialized using a random initialization method. and The values ​​are matched to the spatial dimensions after the cavity correction acoustic feature tensor reconstruction. For the i-th dynamic convolution kernel with the i-th node as the target of interest, its weight modulation coefficients... It is obtained by joint operation of phase gate vector, node distance, and electrical adjacency edge weight: .in This is the normalized distance from the i-th node to the acquisition device. The normalization method is to divide the node distance by the maximum value among all node distances. This represents the total electrical coupling strength between the i-th node and all other nodes. This represents the electrical coupling weight between the i-th node and the j-th node; to , , The modulation weights are trainable parameters, learned through backpropagation during the training phase; σ is a nonlinear activation function used to introduce nonlinear expressive power. The modulation coefficients... Multiplying the i-th dynamic convolutional kernel element-wise with the base convolutional kernel yields the final weights. This process is repeated for all N nodes, generating a total of N dynamic convolutional kernels. Each dynamic convolutional kernel corresponds to an electrical node, and its weight modulation incorporates the spatial location and electrical coupling characteristics of that node.

[0036] The aforementioned dynamic convolutional kernels are used to extract convolutional features from the cavity correction acoustic feature tensor, resulting in a fused feature map. Before being input into the convolutional layer, the cavity correction acoustic feature tensor's correction frequency band feature components are reorganized into a two-dimensional matrix. Temporal abrupt change features and phase angle labels are added as additional channels and stacked with the frequency band feature matrix. Each dynamic convolutional kernel performs a sliding convolution operation on the reorganized feature tensor with a stride of one. A modified linear unit function is applied to the convolution result for nonlinear activation, setting negative values ​​to 0. The output feature maps of all N dynamic convolutional kernels are stacked in channel dimension according to node index order to form a fused feature map. The i-th channel of this fused feature map corresponds to the dynamic convolutional kernel response of the i-th node.

[0037] Finally, the fused feature map is input into the fully connected layer for classification and regression processing, outputting the abnormal discharge candidate probability and initial node distribution. Combined with the switch state data bound to the window, the initial node distribution is constrained by a charged node mask to obtain the suspected discharge node distribution. The fused feature map is first flattened into a one-dimensional vector, then fed into a fully connected network consisting of several hidden layers. The fully connected network has two parallel output branches: a classification branch and a regression branch. The two branches share the flattened input vector and some hidden layers. The last layer of the classification branch contains two neurons, corresponding to the normal state and the abnormal discharge state, respectively. A normalized exponential function maps the output value to two probability values, where the probability value corresponding to the abnormal discharge state is the abnormal discharge candidate probability. The last layer of the regression branch contains N neurons, corresponding to the N electrical nodes inside the distribution box. A normalized exponential function maps the output value of each neuron to the discharge probability distribution of each node, serving as the initial node distribution. The charged node mask constraint refers to correcting the initial node distribution using the switch state vector bound to the current window in step S2. In step S1, when constructing the electrical connection topology, the power supply direction of each electrical connection edge is marked according to the power supply input direction. This allows us to determine the set of downstream nodes that will be de-energized when each circuit breaker is opened. The switch state vector indicates the on / off state of each circuit in binary form. For a circuit in the open state, the probability value of its corresponding downstream node set in the initial node distribution is set to zero, and the probability values ​​of the remaining nodes are renormalized so that the sum of the probabilities of all nodes is 1. This yields the distribution of suspected discharge nodes.

[0038] Through the technical solution in this step, the phase-gated convolutional unit directly integrates AC phase information and electrical topology information into the convolutional kernel weight generation process of feature extraction, realizing adaptive perception of discharge modes under different phases and differentiated response to electrical coupling relationships of different nodes. In addition, the interference of power-off nodes is eliminated by combining switch state mask constraints, thereby significantly improving the accuracy of partial discharge identification inside the distribution box and the reliability of node positioning.

[0039] S5. Based on the distribution of suspected discharge nodes, and combined with the node distance and acoustic reflection weight in the acoustic-electric topology matrix, perform acoustic propagation consistency verification on adjacent periodic phase windows, screen discharge events that conform to the physical laws of cavity acoustic propagation, and determine the abnormal discharge event chain, including: Extract the distribution of suspected discharge nodes in multiple consecutive periodic phase windows, and calculate the time difference of arrival and the frequency band attenuation slope of the same suspected discharge node between adjacent windows. Obtain the node distance and acoustic reflection weight of the corresponding suspected discharge node from the acoustic-electric topology matrix, and calculate the theoretical time difference of arrival and the theoretical frequency band attenuation slope. Calculate the first deviation value between the actual sound arrival time difference and the theoretical sound arrival time difference, and the second deviation value between the actual frequency band attenuation slope and the theoretical frequency band attenuation slope. If both deviation values ​​are less than the corresponding preset threshold, it is determined that the discharge event at this node conforms to the physical law of cavity sound propagation. Multi-cycle discharge data corresponding to suspected discharge nodes that conform to the physical laws of acoustic propagation in a cavity are collected in chronological order to generate an abnormal discharge event chain.

[0040] Specifically, the identification results of a single-cycle phase window may be affected by transient noise or random factors, posing a risk of misjudgment. Real partial discharge events are typically continuous and repetitive, occurring multiple times within consecutive AC cycles, with the spatial location of the discharge source remaining constant. By analyzing the distribution of suspected discharge nodes within consecutive cycle phase windows and using the physical laws of acoustic propagation for consistency verification, it is possible to effectively distinguish between genuine continuous discharges and occasional random noise, thereby significantly improving the reliability of identification.

[0041] First, the distribution of suspected discharge nodes is extracted from multiple consecutive periodic phase windows, and the time difference of arrival and frequency band attenuation slope of the same suspected discharge node between adjacent windows are calculated. Multiple consecutive periodic phase windows refer to several temporally adjacent half-cycle data segments; for example, five consecutive periodic phase windows can be selected, corresponding to 2.5 AC cycles. The aforementioned steps are performed on each window to obtain its respective distribution of suspected discharge nodes. For a given suspected discharge node, its discharge probability in each periodic phase window is extracted. If the discharge probability of this node remains high across multiple consecutive windows, it is considered that the node may be a real discharge source, and its acoustic propagation characteristics between adjacent windows are further analyzed.

[0042] Time difference of arrival (TDOA) refers to the difference in the time it takes for a sound wave generated by the same discharge node to reach the acquisition device between adjacent windows. Since the occurrence time of a discharge event may vary slightly within each cycle, and the sound wave propagation path may change due to environmental disturbances, the TDOA will exhibit minor fluctuations. For a true, fixed-location discharge source, the TDOA should fluctuate within a very small range. If the TDOA is too large, it may indicate that the suspected discharge node is unstable or that the signal is not a true discharge. When calculating the TDOA, within each cycle phase window, the suspected discharge occurrence time is determined based on the location of the short-time energy mutation point. For the same suspected discharge node, its discharge times in two adjacent windows are compared, and the time difference is calculated.

[0043] The band attenuation slope refers to the rate at which the energy of an acoustic signal decreases with frequency across different frequency bands. When sound waves propagate within a cavity, high-frequency components attenuate faster than low-frequency components, causing the spectrum to shift towards lower frequencies as the propagation distance increases. The band attenuation slope can be calculated by comparing the energy ratios of different frequency bands over time. For the same discharge source, since the sound wave propagation path and cavity characteristics remain unchanged, the band attenuation slope should remain consistent between adjacent windows. When calculating the band attenuation slope, the corrected frequency band features are extracted from the cavity's corrected acoustic feature tensor. The average energy of multiple high-frequency bands is then compared to the average energy of multiple low-frequency bands to obtain the high-low frequency energy ratio. For the same suspected discharge node, the high-low frequency ratios are compared between two adjacent windows, and the rate of change of this ratio is the band attenuation slope.

[0044] After calculating the actual time-of-arrival (TOA) and frequency band attenuation slope, the node distance and acoustic reflection weights of the corresponding suspected discharge nodes need to be obtained from the acoustic-electric topology matrix to calculate the theoretical TOA and theoretical frequency band attenuation slope. The physical acoustic propagation model is a theoretical model established based on the acoustic reflection weights and node spatial positions in the acoustic-electric topology matrix. This model describes the sound wave propagation time and spectral attenuation characteristics from a specific node to the acquisition device. The topological distance from the same suspected discharge node to the acquisition device is obtained from the acoustic-electric topology matrix. This distance comprehensively considers the contributions of the direct path and the main reflection path, and is obtained by reading the acoustic reflection weights corresponding to the node and inversely calculating the corresponding optimal propagation path length. Based on this topological distance and sound velocity, the theoretical TOA required for the sound wave to reach the acquisition device from the node can be calculated. For discharge nodes at fixed locations, since the sound wave propagation path is the same in adjacent periods, the theoretical TOA should remain consistent; therefore, the theoretical TOA difference should be close to 0. Considering the small fluctuations in the discharge time within the period and the changes in sound velocity due to environmental influences, a reasonable tolerance range needs to be set for the theoretical TOA difference. The theoretical band attenuation slope is calculated based on the frequency-dependent attenuation characteristics of the topological distance and propagation path. Considering the principle that the high-frequency attenuation coefficient is greater than the low-frequency attenuation coefficient when sound waves propagate in air, and the characteristic that high-frequency absorption loss is greater during wall reflection, a relationship model between the band attenuation slope and the propagation distance is established. This leads to the derivation of the theoretical high-low frequency energy ratio and the theoretical band attenuation slope. The construction process of this relationship model is as follows: using the node distance extracted from the acoustic-electric topology matrix as the basic physical quantity, combined with preset band attenuation characteristic parameters, i.e., the attenuation coefficient of the high-frequency signal is greater than that of the low-frequency signal when sound waves propagate in air... Given the characteristic that the absorption loss of high-frequency signals is more significant when reflected from a metal wall, the total attenuation of the high-frequency and low-frequency bands along the propagation path is calculated separately. Then, based on this frequency band attenuation result, the theoretical attenuation ratio of high-frequency energy to low-frequency energy in the original sound source signal is determined, thereby deriving the theoretical high-low frequency energy ratio characterizing the degree of spectral energy shift to low frequencies. Finally, this ratio is correlated with the propagation distance and frequency span to quantify the rate at which the additional attenuation of high frequencies relative to low frequencies changes with distance, thereby establishing the relationship between the frequency band attenuation slope and the propagation distance, and outputting the theoretical frequency band attenuation slope.

[0045] Subsequently, the first deviation value between the actual and theoretical time-of-arrival (TOA) time difference, and the second deviation value between the actual and theoretical frequency band attenuation slopes are calculated. The first deviation value is the absolute difference between the actual and theoretical TOA time difference; since the theoretical TOA time difference is close to 0, the first deviation value is simply the absolute value of the actual TOA time difference. The second deviation value is the relative difference between the actual and theoretical frequency band attenuation slopes, calculated by comparing the degree of difference between the actual high-low frequency energy ratio and the theoretical high-low frequency energy ratio. If both deviation values ​​are less than the corresponding preset thresholds, the discharge event at that node is determined to conform to the physical laws of cavity sound propagation.

[0046] The first preset threshold is used to determine the consistency of the time difference of arrival. Its setting needs to comprehensively consider the influence of measurement accuracy, sampling rate limitations, and environmental factors, and is usually based on a certain margin according to the system's time resolution. The second preset threshold is used to determine the consistency of the frequency band attenuation slope. Its setting needs to consider factors such as spectrum estimation error, background noise, and cavity resonance fluctuations, and is usually determined based on the normal measurement error range of the frequency band energy ratio. For multiple consecutive periodic phase windows, the first and second deviation values ​​between adjacent windows are checked one by one. Only when the deviation values ​​of all adjacent window pairs meet the threshold conditions is the acoustic propagation characteristics of the suspected discharge node consistent with the physical model throughout the entire time period, confirming that the node is a real abnormal discharge source.

[0047] Finally, the multi-cycle discharge data corresponding to suspected discharge nodes that conform to the physical laws of acoustic propagation in the cavity are collected in chronological order to generate an abnormal discharge event chain. The abnormal discharge event chain is a time-series structure that records discharge events of the corresponding node over multiple consecutive AC cycles. Each event includes information such as the time of occurrence, phase angle, discharge probability, and acoustic characteristics. Through the construction and analysis of the event chain, the development trend, severity, and possible causes of the discharge can be further assessed. For example, the event chain may show that the discharge is concentrated in a specific phase interval, conforming to the typical characteristics of a specific discharge type, while the discharge probability remains at a high level, indicating that the discharge phenomenon is stable and relatively serious. Maintenance personnel can then promptly identify nodes with potential partial discharge hazards and make a preliminary judgment on the discharge type based on the phase distribution and spectral characteristics of the discharge. This provides accurate location information and diagnostic basis for subsequent targeted maintenance, avoiding power outage losses and wasted manpower caused by blind disassembly and inspection. By continuously monitoring the evolution of the abnormal discharge event chain, the rate of insulation degradation can also be assessed, and the occurrence time of potential faults can be predicted, realizing a shift from reactive emergency repairs to proactive prevention.

[0048] S5 further includes: determining whether there are multiple nodes in the suspected discharge node distribution whose discharge probability exceeds a preset multi-source determination threshold; if so, obtaining the theoretical arrival time and theoretical frequency band attenuation characteristics of each candidate node, performing temporal deconvolution on the acoustic signal of the current window based on the theoretical arrival time, and separating the acoustic contribution components corresponding to each candidate node. The acoustic contribution components of the candidate nodes other than the target candidate node are subtracted from the original acoustic signal to obtain the residual signal of the target candidate node; the time difference of arrival and the frequency band attenuation slope are recalculated on the residual signal, and it is verified whether they conform to the cavity acoustic propagation physical law of the corresponding node in the acoustic-electric topology matrix.

[0049] Specifically, in the actual operation of a distribution box, complex situations may arise where multiple electrical nodes experience partial discharge simultaneously or at similar times. When multiple nodes in the suspected discharge node distribution output in step S4 have a high probability of discharge, directly performing propagation consistency verification on each node independently may be subject to aliasing interference from acoustic signals of other discharge sources, leading to deviations in the calculation of the time of arrival and the frequency band attenuation slope, thus affecting the accuracy of the verification results. To address this, this step introduces a time-domain deconvolution separation mechanism based on theoretical time of arrival, using multi-source signal separation and node-by-node residual verification to solve the acoustic signal aliasing problem in multi-discharge scenarios.

[0050] In practice, the process first determines whether multiple nodes in the suspected discharge node distribution have a discharge probability exceeding a preset multi-source determination threshold. This threshold distinguishes between single-discharge and multi-discharge scenarios, and its setting must consider both the confidence distribution characteristics of the model output and the number of nodes within the distribution box. When two or more nodes in the suspected discharge node distribution have a discharge probability exceeding this threshold, it is determined that multiple candidate discharge sources exist within the current period's phase window, triggering the multi-source separation and verification process.

[0051] After identifying multiple candidate nodes, the theoretical time of arrival (TOA) and theoretical band attenuation characteristics of each candidate node are obtained. For each candidate node, the acoustic reflection weights from the node to the acquisition device are read from the acoustic-electric topology matrix, and the corresponding optimal propagation path length is extracted as the node distance. Based on the node distance and sound velocity, the theoretical TOA required for the sound wave to propagate from the node to the acquisition device is calculated. Simultaneously, based on the node distance and the frequency-dependent attenuation characteristics of the propagation path—that is, the physical law that the high-frequency attenuation coefficient is greater than the low-frequency attenuation coefficient when the sound wave propagates in air, and that high-frequency absorption loss is greater when reflected from walls—the theoretical band attenuation characteristics of the corresponding node are calculated, including the theoretical high-low frequency energy ratio and the theoretical band attenuation slope. Therefore, an acoustic propagation parameter template containing theoretical TOA and theoretical band attenuation characteristics is established for each candidate node.

[0052] Next, the acoustic signal in the current window is deconvolved in the time domain based on the theoretical arrival time of each candidate node to separate the acoustic contribution components corresponding to each candidate node. The purpose of time-domain deconvolution is to separate the aliased acoustic signal according to the theoretical arrival time of each sound source, and to restore the acoustic response generated at the acquisition device when each candidate discharge source acts alone. Specifically, the original acoustic signal in the current period phase window is used as the observation signal, and the theoretical arrival time of each candidate node is used as the known time delay parameter to construct a time-domain model of the multi-source mixed signal. In this model, the acoustic signal received by the acquisition device is represented as a linear superposition of the discharge sound source signals of each candidate node after attenuation and time delay through their respective propagation paths. Using deconvolution algorithms, such as time-domain deconvolution based on the least squares criterion or iterative deconvolution based on sparse constraints, the sound source signal component corresponding to each candidate node, i.e., the acoustic contribution component, is estimated from the observation signal. This acoustic contribution component characterizes the independent contribution of a specific candidate node to the acquired signal in the current window.

[0053] After separating the acoustic contribution components, each candidate node is independently verified using a node-by-node residual verification method. For a target candidate node, the acoustic contribution components of all other candidate nodes are subtracted from the original acoustic signal to obtain the residual signal of the target candidate node. Interference components from other discharge sources are then removed from the mixed signal, retaining only the acoustic information relevant to the target candidate node. The residual signal should theoretically primarily contain the discharge acoustic characteristics of the target candidate node and background noise. For example, when verifying the main circuit breaker, the acoustic contribution components of the branch circuit breakers and terminal blocks are subtracted from the original acoustic signal, and the resulting residual signal is the approximate independent acoustic response of the main circuit breaker.

[0054] After obtaining the residual signal, the time of arrival (TOA) and frequency band attenuation slope are recalculated, and their conformity with the cavity acoustic propagation physics of the corresponding node in the aforementioned acoustic-electric topology matrix is ​​verified. The specific verification process is consistent with the single-amplifier power supply scenario: short-time energy mutation localization is re-executed on the residual signal to determine the discharge time, and the actual TOA between the residual signal and the adjacent period phase window is calculated; the high-low frequency energy ratio is extracted from the corrected frequency band characteristics of the residual signal, and the actual frequency band attenuation slope is calculated. Subsequently, the node distance and acoustic reflection weight of the target candidate node are obtained from the acoustic-electric topology matrix, and the theoretical TOA and theoretical frequency band attenuation slope are calculated. The first and second deviation values ​​between the actual and theoretical values ​​are calculated. If both deviation values ​​are less than the corresponding preset thresholds, the discharge event of the target candidate node is determined to conform to the cavity acoustic propagation physics. If the deviation value exceeds the threshold, it indicates that although the discharge probability of the candidate node is high, its acoustic propagation characteristics do not match the physical model, and it may be crosstalk response from other node discharges or random noise interference.

[0055] By performing the aforementioned residual verification on each candidate node that exceeds the multi-source determination threshold, it is possible to effectively distinguish between genuine simultaneous discharge from multiple sources and high-probability multi-point responses caused by a single source, thus avoiding missed detections or misjudgments due to signal aliasing. Ultimately, only candidate nodes that are confirmed to conform to the physical propagation laws through residual verification will have their corresponding multi-cycle discharge data included in the abnormal discharge event chain.

[0056] This invention also provides a partial discharge audio recognition system inside a distribution box, used to implement the above-mentioned method, such as... Figure 3 As shown, the system includes: The acoustic-electric topology construction unit is used to obtain the physical structure parameters and internal electrical connection relationships of the distribution box, determine the spatial position of the internal electrical nodes, calculate the acoustic attenuation coefficient, electrical coupling weight and node distance between nodes, and generate the acoustic-electric topology matrix. The phase window division unit is used to acquire the AC periodic synchronization signal, extract its zero crossover point as the starting reference, divide the acquired internal acoustic signal into acoustic data segments of half-cycle length, attach phase angle labels to each acoustic data segment, and form a periodic phase window sequence. The cavity feature correction unit is used to calculate the short-time energy envelope and locate the energy mutation point of the periodic phase window sequence, extract the frequency band features of the Mel filter bank, use the acoustic reflection side weight of the acoustic-electric topology matrix to perform reflection gain correction on each frequency band feature, and generate the cavity correction acoustic feature tensor by combining the energy mutation point and the phase angle label. The phase-gated convolutional unit is used to input the cavity correction acoustic feature tensor into the phase-gated convolutional unit, and modulate the convolutional kernel weights by using the phase angle label and the node distance and electrical adjacency edge weight in the acoustic-electric topology matrix, and output the abnormal discharge candidate probability and the distribution of suspected discharge nodes. The propagation consistency verification unit is used to perform acoustic propagation consistency verification on adjacent periodic phase windows based on the distribution of suspected discharge nodes and the node distance and acoustic reflection weight in the acoustic-electric topology matrix, to screen discharge events that conform to the physical laws of acoustic propagation in the cavity, and to determine the abnormal discharge event chain.

[0057] The present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the above-described method.

[0058] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0059] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0060] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for partial discharge audio recognition inside a distribution box, characterized in that, The method includes: S1. Obtain the physical structure parameters and internal electrical connection relationships of the distribution box, determine the spatial location of the internal electrical nodes, calculate the acoustic attenuation coefficient, electrical coupling weight and node distance between nodes, and generate an acoustic-electric topology matrix. S2. Acquire the AC periodic synchronization signal, extract its zero crossover point as the starting reference, divide the acquired internal acoustic signal into half-cycle length acoustic data segments, add phase angle labels to each acoustic data segment, and form a periodic phase window sequence. S3. Calculate the short-time energy envelope and locate the energy mutation point for the periodic phase window sequence, extract the frequency band features of the Mel filter bank, use the acoustic reflection side weights of the acoustic-electric topology matrix to perform reflection gain correction on each frequency band feature, and generate the cavity correction acoustic feature tensor by combining the energy mutation point and the phase angle label. S4. Input the cavity correction acoustic feature tensor into the phase-gated convolutional unit, and modulate the convolutional kernel weights by using the phase angle label and the node distance and electrical adjacency edge weights in the acoustic-electric topology matrix, and output the abnormal discharge candidate probability and the distribution of suspected discharge nodes. S5. Based on the distribution of the suspected discharge nodes, and combined with the node distance and sound reflection weight in the acoustic-electric topology matrix, perform acoustic propagation consistency verification on adjacent periodic phase windows, screen out discharge events that conform to the physical laws of cavity sound propagation, and determine the abnormal discharge event chain.

2. The method according to claim 1, characterized in that, Generate the acoustic-electric topology matrix, including: Read the spatial coordinates and voltage levels of each electrical node inside the distribution box, and combine the box cavity dimensions and the position and material of the internal partitions to generate multiple physical spatial sound propagation paths between the nodes. Based on the length and medium properties of each physical space acoustic propagation path, the acoustic attenuation coefficient of the corresponding path is calculated. Based on the charge level and electrical connection relationship of the internal electrical nodes, the electrical coupling weight between the nodes is calculated. The acoustic propagation path with the smallest acoustic attenuation coefficient is selected as the optimal path, and the length of the optimal path is extracted as the node distance. The acoustic attenuation coefficient corresponding to the optimal path is used as the acoustic reflection edge weight, and the electrical coupling weight is used as the electrical adjacency edge weight. Combined with the node distance, an acoustic-electric topology matrix is ​​constructed.

3. The method according to claim 1, characterized in that, Forming a periodic phase window sequence includes: Acquire the AC periodic synchronization signal from the incoming side of the distribution box and the continuously acquired internal acoustic signal, and extract the zero crossover point of the AC periodic synchronization signal as the segmentation starting reference. Using the zero-crossing point as the time node, the continuous internal acoustic signal is divided into multiple half-cycle length acoustic data segments, and a phase angle label is added according to the phase interval corresponding to each acoustic data segment in the AC cycle. The real-time load current data and switch status data of the distribution box are acquired, and the real-time load current data, switch status data and the phase angle label of the corresponding acoustic data segment are bound together to the acoustic data segment to form a periodic phase window sequence.

4. The method according to claim 1, characterized in that, Generate the cavity correction acoustic feature tensor, including: Each window in the periodic phase window sequence is traversed, and the energy mutation threshold is dynamically set in combination with the real-time load current data bound to the window. The frame energy of the acoustic signal in the window is calculated frame by frame to generate a short-time energy envelope. The time point when the energy change rate exceeds the energy mutation threshold is extracted as the short-time energy mutation point. The acoustic signal in each periodic phase window is frequency domain transformed to extract the frequency band features of the Mel filter bank. Based on the acoustic reflection side weights in the acoustic-electric topology matrix, and combined with real-time load current data, attenuation compensation is performed on low-frequency noise. The reflection gain compensation amount corresponding to each frequency band is calculated, and the energy of each frequency band of the Mel filter bank frequency band characteristics is weighted and compensated to obtain the corrected frequency band characteristics. The corrected frequency band characteristics, the feature parameters of short-time energy mutation points, and the phase angle labels of the corresponding windows are concatenated to generate the cavity corrected acoustic feature tensor.

5. The method according to claim 1, characterized in that, The probability of abnormal discharge candidates and the distribution of suspected discharge nodes include: Phase angle labels are extracted from the cavity correction acoustic feature tensor, and the phase angle labels are processed by a sine mapping function to generate a phase gate vector. The phase gate vector is then used to perform tensor operations with the node distances and electrical adjacency weights in the acoustic-electric topology matrix to generate dynamic convolution kernel weights. The dynamic convolution kernels are then used to perform convolution feature extraction on the cavity correction acoustic feature tensor to obtain a fused feature map. The fused feature map is input into a fully connected layer for classification and regression processing, and the abnormal discharge candidate probability and initial node distribution are output. The initial node distribution is then constrained by a charged node mask based on the switch state data bound to the window to obtain the distribution of suspected discharge nodes.

6. The method according to claim 1, characterized in that, Generate an abnormal discharge event chain, including: Extract the distribution of suspected discharge nodes in multiple consecutive periodic phase windows, and calculate the time difference of arrival and the frequency band attenuation slope of the same suspected discharge node between adjacent windows. Obtain the node distance and acoustic reflection weight of the corresponding suspected discharge node from the acoustic-electric topology matrix, and calculate the theoretical time difference of arrival and the theoretical frequency band attenuation slope. Calculate the first deviation value between the actual sound arrival time difference and the theoretical sound arrival time difference, and the second deviation value between the actual frequency band attenuation slope and the theoretical frequency band attenuation slope. If both deviation values ​​are less than the corresponding preset threshold, it is determined that the discharge event at this node conforms to the physical law of cavity sound propagation. Multi-cycle discharge data corresponding to suspected discharge nodes that conform to the physical laws of acoustic propagation in a cavity are collected in chronological order to generate an abnormal discharge event chain.

7. The method according to claim 6, characterized in that, S5 also includes: Determine whether there are multiple nodes in the suspected discharge node distribution whose discharge probability exceeds a preset multi-source judgment threshold; if so, obtain the theoretical time of arrival and theoretical frequency band attenuation characteristics of each candidate node, and perform temporal deconvolution on the acoustic signal of the current window based on the theoretical time of arrival to separate the acoustic contribution components corresponding to each candidate node. The acoustic contribution components of the candidate nodes other than the target candidate node are subtracted from the original acoustic signal to obtain the residual signal of the target candidate node; the time difference of arrival and the frequency band attenuation slope are recalculated on the residual signal, and it is verified whether they conform to the cavity acoustic propagation physical law of the corresponding node in the acoustic-electric topology matrix.

8. The method according to claim 1, characterized in that, The abnormal discharge event chain is a time series structure, which records the discharge events of the corresponding node in multiple consecutive AC cycles. Each event includes at least the time of occurrence, phase angle, discharge probability, and acoustic characteristics.

9. A partial discharge audio recognition system for an internal distribution box, used to implement the method as described in any one of claims 1-8, characterized in that, The system includes: The acoustic-electric topology construction unit is used to obtain the physical structure parameters and internal electrical connection relationships of the distribution box, determine the spatial position of the internal electrical nodes, calculate the acoustic attenuation coefficient, electrical coupling weight and node distance between nodes, and generate the acoustic-electric topology matrix. The phase window division unit is used to acquire the AC periodic synchronization signal, extract its zero crossover point as the starting reference, divide the acquired internal acoustic signal into acoustic data segments of half-cycle length, attach phase angle labels to each acoustic data segment, and form a periodic phase window sequence. The cavity feature correction unit is used to calculate the short-time energy envelope and locate the energy mutation point of the periodic phase window sequence, extract the frequency band features of the Mel filter bank, use the acoustic reflection side weight of the acoustic-electric topology matrix to perform reflection gain correction on each frequency band feature, and generate the cavity correction acoustic feature tensor by combining the energy mutation point and the phase angle label. The phase-gated convolutional unit is used to input the cavity correction acoustic feature tensor into the phase-gated convolutional unit, and modulate the convolutional kernel weights by using the phase angle label and the node distance and electrical adjacency edge weight in the acoustic-electric topology matrix, and output the abnormal discharge candidate probability and the distribution of suspected discharge nodes. The propagation consistency verification unit is used to perform acoustic propagation consistency verification on adjacent periodic phase windows based on the distribution of suspected discharge nodes and the node distance and acoustic reflection weight in the acoustic-electric topology matrix, to screen discharge events that conform to the physical laws of acoustic propagation in the cavity, and to determine the abnormal discharge event chain.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the method as described in any one of claims 1-8.

Images