Power grid equipment hidden danger grading judgment method and system based on deep learning

Abstract

The invention relates to the technical field of neural networks, in particular to a power grid equipment hidden danger grading judgment method and system based on deep learning, and the method comprises the following steps: obtaining continuous periodic voltage data, recognizing a voltage change trend section sequence, tracking the continuous change of multiple parameters to form a response track fluctuation section set, and adjusting a tag time sequence adjustment result sequence according to a change flow direction, comparing channel response differences to identify a channel response non-uniform state section set, and updating tag output in combination with trend departure to obtain a tag continuous output judgment sequence. According to the invention, by identifying the voltage trend and the multi-parameter direction change, constructing the dynamic corresponding relation between the trend section and the label output, the time perception capability and the change response speed of the label to the periodic change are improved, the channel response difference is analyzed to extract the abnormal rhythm section, and the label content is adjusted when the trend direction deviates. The matching degree of label output and actual state evolution is enhanced, and the analysis horizontal force of abnormal continuity and feature fluctuation is improved.

Description

Technical Field

[0001] This invention relates to the field of neural network technology, and in particular to a method and system for classifying and determining potential hazards in power grid equipment based on deep learning. Background Technology

[0002] The field of neural network technology involves mathematical models and algorithmic systems that simulate the connections between neurons in the human brain. Core aspects include neural network structure design, network training methods, learning strategy optimization, weight update mechanisms, and activation function selection. This technology is widely used in various scenarios such as pattern recognition, image processing, speech recognition, natural language processing, and fault diagnosis. It possesses feature extraction, autonomous learning, and modeling capabilities, and is particularly suitable for modeling and discriminant analysis tasks of nonlinear, high-dimensional, and complex systems. In recent years, with the development of deep learning, neural networks have expanded from traditional feedforward structures to various types such as convolutional neural networks, recurrent neural networks, and generative adversarial networks, becoming one of the key supporting technologies for intelligent perception and decision-making. Traditional methods for classifying and determining potential risks in power grid equipment refer to the identification and classification of potential risk levels of equipment during power system operation and maintenance through manual inspections, periodic maintenance, and judgment based on empirical rules. These methods typically rely on the subjective judgment of maintenance personnel regarding the equipment's operating status or on preliminary classification of equipment status using set physical parameter thresholds. These methods generally rely on historical fault data, typical defect characteristics, and changes in equipment operating parameters, combined with a static rule base for logical judgment, and complete the preliminary determination of the risk level through table comparison and empirical value reference.

[0003] Existing technologies rely on setting thresholds for initial state assessment, lacking a mechanism to capture the sequential dependencies between parameter change trends. During cyclical fluctuations, it is difficult to analyze continuous trend information, leading to the continued failure to identify potential risks. The label output cannot be mapped to the actual state evolution rhythm, easily resulting in label update lag. Changes in channel node responses are not associated with abnormal behavior rhythm characteristics, making it difficult to identify potential risks of temporal misalignment. Label trends mostly rely on static comparison methods, lacking dynamic control means for the directional relationship between parameters. Even slight parameter disturbances can easily lead to misjudgments, resulting in a structural deficiency in the ability to identify atypical risks in complex operating environments. Summary of the Invention

[0004] To address the technical problems existing in the prior art, embodiments of the present invention provide a method for classifying and determining potential hazards in power grid equipment based on deep learning;

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for classifying and determining potential hazards in power grid equipment based on deep learning, comprising the following steps:

[0006] S1: Obtain voltage amplitude data of the distribution transformer within a continuous cycle, extract the voltage at the end of the cycle and compare the changing trend during adjacent cycles, identify the duration of the trend, and obtain the voltage change trend segment sequence.

[0007] S2: Based on the voltage change trend segment sequence, extract the amplitude-frequency ratio, current impact offset value and thermal excitation frequency and add them to the period, and track the continuous change direction to obtain the set of response trajectory fluctuation segments;

[0008] S3: Based on the set of response trajectory fluctuation segments, map the direction of voltage, current and thermal excitation changes, use the label content of the previous segment in the same continuous flow direction period segment, and replace the original label to obtain the label timing adjustment result sequence.

[0009] S4: Based on the tag timing adjustment result sequence, extract edge channel activation nodes, compare the timing of continuous period channel response changes, analyze response speed differences, identify difference phenomenon segments, and obtain a set of channel response uneven state segments.

[0010] S5: Based on the set of uneven channel response states, extract the corresponding periodic tag output records and voltage fluctuation data, compare the trend direction with the tag trend, replace the tag content in the period with the difference in trend, and obtain the tag continuous output judgment sequence.

[0011] As a further embodiment of the present invention, the voltage change trend segment sequence includes a period number sequence, tail voltage comparison results, and trend continuous time period identifiers; the response trajectory fluctuation segment set includes an amplitude-frequency ratio sequence, current impact offset changes, and thermal excitation frequency evolution; the tag timing adjustment result sequence includes tag replacement content, period corresponding parameter change direction, and segment tag flow structure; the channel response uneven state segment set includes channel activation node information, response speed differences, and differential step phenomenon segment numbers; and the tag continuous output determination sequence includes tag update content, period number pairing information, and the correspondence between voltage fluctuations and tag trends.

[0012] As a further aspect of the present invention, the voltage at the end of the cycle refers to the voltage amplitude at the end of each power frequency cycle;

[0013] The current surge offset value refers to a parameter that measures the degree to which the current deviates from its average value within a period.

[0014] As a further aspect of the present invention, the parameter timing characteristics refer to a dynamic information sequence that reflects the continuous temporal variation trend of voltage, current, and thermistor parameters;

[0015] The edge channel activation node refers to a node in the neural network structure that is triggered within a period of time in a channel located at the edge of the path.

[0016] As a further aspect of the present invention, the specific steps of S1 are as follows:

[0017] S101: Obtain the voltage amplitude data of the distribution transformer in a continuous cycle, and compare the voltage amplitude at the end of each cycle with the voltage amplitude at the end of the previous cycle to obtain the corresponding sequence of cycle voltages.

[0018] S102: Based on the corresponding sequence of periodic voltages, extract the direction of change between adjacent voltage amplitudes, analyze whether the direction of change is consistent, and obtain a set of trend continuation numbers;

[0019] S103: Based on the trend continuation number set, extract continuous number segments according to the numbering order, and locate the voltage trend of the number segments in the original data to obtain the voltage change trend segment sequence.

[0020] As a further aspect of the present invention, the specific steps of S2 are as follows:

[0021] S201: Based on the voltage change trend segment sequence, monitor the amplitude-frequency ratio, current impulse offset value and thermal excitation frequency corresponding to each cycle, and append the parameters to the cycle data frame in time order to obtain the cycle parameter appending matrix;

[0022] S202: Based on the periodic parameter additional matrix, compare the changes in direction and magnitude of the parameters contained in each period with the corresponding parameters of the previous period, track the trend of numerical changes, and obtain a set of parameter change tracking numbers.

[0023] S203: Based on the parameter change tracking number set, locate the parameter value fluctuation range of the corresponding period in the number segment, analyze the continuous fluctuation sequence segment, and obtain the response trajectory fluctuation segment set.

[0024] As a further aspect of the present invention, the specific steps of S3 are as follows:

[0025] S301: Based on the periodic segments listed in the set of response trajectory fluctuation segments, collect the voltage change trend, current offset amplitude change and thermal excitation point position corresponding to each period, and map the parameters to the time region according to the order before and after the period to obtain the periodic parameter sequence.

[0026] S302: Based on the periodic parameter sequence, trace the direction of parameter change in the period, extract the label content attached to the previous segment, and obtain a set of directional continuous segment labels.

[0027] S303: Based on the continuous segment tag set in the direction, replace the original tag content of the period in numerical order, update the used tags in the period and write them into the current data frame to obtain the tag timing adjustment result sequence.

[0028] As a further aspect of the present invention, the specific steps of S4 are as follows:

[0029] S401: Based on the tag timing adjustment result sequence, collect the activation node information of adjacent cycles in the edge channel in the path structure, and map the response actions of the corresponding nodes of adjacent channels in each cycle to the cycle position in time order to obtain the channel node response sequence;

[0030] S402: Based on the channel node response sequence, compare the response timing of the left and right channels in a continuous period, analyze the relationship between the response start time and the sequence of the continuing action, and obtain the channel response time difference number sequence.

[0031] S403: Based on the channel response time difference number sequence, track the continuous relationship between the numbers and detect the duration range of the inconsistent response speed state to obtain the set of channel response uneven state segments.

[0032] As a further aspect of the present invention, the specific steps of S5 are as follows:

[0033] S501: Based on the set of channel response uneven state segments, call the tag output and voltage fluctuation state data in the corresponding period, and compare the tag trend direction and voltage fluctuation direction in the period in time order to obtain the tag fluctuation order comparison sequence.

[0034] S502: Based on the tag fluctuation sequence comparison, compare the relationship between the tag trend change and the voltage fluctuation direction in the cycle, correct the tag trend direction in the cycle where the trend direction is inconsistent, and obtain the tag direction correction data table.

[0035] S503: Based on the label direction correction data table, the corrected label direction is embedded into the original sequence according to the period number to obtain the label continuous output judgment sequence.

[0036] A deep learning-based system for classifying and determining potential hazards in power grid equipment includes:

[0037] The voltage trend extraction module acquires the voltage amplitude data of the distribution transformer within a continuous cycle, extracts the voltage at the end of the cycle and compares the change trend during adjacent cycles, identifies the duration of the trend, and obtains the voltage change trend segment sequence.

[0038] The parameter trajectory generation module acquires synchronous current and thermal monitoring data, and combines the voltage change trend segment sequence to extract the amplitude-frequency ratio, current impulse offset value and thermal excitation frequency and attach them to the period. Through sequence learning, it analyzes the parameter time series changes to obtain the response trajectory fluctuation segment set.

[0039] The tag timing adjustment module maps voltage, current and thermal excitation direction based on the set of response trajectory fluctuation segments, extracts the initial tag, inherits the tag of the previous segment within the trend-consistent cycle, and obtains the tag timing adjustment result sequence.

[0040] Based on the tag timing adjustment result sequence, the channel response comparison module maps the parameter timing features to the path structure, extracts edge channel activation nodes, compares the periodic channel response changes, analyzes the response speed differences, and obtains a set of channel response uneven state segments.

[0041] The tag output revision module extracts periodic tags and voltage fluctuations based on the set of uneven channel response states, compares the trend direction with the tag trend, and obtains a tag continuous output determination sequence.

[0042] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0043] In this invention, by identifying voltage trends and changes in the direction of multiple parameters, a dynamic correspondence between trend segments and tag output is constructed, which improves the tag's ability to perceive the time of periodic changes and its response speed. By analyzing channel response differences to extract abnormal rhythm segments, the tag content is adjusted when the trend direction deviates, thereby enhancing the matching degree between the tag output and the actual state evolution and improving the analytical level of anomaly continuity and characteristic fluctuations. Attached Figure Description

[0044] 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.

[0045] Figure 1 This is a schematic diagram of the steps of the present invention;

[0046] Figure 2 This is a detailed schematic diagram of S1 of the present invention;

[0047] Figure 3 This is a detailed schematic diagram of S2 of the present invention;

[0048] Figure 4 This is a detailed schematic diagram of S3 of the present invention;

[0049] Figure 5 This is a detailed schematic diagram of S4 of the present invention;

[0050] Figure 6 This is a detailed schematic diagram of S5 of the present invention;

[0051] Figure 7This is a system module diagram of the present invention. Detailed Implementation

[0052] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0053] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0054] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent.

[0055] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0056] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0057] Please see Figure 1 This invention provides a method for classifying and determining potential hazards in power grid equipment based on deep learning, comprising the following steps:

[0058] S1: Obtain the voltage amplitude data of the distribution transformer in a continuous cycle, extract the voltage at the end of each cycle and the voltage at the end of the previous cycle, compare the voltage values ​​between adjacent cycles in chronological order, identify the time period with continuous trend, and extract the cycle number associated with the time period to obtain the voltage change trend segment sequence.

[0059] S2: Acquire synchronous current and thermal monitoring data, and combine them with voltage change trend segment sequences to extract amplitude-frequency ratio, current impulse offset value and thermal excitation frequency. Then, attach them to the cycle according to time. Analyze the time sequence of parameter changes within the cycle through sequence learning feature extraction, extract feature sequences that reflect continuous change trends, and track the continuous continuation of parameter direction and amplitude based on the change trajectory presented by the feature sequences to obtain a set of response trajectory fluctuation segments.

[0060] S3: Based on the periodic segments listed in the set of response trajectory fluctuation segments, the voltage change trend, current offset amplitude change and thermal excitation point position are sequentially mapped into the preceding and following time regions. The initial label content is extracted by analyzing the parameters of the preceding period. The direction of change between the corresponding parameters is traced in the periodic order. The label shown in the previous segment is used in the same continuous flow periodic segment, and the label value is updated periodically to obtain the label timing adjustment result sequence.

[0061] S4: Based on the label time-series adjustment result sequence, the parameter time-series features obtained in the previous sequence learning process are mapped into the path structure, the activation node information in the edge channel of the path structure is extracted, the node response change time sequence in each channel during the continuous period is compared, the difference in response speed between the two channels is compared, the continuous time segments with differential step phenomenon are identified, and the set of channel response uneven state segments is obtained.

[0062] S5: Based on the set of channel response uneven state segments, the tag output records and voltage fluctuation state data of the corresponding extraction period are compared with the order of fluctuation direction and tag trend. The current tag value is replaced in the period with differentiated trend, and the updated content is paired with the original period number to obtain the tag continuous output judgment sequence.

[0063] The voltage change trend segment sequence includes the period number sequence, the tail voltage comparison result, and the trend continuous time period identifier. The response trajectory fluctuation segment set includes the amplitude-frequency ratio sequence, current impact offset change, and thermal excitation frequency evolution. The tag timing adjustment result sequence includes the tag replacement content, the change direction of the period corresponding parameter, and the segment tag flow structure. The channel response uneven state segment set includes the channel activation node information, response speed difference, and the segment number of the differential step phenomenon. The tag continuous output judgment sequence includes the tag update content, period number pairing information, and the correspondence between voltage fluctuation and tag trend.

[0064] Please see Figure 2 The specific steps of S1 are as follows:

[0065] S101: Obtain the voltage amplitude data of the distribution transformer in a continuous cycle, and compare the voltage amplitude at the end of each cycle with the voltage amplitude at the end of the previous cycle to obtain the corresponding sequence of cycle voltages.

[0066] First, instantaneous voltage value sequences within multiple adjacent power frequency cycles are acquired from the voltage acquisition channel on the low-voltage side of the transformer. Each cycle is defined as 50 power frequency cycles. The acquired continuous data is divided into periods. The voltage amplitude of the last 5 data points within each cycle is selected as the feature values ​​of the current cycle's end segment. During execution, the end positions of each cycle are sequentially marked in the data buffer. Starting from the second cycle, the 5 voltage amplitudes at the end of each cycle are compared with the 5 voltage amplitudes at the end of the previous cycle. The difference between the voltage amplitudes at each corresponding point is calculated and the sign of the difference is recorded (e.g., positive, negative, zero), forming a single-cycle voltage change trend indicator group. If the acquisition frequency is set to 1000 times per second... Each power frequency cycle contains 20 data points, with the last 5 points being points 16 to 20. For example, if the voltage amplitude at the end of the first cycle is 220.3, 221.1, 220.6, 220.8, 221.0, and the voltage amplitude at the end of the second cycle is 220.1, 220.9, 220.3, 220.6, 220.9, then the corresponding differences are -0.2, -0.2, -0.3, -0.2, -0.1, with the difference signs being negative, negative, negative, negative, negative, and negative respectively. The corresponding trend indicator is a five-segment continuous decrease of "---". This group of indicators is then used as the difference pattern feature between the current cycle and the previous cycle, and recorded sequentially to finally obtain the cycle voltage sequence.

[0067] S102: Based on the corresponding periodic voltage sequence, extract the direction of change between adjacent voltage amplitudes, analyze whether the direction of change is consistent, and obtain the set of trend continuation numbers;

[0068] First, the voltage difference sign group between each cycle and the previous cycle is read segment by segment from the sequence consisting of multiple cycle difference sign groups obtained from the previous cycle. During the process, when reading any two adjacent cycles, the five-segment sign group of the previous cycle and the five-segment sign group of the current cycle are extracted position by position. Direction extraction refers to calling up the sign at each position and classifying it according to the direction it points to: negative signs are classified as decreasing, positive signs as increasing, and zero as neutral. Then, the direction of the same position in two adjacent cycles is compared. The comparison action refers to comparing the characters of the direction one-to-one. If the characters are the same, the direction is judged to be consistent; if the characters are different, the direction is judged to have changed. When performing the direction consistency judgment, a judgment benchmark value needs to be set. The judgment benchmark value is that the same direction character at a single corresponding position is consistent, and different characters are inconsistent. This judgment benchmark does not require interval division; it only needs to be completely identical as the discrimination condition. For example, if the five-segment sign group of cycle A is negative, negative, negative, negative, negative, and the five-segment sign group of cycle B is negative, negative, zero, negative, negative, negative If the results of the comparison actions at positions 1, 2, 4, and 5 are consistent, and the results at position 3 are different, then a label indicating whether the comparison actions are consistent or not needs to be recorded for each position. The five recorded labels from the comparison actions within the same period are then combined to form a periodic consistency label group. The same comparison action is then performed on the next period and the current period to generate a new label group. By performing direction extraction, direction comparison, and direction judgment actions on all periods, a continuous label group linked list can be formed across all periods. The continuous consistency segment numbering action is then performed on the linked list. This numbering action involves assigning a serial number to consecutively appearing direction-consistent labels, ensuring that each continuous consistent position chain receives a unique number. The number value can be incremented from 1 according to the order of appearance. For example, if the direction consistency results at a fixed position in the first five periods are consistent, consistent, inconsistent, consistent, consistent, then the first two consecutive consistent records are numbered 1, the third is a breakpoint and not numbered, and the fourth and fifth consecutive consistent records are numbered 2. By merging the numbering results formed at each position, a trend continuation number set is obtained.

[0069] S103: Based on the trend continuation number set, extract continuous number segments according to the number order, and locate the voltage trend of the number segments in the original data to obtain the voltage change trend segment sequence;

[0070] First, the number sequence is read sequentially from the number set. During execution, the number sequence is scanned line by line to locate data segments with the same and consecutive number values. Specifically, each number sequence is traversed, and the current position number value is compared with the next position number value for equality comparison. If they are equal, the current position is added to the current consecutive number segment cache, and the process continues to move to the next position and repeat the comparison until the number values ​​are unequal. If the length of the number segment in the cache is greater than or equal to 3, it is considered a consecutive number segment; otherwise, it is discarded. The number segment length comparison operation involves calling the total number of numbers in the current cached number segment and comparing the total number with a set judgment benchmark value. The judgment benchmark value is set to 3 as the starting segment length value. This value is set to ensure that the trend segment has minimum visible continuity. The same process is continued for subsequent numbers until the current number sequence is traversed. Then, the original periodic data mapping is performed on the periodic index position recorded in the number segment, and each number is called... The starting and ending cycle numbers in the numbering segment are extracted using the cycle index in the original voltage data sequence. The extracted cycle segments are then merged to form the voltage data segment corresponding to the trend continuation numbering segment. During the process, the tail voltage value or representative voltage amplitude in each cycle needs to be called as the voltage sequence in the trend segment to form a continuous voltage change data group. For example, if positions 3 to 6 in the numbering sequence are consecutively numbered 2, and number 2 is a valid trend number, then the original voltage tail data in periods 3 to 6 needs to be extracted. If the tail voltages of periods 3 to 6 are 219.8, 219.6, 219.5, and 219.3 respectively, the extracted trend segment voltage sequence is 219.8, 219.6, 219.5, and 219.3. This sequence serves as the voltage trend data corresponding to trend segment number 2. The same data positioning and mapping extraction operation is then performed on all trend numbering segments in sequence to finally obtain the voltage change trend segment sequence.

[0071] Please see Figure 3 The specific steps of S2 are as follows:

[0072] S201: Based on the voltage change trend segment sequence, monitor the amplitude-frequency ratio, current impulse offset value and thermal excitation frequency corresponding to each cycle, and append the parameters to the cycle data frame according to the time sequence to obtain the cycle parameter appending matrix;

[0073] First, the amplitude-to-frequency ratio is calculated. In each cycle, the instantaneous voltage signal is read, and the ratio of the maximum voltage amplitude to the dominant frequency amplitude is calculated. The maximum voltage amplitude is obtained by selecting the sample point with the largest value from the voltage data set within the cycle. The dominant frequency amplitude is obtained from the amplitude corresponding to a frequency of 50 Hz in the Fast Fourier Decomposition results within the cycle. The two values ​​are then divided to obtain the amplitude-to-frequency ratio parameter. If the maximum voltage value within the cycle is 225.4 Hz and the corresponding dominant frequency amplitude is 224.1 Hz, then the amplitude-to-frequency ratio is 1.006. Second, the current impulse offset value is calculated. This requires selecting the instantaneous current sequence within the cycle and calculating the difference between the maximum current value in the sequence and the mean current value within the cycle as the offset value. If the maximum current value in the current sequence is 25.3 Hz and the mean current value is 20.1 Hz, then the offset value is 5.2 Hz. Third, the thermal excitation frequency is calculated by calling the thermal flux change rate sequence at the sampling frequency within the cycle. The system performs a counting operation for values ​​greater than a threshold, which is set to 0.5. This means that when the increase in the thermal parameter at each sampling point is greater than 0.5 compared to the previous point, it is considered an excitation. If such a change occurs 8 times in a cycle, the thermal excitation frequency for that cycle is 8 times. This threshold is set with reference to the sensitivity characteristics of the transformer thermistor. After extracting the three parameters in each cycle, the time tag of the current cycle in the trend sequence is called, and the combination and splicing operation of the parameter value and the cycle tag is performed. The amplitude-frequency ratio, current impulse offset value, and thermal excitation frequency of each cycle are appended to the end of the corresponding cycle data frame to form an expanded cycle data frame. Before each splicing, it is confirmed that the three data are arranged in the order of amplitude-frequency ratio, current impulse offset value, and thermal excitation frequency, and with unit labels. The same operation is performed on all cycles in turn, and all expanded data frame sets are merged to construct the final cycle parameter appended matrix.

[0074] S202: Based on the periodic parameter additional matrix, compare the changes in the direction and magnitude of the parameters contained in each period with the corresponding parameters in the previous period, track the trend of numerical changes, and obtain the parameter change tracking number set.

[0075] First, starting from the second cycle, read the three parameter values ​​of the current cycle sequentially, and simultaneously read the three parameter values ​​of the previous cycle. For each parameter, perform two operations. The first step is direction determination: call the current value and the previous value and calculate the difference between them. If the difference is greater than zero, the direction is upward; if the difference is less than zero, the direction is downward; if the difference is equal to zero, the direction is flat. The output of the direction determination is three sets of direction signs, which record the direction of change of the three parameters. The second step is amplitude change calculation: directly subtract the previous value from the current value. The difference value is obtained as the amplitude change, and then the amplitude level is divided according to the absolute value of the difference. An interval judgment operation is performed on the amplitude level: the absolute value of the amplitude-frequency ratio change is less than or equal to 0.005 and judged as a small change; between 0.005 and 0.02 and judged as a medium change; and greater than 0.02 and judged as a drastic change. The absolute value of the current impulse offset change is less than or equal to 1 and judged as a small change; between 1 and 3 and judged as a medium change; and greater than 3 and judged as a drastic change. The absolute value of the thermal excitation frequency change is less than or equal to 2 and judged as a small change; between 2 and 6 and judged as a medium change; and greater than 6 and judged as a drastic change. The change at point 6 is considered a drastic change. The judgment is based on the normal fluctuation range set in the transformer equipment. The execution result forms three sets of change directions and three sets of change amplitude levels. The direction and amplitude of each parameter are then combined to form a change label. The change label is represented by a combination of characters, such as "up-middle", "hold-slight", "down-drastic", etc. The same operation is performed on each cycle in sequence, and the results of each cycle are recorded in order. By traversing all cycles, a complete parameter change trend record sequence is formed. In this sequence, direction sequences and amplitude level sequences are constructed according to parameter types. Then, the continuity of the labels between adjacent cycles in the sequence is used for tracking and numbering. The tracking and numbering rule is that if the change label of a certain parameter is completely consistent in two consecutive cycles, it is considered as the same number segment, and a unique number is assigned to this segment. The numbering starts from 1 and increments. The numbering terminates when the label changes. For example, if the amplitude-frequency ratio change label in cycles 2 to 4 is "up-middle", the number value is 1. If the change in cycle 5 is "up-drastic", the number value is 2. Finally, the parameter change tracking number set is obtained.

[0076] S203: Based on the parameter change tracking number set, locate the parameter value fluctuation range of the corresponding period in the number segment, analyze the continuous fluctuation sequence segment, and obtain the response trajectory fluctuation segment set;

[0077] First, the start and end cycle indices of each numbered segment are read sequentially. Based on this range, the amplitude-frequency ratio, current impact offset, and thermal excitation frequency are extracted from the period parameter supplementary matrix within the corresponding cycle. For each extracted parameter sequence, a numerical fluctuation range positioning operation is performed. This operation obtains the difference by directly comparing the maximum and minimum values ​​in the sequence; the difference represents the fluctuation range of the corresponding parameter for that numbered segment. Next, the span between the minimum and maximum values ​​is recorded to form the upper and lower boundary markers of the fluctuation segment. Then, a trend judgment operation is performed within the sequence. This operation involves sequentially judging the direction of adjacent values ​​in the parameter sequence. The direction judgment method is to call the current position value and the next position value, performing a subtraction operation. If the difference is greater than zero, it is recorded as an increase; if it is less than zero, it is recorded as a decrease; if it is equal to zero, it is recorded as flat. This constructs a trend direction sequence composed of direction characters. Finally, based on the length of the continuous trend in the direction sequence... Continuous fluctuation segments are extracted. If the length of a continuous upward or downward trend exceeds three cycles, it is considered a stable fluctuation segment, and its start and end cycle indices are recorded. The direction type, fluctuation amplitude range, and cycle span within the segment are also recorded. For example, if segment number 5 corresponds to cycle 10 to cycle 16, and the current impact offset value sequence is 18.1, 18.6, 19.2, 19.7, 20.4, 20.9, 21.3, then its maximum value is 21.3, its minimum value is 18.1, and its fluctuation range is 3.2. If the direction of this sequence increases continuously six times, then this segment is recorded as a response trajectory fluctuation segment with an upward direction. The same operation is performed on the remaining two parameters in the numbered segments to generate their respective response trajectory segment records. Finally, the fluctuation direction, fluctuation range, trend sequence, start and end cycle indices, and cycle span corresponding to the three parameters in each numbered segment are integrated to obtain a set of response trajectory fluctuation segments.

[0078] Please see Figure 4 The specific steps of S3 are as follows:

[0079] S301: Based on the periodic segments listed in the set of response trajectory fluctuation segments, collect the voltage change trend, current offset amplitude change and thermal excitation point location corresponding to each period, and map the parameters to the time region according to the order before and after the period to obtain the periodic parameter sequence.

[0080] First, the starting and ending cycle indices given in the current record are read, and a cycle scan pointer is established. Starting from the starting cycle, the three data items in the cycle parameter appended matrix are called sequentially in ascending order: voltage change trend label, current impact offset value change, and thermal excitation point position. The voltage change trend is called by directly reading the symbol combination in the trend sequence for that cycle, such as "rise," "fall," or "flat." The current offset amplitude change is obtained by reading the difference between the current impact offset value of the current cycle and the previous cycle. If the offset value of cycle a is 5.2 and the offset value of the next cycle is 6.1, then the change is 0.9. The thermal excitation point position is obtained by performing a point-by-point scan of the sampling sequence of the thermistor parameters within that cycle to obtain the index of the position where the change is greater than a set threshold. The threshold is set to 0.5, which is based on the thermistor characteristics of the transformer. If the increase is greater than 0.5 at points 7, 13, and 18 in the sampling sequence, then the thermal excitation point position for that cycle is recorded as points 7, 13, and 18. This acquisition method is repeated in each cycle to ensure that each cycle... Each period generates three corresponding data items. Then, a periodic sequential mapping operation is performed on the data. Periodic sequential mapping refers to using the time sequence number of the period as an index and concatenating the aforementioned three data items into the corresponding index item according to a fixed field order. The mapped period item is then added to a sequential queue, and the same operation is repeated for the next period until the end of the period. After collection, a sequential record chain from the start to the end of the period is obtained. Each item in the chain is presented in the format of "period number, voltage trend label, current offset change, and thermal excitation point location sequence". For example, if the current response trajectory fluctuation range is from period 10 to period 14, after collecting each item from period 10 to 14, we can form the following: period 10 is "10, decrease, -0.3, 5, 12", period 11 is "11, decrease, -0.4, 4, 9, 15", period 12 is "12, flat, 0.0, 11", period 13 is "13, increase, 0.7, 6, 14", and period 14 is "14, increase, 1.1, 8, 17". Then, we can directly arrange this chain according to the period number order to obtain the period parameter sequence.

[0081] S302: Based on the sequential sequence of periodic parameters, trace the direction of parameter change in the period, extract the label content attached to the previous segment, and obtain a set of label for directional continuous segments;

[0082] First, initialize the cycle scan pointer. Then, call the voltage trend label column in the cycle parameter sequence. Perform a character equality check on the trend signs between two adjacent cycles. If the voltage trend of the current cycle is consistent with the previous cycle, it is recorded as a direction continuity marker. If they are different, it is considered a trend breakpoint and a new continuous segment start point is generated. Continue scanning until the direction extraction of the entire voltage trend label column is completed. Next, perform a positive / negative sign check on the change value of each cycle in the current offset change column. If two adjacent cycles are both positive or both are negative, it is recorded as a direction continuity state. If the signs are different, it is recorded as a breakpoint. The directional judgment method does not involve actual amplitude; it only uses directional consistency as the classification criterion. The directionality of the thermal excitation point location is not applicable to sign judgment; it must be judged based on the trend of quantity change. The difference between adjacent periods is calculated for the number of thermal excitation points recorded in each period. If the quantity difference is positive or negative within multiple consecutive periods, it is considered a continuous trend. A turning point from increase to decrease or from decrease to increase is considered a breakpoint. The period index of all breakpoint locations is recorded, and the index of the trend segment preceding the breakpoint is recorded as a valid segment. Subsequently, a label extraction operation is performed. Each directional continuous segment needs to retrieve the label attached to the previous segment within that segment. The tag content is provided by the set of response trajectory fluctuation segments constructed in the previous steps. The extraction method involves searching backwards from the starting period of the current continuous segment to find the number of the previous fluctuation segment to which it belongs. The search rule is that the period number is less than the first period of the current segment and is the most recently matched entry. After obtaining the number, the identifier field of that number record is retrieved from the fluctuation segment set. The identifier field contains a string label, the content of which is set according to the actual application scenario, such as "slight fluctuation," "rapid rise," or "continuous fall." After extraction, this label is appended to the current continuous segment as a classification marker, and the process continues. The same operation is performed on the parameters, and the period index range, trend type, and associated labels with the same direction are recorded. For example, if the voltage trend from period 10 to period 14 is "rising", the current offset changes are 0.8, 1.0, 0.7, 1.1, and 0.9 respectively, all with positive signs, and the number of thermal excitation points are 2, 3, 4, 5, and 6 respectively, with the number increasing continuously, and all three conditions are met, then this segment can be identified as a directional continuous segment. If the label attached to the previous segment is "medium amplitude fluctuation", then this label is used as the belonging label content of the current segment, and finally the directional continuous segment label set is obtained.

[0083] S303: Based on the directional continuous segment label set, replace the original label content of the period in numerical order, update the used labels in the period and write them into the current data frame to obtain the label timing adjustment result sequence.

[0084] First, the label set of continuous directional segments is sorted in ascending order by the number field. Then, a label replacement index table is created, containing three fields: directional segment number, cycle number sequence, and new label content. Each record in this table is read sequentially. First, the cycle number sequence field is called to parse out the start and end cycle numbers covered by the directional segment. All cycle records within the range of these numbers are filtered in the data frame. For each filtered cycle record, its original label field is called, and a string replacement operation is performed, replacing the original label with the new label content provided in the current record. If the original label for a certain cycle is "initial stable," while the current directional continuous segment label content is "moderate upward," then the label field for that cycle is updated to "moderate upward." After the replacement, a label write operation is performed, that is, the updated label value is written to the label field in the current cycle data frame. The writing order remains consistent with the original time order of the cycles. The above process is repeated to perform update operations on all directional segments. If a directional segment is numbered 3, the corresponding cycle number is 20 to 24, and the label content is "continuous rise", then the label field of cycles 20 to 24 in the data frame is uniformly written as "continuous rise". If the label field of the original cycle data frame is empty, the label value is directly inserted without replacement. If the cycle already has historical label content, it is replaced and overwritten. The entire process requires the establishment of an update record cache table. After each update, the cycle number, the new label content, and the content before replacement are stored in this table to preserve the update path. After the replacement writing is completed in all continuous segments in all directions, a copy of the cycle data frame after overwriting the original label is formed. Then, all cycle records are recombined according to the cycle order to form a new data structure arranged in time order, and finally the label time sequence adjustment result sequence is obtained.

[0085] Please see Figure 5 The specific steps of S4 are as follows:

[0086] S401: Based on the label timing adjustment result sequence, collect the activation node information of adjacent cycles in the edge channel of the path structure, and map the response actions of the corresponding nodes of adjacent channels in each cycle to the cycle position in time order to obtain the channel node response sequence;

[0087] First, extract the logical table used for edge channel identification from the path structure. This logical table records the channel number, channel type, period node index, and activation status field. Edge channels are defined as channel types located on both sides of the main path and having unidirectional node triggering attributes. Node index sets are established for records identified as left and right channels respectively. Then, the position number corresponding to each period in the tag timing adjustment result sequence is read in ascending order of period number. Using this number as an index, the current period is checked in the left and right channel node sets to see if an activation marker exists. The activation marker is determined by the activation status field being equal to 1. If the activation status of the corresponding record in the left channel node set for a certain period is 1, then the left channel node for that period is considered activated. The node number, activation status, and channel direction information are combined into a node response action record. The response action record content fields include period number, channel direction, and node number. With the action type fixed as "Active", the same judgment operation is performed on the right channel. If the activation field value of the record for this period in the right channel set is also 1, then another response action record is recorded. For each period, all corresponding active node records need to be added to the temporary response action sequence. After processing all periods, the temporary response action sequence will contain the complete response action events of the left and right channels in each period. Then, all response records need to be sorted in ascending order by the period number field to ensure that the order of response actions strictly corresponds to the period position. Time mapping is performed on the sorted response action records. The time mapping operation is to replace the period number in each record with its sequential position value on the time axis. This position value can be converted by the preset period duration and the initial time offset, or it can be directly substituted using the equally spaced numbering method to obtain the channel node response sequence.

[0088] S402: Based on the channel node response sequence, compare the response timing of the left and right channels in a continuous period, analyze the relationship between the response start time and the sequence of the continuous action, and obtain the channel response time difference number sequence.

[0089] First, the response sequences are sorted in ascending order by period number to construct a period index table. Each entry contains fields such as period number, response time of each phase, and activation node number of each phase. Each period number is traversed, and the response action records of each phase channel within the period are extracted sequentially. The corresponding phase response time value is filtered and filled in. If a phase in a period does not respond, the entry is empty. After constructing the index table for all periods, response time time series analysis is performed. The response times of two consecutive periods under the same phase are compared. If the time of the previous period is less than that of the next period, it is considered a continuation of the action; if they are equal, it is considered a synchronous continuation; if the time of the next period is less than that of the previous period, it is marked as a regression of the action. This rule is used to determine the continuity of the response sequence within a phase. Subsequently, within each period, two phase channels (e.g., two edge phases) are selected to calculate the response time difference. The difference is the time of one channel minus the time of the other channel. A positive value indicates that the former response is lagging, a negative value indicates that the former response is advancing, and a zero value indicates simultaneous response. The response time difference is used to divide the response intervals: an absolute value less than 1 indicates a synchronous response, between 1 and 3 indicates a slight lag, and more than 3 indicates a significant lag. Each period is assigned a corresponding number based on its interval, with numbers increasing sequentially from 1. For example, if the response times of two channels in a certain period are 15 and 14 respectively, the difference is 1, which is classified as a slight early response and numbered 2. After processing all periods, a sequence of channel response time difference numbers is obtained.

[0090] S403: Based on the channel response time difference number sequence, track the continuous relationship between the numbers and detect the duration of the inconsistent response speed state to obtain the set of channel response uneven state segments;

[0091] First, a numbered sequence index table is established. This table uses the period number as the index and records the response time difference classification number value corresponding to the current period. The number value is uniquely identified according to the difference type. For example, synchronous response is marked as 1, slight response lag is marked as 2, and significant response lag is marked as 3. After reading this numbered sequence, a period-by-period traversal operation is performed starting from the first period. A continuous numbering buffer is initialized and the starting number value is recorded. It is determined whether the current period number value is consistent with the previous period number value. If they are consistent, the current period is added to the buffer and the traversal continues downward. If they are inconsistent, the previous continuous segment is marked as ended, and it is determined whether the segment number value is of the asynchronous response type. Asynchronous response refers to the state type where the number is not 1. If the segment number value is 2 or 3, it is marked as a response uneven segment. Then, the starting period number, ending period number, and corresponding number value of the segment are combined into a response uneven segment structure item. A new numbering buffer is initialized and the above process is repeated with the current period number as the new starting number. During the processing, the breakpoint index of the number change point needs to be recorded for subsequent detection of the stability boundary. The processing continues until all period numbers are traversed. After processing, a set of multiple asynchronous response segments is obtained. Each item in this set represents a state where the channel response has inconsistent speed within a continuous period. Then, a minimum length judgment operation is performed on each segment. The minimum response unevenness duration period threshold is set to 3. It is judged whether the number of periods in the segment is greater than or equal to 3. If it is less than this value, it is not recorded. If it meets the condition, it is retained as a formal unevenness segment. For example, if the number values ​​of periods 10 to 13 are all 3, it is a significant response lag, and the number of periods is 4, which meets the condition. Then, the segment item "starting period 10, ending period 13, number 3" is generated. The processing continues to obtain multiple segment records that meet the condition. Finally, all records are arranged in the order of the starting period number and output in a unified format to obtain the set of channel response unevenness state segments.

[0092] Please see Figure 6 The specific steps of S5 are as follows:

[0093] S501: Based on the set of channel response uneven state segments, call the tag output and voltage fluctuation state data in the corresponding period, and compare the tag trend direction and voltage fluctuation direction in the period in time order to obtain the tag fluctuation order comparison sequence.

[0094] First, the set of uneven state segments is read to form a segment index list. Each item in this list records the start cycle number, end cycle number, and number attribute. The segment index list is traversed sequentially, and a cycle scan sequence is established for the start cycle to the end cycle of the current segment. Then, the label field content and voltage fluctuation field content in the cycle parameter sequence are read sequentially. The label field content is a trend character adjusted for rhythm, such as "continuous rise," "moderate decline," "slight fluctuation," etc. The voltage fluctuation field content is the trend direction character and numerical range extracted from the cycle. When scanning the current cycle, its label content must be called first and recorded as the label trend before... The voltage fluctuation direction field is called and recorded as a voltage trend. Then, a direction consistency comparison operation is performed between the label trend and the voltage trend. Specifically, the comparison involves a direct character comparison of the two trend characters. If the label trend direction character and the voltage trend direction character are exactly the same, it is recorded as a consistency flag; otherwise, it is recorded as an inconsistency flag. The same operation is then performed on the next cycle in the cycle scan sequence. This method forms a direction comparison sequence arranged in cycle order within the segment. Subsequently, a continuous tracking operation is performed on the comparison sequence, calling the consistency flag or inconsistency flag corresponding to each cycle in the comparison sequence and performing adjacent cycle flag operations. The equality judgment of segments is as follows: if several consecutive periods have the same sign, a continuous consistent direction chain is formed; if several consecutive periods have the same sign, a continuous deviation direction chain is formed. The starting period number, ending period number, and sign attribute of each chain are recorded. During execution, a mapping structure needs to be established for each record. This structure includes three items: period number, label trend, fluctuation trend, and comparison result. A practical example illustrates this: if a certain uneven segment covers periods 20 to 24, the label trends for these five periods are "rise," "rise," "fall," "fall," and "fall," respectively, and the voltage trends are "rise," "rise," "flat," "fall," and "fall," respectively. Period 20 and... 21. The label trend and voltage trend are consistent, and are recorded as a consistency marker. In period 22, the label is "decreasing" while the voltage trend is "flat", and are recorded as an inconsistency marker. In periods 23 and 24, both trends are decreasing and the label trend is also decreasing, and are recorded as a consistency marker. Therefore, this segment forms a structure of "20-21 consistent chain", "22 inconsistency point", and "23-24 consistent chain" in the direction comparison sequence. The above steps are repeated for all uneven segments in turn. The control sequence items of all segments are integrated into a result list arranged in chronological order. Each period in this list is accompanied by the label trend, fluctuation trend and the comparison result of the two, resulting in the label fluctuation order comparison sequence.

[0095] S502: Based on the tag fluctuation sequence comparison, compare the relationship between the tag trend change and the voltage fluctuation direction in the cycle, correct the tag trend direction in the cycle where the trend direction is inconsistent, and obtain the tag direction correction data table.

[0096] First, a label-matching sequence index table is constructed, indexed by the period number. Each record contains fields for label trend direction, voltage fluctuation direction, and direction comparison result. After sorting by period number in ascending order, a period-by-period judgment operation is performed. The direction comparison result field in the current period record is read. If the field is marked as inconsistent, a correction process is triggered. In the correction process, the character content of the label trend direction and voltage fluctuation direction corresponding to the current period is first obtained, and it is judged whether their specific directions are opposite or have no obvious directional difference based on the character content. The direction judgment rule is: if the label trend direction is "rising" and the voltage fluctuation direction is "falling", or the label trend is "falling" and the voltage trend is "rising", then it is confirmed that the two are completely opposite. At this time, the label trend direction field needs to be replaced with the direction character shown by the voltage fluctuation direction. At the same time, a new label field is added to the current period record, recording the correction source as "direction opposite". If the directions are inconsistent but not opposite, for example, the label trend is "stable" and the voltage trend is "rising" or "falling", then a weight judgment operation is performed, and the judgment rule is as follows. The process involves: using the voltage fluctuation direction of three consecutive cycles to determine the trend continuation. If the voltage direction of the two preceding and following cycles is consistent with the current cycle, the voltage trend is determined to be the dominant direction, and the label trend needs to be corrected accordingly. If the voltage trend is discontinuous or fluctuates frequently within three cycles, the original label trend is retained and marked as "trend unstable". In the actual example, if the label direction of cycle number 36 is "decreasing" while the voltage direction is "increasing", and the voltage direction of the preceding and following cycles is also "increasing", then the label is considered to need to be corrected to "increasing". This result is written to the record item of cycle number 36, and the correction reason is marked as "continuous increase dominance". The corrected trend value is uniformly stored in the label direction correction field. After completing the traversal of all cycles, the original label direction field, the corrected label direction field, the original comparison status field, and the correction basis field are uniformly organized into a data table with a complete structure. Each row of records corresponds to a cycle node, and each column clearly identifies the original label, voltage direction, correction flag, and final trend direction, ultimately resulting in the label direction correction data table.

[0097] S503: Based on the tag direction correction data table, the corrected tag direction is embedded into the original sequence according to the period number to obtain the tag continuous output judgment sequence;

[0098] First, the label direction correction data table is read, extracting four items: cycle number, original label direction, corrected label direction, and correction flag. These are then sorted in ascending order by cycle number to create a correction mapping table. The key of this table is the cycle number, and the value is the corrected label direction. Next, the original label parameter sequence is called. This sequence is a time series structure composed of label fields generated in previous cycles. Each cycle item is read sequentially, and the current cycle number is checked to see if it exists in the correction mapping table. If it exists, it means that the label for that cycle needs to be replaced. The corresponding corrected label direction in the correction mapping table is used to replace the original label direction field, and a "replacement flag" field is added to the current record to record the replacement behavior. If the cycle number does not appear in the correction table, it means that the label direction for that cycle has not been corrected, the original trend direction is retained, and processing continues. The next cycle record is processed, and a new trend output sequence structure is constructed during the process. Each record in this structure contains four items: cycle number, final label direction, whether to replace, and correction source label. The final label direction comes from the correction label field or the original label field. The whether to replace is determined by whether it exists in the correction table. The correction source label comes from the correction flag field in the previous correction data table. In a practical example, if the cycle number is 52, its original label direction is "decreasing", the correction direction is "increasing", and the correction flag is "opposite correction", then the record item "52, increasing, yes, opposite correction" is generated in the output sequence. If the cycle number is 53, it does not appear in the correction table, so the original direction "stable" is retained, and the record item is "53, stable, no, original label". Finally, the label continuous output judgment sequence is obtained.

[0099] Please see Figure 7 A deep learning-based system for classifying and determining potential hazards in power grid equipment includes:

[0100] The voltage trend extraction module acquires the voltage amplitude data of the distribution transformer within a continuous cycle, extracts the voltage at the end of the cycle and compares the change trend during adjacent cycles, identifies the duration of the trend, and obtains the voltage change trend segment sequence.

[0101] The parameter trajectory generation module acquires synchronous current and thermal monitoring data, and combines them with voltage change trend segment sequences to extract the amplitude-frequency ratio, current impulse offset value and thermal excitation frequency and attach them to the period. Through sequence learning, it analyzes the time-series changes of parameters to obtain a set of response trajectory fluctuation segments.

[0102] The tag timing adjustment module is based on the set of response trajectory fluctuation segments, maps voltage, current and thermal excitation direction, extracts initial tags, inherits the previous segment tags within the trend-consistent cycle, and obtains the tag timing adjustment result sequence.

[0103] The channel response comparison module maps the parameter time series features to the path structure based on the tag time series adjustment result sequence, extracts the edge channel activation nodes, compares the periodic channel response changes, analyzes the response speed differences, and obtains the set of channel response uneven state segments.

[0104] The tag output revision module extracts periodic tags and voltage fluctuations based on the set of channel response uneven state segments, compares the trend direction with the tag trend, and obtains the tag continuous output judgment sequence.

[0105] 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 power grid equipment hidden danger grading determination method based on deep learning, characterized in that, The method comprises the following steps: S1: obtaining voltage amplitude data in a continuous period of a power distribution transformer, extracting the voltage at the end of the period and comparing the change trend between adjacent periods, identifying the duration of the trend, and obtaining a sequence of voltage change trend sections; S2: obtaining synchronized current and thermal monitoring data, and combining the voltage change trend section sequence to extract the amplitude-frequency ratio, current shock offset value and thermal excitation frequency and attach them to the period, analyze the parameter time sequence change through sequence learning method, and obtain a response trajectory fluctuation section set; S3: Based on the response trajectory fluctuation section set, map the voltage, current and thermal excitation flow direction, extract the initial label, and inherit the previous section label in the consistent trend period to obtain a label time sequence adjustment result sequence; S4: Based on the label time sequence adjustment result sequence, map the parameter time sequence features to the path structure, extract the edge channel activation node, compare the period channel response change, analyze the response speed difference, and obtain a channel response uneven state section set; S5: Based on the channel response uneven state section set, extract the period label and voltage fluctuation, compare the trend direction and label trend, and obtain a label continuous output judgment sequence. 2.The deep learning-based power grid equipment hidden danger grading determination method according to claim 1, characterized in that, The voltage change trend section sequence includes a period number sequence, a tail voltage comparison result, and a trend continuous time period identifier. The response trajectory fluctuation section set includes an amplitude-frequency ratio sequence, a current shock offset change, and a thermal excitation frequency evolution. The label time sequence adjustment result sequence includes label replacement content, period corresponding parameter change direction, and section label flow structure. The channel response uneven state section set includes channel activation node information, response speed difference, and differential step phenomenon section number. The label continuous output judgment sequence includes label update content, period number pairing information, and voltage fluctuation and label trend correspondence. 3.The deep learning-based power grid equipment hidden danger grading determination method according to claim 1, characterized in that, The voltage at the end of each power frequency period is referred to as the voltage amplitude; The current shock offset value refers to a parameter that measures the degree of deviation of the current from the average value within a period.

4. The deep learning-based power grid equipment hazard grading determination method according to claim 1, characterized in that, The parameter time sequence feature refers to a dynamic information sequence that reflects the continuous change trend of voltage, current and thermal parameters over time. The edge channel activation node refers to a node in the channel at the edge of the path in the neural network structure that is triggered within a period.

5. The deep learning-based power grid equipment hazard grading determination method according to claim 1, characterized in that, The specific steps of S1 are: S101: Obtain the voltage amplitude data in the continuous period of the power distribution transformer, and correspondingly compare the voltage amplitude at the end of each period with the voltage amplitude at the end of the previous period to obtain a period voltage correspondence sequence; S102: Based on the period voltage correspondence sequence, extract the change direction between adjacent voltage amplitudes, analyze whether the change direction is consistent, and obtain a trend continuation number set; S103: Based on the trend continuation number set, extract the continuous number section according to the number order, and locate the voltage trend of the number section in the original data to obtain a voltage change trend section sequence.

6. The deep learning-based power grid equipment hazard grading determination method according to claim 1, characterized in that, The specific steps of S2 are: S201: Based on the voltage change trend section sequence, monitor the amplitude-frequency ratio, current shock offset value and thermal excitation frequency corresponding to each period, and attach the parameters to the period data frame according to the time sequence to obtain a period parameter attachment matrix; S202: Based on the periodic parameter additional matrix, the direction and amplitude of the parameter contained in each period and the corresponding parameter of the previous period are compared, the numerical value change trend is tracked, and a parameter change tracking number set is obtained; S203: Based on the parameter change tracking number set, the parameter value fluctuation range of the corresponding period in the number section is located, the continuous fluctuation sequence section is analyzed, and a response trajectory fluctuation section set is obtained.

7. The deep learning-based power grid equipment hazard grading determination method according to claim 1, characterized in that, The specific steps of S3 are: S301: Based on the period section listed in the response trajectory fluctuation section set, the voltage change trend, current offset amplitude change and thermal excitation point position corresponding to each period are collected, the parameters are mapped to the time region according to the period before and after the sequence, and a period parameter sequence is obtained; S302: Based on the period parameter sequence, the change direction of the parameter in the period is traced back, the label content of the previous section is extracted, and a direction continuous section label set is obtained; S303: Based on the direction continuous section label set, the original label content of the period is replaced according to the number sequence, the used label is updated according to the period, and then written into the current data frame, and a label time sequence adjustment result sequence is obtained.

8. The deep learning-based power grid equipment hazard grading determination method according to claim 1, characterized in that, The specific steps of S4 are: S401: Based on the label time sequence adjustment result sequence, the activation node information in the adjacent period of the edge channel in the path structure is collected, and the response action of the corresponding node of each period in the adjacent channel is mapped to the period position in time sequence, and a channel node response sequence is obtained; S402: Based on the channel node response sequence, the response time sequence of the left and right channels in the continuous period is compared, the response start time and the relationship between the continuation actions are analyzed, and a channel response time difference number sequence is obtained; S403: Based on the channel response time difference number sequence, the continuous relationship between the numbers is tracked, and the duration period range of the inconsistent state of the response speed is detected, and a channel response uneven state section set is obtained.

9. The deep learning-based power grid equipment hazard grading determination method according to claim 1, characterized in that, The specific steps of S5 are: S501: Based on the channel response uneven state section set, the label output and voltage fluctuation state data in the corresponding period are called, and the label trend direction and voltage fluctuation direction in the period are compared in time sequence, and a label fluctuation sequence comparison sequence is obtained; S502: Based on the label fluctuation sequence comparison sequence, the relationship between the label trend change and the voltage fluctuation direction in the period is compared, the label trend direction is corrected in the period where the trend direction is inconsistent, and a label direction correction data table is obtained; S503: Based on the label direction correction data table, the corrected label direction is embedded into the original sequence according to the period number, and a label continuous output judgment sequence is obtained. 10.A power grid equipment hidden danger grading determination system based on deep learning, characterized in that, The system is used to realize the power grid equipment hidden danger grading judgment method based on deep learning in any one of claims 1-9, and the system comprises: The voltage trend extraction module obtains the voltage amplitude data in the continuous period of the distribution transformer, extracts the voltage at the tail of the period, compares the change trend between adjacent periods, identifies the trend duration period, and obtains a voltage change trend section sequence; The parameter trajectory generation module obtains synchronized current and thermal monitoring data, and extracts amplitude-frequency ratio, current shock offset value and thermal excitation frequency in combination with the voltage change trend section sequence and attaches them to the cycle. The parameter time sequence change is analyzed through sequence learning to obtain a response trajectory fluctuation section set, and a response trajectory fluctuation section set is obtained. The label rhythm adjustment module maps the voltage, current and thermal excitation flow based on the response trajectory fluctuation section set, extracts the initial label, inherits the previous section label in the consistent trend cycle, and obtains a label time sequence adjustment result sequence. The channel response comparison module maps the parameter time sequence characteristics to the path structure based on the label time sequence adjustment result sequence, extracts the edge channel activation node, compares the cycle channel response change, analyzes the response speed difference, and obtains a channel response uneven state section set. The label output revision module extracts the cycle label and voltage fluctuation based on the channel response uneven state section set, compares the trend direction and label trend, and obtains a label continuous output judgment sequence.

Images