Power distribution device fault early warning method based on multi-source time sequence AI fusion

Abstract

The present application relates to the technical field of deep learning, in particular to a distribution network equipment fault early warning method based on multi-source time sequence AI fusion, comprising the following steps: dividing the power frequency cycle according to voltage zero-crossing, collecting multi-source data in the same phase section to generate samples, calculating the median according to the equipment identity and the phase section in the reference period, training the gated recurrent neural network with the maintenance fault samples to generate a reference sequence, calculating the difference between the current value and the median to form a precursor residual, analyzing the adjacent cycle precursor memory value and comparing it with the threshold to output warning information. In the present application, through phase alignment sampling and equipment phase reference establishment, the load fluctuation and individual reference difference are separated, the influence of fixed time window smoothing on short-term anomalies is reduced, the residual sequence expresses continuous evolution characteristics in adjacent power frequency cycles, the warning judgment is changed from static state recognition to precursor change recognition, the fault early detection capability is improved, and the risk of lagging warning and false positives and false negatives is inhibited.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method for early warning of faults in power distribution network equipment based on multi-source temporal AI fusion. Background Technology

[0002] The field of deep learning technology involves artificial intelligence computing techniques based on multi-layer neural networks to perform representation learning, training updates, and inference outputs on training samples, temporal features, and model parameters. Among these, traditional methods for early warning of faults in distribution network equipment involve collecting time-series data on current, voltage, temperature, partial discharge, and operating status from distribution transformers, switchgear, feeder terminals, and sensors. This multi-source data is then written into a data cache or historical sample library, normalized, imputed for missing values, segmented into time windows, and feature vectors constructed. Finally, a neural network model reads the sample sequence, performs forward inference, and outputs the equipment status category, fault probability, or alarm identifier.

[0003] Traditional early warning processes rely on multi-source data being cached and preprocessed uniformly. Features are constructed according to fixed time windows, and status categories or alarm labels are directly output. Before entering the model, the data is normalized and imputed. The phase difference within the power frequency cycle is smoothed by the time window. The sequential correlation between current, voltage, temperature, discharge and switching status is compressed into a static vector. It is difficult to maintain the individual operating benchmark of the equipment and the boundary of phase segment fluctuations. Brief precursors are easily masked by historical averages. Alarms are more biased towards the status identification after the fault occurs, resulting in insufficient anticipation and the risk of false alarms and missed alarms. Summary of the Invention

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a method for early warning of faults in distribution network equipment based on multi-source time-series AI fusion, comprising the following steps: The phase start mark formed by the voltage zero crossing is obtained, the power frequency cycle is divided according to the phase start mark, and the current amplitude, voltage amplitude, temperature rise rate, discharge pulse count, switch status information and equipment identity information of the distribution network equipment in the same phase segment are collected to generate multi-source phase time series samples. For the reference establishment period when maintenance is not triggered and the switch state is stable, the median of current amplitude, voltage amplitude, temperature rise rate and discharge pulse number is calculated according to the equipment identity information and phase segment in the multi-source phase time sequence sample. Then, the gated recurrent neural network is trained with the maintenance confirmed fault sample to generate the equipment phase reference sequence. Based on the multi-source phase timing samples and the device phase reference sequence, the current amplitude, voltage amplitude, temperature rise rate, and number of discharge pulses in the current sampled values ​​under the same device identity information and the same phase segment are subtracted from the corresponding median to generate a phase precursor residual sequence. Based on the phase precursor residual sequence and the equipment phase reference sequence, the gated recurrent neural network analyzes the precursor memory value according to adjacent power frequency cycles, compares the precursor memory value with the early warning threshold determined by the quantile of the maintenance fault sample, and generates fault early warning information.

[0005] As a further aspect of the present invention, the generation process of the multi-source phase time series samples includes: The voltage sampling waveform output by the voltage transformer corresponding to the distribution network equipment is obtained. Zero voltage crossing detection is performed on the voltage sampling waveform. Abnormal zero crossing points whose adjacent zero crossing intervals deviate from the rated power frequency half cycle by more than a preset cycle deviation threshold are screened out. A voltage zero crossing point sequence is generated. The cycle duration between adjacent zero crossing points in the same direction is calculated. The phase segment boundary is time-corrected based on the deviation between the cycle duration and the rated power frequency cycle. The corrected zero crossing point in the same direction is determined as the phase start mark. The power frequency cycle is divided according to the time interval between two adjacent phase start marks. Each power frequency cycle is divided into multiple phase segments according to phase sequence. The current amplitude, voltage amplitude, temperature rise rate, number of discharge pulses and switching status information corresponding to the same device identity information are associated in each phase segment to generate the multi-source phase timing sample.

[0006] As a further aspect of the present invention, the process of determining the phase start mark includes: The voltage polarity change relationship between adjacent sampling points in the voltage sampling waveform is obtained, and candidate zero-crossing positions formed by negative to positive changes are screened. Based on the deviation relationship between the time interval between adjacent candidate zero-crossing positions and the rated power frequency period, candidate zero-crossing positions that exceed the preset period deviation threshold are eliminated, and an effective zero-crossing position sequence is generated. Based on the voltage amplitude change trend of sampling points on both sides of multiple effective zero-crossing positions within the effective zero-crossing position sequence, the interpolation time point that crosses zero voltage is determined. A clock synchronization pulse sent by an external time synchronization is received, and the deviation between the interpolation time point and the clock synchronization pulse is checked. When the deviation does not exceed one sampling period, the interpolation time point is determined as the phase start mark.

[0007] As a further aspect of the present invention, the process of generating the device phase reference sequence includes: The maintenance action record and switch status information are obtained. When the maintenance is not triggered and the switch status information does not change state within multiple consecutive power frequency cycles, the corresponding time interval is determined as the reference establishment period. The reference sample record located within the reference establishment period is selected from the multi-source phase timing sample. The reference sample records are grouped according to the device identification information and phase segment. The median of the current amplitude, voltage amplitude, temperature rise rate and discharge pulse number in the same group are calculated to establish a device phase median array. Obtain a fault sample confirmed by inspection, input the equipment phase median array and the fault sample confirmed by inspection into a gated recurrent neural network for training, and associate the phase timing reference output obtained by training with the equipment phase median array to generate the equipment phase reference sequence.

[0008] As a further aspect of the present invention, the process of determining the benchmark establishment time period includes: Obtain the action identifier corresponding to the protection action record, the status code corresponding to the switch status information, and the sampling timestamp corresponding to each status code. When the action identifier is continuously in the inactive state and there is no state switch between opening and closing between adjacent status codes, the corresponding continuous time interval is determined as the candidate reference interval. Calculate the number of power frequency cycles covered by the candidate reference interval. If the number of power frequency cycles reaches a preset threshold for the number of reference cycles, then the candidate reference interval is determined as the reference establishment period.

[0009] As a further aspect of the present invention, the training input generation process for troubleshooting and confirming fault samples includes: Obtain the fault confirmation time, fault equipment identity information and fault type identifier from the maintenance work order, extract continuous power frequency cycle samples located before the fault confirmation time from the multi-source phase time series samples, and establish a fault sample window according to the fault equipment identity information and phase segment. The current amplitude, voltage amplitude, temperature rise rate, and number of discharge pulses within the fault sample window are compared with their corresponding medians to obtain a fault training residual sequence, which is then used as the training input for a gated recurrent neural network.

[0010] As a further aspect of the present invention, the generation process of the phase precursor residual sequence includes: Obtain the current sampling record in the multi-source phase timing sample, read the device identity information, phase segment number, current amplitude, voltage amplitude, temperature rise rate and discharge pulse count corresponding to the current sampling record, and match the median corresponding to the same device identity information and the same phase segment number in the device phase reference sequence; The current amplitude, voltage amplitude, temperature rise rate, and number of discharge pulses in the current sampling record are subtracted from their corresponding medians and written into the residual buffer queue in chronological order according to the power frequency cycle to generate the phase precursor residual sequence.

[0011] As a further aspect of the present invention, the matching process for the device phase median record is as follows: A joint matching key is established based on the device identity information and phase segment number. A reference record that matches the joint matching key is retrieved in the device phase reference sequence. When a unique reference record is found, the median current amplitude, median voltage amplitude, median temperature rise rate, and median number of discharge pulses in the unique reference record are determined as the device phase median record. When no reference record is found or multiple reference records are found, the current sampled record is marked as a reference exception record, and the reference exception record is prevented from being written to the residual cache queue.

[0012] As a further aspect of the present invention, the process of generating the fault early warning information includes: The phase precursor residual sequence and the device phase reference sequence are obtained. An inference window is established according to the device identity information, phase segment number and adjacent power frequency cycle order. The current residual, voltage residual, temperature rise rate residual and discharge pulse residual in the inference window are input into the gated recurrent neural network to generate precursor memory values. The warning threshold is determined by reading the quantile of the fault sample. The precursor memory value is compared with the warning threshold. When the precursor memory value reaches or exceeds the warning threshold, the fault advance warning information containing equipment identity information, phase segment number, sampling timestamp and warning status is generated.

[0013] As a further aspect of the present invention, the process of establishing the inference window includes: According to the time sequence corresponding to the phase start mark, read the phase residual records of consecutive adjacent power frequency cycles under the same device identity information and the same phase segment number. When the number of consecutive phase residual records reaches the length of the input window of the gated recurrent neural network and there are no invalid residual marks, the consecutive phase residual records are determined as candidate inference windows. The candidate inference window is checked for consistency with the device identity information and phase segment number in the device phase reference sequence. If the consistency check passes, the candidate inference window is determined as the inference window.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the phase starting point is determined by the voltage zero crossing and the sampling is divided according to the power frequency cycle, so that the current, voltage, temperature rise rate, discharge pulse and switching state fall into the same phase context. The median is obtained by combining the equipment identity and phase segment during the reference establishment period, forming a phase reference that can be distinguished according to equipment differences. After the current sampled value is converted into phase precursor residual, the load fluctuation and individual reference offset are separated. The continuous change of residual in adjacent power frequency cycles enters the gated recurrent neural network memory analysis. The fault sample quantile of the maintenance confirmation gives the threshold boundary. The early warning judgment is shifted from static state recognition to precursor evolution recognition, reducing the risk of false alarms and missed alarms caused by historical averages masking transient anomalies. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the distribution network equipment fault early warning method of the present invention; Figure 2 This is a schematic diagram illustrating the multi-source phase time-series sample generation effect of the present invention; Figure 3 This is a schematic diagram illustrating the effect of establishing the device phase reference sequence according to the present invention. Figure 4 This is a schematic diagram illustrating the phase precursor residual sequence generation effect of the present invention; Figure 5 This is a schematic diagram illustrating the effect of generating early warning information for faults according to the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] This embodiment provides a method for early warning of faults in distribution network equipment based on multi-source time-series AI fusion. During the continuous operation of distribution network equipment such as distribution lines, switchgear, ring main units, pole-mounted switches or cable accessories, voltage transformers, current sampling records, temperature change records, partial discharge records, switch status records and maintenance confirmation records generate operating data from different sources. This method establishes a continuous processing process around the same distribution network equipment, the same phase segment and adjacent power frequency cycles, so that data from different sources can enter the early warning judgment under the condition that the phase reference is consistent.

[0019] In the process of early warning of faults in distribution network equipment, the original current, voltage, temperature rise rate, and number of discharge pulses have different physical meanings. If they are directly mixed for judgment, phase differences, switch status changes, or protection action status can easily affect the identification of fault precursors. In this embodiment, a phase start mark is first formed by voltage zero crossing, and then a phase reference sequence of equipment is established in the stable operating range. Subsequently, the current sampling record is converted into a phase precursor residual sequence, and the precursor memory value is extracted in adjacent power frequency cycles through a gated recurrent neural network, so that the early warning information comes from the continuous deviation process of the same equipment and the same phase segment.

[0020] Method Implementation Examples Please see Figure 1 and Figure 2 S1: Obtain the phase start mark formed by the voltage zero crossing, divide the power frequency cycle according to the phase start mark, collect the current amplitude, voltage amplitude, temperature rise rate, discharge pulse count, switch status information and equipment identity information of the distribution network equipment in the same phase segment, and generate multi-source phase time sequence samples.

[0021] Among them, the phase start mark refers to the period start mark determined by the zero-crossing position in the same direction in the voltage sampling waveform. Its source is the voltage sampling waveform output by the voltage transformer corresponding to the distribution network equipment, and it is used to divide the continuous time axis into power frequency cycles that can be arranged in phase order. Multi-source phase time sequence samples refer to phased records that carry current amplitude, voltage amplitude, temperature rise rate, discharge pulse count, and switch status information, indexed by equipment identity information, phase segment, sampling timestamp, and power frequency cycle. The generation result serves as the common input for subsequent establishment of reference sequences, generation of residual sequences, and formation of early warning information.

[0022] When generating the phase start mark, the voltage polarity change relationship between adjacent sampling points in the voltage sampling waveform is continuously detected, and candidate zero-crossing positions formed by negative to positive changes are screened. Candidate zero-crossing positions are not directly used as the cycle start point, but are first verified according to the correspondence between the time interval between adjacent candidate zero-crossing positions and the rated power frequency cycle. When the interval between adjacent candidate zero-crossing positions deviates from the preset cycle deviation boundary allowed by the rated power frequency cycle, the candidate zero-crossing position is marked as an abnormal zero-crossing point and excluded from the candidate sequence; when a candidate zero-crossing position meets the cycle consistency requirement, it is added to the valid zero-crossing position sequence.

[0023] For each valid zero-crossing position within the valid zero-crossing position sequence, the voltage amplitude change trend of the sampling points on both sides is read. Based on the direction of voltage amplitude change before and after the zero-crossing and the absolute value ratio of the voltage amplitudes of the sampling points on both sides, a linear interpolation method is used to accurately calculate and determine the interpolation time point that crosses the zero voltage. The interpolation time point is a time marker located between adjacent sampling points before and after the zero-crossing, converted by linear proportion to represent the moment when the voltage crosses the zero voltage. This interpolation time point is also checked for deviation from the clock synchronization pulse. The allowable boundary for deviation check comes from the synchronization allowable error corresponding to the sampling period. When the deviation between the interpolation time point and the clock synchronization pulse does not exceed the synchronization allowable boundary, the interpolation time point is determined as the phase start mark; when the deviation exceeds the synchronization allowable boundary, the zero-crossing position is marked as a clock check anomaly and prevented from participating in the power frequency cycle division.

[0024] After the phase start markers are determined, the power frequency cycle is divided according to the time interval between two adjacent phase start markers, and each power frequency cycle is further divided into multiple phase segments according to phase sequence. A phase segment is a time segment formed in phase sequence within a power frequency cycle, and its boundary is determined by the phase start markers and the cycle duration. If the cycle duration between adjacent zero-crossing points in the same direction deviates from the rated power frequency cycle, the phase segment boundaries are time-corrected based on this deviation, so that the phase segments within the same cycle are still arranged according to the phase progression of the actual voltage waveform. This time correction does not change the original timestamp of the sampling record itself, but rather changes the boundary of the phase segment to which the sampling record is assigned, enabling phase comparability of multi-source data within the same phase segment.

[0025] Within each phase segment, the current amplitude, voltage amplitude, temperature rise rate, number of discharge pulses, and switch status information corresponding to the same distribution network equipment are associated according to the equipment identification information. Equipment identification information refers to the identification field that distinguishes different distribution network equipment; in this method, it is used to group multi-source sampling records within the same phase segment to the same object. Current and voltage amplitudes are derived from electrical sampling records within the same phase segment; the temperature rise rate is derived from the temperature change status of the same equipment between adjacent sampling times; the number of discharge pulses is derived from discharge event count records within the same phase segment; and the switch status information is derived from the status codes corresponding to open, closed, or stable states. After aligning all fields according to the same equipment identification information, the same power frequency cycle, and the same phase segment, multi-source phase time sequence samples are formed.

[0026] In one embodiment, if a phase segment contains missing samples, duplicate samples, zero-crossing anomalies, timestamps that cannot be assigned to the power frequency cycle, or unreadable switch status records, the corresponding records are marked as phase sampling anomaly records. Phase sampling anomaly records retain their device identification information, timestamps, and anomaly type, but are not included in subsequent median calculation and residual cache queues. If multiple available records for the same device identification information exist within the same phase segment, the valid record matching the boundary of that phase segment is selected based on the order of sampling timestamps and included in the multi-source phase time series sample; duplicate records are excluded as redundant records. Through this anomaly handling, data entering subsequent steps all have clear phase attribution, device attribution, and field integrity.

[0027] In this embodiment, multi-source operating data is processed by voltage zero-crossing verification, phase segment boundary correction, and device identity information association, so that current amplitude, voltage amplitude, temperature rise rate, discharge pulse count, and switch status information can form traceable samples within the same phase segment, thereby providing a phase-consistent data foundation for subsequent reference establishment and precursor residual generation.

[0028] Please see Figure 1 and Figure 3 S2: For the reference establishment period when maintenance is not triggered and the switch state is stable, according to the equipment identity information and phase segment in the multi-source phase time sequence sample, the median of current amplitude, voltage amplitude, temperature rise rate and discharge pulse number are calculated, and the gated recurrent neural network is trained with the maintenance confirmed fault sample to generate the equipment phase reference sequence.

[0029] The reference establishment period refers to the time interval during which protection action records remain in an inactive state and switch status information does not switch between opening and closing within a continuous power frequency cycle. It is used to indicate that the distribution network equipment is in a stable operating state that can be used as a reference. The equipment phase reference sequence refers to a reference data sequence that records the median current amplitude, median voltage amplitude, median temperature rise rate, median discharge pulse count, and phase timing reference output, using equipment identity information and phase segments as indexes. The median refers to the centrally located representative value obtained by arranging available records for the same field within the same group according to the field value order. When there are parallel centralized situations among available records within a group, a preset centralization rule is used to determine the representative value, maintaining consistency in the processing rules for the same field across all groups.

[0030] When determining the reference establishment period, the action identifier corresponding to the protection action record, the status code corresponding to the switch status information, and the sampling timestamp corresponding to each status code are read. When the action identifier remains in an inactive state and there is no state transition between opening and closing between adjacent status codes, the corresponding continuous time interval is determined as a candidate reference interval. Subsequently, the number of power frequency cycles covered by the candidate reference interval is read and compared with the judgment condition corresponding to the preset reference cycle number threshold. The preset reference cycle number threshold is derived from the operational stability requirements of the distribution network equipment or historical operation and maintenance rules, and is used to limit the continuity of the reference interval. When the candidate reference interval meets this continuity requirement, it is determined as the reference establishment period; when the candidate reference interval does not meet this continuity requirement, it is only retained as a normal operating interval and does not participate in the establishment of the equipment phase reference sequence.

[0031] After determining the baseline establishment period, baseline sample records within that period are selected from the multi-source phase time series samples. The baseline sample records are grouped according to equipment identity information and phase segment. The median values ​​of current amplitude, voltage amplitude, temperature rise rate, and discharge pulse count within the same group are calculated to obtain the equipment phase median array. The equipment phase median array refers to a stable reference set for each operating field under the same equipment identity information and the same phase segment. It does not include switch status information but retains the source marker corresponding to the switch status stability condition. This source marker indicates that the corresponding median comes from a period when maintenance was not triggered and the switch status was stable. Subsequent matching of the current sampled record uses equipment identity information and phase segment as the joint matching basis.

[0032] Fault samples confirmed during maintenance refer to the source of fault samples defined by the fault confirmation time, faulty equipment identification information, and fault type identifier in the maintenance work order. The fault confirmation time is used to determine the range of continuous power frequency cycle samples before the fault occurred; the faulty equipment identification information is used to filter samples from the same distribution network equipment; and the fault type identifier is used to distinguish the precursor changes corresponding to different fault categories during training. For continuous power frequency cycle samples before the fault confirmation time, a fault sample window is established according to the faulty equipment identification information and phase segment. The current amplitude, voltage amplitude, temperature rise rate, and discharge pulse count within the fault sample window are respectively compared with the records in the corresponding equipment phase median array to obtain the fault training residual sequence. The fault training residual sequence refers to the training input records arranged in adjacent power frequency cycle order, carrying the multi-source deviation state before the fault occurred.

[0033] A gated recurrent neural network (RNN) is a recurrent neural network structure used to process residual sequences of adjacent power frequency cycles. In this method, it includes an input conversion layer, an update gate processing layer, a reset gate processing layer, a candidate memory processing layer, a hidden state propagation layer, and an output layer. The input conversion layer receives current residuals, voltage residuals, temperature rise rate residuals, and discharge pulse residuals from the fault training residual sequence and converts residual fields from different sources into the same timing input format. The update gate processing layer determines the degree of historical memory retention based on the current cycle input and the hidden state of the previous cycle. The reset gate processing layer determines the degree to which historical memory participates in the current candidate state based on the current cycle input and the hidden state of the previous cycle. The candidate memory processing layer forms the current candidate memory based on the current cycle input and the historical state after reset processing. The hidden state propagation layer forms the precursor memory propagation state of the current power frequency cycle under the constraints of the update gate processing result. The output layer converts the hidden state into a phase timing reference output. Data is passed between these layers sequentially according to adjacent power frequency cycles, ensuring that continuous deviations before the fault can be retained as timing memories.

[0034] During training, the equipment phase median array and the fault training residual sequence corresponding to the fault samples confirmed by maintenance are input into a gated recurrent neural network. During training, the fault type identifier in the maintenance work order and the sample window assignment before fault confirmation are used as supervision criteria to compare the phase timing reference result output by the network with the maintenance confirmation result. When the output result is inconsistent with the maintenance confirmation result, the network's internal parameters are adjusted based on error feedback, enabling the network to form different memory states for continuous deviation, recovery deviation, and abrupt deviation within adjacent power frequency cycles. Training is stopped based on preset training completion conditions, which can be determined by the stable state of training error, the completion state of sample traversal, or maintenance confirmation rules. After training is completed, the obtained phase timing reference output is associated with the equipment phase median array to generate the equipment phase reference sequence.

[0035] In one embodiment, if a device's identity information and phase segment lack complete fields during the reference establishment period, or if there are unclear protection action records, discontinuous switch state reversal records, or fault confirmation times that cannot be aligned with multi-source phase timing samples, then the corresponding group will not generate a valid device phase median array. If the fault sample window in the gated recurrent neural network training input lacks continuous power frequency cycles, then the obtained continuous valid cycles are used to form a temporary training window, with a historical insufficiency marker attached; when a complete continuous cycle is subsequently obtained, the complete window replaces the temporary training window for reference updates. The historical insufficiency marker is not used as a fault warning result output, but only to limit the scope of use of training input and reference updates.

[0036] In this embodiment, the benchmark establishment period is defined by the non-triggering of maintenance and the stable switching state, and a median array is established according to the equipment identity information and phase segment, so that the inherent operating differences of the equipment and the inherent fluctuations of the phase are included in the reference sequence. By training the gated recurrent neural network through maintenance confirmation of fault samples, the continuous residual changes before the fault can form a phase timing reference output, thereby providing a callable reference basis for the timing judgment of the current residual.

[0037] Please see Figure 1 and Figure 4 S3: Based on the multi-source phase timing samples and the device phase reference sequence, subtract the corresponding median from the current amplitude, voltage amplitude, temperature rise rate and discharge pulse count in the current sampled values ​​under the same device identity information and the same phase segment to generate a phase precursor residual sequence.

[0038] The phase precursor residual sequence refers to the deviation record of the current sampled record from the corresponding median in the device phase reference sequence, and is written into the residual buffer queue in chronological order according to the power frequency cycle. The residual buffer queue is a time-series queue that stores phase residual records in the order of device identity information, phase segment number, and power frequency cycle, and its contents are used to form the subsequent inference window. The current sampled record refers to the latest available record in the multi-source phase time series samples waiting to enter the early warning judgment, and its fields include device identity information, phase segment number, current amplitude, voltage amplitude, temperature rise rate, and number of discharge pulses.

[0039] When generating the phase precursor residual sequence, the device identification information and phase segment number corresponding to the current sampling record are first read, and a joint matching key is established using the device identification information and phase segment number. The joint matching key refers to the retrieval basis formed by the same device identification information and the same phase segment number, which is used to locate a unique reference record in the device phase reference sequence. If a unique reference record is found, the median current amplitude, median voltage amplitude, median temperature rise rate, and median discharge pulse count are read from the reference record, and the corresponding fields in the current sampling record are subtracted from the median field to obtain the current residual, voltage residual, temperature rise rate residual, and discharge pulse residual.

[0040] The current residual refers to the deviation of the current amplitude in the current phase segment from the stable reference current amplitude in the same phase segment of the same equipment. The voltage residual refers to the deviation of the current voltage amplitude in the current phase segment from the stable reference voltage amplitude. The temperature rise rate residual refers to the deviation of the current temperature change rate from the stable reference temperature change state. The discharge pulse residual refers to the deviation of the current discharge pulse count from the stable reference discharge state. All of the above residual fields retain their original physical meaning and are not added or directly mixed in this step. Instead, they are written into the residual buffer queue as independent channels in the same power frequency cycle order, for the gated recurrent neural network to read in subsequent steps according to the timing relationship.

[0041] During the joint matching process, if no reference record is found, it indicates that the device identity information or phase segment to which the current sampled record belongs has not yet established a valid reference, and the current sampled record is marked as a reference missing record. If multiple reference records are found, it indicates that there are multiple reference records corresponding to the same joint matching key in the device phase reference sequence, and the current sampled record is marked as a reference conflict record. Reference missing records and reference conflict records together constitute reference anomaly records. Reference anomaly records retain the original fields and anomaly flags, but are not written to the residual cache queue to avoid invalid references affecting the establishment of subsequent inference windows.

[0042] In one embodiment, the residual cache queue maintains the power frequency cycle order according to the device identity information and phase segment number. For the startup phase or the initial cycle after the reference sequence has just been established, if the continuous residual records are not yet sufficient to form a complete inference window, the generated valid phase residual records first enter the cache waiting state and do not directly trigger the warning judgment; when the continuous valid residual records under the same device identity information and the same phase segment number meet the integrity condition corresponding to the input window length of the gated recurrent neural network, it then switches to the stable inference state. If an invalid residual mark appears in the cache, the continuous records are reorganized with the invalid mark as the boundary, and the inference window is not formed by crossing the invalid mark.

[0043] In this embodiment, the current sampling record is matched with the unique median record in the device phase reference sequence by a joint matching key, and the sampling values ​​from different sources are converted into residual fields under the same phase and the same device. This transforms the current operating state from an absolute sampling value into a deviation sequence relative to a stable reference, thereby providing continuous and verifiable input for subsequent aura memory analysis.

[0044] Please see Figure 1 and Figure 5 S4: Based on the phase precursor residual sequence and the equipment phase reference sequence, the gated recurrent neural network analyzes the precursor memory value according to adjacent power frequency cycles, compares the precursor memory value with the early warning threshold determined by the quantile of the maintenance fault sample, and generates fault early warning information.

[0045] The equipment phase reference sequence refers to a deep time-series reference record formed by expanding the equipment phase reference sequence. It includes equipment identification information, phase segment number, equipment phase median record, phase time-series reference output obtained from gated recurrent neural network training, and a warning threshold field corresponding to the quantile of the maintenance fault sample. The precursor memory value refers to the fault precursor time-series state value output by the gated recurrent neural network after reading continuous phase residual records. It is used to represent the continuous deviation memory of the same equipment and the same phase segment in adjacent power frequency cycles. The warning threshold is a configurable judgment boundary determined by the quantile of the maintenance fault sample. It originates from the sorting distribution of precursor memory values ​​and the warning strategy in the maintenance confirmed fault samples and is not temporarily generated by the current sampling record.

[0046] When establishing the inference window, phase residual records of consecutive adjacent power frequency cycles under the same device identity information and the same phase segment number are read according to the chronological order corresponding to the phase start marker. The inference window refers to a set of consecutive phase residual records that meets the input window length requirement of the gated recurrent neural network and has no invalid residual markers. During the reading process, if the number of consecutive phase residual records has not yet reached the integrity condition corresponding to the input window length, the system remains in a waiting state; if there are reference anomalies, phase sampling anomalies, timestamp interruptions, or invalid residual markers in the consecutive records, the consecutive record is not determined as an inference window, and the reading of consecutive records restarts from after the invalid marker.

[0047] Once a continuous phase residual record meets the integrity condition, it is used as a candidate inference window and its consistency is checked against the device identity information and phase segment number in the device phase reference sequence. If the device identity information and phase segment number of the candidate inference window are consistent with the reference record in the device phase reference sequence, the candidate inference window is determined as the inference window; if the consistency check fails, the candidate inference window is marked as an inconsistent reference record and is prevented from entering the gated recurrent neural network for inference. This consistency check ensures that the inference window can only call deep reference records trained under the same device and the same phase segment.

[0048] During inference, the residuals of current, voltage, temperature rise rate, and discharge pulse, arranged in order of power frequency cycle within the inference window, are input into a gated recurrent neural network. The input conversion layer first maintains the channel boundaries of each residual field to prevent different physical quantities from being directly mixed before entering the timing processing. The update gate processing layer reads the current cycle residual and the hidden state of the previous cycle to determine the degree to which historical precursor states are retained in the current cycle; the reset gate processing layer reads the current cycle residual and the hidden state of the previous cycle to determine the degree of participation of historical precursor states in the formation of the current candidate memory; the candidate memory processing layer forms the current candidate precursor state based on the current cycle residual and the reset historical state; the hidden state transfer layer performs gated fusion of the current candidate precursor state and the historical state to form the precursor memory transfer state of the current cycle; the output layer generates the precursor memory value based on the final hidden state. The above processing is performed sequentially in adjacent power frequency cycles, so that short-term abrupt changes, continuous deviations, and gradually accumulated deviations enter the output result through different gate states.

[0049] The warning threshold is determined by the quantiles of the precursor memory values ​​corresponding to the fault samples confirmed during maintenance. The quantile determination process uses a numerical sorting rule: first, the precursor memory values ​​corresponding to the fault samples are arranged in order of their absolute values; then, a judgment boundary for triggering the warning is selected according to the preset quantile position, and this judgment boundary is written into the warning threshold field of the equipment phase reference sequence. The warning threshold field includes a threshold source identifier, applicable equipment identification information, applicable phase segment number, applicable fault type identifier, and activation status. The threshold source identifier indicates that the threshold comes from the quantile results of the fault samples confirmed during maintenance, and the activation status indicates whether the threshold can be invoked by the current inference window.

[0050] After the precursor memory value is generated, the warning threshold field corresponding to the same device identity information, the same phase segment number, and the applicable fault type identifier is read, and the precursor memory value is compared with the warning threshold. When the precursor memory value reaches or exceeds the judgment boundary corresponding to the warning threshold field, a fault advance warning information is generated; when the precursor memory value does not reach the judgment boundary, a status record of no warning is generated, and the system continues to wait for the phase residual record of the next power frequency cycle. The fault advance warning information includes device identity information, phase segment number, sampling timestamp, and warning status. The warning status indicates that the current inference window has met the fault advance warning triggering conditions, and the sampling timestamp comes from the timestamp corresponding to the current cycle in the inference window.

[0051] In one embodiment, if the warning threshold field in the device phase reference sequence is in an inactive state, or if the threshold source identifier cannot correspond to the fault sample confirmed by maintenance, the current inference window will not generate a fault advance warning information and will output a threshold unavailable status record. If the gated recurrent neural network cannot generate a precursor memory value during inference due to missing input window fields, interrupted residual order, or failed reference consistency check, the corresponding inference window will be marked as an invalid inference record and will wait for subsequent consecutive valid residual records to re-establish the inference window. The invalid inference record does not change the device phase reference sequence and is only used to indicate that the current warning judgment did not obtain a valid output.

[0052] In this embodiment, an inference window for the same equipment, the same phase segment, and adjacent power frequency cycles is established by using the phase precursor residual sequence and the equipment phase reference sequence, enabling the gated recurrent neural network to form precursor memory values ​​in continuous residual changes. The precursor memory values ​​are judged by the warning threshold determined by the maintenance fault sample quantile, so that the fault advance warning information and the maintenance confirmed fault sample form a source closed loop, thereby outputting a warning result containing equipment identity information, phase segment number, sampling timestamp, and warning status before the protection action.

[0053] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for early warning of faults in distribution network equipment based on multi-source time-series AI fusion, characterized in that, Includes the following steps: The phase start mark formed by the voltage zero crossing is obtained, the power frequency cycle is divided according to the phase start mark, and the current amplitude, voltage amplitude, temperature rise rate, discharge pulse count, switch status information and equipment identity information of the distribution network equipment in the same phase segment are collected to generate multi-source phase time series samples. For the reference establishment period when maintenance is not triggered and the switch state is stable, the median of current amplitude, voltage amplitude, temperature rise rate and discharge pulse number is calculated according to the equipment identity information and phase segment in the multi-source phase time sequence sample. Then, the gated recurrent neural network is trained with the maintenance confirmed fault sample to generate the equipment phase reference sequence. Based on the multi-source phase timing samples and the device phase reference sequence, the current amplitude, voltage amplitude, temperature rise rate, and number of discharge pulses in the current sampled values ​​under the same device identity information and the same phase segment are subtracted from the corresponding median to generate a phase precursor residual sequence. Based on the phase precursor residual sequence and the equipment phase reference sequence, the gated recurrent neural network analyzes the precursor memory value according to adjacent power frequency cycles, compares the precursor memory value with the early warning threshold determined by the quantile of the maintenance fault sample, and generates fault early warning information.

2. The method for early warning of faults in distribution network equipment based on multi-source time-series AI fusion according to claim 1, characterized in that, The generation process of the multi-source phase time series samples includes: The voltage sampling waveform output by the voltage transformer corresponding to the distribution network equipment is obtained. Zero voltage crossing detection is performed on the voltage sampling waveform. Abnormal zero crossing points whose adjacent zero crossing intervals deviate from the rated power frequency half cycle by more than a preset cycle deviation threshold are screened out. A voltage zero crossing point sequence is generated. The cycle duration between adjacent zero crossing points in the same direction is calculated. The phase segment boundary is time-corrected based on the deviation between the cycle duration and the rated power frequency cycle. The corrected zero crossing point in the same direction is determined as the phase start mark. The power frequency cycle is divided according to the time interval between two adjacent phase start marks. Each power frequency cycle is divided into multiple phase segments according to phase sequence. The current amplitude, voltage amplitude, temperature rise rate, number of discharge pulses and switching status information corresponding to the same device identity information are associated in each phase segment to generate the multi-source phase timing sample.

3. The method for early warning of faults in distribution network equipment based on multi-source time-series AI fusion according to claim 2, characterized in that, The process of determining the phase start mark includes: The voltage polarity change relationship between adjacent sampling points in the voltage sampling waveform is obtained, and candidate zero-crossing positions formed by negative to positive changes are screened. Based on the deviation relationship between the time interval between adjacent candidate zero-crossing positions and the rated power frequency period, candidate zero-crossing positions that exceed the preset period deviation threshold are eliminated, and an effective zero-crossing position sequence is generated. Based on the voltage amplitude change trend of sampling points on both sides of multiple effective zero-crossing positions within the effective zero-crossing position sequence, the interpolation time point that crosses zero voltage is determined. A clock synchronization pulse sent by an external time synchronization is received, and the deviation between the interpolation time point and the clock synchronization pulse is checked. When the deviation does not exceed one sampling period, the interpolation time point is determined as the phase start mark.

4. The method for early warning of faults in distribution network equipment based on multi-source time-series AI fusion according to claim 1, characterized in that, The process of generating the device phase reference sequence includes: The maintenance action record and switch status information are obtained. When the maintenance is not triggered and the switch status information does not change state within multiple consecutive power frequency cycles, the corresponding time interval is determined as the reference establishment period. The reference sample record located within the reference establishment period is selected from the multi-source phase timing sample. The reference sample records are grouped according to the device identification information and phase segment. The median of the current amplitude, voltage amplitude, temperature rise rate and discharge pulse number in the same group are calculated to establish a device phase median array. Obtain a fault sample confirmed by inspection, input the equipment phase median array and the fault sample confirmed by inspection into a gated recurrent neural network for training, and associate the phase timing reference output obtained by training with the equipment phase median array to generate the equipment phase reference sequence.

5. The method for early warning of faults in distribution network equipment based on multi-source time-series AI fusion according to claim 4, characterized in that, The process of determining the benchmark establishment period includes: Obtain the action identifier corresponding to the protection action record, the status code corresponding to the switch status information, and the sampling timestamp corresponding to each status code. When the action identifier is continuously in the inactive state and there is no state switch between opening and closing between adjacent status codes, the corresponding continuous time interval is determined as the candidate reference interval. Calculate the number of power frequency cycles covered by the candidate reference interval. If the number of power frequency cycles reaches a preset threshold for the number of reference cycles, then the candidate reference interval is determined as the reference establishment period.

6. The method for early warning of faults in distribution network equipment based on multi-source time-series AI fusion according to claim 1, characterized in that, The process of generating training inputs for fault samples to confirm their repair includes: Obtain the fault confirmation time, fault equipment identity information and fault type identifier from the maintenance work order, extract continuous power frequency cycle samples located before the fault confirmation time from the multi-source phase time series samples, and establish a fault sample window according to the fault equipment identity information and phase segment. The current amplitude, voltage amplitude, temperature rise rate, and number of discharge pulses within the fault sample window are compared with their corresponding medians to obtain a fault training residual sequence, which is then used as the training input for a gated recurrent neural network.

7. The method for early warning of faults in distribution network equipment based on multi-source time-series AI fusion according to claim 1, characterized in that, The generation process of the phase precursor residual sequence includes: Obtain the current sampling record in the multi-source phase timing sample, read the device identity information, phase segment number, current amplitude, voltage amplitude, temperature rise rate and discharge pulse count corresponding to the current sampling record, and match the median corresponding to the same device identity information and the same phase segment number in the device phase reference sequence; The current amplitude, voltage amplitude, temperature rise rate, and number of discharge pulses in the current sampling record are subtracted from their corresponding medians and written into the residual buffer queue in chronological order according to the power frequency cycle to generate the phase precursor residual sequence.

8. The method for early warning of faults in distribution network equipment based on multi-source time-series AI fusion according to claim 7, characterized in that, The matching process for the median phase record of the device is as follows: A joint matching key is established based on the device identity information and phase segment number. A reference record that matches the joint matching key is retrieved in the device phase reference sequence. When a unique reference record is found, the median current amplitude, median voltage amplitude, median temperature rise rate, and median number of discharge pulses in the unique reference record are determined as the device phase median record. When no reference record is found or multiple reference records are found, the current sampled record is marked as a reference exception record, and the reference exception record is prevented from being written to the residual cache queue.

9. The method for early warning of faults in distribution network equipment based on multi-source time-series AI fusion according to claim 1, characterized in that, The process of generating the fault early warning information includes: The phase precursor residual sequence and the device phase reference sequence are obtained. An inference window is established according to the device identity information, phase segment number and adjacent power frequency cycle order. The current residual, voltage residual, temperature rise rate residual and discharge pulse residual in the inference window are input into the gated recurrent neural network to generate precursor memory values. The warning threshold is determined by reading the quantile of the fault sample. The precursor memory value is compared with the warning threshold. When the precursor memory value reaches or exceeds the warning threshold, the fault advance warning information containing equipment identity information, phase segment number, sampling timestamp and warning status is generated.

10. The method for early warning of faults in distribution network equipment based on multi-source time-series AI fusion according to claim 9, characterized in that, The process of establishing the inference window includes: According to the time sequence corresponding to the phase start mark, read the phase residual records of consecutive adjacent power frequency cycles under the same device identity information and the same phase segment number. When the number of consecutive phase residual records reaches the length of the input window of the gated recurrent neural network and there are no invalid residual marks, the consecutive phase residual records are determined as candidate inference windows. The candidate inference window is checked for consistency with the device identity information and phase segment number in the device phase reference sequence. If the consistency check passes, the candidate inference window is determined as the inference window.

Images