A power failure intelligent analysis method and device

Abstract

The method comprises the following steps: removing the fault text data to obtain a fault text data list, wherein the fault text data list comprises fault reasons and treatment methods; segmenting the fault text data according to the text length of the fault reasons, combining the first segment data with a large model to generate a classification framework, and according to the text similarity between each new fault reason in the subsequent segment data and the existing fault reasons, the new fault reason is classified into an existing fault category or a new fault category is generated to obtain a classification result data set; according to the fault occurrence frequency of each fault category in the classification result data set, the high-frequency screening of the fault reasons and the treatment methods is performed; and the fault category names in the screened classification result data set are combined to generate a fault maintenance report. The problems that the classification framework cannot be dynamically adjusted, the classification precision is low, and the classification efficiency is low are effectively solved, and the accuracy and efficiency of fault analysis are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power failure analysis, in particular to a power failure intelligent analysis method and device. BACKGROUND

[0002] In the field of power equipment failure analysis and maintenance management, the existing technology mainly focuses on the classification, statistics and report generation of failure records, and has experienced a development process from manual dominance to semi-automation and automation. Specifically, the development of technology can include the following stages:

[0003] The traditional manual classification and statistical method relies on the experience and judgment of professionals, and classifies and counts failure records manually, which is low in efficiency, strong in subjectivity, and easy to lead to non-uniform classification standards and statistical errors.

[0004] The automatic classification method based on pre-defined rules classifies failure records automatically through rules pre-set by humans, such as the correspondence between fault codes and fault types. Although it improves the classification efficiency, the classification framework is rigid and lacks adaptability.

[0005] The classification method based on traditional machine learning uses machine learning algorithms such as decision trees and support vector machines to extract features and train models for failure records to achieve automatic fault classification. To some extent, it reduces the dependence on manual rules, but the data labeling cost of machine learning model training is high, the generalization ability is limited, which leads to a decrease in classification accuracy, and the inability to handle multi-value fields leads to information loss.

[0006] The static classification method based on early large models uses the text understanding ability of large models to classify all failure records at once without the need for manual feature extraction, and can understand complex semantics. However, when the data volume is too large, the attention allocation for long text is imbalanced, leading to classification bias, and the classification framework cannot be dynamically adjusted. Secondly, the current maintenance report generation technology still relies on manual analysis of core content, which is low in report generation efficiency and prone to content bias. SUMMARY

[0007] The present application provides a power failure intelligent analysis method and device, which can solve the technical problems of existing large model failure analysis technology, such as the inability of the classification framework to dynamically adjust, low classification accuracy, and low classification efficiency.

[0008] To achieve the above-mentioned purpose, in a first aspect, the present application provides a power failure intelligent analysis method, which comprises:

[0009] De-duplicate the failure text data to obtain a failure text data list, wherein the failure text data list includes failure causes and treatment methods.

[0010] According to the text length of the fault reason, the fault text data is segmented, the first segment data is combined with the large model to generate a classification framework, according to the text similarity of each new fault reason in the subsequent segment data and the existing fault reason, the new fault reason is classified into the existing fault category or a new fault category is generated, and a classification result data set is obtained.

[0011] According to the fault occurrence frequency of each fault category in the classification result data set, the fault reason and the processing method are high-frequency filtered.

[0012] The fault category names in the filtered classification result data set are merged to generate a fault maintenance report.

[0013] Further, in an embodiment, the fault reason and the processing method are obtained by semantic extraction of the fault text data without deduplication through a large model.

[0014] If the number of fault reasons and processing methods in the fault text data list is inconsistent, the number is aligned through a list completion mechanism.

[0015] Further, in an embodiment, the deduplication of the fault text data comprises:

[0016] The fault reason character similarity and the processing method character similarity are calculated according to the string edit distance similarity formula, and the fault reason semantic similarity and the processing method semantic similarity are calculated according to the cosine similarity formula.

[0017] The fault reason text similarity is obtained by taking the mean of the fault reason character similarity and the fault reason semantic similarity, and the processing method text similarity is obtained by taking the mean of the processing method character similarity and the processing method semantic similarity.

[0018] If the fault reason text similarity of two fault reasons is greater than or equal to a preset fault reason text similarity threshold, the processing method text similarity is greater than or equal to a preset processing method text similarity threshold, and the fault keywords belong to the same physical system, the fault reason name and the processing method name with the shortest text length are retained, and the two faults and their processing methods are merged.

[0019] Further, in an embodiment, the fault text data is segmented according to the text length of the fault reason, comprising:

[0020] The average text length of all fault reasons in the fault text data list is calculated.

[0021] If the average text length is less than or equal to the first length threshold, the segmentation threshold is the base threshold; if the average text length is greater than the first length threshold and less than or equal to the second length threshold, the segmentation threshold is 2 / 3 of the base threshold; and if the average text length is greater than the second length threshold, the segmentation threshold is 1 / 2 of the base threshold.

[0022] Segment the fault text data according to the segmentation threshold.

[0023] Further, in an embodiment, according to the text similarity between each new fault reason in the subsequent segment data and the existing fault reasons, the new fault reason is classified into an existing fault category or a newly generated fault category, including:

[0024] If the text similarity between the new fault reason and at least one of the existing fault reasons in the existing fault category is greater than or equal to a preset text similarity threshold, the new fault reason is classified into the fault category in which the fault reason with the highest similarity is located.

[0025] If the text similarity between the new fault reason and all existing fault reasons is lower than the preset text similarity threshold, and none of them belongs to the preset similarity fuzzy matching interval, the new fault reason is generated into a newly generated fault category according to the professional term library.

[0026] Further, in an embodiment, after obtaining the classification result data set, further including:

[0027] Calculate the name semantic similarity of each newly generated fault category, if the name semantic similarity of two newly generated fault categories is greater than or equal to a preset name semantic similarity threshold, retain the name of the newly generated fault category generated earliest, and merge the fault text data in the two newly generated fault categories.

[0028] Further, in an embodiment, according to the fault occurrence frequency of each fault category in the classification result data set, high-frequency screening is performed on the fault reasons and their processing methods, including:

[0029] Statistically count the fault occurrence frequency corresponding to each fault reason in each single fault category, and sort the fault reasons according to the fault occurrence frequency, and retain a preset number of high-frequency fault reasons and their processing methods in the single fault category.

[0030] Further, in an embodiment, the fault category name merging of each fault category in the screened classification result data set includes:

[0031] According to the name text similarity, fault keyword constraint, name text length and fault type coverage of each fault category in the screened classification result data set, the fault category name merging is performed, and the merged fault text data is arranged in order of the fault occurrence frequency of a single fault category.

[0032] Further, in an embodiment, the name text similarity of each fault category in the screened classification result data set, the fault keyword constraint, the name text length, and the fault type coverage are used to perform fault category name merging, including:

[0033] If the name text similarity of two fault categories is greater than or equal to a preset name text similarity threshold and the fault keywords belong to the same physical system, the fault text data in the two fault categories are merged, and the fault category name with the shortest name text length and the highest fault type coverage is retained.

[0034] The fault coverage is the ratio of the number of fault cause types of the fault category to the total number of fault cause types of the merged fault category.

[0035] In a second aspect, the present application provides an intelligent power failure analysis device, which includes:

[0036] A deduplication module is configured to deduplicate the fault text data to obtain a fault text data list, and the fault text data list includes fault causes and treatment methods.

[0037] A classification module is configured to segment the fault text data according to the text length of the fault causes, combine the first segment data with a large model to generate a classification framework, and according to the text similarity of each new fault cause and the existing fault causes in the subsequent segment data, classify the new fault cause into an existing fault category or generate a new fault category to obtain a classification result data set.

[0038] A screening module is configured to screen high-frequency fault causes and treatment methods according to the fault occurrence frequency of each fault category in the classification result data set.

[0039] A report generation module is configured to merge the fault category names in the screened classification result data set to generate a fault maintenance report.

[0040] The technical solutions provided by the embodiments of the present application have the following beneficial effects:

[0041] By deduplication of fault text data, data redundancy is effectively reduced, and data processing efficiency is improved. Further, according to the text length of the fault reason, the fault text data is segmented, and a dynamic classification framework is generated by using a large model. Through text similarity calculation, new fault reasons are classified into existing fault categories or new fault categories are generated, which can dynamically adjust the classification structure according to new data and improve the classification accuracy. At the same time, according to the frequency of fault occurrence, the fault reason and its processing method are high-frequency filtered, which ensures the practicality and pertinence of the fault maintenance report. Finally, the fault category names in the classified result data set are merged, which optimizes the fault category structure and further improves the classification efficiency. Therefore, the technical scheme of the present application effectively solves the technical problems of the existing large model fault analysis technology, such as the classification framework cannot be dynamically adjusted, the classification accuracy is low, and the classification efficiency is low, which significantly improves the accuracy and efficiency of fault analysis. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 The flowchart of the power fault intelligent analysis method of the embodiment of the present application.

[0043] Figure 2 The block diagram of the power fault intelligent analysis device of the embodiment of the present application.

[0044] Figure 3 The full-process block diagram of the fault maintenance report generation of the present application. DETAILED DESCRIPTION

[0045] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0046] First, some technical terms in the present application are explained and described to facilitate understanding of the present application by those skilled in the art.

[0047] TF-IDF: Term Frequency-Inverse Document Frequency, word frequency-inverse document frequency, word frequency represents the frequency of a word appearing in a text, and inverse document frequency measures the importance of a word. The process of generating a TF-IDF vector is as follows:

[0048] The word frequency is calculated by formula (1) and formula (2) and the inverse document frequency :

[0049] (1),

[0050] (2).

[0051] The TF-IDF weight is calculated by formula (3):

[0052] (3).

[0053] For a text T , a vector V is generated, where each element （T） TF-IDF represents the TF-IDF weight of each word in the text T .

[0054] In order to make the purpose, technical solutions and advantages of the present application clearer, the embodiments of the present application will be described in further detail below with reference to the accompanying drawings.

[0055] In a first aspect, the embodiments of the present application provide an intelligent power failure analysis method.

[0056] In an embodiment, referring to FIG. 1, the intelligent power failure analysis method comprises the following steps. Figure 1

[0057] S1, de-duplicate the fault text data to obtain a fault text data list, wherein the fault text data list includes fault reasons and treatment methods.

[0058] S2, segment the fault text data according to the text length of the fault reasons, combine the first segment data with a large model to generate a classification framework, and according to the text similarity between each new fault reason in the subsequent segment data and the existing fault reasons, classify the new fault reason into an existing fault category or generate a new fault category to obtain a classification result data set.

[0059] S3, according to the fault occurrence frequency of each fault category in the classification result data set, high-frequency screen the fault reasons and their treatment methods.

[0060] S4, merge the fault category names of each fault category in the screened classification result data set to generate a fault maintenance report.

[0061] ​In this embodiment, by performing deduplication operation on the fault text data, a fault text data list containing fault causes and treatment methods is generated, reducing the interference of repeated data on analysis and improving the efficiency of subsequent processing. The fault text data is segmented according to the text length of the fault cause, and a classification framework is generated by combining the first segment data with the large model. This segmentation processing method can effectively deal with the problem of uneven attention allocation that may occur when processing long text data, ensuring that each paragraph is fully and accurately analyzed. The dynamically generated classification framework can be continuously adjusted and optimized according to new data, improving the adaptability and flexibility of classification. By calculating the text similarity between the new fault cause and the existing fault cause, the new fault cause is classified into an existing fault category or a new fault category. This process not only considers the surface character similarity of the text, but also combines semantic similarity for comprehensive judgment, so that the fault cause can be more accurately identified and classified. According to the fault occurrence frequency of each fault category in the classification result data set, high-frequency filtering is performed on the fault cause and its treatment method, and the fault category names in the filtered classification result data set are merged, finally generating a fault maintenance report. High-frequency filtering can highlight frequently occurring fault causes and their treatment methods, helping maintenance personnel quickly locate and solve major problems, while fault category name merging further optimizes the structure and content of the report, making it clearer, easier to understand and use, thereby improving the quality and practicality of the fault maintenance report.

[0062] Further, in an embodiment, the fault cause and treatment method in step S1 are obtained by semantic extraction of the fault text data without deduplication by a large model, specifically:

[0063] The database connection module interacts with the fault text database, performs structured queries based on target fault codes, time ranges, or fault names, and extracts fault text data containing loss record IDs, downtime start times, manufacturers, model series, standard library fault names, fault codes, loss descriptions (including fault causes), loss handling situations (including treatment methods), and other fields from the fault text database.

[0064] The loss description and loss handling situation in the fault text data are combined into a text input for the semantic task of the large model, with additional professional term explanations and output format constraints. The large model extracts the fault cause and treatment method from the combined text, forms a semantic-level structured output in the format of "fault cause + treatment method", and performs field number matching, null value and illegal character checking on the output result. If the number of fault cause and treatment method fields in the output is inconsistent, or there are null values or illegal characters in the output result, a retry or manual review is triggered.

[0065] Since the fault reason or processing mode output by the large model may contain multiple values, further multi-value field splitting is performed on the output result, the strings are divided by commas as separators, the spaces at the beginning and end of each string element are removed, and independent fault reason lists and processing mode lists are formed. If the number of fault reasons and processing modes after splitting is inconsistent, the number is aligned through a list padding mechanism. Specifically:

[0066] If the number difference between the fault reason and the processing mode is 1, the short list padding mechanism is automatically corrected. If the fault reason list is longer, the last processing mode is copied to the same length. If the number difference between the fault reason and the processing mode is greater than 1, an artificial alignment workflow is triggered, and a pop-up window prompts the operation and maintenance personnel to supplement the missing data. The padding mechanism can ensure that the “fault reason” list and the “processing mode” list have a 1:1 causal relationship, avoiding analysis deviation caused by misplacement of multi-value fields.

[0067] After alignment, the “fault reason-processing mode” pair is encoded using a TF-IDF vector, and the cosine similarity algorithm is used to verify the semantic association of each “reason-processing” pair. When the similarity is less than the preset similarity threshold (i.e., the semantic association is very weak, such as a sensor failure corresponding to a gear box replacement), it is marked as a semantic mismatch and pushed to manual confirmation to ensure the rationality of the causal relationship.

[0068] The fault text data after multi-value field splitting is subjected to term standardization, and the fault reason and processing mode are unified to standard term expression to eliminate semantic noise caused by synonymy and isomorphism or human entry differences. Term standardization is achieved through a mapping table and a similarity matching rule to ensure consistency of expression across devices and personnel.

[0069] Fields with no valid content in the term-standardized fault text data are labeled “no record”, and the Z-score algorithm is used to identify unusually long fields, as follows:

[0070] (4),

[0071] wherein, Z represents an outlier, represents the field length, represents the mean of the field length of the same batch of data, and represents the standard deviation. In this embodiment, if an abnormality is determined, triggering an artificial verification process.

[0072] The term-standardized fault text data is subjected to information density verification by text entropy, as follows:

[0073] (5),

[0074] wherein, represents the text entropy, represents the firsti The probability of each character appearing. m This represents the total number of distinct characters in the text. In this embodiment, if... If the percentage of repeated characters exceeds 80%, the data is considered to have low information density and will be subject to manual review.

[0075] Furthermore, in one embodiment, the deduplication of faulty text data in step S1 above specifically involves the following steps:

[0076] For fault causes with completely duplicate text, perform initial deduplication, accumulate the fault occurrence counts, retain the processing method with the shortest text length, and generate a fault text data list containing the fault cause, processing method, and fault occurrence count, as shown in the example below:

[0077]

[0078] The character similarity of the fault cause and the character similarity of the processing method are calculated using the string edit distance similarity formula, and the semantic similarity of the fault cause and the semantic similarity of the processing method are calculated using the cosine similarity formula.

[0079] String edit distance similarity measures the degree of direct matching at the character level, and the calculation formula is as follows:

[0080] (6),

[0081] in, This represents the similarity of the edit distance between strings. and Represents text, indicating that Convert to The minimum number of insert, delete, and replace operations required.

[0082] Cosine similarity measures the degree of semantic matching, and is calculated as follows:

[0083] Use jieba (a Python library for Chinese text segmentation) to segment Chinese text, such as... T 1 = “IGBT overheating” is segmented into [IGBT, overheating], T 2 = “IGBT module overheating” is segmented into [IGBT, module, overheating].

[0084] The formula for cosine similarity is as follows:

[0085] (7),

[0086] in, Represents cosine similarity. Representing text TF-IDF vector, TF-IDF vector representing the text total dimension of the vocabulary, the first component of the TF-IDF vector, i the first component of the TF-IDF vector. i the first component of the TF-IDF vector.

[0087] The fault reason text similarity is obtained by averaging the fault reason character similarity and the fault reason semantic similarity, and the processing method text similarity is obtained by averaging the processing method character similarity and the processing method semantic similarity.

[0088] The text similarity is as follows:

[0089] (8).

[0090] If the fault reason text similarity of two fault reasons is greater than or equal to a preset fault reason text similarity threshold, the processing method text similarity is greater than or equal to a preset processing method text similarity threshold, and the fault keywords belong to the same physical system, the fault reason name and the processing method name with the shortest text length are retained, and the two faults and the processing methods thereof are merged.

[0091] The constraint condition that the fault keywords belong to the same physical system is to force the same fault reason to belong to the same physical system, match the fault keywords in the fault reason with a pre-constructed keyword library to obtain a uniquely belonging physical system, and the physical systems to which two same fault keywords belong are consistent.

[0092] Further, in an embodiment, in the step S2, the fault text data is segmented according to the text length of the fault reason, and the specific steps are as follows:

[0093] The average text length of all fault reasons in the fault text data list is calculated, and the formula is as follows:

[0094] (9),

[0095] wherein, the average text length, the total number of fault reasons.

[0096] If the average text length is less than or equal to the first length threshold, the segmentation threshold is the basic threshold; if the average text length is greater than the first length threshold and less than or equal to the second length threshold, the segmentation threshold is 2 / 3 of the basic threshold; if the average text length is greater than the second length threshold, the segmentation threshold is 1 / 2 of the basic threshold. ​​​​

[0097] In this embodiment, the basic threshold is 30 pieces per segment, the number of segments is calculated based on the basic threshold, and the average text length of the fault reason is adjusted. The specific formula is as follows:

[0098] (10),

[0099] and satisfies to ensure that the amount of data in each segment is moderate.

[0100] According to the segmentation threshold, the fault text data is segmented, and the segmentation threshold is K The full amount of fault data is divided into M segments, avoiding single input overload, and each segment of data is continuous fault data, recording the segment index start i and segment length M i ， The index of the first segment of data is i=0 , and the index of all segment data is 0~K-1 .

[0101] (11),

[0102] (12).

[0103] Further, in an embodiment, in the above step S2, the first segment of data is combined with the large model to generate a classification framework.

[0104] The classification benchmark is "core component + fault performance" (such as "variable pitch cabinet + electrical element fault"), each category contains multiple "fault reason + treatment method", each "fault reason + treatment method" can only be classified into one category, and cannot belong to multiple categories at the same time. The category output format is {"fault category":["fault 1","fault 2",...]}, and the initial category set G0={g1,g2,...,g p} (in this embodiment, p≤15) is obtained, and the category is excluded, and the category with a coverage of ≥8% is retained (category coverage=total fault occurrence number of single category / total fault occurrence number of all categories), and the classification framework G int =G0' (G0' is the filtered valid category).

[0105] Further, in an embodiment, in the above step S2, according to the text similarity between each new fault reason in the subsequent segment data and the existing fault reason, the new fault reason is classified into the existing fault category or a new fault category is generated. The specific steps are as follows:

[0106] For the first i ( i≥1From the second paragraph to the last paragraph) paragraph data, based on the above classification framework to perform the following operations:

[0107] By paragraph index range start i , end i ) (end i = min ((i+1) x K, N) extract data, collate into local index associated structured text format, that is: k. Cause of failure: T local1global[k] ; processing method: M local1global[k] .

[0108] If the new fault reason and at least one of the existing fault reasons in the existing fault category have a text similarity greater than or equal to a preset text similarity threshold, the new fault reason is classified into the fault category with the highest similarity to the fault reason. In this embodiment, the preset text similarity threshold is 0.7.

[0109] If the text similarity between the new fault reason and all existing fault reasons is less than the preset text similarity threshold, and none of them belongs to the preset similarity fuzzy matching interval, a new fault category is generated according to the professional term library. In this embodiment, the preset similarity fuzzy matching interval is [0.5, 0.7). If the text similarity between the new fault reason and all existing fault reasons is less than the preset text similarity threshold, and belongs to the preset similarity fuzzy matching interval, it is automatically marked as "to be reviewed" and pushed to the professional personnel to confirm the category.

[0110] When a new fault category needs to be added, semantic matching is performed with the power industry term library (including equipment components, fault types, and processing measures professional terms). The similarity between the new fault category name and the term library is required to be greater than or equal to a preset term similarity. The term vector similarity is calculated by the TF-IDF vectorization method. In this embodiment, the preset term similarity is 0.6. Otherwise, it is determined as a non-professional term and needs to be manually adjusted. The similarity between the new fault category and all existing fault categories is calculated by the longest common subsequence (LCS) algorithm. If the similarity with any existing category is greater than or equal to a preset category similarity, that is, the semantics are highly overlapped, such as "slip ring pollution" and "slip ring oil accumulation", the category merging suggestion or manual decision is triggered to ensure the professional term exclusivity of the classification system.

[0111] According to the above classification method, subsequent paragraph data is classified to generate the first i Paragraph mapping relationship map i ={g:[k 1 , k 2 ,...]} ,map i The local index list corresponding to each category, according to the formula global_idx=start i +k Convert to global index, calculate the fault occurrence frequency of the new category for each category mapped by the current segment, and the calculation formula is as follows:

[0112] (13),

[0113] Wherein, Indicates the fault occurrence frequency of the new category, Indicates the first The fault occurrence frequency of the fault data.

[0114] After all the segmented processing is completed, the classification results of the same fault reason in different segments are checked for consistency through hash check + distribution statistics. If there is inconsistency (such as "sensor communication exception" is classified as "communication system failure" in the first segment and "sensor hardware failure" in the second segment), backtrack the segmented matching log, and recommend the optimal classification by the large model or manually intervene to correct it, to ensure the unity of the global classification logic.

[0115] Further, in an embodiment, after obtaining the classification result data set in the above step S2, the following steps are further included:

[0116] Calculate the semantic similarity of the names of each new fault category. If the semantic similarity of the names of two new fault categories is greater than or equal to a preset name semantic similarity threshold, retain the name of the earliest generated new fault category, and merge the fault text data in the two new fault categories.

[0117] Further, in an embodiment, in the above step S3, according to the fault occurrence frequency of each fault category in the classification result data set, high-frequency filtering is performed on the fault reason and its processing method, and the specific steps are as follows:

[0118] Statistical analysis of the fault occurrence frequency of each fault reason in each single fault category, and sorting of the fault reasons according to the fault occurrence frequency, and retaining a preset number of high-frequency fault reasons and their processing methods in the single fault category.

[0119] Further, in an embodiment, in the above step S4, the classification result data set after filtering is subjected to fault category name merging, and the specific steps are as follows:

[0120] According to the name text similarity, fault keyword constraint, name text length and fault type coverage of each fault category in the classification result data set after filtering, the fault category name merging is performed, and the merged fault text data is arranged in order according to the fault occurrence frequency of a single fault category.

[0121] If the name text similarity of two fault categories is greater than or equal to the preset name text similarity threshold and the fault keywords belong to the same physical system, the fault text data in the two fault categories are merged, the name text with the shortest length and the highest fault type coverage is retained, and the index list is merged, that is, , G is the merged category name set.

[0122] Wherein, the fault coverage is the ratio of the number of fault cause types of the fault category to the total number of fault cause types of the merged fault category.

[0123] If the number of fault causes under the fault category is greater than the preset number of faults, which is 20 in this embodiment, the fault causes are further split according to the core components or fault types (such as "sensor abnormality" is split into "temperature sensor abnormality" or "position sensor abnormality"), to ensure that each fault category focuses on a single core fault feature.

[0124] Check whether all the indexes are included in the valid fault category to ensure that the omission rate O is less than or equal to the preset omission rate threshold, which is 3% in this embodiment. If it does not meet the requirement, manual review and supplementary classification are performed. The calculation formula of the omission rate O is as follows:

[0125] (14),

[0126] (15),

[0127] Wherein, represents the total number of fault occurrences of the fault cause, represents the unclassified fault cause index set.

[0128] According to the order of the fault occurrence number of a single fault category from high to low, a classification report containing "category name, core sub-fault, treatment method, total number, and fault proportion" is formed. The fault proportion is the ratio of the fault occurrence number of a single fault category to the fault occurrence number of all fault categories, and the calculation formula is as follows:

[0129] (16),

[0130] (17),

[0131] Wherein, represents the total number of fault occurrences of the fault cause of a single category, represents the fault proportion, and the calculation result is rounded to two decimal places.

[0132] The calculation logic of the statistical results in the automatic verification report automatically re-executes the statistical process by reverse derivation method from the statistical results to the fault text data merging process, if there are repeated calculation, numerical overflow, or proportion and not 100% and other problems.

[0133] The historical repair effect database is called to match the historical effectiveness of the processing mode in the report, if the repair success rate is less than the preset success rate threshold, or the average working hours exceed the preset multiple of the industry benchmark, the historical better scheme is automatically recommended to ensure the feasibility and economy of the repair suggestion, in the embodiment, the preset success rate threshold is 60%, and the preset multiple is 2.

[0134] In a second aspect, an embodiment of an electric power fault intelligent analysis device is provided. Referring to Figure 2 As shown in the figure, the device includes a deduplication module, a classification module, a screening module, and a report generation module, specifically:

[0135] The deduplication module is used to deduplicate the fault text data to obtain a fault text data list, and the fault text data list includes fault causes and processing methods.

[0136] The classification module is used to segment the fault text data according to the text length of the fault causes, combine the first segment data with the large model to generate a classification framework, and according to the text similarity between each new fault cause in the subsequent segment data and the existing fault causes, the new fault cause is classified into the existing fault category or a new fault category is generated to obtain a classification result data set.

[0137] The screening module is used to screen the fault causes and their processing methods according to the fault occurrence times of each fault category in the classification result data set.

[0138] The report generation module is used to merge the fault category names of each fault category in the screened classification result data set to generate a fault repair report.

[0139] Referring to Figure 3 As shown in the figure, a fault repair report generation full-process diagram from data input to fault repair report output is given, which is divided into four layers of structure, and each layer of module cooperates to realize the closed-loop process of "data preprocessing-intelligent classification-report generation", and the specific description is as follows:

[0140] Data Source Layer: Serving as the "data foundation," this layer includes a fault record database and a power industry terminology library. The fault record database stores raw fault data throughout the entire lifecycle of power equipment, covering fields such as fault time, fault description, handling measures, and downtime, providing "raw material" for analysis. The power industry terminology library collects professional terms and standardized definitions in the power field (e.g., pitch safety chain: emergency shutdown protection device), ensuring consistency in professional terminology throughout the process and avoiding ambiguity that could interfere with analysis results.

[0141] The data processing layer includes a data cleaning module and a format conversion module. These modules clean and convert the data to provide high-quality, standardized, and segmentable data for intelligent classification. The data cleaning module addresses the issues of "dirty, messy, and incomplete" data by performing operations such as null value marking, outlier detection (e.g., Z-score algorithm, triggering manual verification when |Z|>3), and deduplication, thereby improving data quality. The format conversion module unifies multi-source heterogeneous data into a standard format that can be parsed by large models, breaking down data format barriers.

[0142] Intelligent classification layer: Based on the "large model + self-growing classification" technology, it realizes the "dynamic construction and optimization" of the fault classification system, including an adaptive segmentation module, a classification framework generation module, a dynamic matching growth module, and a category merging optimization module.

[0143] Adaptive segmentation module: Dynamically adjusts the segmentation strategy based on real-time data volume and large model token efficiency (e.g., when the data volume surges, automatically optimizes the number of records in each segment to ensure token utilization ≥ 80%), improves processing efficiency, and avoids information loss or model performance degradation when processing long texts.

[0144] Classification framework generation module: Uses a large model to perform semantic analysis on the first segment of data to generate a fault classification framework (such as a fault category structure of "core components + fault manifestations" to ensure appropriate classification granularity).

[0145] Dynamic matching and growth module: For data segments 2 to N, the cosine similarity algorithm is used to match existing fault categories; if the matching degree is <0.7, a new category is added, and the professionalism of the fault category and its similarity with existing fault categories are verified by combining the "power industry terminology library", so as to realize the "self-growth" of the classification system.

[0146] Category merging optimization module: Calculates the similarity of faulty categories using the longest common subsequence algorithm; if the similarity is ≥0.7, merges semantically similar categories, retains high-frequency category names, and avoids category redundancy.

[0147] The report summary layer transforms the intelligent classification results into reports that directly guide maintenance. This includes a statistical indicator calculation module, a batch report generation module, and a full report integration module. The statistical indicator calculation module automatically calculates the frequency and percentage of each fault category, quantifying fault distribution patterns. The batch report generation module generates Markdown-formatted phased reports according to classification segments. Each segment covers a preset number of fault categories, including fault causes and handling methods. The full report integration module merges all segmented reports, supplementing them with classification logic explanations, high-frequency faults, and maintenance priority suggestions, ultimately forming a complete maintenance handover report that directly supports power equipment maintenance decisions.

[0148] The flowchart for generating fault repair reports clearly presents the flow logic of data from the bottom layer to the top layer: the data source layer provides "raw materials" → the data processing layer performs "preprocessing" → the intelligent classification layer realizes "core analysis and dynamic classification" → the report summary layer outputs "decision results". The entire process highlights the innovation of "large model empowerment + self-growing classification", solving the pain points of traditional fault classification such as "fixed and difficult to expand" and "limited analysis of long texts", and providing systematic support for intelligent analysis and maintenance optimization of power equipment faults.

[0149] This application addresses the issues of high noise levels and low reliability in direct analysis of raw data through a comprehensive data cleaning mechanism. The Z-score algorithm identifies abnormally long fields and filters out erroneous data. Multi-valued fields are split and aligned, and "short list completion" ensures a one-to-one correspondence between cause and treatment method, avoiding causal misalignment. Duplicates are removed and frequencies are merged based on "standard fault names" to ensure accurate statistical data and provide a reliable data foundation for subsequent classification.

[0150] The dynamic growth classification mechanism solves the problem of rigidity in traditional classification systems. By segmenting the entire dataset into small batches, it avoids the rigidity of a framework caused by one-time classification. New data is matched with existing categories using cosine similarity; when the matching degree falls below 50%, a new category is automatically added, ensuring the classification system can expand as fault types evolve. Simultaneously, redundant categories are quantified and merged to resolve the issue of "different names for the same cause."

[0151] By employing a segmentation and iterative processing strategy, the bottleneck of traditional large-scale model data processing capabilities is overcome. The segmentation threshold is dynamically set according to the model's token limit (default 50 records / segment, with the number of tokens in a single segment less than 80% of the model's limit), ensuring that each batch of data is fully understood by the model. A "first segment framework building + subsequent segment matching and growth" model is adopted: the first segment of data generates an initial classification, and subsequent segments are matched with the existing framework based on cosine similarity, ensuring classification consistency while avoiding excessively large data volumes processed at once. Global validation (comparing with all historical categories) is performed on each newly added category in each segment, ensuring that segmentation processing does not affect the coherence of the overall classification logic, thereby improving the model's ability to process long texts and batches of data.

[0152] The automated processing workflow solves the problems of excessive manual intervention, low efficiency, and strong subjectivity. The entire classification process, from data segmentation and similarity calculation to category merging, is algorithm-driven, eliminating the need for manual definition of classification rules. The statistical analysis automatically calculates the fault percentage, avoiding errors from manual calculations. The reporting process automatically generates structured tables in batches, reducing the report generation time for 1000 records from one week to 10 minutes, while ensuring consistent formatting and reproducible results, significantly improving the efficiency and objectivity of fault analysis.

[0153] By using quantitative indicators and process traceability, the problems of poor interpretability of analysis results and difficulty in tracing the classification basis were solved. Each classification step outputs a quantifiable matching degree, clearly defining the classification basis. During segmented processing, classification logs for each batch are retained, including the verification results of newly added categories. When merging similar categories, the similarity calculation process is output synchronously, ensuring that the entire process is traceable and interpretable, improving maintenance personnel's understanding and trust in the classification logic.

[0154] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0155] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0156] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0157] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0158] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0159] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A power failure intelligent analysis method, characterized in that, The method comprises: The fault text data is de-duplicated to obtain a fault text data list, and the fault text data list comprises fault reasons and treatment methods; According to the text length of the fault reasons, the fault text data is segmented, a classification framework is generated by combining the first segment data and a large model, and according to the text similarity between each new fault reason in the subsequent segment data and the existing fault reasons, the new fault reason is classified into an existing fault category or a newly generated fault category to obtain a classification result data set; According to the fault occurrence frequency of each fault category in the classification result data set, high-frequency screening is performed on the fault reasons and the treatment methods; The fault category names in the screened classification result data set are merged to generate a fault maintenance report; The fault text data is segmented according to the text length of the fault reasons, comprising: The average text length of all fault reasons in the fault text data list is calculated; If the average text length is less than or equal to a first length threshold, the segmentation threshold is a basic threshold; if the average text length is greater than the first length threshold and less than or equal to a second length threshold, the segmentation threshold is 2 / 3 of the basic threshold; if the average text length is greater than the second length threshold, the segmentation threshold is 1 / 2 of the basic threshold; The fault text data is segmented according to the segmentation threshold.

2. The power failure intelligent analysis method of claim 1, wherein, The fault reasons and the treatment methods are obtained by performing semantic extraction on the fault text data that has not been de-duplicated by a large model; If the number of fault reasons and the number of treatment methods in the fault text data list are inconsistent, the number is aligned through a list completion mechanism.

3. The power failure intelligent analysis method of claim 1, wherein, The fault text data is de-duplicated, comprising: The fault reason character similarity and the treatment method character similarity are calculated according to the string edit distance similarity formula, and the fault reason semantic similarity and the treatment method semantic similarity are calculated according to the cosine similarity formula; The fault reason text similarity is obtained by taking the average of the fault reason character similarity and the fault reason semantic similarity, and the treatment method text similarity is obtained by taking the average of the treatment method character similarity and the treatment method semantic similarity; If the fault reason text similarity of two fault reasons is greater than or equal to a preset fault reason text similarity threshold, the treatment method text similarity is greater than or equal to a preset treatment method text similarity threshold, and the fault keywords belong to the same physical system, the fault reason name and the treatment method name with the shortest text length are retained, and the two faults and the treatment methods are combined.

4. The power failure intelligent analysis method of claim 1, wherein, According to the text similarity between each new fault reason in the subsequent segment data and the existing fault reasons, the new fault reason is classified into an existing fault category or a newly generated fault category, comprising: If the text similarity between the new fault reason and at least one existing fault reason in the existing fault category is greater than or equal to a preset text similarity threshold, the new fault reason is classified into the fault category where the fault reason with the highest similarity is located; If the text similarity between the new fault reason and all existing fault reasons is lower than the preset text similarity threshold, and none of them belongs to the preset similarity fuzzy matching interval, a new fault category is generated for the new fault reason according to a professional term library.

5. The power failure intelligent analysis method of claim 4, wherein, After obtaining the classification result data set, further comprising: Calculate the name semantic similarity of each newly added fault category. If the name semantic similarity of two newly added fault categories is greater than or equal to a preset name semantic similarity threshold, retain the name of the earliest generated newly added fault category, and merge the fault text data in the two newly added fault categories.

6. The power failure intelligent analysis method of claim 1, wherein, The high-frequency screening of fault reasons and their processing methods according to the fault occurrence frequency of each fault category in the classification result data set comprises: Count the fault occurrence frequency of each fault reason in each single fault category, and sort the fault reasons according to the fault occurrence frequency. In a single fault category, retain a preset number of high-frequency fault reasons and their processing methods.

7. The power failure intelligent analysis method of claim 1, wherein, The name merging of each fault category in the screened classification result data set comprises: According to the name text similarity, fault keyword constraint, name text length, and fault type coverage of each fault category in the screened classification result data set, the name merging of each fault category is performed, and the merged fault text data is arranged in order of the fault occurrence frequency of a single fault category.

8. The power failure intelligent analysis method of claim 7, wherein, The name merging of each fault category in the screened classification result data set comprises: If the name text similarity of two fault categories is greater than or equal to a preset name text similarity threshold and the fault keywords belong to the same physical system, the fault text data in the two fault categories is merged, and the name of the fault category with the shortest name text length and the highest fault type coverage is retained. The fault coverage is the ratio of the number of fault reason types of a fault category to the total number of fault reason types of the merged fault categories.

9. An electric power failure intelligent analysis apparatus characterized by comprising: The device comprises: A deduplication module for deduplicating fault text data to obtain a fault text data list, the fault text data list comprising fault reasons and processing methods; A classification module for segmenting the fault text data according to the text length of the fault reasons, generating a classification framework by combining the first segment data with a large model, and according to the text similarity of each new fault reason in the subsequent segment data with the existing fault reasons, classifying the new fault reason into an existing fault category or generating a newly added fault category to obtain a classification result data set; A screening module for high-frequency screening of fault reasons and their processing methods according to the fault occurrence frequency of each fault category in the classification result data set; A report generation module for name merging of each fault category in the screened classification result data set to generate a fault maintenance report; Segmenting the fault text data according to the text length of the fault reasons, comprising: Calculating the average text length of all fault reasons in the fault text data list; If the average text length is less than or equal to a first length threshold, the segmentation threshold is a base threshold; if the average text length is greater than the first length threshold and less than or equal to a second length threshold, the segmentation threshold is 2 / 3 of the base threshold; if the average text length is greater than the second length threshold, the segmentation threshold is 1 / 2 of the base threshold; Segmenting the fault text data according to the segmentation threshold. Segmenting the fault text data according to the segmentation threshold.

Images