A high-voltage partial discharge monitoring method and system based on edge AI voiceprint recognition

Abstract

The application relates to the technical field of high-voltage partial discharge monitoring, and discloses a high-voltage partial discharge monitoring method and system based on edge AI voiceprint recognition, which comprises the following steps: establishing an electromagnetic fingerprint baseline database according to field electromagnetic signals collected in the initial stage of high-voltage equipment operation; collecting online electromagnetic signals of template high-voltage equipment in the running process in real time, comparing the online electromagnetic signals with the electromagnetic fingerprint baseline database, and obtaining a suspected partial discharge signal set; performing time-frequency domain feature deep extraction on each signal in the suspected partial discharge signal set on the edge side, and classifying and identifying the signals according to the time-frequency domain features, so as to distinguish real partial discharge signals from various interference signals; generating alarm information and reporting only when the classification and identification result is confirmed as a real partial discharge signal; and updating the newly confirmed interference signal features to the electromagnetic fingerprint baseline database at a preset period; and the application can improve the recognition degree of high-voltage partial discharge identification.

Description

Technical Field

[0001] This invention relates to the field of high-voltage partial discharge monitoring technology, and in particular to a high-voltage partial discharge monitoring method and system based on edge AI voiceprint recognition. Background Technology

[0002] In the field of online monitoring of partial discharge in high-voltage equipment, voiceprint recognition-based diagnostic methods are one of the mainstream technical approaches. These methods typically deploy voiceprint acquisition devices at the equipment side to obtain operating acoustic signals. After preprocessing such as filtering and noise reduction, time-domain, frequency-domain, or time-frequency-domain features are extracted. Then, relying on a pre-built voiceprint feature database or a trained recognition model, the presence of partial discharge is determined through feature matching or pattern classification. Some solutions further incorporate edge computing technology, moving feature extraction and preliminary diagnostic capabilities to the terminal device to reduce data transmission latency. The effective operation of these methods highly depends on a relatively complete and stable voiceprint feature database or diagnostic model. The construction of the database and model is usually based on field data collected during the initial operation of the equipment or laboratory simulation data.

[0003] However, the operating environment of high-voltage equipment is not static. For example, the composition of environmental noise, equipment load conditions, and various electromagnetic interferences all change continuously over time and under operating conditions. Under such dynamic and complex conditions, methods based on fixed databases and static models are difficult to effectively adapt to the constantly evolving interference environment. On the one hand, if newly emerging interference patterns are not pre-recorded in the database or model, they are easily misjudged as partial discharge signals, leading to an increased false alarm rate. On the other hand, the natural drift of equipment operating conditions may also cause normally occurring signal characteristics to deviate from the baseline range in the database, resulting in missed alarms. This degradation in recognition performance due to the lack of dynamic updating capabilities of the database and model is the fundamental problem restricting the long-term stable operation of existing methods. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides a high-voltage partial discharge monitoring method based on edge AI voiceprint recognition, comprising: Step 110: Establish an electromagnetic fingerprint baseline database based on the on-site electromagnetic signals collected during the initial commissioning of the high-voltage equipment; Step 120: Real-time acquisition of online electromagnetic signals during the operation of the template high-voltage equipment, and comparison of the online electromagnetic signals with the electromagnetic fingerprint baseline database to obtain a set of suspected partial discharge signals; Step 130: On the edge side, perform time-frequency domain feature extraction on each signal in the suspected partial discharge signal set, and classify and identify the signals according to the time-frequency domain features to distinguish between real partial discharge signals and various interference signals; Step 140: Only when the classification and identification result confirms that it is a real partial discharge signal, generate alarm information and report it; Step 150: Update the newly confirmed interference signal features to the electromagnetic fingerprint baseline database at a preset period.

[0005] To address the aforementioned problems, this invention also provides a high-voltage partial discharge monitoring system based on edge AI voiceprint recognition, the system comprising: The fingerprint filing module is used to establish an electromagnetic fingerprint baseline database based on the on-site electromagnetic signals collected during the initial commissioning of high-voltage equipment. The signal comparison module is used to collect online electromagnetic signals during the operation of the template high-voltage equipment in real time, and compare the online electromagnetic signals with the electromagnetic fingerprint baseline database to obtain a set of suspected partial discharge signals; The feature discrimination module is used to perform time-frequency domain feature extraction on each signal in the suspected partial discharge signal set at the edge, and classify and identify the signals according to the time-frequency domain features to distinguish between real partial discharge signals and various interference signals. The anomaly alarm module is used to generate and report alarm information only when the classification and identification results confirm that it is a real partial discharge signal; The baseline iteration module is used to update the electromagnetic fingerprint baseline database with newly confirmed interference signal features at a preset period.

[0006] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs an electromagnetic fingerprint baseline database based on the electromagnetic signals from the initial stage of high-voltage equipment commissioning. It combines this database with real-time online signal comparison and screening, in-depth extraction and classification of edge-side time-frequency domain features, and a closed-loop mechanism that updates the baseline database with newly confirmed interference signal features at preset intervals. This achieves the system's ability to maintain the adaptability of the identification benchmark in a dynamic operating environment. This solution fundamentally overcomes the problem described in the prior art where fixed databases and static models are difficult to adapt to the continuously changing field interference environment and equipment operating status. It ensures that the identification benchmark is no longer fixed in the limited state of the initial commissioning stage but evolves synchronously with the operation process, providing a fundamental guarantee for the long-term stable differentiation between real partial discharge signals and various interference signals.

[0007] 2. This invention compares the amplitude fluctuation of each frequency component in the same time-frequency domain deep feature vector across different sub-signal segments, and deletes frequency components with fluctuations exceeding a set range as interfering feature components, thus obtaining a purified time-frequency domain deep feature vector. This additional processing ensures that the feature components participating in classification and recognition have stable consistency across time windows, further improving the accuracy of feature template matching and the confidence of the recognition results. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating a high-voltage partial discharge monitoring method based on edge AI voiceprint recognition, provided in an embodiment of the present invention. Figure 2 This is a functional block diagram of a high-voltage partial discharge monitoring system based on edge AI voiceprint recognition, provided in an embodiment of the present invention. Figure 3 A 24-hour time-domain amplitude variation curve of a normal electromagnetic fingerprint baseline database for a high-voltage partial discharge monitoring system based on edge AI voiceprint recognition is provided as an embodiment of the present invention. Figure 4 This is a diagram illustrating a real-time electromagnetic signal online comparison and dynamic over-limit judgment mechanism for a high-voltage partial discharge monitoring system based on edge AI voiceprint recognition, provided in an embodiment of the present invention. Figure 5 This invention provides a refined multi-category interference attribution graph based on deviation direction vector space for a high-voltage partial discharge monitoring system based on edge AI voiceprint recognition, as an embodiment of the present invention. Figure 6 This is a diagram showing the self-evolution and false alarm rate degradation trend of a closed-loop iterative adaptive evolution feature library for a high-voltage partial discharge monitoring system based on edge AI voiceprint recognition, provided as an embodiment of the present invention.

[0009] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0010] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0011] This application provides a high-voltage partial discharge monitoring method based on edge AI voiceprint recognition. The executing entity of this high-voltage partial discharge monitoring method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the high-voltage partial discharge monitoring method based on edge AI voiceprint recognition can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0012] Reference Figure 1The diagram shown is a flowchart illustrating a high-voltage partial discharge monitoring method based on edge AI voiceprint recognition according to an embodiment of the present invention. In this embodiment, the high-voltage partial discharge monitoring method based on edge AI voiceprint recognition includes: Step 110: Establish an electromagnetic fingerprint baseline database based on the on-site electromagnetic signals collected during the initial commissioning of the high-voltage equipment; In the embodiments of this application, the high-voltage equipment is in relatively good insulation condition and the operating conditions are relatively stable in the initial stage of commissioning. The electromagnetic environment at the site is not significantly affected by equipment aging or sudden failure. Therefore, the electromagnetic signals collected in this stage can more realistically reflect the background characteristics of electromagnetic radiation of the equipment under normal operating conditions.

[0013] However, the electromagnetic environment at the actual operating site of high-voltage equipment is not constant. As the equipment operates over time, fluctuations in load current, changes in ambient temperature and humidity, and the start-up and shutdown of other surrounding power equipment will all cause a slow drift in the spectrum distribution and time-domain waveform amplitude of the electromagnetic signal.

[0014] If a limited number of samples collected at a specific moment are used as the sole criterion for subsequent judgment, the criterion will gradually deviate from the current actual operating state of the device over time, resulting in a significant decrease in the accuracy of partial discharge identification.

[0015] Therefore, this application proposes to construct an electromagnetic fingerprint baseline database that can characterize the statistical laws of electromagnetic signals under normal operating conditions of equipment by collecting data in different time periods and extracting features in multiple dimensions.

[0016] In one feasible implementation, the acquisition period of the on-site electromagnetic signal is preferably 7 to 30 consecutive calendar days, and the sampling rate is preferably set to 100 MS / s to 1 GS / s to cover the main spectral range of the partial discharge electromagnetic signal.

[0017] The sampling rate is selected based on the Nyquist sampling theorem, which states that the sampling frequency should be at least twice the highest frequency component of the acquired signal. For electromagnetic signals generated by partial discharge, their energy is mainly concentrated in the ultra-high frequency band of 300MHz to 1.5GHz. Therefore, setting the sampling rate to 100MS / s to 1GS / s can ensure effective acquisition of signal components in this frequency band without generating spectral aliasing.

[0018] The acquired raw electromagnetic signals are converted from analog to digital and stored in digital form. The amplitude quantization accuracy of each sampling point is preferably 12 bits to 16 bits, so as to control the amount of data storage while ensuring the dynamic range of the signal.

[0019] By configuring the above acquisition parameters, we can obtain a raw signal dataset that reflects the background characteristics of the electromagnetic environment during the initial operation of the equipment.

[0020] In one feasible implementation, the acquired field electromagnetic signals are divided into multiple time-segment signal groups according to the acquisition time sequence.

[0021] The basis for dividing the time period is that the operating status of high-voltage equipment and the on-site electromagnetic environment have regular changes on a daily cycle—for example, the load is higher during the day and lower at night, and there are also significant differences in the load curves between weekdays and holidays.

[0022] If signals from all time periods are processed together, the difference between the normal electromagnetic radiation characteristics under high load conditions during the day and the characteristics under low load conditions at night may be misjudged as abnormal signals, thus introducing unnecessary false alarms.

[0023] Therefore, in one feasible implementation, the continuously acquired electromagnetic signals are divided into 24 time period groups, with each hour serving as a time period. In another feasible implementation, considering that the actual load curve of the substation usually shows peak and valley values ​​at several fixed times each day, it can also be divided non-uniformly according to the actual operating conditions of the equipment. For example, the day can be divided into three time period groups: peak load period, off-peak load period, and low-peak load period.

[0024] In another feasible implementation, the length of the time window for dividing the time period can be adjusted according to the equipment type and the stability of the on-site electromagnetic environment. For example, for gas-insulated substation (GIS) equipment with a relatively stable electromagnetic environment, a time period of 2 hours can be used; for open equipment with a more drastic change in electromagnetic environment, a time period of 30 minutes can be used.

[0025] Regardless of which time period division method is used, each original signal sample is classified into the corresponding time period signal group based on the collection timestamp.

[0026] In one feasible implementation, time-domain waveform amplitude variation features and frequency-domain energy concentration band features are extracted for each time period signal group.

[0027] The extraction method for the amplitude variation characteristics of the time-domain waveform is as follows: perform time-domain statistical analysis on each signal sample in the signal group of the time period, calculate the peak value, root mean square value and standard deviation of each signal sample, and calculate the mean and variance of the above statistics in units of time period as the time-domain amplitude statistical characteristics of that time period.

[0028] The extraction method for frequency domain energy concentration band features is as follows: perform Fast Fourier Transform on each signal sample within the signal group of the time period to convert the time domain signal to the frequency domain, and calculate the energy amplitude corresponding to each frequency component; then accumulate the energy amplitude from low to high frequency, and determine the frequency interval corresponding to when the accumulated energy reaches 90% of the total energy as the energy concentration band of the signal sample; and statistically analyze the mean and standard deviation of the upper and lower limit frequencies of the energy concentration band of each signal sample in a time period as the frequency domain energy concentration band feature of that time period.

[0029] By extracting features in both the time and frequency domains as described above, the background characteristics of electromagnetic signals of the equipment under normal operating conditions can be characterized from two aspects: amplitude statistical regularity and spectrum distribution regularity.

[0030] In one feasible implementation, the time-domain waveform amplitude variation features and frequency-domain energy concentration band features extracted for each time period are stored as the baseline sub-library for that time period.

[0031] Specifically, the baseline sub-library for each time period includes the time period identifier (e.g., "1st hour" or "peak load period"), the time domain amplitude statistical feature vector (including the mean and variance of peak-to-peak values, the mean and variance of root mean square values, and the mean and variance of standard deviations), and the frequency domain energy concentration band feature vector (including the mean and variance of the lower limit frequency and the mean and variance of the upper limit frequency of the energy concentration band).

[0032] The baseline sub-libraries for each time period are stored in a structured data format on the local storage medium of the edge computing nodes to facilitate rapid retrieval and comparison in subsequent steps.

[0033] By storing the features from different time periods as independent baseline sub-libraries, the confusion between normal signal features under different operating conditions is avoided, providing a fine-grained comparison benchmark for subsequent screening of suspected partial discharge signals.

[0034] In one feasible implementation, after initially establishing the baseline sub-databases for each time period, the method also includes a step of performing cross-time period consistency comparison on the same type of features in the baseline sub-databases for each time period.

[0035] The purpose of this step is to eliminate abnormal sub-databases caused by factors such as abnormal acquisition equipment, sudden strong external interference, or abnormal equipment operation, so as to ensure the purity of the baseline database.

[0036] Specifically, for the time-domain amplitude statistical characteristics, the overall mean and overall standard deviation of the peak-to-peak average values ​​in the baseline sub-database for each time period are calculated, and the sub-database for each time period whose peak-to-peak average value deviates from the overall mean by more than 3 times the overall standard deviation is marked as an abnormal sub-database and removed; for the frequency-domain energy concentration band characteristics, the overall mean and overall standard deviation of the upper limit frequency of the energy concentration band in the baseline sub-database for each time period are calculated, and the sub-database for each time period whose peak-to-peak average value deviates from the overall mean by more than 3 times the overall standard deviation is marked as an abnormal sub-database and removed.

[0037] In another feasible implementation, the threshold for judging the abnormal sub-database can also be set to 2.5 times the standard deviation or 2 times the standard deviation, depending on the actual data distribution. The specific threshold can be preset by the operator based on a large number of experiments or statistical results.

[0038] Through the above cross-time period consistency comparison and anomaly removal process, contaminated data introduced by occasional interference or collection anomalies can be effectively filtered out, ensuring that the baseline sub-libraries of each time period can truly reflect the electromagnetic signal characteristics of the equipment under normal operating conditions.

[0039] In one feasible implementation, the baseline sub-databases for each time period are arranged and combined in chronological order to form a complete electromagnetic fingerprint baseline database.

[0040] The database uses time period identifiers as indexes, and each time period entry stores the time-domain amplitude statistical characteristics and frequency-domain energy concentration band characteristics corresponding to that time period.

[0041] Once the electromagnetic fingerprint baseline database is built, it is stored in the non-volatile storage medium of the edge computing node for online signal comparison during subsequent real-time monitoring.

[0042] The electromagnetic fingerprint baseline database constructed in step 110 above enables a comprehensive and fine-grained characterization of electromagnetic signal features under normal operating conditions of high-voltage equipment, providing a reliable data foundation for the real-time online signal and baseline feature comparison in subsequent steps.

[0043] Step 120: Real-time acquisition of online electromagnetic signals during the operation of the template high-voltage equipment, and comparison of the online electromagnetic signals with the electromagnetic fingerprint baseline database to obtain a set of suspected partial discharge signals; In this embodiment of the invention, after the construction of the electromagnetic fingerprint baseline database is completed, it is necessary to continuously collect and analyze the online electromagnetic signals during the actual operation of the high-voltage equipment in real time.

[0044] During long-term operation, high-voltage equipment may experience partial discharge due to manufacturing defects, mechanical vibration, thermal aging, or moisture. These partial discharge signals are usually superimposed on the electromagnetic background noise generated during normal operation of the equipment in the form of pulses. They have relatively small amplitudes and short durations, but they have specific time-domain waveform characteristics and frequency-domain energy distribution characteristics.

[0045] However, various electromagnetic signals from sources other than partial discharge also exist at the equipment operation site—including operating noise generated by the switching action of power electronic devices, radiation interference generated by corona discharge of adjacent high-voltage equipment, and external electromagnetic disturbances such as wireless communication signals. Some characteristics of these interference signals in the time and frequency domains are similar to those of real partial discharge signals. If all of them are reported without distinction, it will lead to a large number of false alarms, which will seriously reduce the reliability of the monitoring system and the efficiency of operation and maintenance.

[0046] To this end, this application proposes to compare online electromagnetic signals with the corresponding time-period baseline features in the electromagnetic fingerprint baseline database item by item, first filtering out suspected signals that significantly deviate from normal operation characteristics from massive real-time signals, thereby narrowing the scope of objects for subsequent in-depth identification.

[0047] In one feasible implementation, the sampling rate of the real-time online electromagnetic signal acquisition is consistent with the sampling rate used when establishing the baseline database in step 110, preferably set to 100MS / s to 1GS / s, and the amplitude quantization accuracy of the sampling point is preferably 12 bits to 16 bits, so as to ensure that the online acquisition signal is comparable to the historical signal in the baseline database.

[0048] Real-time data acquisition is performed by the front-end acquisition unit controlled by the edge computing node deployed near the high-voltage equipment. The acquisition unit preferably uses a high-frequency current transformer or an ultra-high-frequency antenna as a sensor and is installed on the grounding lead of the high-voltage equipment or at the opening of the equipment shell to effectively couple the electromagnetic wave signal generated by partial discharge.

[0049] The edge computing node continuously receives the digital signal stream output by the acquisition unit. The duration of each frame of data is preferably 1 to 10 milliseconds, and the overlap rate between frames is preferably maintained at 50% to avoid feature loss due to the signal being located at the frame boundary.

[0050] The signals acquired in real time are digitized and temporarily stored in the buffer of the edge computing node, awaiting subsequent comparison with the baseline database.

[0051] The above real-time acquisition configuration ensures that the online signal remains consistent with the baseline database in terms of time resolution, spectrum coverage, and amplitude quantization accuracy, laying the foundation for the accuracy of subsequent difference comparisons.

[0052] In one feasible implementation, before comparing the online electromagnetic signal with the electromagnetic fingerprint baseline database, the time period to which the current online electromagnetic signal belongs is first determined based on the acquisition time of the signal.

[0053] Since the electromagnetic fingerprint baseline database has already divided the historical signals into multiple time period baseline sub-databases according to the acquisition time in step 110, and each time period baseline sub-database corresponds to the normal signal characteristics within a specific time period, the acquisition time of the online signal determines which time period baseline sub-database it should be compared with.

[0054] Specifically, the edge computing node reads the acquisition timestamp of the current online signal frame, extracts the time information in the timestamp—for example, if the acquisition time is 14:35:22 on a certain day, then the signal belongs to the category of "14th time period" or "peak load period"—and then uses the time period identifier as an index to retrieve the corresponding time period baseline sub-library from the electromagnetic fingerprint baseline database.

[0055] In another feasible implementation, considering that the actual operating conditions of the equipment are not only related to the clock time, but also to operating parameters such as equipment load rate and ambient temperature, when determining the time period, the equipment load rate data at the current time can also be comprehensively referenced, and the online signal can be classified into the historical time period baseline sub-library closest to the load rate for comparison. For example, when the current load rate of the equipment is 80% of the rated load, the time period baseline sub-library in the historical database with a load rate between 75% and 85% is preferentially retrieved as the comparison benchmark.

[0056] In another feasible implementation, both clock-based and load-rate-based time-period matching methods can be retained simultaneously, with the edge computing node dynamically selecting the matching strategy based on the current device operating status.

[0057] Through the above-mentioned time period matching process, it is ensured that the online electromagnetic signal is compared with the historical normal signal under the same operating conditions, thus avoiding the misjudgment of the normal signal deviation caused by the difference in operating conditions as abnormal.

[0058] In one feasible implementation, time-domain amplitude fluctuation features and frequency-domain dominant frequency offset features are extracted from the online electromagnetic signal, and the degree of difference between the extracted features and the corresponding baseline features in the retrieved time-period baseline sub-library is determined.

[0059] The method for extracting the time-domain amplitude fluctuation characteristics of online signals is as follows: statistical analysis is performed on the amplitude of each sampling point within the current signal frame, and the peak-to-peak value, root mean square value, and standard deviation of the signal in that frame are calculated, which are denoted as the online time-domain feature vector. .

[0060] The method for extracting the dominant frequency offset feature of the online signal in the frequency domain is as follows: Perform a Fast Fourier Transform on the current signal frame, calculate the energy amplitude of each frequency component, and record the three frequency components with the largest energy amplitudes as the dominant frequency components. Record their frequency values ​​and corresponding energy amplitudes, and denote them as the online frequency domain feature vector. .

[0061] After extracting the above online features, the difference between the online time-domain feature vector and the mean of the time-domain amplitude statistical features in the retrieved time-domain baseline sub-library is calculated item by item, and the deviation between each main frequency value in the online frequency-domain feature vector and the mean of the upper and lower limit frequencies of the energy concentration band in the baseline sub-library is calculated.

[0062] The specific calculation method for determining the degree of difference is as follows: For time domain characteristics, calculate the absolute value of the difference between the online value and the baseline mean value for that period, and then divide it by the baseline standard deviation for that period to obtain the normalized deviation in units of standard deviation; for frequency domain dominant frequency offset characteristics, calculate the absolute value of the frequency difference between the online dominant frequency value and the center frequency of the baseline energy concentration band.

[0063] In one feasible implementation, when the normalization deviation of a certain feature exceeds a preset allowable range, the signal is marked as a signal suspected of deviating from normal operating characteristics.

[0064] In another feasible implementation, the degree of difference can also be determined by multi-feature comprehensive discrimination, that is, the time domain amplitude deviation and the frequency domain main frequency offset deviation are weighted and summed according to preset weights to obtain a comprehensive deviation index. When the comprehensive deviation exceeds the preset threshold, it is marked as a suspected signal.

[0065] In another feasible implementation, the degree of deviation of the frequency domain characteristics can be used as the main criterion for judgment, because the generation of partial discharge signals often first manifests as the appearance of new frequency components or a significant change in the energy distribution of the original frequency components, while the change in the time domain amplitude is relatively lagging or insignificant.

[0066] The specific values ​​of the aforementioned preset allowable range, weighted weight value, or preset threshold can be preset by the operator based on a large amount of field experimental data or statistical results. For example, the preset allowable range is set to 2 to 3 times the standard deviation, the weighted weight value is set to 1, and the comprehensive deviation threshold is set to 4.

[0067] Through the above feature extraction and difference determination, the quantitative deviation assessment of the online signal relative to the normal baseline characteristics was achieved.

[0068] In one feasible implementation, online electromagnetic signals whose differences exceed a preset allowable range are marked as suspected partial discharge signals and included in a suspected partial discharge signal set.

[0069] Specifically, for each frame of online electromagnetic signal, a binary labeling result is generated after the degree of difference is determined. If the normalized deviation of any time-domain feature or frequency-domain feature of the signal exceeds the preset allowable range, or the comprehensive deviation exceeds the preset threshold, the signal frame is labeled as "suspected" and stored together with its acquisition timestamp, the original signal data frame, and the deviation of each feature in the suspected partial discharge signal set buffer area. If all feature deviations are within the preset allowable range, the signal frame is labeled as "normal" and no further in-depth processing is performed.

[0070] In another feasible implementation, the determination of the difference exceeding the preset allowable range can also adopt the method of joint determination of consecutive frames. That is, it is only marked as a suspected signal when the deviation of the online signal of N consecutive frames exceeds the preset threshold, so as to avoid false triggering of single frame signal due to occasional noise spikes.

[0071] The number of consecutive frames N is preferably between 3 and 5 frames, and the specific value can be set by those skilled in the art based on the intensity of electromagnetic noise and the acquisition frame rate at the site.

[0072] Through the above screening mechanism, during most normal operating periods, a large number of online signals that match the baseline features well are filtered out in step 120, without needing to enter the subsequent edge-side deep time-frequency domain feature extraction and classification recognition stages.

[0073] The processing in step 120 significantly reduces the amount of signal data that edge computing nodes need to analyze in depth, greatly saving computing and storage resources on the edge side. This allows limited edge computing power to be concentrated on processing suspected signals that truly deviate from normal operating characteristics, thereby improving the processing efficiency and resource utilization of the entire monitoring system.

[0074] Step 130: On the edge side, perform time-frequency domain feature extraction on each signal in the suspected partial discharge signal set, and classify and identify the signals according to the time-frequency domain features to distinguish between real partial discharge signals and various interference signals; In this embodiment of the invention, the deviation between each signal in the suspected partial discharge signal set selected in step 120 and the baseline features only indicates that the signal deviates from the statistical range of the normal operating state of the device in one or more feature dimensions. However, the deviation itself cannot distinguish whether the deviation is caused by actual partial discharge or by various interference signals on site.

[0075] Both real partial discharge signals and various interference signals exhibit transient pulse patterns superimposed on background noise in the time domain waveform, and may show energy accumulation phenomena in specific frequency bands in the frequency domain, making it difficult to effectively distinguish between the two in a single domain.

[0076] However, the real partial discharge signal originates from the air gap breakdown or surface discharge inside the insulation of the equipment. Its physical nature is a steep pulse current in the picosecond to nanosecond range. This pulse current excites broadband electromagnetic radiation, and its time-frequency distribution has specific initial phase characteristics, pulse repetition frequency characteristics, and damped oscillation characteristics. Operational noise interference originates from the transient overvoltage generated by the opening and closing of the switch. Its time spectrum is characterized by low-frequency components dominating and a wide pulse width. Electromagnetic interference originates from the high-frequency switching ripple of power electronic devices or wireless communication signals. Its time spectrum exhibits narrow-band continuous spectrum or fixed carrier frequency characteristics. Corona discharge interference originates from the air ionization on the surface of high-voltage conductors. Its time spectrum exhibits half-wave symmetry and the energy is mainly distributed in a specific power frequency phase range.

[0077] Based on the differentiated physical characteristics of signals from different sources in the time-frequency domain, this application proposes to perform refined time-frequency domain feature depth extraction on the edge side of suspected signals, and to perform classification and identification based on the extracted feature vectors.

[0078] In one feasible implementation, each suspected signal in the suspected partial discharge signal set is divided into multiple consecutive sub-signal segments according to a fixed time window.

[0079] Since partial discharge signals and various interference signals all have transient pulse properties, their energy is not uniformly distributed on the time axis but concentrated in a relatively short time interval. If the entire signal frame is used as a unit for overall analysis, the specific time information of the pulse position within the frame will be submerged in the overall statistical average, resulting in insufficient time resolution to characterize the temporal details of the pulse.

[0080] Therefore, each suspected signal frame is divided into fixed time windows. The time window length of each sub-signal segment is preferably 1 microsecond to 10 microseconds. Adjacent sub-signal segments preferably maintain a 50% overlap rate, that is, the start time of each sub-signal segment is slid backward by half a window length relative to the start time of the previous sub-signal segment.

[0081] The selection of the time window length is based on the following: The rise time of a typical partial discharge pulse is on the order of nanoseconds to tens of nanoseconds, and the total duration of the pulse is usually on the order of hundreds of nanoseconds to several microseconds. Setting the window length to 1 microsecond to 10 microseconds can ensure that a single partial discharge pulse falls completely within one or more sub-signal segments without multiple pulses overlapping due to an excessively wide window, or a single pulse being truncated due to an excessively narrow window.

[0082] A 50% overlap rate ensures that pulse signals located near the boundary of two adjacent sub-signal segments are fully captured within at least one sub-signal segment without being truncated by the boundary.

[0083] In another feasible implementation, the time window length can also be adjusted according to the statistical results of the pulse width measured on site. For example, when the pulse width in the field signal is generally narrow, a window length of 1 microsecond can be used, and when the pulse width is wide, a window length of 5 microseconds or 10 microseconds can be used.

[0084] In another feasible implementation, the overlap rate between adjacent sub-signal segments can also be set to 75% or 33%. The higher the overlap rate, the higher the time resolution, but the computational load increases accordingly. The specific value can be set by those skilled in the art based on the processing capability and real-time requirements of the edge computing node.

[0085] In one feasible implementation, time-frequency conversion is performed on each sub-signal segment to generate a time-frequency distribution representation corresponding to the sub-signal segment.

[0086] The purpose of time-frequency conversion is to transform the sub-signal segment from a pure time domain representation to a two-dimensional joint representation of time and frequency, thereby revealing the variation law of each frequency component in the signal over time.

[0087] Specifically, a short-time Fourier transform is performed on the discrete sampling point sequence within each sub-signal segment, and the transform formula is as follows: ; in, This is the sequence number of the sub-signal segment within its suspected signal frame, with a value ranging from 1 to M. This represents the total number of sub-signal segments divided from the suspected signal frame; is the sampling point number within the sub-signal segment, with a value range from 0 to N-1, where N is the total number of sampling points within each sub-signal segment; The amplitude of the nth sampling point within the mth sub-signal segment; For the window function in the first The weighting coefficients at each sampling point are preferably obtained using the Hanning window as the window function, which has a moderate main lobe width and a large side lobe attenuation, and can effectively suppress spectral leakage. This is a discrete frequency index in the frequency domain, with values ​​ranging from 0 to N-1; For the m-th sub-signal segment in the... Complex spectral values ​​at each frequency point, for The energy amplitude at that frequency point is obtained by taking the modulus and squaring it. .

[0088] The total number of sampling points N in each sub-signal segment is determined by the time window length and the sampling rate. When the sampling rate is 500 MS / s and the time window length is 2 microseconds, N equals 1000 sampling points. The number of points in the Fast Fourier Transform is preferably set to the same value as N. When N is not an integer power of 2, zeros are padded to the next integer power of 2 to improve computational efficiency.

[0089] In another feasible implementation, the time-frequency conversion can also use wavelet transform instead of short-time Fourier transform. Wavelet transform has high frequency resolution in the low-frequency band and high time resolution in the high-frequency band, and is also applicable to the analysis of partial discharge signals containing broadband components.

[0090] Through the above time-frequency conversion process, each sub-signal segment is mapped to a two-dimensional time-frequency energy matrix. The rows of the matrix correspond to time sampling points, the columns correspond to frequency components, and each element value in the matrix represents the energy amplitude at a specific frequency at a specific time.

[0091] In one feasible implementation, the frequency components whose energy amplitude exceeds a preset threshold and their occurrence time are extracted from the time-frequency distribution representation and used as the time-frequency characteristics of the sub-signal segment.

[0092] In the time-frequency energy matrix, the energy amplitude of the real signal components is usually significantly higher than the background noise floor, while the amplitude level of the noise floor is relatively stable between different sub-signal segments.

[0093] Based on this characteristic, the statistical distribution of all element values ​​in the time-frequency energy matrix of each sub-signal segment is first calculated, and the mean of the energy amplitude plus K times the standard deviation is used as the preset threshold value, where the value of K is preferably 3 to 5.

[0094] When K is 3, under the assumption of normal distribution, the probability of exceeding this threshold is about one in a thousand, which can effectively filter out most random noise components without affecting the detection of real signal components.

[0095] Another way to determine the preset threshold is to sort all the element values ​​in the time-frequency energy matrix of each sub-signal segment from largest to smallest in amplitude, and multiply the energy value of the 95th percentile after sorting by a coefficient less than 1 as the threshold value, such as multiplying by 0.5, so that only the frequency components with energy amplitude at the top of the matrix are retained.

[0096] After determining the preset threshold, each element in the time-frequency energy matrix is ​​traversed, and the frequency component corresponding to the element whose energy amplitude exceeds the preset threshold and the time position of the sub-signal segment in which the frequency component appears are recorded.

[0097] Specifically, for the m-th sub-signal segment, record the set of indices k of all frequency points whose energy amplitude exceeds a preset threshold. The occurrence times of these frequency points within the sub-signal segments are recorded; then, the occurrence times of the same frequency component in different sub-signal segments are recorded as the timing occurrence positions of that frequency component.

[0098] In another feasible implementation, the preset threshold can also be determined by a fixed value. For example, a fixed energy amplitude threshold can be preset based on the background noise level of the analog-to-digital converter. This fixed threshold value remains unchanged after being set once during the equipment commissioning and calibration phase.

[0099] Regardless of which threshold determination method is used, by extracting the frequency components whose energy amplitude exceeds the preset threshold, the background noise components in the time-frequency energy matrix can be effectively filtered out, and only the significant energy components in the signal are retained.

[0100] In one feasible implementation, the time-frequency features of all sub-signal segments of the same suspected signal are combined in chronological order to form the time-frequency domain depth feature vector of the suspected signal.

[0101] Specifically, for M sub-signal segments belonging to the same suspected signal frame, the frequency values ​​of the frequency components exceeding the threshold extracted from each sub-signal segment and their corresponding normalized energy amplitudes are arranged sequentially according to the sub-signal segment number m from 1 to M, forming a multi-dimensional feature vector.

[0102] The dimension of the feature vector is determined by the total number of frequency components exceeding the threshold in each sub-signal segment. The feature vector dimension of different suspected signals may vary depending on the complexity of the signal.

[0103] Each element in the feature vector is stored in the form of a frequency-energy tuple, that is, the element form is ( , ),in This represents the frequency value of the frequency component. This is the normalized ratio of the energy amplitude of the frequency component to the total energy of the sub-signal segment.

[0104] Through the above combined processing, each suspected signal is characterized as a multi-dimensional time-frequency feature sequence carrying time sequence information. This feature sequence not only retains the occurrence position and order information of each frequency component on the time axis, but also retains the relative energy relationship of each frequency component, providing a rich information foundation for subsequent classification and recognition.

[0105] Through the aforementioned processing of time-frequency domain depth feature extraction in step 130, the suspected signal is transformed from the original sampling point sequence into a structured time-frequency domain depth feature vector, achieving efficient compression of signal information and enhancement of key discriminative features, and significantly improving the information input quality for subsequent classification and recognition.

[0106] Step 140: Only when the classification and identification result confirms that it is a real partial discharge signal, generate alarm information and report it; In this embodiment of the invention, after the classification and identification of each signal in the suspected partial discharge signal set is completed in step 130, each suspected signal frame has been assigned a signal category label, which indicates that the signal belongs to a certain category among real partial discharge signal, operation noise interference, electromagnetic interference, or corona discharge interference.

[0107] However, not all frames identified as genuine partial discharge signals need to trigger an alarm. In actual operation, occasional genuine partial discharge pulses may be generated due to the self-recovering breakdown of tiny defects inside the equipment insulation. Although such single pulses fall under the category of partial discharge, their energy is extremely low and they are not continuous. If an alarm is generated for every isolated pulse, the number of alarm messages will surge, overwhelming the truly important continuous discharge events.

[0108] Therefore, this application sets up a condition triggering mechanism, that is, an alarm message is generated and reported only when the classification and identification result is confirmed as a real partial discharge signal. Signals that are not real partial discharge categories or signals that are real partial discharges but have not yet met the alarm triggering accumulation conditions will not trigger an alarm.

[0109] In one feasible implementation, the criterion for confirming that the classification and identification result is a real partial discharge signal is: the signal category label output by the classification and identification in step 130 is a real partial discharge category, and the matching confidence level corresponding to the category label reaches the preset alarm trigger threshold.

[0110] Confidence level represents the degree of credibility of classification and recognition results. Its value ranges from 0 to 1, and the higher the value, the more reliable the recognition result.

[0111] When a signal frame is classified as a true partial discharge and the confidence level reaches the preset alarm trigger threshold, the signal frame is marked as a valid alarm candidate frame.

[0112] The preset alarm trigger threshold is preferably between 0.85 and 0.95. The determination of this range is based on the following: when the confidence level is below 0.85, the classifier has a large uncertainty in classifying the signal frame, and its position in the feature space may be in the ambiguous region between the real partial discharge category and the interference category. Triggering an alarm in this case carries a high risk of false alarms. When the confidence level is above 0.95, the classifier's judgment has extremely high reliability, but some real partial discharge signals with weak energy or slight feature variations may be missed due to slightly low confidence. Therefore, setting the threshold between 0.85 and 0.95 can achieve a reasonable balance between the false alarm rate and the missed alarm rate.

[0113] In another feasible implementation, the preset alarm trigger threshold can be dynamically adjusted based on on-site operating experience. For example, when the on-site electromagnetic environment is relatively clean and there are few interference signals, the threshold can be appropriately reduced to 0.80 to improve the detection rate of weak partial discharge signals. When the on-site electromagnetic environment is complex and there are many interference signals, the threshold can be appropriately increased to 0.95 to prioritize ensuring the accuracy of the alarm.

[0114] In another feasible implementation, the alarm triggering condition can also adopt a joint determination method of multiple consecutive frames, that is, the alarm is triggered only when multiple consecutive frames of signals are classified as real partial discharge and the confidence level reaches the preset alarm triggering threshold. The number of consecutive frames is preferably 3 to 5 frames.

[0115] This multi-frame joint judgment method can effectively filter out occasional false triggers caused by a single isolated pulse or transient noise spike, further improving the reliability of the alarm.

[0116] In one feasible implementation, the alarm information includes at least the device identification information corresponding to the signal, the collection timestamp, the classification and identification confidence level, and the location data.

[0117] The equipment identification information is used to indicate the specific number or name of the high-voltage equipment that emitted the partial discharge signal, so that maintenance personnel can accurately locate the faulty equipment; the acquisition timestamp is used to record the specific time when the signal was acquired, providing a time reference for subsequent trend analysis and fault evolution tracking; the classification and identification confidence level is used to assist maintenance personnel in assessing the credibility of the alarm, and the higher the confidence level, the greater the reference value of the alarm; the location data is used to indicate the approximate location of the partial discharge signal on the high-voltage equipment, and its location accuracy depends on the sensor layout density and the performance of the time difference of arrival (TDOA) positioning algorithm.

[0118] In one feasible implementation, the alarm information can further include key feature components in the time-frequency domain deep feature vector of the actual partial discharge signal, such as the main frequency point with the highest energy amplitude and its corresponding normalized energy value, as well as the amplitude stability information of the main frequency point in different sub-signal segments, so that maintenance personnel can comprehensively judge the severity and development trend of the discharge.

[0119] In one feasible implementation, the reporting operation is performed via a communication link between the edge computing node and the remote monitoring center.

[0120] After generating alarm information, the edge computing node encapsulates the alarm information into a data message according to the preset data reporting format, and preferably sends it to the remote monitoring center through a wired network or wireless cellular network using a message queue telemetry transmission protocol or a hypertext transfer security protocol.

[0121] The message queue telemetry transmission protocol has the advantages of being lightweight, low-power, and suitable for unstable network environments, and can ensure the reliable transmission of alarm information under network bandwidth-limited conditions; the hypertext transfer security protocol has broad compatibility and the ability to penetrate firewalls, and is suitable for scenarios where monitoring centers are deployed on enterprise intranets or cloud servers.

[0122] While reporting alarm information, edge computing nodes preferably back up and store the alarm information on local storage media to prevent the loss of alarm information due to temporary interruption of communication links.

[0123] In another feasible implementation, the reporting operation can also adopt a hierarchical reporting strategy. That is, when the signal is confirmed to be a real partial discharge with extremely high confidence, a high-priority alarm is reported immediately. When the confidence is in the range near the preset alarm trigger threshold, the signal is first cached on the edge side and the recognition results of subsequent signal frames are observed before a reporting decision is made. This reduces unnecessary alarm reporting traffic while ensuring the real-time nature of emergency alarms.

[0124] In one feasible implementation, no alarm information is generated or reporting is performed for signal frames that are classified as real partial discharges but whose confidence level does not reach the preset alarm trigger threshold, as well as for all signal frames that are classified as various types of interference signals.

[0125] These signal frames are released after classification and recognition are completed, and the edge computing node cache space they occupy is cleared to receive new signals acquired in real time.

[0126] Through the above-mentioned conditional triggering mechanism, a large number of signals identified as interference categories and real partial discharge candidate signals with insufficient confidence are effectively intercepted before the alarm triggering stage, and only the signal with the highest confirmation will trigger the alarm.

[0127] Step 140 enables strict constraints on alarm information generation conditions, fundamentally avoiding false alarms caused by various interference signals, significantly reducing the total amount of alarm information received by the remote monitoring center, and allowing maintenance personnel to focus their limited energy on handling partial discharge events that actually pose a risk of failure, thus significantly improving the actual availability of the monitoring system and the efficiency of maintenance work.

[0128] Step 150: Update the newly confirmed interference signal features to the electromagnetic fingerprint baseline database at a preset period.

[0129] In this embodiment of the invention, in step 140, the signal frames that are classified and identified as various types of interference signals do not trigger an alarm, but they themselves carry valuable information about the on-site electromagnetic environment—these interference signals reflect the types, intensity and spectral distribution characteristics of various non-partial discharge electromagnetic disturbances that actually exist at the equipment operating site.

[0130] As the high-voltage equipment continues to operate, new sources of interference may appear on site, such as newly commissioned power electronic devices, the addition of mobile communication base stations, and changes in corona activity of nearby transmission lines. These newly emerging interference patterns were not included in the initial electromagnetic fingerprint baseline database constructed in step 110.

[0131] If these newly added interference patterns are not included in the baseline database for a long time, they will continue to be marked as suspected signals that deviate from normal characteristics in the comparison stage of step 120 during the subsequent real-time monitoring process, and will repeatedly enter the deep identification stage of step 130, resulting in continuous waste and ineffective occupation of edge computing resources.

[0132] To this end, this application proposes to update the characteristics of newly confirmed interference signals to the electromagnetic fingerprint baseline database at a preset period, so that the baseline database can evolve synchronously with the evolution of the on-site electromagnetic environment, ensuring that the subsequent suspected signal screening process always maintains high screening efficiency and accuracy.

[0133] In one feasible implementation, the preset cycle is 24 to 168 hours, that is, once a day to once a week.

[0134] The specific value of the preset period needs to take into account two factors: the rate of change of the on-site electromagnetic environment and the load capacity of the edge computing nodes. If the period is too short, the edge computing nodes will need to frequently perform database update operations, consuming additional computing and storage resources. Moreover, the number of interference signal samples collected in a short period of time is limited, and the statistical representativeness is insufficient. If the period is too long, the newly emerging interference patterns will not be included in the baseline database for a long time, continuously consuming the computing resources of deep recognition.

[0135] Setting the preset period to 24 to 168 hours ensures that edge computing nodes have enough time to accumulate a sufficient number of interference signal samples to form statistically valid feature data, while also ensuring that the baseline database can track the evolution of the on-site electromagnetic environment at a reasonable time granularity.

[0136] In another feasible implementation, the preset cycle can also be set differently according to the equipment type and operating environment. For example, for urban substations where the electromagnetic environment changes frequently, a short cycle of 24 hours can be used for updates; for substations in remote areas where the electromagnetic environment is relatively stable, a long cycle of 168 hours can be used for updates.

[0137] In another feasible implementation, the preset period can also be dynamically adjusted. That is, the edge computing node monitors the daily trend of the number of frames classified as interference signals. When the number of interference signal frames increases by more than a preset ratio compared with the average of the previous period, an automatic non-periodic emergency update is triggered.

[0138] Regardless of which of the above-mentioned period determination methods is used, the database update operation is triggered based on a preset time base.

[0139] In one feasible implementation, at the arrival of each update cycle, the time-frequency domain features of all signals classified as interference in the previous cycle are obtained, and each interference signal is classified into the corresponding time-based baseline sub-library according to the acquisition time of each interference signal.

[0140] Specifically, when the edge computing node starts each update cycle, it traverses the classification and identification records of all signal frames in the previous cycle in the local cache, and filters out all data records that are labeled as various interference signals rather than real partial discharge signals.

[0141] Each interference signal data record contains the timestamp information of its acquisition time and the time-frequency domain depth feature vector extracted in step 130.

[0142] The edge computing node assigns the time-frequency domain depth feature vector of each interference signal to the corresponding time-period baseline sub-database to be updated, based on the time period identifier corresponding to the acquisition time of each interference signal. For example, the interference signal acquired at 14:35 is assigned to the data set to be updated for the 14th time period or the peak load time period.

[0143] Through the above classification process, each time period baseline sub-database obtains a set of characteristic data of all interference signals belonging to the same time period in the previous cycle, providing a data foundation for subsequent targeted updates.

[0144] In one feasible implementation, the time-frequency domain features of interference signals classified into the same time period baseline sub-library are compared with the baseline features already stored in the current time period baseline sub-library to determine their feature overlap.

[0145] The purpose of this comparison is to determine whether the interference feature belongs to a known interference pattern already included in the baseline database, or to a newly added interference pattern that has not been included before.

[0146] Specifically, for each baseline feature vector already stored in the time period baseline sub-library, it includes the upper and lower limits of the energy concentration frequency band range of the baseline feature, the statistical mean of the energy amplitude within the frequency band range, and the time stability index of the feature.

[0147] The feature vector of the interference signal to be updated also includes the upper and lower limits of its energy concentration band and the corresponding energy amplitude.

[0148] The mathematical calculation method for feature overlap comparison is as follows: calculate the frequency overlap ratio between the energy concentration band of the interference signal features and the energy concentration band of the baseline features, using the following formula: ; in, It is the larger of the lower limit frequency of the interference signal energy concentration band and the lower limit frequency of the baseline characteristic energy concentration band. The smaller of the two upper frequency values. It is the smaller of the two lower limit frequencies. The larger of the two upper frequency limits is denoted as 1; the numerator is the width of the intersection interval of the two frequency bands, and the denominator is the width of the union interval of the two frequency bands. Overlap is the frequency overlap ratio, which ranges from 0 to 1.

[0149] When the overlap value is close to 1, it indicates that the two frequency bands basically overlap, that is, the characteristics of the interference signal are highly similar to the characteristics of the existing baseline; when the overlap value is close to 0, it indicates that the two frequency bands have almost no overlap, that is, the interference signal has new spectral characteristics that are significantly different from the characteristics of the existing baseline.

[0150] In another feasible implementation, feature overlap comparison can also add the comparison dimension of time domain amplitude statistical features on the basis of frequency domain overlap ratio. That is, the difference between the peak-to-peak mean of the interference signal and the peak-to-peak mean of the baseline feature is calculated. When the absolute value of the difference is less than the preset tolerance range, it is also regarded as feature overlap.

[0151] In another feasible implementation, the similarity measure for feature overlap comparison can also use alternative algorithms such as cosine similarity or Euclidean distance. Cosine similarity is suitable for measuring the directional consistency of two feature vectors in high-dimensional space, while Euclidean distance is suitable for measuring the degree of proximity of the absolute positions of two feature vectors in space. The specific measurement method can be selected by those skilled in the art based on the specific dimension of the feature vectors and the data distribution characteristics.

[0152] In one feasible implementation, the time-frequency domain features of interference signals with feature overlap below a set threshold are added as new features and incorporated into the current time-period baseline sub-library to form an updated time-period baseline sub-library.

[0153] The threshold value is preferably between 0.6 and 0.8. That is, when the frequency overlap ratio between the interference signal feature and the baseline feature is lower than the threshold, the interference signal feature is considered to be a newly added interference mode that has not yet been included in the baseline database.

[0154] The threshold value is determined as follows: when the overlap ratio is less than 0.6, the spectral range of the interference signal is significantly different from the spectral range of the existing baseline features, and has clear incremental information value, so it should be included; when the overlap ratio is greater than 0.8, the spectral characteristics of the interference signal are very close to the existing baseline features, and adding this feature will not bring significant incremental information to the database, so it does not need to be stored repeatedly; when the overlap ratio is between 0.6 and 0.8, whether to include it can be determined comprehensively based on the degree of deviation of the frequency or energy amplitude of the interference signal.

[0155] In another feasible implementation, the specific value of the threshold can be adjusted according to the number of features already stored in the baseline sub-database for the current time period. When the number of stored features is small, the threshold can be appropriately relaxed to enrich the coverage of the database, for example, the threshold can be set to 0.5. When the number of stored features is large, the threshold can be appropriately tightened to control the database size, for example, the threshold can be set to 0.8.

[0156] When the characteristics of the interference signal meet the inclusion conditions, the time-frequency domain characteristics of the interference signal are written as new baseline feature entries into the corresponding time period baseline sub-library. The new entries include metadata such as the energy concentration frequency band range, amplitude statistical average, and first entry time of the interference signal, thereby completing an incremental update of the time period baseline sub-library.

[0157] Through the cyclical processing of feature overlap comparison and incremental merging, all signals in the set of interference signals to be updated in the baseline sub-database for each time period are processed one by one. Only newly added interference patterns are included in the database, and existing patterns are no longer stored repeatedly. This effectively controls the data redundancy of the database while ensuring the integrity of the baseline database information.

[0158] Through the processing in step 150, the electromagnetic fingerprint baseline database achieves continuous self-improvement and dynamic evolution. Various interference modes newly added at the equipment operation site can be absorbed into the baseline database in the next update cycle after being identified for the first time. When the interference of this type reappears, it will match the updated baseline features in the comparison stage in step 120 and will no longer be marked as a suspected signal, thus forming a closed-loop self-optimization mechanism of "identifying interference, collecting interference, and filtering interference".

[0159] like Figure 2 The diagram shown is a functional block diagram of a high-voltage partial discharge monitoring system based on edge AI voiceprint recognition provided in an embodiment of the present invention.

[0160] The high-voltage partial discharge monitoring system 100 based on edge AI voiceprint recognition described in this invention can be installed in an electronic device. Depending on the functions implemented, the high-voltage partial discharge monitoring system 100 may include a fingerprint registration module 101, a signal comparison module 102, a feature identification module 103, an anomaly alarm module 104, and a baseline iteration module 105. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0161] In this embodiment, the functions of each module / unit are as follows: The fingerprint filing module 101 is used to establish an electromagnetic fingerprint baseline database based on the on-site electromagnetic signals collected during the initial operation of the high-voltage equipment. The signal comparison module 102 is used to collect online electromagnetic signals during the operation of the template high-voltage equipment in real time, and compare the online electromagnetic signals with the electromagnetic fingerprint baseline database to obtain a set of suspected partial discharge signals. The feature discrimination module 103 is used to perform time-frequency domain feature depth extraction on each signal in the suspected partial discharge signal set at the edge side, and classify and identify the signals according to the time-frequency domain features to distinguish between real partial discharge signals and various interference signals. The abnormal alarm module 104 is used to generate alarm information and report it only when the classification and identification result is confirmed to be a real partial discharge signal; The baseline iteration module 105 is used to update the electromagnetic fingerprint baseline database with newly confirmed interference signal features at a preset period.

[0162] like Figure 3 The figure shown is a curve illustrating the 24-hour time-domain amplitude variation of a normal electromagnetic fingerprint baseline database for a high-voltage partial discharge monitoring system based on edge AI voiceprint recognition, provided in an embodiment of the present invention.

[0163] In the initial stage of high-voltage equipment commissioning and background noise characterization, establishing a time-domain amplitude baseline database is the core foundation for achieving subsequent differentiated over-limit triggering. Specifically, the acquisition system uses an ultra-high frequency antenna deployed on the grounding down conductor, with a fixed... High-speed sampling rate was used to continuously capture on-site electromagnetic signals for 7 to 14 consecutive natural days. The acquired raw sequences were divided into 24 time-segment signal groups with a basic physical interval of 1 hour.

[0164] For each time-segment signal group, the time-domain feature extraction unit first calculates the peak value of each signal sample using a sliding window. This data was used to construct a statistical sample set for each time period. The dark blue solid line in the figure represents the time period within each time period. Mean Statistical Curve (in Its waveform shows a regular rise during the day as the transformer load in the station increases, the wind turbines start and stop, and the power grid is operated (reaching approximately [a certain value] between 11:00 and 15:00). The peak load (in hours) falls back to approximately [a certain value] during the nighttime off-peak period. .

[0165] To achieve a refined description of the temporal random fluctuations of background noise, the system introduces multi-confidence interval bands. The dark blue semi-transparent filled area in the figure represents the core distribution interval, whose boundaries are defined by one standard deviation. ), representing approximately The normal background noise concentration area; the light blue semi-transparent filled area is the normal fluctuation range. ), which corresponds to statistically The probability boundary. Through this dynamic time envelope design, the time-domain baseline sub-library not only stores a single static threshold, but also stores a continuous time function containing the mean and higher-order statistical variance, thereby avoiding the misjudgment of normal noise electromagnetic radiation caused by the alternation of day and night loads and changes in meteorological temperature in the substation environment as abnormal partial discharge.

[0166] like Figure 4 The diagram shown is a real-time electromagnetic signal online comparison and dynamic over-limit judgment mechanism of a high-voltage partial discharge monitoring system based on edge AI voiceprint recognition, provided by an embodiment of the present invention.

[0167] In the real-time monitoring phase of high-voltage equipment in substations during operation, the comparison and screening process of online electromagnetic signals is crucial for filtering massive amounts of background data and ensuring the rational allocation of edge AI computing power. The front-end high-speed digital-to-analog converter (ADC) captures the online electromagnetic signals in operation in real time using a continuously sliding time window; the black undulating curve in the figure represents the real-time acquired discrete voltage waveform in the time domain.

[0168] To address the dynamic drift of background noise due to substation load and day / night cycles, the signal comparison module at the current moment... After acquiring the online signal, the system first extracts the timestamp corresponding to the signal frame to determine its time period (e.g., 2 PM to 3 PM). Then, it retrieves the baseline sub-library corresponding to that time period from the local non-volatile memory in real time and extracts the mean temporal amplitude and three times the standard deviation boundary of that time period. Combined with the environmental correction coefficient under the current operating conditions, the system generates a pink dashed line above the real-time waveform, representing the "temporal dynamic difference trigger threshold line".

[0169] During normal operation, the black wavy line remains stable below the dynamic threshold line. However, when partial discharge breakdown occurs inside the equipment insulation, or when high-voltage operation takes place on-site, the time-domain waveform changes. and A high-amplitude, narrow pulse spike erupted nearby. The comparison unit detected that the real-time sampling point voltage directly broke through the dynamic limit (shown by the red dots in the figure, which are the "characteristic deviation anomaly trigger points"), and this over-limit behavior satisfies continuous time-axis... The common decision constraint of the sliding window. The trigger point signal is immediately marked as a suspected partial discharge signal and quickly included in the "suspected partial discharge signal set" buffer area to be analyzed in depth, thus realizing high-sensitivity capture of transient anomalies with minimal computing power overhead.

[0170] like Figure 5 As shown, this is a refined multi-category interference attribution graph based on the deviation direction vector space of a high-voltage partial discharge monitoring system based on edge AI voiceprint recognition, provided by an embodiment of the present invention.

[0171] For classification and identification results that do not exceed the partial discharge confirmation alarm threshold ( For samples from the edge AI classifier, instead of discarding them as useless data, the off-direction vector space analysis is introduced to achieve refined multi-class attribution diagnosis of the interference environment within the station. In this algorithm, the classifier uses the "true partial discharge feature center" as the geometric reference origin of the multi-dimensional feature space (marked by the red pentagram in the figure).

[0172] When the classification matching degree of a suspected signal does not reach the partial discharge alarm threshold, the algorithm calculates the deviation vector of the corresponding time-frequency depth feature vector from the origin of the partial discharge center. And using principal component analysis (PCA) projection technology, it is projected onto the two-dimensional deviation space feature surface, as shown in the figure. Axis (characteristic shape deviation component) and A coordinate system composed of axes (characteristic frequency domain deviation components).

[0173] Electromagnetic interference with different physical properties exhibits distinct directional clustering characteristics in the deviation direction space due to differences in physical radiation mechanisms: the deviation vector pointing to the upper left corner "Direction A" (blue dashed arrow and "Direction A" label in the figure) corresponds to the "shutdown operation noise interference zone," the physical mechanism of which lies in the slow attenuation of the pulse generated by the circuit breaker arcing in the time-frequency domain; the deviation vector pointing to the upper right corner "Direction B" (purple dashed arrow and "Direction B" label in the figure) corresponds to the "continuous high-frequency electromagnetic interference zone"; and the deviation vector pointing to the lower "Direction C" (orange dashed arrow and "Direction C" label in the figure) corresponds to the "electromagnetic radiation peak corona discharge zone," reflecting the characteristic deviation of air ionization in a fixed positive and negative phase. By introducing the multi-directional discrimination diagram of the "deviation characteristic physical direction vector" shown by the gray arrow, the system can finely subdivide the complex field interference into three categories at the edge and label them with specific types, providing accurate data attribute classification for the closed-loop self-evolution of the fingerprint database.

[0174] like Figure 6 The figure shown is a chart illustrating the self-evolution and false alarm rate degradation trend of a closed-loop iterative adaptive evolution feature library for a high-voltage partial discharge monitoring system based on edge AI voiceprint recognition, provided in an embodiment of the present invention.

[0175] The physical reason why the high-voltage partial discharge monitoring system can maintain high accuracy in dynamic electromagnetic fields for a long time lies in its closed-loop self-updating adaptive evolution mechanism based on edge AI feedback. This evolution process is an incremental learning process of the electromagnetic fingerprint baseline database with a preset update cycle (weekly in this example) as the step size.

[0176] The dual-axis curves in the figure quantitatively demonstrate the evolution trend of the feature database capacity and the false alarm rate during the operation of the system over seven update cycles. The left Y-axis (solid blue line and circular markers) represents the number of feature entries stored in the feature sub-database of the electromagnetic fingerprint baseline library, while the right Y-axis (dashed red line and square markers) represents the probability of the system triggering false alarms due to interference signals at the edge.

[0177] In the first week, the system relied solely on the 24 basic background features established during the initial commissioning phase. Due to the emergence of electromagnetic noise from surrounding construction at the substation site, the false alarm rate of interference signals at the system's edge side was low. The system achieves a relatively high level of overlap. At the start of each cycle, the system retrieves all subdivided and tagged interference signal features (such as operational noise, corona discharge, etc.) from the previous cycle and calculates their overlap with the features already stored in the corresponding time period's sub-database. For overlap levels below a certain threshold... Newly added interference patterns are incorporated into the baseline iteration module as new feature entries into the corresponding baseline sub-library. As the runtime progresses, the number of entries in the baseline feature sub-library gradually increases from 24 to 79 by week 7. Correspondingly, as newly emerging field interference patterns are successively included, they are incorporated into Phase Two (…). Figure 4 The comparison step in the real-time over-limit judgment (as shown) is perfectly matched and intercepted by the baseline library, thus reducing the system's interference false alarm rate from It exhibited exponential decay, completely degrading to its lowest value in week 7. This closed-loop evolutionary characteristic fully demonstrates the system's long-term autonomous survival and highly reliable identification capabilities in a variable and dynamic electromagnetic environment.

[0178] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0179] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0180] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0181] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0182] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0183] Finally, it should be noted that the above 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A high-voltage partial discharge monitoring method based on edge AI voiceprint recognition, characterized in that, The method includes: Step 110: Establish an electromagnetic fingerprint baseline database based on the on-site electromagnetic signals collected during the initial commissioning of the high-voltage equipment; Step 120: Real-time acquisition of online electromagnetic signals during the operation of the template high-voltage equipment, and comparison of the online electromagnetic signals with the electromagnetic fingerprint baseline database to obtain a set of suspected partial discharge signals; Step 130: On the edge side, perform time-frequency domain feature extraction on each signal in the suspected partial discharge signal set, and classify and identify the signals according to the time-frequency domain features to distinguish between real partial discharge signals and various interference signals; Step 140: Only when the classification and identification result confirms that it is a real partial discharge signal, generate alarm information and report it; Step 150: Update the newly confirmed interference signal features to the electromagnetic fingerprint baseline database at a preset period.

2. The high-voltage partial discharge monitoring method based on edge AI voiceprint recognition as described in claim 1, characterized in that, Based on the electromagnetic signals collected during the initial commissioning of high-voltage equipment, an electromagnetic fingerprint baseline database is established, including: The on-site electromagnetic signals were divided into multiple time-segment signal groups according to the acquisition time sequence; For each time period signal group, extract the time domain waveform amplitude variation characteristics and the frequency domain energy concentration band characteristics respectively; The time-domain waveform amplitude variation features and frequency domain energy concentration band features extracted for each time period are stored as the baseline sub-library for that time period.

3. The high-voltage partial discharge monitoring method based on edge AI voiceprint recognition as described in claim 2, characterized in that, Establishing an electromagnetic fingerprint baseline database also includes: Cross-time consistency comparison is performed on the same type of features in the baseline sub-database of each time period, and abnormal sub-databases in each time period whose deviation from other sub-databases in the corresponding time period exceeds the set range are removed; The baseline sub-databases for each time period are arranged and combined in chronological order to form an electromagnetic fingerprint baseline database.

4. The high-voltage partial discharge monitoring method based on edge AI voiceprint recognition as described in claim 1, characterized in that, The online electromagnetic signals are compared with the electromagnetic fingerprint baseline database to obtain a set of suspected partial discharge signals, including: Based on the current online electromagnetic signal acquisition time, retrieve the time period baseline sub-database corresponding to the acquisition time from the electromagnetic fingerprint baseline database; Extract the time-domain amplitude fluctuation characteristics and frequency-domain dominant frequency offset characteristics of online electromagnetic signals, and determine the degree of difference between the time-domain amplitude fluctuation characteristics and frequency-domain dominant frequency offset characteristics and the corresponding baseline characteristics in the time period baseline sub-library; Online electromagnetic signals whose differences exceed the preset allowable range are marked as suspected partial discharge signals and included in the suspected partial discharge signal set.

5. The high-voltage partial discharge monitoring method based on edge AI voiceprint recognition as described in claim 1, characterized in that, On the edge side, time-frequency domain feature depth extraction is performed on each signal in the suspected partial discharge signal set, including: Each suspected signal in the suspected partial discharge signal set is divided into multiple consecutive sub-signal segments according to a fixed time window; Perform time-frequency conversion on each sub-signal segment to generate the time-frequency distribution representation corresponding to the sub-signal segment; The frequency components whose energy amplitude exceeds a preset threshold and their occurrence time are extracted from the time-frequency distribution representation and used as the time-frequency characteristics of the sub-signal segment. The time-frequency features of all sub-signal segments of the same suspected signal are combined in chronological order to form the time-frequency domain depth feature vector of the suspected signal.

6. The high-voltage partial discharge monitoring method based on edge AI voiceprint recognition as described in claim 5, characterized in that, The time-frequency domain feature depth extraction of each signal in the suspected partial discharge signal set at the edge also includes: For each frequency component in the time-frequency domain deep feature vector, compare the amplitude fluctuation between different sub-signal segments; Frequency components whose amplitude fluctuations exceed the set range are marked as interference feature components, and the interference feature components are deleted from the time-frequency domain depth feature vector to obtain the updated time-frequency domain depth feature vector.

7. The high-voltage partial discharge monitoring method based on edge AI voiceprint recognition as described in claim 6, characterized in that, Signals are classified and identified based on their time-frequency domain characteristics to distinguish between real partial discharge signals and various interference signals, including: From the pre-set classification and discrimination rule library on the edge side, retrieve the feature matching templates corresponding to various types of signals. The feature matching templates include real partial discharge templates and various interference templates. The time-frequency domain depth feature vector of each suspected partial discharge signal is compared with the matching degree of the real partial discharge template and various interference templates. Based on the matching degree comparison results, the signal category corresponding to the template with the highest matching degree is taken as the classification and identification result of the suspected signal.

8. The high-voltage partial discharge monitoring method based on edge AI voiceprint recognition as described in claim 7, characterized in that, Classifying and identifying signals based on time-frequency domain characteristics to distinguish between real partial discharge signals and various interference signals also includes: When the classification result is an interference signal category, further obtain the deviation direction information between the time-frequency domain depth feature vector of the interference signal and the real partial discharge template; Based on the deviation direction information, the interference signals are subdivided into corresponding types among operation noise interference, electromagnetic disturbance interference, or corona discharge interference, and each type is labeled with an interference type label.

9. The high-voltage partial discharge monitoring method based on edge AI voiceprint recognition as described in claim 1, characterized in that, The newly identified interference signal characteristics are updated to the electromagnetic fingerprint baseline database at a preset period, including: At the arrival of each update cycle, the time-frequency domain features of all interference signals classified in the previous cycle are obtained, and they are classified into the corresponding time-based baseline sub-library according to the acquisition time of each interference signal. The time-frequency domain features of interference signals classified into the same time period baseline sub-library are compared with the baseline features already stored in the current time period baseline sub-library for feature overlap. Interference signals with time-frequency domain features whose feature overlap is below a set threshold are added as new features and incorporated into the current time period baseline sub-library to form an updated time period baseline sub-library.

10. A high-voltage partial discharge monitoring system based on edge AI voiceprint recognition, characterized in that, The system for implementing the high-voltage partial discharge monitoring method based on edge AI voiceprint recognition as described in claim 1 includes: The fingerprint filing module is used to establish an electromagnetic fingerprint baseline database based on the on-site electromagnetic signals collected during the initial commissioning of high-voltage equipment. The signal comparison module is used to collect online electromagnetic signals during the operation of the template high-voltage equipment in real time, and compare the online electromagnetic signals with the electromagnetic fingerprint baseline database to obtain a set of suspected partial discharge signals; The feature discrimination module is used to perform time-frequency domain feature extraction on each signal in the suspected partial discharge signal set at the edge, and classify and identify the signals according to the time-frequency domain features to distinguish between real partial discharge signals and various interference signals. The anomaly alarm module is used to generate and report alarm information only when the classification and identification results confirm that it is a real partial discharge signal; The baseline iteration module is used to update the electromagnetic fingerprint baseline database with newly confirmed interference signal features at a preset period.

Images