Power distribution network dispatching voice accurate transcription method based on topological relation constraint

Abstract

This invention discloses a method for accurate transcription of distribution network dispatch voice based on topology constraints, belonging to the field of distribution network dispatch voice transcription technology. The method includes the following steps: acquiring distribution network dispatch voice and a distribution network topology model, generating candidate transcription segments with timestamps; determining the segments to be confirmed and retaining multiple candidate device objects; parsing subsequent associated candidate transcription segments and extracting topology relationship information; applying backtracking constraints to the candidate device objects based on the topology relationship information to determine the target device object; and backfilling the target device object to the corresponding timestamp to generate the transcribed text. This invention can reduce the risk of mistranscription of device names, numbers, or abbreviations, and improve the accuracy of dispatch records and instruction parsing.

Description

Technical Field

[0001] This invention relates to the field of voice transcription technology for distribution network dispatching, and more specifically, to a method for accurate voice transcription of distribution network dispatching based on topological constraints. Background Technology

[0002] Voice transcription of distribution network dispatching is a crucial step in distribution network dispatching automation, dispatching record archiving, and auxiliary parsing of dispatching instructions. Current technologies typically use speech recognition models to convert dispatching calls or instructions into text. This text is then combined with power industry lexicons, equipment name libraries, dispatching terminology lists, or pre-set semantic templates to correct equipment names, operational verbs, and dispatching terminology in the transcribed text, thereby improving the completeness and readability of the dispatching voice recordings.

[0003] However, in distribution network dispatching voice scenarios, dispatchers often speak the equipment name, equipment number, abbreviation, or pronoun in a conversational order, and then supplement the topological constraints such as the upstream and downstream range, connection relationship, transfer path, isolation boundary, or power outage / restoration range corresponding to the equipment. Existing speech-to-text methods usually determine the equipment name during the text generation stage and then perform dictionary correction or topological verification after text generation. They lack mechanisms for processing the expression fragments of equipment objects for confirmation and for retaining multiple candidate equipment objects. When subsequent speech contains topological relationship information that can define the preceding equipment objects, existing methods often can only determine whether the output text is abnormal. It is difficult to backtrack and fill in the positions of the preceding equipment object candidates. This can easily lead to the equipment objects in the transcribed text being real distribution network equipment, but inconsistent with the topological relationship actually expressed in the dispatching voice, thus affecting the accuracy of dispatching records, command parsing, and operation ticket generation.

[0004] The above-disclosed technical solutions have at least the following technical problems: the existing distribution network dispatch voice transcription method usually determines the device name first and then performs post-verification. It does not retain multiple candidate device objects for the case where the device object expression appears earlier than the topology constraint information, which makes it difficult to backtrack and correct the position of the preceding device object based on the topology relationship information. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for accurate transcription of distribution network dispatch voice based on topology constraints. By retaining multiple candidates of preceding device objects and using subsequent topology information for backtracking constraints, this method solves the problem in existing distribution network dispatch voice transcription where device objects are determined too early and are difficult to correct later, resulting in inconsistencies between the transcription results and the actual topology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for accurate transcription of distribution network dispatch voice based on topology constraints includes the following steps: acquiring distribution network dispatch voice and distribution network topology model; segmenting and recognizing the distribution network dispatch voice to generate candidate transcription segments with timestamps; identifying segments involving distribution network equipment or equipment numbers from the candidate transcription segments and reserving multiple candidate equipment objects for the segments to be confirmed; using the timestamp of the segment to be confirmed as a reference, parsing its subsequent associated candidate transcription segments and extracting topology relationship information used to limit the candidate equipment objects; applying backtracking constraints to the multiple candidate equipment objects based on the topology relationship information and distribution network topology model to determine the target equipment object; and backfilling the target equipment object to the corresponding timestamp position to generate the transcribed text of the distribution network dispatch voice.

[0007] In a preferred embodiment, the step of segmenting and recognizing the distribution network dispatch voice to generate candidate transcription segments with timestamps includes: preprocessing the distribution network dispatch voice to obtain a voice stream to be recognized; determining initial voice segments based on the voice stream to be recognized, and merging relevant initial voice segments into a recognition window according to the continuity of the dispatch voice; performing multi-candidate recognition on the recognition window, retaining candidate recognition results containing device identifiers or numbers, and generating candidate transcription segments.

[0008] In a preferred embodiment, determining the initial speech segments based on the speech stream to be identified includes: determining candidate segmentation positions based on the speech features of the speech stream to be identified, wherein the speech features include at least one of pause features, speech rate variation features, and energy variation features; determining whether the candidate segmentation position disrupts the expression of scheduling operation terms or device names; if not, determining the candidate segmentation position as the initial segmentation boundary, and obtaining multiple initial speech segments based on the initial segmentation boundary.

[0009] In a preferred embodiment, determining the segment to be confirmed involving distribution network equipment or equipment number from the candidate transcribed segments includes: performing distribution network equipment entity recognition on the candidate text in the candidate transcribed segments and extracting the equipment information to be matched; matching the equipment information to be matched with the equipment ledger and the distribution network topology model to obtain the equipment matching result; when the equipment matching result corresponds to multiple equipment objects, or when the candidate text has ambiguity in equipment reference, the corresponding candidate transcribed segment is determined as the segment to be confirmed.

[0010] In a preferred embodiment, reserving multiple candidate device objects for the segment to be confirmed includes: recalling matching device objects from the device ledger and distribution network topology model based on the candidate text of the segment to be confirmed, forming an initial set of device objects; filtering the initial set of device objects based on the contextual semantics of the segment to be confirmed and the current scheduling scenario; and determining the filtered and retained device objects as candidate device objects.

[0011] In a preferred embodiment, the step of parsing subsequent associated candidate transcribing segments based on the timestamp of the segment to be confirmed, and extracting topological relationship information for defining candidate device objects, includes: determining the starting point of the time sequence retrieval based on the timestamp of the segment to be confirmed, obtaining subsequent candidate transcribing segments within a preset association window; filtering associated candidate transcribing segments from the candidate transcribing segments that belong to the same operation statement or the same scheduling task as the segment to be confirmed; parsing the topological relationship semantics contained in the associated candidate transcribing segments, and generating topological relationship information, wherein the topological relationship information includes at least one of the following: power supply direction relationship between nodes, path connectivity relationship, connection switching relationship, power supply boundary relationship, and power supply range relationship.

[0012] In a preferred embodiment, the step of backtracking constraints on multiple candidate device objects to determine the target device object includes: generating corresponding topology constraints based on topology relationship information; matching the topology relationships of each candidate device object in the distribution network topology model with the topology constraints to obtain the topology matching results of each candidate device object; determining the topology support status of each candidate device object based on the topology matching results, and combining the speech recognition confidence score to determine the target device object from multiple candidate device objects.

[0013] In a preferred embodiment, determining the target device object from multiple candidate device objects includes: when a candidate device object with high acoustic confidence conflicts with topological relationship information, and other candidate device objects satisfy the topological relationship information, the candidate device object satisfying the topological relationship information is taken as the priority candidate; when multiple candidate device objects satisfy the topological relationship information, the target device object is determined based on the degree of topological matching, speech recognition confidence, and device ledger matching; when the target device object cannot be uniquely determined, the pending confirmation status is retained and a review prompt is output.

[0014] In a preferred embodiment, the step of backfilling the target device object to the corresponding timestamp position to generate the transcribed text of the distribution network dispatch voice includes: determining the backfill timestamp of the target device object according to the correspondence between the target device object and the segment to be confirmed; replacing or completing the standard device information of the target device object to the segment to be confirmed corresponding to the backfill timestamp; generating transcribed text according to the chronological order of the candidate transcribed segments, and associating and saving the backtracking constraint result with the transcribed text; when the target device object cannot be uniquely determined, retaining the segment to be confirmed at the corresponding timestamp position, and outputting the candidate device object and its topology matching result.

[0015] In a preferred embodiment, determining the backfill timestamp of the target device object based on the correspondence between the target device object and the segment to be confirmed includes: determining the segment to be confirmed that has a matching relationship with the source candidate device object corresponding to the target device object, and determining the timestamp of the segment to be confirmed as the backfill timestamp; when the target device object is determined by multiple candidate transcription segments, the segment to be confirmed that directly expresses the device object is determined as the backfill segment, and the associated candidate transcription segment used for topological constraints is retained as the basis for backtracking.

[0016] The technical effects and advantages of the present invention's method for accurate voice transcription in distribution network scheduling based on topological constraints are as follows: This invention retains multiple candidate devices for preceding equipment during voice transcription in distribution network dispatching and uses subsequent topology information for backtracking and confirmation, avoiding the loss of correct candidates due to premature device name determination. This reduces the risk of mistranscription caused by similar device names, partial numbers, or abbreviations, minimizes the problem of valid device names but mismatched topology relationships, and improves the accuracy of dispatching records, instruction parsing, and operation ticket generation. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the method for accurate voice transcription of distribution network scheduling based on topological constraints according to the present invention. Figure 2 This is a schematic diagram of the topological relationship information constraint results; Figure 3 This is a diagram illustrating the comparison of miswrite rates for device objects. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1, Figure 1 This invention presents a method for accurate voice transcription of distribution network scheduling based on topological constraints, comprising the following steps: S1. Obtain the distribution network dispatch voice and distribution network topology model, segment the distribution network dispatch voice, and generate candidate transcription segments with timestamps. In this embodiment, the distribution network topology model refers to a data model generated based on the distribution network equipment ledger, feeder wiring relationships, switch on / off status, and dispatch operation mode, used to characterize the connection relationships and power supply relationships of distribution network equipment. The distribution network topology model includes at least one of the following: distribution network equipment nodes, equipment connection relationships, equipment standard names, equipment aliases, associated feeders, switch on / off status, upstream and downstream relationships, and interconnection relationships. This model is used in subsequent steps to determine whether candidate equipment objects conform to the topology relationship information in the dispatch voice.

[0020] The candidate transcription segment refers to an intermediate transcription unit obtained after recognizing continuous speech segments in the distribution network dispatch voice. It is not the final confirmed transcribed text, but includes at least a segment timestamp, candidate text content, candidate confidence level, and segment type label. The timestamp is used to indicate the position of the candidate transcription segment in the distribution network dispatch voice, and the segment type label is used to distinguish between ordinary speech segments, operation word segments, equipment object segments, directional relationship segments, range boundary segments, or transfer relationship segments.

[0021] Unlike existing speech-to-text methods that directly output the text with the highest acoustic confidence, this embodiment retains device-related and topology-related segments that may affect the determination of scheduling objects during the segmentation and recognition stage. For candidate texts that are acoustically similar but may correspond to different distribution network equipment objects, they are not immediately uniquely confirmed, but are retained as candidate transcription segments with timestamps, so that subsequent back-tracking constraints on previous candidate results can be applied in conjunction with upstream and downstream relationships, communication relationships, transfer directions, or isolation boundaries.

[0022] The specific method for obtaining the distribution network topology model is as follows: Obtain information on distribution network equipment ledgers, feeder wiring relationships, and switch status; Based on the distribution network equipment ledger, determine the distribution network equipment nodes and their corresponding standard names, aliases, types, and feeders; Determine the device connection relationships between distribution network equipment nodes based on feeder wiring relationships; The conduction status of device connections is determined based on switch status information; Based on the device connection relationships and their conduction status, determine the upstream and downstream relationships and interconnection relationships under the current operating mode, and generate a distribution network topology model.

[0023] The step of segmenting and recognizing the voice dispatching data to generate candidate transcription segments with timestamps includes: The acquired distribution network dispatch voice is preprocessed to obtain the voice stream to be identified; the preprocessing includes at least one of noise reduction, endpoint detection, silence removal and volume normalization; An initial speech segment is determined based on pause locations, speech rate variations, and speech energy variations in the speech stream to be identified; the initial speech segment is used to limit the basic time range for candidate transcription. Perform scheduling script recognition on the initial voice segment to identify operation trigger words, suspected device name words, suspected number words, directional relation words, and range boundary words; When there is an operation object dependency or topological dependency between adjacent initial speech segments, the adjacent initial speech segments are merged into the same recognition window; For each recognition window, perform multi-candidate speech recognition to generate multiple candidate text results, and record the speech start and end time and confidence level corresponding to each candidate text result; Based on the word types in the candidate text results, the candidate text results are divided into ordinary speech fragments, operation word fragments, equipment object fragments, directional relationship fragments, range boundary fragments, or transfer relationship fragments; When the candidate text results involve distribution network equipment or equipment numbers, the candidate text with the highest acoustic confidence is taken as the provisional candidate text, and other candidate texts that are similar in pronunciation, text, or equipment ledger are taken as reserved candidate texts, generating candidate transcription fragments with timestamps.

[0024] In this embodiment, the acoustic confidence score can be obtained by using the candidate text confidence score output by an existing speech recognition model, or by normalizing the acoustic model score and language model score corresponding to the candidate text. This invention does not limit the specific calculation method of the acoustic confidence score; the acoustic confidence score is only used to represent the degree of matching between the candidate text and the speech segment to be recognized.

[0025] The process of determining the initial speech segment based on pause positions, speech rate changes, and speech energy changes in the speech stream to be identified includes: Frame-level analysis is performed on the speech stream to be recognized to obtain the short-time energy, zero-crossing rate and speech presence probability of each speech frame, and the speech interval and non-speech interval are determined based on the speech presence probability. Non-speech intervals whose duration exceeds the first pause threshold are identified as candidate pause locations; Calculate the speech rate change within adjacent speech intervals; the speech rate change is determined based on the number of syllables, words, or acoustic units identified per unit time; when the speech rate change of adjacent speech intervals exceeds the speech rate change threshold, the corresponding position is determined as a candidate speech rate boundary position. Calculate the average short-time energy difference and energy drop duration between adjacent speech intervals; when the average short-time energy difference exceeds the energy change threshold and the energy drop duration exceeds the second pause threshold, the corresponding position is determined as a candidate energy boundary position. The candidate pause positions, candidate speech rate boundary positions, and candidate energy boundary positions are fused and judged to generate candidate segmentation positions; The continuity of the voice content before and after the candidate segmentation position is determined. When the candidate segmentation position is located inside the equipment name, line name, switch number, tower number, ring network cabinet number or operation phrase, the candidate segmentation position is cancelled. The candidate segmentation positions that have not been cancelled are used as the initial segmentation boundaries, and the speech stream to be recognized is divided into multiple initial speech segments according to the initial segmentation boundaries.

[0026] The specific formula for calculating the change in speech rate is as follows:

[0027]

[0028] in, Let be the change in speech rate between the i-th speech interval and the (i-1)-th speech interval. Let i be the speech rate of the i-th speech interval. Let i be the speech rate of the (i-1)th speech interval. This represents the number of syllables, words, or acoustic units identified within the i-th speech interval. Let be the end time of the i-th speech interval. Let be the start time of the i-th speech interval.

[0029] The average short-time energy difference is calculated using the following formula:

[0030]

[0031] in, This represents the average short-time energy difference. Let be the average short-time energy of the i-th speech interval. Let be the average short-time energy of the (i-1)th speech interval. Let be the number of speech frames contained in the i-th speech interval. Let be the short-time energy of the m-th frame in the i-th speech interval.

[0032] The process of performing fusion judgment and generating candidate segmentation positions is as follows:

[0033] in, Let be the segmentation support value for the i-th candidate position. This is an indicator function that takes the value 1 when the condition within the parentheses is true, and 0 otherwise. The non-speech duration corresponding to the i-th candidate position. The first pause threshold, The change in speech rate The threshold for speech rate variation. This represents the average short-time energy difference. The energy change threshold, The duration of energy decrease, This is the second pause threshold.

[0034] When satisfied If the candidate segmentation position is not located within the equipment name, line name, switch number, tower number, ring main unit number, or operation phrase, the i-th candidate position is determined as the initial segmentation boundary.

[0035] In this embodiment, the first pause threshold, speech rate change threshold, energy change threshold, and second pause threshold are not fixed constants, but are pre-calibrated based on historical dispatch speech samples of the target distribution network dispatch scenario. Specifically, a preset number of historical distribution network dispatch speech samples are obtained, and the actual semantic segmentation positions and non-semantic segmentation positions are labeled; the non-speech duration, speech rate change, average short-term energy difference, and energy decrease duration corresponding to the labeled positions are extracted respectively; and the corresponding thresholds are determined based on the distribution differences between the feature values ​​corresponding to the actual semantic segmentation positions and the feature values ​​corresponding to the non-semantic segmentation positions.

[0036] The first pause threshold is used to distinguish between short pauses during the scheduler's continuous expression and pauses between different scheduling semantics; the speech rate change threshold is used to identify sudden changes in speech rate during scheduling statement transitions; the energy change threshold is used to identify candidate boundaries formed by changes in speech expression intensity; and the second pause threshold is used to limit the duration of energy decrease to avoid misjudging instantaneous energy fluctuations as semantic segmentation locations.

[0037] In this embodiment, the continuity judgment of the scheduling object is used to avoid incorrectly segmenting objects with complete scheduling meaning. When the voice content before and after the candidate segmentation position corresponds to the line name and switch number, or to the equipment number and equipment type word, respectively, it is determined that the candidate segmentation position is within the same scheduling object and is not used as the initial segmentation boundary. This avoids the common silent segmentation method from splitting continuous scheduling objects such as "Chengbei Line No. 3 Ring Network Cabinet Switch," "Load Below a Certain Switch," and "Close the Connecting Switch" into independent segments.

[0038] S2, identify the segments to be confirmed from the candidate transcribed segments that involve distribution network equipment or equipment numbers, and reserve multiple candidate equipment objects for the segments to be confirmed; In this embodiment, the segment to be confirmed refers to a candidate transcription segment that contains the name, abbreviation, number, or location description of a distribution network device, and whose corresponding device cannot be uniquely identified based solely on the current voice segment. The segment to be confirmed is not the final transcribed text, but rather an intermediate segment that requires further confirmation by combining topological relationship information from subsequent voice recordings.

[0039] The candidate equipment objects refer to real equipment entities matched from the equipment ledger and distribution network topology model based on the candidate text content of the fragment to be confirmed. The candidate equipment objects include at least one of the following: lines, switches, ring main units, branch boxes, distribution transformers, transformer substations, poles, tie points, or feeder outgoing line bays. Each candidate equipment object must have at least an equipment identifier, a standard equipment name, an equipment type, its associated feeder, and its topology node location.

[0040] It should be noted that candidate equipment objects are different from candidate text. Candidate text is the text result obtained from speech recognition, such as "Switch No. 3" or "Switch No. 5"; candidate equipment objects are specific equipment entities in the distribution network ledger or topology model, such as "Switch No. 3 on the 10kV Chengbei Line", "Switch No. 3 on the 10kV Chengnan Line", or "Switch No. 5 on the 10kV Chengbei Line".

[0041] The step of determining the segments to be confirmed from the candidate transcription segments that involve distribution network equipment or equipment numbers includes: The candidate text content in the candidate transcription fragments is subjected to distribution network object word recognition, and the equipment type words, line name words, station name words, number words, location description words and range indicator words are extracted; Combine equipment type terms, line name terms, station name terms, number terms, location description terms, and range indicator terms to generate information on the objects to be matched; Based on the information of the object to be matched, a matching is performed in the equipment ledger and the distribution network topology model to obtain the equipment matching result corresponding to the information of the object to be matched. When there are multiple device objects in the device matching result, or when the device objects in the device matching result need to rely on subsequent topological relationship information to be uniquely identified, the corresponding candidate transcription fragments are identified as fragments to be confirmed. When a candidate transcription fragment does not fully contain the standard name of the device, but contains the device abbreviation, local number, upstream and downstream description, connection description or range description, and can form a corresponding relationship with multiple device objects in the distribution network topology model, the candidate transcription fragment is identified as a fragment to be confirmed.

[0042] The method of reserving multiple candidate device objects for the segment to be confirmed includes: Based on the candidate text content corresponding to the segment to be confirmed, obtain the device name that is similar in pronunciation, text, or number to the candidate text content; The device names are matched with the standard device names, device aliases, device abbreviations, and historical scheduling names in the device ledger to obtain an initial set of device objects; Based on the context fragment of the segment to be confirmed, the initial set of device objects is filtered by scene to remove device objects that are obviously unrelated to the current scheduling area, the current feeder, or the current scheduling task; wherein, the scene filtering is performed based on at least one of the following: current scheduling area, call-related feeder, scheduling task type, previously confirmed device objects, and connectivity relationships in the distribution network topology model; The filtered device objects are used as candidate device objects, and the matching source, matching confidence, and timestamp relationship between each candidate device object and the fragment to be confirmed are retained.

[0043] The confidence level of the match between the candidate device and the fragment to be confirmed is expressed as follows:

[0044] in, To match the confidence level, Let be the speech similarity between the device name corresponding to the k-th candidate device and the segment to be confirmed. For text similarity, To ensure the accuracy of equipment inventory matching, To determine the degree of contextual matching, This is the normalized value.

[0045] S3, using the timestamp of the segment to be confirmed as a basis, parse the associated candidate transcribing segments that follow its time sequence, and extract the topological relationship information used to limit the candidate device objects; In this embodiment, the subsequent associated candidate transcription segment refers to a candidate transcription segment located after the segment to be confirmed, and which is associated with the segment to be confirmed in the same scheduling statement, the same operation intention, or the same scheduling task. The subsequent associated candidate transcription segment is not all the speech content after the segment to be confirmed, but rather a subsequent segment that can limit the device object corresponding to the segment to be confirmed.

[0046] The topology relationship information refers to information extracted from subsequent associated candidate transcription fragments that can be used to determine whether candidate device objects conform to the current distribution network topology. The topology relationship information includes at least one of the following: upstream and downstream direction clues, power supply path clues, interconnection relationship clues, isolation boundary clues, power outage and restoration range clues, and power transfer target clues.

[0047] The process of parsing subsequent candidate transcribing segments based on the timestamp of the segment to be confirmed, and extracting topological relationship information to limit candidate device objects, includes: Determine the starting point for subsequent searches based on the end time or segment number of the segment to be confirmed; Starting from the subsequent retrieval starting point, candidate transcription fragments are obtained within a preset association window; the preset association window is determined by time length, number of fragments, or operation statement boundaries; Identify the fragment type of the candidate transcription fragments within the preset association window, and extract the operation words, direction words, range words, contact words, and target power supply path words; When the operation word, direction word, range word, contact word or target power supply path word in the candidate transcription fragment belongs to the same operation statement or the same scheduling task as the fragment to be confirmed, the candidate transcription fragment is identified as the subsequent associated candidate transcription fragment. Based on the operation words, direction words, range words, connection words and target power supply path words in the subsequent associated candidate transcription fragments, generate topological relationship information related to the candidate device object.

[0048] In this embodiment, whether the candidate transcription fragment and the fragment to be confirmed belong to the same operation statement or the same scheduling task is determined based on at least one of the following conditions: No new independent operation initiation words were found between the candidate transcription fragment and the fragment to be confirmed; The candidate transcribed fragment contains pronouns such as "the switch," "its downstream," "the aforementioned line," and "the branch." The candidate transcribed fragment contains directional or range terms related to the candidate device object corresponding to the fragment to be confirmed; The candidate transcription fragment contains another device object that can form a power supply path, communication relationship or isolation boundary with the candidate device object; The time interval between the candidate transcription fragment and the fragment to be confirmed did not exceed the preset association time threshold.

[0049] S4, based on topology relationship information and distribution network topology model, performs backtracking constraints on multiple candidate device objects to determine the target device object; In this embodiment, the backtracking constraint refers to, after extracting topology relationship information following the segment to be confirmed, applying that topology relationship information in reverse to multiple candidate device objects corresponding to the segment to be confirmed, in order to determine whether each candidate device object can satisfy the distribution network connection relationship, power supply direction relationship, interconnection relationship, or range boundary relationship defined by the topology relationship information. The backtracking constraint is not a post-hoc verification of the final transcribed text, but rather a screening, elimination, and rearrangement of multiple candidate device objects before the device objects for the segment to be confirmed are uniquely determined.

[0050] The step of applying backtracking constraints to multiple candidate device objects to determine the target device object includes: Based on the type of topology relationship information, the corresponding topology constraints are determined; wherein, the topology constraints include at least one of upstream and downstream constraints, co-feeder constraints, tie path constraints, isolation boundary constraints, power outage and restoration range constraints, and power transfer target constraints. Starting with each candidate device object, query the topology location, feeder, adjacent devices, switch status, upstream and downstream range, and available power supply path of the candidate device object in the distribution network topology model; The topological location, feeder, adjacent devices, switch status, upstream and downstream range, and power supply path of each candidate device are matched with the topological constraints corresponding to the topological relationship information to generate the topological matching degree of each candidate device. A topology support flag is generated when a candidate device object satisfies the topology constraints corresponding to the topology relationship information; a topology conflict flag is generated when a candidate device object does not satisfy the topology constraints corresponding to the topology relationship information. The target device is determined from multiple candidate device objects based on the topology support flag, topology conflict flag, and corresponding speech recognition confidence score of each candidate device object.

[0051] The step of determining the target device object from multiple candidate device objects based on the topology support marker, topology conflict marker, and corresponding speech recognition confidence score of each candidate device object includes: When the candidate device with the highest acoustic confidence has a topology conflict flag, and only one candidate device with a lower acoustic confidence has a topology support flag, the candidate device with the topology support flag is identified as the target device. When multiple candidate device objects have topology support tags, the target device object is determined based on the degree of matching between the candidate device object and the topology relationship information, the acoustic confidence level, and the degree of matching between the device ledger and the candidate device object. When all candidate device objects have topology conflict markers, retain the candidate status of the fragment to be confirmed and output a manual review prompt; When the topology information is insufficient to distinguish multiple candidate device objects, supplementary topology information is obtained within the preset extended window and backtracking constraints are performed again; if the target device object still cannot be uniquely determined within the preset extended window, the candidate status of the segment to be confirmed is retained and a manual review prompt is output.

[0052] For example, if the candidate text corresponding to the segment to be confirmed is "Switch No. 3", the first candidate equipment object "Chengbei Line No. 3 Section Switch" and the second candidate equipment object "Chengnan Line No. 3 Interconnect Switch" are obtained based on the speech recognition results and equipment ledger matching. If the subsequent associated candidate transcribed segments contain topological relationship information such as "its downstream load is transferred to Chengdong Line", the downstream range and transferable paths of the first and second candidate equipment objects are queried in the distribution network topology model. If the first candidate equipment object does not have a transferable path to Chengdong Line, but the second candidate equipment object has an interconnection transfer relationship with Chengdong Line, then even if the acoustic confidence of the first candidate equipment object is high, the second candidate equipment object will be identified as the target equipment object.

[0053] The specific formula for calculating the degree of matching between the candidate device object and the topology relationship information is as follows:

[0054] in, The degree of topology matching for the k-th candidate device object. This is the set of topological constraints corresponding to the topological relationship information. Set of topological constraints Any topological constraint in the, Let q be the weight corresponding to the topological constraint condition q. This represents the result of the k-th candidate device object satisfying the topological constraint q; it is 1 if satisfied and 0 if not.

[0055] when At that time, generate topology support tags; when At that time, generate topology conflict markers; when This generates topologically uncertain markers. Topology support threshold; This is the topology conflict threshold, set based on historical data.

[0056] Compared with the prior art, this embodiment does not perform device name verification after the speech transcription result is generated. Instead, it retains multiple candidate device objects in the stage of unconfirmed segments and uses the topological relationship information that appears later to back-constrain the preceding candidate device objects. This allows the upstream and downstream ranges, communication relationships, transfer paths, or isolation boundaries in the subsequent speech to participate in the determination of the preceding device objects, thereby reducing erroneous transcription results that are acoustically similar but have no valid topological relationship.

[0057] S5, fill the target device object back into the corresponding timestamp position, and generate the transcribed text of the distribution network dispatch voice.

[0058] In this embodiment, backfilling refers to replacing or supplementing the target device object determined by backtracking constraints with the original timestamp of the segment to be confirmed in the distribution network dispatch voice, so that the generated transcribed text maintains the expression order of the original dispatch voice and uses the standard device name or dispatch specification name corresponding to the target device object.

[0059] The step of backfilling the target device object to the corresponding timestamp position and generating the transcribed text of the distribution network dispatch voice includes: Based on the correspondence between the target device object and the segment to be confirmed, determine the segment timestamp that the target device object needs to backfill; Obtain the standard equipment name, equipment type, feeder to which it belongs, and equipment number corresponding to the target equipment object; Determine whether the candidate text content of the segment to be confirmed is a complete device name. If so, replace the candidate text content of the segment to be confirmed with the standard device name corresponding to the target device object. If the candidate text content of the segment to be confirmed is a device abbreviation, local number, pronoun, or scope expression, then the segment to be confirmed is completed according to the standard device name corresponding to the target device object; According to the chronological order of the candidate transcription segments, the standard language segments, operation word segments, backfilled equipment object segments, directional relationship segments, range boundary segments, and transfer relationship segments are spliced ​​together to generate the transcribed text of the distribution network dispatch voice. The backtracking constraint results corresponding to the target device object are associated with and saved with the transcribed text. The backtracking constraint results include at least one of the following: the determination source of the target device object, topological relationship information, and topological matching degree. When the target device object cannot be uniquely determined, the pending confirmation identifier is retained at the corresponding timestamp, and multiple candidate device objects and their topology matching degree are output. The candidate device object with the highest acoustic confidence is not directly used as the final backfill result.

[0060] The step of determining the timestamp of the segment that needs to be filled back for the target device object based on the correspondence between the target device object and the segment to be confirmed includes: Obtain the source candidate device object corresponding to the target device object; Based on the pre-saved matching relationship between the source candidate device object and the segment to be confirmed, determine the segment to be confirmed corresponding to the target device object; Read the segment number, segment start time, and segment end time of the segment to be confirmed, and determine them as the backfill timestamp of the target device object; When the target device object is determined by a segment to be confirmed, the timestamp of the segment to be confirmed is used as the backfill timestamp. When the target device object is determined by multiple unconfirmed segments and subsequent topology relationship information, the unconfirmed segments that directly contain the device name, device abbreviation, device number, pronoun, or scope expression are determined as backfill segments, and the segments containing the subsequent topology relationship information are retained as constraint basis and are not used as backfill positions for the target device object. When multiple unconfirmed segments may be used as backfill segments and cannot be uniquely determined, the unconfirmed identifier is retained at the corresponding segment position, and the candidate device object and its topology matching degree are output.

[0061] In this embodiment, when the segment to be confirmed already contains a replaceable candidate device name, a replacement method is used for backfilling. For example, if the candidate text of the segment to be confirmed is "Chengbei Line No. 3 Switch", and the target device object is determined to be "10kV Chengbei Line No. 3 Section Switch" through backtracking constraints, the original candidate text is replaced with the standard device name.

[0062] When the segment to be confirmed only contains a partial number, abbreviation, or pronoun, it is filled in using a completion method. For example, if the segment to be confirmed is "Switch No. 3" or "this switch", after determining the target equipment object through backtracking constraints, it is completed as "10kV Chengbei Line No. 3 Section Switch". In this way, the generated transcribed text can avoid the problem of subsequent systems being unable to accurately identify equipment objects due to only retaining the dispatcher's colloquial expressions.

[0063] After generating the transcribed text of the distribution network dispatch speech, the following steps are also included: Extract target device object, operation type, and topology information from the transcribed text; Generate consistency check items based on the type of topology relationship information; Based on the distribution network topology model, obtain the feeder, upstream and downstream range, connection path, isolation boundary and power supply path corresponding to the target equipment object; The topology information corresponding to the target device object is compared with the consistency check item to generate a transcribing consistency result; the transcribing consistency result includes one of the following: a consistent result, an inconsistent result, and an uncertain result; When the target device object satisfies the corresponding consistency check item, a consistent result is generated; When the target device object does not meet the corresponding consistency check item, an inconsistent result is generated; When the distribution network topology model lacks the switch status, connection relationship or device connection relationship required for judgment, or when multiple candidate device objects all meet the consistency check items and cannot be uniquely determined, an uncertain result is generated; When inconsistent or uncertain results are generated, a review identifier is generated at the timestamp corresponding to the target device object, and the target device object, original candidate text, other candidate device objects, topology relationship information and reasons for inconsistency are output. When a consistent result is generated, the transcribed text, target device object, topology relationship information, and consistent result are associated and saved as input for scheduling records, scheduling instruction parsing, or operation ticket generation.

[0064] Through the above methods, the transcribed text generated in this embodiment not only maintains the original expression order of the distribution network dispatch voice, but also replaces or completes key equipment objects with standard equipment objects confirmed by topological relationship constraints. Compared with voice transcription methods that only output text based on acoustic confidence, this embodiment can reduce the problem of miswriting objects caused by equipment abbreviations, partial numbers, or similar equipment names, and improve the accuracy of subsequent dispatch record archiving, dispatch instruction parsing, and operation ticket generation.

[0065] Example 2, Figure 2 This demonstrates the process of backtracking constraints using subsequent topological relationship clues when multiple candidate device objects correspond to a segment to be confirmed. The diagram uses "downstream range, transfer target, and connection path" as topological constraints to match candidate device objects C01, C02, and C03. Although C01 has the highest acoustic confidence score of 0.87, it fails to match any of the downstream range, transfer target, and connection path constraints, resulting in a topological match score of only 0.25, and is therefore judged as a conflict. C02 has an acoustic confidence score of 0.73, lower than C01, but it matches all three topological constraints, resulting in a topological match score of 1.00, and is therefore identified as the target device object. C03 also fails to match any topological constraints, resulting in a topological match score of 0.00, and is therefore eliminated. This figure illustrates that the present invention does not simply determine the device name based on acoustic confidence, but rather, based on retaining multiple candidate device objects, uses subsequent topological relationship clues to back-constrain previous candidates, thereby being able to identify and correct erroneous transcription results with high acoustic confidence but invalid topological relationships.

[0066] Figure 3 The figure shows a comparison of the error transcription rates between traditional transcription verification methods and the backtracking constraint method of this invention in different distribution network dispatching voice scenarios. As can be seen from the figure, the error transcription rate of the method of this invention is lower than that of the traditional method in scenarios involving similar device names, local numbers or abbreviations, upstream and downstream ranges, communication and transfer paths, and isolation boundary expressions. Specifically, in the communication and transfer path scenario, the error transcription rate decreased from 15.0% to 3.8%; in the local number or abbreviation scenario, the error transcription rate decreased from 12.3% to 3.8%; and in the similar device name scenario, the error transcription rate decreased from 14.2% to 5.0%. This figure illustrates that by retaining candidate device objects during the speech recognition stage and combining them with subsequent topological relationship information for backtracking confirmation, this invention can effectively reduce the risk of error transcription caused by device names, device numbers, abbreviations, and topological constraint expressions, and improve the consistency between the transcribed text of distribution network dispatching voice and the actual topological relationship.

[0067] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0068] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0069] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0070] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0071] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0072] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for accurate voice transcription of distribution network scheduling based on topological constraints, characterized in that, Includes the following steps: Obtain the distribution network dispatch voice and distribution network topology model, segment the distribution network dispatch voice, and generate candidate transcription segments with timestamps; Identify the segments to be confirmed from the candidate transcription segments that involve distribution network equipment or equipment numbers, and reserve multiple candidate equipment objects for the segments to be confirmed; Based on the timestamp of the segment to be confirmed, the associated candidate transcribing segments following its time sequence are parsed to extract the topological relationship information used to limit the candidate device objects; Based on topological relationship information and distribution network topology model, backtracking constraints are applied to multiple candidate device objects to determine the target device object; The target device object is backfilled to the corresponding timestamp position to generate the transcribed text of the distribution network dispatch voice.

2. The method for accurate voice transcription of distribution network scheduling based on topological constraints according to claim 1, characterized in that, The step of segmenting and recognizing the voice dispatching data to generate candidate transcription segments with timestamps includes: Preprocess the voice of the power distribution network dispatch to obtain the voice stream to be recognized; The initial speech segments are determined based on the speech stream to be recognized, and the relevant initial speech segments are merged into a recognition window according to the continuity of the scheduled speech. Perform multi-candidate recognition on the recognition window, retain the candidate recognition results containing device identifiers or numbers, and generate candidate transcription fragments.

3. The method for accurate voice transcription of distribution network scheduling based on topological constraints according to claim 2, characterized in that, The process of determining the initial speech segment based on the speech stream to be identified includes: Candidate segmentation positions are determined based on the speech features of the speech stream to be identified, wherein the speech features include at least one of pause features, speech rate variation features, and energy variation features; Determine whether the candidate segmentation position has disrupted the expression of scheduling operation terms or device names. If not, determine the candidate segmentation position as the initial segmentation boundary and obtain multiple initial speech segments based on the initial segmentation boundary.

4. The method for accurate voice transcription of distribution network scheduling based on topological constraints according to claim 3, characterized in that, The step of determining the segments to be confirmed from the candidate transcription segments that involve distribution network equipment or equipment numbers includes: Perform distribution network device entity recognition on candidate text in candidate transcription fragments and extract information of the device to be matched; The device information to be matched is matched with the device ledger and the distribution network topology model to obtain the device matching results; When the device matching result corresponds to multiple device objects, or when the candidate text contains ambiguity in device reference, the corresponding candidate transcription fragment will be determined as the fragment to be confirmed.

5. The method for accurate voice transcription of distribution network scheduling based on topological constraints according to claim 4, characterized in that, The method of reserving multiple candidate device objects for the segment to be confirmed includes: Based on the candidate text of the fragment to be confirmed, the device objects that match it are recalled from the device ledger and the distribution network topology model to form an initial set of device objects; Based on the contextual semantics of the fragments to be confirmed and the current scheduling scenario, the initial set of device objects is filtered; The device objects retained after filtering are identified as candidate device objects.

6. The method for accurate voice transcription of distribution network scheduling based on topological constraints according to claim 5, characterized in that, The process of parsing subsequent candidate transcribing segments based on the timestamp of the segment to be confirmed, and extracting topological relationship information to limit candidate device objects, includes: The starting point of the time-series retrieval is determined by the timestamp of the segment to be confirmed, and subsequent candidate transcribed segments are obtained within the preset association window; Filter candidate transcription fragments from the candidate transcription fragments that belong to the same operation statement or the same scheduling task as the fragment to be confirmed; The topological relationship semantics contained in the associated candidate transcription fragments are parsed to generate topological relationship information, which includes at least one of the following: power supply direction relationship, path connectivity relationship, connection switching relationship, power supply boundary relationship, and power supply range relationship between nodes.

7. The method for accurate voice transcription of distribution network scheduling based on topological constraints according to claim 6, characterized in that, The step of applying backtracking constraints to multiple candidate device objects to determine the target device object includes: Generate corresponding topological constraints based on topological relationship information; The topological relationships of each candidate device in the distribution network topology model are matched with the topological constraints to obtain the topology matching results of each candidate device. The topology support status of each candidate device is determined based on the topology matching results, and the target device is determined from multiple candidate device objects by combining the speech recognition confidence score.

8. The method for accurate voice transcription of distribution network scheduling based on topological constraints according to claim 7, characterized in that, The step of determining the target device object from multiple candidate device objects includes: When a candidate device with high acoustic confidence conflicts with topology information, and other candidate device objects satisfy the topology information, the candidate device object that satisfies the topology information will be the preferred candidate. When multiple candidate device objects all satisfy the topological relationship information, the target device object is determined based on the degree of topological matching, speech recognition confidence, and device ledger matching. When the target device cannot be uniquely identified, retain the pending confirmation status and output a verification prompt.

9. The method for accurate voice transcription of distribution network scheduling based on topological constraints according to claim 8, characterized in that, The step of backfilling the target device object to the corresponding timestamp position and generating the transcribed text of the distribution network dispatch voice includes: Based on the correspondence between the target device object and the segment to be confirmed, determine the backfill timestamp of the target device object; Replace or complete the standard equipment information of the target equipment object into the unconfirmed segment corresponding to the backfill timestamp; The transcribed text is generated according to the temporal arrangement of the candidate transcribed fragments, and the backtracking constraint results are associated with and saved with the transcribed text. When the target device object cannot be uniquely identified, a confirmation flag is retained at the corresponding timestamp position, and the candidate device object and its topology matching result are output.

10. The method for accurate voice transcription of distribution network scheduling based on topological constraints according to claim 9, characterized in that, The step of determining the backfill timestamp of the target device object based on the correspondence between the target device object and the segment to be confirmed includes: Based on the source candidate device object corresponding to the target device object, determine the segment to be confirmed that is matched with it, and determine the timestamp of the segment to be confirmed as the backfill timestamp. When the target device object is determined by multiple candidate transcription fragments, the fragment to be confirmed that directly expresses the device object is determined as the backfill fragment, and the associated candidate transcription fragment used for topological constraints is retained as the basis for backtracking.

Images