A method for merging and processing low-voltage user fault reports based on event extraction

Abstract

The present invention discloses a method for merging and processing low-voltage user fault reports based on event extraction, wherein the fault report texts to be identified are batch-inputted into a fault classifier, and the fault classifier determines whether all the fault report texts to be identified are of the same fault type, and if so, all the fault report texts to be identified are merged and processed. The present invention is directed to an event extraction method for fault report texts, extracts fault event elements and identifies fault types from the fault report texts reported by users, and merges and processes repeated fault reports of the same type from multiple users in the same substation, thereby reducing the manual processing process and workload, improving the fault repair response rate, and further improving power supply reliability and user satisfaction.

Description

Technical Field

[0001] The invention relates to a low-voltage user fault repair merging processing method based on event extraction, belonging to the technical field of fault repair. Background Art

[0002] With the rapid economic development, the society's demand for electric energy continues to grow, and the number of electricity users continues to increase, which puts higher requirements on the reliability of the power grid. Low-voltage user fault reporting is an important business for power grid companies to users, and it is also an important source for the establishment of subsequent distribution and rescue work orders. At present, low-voltage user fault reporting is mainly done by filling in faults on the PC / mobile terminal or reporting through the service hotline. The number of reports is large, and the content and form of the fault description are diverse. The existing fault reporting business mainly relies on manual processing one by one, and it is also necessary to merge the work orders for repeated reports from users caused by the same fault point, which is a large workload and inefficient. Due to the diversity of user fault reporting content descriptions and multiple users reporting the same fault point, the efficiency of manual processing is limited.

[0003] At present, the research on optimizing user fault reporting and repair processing mainly focuses on process optimization and statistical analysis. Existing technical research either analyzes from a text perspective or from the perspective of distribution network operation status, and is not very targeted in identifying low-voltage user fault events. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a method for merging and processing low-voltage user fault reports based on event extraction.

[0005] To achieve the above-mentioned purpose, the present invention provides a method for merging and processing low-voltage user fault reports based on event extraction, in which the fault report texts to be identified are batch-inputted into a fault classifier constructed, and the fault classifier determines whether all the fault report texts to be identified are of the same fault type. If so, all the fault report texts to be identified are merged and processed.

[0006] Preferably, in step 1, the fault repair texts to be identified are grouped according to the substations, and there are L groups in total, where L is a positive integer; and a value a=1 is assigned;

[0007] Step 2, determine whether a is less than or equal to L, if yes, proceed to step 3, if not, end the operation;

[0008] Step 3, determine whether there are more than three users in the same area reporting the fault repair text to be identified, if yes, go to step 4, if not, go to step 2 and increase the value of a by 1;

[0009] Step 4, determine whether the fault repair texts to be identified reported by all users in step 3 are of the same fault type. If so, they are processed together. If not, go to step 2 and increase the value of a by 1. If not, go to step 2 and increase the value of a by 1.

[0010] Prioritize building a fault classifier, including:

[0011] Construct a fault event extraction model;

[0012] Extract fault event elements based on the fault event extraction model;

[0013] Based on the fault event elements, feature selection is performed on the fault report text in the existing fault report record, and the real-time operating status of the distribution transformer and line to which the repair user belongs within a period of time before the fault time is extracted to obtain a feature set;

[0014] The feature set is input into the Naive Bayes algorithm and trained to obtain the fault classifier.

[0015] Prioritize building a fault event extraction model, including:

[0016] Annotate the fault report texts in the existing fault report records according to the feature template to obtain a training corpus;

[0017] The training corpus is input into the CRF model for training, and the mapping relationship between the training corpus and the fault event elements is constructed to obtain the fault event extraction model.

[0018] Preferably, based on the fault event extraction model, the fault event elements are extracted, including:

[0019] Fault event trigger words are filtered from the fault report text in the existing fault report records, and the fault description sentences in the fault report text are located using the fault event trigger words. The fault description sentences are brought into the fault event extraction model to identify and obtain the fault event elements.

[0020] Preferably, a feature set is obtained, including:

[0021] Based on the improved TF-DFI algorithm, feature selection is performed on the fault actions in the fault report text in the existing fault report records and the fault devices in the fault report text in the existing fault report records;

[0022] The real-time operating status of the distribution transformer and line to which the repair user belongs within a period of time before the fault time is extracted to obtain a feature set including fault action, faulty equipment and real-time operating status.

[0023] Preferably, feature selection is performed on the fault actions in the fault report text in the existing fault report record and the fault devices in the fault report text in the existing fault report record based on the improved TF-DFI algorithm, including:

[0024] Select historical fault texts from existing fault repair records as training samples;

[0025] The TF-IDF algorithm is used to calculate the weight of each fault action in each fault type;

[0026] The improved TF-IDF algorithm is used to calculate the weight of each faulty device in each fault type;

[0027] Sort by weight from large to small, select some fault actions in each fault type, and obtain the fault action feature set {V};

[0028] Sort by weight from large to small, select some faulty devices in each fault type, and obtain a faulty device feature set {D};

[0029] Among them, the TF-IDF algorithm is used to calculate the weight of each fault action in each fault type, including:

[0030] Calculate the word frequency of a single fault action in the fault type:

[0031]

[0032] Where V i is the ith fault action, C k is the kth fault type to which the training sample belongs, and the historical fault text is the fault repair text in the existing fault repair record;

[0033] Calculate the reverse frequency of a single fault action:

[0034]

[0035] Calculate the weight of a single fault action and normalize it:

[0036]

[0037] Preferably, the improved TF-IDF algorithm is used to calculate the weight of each faulty device in each fault type, including:

[0038] Calculate the word frequency of a single faulty device in the fault type:

[0039]

[0040] Where D i is the i-th faulty device, V jis the jth fault action, j∈[1,m], m is a positive integer, j=i, WTFV j C k is the weight of a single fault action after normalization in the previous step WTFV i C k :

[0041]

[0042] Calculate the weight of a single faulty device:

[0043] TIFD i C k =TFD i C k ×IDFD i .

[0044] Preferably, the feature set is input into the Naive Bayes algorithm to train a fault classifier, including:

[0045] The naive Bayes algorithm is used to construct a fault classifier and calculate the probability of each feature item in the feature set appearing in each fault type:

[0046]

[0047] In the formula, the total number of all feature items in the feature set is N,w i is the i-th feature item in the feature set {W};

[0048] If the Tth training sample contains m feature items, then w i Fault type C k The probability is:

[0049]

[0050] Based on P(T|C k ) confirms the fault type to which the Tth training sample belongs.

[0051] Prioritize the merging of all fault report texts to be identified, including:

[0052] All the fault report texts to be identified are deleted until the number of the fault report texts to be identified is one.

[0053] The beneficial effects achieved by the present invention are:

[0054] The present invention provides a method for merging and processing low-voltage user fault reports based on event extraction, which makes full use of the effective information of the existing power grid system, adopts an event extraction method for fault report text, extracts fault event elements and identifies fault types from the user's fault report text, and merges and processes repeated fault reports of the same type from multiple users in the same substation, thereby reducing the manual processing flow and workload, improving the fault repair response rate, and further improving power supply reliability and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a flow chart of the present invention;

[0056] Figure 2 is a flow chart of constructing a fault classifier according to the present invention;

[0057] Figure 3 It is a logic diagram of the fault classification implementation of the present invention. DETAILED DESCRIPTION

[0058] The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and are not intended to limit the protection scope of the present invention.

[0059] A method for merging and processing low-voltage user fault reports based on event extraction includes the following steps: Figure 1 :

[0060] Step SS1: Build a fault extraction model

[0061] Based on the existing business knowledge accumulation, the fault repair texts reported by users are sorted into four fault types: equipment failure, power quality problems, safety hazards and other power outages. Secondly, a feature template including context features, part-of-speech features, equipment features, time features and fault action features is defined. The feature template is input into the CRF model training to form a fault event extraction model, and the fault event extraction model outputs fault event elements.

[0062] Fault event elements include fault time, fault action, fault equipment and fault type.

[0063] Step SS2: Fault event element extraction

[0064] Fault event trigger words are filtered from historical fault report texts, and fault description sentences in the fault report texts are located using the fault event trigger words. The fault description sentences are then introduced into the fault event extraction model to identify fault event elements.

[0065] Step SS3: Fault type identification

[0066] Select historical fault repair texts and use fault event elements to perform feature selection on the fault repair texts;

[0067] Feature selection includes, based on the improved TF-DFI algorithm, performing feature selection on the fault actions and faulty equipment in the fault report text, and extracting the real-time operating status of the distribution transformer and line to which the reporting user belongs within 1 hour before the fault time, to form a feature set including fault actions, faulty equipment and real-time operating status. The feature set is input into the naive Bayes algorithm to train a fault classifier, thereby realizing fault type identification.

[0068] The real-time operating status includes line switch position, line overload, distribution transformer operation status stop, distribution transformer three-phase imbalance rate exceeding the standard, distribution transformer voltage qualification rate exceeding the standard and distribution transformer overload.

[0069] Step SS4: Fault event merging

[0070] like Figure 3 As shown, in step 1, the fault repair texts to be identified are grouped according to the substations, and there are L groups in total, where L is a positive integer;

[0071] Assign a=1;

[0072] Step 2, determine whether a is less than or equal to L, if yes, proceed to step 3, if not, end the operation;

[0073] Step 3, determine whether there are more than three users in the same area reporting the fault repair text to be identified, if yes, go to step 4, if not, go to step 2 and increase the value of a by 1;

[0074] Step 4, determine whether the fault report texts to be identified reported by all users in step 3 are of the same fault type. If so, merge them (all fault report texts to be identified reported by all users are merged into one fault report text); if not, go to step 2 and increase the value of a by 1; if not, go to step 2 and increase the value of a by 1.

[0075] Batch input fault classifier to determine whether the fault report text belongs to the same type of fault type. If so, the same type of fault types will be merged for processing.

[0076] Fault types are divided into equipment failure, power quality issues, safety hazards and other power outages. Each type has a related typical fault report text, see Table 1.

[0077] Table 1

[0078]

[0079] The feature template includes context features, part-of-speech features, faulty equipment features, fault time features, and fault action features; the feature template is a CRF that combines the length of the fault repair text to be identified, and uses the information of the preceding and following characters and feature information as a combination probability information set. The feature template is constructed by using the common context features and part-of-speech features of the CRF feature template, combined with the faulty equipment features, fault time features, and fault action features unique to the fault repair text.

[0080] Construct faulty device features, including:

[0081] In order to reduce the complexity of recognition and improve the recognition accuracy, a distribution network general equipment dictionary was constructed and the equipment in the fault text was annotated. The THULAC Chinese lexical analysis toolkit was called to obtain a list of nouns from the fault repair text in the existing fault repair records, and the distribution network general equipment dictionary was obtained by screening. In the distribution network general equipment dictionary, the parts of speech such as electric meters, transformers, wires and switches are separately annotated as Device, and the equipment outside the maintenance scope such as household appliances is annotated as outOfDevice to obtain the characteristics of the faulty equipment.

[0082] Construct fault action features, including:

[0083] The THULAC Chinese lexical analysis toolkit is called to obtain a verb list and an adjective list from the fault report text in the existing fault report records, and the commonly used descriptive verbs related to the fault are screened and marked, such as power outage, tripping, fire, smoke, abnormality and instability, etc. These commonly used descriptive verbs are summarized and sorted to construct a trigger word set and obtain the fault action characteristics.

[0084] Construct failure time features, including:

[0085] The fault time is generally reported immediately. Regular expressions are used to extract and annotate common time expressions. The extraction range includes weekday expressions (such as Monday and last Monday), date expressions (such as today, yesterday, morning, afternoon and evening), and time expressions (such as 7:45 and ten minutes ago). The annotated word is Time to obtain the fault time feature.

[0086] Perform CRF model training

[0087] The fault report text in the existing fault report record is annotated according to the aforementioned feature template to obtain a training corpus, and the training corpus is input into CRF for model training. In this embodiment, the CRF model training process follows the classic algorithm and is supported by open source software such as CRF++ and TensorFlow. The present invention adopts CRF++ training, which will not be described in detail here.

[0088] Based on the fault event extraction model, extract the fault event elements, including:

[0089] For the fault report text in the existing fault report record, the target sentence to be extracted is distinguished, so as to facilitate more accurate identification of fault event elements. The positioning method adopts fault event trigger word positioning, that is, the fault action feature in the feature template in step SS1 is used as the fault event trigger word, see Table 2, and the sentence containing the specific fault event trigger word in the fault report text is used as the target sentence of the fault event element to be extracted. The fault event elements include fault time, fault action, fault equipment and fault type.

[0090] Table 2

[0091]

[0092] Use the trained fault event extraction model to identify fault event elements in historical fault report texts, remove common stop words in advance, and convert the fault time to a unified time standard. The extraction results are described in JSON format. The example is as follows: At 8 o'clock in the evening, the 2# branch line circuit breaker tripped.

[0093] {"trigger":["trip"],"argument":{"device":"2# branch line circuit breaker","time","2020-3-2 20:00"}Select the historical fault text collection

[0094] Select the historical fault texts in the existing fault repair records as training samples and label them as C according to the four fault types. k (k=1, 2, 3, 4), use step SS2 to extract fault event elements including fault time, fault action, fault equipment and fault type.

[0095] The TF-IDF algorithm is used to calculate the weight of each fault action in each fault type.

[0096] Using V i Represents the i-th fault action, and calculates the word frequency of a single fault action in the fault type:

[0097]

[0098] Calculate the reverse frequency of a single fault action:

[0099]

[0100] Calculate the weight of a single fault action and normalize it:

[0101]

[0102] The improved TF-IDF algorithm is used to calculate the weight of each faulty device in each fault type.

[0103] Using D j represents the jth faulty device, C k represents the kth fault type.

[0104] Calculate the word frequency of a single faulty device in the faulty device type:

[0105]

[0106] D i is the i-th faulty device, V j is the jth fault action, j∈[1,m], m is a positive integer, j=i, WTFV j C k is the weight of a single fault action after normalization in the previous step WTFV i C k ;

[0107] Calculate the reverse frequency of a single faulty device:

[0108]

[0109] Calculate the feature weight of a single faulty device:

[0110] TIFD i C k =TFD i C k ×IDFD i .

[0111] Obtain feature set: Sort by weight from large to small, and obtain the top 70% fault action feature set {V} and fault device feature set {D} of each fault type.

[0112] Taking the mapping relationship between the substation, distribution transformer and user number in the power grid marketing system and the load control system as the basic data, the distribution transformer and line to which the user who reported the fault belongs are obtained, and the real-time operating status of the distribution transformer and line within 1 hour before the fault occurs is obtained as the equipment status feature set {ET}. The real-time operating status includes line switch position ET1, line overload ET2, distribution transformer operation status stop ET3, distribution transformer three-phase imbalance rate exceeds the standard ET4, distribution transformer voltage qualification rate exceeds the standard ET5 and distribution transformer overload ET6. If there is an abnormal real-time operating status before the fault time, the corresponding feature annotation is performed on the fault repair text. The fault action feature set {V}, the fault equipment feature set {D} and the equipment status feature set {ET} are merged to obtain the complete feature set {W}. Construct a fault classifier:

[0113] The feature set in the previous step is used as input, and the fault classifier is constructed using the naive Bayes algorithm. The total number of features is defined as N, and the probability of each feature item in the feature set appearing in each fault type is calculated.

[0114]

[0115] If the Tth training sample contains m feature items, then w i Fault type C k The probability is:

[0116]

[0117] Calculate P(T|C k ) maximum value can confirm the fault type category to which the fault text T belongs.

[0118] The specific fault classification execution process of the present invention extracts faulty equipment and fault actions, obtains abnormal characteristics of the distribution transformer and line to which the user belongs one hour ago according to the fault time, and then confirms the feature set of the fault report text, inputs the feature set to the fault classifier, and obtains the classification to which the fault type belongs.

[0119] If multiple users in the same area repeatedly report faults, the following steps are used to merge them:

[0120] See the flowchart Figure 3 , taking 1 hour as the time window, cyclically read the fault repair texts within 1 hour. Based on the mapping relationship between the substation, distribution transformer and user number in the power grid marketing system and load control system, the acquired data within 1 hour is grouped by substation. If there are more than 3 fault repair texts in the same substation and the fault types identified by the fault classifier are the same, the fault repair texts are merged into the same fault.

[0121] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for merging and processing low-voltage user fault reports based on event extraction, characterized in that: The fault repair texts to be identified are input into the constructed fault classifier in batches, and the fault classifier determines whether all the fault repair texts to be identified are of the same fault type. If so, all the fault repair texts to be identified are merged for processing; The construction of the fault classifier comprises: Construct a fault event extraction model; Extract fault event elements based on the fault event extraction model; Based on the fault event elements, feature selection is performed on the fault report text in the existing fault report record, and the real-time operating status of the distribution transformer and line to which the repair user belongs within a period of time before the fault time is extracted to obtain the feature set {W}; Input the feature set into the Naive Bayes algorithm and train to obtain a fault classifier; The construction of the fault event extraction model includes: Annotate the fault report texts in the existing fault report records according to the feature template to obtain a training corpus; The training corpus is input into the CRF model for training, a mapping relationship between the training corpus and the fault event elements is constructed, and a fault event extraction model is obtained; The extraction of fault event elements based on the fault event extraction model includes: Fault event trigger words are filtered from the fault report text in the existing fault report records, and the fault description sentences in the fault report text are located using the fault event trigger words. The fault description sentences are brought into the fault event extraction model to identify and obtain the fault event elements.

2. According to the method for merging and processing low-voltage user fault reports based on event extraction according to claim 1, it is characterized in that: Step 1: The fault repair texts to be identified are grouped according to the substations, and there are L groups in total, where L is a positive integer; Assign a=1; Step 2, determine whether a is less than or equal to L, if yes, proceed to step 3, if not, end the operation; Step 3, determine whether there are more than three users in the same area reporting the fault repair text to be identified, if yes, go to step 4, if not, go to step 2 and increase the value of a by 1; Step 4, determine whether the fault repair texts to be identified reported by all users in step 3 are of the same fault type. If so, they are processed together. If not, go to step 2 and increase the value of a by 1. If not, go to step 2 and increase the value of a by 1.

3. According to the method for merging and processing low-voltage user fault reports based on event extraction according to claim 1, it is characterized in that: Get a feature set, including: Based on the improved TF-DFI algorithm, feature selection is performed on the fault actions in the fault report text in the existing fault report records and the fault devices in the fault report text in the existing fault report records; The real-time operating status of the distribution transformer and line to which the repair user belongs within a period of time before the fault time is extracted to obtain a feature set including fault action, faulty equipment and real-time operating status.

4. A method for merging and processing low-voltage user fault reports based on event extraction according to claim 3, characterized in that: Based on the improved TF-DFI algorithm, feature selection is performed on the fault actions in the fault report text in the existing fault report record and the fault devices in the fault report text in the existing fault report record, including: Select historical fault texts from existing fault repair records as training samples; The TF-IDF algorithm is used to calculate the weight of each fault action in each fault type; The improved TF-IDF algorithm is used to calculate the weight of each faulty device in each fault type; Sort by weight from large to small, select some fault actions in each fault type, and obtain the fault action feature set {V}; Sort by weight from large to small, select some faulty devices in each fault type, and obtain a faulty device feature set {D}; Among them, the TF-IDF algorithm is used to calculate the weight of each fault action in each fault type, including: Calculate the word frequency of a single fault action in the fault type: ； In the formula, V i For the i Fault action, C k The training sample belongs to k The fault type, the historical fault text is the fault repair text in the existing fault repair record; Calculate the reverse frequency of a single fault action: ； Calculate the weight of a single fault action and normalize it: 。 5. A method for merging and processing low-voltage user fault reports based on event extraction according to claim 4, characterized in that: The improved TF-IDF algorithm is used to calculate the weight of each faulty device in each fault type, including: Calculate the word frequency of a single faulty device in the fault type: ； In the formula, D i For the i Faulty equipment, V j It is j fault actions, j∈[1,m], m is a positive integer, j = i , WTF j C k is the weight of a single fault action after normalization in the previous step WTF i C k ; Calculate the reverse frequency of a single faulty device: ； Calculate the weight of a single faulty device: TIFD i C k = TFD i C k × IDFD i 。 6. The method for merging and processing low-voltage user fault reports based on event extraction according to claim 1 is characterized in that: The feature set is input into the Naive Bayes algorithm to train the fault classifier, including: The naive Bayes algorithm is used to construct a fault classifier and calculate the probability of each feature item in the feature set appearing in each fault type: ； In the formula, the total number of feature items in the feature set is N, w i is the first i Feature item; If the Tth training sample contains m feature items, then w i Fault type C k The probability is: ， based on The maximum value of confirms the fault type to which the Tth training sample belongs.

7. The method for merging and processing low-voltage user fault reports based on event extraction according to claim 1 is characterized in that: All the fault repair texts to be identified are merged and processed, including: Deleting all the fault report texts to be identified until the number of the fault report texts to be identified is one; The construction of the fault classifier comprises: Construct a fault event extraction model; Extract fault event elements based on the fault event extraction model; Based on the fault event elements, feature selection is performed on the fault report text in the existing fault report record, and the real-time operating status of the distribution transformer and line to which the repair user belongs within a period of time before the fault time is extracted to obtain the feature set {W}; Input the feature set into the Naive Bayes algorithm and train to obtain a fault classifier; The construction of the fault event extraction model includes: Annotate the fault report texts in the existing fault report records according to the feature template to obtain a training corpus; The training corpus is input into the CRF model for training, a mapping relationship between the training corpus and the fault event elements is constructed, and a fault event extraction model is obtained; The extraction of fault event elements based on the fault event extraction model includes: Fault event trigger words are filtered from the fault report text in the existing fault report records, and the fault description sentences in the fault report text are located using the fault event trigger words. The fault description sentences are brought into the fault event extraction model to identify and obtain the fault event elements.

Images