Power operation and maintenance information mining method and system based on inspection business

Abstract

The embodiment of the invention relates to the technical field of electric power data analysis, in particular to an inspection service-based electric power operation and maintenance information mining method and system, which can comprehensively and accurately integrate operation and maintenance information by distinguishing structured and unstructured parts in target electric power operation and maintenance information. Operation and maintenance inspection semantic vectors and equipment operation index vectors are mined for different types of information, and respective information values are fully mined. After vector interaction, operation and maintenance inspection decision vectors are obtained, then fault weights are obtained through discriminant analysis, and deep fusion and analysis of multi-source information are achieved. According to the mode, the abnormal operation and maintenance information is accurately judged, the operation and maintenance efficiency and precision are improved, the manual troubleshooting workload is reduced, the fault risk of a power system is effectively reduced, the intelligence and reliability of power operation and maintenance are enhanced, and the stability of power supply is guaranteed.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of power data analysis, and in particular, to a method and system for mining power operation and maintenance information based on audit services. Background Art

[0002] In the field of power operation and maintenance, traditional operation and maintenance methods face many challenges. Power operation and maintenance information contains structured data (such as equipment operation values, etc.) and unstructured data (such as inspection text descriptions, etc.). Traditional methods are difficult to effectively process these two types of data at the same time. On the one hand, the lack of effective means to integrate and analyze the two types of data leads to information fragmentation and the inability to fully and accurately reflect the operating status of power equipment. On the other hand, when judging abnormal power operation and maintenance situations, they mostly rely on manual experience or simple indicators, which are of low accuracy and low efficiency, and it is difficult to quickly locate potential faults. This not only increases labor costs, but also may affect the stability of power supply due to untimely fault detection, and increase the risk of power system failure. Summary of the invention

[0003] In order to at least overcome the above-mentioned deficiencies in the prior art, one of the purposes of the embodiments of the present invention is to provide a method and system for mining power operation and maintenance information based on audit business.

[0004] An embodiment of the present invention provides a method for mining electric power operation and maintenance information based on audit services, the method being applied to an electric power operation and maintenance information mining system, the method comprising: Acquire first power operation and maintenance information and second power operation and maintenance information corresponding to the target power operation and maintenance information, wherein the first power operation and maintenance information includes an operation and maintenance information set that does not include structured data in the target power operation and maintenance information, and the second power operation and maintenance information includes an operation and maintenance information set that includes structured data in the target power operation and maintenance information; Performing data mining on the first power operation and maintenance information to obtain an operation and maintenance inspection semantic vector corresponding to the first power operation and maintenance information; performing data mining on the second power operation and maintenance information to obtain an equipment operation indicator vector corresponding to the second power operation and maintenance information; Performing vector interaction on the equipment operation index vector and the operation and maintenance inspection semantic vector to obtain an operation and maintenance inspection decision vector corresponding to the target power operation and maintenance information; performing discriminant analysis on the operation and maintenance inspection decision vector to obtain a power operation and maintenance inspection fault weight corresponding to the target power operation and maintenance information; In response to the power operation and maintenance audit fault weight falling into a set weight interval, the target power operation and maintenance information is determined as abnormal operation and maintenance information.

[0005] Optionally, the acquiring the first power operation and maintenance information and the second power operation and maintenance information corresponding to the target power operation and maintenance information includes: Performing structured feature recognition on the target power operation and maintenance information to obtain a first operation and maintenance information set that does not contain structured data in the target power operation and maintenance information, and performing information splitting on the first operation and maintenance information set in the target power operation and maintenance information as the first power operation and maintenance information; Based on the first operation and maintenance information set, a second operation and maintenance information set containing structured data in the target power operation and maintenance information is obtained, and the second operation and maintenance information set in the target power operation and maintenance information is split into information as the second power operation and maintenance information.

[0006] Optionally, the acquiring the first power operation and maintenance information and the second power operation and maintenance information corresponding to the target power operation and maintenance information includes: Performing structured feature recognition on the target power operation and maintenance information to obtain a third operation and maintenance information set containing structured data in the target power operation and maintenance information, and performing information splitting on the third operation and maintenance information set in the target power operation and maintenance information to obtain the second power operation and maintenance information; Based on the third operation and maintenance information set, a fourth operation and maintenance information set that does not contain structured data in the target power operation and maintenance information is obtained, and the fourth operation and maintenance information set in the target power operation and maintenance information is split into information as the first power operation and maintenance information.

[0007] Optionally, performing data mining on the first power operation and maintenance information to obtain an operation and maintenance inspection semantic vector corresponding to the first power operation and maintenance information includes: performing data mining on the first power operation and maintenance information through a first data mining module in a deep learning algorithm to obtain an operation and maintenance inspection semantic vector corresponding to the first power operation and maintenance information; The performing data mining on the second power operation and maintenance information to obtain the equipment operation index vector corresponding to the second power operation and maintenance information includes: performing data mining on the second power operation and maintenance information through a second data mining module in the deep learning algorithm to obtain the equipment operation index vector corresponding to the second power operation and maintenance information; The performing vector interaction on the equipment operation index vector and the operation and maintenance inspection semantic vector to obtain the operation and maintenance inspection decision vector corresponding to the target power operation and maintenance information comprises: performing vector interaction on the equipment operation index vector and the operation and maintenance inspection semantic vector through a vector interaction module in the deep learning algorithm to obtain the operation and maintenance inspection decision vector corresponding to the target power operation and maintenance information; The discriminant analysis is performed on the operation and maintenance audit decision vector to obtain the power operation and maintenance audit fault weight corresponding to the target power operation and maintenance information, including: through the discriminant analysis module in the deep learning algorithm, the discriminant analysis is performed on the operation and maintenance audit decision vector to obtain the power operation and maintenance audit fault weight corresponding to the target power operation and maintenance information.

[0008] Optionally, in response to the power operation and maintenance audit fault weight falling into a set weight interval, the target power operation and maintenance information is determined as abnormal operation and maintenance information, including: on the basis that the power operation and maintenance audit fault weight is greater than a first set fault weight, the target power operation and maintenance information is determined as abnormal operation and maintenance information.

[0009] Optionally, in response to the power operation and maintenance audit fault weight falling into a set weight interval, the target power operation and maintenance information is determined as abnormal operation and maintenance information, including: on the basis that the power operation and maintenance audit fault weight is less than a second set fault weight, the target power operation and maintenance information is determined as abnormal operation and maintenance information.

[0010] Optionally, before performing data mining on the first power operation and maintenance information through a first data mining module in a deep learning algorithm to obtain an operation and maintenance inspection semantic vector corresponding to the first power operation and maintenance information, the method further includes: Obtaining a first power operation and maintenance information sample and a second power operation and maintenance information sample corresponding to the power operation and maintenance information sample, wherein the first power operation and maintenance information sample includes an operation and maintenance information set that does not include structured data in the power operation and maintenance information sample, and the second power operation and maintenance information sample includes an operation and maintenance information set that includes structured data in the power operation and maintenance information sample; Performing data mining on the first power operation and maintenance information sample through a first data mining module in the deep learning algorithm to obtain an operation and maintenance inspection semantic vector sample corresponding to the first power operation and maintenance information sample; Performing data mining on the second power operation and maintenance information sample through a second data mining module in the deep learning algorithm to obtain a first equipment operation indicator vector sample corresponding to the second power operation and maintenance information sample; Through the vector interaction module in the deep learning algorithm, the operation and maintenance inspection semantic vector sample and the first equipment operation indicator vector sample are vector-interacted to obtain the operation and maintenance inspection decision vector sample corresponding to the power operation and maintenance information sample; Through the discriminant analysis module in the deep learning algorithm, the operation and maintenance audit decision vector sample is discriminantly analyzed to obtain the power operation and maintenance fault training weight corresponding to the power operation and maintenance information sample; The deep learning algorithm is debugged based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight corresponding to the power operation and maintenance information sample.

[0011] Optionally, the debugging of the deep learning algorithm based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight corresponding to the power operation and maintenance information sample includes: Through the first fully connected module in the deep learning algorithm, the operation and maintenance inspection semantic vector sample is fully connected to obtain the operation and maintenance inspection fully connected prediction vector corresponding to the operation and maintenance inspection semantic vector sample; The deep learning algorithm is debugged based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight, as well as the operation and maintenance inspection full-connection prediction vector and the operation and maintenance inspection full-connection authentication vector corresponding to the power operation and maintenance information sample.

[0012] Optionally, the debugging of the deep learning algorithm based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight corresponding to the power operation and maintenance information sample includes: Performing full-connection processing on the first device operation indicator vector sample through the second full-connection module in the deep learning algorithm to obtain a device operation indicator full-connection training vector corresponding to the first device operation indicator vector sample; The deep learning algorithm is debugged based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight, as well as the equipment operation indicator full connection training vector and the equipment operation indicator full connection authentication vector corresponding to the power operation and maintenance information sample.

[0013] An embodiment of the present invention also provides an electric power operation and maintenance information mining system, comprising a processor and a memory and a bus connected to the processor; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the above-mentioned electric power operation and maintenance information mining method based on audit business.

[0014] An embodiment of the present invention further provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the above-mentioned power operation and maintenance information mining method based on audit business.

[0015] The embodiment of the present invention provides a method and system for mining electric power operation and maintenance information based on audit services, which can comprehensively and accurately integrate operation and maintenance information by distinguishing between structured and unstructured parts in the target electric power operation and maintenance information. Operation and maintenance inspection semantic vectors and equipment operation indicator vectors are mined for different types of information respectively, and the value of each information is fully mined. After vector interaction, the operation and maintenance audit decision vector is obtained, and then the fault weight is obtained through discriminant analysis to achieve deep fusion and analysis of multi-source information. This method accurately judges abnormal operation and maintenance information, improves operation and maintenance efficiency and accuracy, reduces the workload of manual troubleshooting, effectively reduces the risk of power system failures, enhances the intelligence and reliability of power operation and maintenance, and ensures stable power supply. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 A flowchart of a method for mining power operation and maintenance information based on audit services provided in an embodiment of the present invention.

[0018] Figure 2 A block diagram of a power operation and maintenance information mining system provided by an embodiment of the present invention.

[0019] Reference numerals: 100-Power operation and maintenance information mining system; 101 - processor; 102 - memory; 103 - bus. DETAILED DESCRIPTION

[0020] The exemplary embodiments disclosed in the embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the embodiments of the present invention are shown in the accompanying drawings, it should be understood that the embodiments of the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the embodiments of the present invention and to fully convey the scope of the embodiments of the present invention to those skilled in the art.

[0021] In order to better understand the above technical scheme, the technical scheme of the embodiment of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiment of the present invention and the specific features in the embodiments are detailed descriptions of the technical scheme of the embodiment of the present invention, rather than limitations on the technical scheme of the embodiment of the present invention. In the absence of conflict, the embodiment of the present invention and the technical features in the embodiments can be combined with each other.

[0022] Figure 1 This is a flowchart of a method for mining electric power operation and maintenance information based on audit services according to an embodiment of the present invention, which is applied to an electric power operation and maintenance information mining system and includes steps 110 to 140.

[0023] Step 110: Obtain first power operation and maintenance information and second power operation and maintenance information corresponding to the target power operation and maintenance information, the first power operation and maintenance information includes an operation and maintenance information set that does not contain structured data in the target power operation and maintenance information, and the second power operation and maintenance information includes an operation and maintenance information set that contains structured data in the target power operation and maintenance information.

[0024] Step 120: Perform data mining on the first power operation and maintenance information to obtain an operation and maintenance inspection semantic vector corresponding to the first power operation and maintenance information; perform data mining on the second power operation and maintenance information to obtain an equipment operation indicator vector corresponding to the second power operation and maintenance information.

[0025] Step 130: Perform vector interaction on the equipment operation index vector and the operation and maintenance inspection semantic vector to obtain the operation and maintenance inspection decision vector corresponding to the target power operation and maintenance information; perform discriminant analysis on the operation and maintenance inspection decision vector to obtain the power operation and maintenance inspection fault weight corresponding to the target power operation and maintenance information.

[0026] Step 140: In response to the power operation and maintenance audit fault weight falling into a set weight interval, the target power operation and maintenance information is determined as abnormal operation and maintenance information.

[0027] In step 110, the power operation and maintenance information mining system first processes the target power operation and maintenance information. This information covers various types of data in the power system operation and maintenance process. For example, the operation and maintenance information of a medium-sized power distribution network may include substation equipment operation data, transmission line inspection records, power equipment fault maintenance records, and other data within a certain period of time (such as one month).

[0028] The target power operation and maintenance information may be divided into two parts: first power operation and maintenance information and second power operation and maintenance information.

[0029] The first power operation and maintenance information is an operation and maintenance information set that does not contain structured data in the target power operation and maintenance information. In the power operation and maintenance scenario, this part of information may include the description of the appearance of the equipment recorded by the inspectors on site (such as the transformer has no obvious damage on the appearance, but has slight rust), the description of the on-site environment when the fault occurred (such as trees falling near the substation, which may affect the power transmission line), etc. This information is often described in natural language and lacks a clear structured format. For example, an inspector recorded that "when inspecting the 10kV line L1, it was found that some branches along the line were close to the line, which may pose a safety hazard." This is typical first power operation and maintenance information.

[0030] The second power operation and maintenance information is an operation and maintenance information set containing structured data in the target power operation and maintenance information. This part of the information has a clear structure and format, and can usually be represented by a table or a specific data structure. Taking the operation data of substation equipment as an example, it may contain numerical data such as the voltage value, current value, and power factor of the equipment, and these data are recorded at a certain time interval (such as once every 15 minutes). For example, for a transformer, there may be such structured data: at 10:00 on May 1, 2023, the voltage is 10.5kV, the current is 50A, the power factor is 0.9, etc.

[0031] In step 120, for the first power operation and maintenance information, since it is unstructured data, the power operation and maintenance information mining system can use natural language processing technology to perform data mining. For example, a word vector model, such as a Word2Vec model, can be used. For example, we have a corpus containing a large amount of power operation and maintenance related texts, and this corpus contains various inspection records, fault descriptions and other texts. In the Word2Vec model, each word is mapped to a vector space of a fixed dimension (for example, 100 dimensions).

[0032] Take the inspection record "When inspecting the 10kV line L1, it was found that some tree branches along the line were close to the line, which may pose a safety hazard" as an example. First, the sentence is segmented to obtain words such as "at", "to", "10kV", "line", "L1", "inspection", "when", "found", "along the line", "some", "tree branches", "close to", "line", "may", "exist", "safety hazard", etc. Then, according to the Word2Vec model, each word can get a 100-dimensional vector.

[0033] Next, these word vectors can be combined into a sentence vector by some methods, such as simple averaging or weighted averaging. If the simple averaging method is used, all word vectors are added together and then divided by the number of words to obtain the operation and maintenance inspection semantic vector corresponding to the sentence. For example, the operation and maintenance inspection semantic vector obtained after processing this sentence is (v_1=(x_1, x_2, cdots, x_{100})).

[0034] For the second power operation and maintenance information, since it is structured data, data standardization and feature engineering methods can be used for mining. For example, for the voltage, current, power factor and other data of the substation equipment, data standardization processing is required first. For example, the voltage value range is (9kV-11kV), the current value range is (30A-70A), and the power factor range is (0.8-1).

[0035] The Z-score standardization method can be used. For the voltage value (V), the standardization formula is (Z_V=frac{V-mu_V}{sigma_V}), where (mu_V) is the mean value of the voltage value and (sigma_V) is the standard deviation of the voltage value. The current and power factor are also standardized.

[0036] Then, these standardized values ​​are combined into an equipment operation index vector. For example, after standardization, the standardized value corresponding to the voltage value is (z_1), the standardized value corresponding to the current value is (z_2), and the standardized value corresponding to the power factor is (z_3), then the equipment operation index vector is (v_2=(z_1, z_2, z_3)).

[0037] In step 130, the power operation and maintenance information mining system interacts the equipment operation index vector (v_2=(z_1, z_2, z_3)) and the operation and maintenance inspection semantic vector (v_1=(x_1, x_2, cdots, x_{100})). A common vector interaction method is to concatenate the two vectors to obtain a new vector (v=(x_1, x_2, cdots, x_{100}, z_1, z_2, z_3)), which is the operation and maintenance inspection decision vector.

[0038] For the obtained operation and maintenance audit decision vector (v), the linear discriminant analysis (LDA) method can be used for discriminant analysis. For example, we have a known sample set of normal and faulty power operation and maintenance, of which there are (n_1=100) normal samples and (n_2=50) faulty samples.

[0039] First, the mean vector of each class is calculated, the mean vector of normal samples (mu_1) and the mean vector of faulty samples (mu_2), as well as the intra-class scatter matrix (S_w) and the inter-class scatter matrix (S_b).

[0040] According to the formula of linear discriminant analysis (J(w)=frac{w^TS_bw}{w^TS_ww}), find the projection vector (w) that maximizes (J(w)). Then project the operation and maintenance audit decision vector (v) onto (w) to obtain a scalar value (y=w^Tv).

[0041] According to the classification of known samples, a classification function (f(y)) is established. For example, it can be a simple threshold function. If (y>t) ((t) is the threshold), it is judged as normal, otherwise it is judged as a fault. At the same time, according to the distribution of samples, the fault weight of each sample can be calculated. For the operation and maintenance inspection decision vector (v) corresponding to the target power operation and maintenance information, the corresponding power operation and maintenance inspection fault weight (w_f) is obtained after discriminant analysis.

[0042] Finally, in step 140, the weight interval is set based on historical data and operation and maintenance experience. For example, through the analysis of a large amount of power operation and maintenance data, it is found that when the power operation and maintenance audit fault weight is less than (0.3), the probability of normal operation of the power system is high; when the fault weight is greater than (0.7), the power system is likely to have a fault; and when the fault weight is between (0.3-0.7), further analysis is required. Therefore, the weight interval can be set to ((0, 0.3)) as the normal interval, ((0.7, 1)) as the fault interval, and ((0.3, 0.7)) as the interval to be further analyzed.

[0043] When the power operation and maintenance information mining system obtains the power operation and maintenance inspection fault weight (w_f) corresponding to the target power operation and maintenance information, if (w_f) falls into the set fault interval (such as ((0.7, 1))), the target power operation and maintenance information is determined as abnormal operation and maintenance information. This means that in this case, the power system may have a fault or potential risk, and further operation and maintenance operations are required, such as detailed inspection, maintenance or adjustment of related equipment.

[0044] In summary, the embodiments of the present invention can comprehensively and accurately integrate operation and maintenance information by distinguishing between structured and unstructured parts in the target power operation and maintenance information. Operation and maintenance inspection semantic vectors and equipment operation indicator vectors are mined for different types of information respectively, and the value of each information is fully mined. After vector interaction, the operation and maintenance inspection decision vector is obtained, and then the fault weight is obtained through discriminant analysis to achieve deep fusion and analysis of multi-source information. This method accurately judges abnormal operation and maintenance information, improves operation and maintenance efficiency and accuracy, reduces the workload of manual troubleshooting, effectively reduces the risk of power system failures, enhances the intelligence and reliability of power operation and maintenance, and ensures stable power supply.

[0045] In an optional embodiment, the obtaining of the first power operation and maintenance information and the second power operation and maintenance information corresponding to the target power operation and maintenance information includes: performing structured feature recognition on the target power operation and maintenance information to obtain a first operation and maintenance information set that does not contain structured data in the target power operation and maintenance information, and performing information splitting on the first operation and maintenance information set in the target power operation and maintenance information as the first power operation and maintenance information; based on the first operation and maintenance information set, obtaining a second operation and maintenance information set that contains structured data in the target power operation and maintenance information, and performing information splitting on the second operation and maintenance information set in the target power operation and maintenance information as the second power operation and maintenance information.

[0046] Alternatively, in another optional embodiment, the obtaining of the first power operation and maintenance information and the second power operation and maintenance information corresponding to the target power operation and maintenance information includes: performing structured feature recognition on the target power operation and maintenance information to obtain a third operation and maintenance information set containing structured data in the target power operation and maintenance information, and performing information splitting on the third operation and maintenance information set in the target power operation and maintenance information as the second power operation and maintenance information; based on the third operation and maintenance information set, obtaining a fourth operation and maintenance information set that does not contain structured data in the target power operation and maintenance information, and performing information splitting on the fourth operation and maintenance information set in the target power operation and maintenance information as the first power operation and maintenance information.

[0047] In the first optional embodiment, the power operation and maintenance information mining system first performs structured feature recognition when processing the target power operation and maintenance information. Structured feature recognition aims to distinguish the characteristics of structured data and unstructured data in the target power operation and maintenance information. For power operation and maintenance information, structured data usually has a clear format, such as numerical data recorded at specific time intervals (for example, the voltage and current values ​​of substation equipment are recorded every 5 minutes), and these data can often be presented in a table. Unstructured data is mostly text descriptions, such as the inspection personnel's text description of the appearance of the equipment (such as "the transformer casing has slight corrosion marks").

[0048] When performing structured feature recognition, a pattern matching algorithm can be used. For example, a pattern is defined. For numerical data, its format may be a number plus a specific unit (such as "10kV" or "50A"), and it may have a fixed field name (such as "voltage" or "current") in the data storage. For time series data, it may have a specific time format (such as "YYYY-MM-DDHH:MM:SS"). By searching for data that meets these patterns in the target power operation and maintenance information, data that does not meet these patterns can be preliminarily determined as unstructured data.

[0049] After structured feature recognition, the first operation and maintenance information set that does not contain structured data in the target power operation and maintenance information is obtained. For example, in a target information set containing 1,000 power operation and maintenance records, after structured feature recognition, 300 records may be determined as the first operation and maintenance information set, which are mainly unstructured information such as text descriptions of inspection personnel and descriptions of the environment when the fault occurred.

[0050] The purpose of information splitting of the first operation and maintenance information set is to further process and mine the information therein. Information splitting can be performed based on semantic units. For example, for an inspection record "When inspecting the transformer of substation A, it was found that its oil temperature was slightly high and dust accumulated on the heat sink", it can be split into units such as "substation A", "transformer", "inspection", "slightly high oil temperature", "heat sink", and "dust accumulation" according to semantic units.

[0051] These split units will be used as the first power operation and maintenance information. The split first power operation and maintenance information is more convenient for subsequent data mining. For example, when constructing the operation and maintenance inspection semantic vector, operations such as word vector conversion can be performed on these semantic units.

[0052] After obtaining the first operation and maintenance information set, the second operation and maintenance information set is obtained based on it. Since the first operation and maintenance information set is an unstructured data part, the second operation and maintenance information set is the structured data part of the target power operation and maintenance information except the first operation and maintenance information set. For example, among the above 1000 power operation and maintenance records, the remaining 700 records are the second operation and maintenance information set.

[0053] For the second operation and maintenance information set, the information is also split. The information splitting of structured data can be based on data fields. For example, for the operation data record of substation equipment, it may contain fields such as "equipment name", "time", "voltage value", "current value", and "power factor". The value of each field is used as the unit after splitting, and the set of these units is the second power operation and maintenance information. For example, for a record "2023-05-0110:00, substation B, 10.5kV, 50A, 0.9", after splitting, the second power operation and maintenance information units such as "2023-05-0110:00", "substation B", "10.5kV", "50A", and "0.9" are obtained.

[0054] As mentioned above, for the first power operation and maintenance information, natural language processing technology is used for data mining. Take the word vector model as an example, for example, the Skip-Gram model is used. The goal of the Skip-Gram model is to predict context words based on the central word. For each semantic unit in the first power operation and maintenance information (such as "the oil temperature is slightly high"), it is regarded as a word.

[0055] The objective function of the Skip-Gram model is: (J=frac{1}{T}sum_{t=1}^{T}sum_{-cleqjleqc,jneq0}logp(w_{t+j}|w_t)), where (T) is the total number of words in the corpus, (c) is the context window size, (w_t) is the center word, and (w_{t+j}) are the context words. (p(w_{t+j}|w_t)=frac{exp(v_{w_{t+j}}^Tv_{w_t})}{sum_{w=1}^{W}exp(v_w^Tv_{w_t})}), where (v_w) is the word vector of word (w) and (W) is the vocabulary size.

[0056] By training the Skip-Gram model for all semantic units in the first power operation and maintenance information, the word vector of each semantic unit can be obtained. Then the weighted average method is used to construct the operation and maintenance inspection semantic vector. For example, the word vector of the semantic unit "the oil temperature is slightly high" is (v_1=(x_1, x_2, cdots, x_{100})), and the word vector of "heat sink" is (v_2=(y_1, y_2, cdots, y_{100})). If the weight of "the oil temperature is slightly high" is (a=0.6) and the weight of "heat sink" is (b=0.4) (the weight can be set according to semantic importance or experience), then the operation and maintenance inspection semantic vector is (v=(atimesv_1+btimesv_2)=(z_1, z_2, cdots, z_{100})).

[0057] For the second power operation and maintenance information, principal component analysis (PCA) is used for data mining. For example, the second power operation and maintenance information contains (n) data fields (such as (n=5), which are voltage, current, power factor, equipment operation time, and equipment load rate), and has (m) records (such as (m=1000)).

[0058] First, construct the data matrix: (X=begin{bmatrix}x_{11}&x_{12}&cdots&x_{1n}x_{21}&x_{22}&cdots&x_{2n}vdots&vdots&ddots&vdotsx_{m1}&x_{m2}&cdots&x_{mn}end{bmatrix}), where (x_{ij}) represents the value of the (j)th data field in the (i)th record.

[0059] Calculate the covariance matrix (C=frac{1}{m-1}X^TX) of the data matrix (X). Then solve the eigenvalues ​​(lambda_1, lambda_2, cdots, lambda_n) and corresponding eigenvectors (e_1, e_2, cdots, e_n) of the covariance matrix (C). Sort by eigenvalue size and select the eigenvectors (e_{i_1}, e_{i_2}, e_{i_3}) corresponding to the first (k) (e.g. (k=3)) largest eigenvalues.

[0060] Project the original data matrix (X) onto these (k) feature vectors to obtain the device operation index vector. For example, for a record (x=(x_1, x_2, x_3, x_4, x_5)), its device operation index vector (v=(e_{i_1}^Tx, e_{i_2}^Tx, e_{i_3}^Tx)).

[0061] Vector interaction: Concatenate the equipment operation indicator vector (v_2=(z_1, z_2, z_3)) and the operation and maintenance inspection semantic vector (v_1=(x_1, x_2, cdots, x_{100})) to obtain the operation and maintenance inspection decision vector (v=(x_1, x_2, cdots, x_{100}, z_1, z_2, z_3)).

[0062] Discriminant analysis: Support vector machine (SVM) is used for discriminant analysis. For a binary classification problem (normal and faulty), the goal of SVM is to find a hyperplane (w^Tx+b=0) that maximizes the distance between the two types of data points and the hyperplane.

[0063] For the training data ((x_i, y_i)), (i=1, 2, cdots, m) ((y_iin{-1, 1}) represents the category), the optimization problem of SVM is (min_{w, b}frac{1}{2}w^Tw+Csum_{i=1}^{m}xi_i), subjectto(y_i(w^Tx_i+b)geq1-xi_i), (xi_igeq0), where (C) is the penalty parameter and (xi_i) is the slack variable.

[0064] By solving this optimization problem, we get (w) and (b). Then substitute the operation and maintenance inspection decision vector (v) into (f(v)=w^Tv+b), determine its category according to the value of (f(v)), and calculate the power operation and maintenance inspection fault weight. For example, if (f(v)>0), it is judged as normal, and the fault weight is (w_f=0.1); if (f(v)<0), it is judged as a fault, and the fault weight is (w_f=0.9).

[0065] In a second optional embodiment, similarly, the electric power operation and maintenance information mining system performs structured feature recognition on the target electric power operation and maintenance information. However, in this embodiment, the third operation and maintenance information set containing structured data in the target electric power operation and maintenance information is first identified. The recognition process is similar to the previous one, and the data part with a clear format is identified through a pattern matching algorithm. For example, for the telemetry data of power equipment (such as real-time values ​​of voltage and current), which has a fixed numerical format and time stamp, these data are identified as the third operation and maintenance information set. For example, in a target information set containing 800 electric power operation and maintenance records, after structured feature recognition, 400 records are determined to be the third operation and maintenance information set.

[0066] The third operation and maintenance information set is split. For the information splitting of structured data, taking the monitoring data of substation equipment as an example, it may include fields such as equipment number, collection time, and measurement value. The value of each field is split out, and the set of these split values ​​is the second power operation and maintenance information. For example, for a monitoring record "equipment number 1001, 2023-05-0211:00, 10.3kV", after splitting, the second power operation and maintenance information units such as "1001", "2023-05-0211:00", and "10.3kV" are obtained.

[0067] After obtaining the third operation and maintenance information set, the fourth operation and maintenance information set is obtained based on it. The fourth operation and maintenance information set is the unstructured data part of the target power operation and maintenance information except the third operation and maintenance information set. Among the above 800 power operation and maintenance records, the remaining 400 records are the fourth operation and maintenance information set.

[0068] The fourth operation and maintenance information set is split. The information splitting is based on semantic units. For example, for an inspection record "a strange smell was found in the switch cabinet of substation C", it can be split into units such as "substation C", "switch cabinet", and "smell". These split units constitute the first power operation and maintenance information.

[0069] A language model based on a neural network, such as a long short-term memory network (LSTM), is used to process the first power operation and maintenance information. For example, the text sequence of the first power operation and maintenance information is (x=(x_1, x_2, cdots, x_T)), where (T) is the sequence length.

[0070] The calculation process of the LSTM unit is as follows: first calculate the input gate (i_t=sigma(W_{xi}x_t+W_{hi}h_{t-1}+b_i)), the forget gate (f_t=sigma(W_{xf}x_t+W_{hf}h_{t-1}+b_f)), and the output gate (o_t=sigma(W_{xo}x_t+W_{ho}h_{t-1}+b_o)), where (sigma) is the sigmoid function, (W) is the weight matrix, (b) is the bias term, and (h_{t-1}) is the hidden state of the previous moment.

[0071] Then calculate the cell state (c_t=f_tc_{t-1}+i_ttanh(W_{xc}x_t+W_{hc}h_{t-1}+b_c)) and the hidden state (h_t=o_ttanh(c_t)).

[0072] By performing LSTM processing on the entire first electric power operation and maintenance information text sequence, a hidden state vector of a fixed dimension (for example, 200 dimensions) can be obtained as an operation and maintenance inspection semantic vector (v=(y_1, y_2, cdots, y_{200})).

[0073] The cluster analysis method is used to mine the second power operation and maintenance information. For example, for the operation data of substation equipment (such as voltage, current, power factor, etc.), the k-means clustering algorithm is used. For example (k=3), the data is divided into 3 categories.

[0074] First, randomly select (k) initial cluster centers (mu_1, mu_2, mu_3). Then for each data point (xi), calculate its distance to each cluster center (d(xi, mu_j) = sqrt{sum_{l=1}^{n}(xi_{il}-mu_{jl})^2}), where (n) is the dimension of the data (such as for voltage, current, power factor, (n=3)).

[0075] Assign the data point to the class to which the nearest cluster center belongs. Then recalculate the cluster center of each class (mu_j=frac{1}{|C_j|}sum_{x_iinC_j}x_i), where (C_j) is the set of data points in the (j)th class.

[0076] Repeat this process until the cluster center no longer changes. Finally, according to the cluster category to which the data point belongs, it is mapped to a low-dimensional space (for example, 3 dimensions) to obtain the equipment operation index vector (v=(z_1, z_2, z_3)).

[0077] Vector interaction: Concatenate the equipment operation indicator vector (v_2=(z_1, z_2, z_3)) and the operation and maintenance inspection semantic vector (v_1=(y_1, y_2, cdots, y_{200})) to obtain the operation and maintenance inspection decision vector (v=(y_1, y_2, cdots, y_{200}, z_1, z_2, z_3)).

[0078] Discriminant analysis: Decision tree algorithm is used for discriminant analysis. Decision tree constructs a classification model by dividing data features. For example, for the operation and maintenance inspection decision vector (v), a decision tree is constructed based on the elements in the vector (such as the values ​​corresponding to certain semantic units in the operation and maintenance inspection semantic vector and the values ​​in the equipment operation indicator vector).

[0079] For example, a node judgment condition of the decision tree is "If the value corresponding to the 'smell' semantic unit in the operation and maintenance inspection semantic vector is greater than 0 and the voltage value in the equipment operation index vector is less than 10kV, it is judged as a fault." The power operation and maintenance inspection fault weight is calculated based on the classification results of the decision tree. For example, if it is judged as a fault, the fault weight is (w_f=0.8); if it is judged to be normal, the fault weight is (w_f=0.2).

[0080] The above technical solution proposes two ways to obtain the first power operation and maintenance information and the second power operation and maintenance information. These two ways provide flexible processing means for the mining of power operation and maintenance information. In the data mining process, a variety of suitable algorithms are used for different types of power operation and maintenance information (structured and unstructured). For example, the Skip-Gram model, LSTM and other natural language processing algorithms are used for the unstructured first power operation and maintenance information, which can effectively extract semantic information from the text description; PCA, cluster analysis and other algorithms are used for the structured second power operation and maintenance information, which can mine the inherent structural characteristics of the data. Through vector interaction and discriminant analysis, structured and unstructured information can be comprehensively considered to accurately calculate the power operation and maintenance inspection fault weight, so as to effectively determine whether the power operation and maintenance information is abnormal, which helps to improve the efficiency and accuracy of power operation and maintenance, reduce the risk of power system failure, and ensure the stable operation of the power system.

[0081] In the actual application process, the above technical solution can be implemented based on a deep learning algorithm. In detail, the data mining of the first power operation and maintenance information to obtain the operation and maintenance inspection semantic vector corresponding to the first power operation and maintenance information includes: through the first data mining module in the deep learning algorithm, the data mining of the first power operation and maintenance information is performed to obtain the operation and maintenance inspection semantic vector corresponding to the first power operation and maintenance information. The data mining of the second power operation and maintenance information to obtain the equipment operation index vector corresponding to the second power operation and maintenance information includes: through the second data mining module in the deep learning algorithm, the data mining of the second power operation and maintenance information is performed to obtain the equipment operation index vector corresponding to the second power operation and maintenance information. The vector interaction of the equipment operation index vector and the operation and maintenance inspection semantic vector to obtain the operation and maintenance audit decision vector corresponding to the target power operation and maintenance information includes: through the vector interaction module in the deep learning algorithm, the vector interaction of the equipment operation index vector and the operation and maintenance inspection semantic vector is performed to obtain the operation and maintenance audit decision vector corresponding to the target power operation and maintenance information. The discriminant analysis is performed on the operation and maintenance audit decision vector to obtain the power operation and maintenance audit fault weight corresponding to the target power operation and maintenance information, including: through the discriminant analysis module in the deep learning algorithm, the discriminant analysis is performed on the operation and maintenance audit decision vector to obtain the power operation and maintenance audit fault weight corresponding to the target power operation and maintenance information.

[0082] In actual application, the technical solution based on deep learning algorithm is implemented as follows.

[0083] 1. Mining of the First Power Operation and Maintenance Information by the First Data Mining Module 1) Selection and principle of the first data mining module: For the first power operation and maintenance information, since it is unstructured data, such as text descriptions of inspection personnel, etc., a suitable deep learning algorithm is selected to construct the first data mining module. For example, the text processing method of convolutional neural network (CNN) can be used. In text processing, CNN extracts features by sliding the convolution kernel on the text vector.

[0084] For example, after preprocessing (such as word segmentation, vectorization, etc.), the first power operation and maintenance information obtains a matrix with a shape of ((n, d)), where (n) is the length of the text sequence (for example, 100 words) and (d) is the dimension of each word vector (for example, 100 dimensions).

[0085] Define a convolution kernel size as ((k, d)), where (k) is the window size of the convolution kernel (for example, (k=3)). The calculation formula for the convolution operation is (c_{i}=tanh(sum_{j=1}^{d}w_jx_{i:j+k-1}+b)), where (w_j) is the weight of the convolution kernel, (x_{i:j+k-1}) is the submatrix from the (i)th position to the (i+k-1)th position in the input matrix, (b) is the bias term, and (tanh) is the activation function.

[0086] By performing convolution operations with multiple different convolution kernels, multiple feature maps can be obtained. For example, using 10 different convolution kernels will result in 10 feature maps, each of which has a size of ((n-k+1, 1)).

[0087] 2) Constructing the operation and maintenance inspection semantic vector: Perform pooling operations on the obtained feature maps, such as maximum pooling. The maximum pooling operation selects the maximum value from each feature map as the representative value of the feature map. For example, after maximum pooling, a vector with a length of 10 is obtained (because there are 10 feature maps).

[0088] This vector is then mapped through a fully connected layer, the weight matrix of the fully connected layer is (W_f) (for example, the shape of (W_f) is ((10, m)), where (m) is the dimension of the final operation and maintenance inspection semantic vector, for example (m=50)), and the bias term is (b_f). The calculation formula is (y=tanh(W_fx+b_f)), where (x) is the vector after pooling. The obtained (y) is the operation and maintenance inspection semantic vector corresponding to the first power operation and maintenance information, with a shape of ((m,)) or ((50,)).

[0089] (II) Mining of the Second Power Operation and Maintenance Information by the Second Data Mining Module 1) Selection and principle of the second data mining module: For the second power operation and maintenance information, since it is structured data, such as numerical data such as voltage and current of equipment, a deep neural network (DNN) can be used to construct the second data mining module.

[0090] For example, the second power operation and maintenance information contains (p) features (such as voltage, current, power factor, (p=3)), and has (q) data records (for example, (q=1000)). First, the data is normalized. For example, for the voltage value (V), the minimum-maximum normalization formula (V_{norm}=frac{V-V_{min}}{V_{max}-V_{min}}) is used, where (V_{min}) and (V_{max}) are the minimum and maximum values ​​of the voltage value.

[0091] Construct a deep neural network with (l) hidden layers (e.g., (l=2)). For the connection from the input layer to the first hidden layer, the weight matrix is ​​(W_1) (e.g., the shape of (W_1) is ((p, n_1)), where (n_1) is the number of neurons in the first hidden layer, e.g., (n_1=10)), and the bias term is (b_1). The calculation formula is (h_1=tanh(W_1x+b_1)), where (x) is the normalized second power operation and maintenance information vector.

[0092] For the connection from the first hidden layer to the second hidden layer, the weight matrix is ​​(W_2) (for example, the shape of (W_2) is ((n_1, n_2)), where (n_2) is the number of neurons in the second hidden layer, for example (n_2=5)), and the bias term is (b_2). The calculation formula is (h_2=tanh(W_2h_1+b_2)).

[0093] 2) Construct the equipment operation index vector: The connection from the last hidden layer to the output layer, the weight matrix is ​​(W_o) (for example, the shape of (W_o) is ((n_2, r)), where (r) is the dimension of the equipment operation index vector, for example (r=3)), and the bias term is (b_o). The calculation formula is (y=tanh(W_oh_2+b_o)), and the obtained (y) is the equipment operation index vector corresponding to the second power operation and maintenance information, with a shape of ((r,)) or ((3,)).

[0094] 3. Operation of the vector interaction module 1) Principle and operation of the vector interaction module: The vector interaction module receives the operation and maintenance inspection semantic vector from the first data mining module and the equipment operation indicator vector from the second data mining module.

[0095] For example, the operation and maintenance inspection semantic vector is (v_1=(x_1, x_2, cdots, x_{50})), and the equipment operation indicator vector is (v_2=(y_1, y_2, y_3)). The vector interaction module uses a simple splicing operation to combine the two vectors into a new vector. The new operation and maintenance inspection decision vector (v=(x_1, x_2, cdots, x_{50}, y_1, y_2, y_3)) has a dimension of (50+3=53).

[0096] (IV) Operation of the discriminant analysis module 1) Selection and principle of the discriminant analysis module: The discriminant analysis module uses a multi-layer perceptron (MLP) for discriminant analysis. MLP is a feedforward neural network.

[0097] For the input operation and maintenance audit decision vector (v), for example, the weight matrix from the input layer to the first hidden layer is (W_d1) (for example, the shape is ((53, m_1)), where (m_1=20)), and the bias term is (b_d1). The calculation formula is (h_d1=tanh(W_d1v+b_d1)).

[0098] The weight matrix from the first hidden layer to the second hidden layer is (W_d2) (for example, the shape is ((m_1, m_2)), where (m_2=10)), and the bias term is (b_d2). The calculation formula is (h_d2=tanh(W_d2h_d1+b_d2)).

[0099] The connection from the last hidden layer to the output layer has only one neuron, the weight matrix is ​​(W_do) (for example, the shape is ((m_2, 1))), and the bias term is (b_do). The calculation formula is (z=tanh(W_doh_d2+b_do)).

[0100] The obtained (z) is a value between ((-1, 1)), which is mapped to the interval ((0, 1)) as the fault weight of power operation and maintenance inspection. For example, the mapping formula (w_f=frac{z+1}{2}) can be used. If (w_f) is greater than a certain threshold (for example, the threshold is (0.5)), it is determined that there may be a fault; if (w_f) is less than or equal to (0.5), it is determined to be normal.

[0101] It can be seen that this embodiment constructs various modules based on deep learning algorithms to implement the technical solution, which has many beneficial effects. For the first power operation and maintenance information, the data mining module constructed by CNN can effectively extract semantic features in the text and convert unstructured data into meaningful operation and maintenance inspection semantic vectors. For the second power operation and maintenance information, the data mining module constructed by DNN can process structured data and mine the complex relationship between data to obtain equipment operation indicator vectors. The vector interaction module simply and effectively combines the two types of vectors, taking into account various aspects of operation and maintenance information. The discriminant analysis module uses MLP for discriminant analysis, which can accurately calculate the fault weights of power operation and maintenance inspections, which helps to accurately determine the power operation and maintenance status, improve the intelligence level, accuracy and efficiency of power operation and maintenance, and ensure the stable operation of the power system.

[0102] In one possible example, in response to the power operation and maintenance audit fault weight falling into a set weight interval, the target power operation and maintenance information is determined as abnormal operation and maintenance information, including: on the basis that the power operation and maintenance audit fault weight is greater than a first set fault weight, the target power operation and maintenance information is determined as abnormal operation and maintenance information.

[0103] In another possible example, in response to the power operation and maintenance audit fault weight falling into a set weight interval, the target power operation and maintenance information is determined as abnormal operation and maintenance information, including: on the basis that the power operation and maintenance audit fault weight is less than a second set fault weight, the target power operation and maintenance information is determined as abnormal operation and maintenance information.

[0104] In the embodiment of the present invention, the exemplary introduction of the above two examples is described as follows.

[0105] (I) The power operation and maintenance inspection fault weight is greater than the first set fault weight to determine abnormal operation and maintenance information 1) Basis for determining the first set fault weight: In the power operation and maintenance system, the first set fault weight is determined based on a large amount of historical operation and maintenance data, equipment operation characteristics, and operation and maintenance experience. For example, through the analysis of operation and maintenance information related to the past 1,000 power equipment failure events, it was found that when the power operation and maintenance audit fault weight reaches a certain value, the probability of equipment failure or serious fault hidden dangers increases significantly. For example, this value is determined to be 0.7 after analysis and statistical calculation. This 0.7 is the first set fault weight.

[0106] 2) Judgment logic and principle: After obtaining the power operation and maintenance inspection fault weight corresponding to the target power operation and maintenance information through the discriminant analysis module in the previous deep learning algorithm, the following judgment is made. If this power operation and maintenance inspection fault weight is greater than 0.7 (the first set fault weight), the target power operation and maintenance information is determined as abnormal operation and maintenance information.

[0107] The principle behind this judgment is that in previous in-depth studies of power operation and maintenance data, it was found that when the fault weight reaches 0.7 or above, the equipment in the power system is often in a state that deviates more from normal operation. For example, for substation equipment, it may be the key operating indicators of the equipment (such as excessive voltage fluctuations, current overload, etc.), which are reflected in the operation and maintenance audit decision vector after discriminant analysis. The result is a higher fault weight. From an algorithmic point of view, in the discriminant analysis module, such as the result (z) output by the multi-layer perceptron (MLP) is obtained by the mapping formula (w_f=frac{z+1}{2}) to obtain (w_f). If (w_f>0.7), it means that under the previously constructed discriminant model, various operation and maintenance information (including the information in the operation and maintenance inspection semantic vector and the equipment operation indicator vector) are combined to indicate that the equipment is more likely to fail or is already in a faulty state.

[0108] Take a specific device operation data as an example. For example, when processing the operation and maintenance information of a certain transmission line, the operation and maintenance inspection semantic vector contains the description information that the inspectors found that there were leaning trees near the line that may affect the safety of the line, and the equipment operation index vector contains the data that the line current has a slight fluctuation (close to the overload critical value). The power operation and maintenance inspection fault weight obtained after vector interaction and discriminant analysis is 0.8. Since 0.8>0.7 (the first set fault weight), the target power operation and maintenance information of this transmission line is determined to be abnormal operation and maintenance information, and further operation and maintenance measures need to be taken, such as arranging personnel to deal with trees near the line, monitoring and adjusting the line current, etc.

[0109] (II) The power operation and maintenance inspection fault weight is less than the second set fault weight to determine abnormal operation and maintenance information 1) Basis for determining the second set fault weight: The second set fault weight is also determined based on the historical data of power operation and maintenance, equipment operation specifications, and related operation and maintenance knowledge. For example, after analyzing the operation and maintenance information of a large number of normally operating equipment, it is found that when the power operation and maintenance audit fault weight is lower than a certain value, although the equipment seems to be operating normally, there may actually be some hidden and not yet fully identified abnormal conditions. For example, this value is determined to be 0.3 after detailed analysis and verification. This 0.3 is the second set fault weight.

[0110] 2) Judgment logic and principle: After obtaining the power operation and maintenance inspection fault weight, if this weight is less than 0.3 (the second set fault weight), the target power operation and maintenance information is also determined as abnormal operation and maintenance information. This is because in some cases, a lower fault weight may imply abnormal data, which may be due to a failure of the data collection equipment, a deviation in the construction of the operation and maintenance inspection semantic vector, or unreasonable data in the equipment operation indicator vector.

[0111] For example, in the operation and maintenance information processing of a substation equipment, the operation and maintenance inspection semantic vector may contain some unclear or inconsistent descriptions with the actual situation (such as the inspector mistakenly recorded that the equipment looks normal but there are actually small cracks), and the data in the equipment operation index vector may show that everything is normal due to sensor failure. The power operation and maintenance inspection fault weight obtained after vector interaction and discriminant analysis is 0.2. Since 0.2<0.3 (the second set fault weight), the target power operation and maintenance information of this substation equipment is judged as abnormal operation and maintenance information. This requires a re-examination of the data, such as re-inspecting the equipment, checking the data acquisition sensors, etc., to ensure the normal operation of the equipment.

[0112] It can be seen that determining abnormal operation and maintenance information by setting different fault weights has significant beneficial effects. For the first case where the fault weight is set to be greater than the abnormality determined, it can be discovered in time that the equipment may be in or about to be in a faulty state, which is conducive to quickly taking operation and maintenance measures to avoid further deterioration of equipment failures and affecting the stable operation of the power system. For the second case where the fault weight is set to be less than the abnormality determined, some hidden and possibly overlooked abnormalities can be discovered, which helps to improve the accuracy and reliability of operation and maintenance data, avoid potential risks caused by data deviations or hidden problems, thereby comprehensively improving the quality and efficiency of power operation and maintenance, and ensuring the safe and reliable operation of the power system.

[0113] In a preferred embodiment, before the first data mining module in the deep learning algorithm is used to perform data mining on the first power operation and maintenance information to obtain the operation and maintenance inspection semantic vector corresponding to the first power operation and maintenance information, the method also includes: obtaining a first power operation and maintenance information sample and a second power operation and maintenance information sample corresponding to the power operation and maintenance information sample, the first power operation and maintenance information sample including an operation and maintenance information set that does not contain structured data in the power operation and maintenance information sample, and the second power operation and maintenance information sample including an operation and maintenance information set that contains structured data in the power operation and maintenance information sample; performing data mining on the first power operation and maintenance information sample through the first data mining module in the deep learning algorithm to obtain the operation and maintenance inspection semantic vector sample corresponding to the first power operation and maintenance information sample ; Through the second data mining module in the deep learning algorithm, data mining is performed on the second power operation and maintenance information sample to obtain the first equipment operation indicator vector sample corresponding to the second power operation and maintenance information sample; through the vector interaction module in the deep learning algorithm, vector interaction is performed on the operation and maintenance inspection semantic vector sample and the first equipment operation indicator vector sample to obtain the operation and maintenance inspection decision vector sample corresponding to the power operation and maintenance information sample; through the discriminant analysis module in the deep learning algorithm, discriminant analysis is performed on the operation and maintenance inspection decision vector sample to obtain the power operation and maintenance fault training weight corresponding to the power operation and maintenance information sample; based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight corresponding to the power operation and maintenance information sample, the deep learning algorithm is debugged.

[0114] In the next step, the deep learning algorithm is debugged based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight corresponding to the power operation and maintenance information sample, including: performing full-connection processing on the operation and maintenance inspection semantic vector sample through the first fully-connected module in the deep learning algorithm to obtain the operation and maintenance inspection fully-connected prediction vector corresponding to the operation and maintenance inspection semantic vector sample; debugging the deep learning algorithm based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight, as well as the operation and maintenance inspection fully-connected prediction vector and the operation and maintenance inspection fully-connected authentication vector corresponding to the power operation and maintenance information sample.

[0115] In addition, the debugging of the deep learning algorithm based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight corresponding to the power operation and maintenance information sample includes: performing full-connection processing on the first device operation indicator vector sample through the second fully-connected module in the deep learning algorithm to obtain the device operation indicator fully-connected training vector corresponding to the first device operation indicator vector sample; debugging the deep learning algorithm based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight, as well as the device operation indicator fully-connected training vector and the device operation indicator fully-connected authentication vector corresponding to the power operation and maintenance information sample.

[0116] The technical solution for debugging a deep learning algorithm based on sample data involved in the above embodiment is described in detail as follows.

[0117] 1. Obtaining and classifying power operation and maintenance information samples 1) Acquisition and composition of sample data: First, we need to obtain power operation and maintenance information samples, which are representative data sets of actual power operation and maintenance conditions. The power operation and maintenance information samples contain two parts: the first power operation and maintenance information sample and the second power operation and maintenance information sample. The first power operation and maintenance information sample is an operation and maintenance information set that does not contain structured data in the power operation and maintenance information sample. This part may include the inspection personnel's text description of the equipment appearance, operating environment, etc. For example, descriptions such as "there are oil stains on the surface of the transformer and a small amount of weeds around it." The second power operation and maintenance information sample is an operation and maintenance information set that contains structured data in the power operation and maintenance information sample, such as numerical data such as the voltage value, current value, and power factor of the equipment, in the form of data such as "voltage: 10.5kV, current: 50A, power factor: 0.9".

[0118] 2. Preliminary mining and interaction based on sample data 1) Data mining of the first power operation and maintenance information sample: The first data mining module in the deep learning algorithm is used to perform data mining on the first power operation and maintenance information sample to obtain an operation and maintenance inspection semantic vector sample. For example, if a recurrent neural network (RNN) is used as part of the first data mining module. For the text sequence (x=(x_1, x_2, cdots, x_T)) in the first power operation and maintenance information sample, the calculation formula of the RNN is (h_t=tanh(W_{xh}x_t+W_{hh}h_{t-1}+b_h)), where (W_{xh}) is the weight matrix input to the hidden layer, (W_{hh}) is the weight matrix from the hidden layer to the hidden layer, (b_h) is the bias term, and (h_t) is the hidden state at time (t). After processing the entire text sequence, the final hidden state can be used as an example of an operation and maintenance inspection semantic vector, for example, its dimension is (d_1=100), that is, (v_{s1}=(y_1, y_2, cdots, y_{100})).

[0119] 2) Data mining of the second power operation and maintenance information sample: The second data mining module in the deep learning algorithm is used to perform data mining on the second power operation and maintenance information sample to obtain the first equipment operation index vector sample. For example, a multi-layer perceptron (MLP) is used to construct the second data mining module. If the second power operation and maintenance information sample has (n) features (e.g., (n=3) features: voltage, current, power factor), the input vector is (x=(x_1, x_2, x_3)). For the input layer to the first hidden layer, the weight matrix is ​​(W_1) (e.g., the shape is ((3, m_1)), where (m_1=10)), the bias term is (b_1), and the calculation formula is (h_1=tanh(W_1x+b_1)). After processing through multiple hidden layers, the first equipment operation index vector sample is obtained, for example, its dimension is (d_2=5), that is, (v_{s2}=(z_1, z_2, z_3, z_4, z_5)).

[0120] 3) Vector interaction to obtain the operation and maintenance inspection decision vector sample: Through the vector interaction module in the deep learning algorithm, the operation and maintenance inspection semantic vector sample (v_{s1}) and the first equipment operation indicator vector sample (v_{s2}) are vector-interacted. Here, a simple splicing method is used to obtain the operation and maintenance inspection decision vector sample (v_{s}=(y_1, y_2, cdots, y_{100}, z_1, z_2, z_3, z_4, z_5)), whose dimension is (100+5=105).

[0121] 4) Discriminant analysis to obtain the training weights of power operation and maintenance faults: Use the discriminant analysis module in the deep learning algorithm to perform discriminant analysis on the operation and maintenance audit decision vector sample (v_{s}) to obtain the power operation and maintenance fault training weights corresponding to the power operation and maintenance information sample. For example, the discriminant analysis module uses a support vector machine (SVM). For a binary classification problem (normal and faulty), the goal of the SVM is to find a hyperplane (w^Tx+b=0) that maximizes the interval between the two types of data points and the hyperplane. For the training data ((x_i, y_i)), (i=1,2,cdots,m) ((y_iin{-1,1}) represents the category), the optimization problem of the SVM is (min_{w,b}frac{1}{2}w^Tw+Csum_{i=1}^{m}xi_i), subjectto(y_i(w^Tx_i+b)geq1-xi_i), (xi_igeq0), where (C) is the penalty parameter and (xi_i) is the slack variable. By solving this optimization problem, we get (w) and (b), and then substitute the operation and maintenance audit decision vector sample (v_{s}) into (f(v)=w^Tv+b), and determine its category based on the value of (f(v)), and get the power operation and maintenance fault training weight (w_{t}). For example, the value of (w_{t}) is (0.6) (this value is obtained through discriminant analysis based on the sample data).

[0122] 3. Debugging Deep Learning Algorithms Based on Fully Connected Modules 1) Debugging through the first fully connected module: The operation and maintenance inspection semantic vector sample is fully connected through the first fully connected module in the deep learning algorithm to obtain the operation and maintenance inspection fully connected prediction vector. For example, the first fully connected module has a weight matrix (W_{fc1}) from the input layer to the hidden layer (for example, the shape is ((100, m_{fc1})), where (m_{fc1}=50)), and the bias term is (b_{fc1}). For the operation and maintenance inspection semantic vector sample (v_{s1}=(y_1, y_2, cdots, y_{100})), the calculation formula is (h_{fc1}=tanh(W_{fc1}v_{s1}+b_{fc1})), and the obtained (h_{fc1}) is the operation and maintenance inspection fully connected prediction vector.

[0123] The deep learning algorithm is debugged based on the power operation and maintenance fault training weight (w_{t}) and the power operation and maintenance fault authentication weight (w_{a}) corresponding to the power operation and maintenance information sample (for example, (w_{a}=0.5), which is a known correct classification weight), as well as the operation and maintenance inspection full connection prediction vector (h_{fc1}) and the operation and maintenance inspection full connection authentication vector (h_{ac1}) corresponding to the power operation and maintenance information sample (for example, (h_{ac1}) is a vector obtained based on accurately labeled data). The error back propagation algorithm can be used to adjust the parameters in the first data mining module. For example, the mean square error (MSE=frac{1}{n}sum_{i=1}^{n}(h_{fc1i}-h_{ac1i})^2) between the prediction vector and the authentication vector is calculated, and then the weight matrix (W_{fc1}) and the bias term (b_{fc1}) and other parameters are adjusted according to this error, so that the error gradually decreases.

[0124] 2) Debugging through the second fully connected module: The first device operation indicator vector sample is fully connected through the second fully connected module in the deep learning algorithm to obtain the device operation indicator fully connected training vector corresponding to the first device operation indicator vector sample. For example, the second fully connected module has a weight matrix (W_{fc2}) from the input layer to the hidden layer (for example, the shape is ((5, m_{fc2})), where (m_{fc2}=10)), and the bias term is (b_{fc2}). For the first device operation indicator vector sample (v_{s2}=(z_1, z_2, z_3, z_4, z_5)), the calculation formula is (h_{fc2}=tanh(W_{fc2}v_{s2}+b_{fc2})), and the obtained (h_{fc2}) is the device operation indicator fully connected training vector.

[0125] Similarly, the deep learning algorithm is debugged based on the power operation and maintenance fault training weight (w_{t}) and the power operation and maintenance fault authentication weight (w_{a}), as well as the equipment operation indicator full connection training vector (h_{fc2}) and the equipment operation indicator full connection authentication vector (h_{ac2}) corresponding to the power operation and maintenance information sample (for example, (h_{ac2}) is a vector obtained based on accurately labeled data). The mean square error (MSE=frac{1}{n}sum_{i=1}^{n}(h_{fc2i}-h_{ac2i})^2) between the prediction vector and the authentication vector is calculated, and the weight matrix (W_{fc2}) and bias term (b_{fc2}) and other parameters are adjusted according to this error to optimize the deep learning algorithm and improve its accuracy in mining power operation and maintenance information.

[0126] It can be seen that debugging the deep learning algorithm by using the power operation and maintenance information samples has many beneficial effects. First, by mining, interacting and discriminant analysis of different modules on the sample data to obtain the power operation and maintenance fault training weights, the algorithm can initially establish a processing model for the power operation and maintenance information. Then, the algorithm is further debugged based on the fully connected module. By comparing the error between the prediction vector and the certification vector, the parameters in the algorithm can be adjusted to optimize the performance of the deep learning algorithm. This helps to improve the accuracy of the deep learning algorithm when mining power operation and maintenance information, and both the unstructured first power operation and maintenance information and the structured second power operation and maintenance information can be processed more accurately, thereby improving the reliability and effectiveness of the entire power operation and maintenance information mining system, and better ensuring the stable operation of the power system.

[0127] An embodiment of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the power operation and maintenance information mining method based on audit business.

[0128] An embodiment of the present invention provides a processor, which is used to run a program, wherein the power operation and maintenance information mining method based on the audit business is executed when the program is running.

[0129] In the embodiment of the present invention, Figure 2 As shown, the electric power operation and maintenance information mining system 100 includes at least one processor 101, and at least one memory 102 and a bus 103 connected to the processor 101; wherein the processor 101 and the memory 102 communicate with each other through the bus 103; the processor 101 is used to call the program instructions in the memory 102 to execute the above-mentioned electric power operation and maintenance information mining method based on the audit business.

[0130] The embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of the methods, power operation and maintenance information mining systems (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0131] In a typical configuration, the power operation and maintenance information mining system includes one or more processors (CPU), memory and bus. The power operation and maintenance information mining system may also include input / output interface, network interface and the like.

[0132] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip. The memory is an example of a computer-readable medium.

[0133] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage computer-readable storage media or any other non-transmission media that can be used to store information that can be accessed by the power operation and maintenance information mining system. As defined in this article, computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carriers.

[0134] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or computer-readable storage medium that includes a series of elements includes not only those elements, but also other elements that are not explicitly listed, or also includes elements inherent to such process, method, commodity or computer-readable storage medium. In the absence of further restrictions, the elements defined by the sentence "comprises a ..." do not exclude the presence of other identical elements in the process, method, commodity or computer-readable storage medium that includes the elements.

[0135] It should be understood by those skilled in the art that the embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the embodiments of the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the embodiments of the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0136] The above are only embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for mining power operation and maintenance information based on auditing business, characterized in that: The method is applied to a power operation and maintenance information mining system, and the method comprises: Acquire first power operation and maintenance information and second power operation and maintenance information corresponding to the target power operation and maintenance information, wherein the first power operation and maintenance information includes an operation and maintenance information set that does not include structured data in the target power operation and maintenance information, and the second power operation and maintenance information includes an operation and maintenance information set that includes structured data in the target power operation and maintenance information; Performing data mining on the first power operation and maintenance information to obtain an operation and maintenance inspection semantic vector corresponding to the first power operation and maintenance information; performing data mining on the second power operation and maintenance information to obtain an equipment operation indicator vector corresponding to the second power operation and maintenance information; Performing vector interaction on the equipment operation index vector and the operation and maintenance inspection semantic vector to obtain an operation and maintenance inspection decision vector corresponding to the target power operation and maintenance information; performing discriminant analysis on the operation and maintenance inspection decision vector to obtain a power operation and maintenance inspection fault weight corresponding to the target power operation and maintenance information; In response to the power operation and maintenance audit fault weight falling into a set weight interval, the target power operation and maintenance information is determined as abnormal operation and maintenance information.

2. The method according to claim 1, characterized in that The obtaining of the first power operation and maintenance information and the second power operation and maintenance information corresponding to the target power operation and maintenance information includes: Performing structured feature recognition on the target power operation and maintenance information to obtain a first operation and maintenance information set that does not contain structured data in the target power operation and maintenance information, and performing information splitting on the first operation and maintenance information set in the target power operation and maintenance information as the first power operation and maintenance information; Based on the first operation and maintenance information set, a second operation and maintenance information set containing structured data in the target power operation and maintenance information is obtained, and the second operation and maintenance information set in the target power operation and maintenance information is split into information as the second power operation and maintenance information.

3. The method according to claim 1, characterized in that The obtaining of the first power operation and maintenance information and the second power operation and maintenance information corresponding to the target power operation and maintenance information includes: Performing structured feature recognition on the target power operation and maintenance information to obtain a third operation and maintenance information set containing structured data in the target power operation and maintenance information, and performing information splitting on the third operation and maintenance information set in the target power operation and maintenance information to obtain the second power operation and maintenance information; Based on the third operation and maintenance information set, a fourth operation and maintenance information set that does not contain structured data in the target power operation and maintenance information is obtained, and the fourth operation and maintenance information set in the target power operation and maintenance information is split into information as the first power operation and maintenance information.

4. The method according to claim 1, characterized in that: The performing data mining on the first power operation and maintenance information to obtain the operation and maintenance inspection semantic vector corresponding to the first power operation and maintenance information includes: performing data mining on the first power operation and maintenance information through a first data mining module in a deep learning algorithm to obtain the operation and maintenance inspection semantic vector corresponding to the first power operation and maintenance information; The performing data mining on the second power operation and maintenance information to obtain the equipment operation index vector corresponding to the second power operation and maintenance information includes: performing data mining on the second power operation and maintenance information through a second data mining module in the deep learning algorithm to obtain the equipment operation index vector corresponding to the second power operation and maintenance information; The performing vector interaction on the equipment operation index vector and the operation and maintenance inspection semantic vector to obtain the operation and maintenance inspection decision vector corresponding to the target power operation and maintenance information comprises: performing vector interaction on the equipment operation index vector and the operation and maintenance inspection semantic vector through a vector interaction module in the deep learning algorithm to obtain the operation and maintenance inspection decision vector corresponding to the target power operation and maintenance information; The discriminant analysis is performed on the operation and maintenance audit decision vector to obtain the power operation and maintenance audit fault weight corresponding to the target power operation and maintenance information, including: through the discriminant analysis module in the deep learning algorithm, the discriminant analysis is performed on the operation and maintenance audit decision vector to obtain the power operation and maintenance audit fault weight corresponding to the target power operation and maintenance information.

5. The method according to claim 1, characterized in that In response to the power operation and maintenance audit fault weight falling into a set weight interval, the target power operation and maintenance information is determined as abnormal operation and maintenance information, including: on the basis that the power operation and maintenance audit fault weight is greater than a first set fault weight, the target power operation and maintenance information is determined as abnormal operation and maintenance information.

6. The method according to claim 1, characterized in that In response to the power operation and maintenance audit fault weight falling into a set weight interval, the target power operation and maintenance information is determined as abnormal operation and maintenance information, including: on the basis that the power operation and maintenance audit fault weight is less than a second set fault weight, the target power operation and maintenance information is determined as abnormal operation and maintenance information.

7. The method according to claim 4, characterized in that Before performing data mining on the first power operation and maintenance information through a first data mining module in a deep learning algorithm to obtain an operation and maintenance inspection semantic vector corresponding to the first power operation and maintenance information, the method further includes: Obtaining a first power operation and maintenance information sample and a second power operation and maintenance information sample corresponding to the power operation and maintenance information sample, wherein the first power operation and maintenance information sample includes an operation and maintenance information set that does not include structured data in the power operation and maintenance information sample, and the second power operation and maintenance information sample includes an operation and maintenance information set that includes structured data in the power operation and maintenance information sample; Performing data mining on the first power operation and maintenance information sample through a first data mining module in the deep learning algorithm to obtain an operation and maintenance inspection semantic vector sample corresponding to the first power operation and maintenance information sample; Performing data mining on the second power operation and maintenance information sample through a second data mining module in the deep learning algorithm to obtain a first equipment operation indicator vector sample corresponding to the second power operation and maintenance information sample; Through the vector interaction module in the deep learning algorithm, the operation and maintenance inspection semantic vector sample and the first equipment operation indicator vector sample are vector-interacted to obtain the operation and maintenance inspection decision vector sample corresponding to the power operation and maintenance information sample; Through the discriminant analysis module in the deep learning algorithm, the operation and maintenance audit decision vector sample is discriminantly analyzed to obtain the power operation and maintenance fault training weight corresponding to the power operation and maintenance information sample; The deep learning algorithm is debugged based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight corresponding to the power operation and maintenance information sample.

8. The method according to claim 7, characterized in that The debugging of the deep learning algorithm based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight corresponding to the power operation and maintenance information sample includes: Through the first fully connected module in the deep learning algorithm, the operation and maintenance inspection semantic vector sample is fully connected to obtain the operation and maintenance inspection fully connected prediction vector corresponding to the operation and maintenance inspection semantic vector sample; Debugging the deep learning algorithm based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight, as well as the operation and maintenance inspection full-connection prediction vector and the operation and maintenance inspection full-connection authentication vector corresponding to the power operation and maintenance information sample; Alternatively, debugging the deep learning algorithm based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight corresponding to the power operation and maintenance information sample includes: Performing full-connection processing on the first device operation indicator vector sample through the second full-connection module in the deep learning algorithm to obtain a device operation indicator full-connection training vector corresponding to the first device operation indicator vector sample; The deep learning algorithm is debugged based on the power operation and maintenance fault training weight and the power operation and maintenance fault authentication weight, as well as the equipment operation indicator full connection training vector and the equipment operation indicator full connection authentication vector corresponding to the power operation and maintenance information sample.

9. A power operation and maintenance information mining system, characterized in that: It includes a processor and a memory and a bus connected to the processor; wherein the processor and the memory communicate with each other through the bus; the processor is used to call the program instructions in the memory to execute the power operation and maintenance information mining method based on the audit business as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the method for mining electric power operation and maintenance information based on audit business as described in any one of claims 1 to 8 is implemented.

Images