Verification method of intelligent comprehensive protection device

Abstract

The invention relates to the technical field of protection device testing, in particular to a verification method of an intelligent comprehensive protection device, which comprises the following steps of: extracting starting time and response time in a tripping state transition sequence of the protection device, calculating execution delay and action segment length, and comparing and identifying an abnormal link based on a state number sequence; the method comprises the following steps: realizing analysis and standard reconstruction of a state logic structure, forming an ordered tripping path, introducing a long short-term memory neural network to carry out time alignment on a state change sequence and a current signal segment, identifying a potential mismatch or abnormal change trend in a state chain through linkage analysis of a state-current abrupt change position, and reconstructing a state path. According to the method, path continuity and dynamic adaptability are enhanced, a random forest is adopted to take device numbers, the number of jump segments and response residual errors as multiple inputs for training modeling, device scores are output through an integrated tree voting mechanism, and recognition sorting of abnormal frequency significant devices is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of protection device testing, and in particular to a calibration method for an intelligent integrated protection device. Background Art

[0002] The focus of the protection device testing technology field is to conduct systematic and standardized functional verification and performance evaluation of various protection devices in the power system to ensure that the devices can operate promptly and accurately according to predetermined logic under abnormal operating conditions such as faults, voltage fluctuations, and short circuits, thereby effectively eliminating faults, protecting equipment from safe operation, and maintaining grid stability.

[0003] A verification method for intelligent integrated protection devices aims to comprehensively test the response behavior of protection devices under simulated faults, normal operation, and boundary conditions. The purpose is to determine whether the action of the protection device is consistent with the set parameters, verify the correctness and timeliness of the response, and identify potential misoperation or refusal to operate hazards, thereby improving the reliability and fault tolerance of the protection device in the actual power grid environment, enhancing the security defense capabilities of the system, and ensuring the timeliness and accuracy of fault handling.

[0004] Existing technologies lack in-depth modeling of state logic chains and sequence response flows during protection device verification. Fault response behavior is judged based on start and end states or delay parameters, making it difficult to cover structural anomalies such as jumps, divergences, and sequence disruptions that may exist during state transitions. Under the multi-state drive and multi-conditional blocking logic of protection devices, traditional verification methods struggle to distinguish between behavioral patterns that appear normal on the surface and abnormal internal paths. The sudden change characteristics of current responses are not fully quantified and analyzed, and there is a lack of a linkage identification mechanism for coupled changes between states and electrical quantities, making it easy to miss short-term, non-continuous false trips. Response results are often judged based on single events, lacking systematic aggregation and statistical ranking of similar devices and similar anomalies, making it difficult to identify potential high-risk devices through early warning. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a calibration method for an intelligent integrated protection device.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a calibration method for an intelligent integrated protection device, comprising the following steps: S1: Based on the event change sequence and switch status records recorded during the operation of the protection device, the start time and trip response time are extracted, the execution delay and action duration are calculated, and the state numbers are compared for sequential consistency to obtain the trip logic transfer path set; S2: Based on the trip logic transfer path set, filter the transfer nodes with the same trigger condition value but different state successors in the blocking logic module, extract the jump time and path structure differences, and generate a state branch point bitmap set; S3: Based on the state divergence point atlas, a long short-term memory neural network is used to extract the state change sequence and current data fragments in the alarm trigger logic, perform state sequence comparison and mutation detection, insert the reconstructed fragments to form a sequence path flow, and obtain the abnormal state reconstructed path group; S4: Reconstructing the path group based on the abnormal state, identifying the jump segment position in the restoration response mechanism, calculating the response residual and determining the over-limit segment, using random forest to locate the state starting point corresponding to the jump segment and classify and summarize it to obtain the jump behavior risk index cluster; S5: Based on the jump behavior risk index cluster, extract the device number and time information corresponding to the jump segment, count the response residual and the number of jumps, perform scoring and sorting, and establish a high-frequency jump device risk sequence table.

[0007] As a further solution of the present invention, the specific steps of generating the trip logic transfer path set are: Based on the event change sequence and switch status records recorded during the operation of the protection device, the time of the first and last nodes in each sequence is read. By performing a difference calculation between the start time and the completion time and uniformly numbering the node labels, the time units of each record are checked and converted to a unified reference time format. After eliminating time abnormality records and completing the cleaning, the duration of the tripping segment is calculated to generate a trip response time parameter group; Based on the trip response time parameter group, the state numbers of each section are sorted, and the index difference of the continuous number sequence is determined to be the minimum single-step value by performing a comparison item by item, and a reverse matching check is performed. The numbers with gaps are marked and offset compensated. The number streams that pass the consistency check are structured and organized to generate a state number sequence structure set; Based on the state number sequence structure set, the number sequence is bound to the time parameter, each group of numbers is sequentially matched to the response time period and rearranged into a time stream, the action triggering phase and the recovery phase in the time stream are divided, the sequence is spliced ​​and combined in segments, and the action chain is reorganized according to the triggering order to obtain a set of tripping logic transfer paths.

[0008] As a further solution of the present invention, the specific steps of generating the state bifurcation point bitmap set are: Based on the trip logic transfer path set, extract the input trigger conditions of all state segments in the blocking logic module, group them by input condition values ​​and filter out state pairs with exactly the same input values, obtain all subsequent state information in the state pairs and perform type annotation, identify state pairs with structural differences and classify them into the same group set, and generate trigger condition consistency and divergence groups; Based on the trigger condition consistency and disagreement groups, the difference between the start time of each state segment and the response time span is calculated, the trigger time is matched with the sequential number of the jump node, the path length between each pair of nodes is calculated and the jump direction and type are summarized, the path difference attributes are recorded in groups and the missing segments are supplemented to generate a path jump structure feature set; Based on the path jump structure feature set, a mapping sequence is constructed between the encoding of all jump nodes and the response interval time, a unified numbering format is set and a unified numbering conversion is performed, the path type, response period and numbering position are integrated into a structure matrix, the offset value analysis is completed for each jump segment in the matrix, and a state divergence point bitmap set is obtained.

[0009] As a further solution of the present invention, the specific steps of generating the abnormal state reconstruction path group are: Based on the state divergence point atlas, a long short-term memory neural network is used to perform a time alignment operation on the state change sequence and the current sequence. The interception range is determined by extracting the state start and end indexes and the corresponding current value interval is extracted. The amplitude change of adjacent data points is calculated and the mutation position index is marked. The state synchronization mutation position is extracted and an associated index table is constructed to generate a state current mutation corresponding group. Based on the state current mutation corresponding group, a one-to-one matching operation is performed between the state and the current mutation point position, the index difference between the two types of mutation points is detected and a threshold is set to filter out the unaligned segments, the unmatched state segments are re-labeled and numbered and stored separately from the original sequence, and index offset processing is performed on all inserted fragments to generate a recombined state inserted fragment set; Based on the recombinant state insertion fragment set, the original state sequence and the insertion fragment are spliced ​​together. By extracting the front and back numbers of each splicing point, sequence deviation judgment is performed and discontinuous sections are processed. The merged state streams are sorted according to the triggering order and uniformly indexed and named to obtain the abnormal state reconstruction path group.

[0010] As a further embodiment of the present invention, the long short-term memory neural network is according to the formula:

[0011] in: Indicates time The hidden state vector of Represents the activation function, which is used for nonlinear transformation. represents the weight adjustment coefficient of the input path, represents the weight matrix input to the hidden layer, Indicates time The composite input feature vector of represents the weight adjustment coefficient of the recursive path, represents the recursive weight matrix from the previous hidden state to the current state, represents the hidden state vector at the previous moment, represents the weight factor of the mutation density channel, Indicates time The state mutation density within the time window, represents the weight factor of the stability channel, Indicates time The stability index of the current signal, represents the bias term of the input path, Represents the bias term of the recursive path.

[0012] As a further solution of the present invention, the long short-term memory neural network first performs time alignment processing on the extracted state change sequence and the corresponding current data segment to form an input sequence pair, and then inputs the sequence pair into the long short-term memory neural network unit, and controls the transmission and update of the state information through the input gate, forget gate and output gate, and gradually processes each time step data in the time series. At the same time, the network records the historical state information and combines it with the current input to output the predicted state change trend. By comparing the predicted results with the actual state mutation position, the corresponding time point of the mutation is extracted and the synchronous mutation index of the state and current is output to construct a state current mutation corresponding group.

[0013] As a further solution of the present invention, the specific steps of generating the jumping behavior risk index cluster are: Based on the abnormal state reconstructed path group, a jump segment identification operation is performed, non-single-step number pairs are extracted by calculating the difference between consecutive state numbers and marking the start and end point indexes, corresponding sequence slices are generated for each group of number jump segments and the position relationship within the sequence is recorded to generate a jump segment index set; Based on the jump segment index set, a response residual calculation operation is performed by extracting the response time corresponding to the start and end states of each segment and calculating the difference, then comparing it with the fixed action tolerance limit, marking all segments that exceed the tolerance value and associating the state number with the trigger time, and generating a response abnormal segment list; Based on the response abnormal segment list, the abnormal segment device number and jump position classification operation are performed, and a two-dimensional grouping structure is constructed by extracting the starting number and device number of each state segment. Random forest is used to model the jump risk with the device number, the number of abnormal segments, and the residual distribution as multiple inputs. The behavior score is output and the abnormal frequency index matrix is ​​generated to obtain the jump behavior risk index cluster.

[0014] As a further solution of the present invention, the random forest is according to the formula:

[0015] in: represents the jump risk score value, represents the total number of decision trees in the random forest, Indicates the The weight coefficient of a decision tree, Indicates the The prediction function of a decision tree, Indicates the device number, Indicates the number of abnormal segments of the corresponding device, Represents the residual distribution feature set of the device, represents the coefficient of variation of the state switching period, Indicates the jump frequency per unit time.

[0016] As a further solution of the present invention, the random forest first constructs an input sample set characterized by device number, number of abnormal segments and response residual distribution parameters, forms a multidimensional feature vector for each device, and then uses the training set to generate multiple decision trees. Each tree is constructed by random sampling with replacement from the original sample, and nodes are divided by randomly selecting some features to generate a structurally differentiated classification model. After all decision trees are constructed, the majority voting mechanism is used to classify the risk scores of each jump behavior, output the jump risk level corresponding to each device, and generate an abnormal frequency index matrix.

[0017] As a further solution of the present invention, the specific steps of generating the high-frequency jump device risk sequence table are: Based on the jump behavior risk index cluster, an operation of extracting the device number in the jump segment is performed. By reading the device bit code of the index field and sorting it by segment group, the corresponding start time and end time of each group of jump segments are extracted and converted into a unified time unit. The jump segment identification index is extracted and bound to the time point for mapping, and a jump segment device index information table is generated; Based on the jump segment device index information table, a jump segment count operation is performed under each group of device numbers. A jump frequency table is formed by accumulating the number of segments corresponding to each number, and the start and end times are subtracted to obtain the response duration of each segment. After normalization conversion of each segment response duration, the response duration is aligned with the jump frequency by column to form a two-dimensional matrix, thereby generating a jump behavior response parameter matrix. Based on the jump behavior response parameter matrix, the jump frequency corresponding to each device number is multiplied by the response residual value, and the total score is obtained by summarizing the product values ​​by device number. The score values ​​are sorted by size and the result table is renumbered by serial number. The sorting results are output in association with the device number to obtain a high-frequency jump device risk sequence table.

[0018] Compared with the prior art, the advantages and positive effects of the present invention are: In the present invention, the starting time and response time in the tripping state transition sequence of the protection device are extracted to calculate the execution delay and action segment length. The abnormal link is identified based on the state number sequence comparison, and the state logic structure is analyzed and reconstructed in a standardized manner to form an orderly tripping path. At the same time, the transfer nodes with the same triggering conditions but different state successors are screened, and the deviation relationship between the state structure change and the consistency of the triggering conditions is clarified to generate a state jump graph with the ability to describe behavioral divergence. In this invention, a long short-term memory neural network is introduced to time-align the state change sequence with the current signal segment. Through the linkage analysis of the state-current mutation position, potential mismatches or abnormal change trends in the state chain are identified, and the state path is reconstructed to enhance the path continuity and dynamic adaptability. In the present invention, random forest is used to take the device number, number of jump segments and response residual as multiple inputs for training modeling, and the device score is output through the integrated tree voting mechanism to realize the identification and sorting of devices with significant abnormal frequency, effectively supporting the system to actively check, prioritize and sort the risk levels of high-risk devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the main steps of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0021] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0022] Example 1 See also Figure 1 The present invention provides a technical solution: a calibration method for an intelligent integrated protection device, comprising the following steps: S1: Based on the event change sequence and switch status records recorded during the operation of the protection device, the start time and trip response time are extracted, the execution delay and action duration are calculated, and the state numbers are compared for sequential consistency to obtain the trip logic transfer path set; S2: Based on the trip logic transfer path set, the transfer nodes with the same trigger condition value but different state successors in the blocking logic module are screened, the jump time and path structure differences are extracted, and the state divergence point bitmap set is generated; S3: Based on the state divergence point atlas, a long short-term memory neural network is used to extract the state change sequence and current data fragments in the alarm trigger logic, perform state sequence comparison and mutation detection, insert the reconstructed fragments to form a sequence path flow, and obtain the abnormal state reconstructed path group; S4: Reconstruct the path group based on the abnormal state, identify the jump segment position in the reversion response mechanism, calculate the response residual and determine the over-limit segment, use random forest to locate the state starting point corresponding to the jump segment and classify and summarize it to obtain the jump behavior risk index cluster; S5: Based on the jumping behavior risk index cluster, extract the device number and time information corresponding to the jumping segment, count the response residual and the number of jumps, perform scoring and sorting, and establish a high-frequency jumping device risk sequence table.

[0023] The specific steps for generating a trip logic transfer path set are: Based on the event change sequence and switch status records recorded during the operation of the protection device, the time of the first and last nodes in each sequence is read. By performing a difference calculation between the start time and the completion time and uniformly numbering the node labels, the time units of each record are checked and converted to a unified reference time format. After eliminating time abnormality records and completing the cleaning, the duration of the tripping segment is calculated to generate a trip response time parameter group; Based on the trip response time parameter group, the state numbers of each section are sorted. By performing a one-by-one comparison of the continuous number sequence, it is determined whether the index difference is the minimum single-step value. A reverse matching check is then performed. The numbers with gaps are marked and offset compensated. The number streams that pass the consistency check are structured and organized to generate a state number sequence structure set. Based on the state number sequence structure set, the number sequence is bound to the time parameter, each group of numbers is sequentially matched to the response time period and rearranged into a time stream. The time stream is divided into the action triggering phase and the recovery phase. The sequence is spliced ​​and combined in sections and the action chain is reorganized according to the triggering order to obtain the tripping logic transfer path set. Based on the event change sequence and switch status records recorded during the operation of the protection device, the time field parsing and difference calculation method is used to manually set the standard format of the first and last node time information in each event record. It is read in the form of year-month-day hour: minute: second. millisecond and converted into a unified time unit of milliseconds. The difference operation between the start time and the completion time is performed, and the state segment time span is calculated using the integer timestamp method. The event tags are mapped through the numbering table and unified numbering is performed. The time fields of records from different sources are proofread to the millisecond benchmark format, with a unified time granularity of 1 millisecond. Records that span days or have time sequence reversals are eliminated, and records with missing or abnormal time fields are deleted. Then, the start and end time difference corresponding to each jump state is counted, the duration of each jump segment is calculated, and a trip response time parameter group is generated. The time information in the original event data is standardized, unified, and the data legitimacy is verified, providing a complete time constraint foundation for the subsequent state sequence construction. Based on the trip response time parameter group, a state number continuity comparison and compensation method is adopted. Each state number sequence in the jump segment is arranged in ascending order. The difference judgment operation is performed on every two adjacent numbers. The index difference between the numbers is calculated to see if it is a single-step increment. The number difference threshold is set to 1, and the jump number points are screened out as interruption marks. Subsequently, offset compensation is performed on the records with number jump segments. The missing numbers are inserted into the original number stream in sequence and compensation marks are established. The logical consistency of the compensated number stream is checked and residual conflicting items are eliminated. A structured number stream is constructed for all number sequences that pass the check, generating a state number sequence structure set. A continuous state number sequence with no missing items is constructed, which helps to accurately map the action phase and improve path consistency. Based on the state number sequence structure set, the number-time period correspondence reorganization method is adopted to bind the state numbers and tripping response time periods one by one, establish the start time and end time period structure corresponding to each state number, and rearrange the bound number time pairs in the number sequence, dividing them into the action triggering stage and the state recovery stage. The paragraphs are segmented according to the stage boundary labels in the number flow, and the stages are spliced ​​and combined in sequence to merge into a complete action path chain. The empty or repeated segments in the spliced ​​path are removed and re-sequenced to obtain the tripping logic transfer path set. The discrete numbers and time parameters are integrated into a complete path flow to achieve the reconstructible and serialized expression of the action chain.

[0024] The specific steps to generate the state bifurcation point atlas are as follows: Based on the trip logic transfer path set, the trigger conditions of all state segments in the blocking logic module are extracted. The input condition values ​​are grouped and state pairs with exactly the same input values ​​are screened. All subsequent state information in the state pairs is obtained and type-labeled. State pairs with structural differences are identified and grouped into the same set to generate groups with consistent and divergent trigger conditions. Based on the trigger condition consistency and disagreement groups, the difference between the start time of each state segment and the response time span is calculated. The trigger time is matched with the sequence number of the jump node. The path length between each pair of nodes is calculated and the jump direction and type are summarized. The path difference attributes are recorded in groups and the missing segments are filled in to generate the path jump structure feature set. Based on the path transition structure feature set, a mapping sequence is constructed between all transition node codes and response intervals. A unified numbering format is set and a unified numbering conversion is performed. The path type, response period, and number position are integrated into a structure matrix. The offset value of each transition segment in the matrix is ​​analyzed to obtain a state divergence point bitmap set. Based on the tripping logic transfer path set, the condition grouping and state pair matching method is adopted to extract the input trigger conditions of all state segments in the path set one by one, and the value of each input condition field is matched using the character field comparison method. The judgment rule is set to the completeness of the input fields as the screening basis. Then, the input field of each state segment is used as the primary key through the key-value pair mapping mechanism. Grouped by the same input value, the state segment numbers in the group are extracted to form a state pair, and all subsequent state records in the state pair are extracted. The subsequent states are classified into adjacent type and jump type according to the size of the number. The type tag value of each type of subsequent state is set to 1 and 2 respectively. The marked state pairs are classified and archived by group. The structural difference state groups are summarized by field association to generate trigger condition consistency and divergence groups. Based on the trigger condition consistency and divergence groups, the path difference calculation and structural difference clustering methods are used to perform difference calculation on the start time of each state segment and the corresponding response time period. The time field is set to milliseconds, the start field is start_time, the response field is end_time, and the difference field is set to duration. The calculation operation is end_time minus start_time. Then, the state segment trigger time and the jump node number are paired. The number difference calculation rule is set to the next number minus the previous number. The matching rule is that if the number difference is not equal to 1, it is marked as a jump segment. After extracting the jump segment, the path length calculation method is used to count the number of hops between each pair of nodes. The path length is counted in absolute value. Then, the jump direction is marked forward and backward, with the forward direction marked as F and the backward direction marked as B. Finally, the path difference type and direction of each state pair are annotated and aggregated by number. The missing number segments in the discontinuous jump path are supplemented by interpolation to generate a path jump structure feature set. Based on the path transition structure feature set, a number coding conversion and response mapping construction method is adopted. The numbers of all transition nodes are uniformly coded and the unified number format is set to 6-bit integer. Insufficient digits are padded with zeros. For example, the original number 23 is converted to 000023. The response time field is calculated based on the millisecond format and a correspondence is established between each number and time period. A mapping sequence with the structure of number-response interval pairs is constructed. The value range of the path type field is set to {adjacent, jump}, and the value range of the number position is set to {first, middle, last}. The path type, response interval period, and number position are constructed in a triple format and uniformly combined into a structure matrix. After sorting by the number field, the matrix rows are traversed. For each transition segment, the absolute difference between the current number position and the previous number position is recorded as the offset value. The record field is set to offset. The offset values ​​of all transition segments are written into the corresponding structure matrix rows to obtain the state branch point bitmap set.

[0025] The specific steps for generating an abnormal state reconstruction path group are: Based on the state divergence point atlas, a long short-term memory neural network is used to perform time alignment operations on the state change sequence and the current sequence. The interception range is determined by extracting the state start and end indexes and the corresponding current value interval is extracted. The amplitude changes of adjacent data points are calculated and the mutation position indexes are marked. The state synchronization mutation positions are extracted and an associated index table is constructed to generate the corresponding group of state and current mutations. Based on the state-current mutation correspondence group, a one-to-one matching operation is performed between the state and the current mutation point position. The index difference between the two types of mutation points is detected and a threshold is set to filter out the misaligned segments. The unmatched state segments are re-labeled and stored separately from the original sequence. The index offset processing is performed on all inserted fragments to generate a recombined state inserted fragment set; Based on the recombinant state insertion fragment set, the original state sequence and the insertion fragment are spliced ​​together. By extracting the front and back numbers of each splicing point, the sequence deviation is judged and discontinuous segments are processed. The merged state streams are sorted according to the triggering order and uniformly indexed and named to obtain the abnormal state reconstruction path group. Based on the state divergence point atlas, a long short-term memory neural network is used to perform time alignment on the state change sequence and the current sequence. The input format is set to a two-dimensional tensor shape [t, f], where t is the number of time steps and f is the number of features input at each step. The state sequence and current sequence are constructed as two independent input tensors respectively. The state start index and end index are extracted by sliding window method. The window length is set to 20 steps and the step size is 5 steps. The original current data interval is intercepted as input. The input current value is normalized and the normalization rule is min-max scaling. The range is set to [0, 1]. Then, the amplitude change of adjacent data points is calculated in each window by subtracting the previous value from the current value. The position of the data point whose change exceeds the set threshold is extracted and the corresponding index is recorded as the mutation point. The time index of the state change point and the current mutation point is extracted. The mutation position association mapping is established by using the corresponding position coincidence judgment. An index table containing the state number, current mutation index, and synchronization time field is constructed to generate the state and current mutation corresponding group. Based on the state current mutation corresponding group, a mutation point matching and number reconstruction method is used to perform a one-to-one matching operation on the state mutation point and current mutation point index. The index difference threshold is set to ±2, and the index difference of each group of state points and current points is calculated. The records with the absolute value of the difference greater than 2 are marked as unaligned segments. The unmatched state number segments are renumbered by adding an offset value of 1000 to the original number to distinguish them. The renumbered segments are stored separately from the main sequence and the new number segments are set as the insertion segment area. Index offset processing is performed on the insertion segments. The offset value is generated incrementally in the renumbering order. The insertion number stream is cumulatively updated starting from the offset starting point plus 1 to ensure number continuity and generate a reorganized state insertion segment set. Based on the recombinant state insertion fragment set, the state splicing and number sorting method is used to merge the original state sequence and the insertion fragment, extract the front and back numbers of each splicing point, and perform a sequence deviation judgment operation. The numbering sequence judgment condition is that the back number minus the front number is 1. If it is not met, the location will be marked as a discontinuous segment, and its number pair will be recorded and the correction strategy will be called. The correction method is to insert a virtual number segment to fill the number breakpoint. The virtual number segment is named in the form of "999xxx" and added to the number stream. Then, all numbers are merged to construct a complete state stream. The number fields in the merged sequence are re-sorted in ascending order, and a unified index naming rule is reset for each number. The number field plus the serial number is named in the format of "ID_serial number". Finally, the number reconstruction after splicing is completed, and the abnormal state reconstruction path group is obtained.

[0026] Long short-term memory neural network, according to the formula:

[0027] in: Indicates time The hidden state vector of Represents the activation function, which is used for nonlinear transformation. represents the weight adjustment coefficient of the input path, represents the weight matrix input to the hidden layer, Indicates time The composite input feature vector of represents the weight adjustment coefficient of the recursive path, represents the recursive weight matrix from the previous hidden state to the current state, represents the hidden state vector at the previous moment, represents the weight factor of the mutation density channel, Indicates time The state mutation density within the time window, represents the weight factor of the stability channel, Indicates time The stability index of the current signal, represents the bias term of the input path, Represents the bias term of the recursive path; The long short-term memory neural network first time-aligns the extracted state change sequence with the corresponding current data fragment to form an input sequence pair. The sequence pair is then input into the long short-term memory neural network unit. The input gate, forget gate, and output gate control the transmission and update of state information, gradually processing each time step in the time series. At the same time, the network records historical state information and combines it with the current input to output the predicted state change trend. By comparing the predicted result with the actual state mutation position, the corresponding time point of the mutation is extracted and the synchronous mutation index of the state and current is output to construct the state and current mutation corresponding group; Execution process: First, obtain the composite input feature vector at the current moment , which consists of state code, current value, current difference and change direction, represents the multi-dimensional input signal of the device under the current working condition. Then, and the hidden state vector at the previous moment The input paths and recursive paths are respectively fed into the network, and the weight coefficients corresponding to each path are and , which indicates the degree of dependence of the model on the current input features and historical states, is obtained by automatic adjustment through error minimization during the training phase, and then the current mutation density parameter is introduced , which represents the number of mutation points per unit time in the current time window, reflecting the frequency of current system state fluctuations, and at the same time introduces the current stability parameter , defined as the ratio of the current standard deviation to the average value in the current time window, which describes the fluctuation amplitude of the current signal. The two are multiplied by the channel weight coefficient and Then it is incorporated into the calculation path, and the information gain score and loss function regularization term are dynamically adjusted to ensure the model discrimination accuracy. Finally, all weighted terms and bias terms are added together and then activated by the function Perform nonlinear mapping and output the hidden state vector at the current moment , used to support subsequent state-current synchronous mutation detection and verification judgment, and realize synchronous consistency verification of protection device state logic and physical electrical quantities.

[0028] The specific steps to generate the jumping behavior risk index cluster are: Based on the abnormal state, the path group is reconstructed and the jump segment identification operation is performed. The non-single-step number pairs are extracted by calculating the difference between the consecutive state numbers and the start and end point indexes are marked. For each group of number jump segments, a corresponding sequence slice is generated and the position relationship within the sequence is recorded to generate a jump segment index set; Based on the jump segment index set, the response residual calculation operation is performed by extracting the response time corresponding to the start and end states of each segment and calculating the difference. Then, it is compared with the fixed action tolerance limit, marking all segments that exceed the tolerance value and associating the state number with the trigger time to generate a list of response abnormal segments; Based on the list of response abnormal sections, the abnormal section device number and jump position are classified. A two-dimensional grouping structure is constructed by extracting the starting number and device number of each section state. A random forest is used to model the jump risk with the device number, number of abnormal sections, and residual distribution as multiple inputs. The behavior score is output and an abnormal frequency index matrix is ​​generated to obtain the jump behavior risk index cluster. Based on the abnormal state reconstruction path group, the state number difference screening method is adopted to perform the difference calculation operation between the numbers on all state numbers. The difference calculation method is set to the current number minus the previous number, and the single-step threshold is set to 1. The number pairs with number differences not equal to 1 are screened and their index positions are recorded. All discontinuous number segments are extracted and the starting and ending numbers are marked. The corresponding fragments in the original state stream are intercepted according to the recorded index start and end values. Each fragment is constructed as a set of state number sequence slices and recorded. The offset format is the distance from the starting number position to the end number position. Finally, all jump segments are sorted out in the order of the fragments and output as a number jump segment index group to generate a jump segment index set. Based on the jump segment index set, the response time residual comparison method is adopted to extract the corresponding response time field for each start state and end state in the jump segment. The format is unified as millisecond timestamp, and the calculation method is to subtract the start time from the end time. The calculation result is defined as the response time difference residual. The tolerance threshold is set to 500 milliseconds. Segments greater than the threshold are used for screening. The screening rule is that if the time difference residual is greater than 500, it is judged as an out-of-limit segment. The start state number, residual value and corresponding trigger time of all out-of-limit segments are recorded. All out-of-limit segments are summarized in order of number and output as a time anomaly segment record list to generate a response anomaly segment list; Based on the response anomaly segment list, the random forest algorithm is used. The device number, number of abnormal segments and response residual value in each abnormal record are used as feature inputs to construct the input feature vector format of [X1, X2, X3], where X1 is the value corresponding to the device number, X2 is the number of abnormal segments corresponding to the number, and X3 is the average of all residual values. The training set grouping ratio is set to 7:3. The random forest classification method is used to perform modeling operations on the training data. The number of trees is set to 100 and the maximum depth is 10. Information gain is used as the partitioning criterion. The random forest model is generated through training. The model outputs the anomaly score value corresponding to each device, and then the score values ​​are sorted. The sorting field is set to the score value from high to low. Finally, all the device numbers and jump frequencies corresponding to the scores are summarized to form a two-dimensional score matrix to obtain the jumping behavior risk index cluster.

[0029] Random forest, according to the formula:

[0030] in: represents the jump risk score value, represents the total number of decision trees in the random forest, Indicates the The weight coefficient of a decision tree, Indicates the The prediction function of a decision tree, Indicates the device number, Indicates the number of abnormal segments of the corresponding device, Represents the residual distribution feature set of the device, represents the coefficient of variation of the state switching period, Indicates the jump frequency per unit time; Random forests first construct an input sample set characterized by device number, number of abnormal segments, and response residual distribution parameters. A multidimensional feature vector is formed for each device. Multiple decision trees are then generated using the training set. Each tree is constructed by random sampling with replacement from the original sample and node partitioning is performed by randomly selecting some features to generate a structurally differentiated classification model. After all decision trees are constructed, a majority voting mechanism is used to classify the risk scores of each jump behavior, outputting the jump risk level corresponding to each device and generating an abnormality frequency index matrix. Execution process: First, extract the number of each device , count the number of corresponding abnormal segments , and then calculate the residual distribution characteristics based on historical behavior data , including the maximum deviation, mean deviation, standard deviation and skewness between the observed value and the theoretical model in the abnormal section, which is used to measure the stability of the response anomaly. Then the state switching time series is extracted to calculate the periodic variation coefficient , as a quantitative indicator of whether the state switching rhythm fluctuates abnormally, and the jump frequency is calculated at the same time , that is, the number of discontinuous state transitions per unit time, is used to measure the degree of dynamic instability of the device, and the input features constitute the feature vector input to each decision tree Risk prediction is performed using the weight coefficient Weight the prediction output of each tree, According to the normalization of the prediction accuracy of each tree on the training set, the final The weighted outputs of the trees are summed to obtain the jump risk score , which is used to establish an abnormal frequency index matrix and divide jumping behavior risk index clusters.

[0031] The specific steps to generate the high-frequency jump device risk sequence table are: Based on the jump behavior risk index cluster, the device number extraction operation in the jump segment is performed. By reading the device bit code of the index field and sorting it by segment group, the corresponding start time and end time of each group of jump segments are extracted and converted into a unified time unit. The jump segment identification index is extracted and bound to the time point for mapping, and the jump segment device index information table is generated; Based on the jump segment device index information table, the number of jump segments under each group of device numbers is counted. The jump frequency table is formed by accumulating the number of segments corresponding to each number, and the start and end times are subtracted to obtain the response duration of each segment. After normalization, each response duration is aligned with the jump frequency by column to form a two-dimensional matrix, generating a jump behavior response parameter matrix. Based on the jump behavior response parameter matrix, the jump frequency corresponding to each device number is multiplied by the response residual value. The total score is obtained by summing up the product values ​​by device number. The score values ​​are sorted and the result table is renumbered by sequence number. The sorted results are output in association with the device number to obtain the risk sequence table of high-frequency jump devices. Based on the jump behavior risk index cluster, the field parsing and time standardization methods are used to traverse and extract the device number field recorded in the jump segment. The device code indicated in each record is used as the grouping basis for classification and sorting. The start time and end time marked for each group of jump segment records are extracted, and the data is uniformly read in the format of year-month-day hour: minute: second. millisecond and then transcribed into a unified pure digital format. The start and end time values ​​are recorded in milliseconds. The jump identification number is extracted from each jump segment information according to the original index field, and the corresponding relationship between it and the time field is established. Finally, the jump segment index, device number, and start and end time are encapsulated into a field structure to generate a jump segment device index information table. This completes the standardized device number extraction and unified time binding of the jump segment structure, forming a statistical data infrastructure. Based on the jump segment device index information table, the number of counts and response time normalization method are used to count the jump segment entries under each device number group, and the jump number results corresponding to each device number are generated by accumulating them one by one. At the same time, the start time and end time in each jump segment record are directly subtracted to obtain the response time, and all response time values ​​are uniformly linearly mapped and converted to a normalized value between 0 and 1. The normalized response time and jump number are merged into a double-field record of the same device in a row matching manner. Each record is output as a ternary table item containing the device number, jump number, and normalized response time. The jump behavior response parameter matrix is ​​generated to clarify the frequency and persistence characteristics of each device in the response behavior, and to construct the input basis for subsequent scoring; Based on the jump behavior response parameter matrix, a behavior score calculation and hierarchical sorting method is adopted. The jump frequency value corresponding to the device number in each record is multiplied by the normalized value of the response time to obtain the corresponding behavior score value. Then, the device number is grouped and all the score values ​​are summed to obtain the total score result for each device number. The score values ​​are sorted and renumbered in descending order to generate a sorting field. The sorting field is paired with the corresponding device number. Finally, the device number, total score and sorting number in each record are output to obtain the risk sequence list of high-frequency jump devices. A device grading system based on the quantitative superposition of abnormal characteristics is constructed to form an orderly risk identification list.

[0032] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A calibration method for an intelligent integrated protection device, characterized in that: The following steps are involved: S1: Based on the event change sequence and switch status records recorded during the operation of the protection device, the start time and trip response time are extracted, the execution delay and action duration are calculated, and the state numbers are compared for sequential consistency to obtain the trip logic transfer path set; S2: Based on the trip logic transfer path set, filter the transfer nodes with the same trigger condition value but different state successors in the blocking logic module, extract the jump time and path structure differences, and generate a state branch point bitmap set; S3: Based on the state divergence point atlas, a long short-term memory neural network is used to extract the state change sequence and current data fragments in the alarm trigger logic, perform state sequence comparison and mutation detection, insert the reconstructed fragments to form a sequence path flow, and obtain the abnormal state reconstructed path group; S4: Reconstructing the path group based on the abnormal state, identifying the jump segment position in the restoration response mechanism, calculating the response residual and determining the over-limit segment, using random forest to locate the state starting point corresponding to the jump segment and classify and summarize it to obtain the jump behavior risk index cluster; S5: Based on the jump behavior risk index cluster, extract the device number and time information corresponding to the jump segment, count the response residual and the number of jumps, perform scoring and sorting, and establish a high-frequency jump device risk sequence table.

2. The verification method of the intelligent integrated protection device according to claim 1, characterized in that: The specific steps of generating the trip logic transfer path set are: Based on the event change sequence and switch status records recorded during the operation of the protection device, the time of the first and last nodes in each sequence is read. By performing a difference calculation between the start time and the completion time and uniformly numbering the node labels, the time units of each record are checked and converted to a unified reference time format. After eliminating time abnormality records and completing the cleaning, the duration of the tripping segment is calculated to generate a trip response time parameter group; Based on the trip response time parameter group, the state numbers of each section are sorted, and the index difference of the continuous number sequence is determined to be the minimum single-step value by performing a comparison item by item, and a reverse matching check is performed. The numbers with gaps are marked and offset compensated. The number streams that pass the consistency check are structured and organized to generate a state number sequence structure set; Based on the state number sequence structure set, the number sequence is bound to the time parameter, each group of numbers is sequentially matched to the response time period and rearranged into a time stream, the action triggering phase and the recovery phase in the time stream are divided, the sequence is spliced ​​and combined in segments, and the action chain is reorganized according to the triggering order to obtain a set of tripping logic transfer paths.

3. The verification method of the intelligent integrated protection device according to claim 1, characterized in that: The specific steps of generating the state divergence point bitmap set are: Based on the trip logic transfer path set, extract the input trigger conditions of all state segments in the blocking logic module, group them by input condition values ​​and filter out state pairs with exactly the same input values, obtain all subsequent state information in the state pairs and perform type annotation, identify state pairs with structural differences and classify them into the same group set, and generate trigger condition consistency and divergence groups; Based on the trigger condition consistency and disagreement groups, the difference between the start time of each state segment and the response time span is calculated, the trigger time is matched with the sequential number of the jump node, the path length between each pair of nodes is calculated and the jump direction and type are summarized, the path difference attributes are recorded in groups and the missing segments are supplemented to generate a path jump structure feature set; Based on the path jump structure feature set, a mapping sequence is constructed between the encoding of all jump nodes and the response interval time, a unified numbering format is set and a unified numbering conversion is performed, the path type, response period and numbering position are integrated into a structure matrix, the offset value analysis is completed for each jump segment in the matrix, and a state divergence point bitmap set is obtained.

4. The verification method of the intelligent integrated protection device according to claim 1, characterized in that: The specific steps of generating the abnormal state reconstruction path group are: Based on the state divergence point atlas, a long short-term memory neural network is used to perform a time alignment operation on the state change sequence and the current sequence. The interception range is determined by extracting the state start and end indexes and the corresponding current value interval is extracted. The amplitude change of adjacent data points is calculated and the mutation position index is marked. The state synchronization mutation position is extracted and an associated index table is constructed to generate a state current mutation corresponding group. Based on the state current mutation corresponding group, a one-to-one matching operation is performed between the state and the current mutation point position, the index difference between the two types of mutation points is detected and a threshold is set to filter out the unaligned segments, the unmatched state segments are re-labeled and numbered and stored separately from the original sequence, and index offset processing is performed on all inserted fragments to generate a recombined state inserted fragment set; Based on the recombinant state insertion fragment set, the original state sequence and the insertion fragment are spliced ​​together. By extracting the front and back numbers of each splicing point, sequence deviation judgment is performed and discontinuous sections are processed. The merged state streams are sorted according to the triggering order and uniformly indexed and named to obtain the abnormal state reconstruction path group.

5. The verification method of the intelligent integrated protection device according to claim 4, characterized in that: The long short-term memory neural network is based on the formula: in: Indicates time The hidden state vector of Represents the activation function, which is used for nonlinear transformation. represents the weight adjustment coefficient of the input path, represents the weight matrix input to the hidden layer, Indicates time The composite input feature vector of represents the weight adjustment coefficient of the recursive path, represents the recursive weight matrix from the previous hidden state to the current state, represents the hidden state vector at the previous moment, represents the weight factor of the mutation density channel, Indicates time The state mutation density within the time window, represents the weight factor of the stability channel, Indicates time The stability index of the current signal, represents the bias term of the input path, Represents the bias term of the recursive path.

6. The verification method of the intelligent integrated protection device according to claim 4, characterized in that: The long short-term memory neural network first performs time alignment processing on the extracted state change sequence and the corresponding current data segment to form an input sequence pair, and then inputs the sequence pair into the long short-term memory neural network unit. The transmission and update of the state information are controlled by the input gate, the forget gate and the output gate, and each time step data in the time series is processed step by step. At the same time, the network records the historical state information and combines it with the current input to output the predicted state change trend. By comparing the predicted result with the actual state mutation position, the corresponding time point of the mutation is extracted and the synchronous mutation index of the state and current is output to construct the state current mutation corresponding group.

7. The verification method of the intelligent integrated protection device according to claim 1, characterized in that: The specific steps of generating the jumping behavior risk index cluster are: Based on the abnormal state reconstructed path group, a jump segment identification operation is performed, non-single-step number pairs are extracted by calculating the difference between consecutive state numbers and marking the start and end point indexes, corresponding sequence slices are generated for each group of number jump segments and the position relationship within the sequence is recorded to generate a jump segment index set; Based on the jump segment index set, a response residual calculation operation is performed by extracting the response time corresponding to the start and end states of each segment and calculating the difference, then comparing it with the fixed action tolerance limit, marking all segments that exceed the tolerance value and associating the state number with the trigger time, and generating a response abnormal segment list; Based on the response abnormal segment list, the abnormal segment device number and jump position classification operation are performed, and a two-dimensional grouping structure is constructed by extracting the starting number and device number of each state segment. Random forest is used to model the jump risk with the device number, the number of abnormal segments, and the residual distribution as multiple inputs. The behavior score is output and the abnormal frequency index matrix is ​​generated to obtain the jump behavior risk index cluster.

8. The verification method of the intelligent integrated protection device according to claim 7, characterized in that: The random forest is based on the formula: in: represents the jump risk score value, represents the total number of decision trees in the random forest, Indicates the The weight coefficient of a decision tree, Indicates the The prediction function of a decision tree, Indicates the device number, Indicates the number of abnormal segments of the corresponding device, Represents the residual distribution feature set of the device, represents the coefficient of variation of the state switching period, Indicates the jump frequency per unit time.

9. The verification method of the intelligent integrated protection device according to claim 7, characterized in that: The random forest method first constructs an input sample set characterized by device number, number of abnormal segments, and response residual distribution parameters, forms a multidimensional feature vector for each device, and then uses the training set to generate multiple decision trees. Each tree is constructed by random sampling with replacement from the original sample, and nodes are divided by randomly selecting some features to generate a structurally differentiated classification model. After all decision trees are constructed, a majority voting mechanism is used to classify the risk scores of each jump behavior, output the jump risk level corresponding to each device, and generate an abnormal frequency index matrix.

10. The verification method of the intelligent integrated protection device according to claim 1, characterized in that: The specific steps for generating the high-frequency jump device risk sequence table are: Based on the jump behavior risk index cluster, an operation of extracting the device number in the jump segment is performed. By reading the device bit code of the index field and sorting it by segment group, the corresponding start time and end time of each group of jump segments are extracted and converted into a unified time unit. The jump segment identification index is extracted and bound to the time point for mapping, and a jump segment device index information table is generated; Based on the jump segment device index information table, a jump segment count operation is performed under each group of device numbers. A jump frequency table is formed by accumulating the number of segments corresponding to each number, and the start and end times are subtracted to obtain the response duration of each segment. After normalization conversion of each segment response duration, the response duration is aligned with the jump frequency by column to form a two-dimensional matrix, thereby generating a jump behavior response parameter matrix. Based on the jump behavior response parameter matrix, the jump frequency corresponding to each device number is multiplied by the response residual value, and the total score is obtained by summarizing the product values ​​by device number. The score values ​​are sorted by size and the result table is renumbered by serial number. The sorting results are output in association with the device number to obtain a high-frequency jump device risk sequence table.

Images