Power equipment operation fault identification and reasoning method and system

Abstract

The invention discloses a power equipment operation fault identification and reasoning method and system, and the method comprises the steps: obtaining historical equipment sensor data, historical audio data and historical video data, and carrying out the preprocessing of the data; the preprocessed features serve as input, historical data serve as output, classification is carried out according to divided time periods, a deep belief network is adopted for model training, and a plurality of models are obtained after training; preprocessing the real-time equipment sensor data, the real-time audio data and the real-time video data, and inputting the data into a plurality of models to obtain a plurality of fault identification results; and if the plurality of fault identification results are consistent, the determined fault information can be obtained, and if the plurality of fault identification results are inconsistent, the probability of occurrence of each fault is obtained through Bayesian reasoning. The method can improve the fault detection precision, and greatly improves the safety and reliability of a power system.

Description

Technical Field

[0001] The present invention relates to the field of intelligent maintenance of electric power equipment, and in particular to multi-source input fault identification and reasoning technology in the operation of electric power equipment. Background Art

[0002] The power equipment operation fault identification and reasoning system is a key technology designed to ensure the stability and reliability of the power system, involving comprehensive analysis and real-time monitoring of the power system. The system mainly integrates multiple technical means to achieve rapid detection, accurate identification and effective positioning of faults. The traditional method is to use sensors to continuously monitor key operating parameters such as current, voltage, and frequency in order to promptly detect and respond to any abnormal conditions in the power grid. However, the detection results of traditional methods are often not accurate and timely enough. Summary of the invention

[0003] Purpose of the invention: In order to solve the above problems, this application uses sensor data, audio data and video data as identification inputs, and runs the identification and reasoning calculations on the on-site edge computer to avoid delays in the transmission of identification information.

[0004] Technical solution: A method for identifying and reasoning faults in power equipment operation, comprising the following steps:

[0005] Obtain historical device sensor data, historical audio data, and historical video data, and obtain digital features after preprocessing them respectively;

[0006] The digital features are used as input, the input is classified according to the divided time periods, a deep belief network is used for model training, and the confirmed fault type is used as output, and several models are obtained after training;

[0007] Real-time device sensor data, real-time audio data and real-time video data are pre-processed and input into the several models to obtain several fault identification results; if the several fault identification results are consistent, the determined fault information can be obtained; if the several fault identification results are inconsistent, the probability of occurrence of each fault in the fault identification results is obtained through Bayesian reasoning.

[0008] Further, preprocessing the historical device sensor data includes: normalizing the historical device sensor data and classifying them according to the divided time periods, wherein the normalization method is:

[0009] Normalized value = (sampling value – minimum value of sensor range) / (maximum value of sensor range – minimum value of sensor range)

[0010] After normalization, a multi-dimensional measurement value vector is obtained.

[0011] Furthermore, preprocessing the historical audio data includes: extracting digital features from the historical audio data and classifying them according to the divided time periods, and the digital feature extraction method is as follows:

[0012] Losslessly compress the sound recorded by the on-site microphone to obtain an audio file;

[0013] Extract the Mel-frequency cepstral coefficients of the audio file as the feature expression vector.

[0014] Furthermore, preprocessing the historical video data includes: extracting digital features from the historical video data and classifying them according to the divided time periods. The digital feature extraction method is as follows:

[0015] Extract image frames from the video stream at uniform time intervals, and extract color features and texture features from all image frames;

[0016] The color aggregation vector method is used to extract color features from the image: the fixed threshold method is used to analyze the connected regions of the color interval, and the root mean square difference between each pixel and the typical RGB value of the color interval is calculated. If it is less than the fixed threshold, it is the aggregation pixel corresponding to the color interval, otherwise it is a non-aggregate pixel, and the following aggregation vector is obtained:

[0017] {Color range N＝1……16 (R,G,B),(number of aggregated pixels, number of non-aggregated pixels)}

[0018] The grayscale binary value method is used to extract texture features of the image: the image is converted into grayscale using the averaging method: gray = (R+G+B) / 3; for the divided color interval, the pixel point closest to the typical RGB value of the color interval is selected, and its local binary value is calculated to form a multidimensional vector and normalized.

[0019] Furthermore, if the fault identification results are inconsistent, the probability of each fault occurring is obtained by Bayesian reasoning, specifically:

[0020] According to historical data, the probability of occurrence of different defined failure modes in all failure times is obtained, and the total probability of failure basis and the probability of fault misidentification are obtained;

[0021] Setting identification weights of several fault identification results to obtain the fault identification accuracy;

[0022] According to the Bayesian inference formula:

[0023] Reasoning result = fault identification accuracy * fault misidentification probability / total basic fault probability

[0024] Get the final result.

[0025] A power equipment operation fault identification and reasoning system, comprising: a data acquisition module and a data processing module, wherein the data acquisition module comprises a power grid equipment sensor, a power grid field recorder, and a power grid field monitoring camera; the data processing module comprises a power grid field edge computer;

[0026] Acquire historical device sensor data, historical audio data, and historical video data as well as real-time device sensor data, real-time audio data, and real-time video data in the data acquisition module for use in the data processing module;

[0027] The data processing module obtains historical equipment sensor data, historical audio data and historical video data, and obtains digital features after preprocessing them respectively; the digital features are used as input, the input is classified according to the divided time periods, a deep belief network is used for model training, and the confirmed fault type is used as output, and several models are obtained after training; the real-time equipment sensor data, real-time audio data and real-time video data are preprocessed and input into the several models to obtain several fault identification results; if the several fault identification results are consistent, the confirmed fault information can be obtained, if the several fault identification results are inconsistent, the probability of occurrence of each fault in the fault identification results is obtained through Bayesian reasoning.

[0028] Further, preprocessing the historical device sensor data includes: normalizing the historical device sensor data and classifying them according to the divided time periods, wherein the normalization method is:

[0029] Normalized value = (sampling value – minimum value of sensor range) / (maximum value of sensor range – minimum value of sensor range)

[0030] After normalization, a multi-dimensional measurement value vector is obtained.

[0031] Furthermore, preprocessing the historical audio data includes: extracting digital features from the historical audio data and classifying them according to the divided time periods, and the digital feature extraction method is as follows:

[0032] Losslessly compress the sound recorded by the on-site microphone to obtain an audio file;

[0033] Extract the Mel-frequency cepstral coefficients of the audio file as the feature expression vector.

[0034] Furthermore, preprocessing the historical video data includes: extracting digital features from the historical video data and classifying them according to the divided time periods. The digital feature extraction method is as follows:

[0035] Extract image frames from the video stream at uniform time intervals, and extract color features and texture features from all image frames;

[0036] The color aggregation vector method is used to extract color features from the image: the fixed threshold method is used to analyze the connected regions of the color interval, and the root mean square difference between each pixel and the typical RGB value of the color interval is calculated. If it is less than the fixed threshold, it is the aggregation pixel corresponding to the color interval, otherwise it is a non-aggregate pixel, and the following aggregation vector is obtained:

[0037] {Color range N＝1……16 (R,G,B),(number of aggregated pixels, number of non-aggregated pixels)}

[0038] The grayscale binary value method is used to extract texture features of the image: the image is converted into grayscale using the averaging method: gray = (R+G+B) / 3; for the divided color interval, the pixel point closest to the typical RGB value of the color interval is selected, and its local binary value is calculated to form a multidimensional vector and normalized.

[0039] Furthermore, if the fault identification results are inconsistent, the probability of each fault occurring is obtained by Bayesian reasoning, specifically:

[0040] According to historical data, the probability of occurrence of different defined failure modes in all failure times is obtained, and the total probability of failure basis and the probability of fault misidentification are obtained;

[0041] Setting identification weights of several fault identification results to obtain the fault identification accuracy;

[0042] According to the Bayesian inference formula:

[0043] Reasoning result = fault identification accuracy * fault misidentification probability / total basic fault probability

[0044] Get the final result.

[0045] A computing device comprises: one or more processors, one or more memories and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing a method for identifying and reasoning faults in operation of an electric power device.

[0046] A computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by a computing device, the computing device executes a method for identifying and reasoning faults in power equipment operation.

[0047] Beneficial effects: This application uses power grid equipment sensor data, power grid site video images and site sound information to identify faults with multiple input sources. Using deep belief network models and Bayesian reasoning, the complex patterns of faults are learned. In this way, the accuracy of fault detection can be improved, greatly improving the safety and reliability of the power system. DETAILED DESCRIPTION

[0048] The technical solution of the present invention is further described below.

[0049] Embodiment 1:

[0050] A method for identifying and reasoning an operation fault of an electric power equipment comprises the following steps:

[0051] (1) Obtain historical device sensor data, historical audio data, and historical video data, and pre-process them to obtain digital features;

[0052] (2) The digital features are used as input, the input is classified according to the divided time periods, a deep belief network is used for model training, and the confirmed fault type is used as output, and several models are obtained after training;

[0053] (3) Preprocessing the real-time device sensor data, real-time audio data, and real-time video data and inputting them into the several models obtained in step (2) to obtain several fault identification results; if the several fault identification results are consistent, the determined fault information can be obtained; if the several fault identification results are inconsistent, the probability of occurrence of each fault in the fault identification results is obtained through Bayesian reasoning.

[0054] The confirmed fault types belong to historical data, which include:

[0055] 1) Historical equipment sensor data, historical audio data, and historical video data before, during, and after a confirmed fault type occurs;

[0056] 2) The number of times all fault types have occurred;

[0057] 3) The probability of a certain fault type occurring in all the fault statistics, that is, the basic total probability of the occurrence of the fault type. For example, if a certain power equipment has a total of n faults of multiple types in history, and a certain type of fault has occurred m times, then the basic total probability of the fault is m / n;

[0058] 4) The probability of a certain fault type being misdiagnosed. For example, historically, the method proposed in this application deduced that the fault occurred a total of p times and was manually confirmed after the accident. In fact, the fault only occurred q times. Then the probability of the fault being misdiagnosed is: (pq) / p.

[0059] The time periods before, during and after the occurrence of the fault type are divided into time periods, for example:

[0060] 1) 10 minutes before the fault occurs; 2) From the beginning of the fault until the system stabilizes; 3) 10 minutes after the system stabilizes after the fault occurs.

[0061] Preferably, an initialization fault mode is defined. The fault modes include: insulation fault (F1); winding fault (F2); cooling system fault (F3); harmonic fault (F4); mechanical fault (F5); connector and wiring fault (F6); protection system fault (F7); overvoltage and overcurrent fault (F8); insulating oil and core fault (F9); and other faults (F10).

[0062] Preferably, preprocessing the historical device sensor data includes: normalizing the historical device sensor data and classifying them according to the divided time periods, wherein the normalization method is:

[0063] Normalized value = (sampling value – minimum value of sensor range) / (maximum value of sensor range – minimum value of sensor range)

[0064] After normalization, a multi-dimensional measurement value vector is obtained.

[0065] Preferably, the normalized sensor data is classified according to the following time periods:

[0066] 1) 10 minutes before the fault occurs; 2) From the beginning of the fault until the system stabilizes; 3) 10 minutes after the system stabilizes after the fault occurs.

[0067] Preferably, preprocessing the historical audio data includes: extracting digital features from the historical audio data and classifying the data according to divided time periods.

[0068] Preferably, the method for extracting digital features from audio data is as follows:

[0069] (1) Losslessly compress the sound recorded by the on-site microphone to obtain a WMA format audio file;

[0070] (2) Extracting the Mel-frequency cepstral coefficients of the audio file as the feature expression vector, wherein: when designing the Mel filter bank, select the following 20 sampling frequency points Mf and convert them into Mel scale, calculate and obtain 20 Mel-frequency cepstral coefficients to form a 20-dimensional audio feature vector and normalize it.

[0071] Mf=[62.5,125,250,500,1000,2000,3000,4000,5120,6340,7780,9510,11700,14470,18100,22880,29290,37930,49630,65535]Hz;

[0072] Similarly, the normalized audio feature vectors are classified according to the following time periods:

[0073] 1) 10 minutes before the fault occurs; 2) From the beginning of the fault until the system stabilizes; 3) 10 minutes after the system stabilizes after the fault occurs.

[0074] Preferably, preprocessing the historical video data includes: extracting digital features from the historical video data and classifying them according to the divided time periods.

[0075] Preferably, the method for extracting digital features of video data is as follows:

[0076] (1) Extract image frames from RTSP or RTMP video streams at uniform time intervals. Each image frame contains 256×256 pixels (each pixel can be represented by a three-dimensional vector [R, G, B]). The most common colors in the power grid operation monitoring screen (such as gray-black, blue-gray, etc.) are divided into 16 color intervals. Next, color features and texture features are extracted from all image frames.

[0077] (2) The color feature extraction of the image is performed using the color aggregation vector method. The method is as follows: 1. The connected region analysis is performed on the 16 color intervals using the fixed threshold method. The root mean square difference between each pixel and the typical RGB value of the color interval is calculated. If the difference is less than the fixed threshold, it is an aggregate pixel corresponding to the color interval. Otherwise, it is a non-aggregate pixel. The following aggregation vector is obtained:

[0078] {Color range N＝1……16 (R,G,B),(number of aggregated pixels, number of non-aggregated pixels)}

[0079] (3) Use the grayscale binary value method to extract texture features from the image. The method is as follows: 1. Use the average method to convert the image into grayscale: gray = (R + G + B) / 3. 2. For the 16 color intervals, select the pixel point closest to the typical RGB value of the color interval, calculate its local binary (LPB) value, form a 16-dimensional vector and normalize it.

[0080] As above, the normalized video feature vectors are classified according to the following time periods:

[0081] 1) 10 minutes before the fault occurs; 2) From the beginning of the fault until the system stabilizes; 3) 10 minutes after the system stabilizes after the fault occurs.

[0082] Preferably, the feature vector obtained after preprocessing is used as input, and the historical data of the fault mode is used as output. According to the divided time periods, a deep belief network (DBN) is used for model training, and a total of 9 trained models are obtained, as shown in the following Table 1:

[0083]

[0084] Table 1

[0085] Preferably, based on the sensor, sound and image information acquired in real time, the above-mentioned feature value extraction method is used to obtain feature vectors, and these features are used as inputs into the trained deep belief network model to obtain 9 fault identification results respectively. If the 9 discrimination results are consistent, a definite conclusion can be obtained. If the discrimination results are inconsistent, the probability of each fault appearing in the discrimination results is obtained through Bayesian reasoning. As shown in Table 2 below:

[0086] <![CDATA[DBN 11 :F3]]> <![CDATA[DBN 12 :F3]]> <![CDATA[DBN 13 :F7]]> <![CDATA[DBN 21 :F3]]> <![CDATA[DBN 22 :F1]]> <![CDATA[DBN 23 :F3]]> <![CDATA[DBN 31 :F7]]> <![CDATA[DBN 32 :F3]]> <![CDATA[DBN 33 :F3]]>

[0087] Table 2

[0088] Preferably, if the several fault identification results are inconsistent, the probability of each fault occurring is obtained by Bayesian reasoning, specifically:

[0089] a. Obtain the probability of occurrence of different defined failure modes in all failure times based on historical data, and obtain the total probability of failure basis and the probability of fault misidentification;

[0090] Preferably, the probability of occurrence of the fault modes F1, F3, and F7 in all fault times can be obtained based on historical data, namely: the basic total probability p(F1), p(F3), p(F7), and the probability of fault misdiagnosis p(F1-), p(F3-), p(F7-);

[0091] b. Set the identification weights of several fault identification results to obtain the fault identification accuracy;

[0092] Preferably, set the recognition weight of each model:

[0093] The accuracy of sensor data recognition> the accuracy of video data recognition> the accuracy of audio data recognition> the accuracy of the time period from when the fault occurs to when the system is stable> the accuracy of fault recognition in the 10-minute period before the fault occurs> the accuracy of the time period after the system is stable after the fault occurs. As shown in Table 3:

[0094]

[0095] Table 3

[0096] Therefore, according to the model identification results this time, the identification accuracy of fault modes F1, F3, and F7 is:

[0097] p(F1+)=1 / 9*75%=8.3%;

[0098] p(F3+)=6 / 9*(90%+95%+70%+65%+85%+75%) / 6=53.3%;

[0099] p(F7+)=2 / 9*(80%+85%) / 2=18.3%;

[0100] c. According to the Bayesian inference formula:

[0101] Reasoning result = fault identification accuracy * fault misidentification probability / total basic fault probability

[0102] The final result is obtained. Preferably, P(F1) = p(F1+)*p(F1-) / p(F1); P(F3) = p(F3+)*p(F3-) / p(F3); P(F7) = p(F7+)*p(F7-) / p(F7).

[0103] Embodiment 2:

[0104] A power equipment operation fault identification and reasoning system, comprising: a data acquisition module and a data processing module, wherein the data acquisition module comprises a power grid equipment sensor, a power grid field recorder, and a power grid field monitoring camera; the data processing module comprises a power grid field edge computer;

[0105] Acquire historical device sensor data, historical audio data, and historical video data as well as real-time device sensor data, real-time audio data, and real-time video data in the data acquisition module for use in the data processing module;

[0106] The data processing module obtains historical equipment sensor data, historical audio data and historical video data, and obtains digital features after preprocessing them respectively; the digital features are used as input, the input is classified according to the divided time periods, a deep belief network is used for model training, and the confirmed fault type is used as output, and several models are obtained after training; the real-time equipment sensor data, real-time audio data and real-time video data are preprocessed and input into the several models to obtain several fault identification results; if the several fault identification results are consistent, the confirmed fault information can be obtained, if the several fault identification results are inconsistent, the probability of occurrence of each fault in the fault identification results is obtained through Bayesian reasoning.

[0107] The confirmed fault types belong to historical data, which include:

[0108] 1) Historical equipment sensor data, historical audio data, and historical video data before, during, and after a confirmed fault type occurs;

[0109] 2) The number of times all fault types have occurred;

[0110] 3) The probability of a certain fault type occurring in all the fault statistics, that is, the basic total probability of the occurrence of the fault type. For example, if a certain power equipment has a total of n faults of multiple types in history, and a certain type of fault has occurred m times, then the basic total probability of the fault is m / n;

[0111] 4) The probability of a certain fault type being misdiagnosed. For example, historically, the method proposed in this application deduced that the fault occurred a total of p times and was manually confirmed after the accident. In fact, the fault only occurred q times. Then the probability of the fault being misdiagnosed is: (pq) / p.

[0112] The time periods before, during and after the occurrence of the fault type are divided into time periods, for example:

[0113] 1) 10 minutes before the fault occurs; 2) From the beginning of the fault until the system stabilizes; 3) 10 minutes after the system stabilizes after the fault occurs.

[0114] Preferably, an initialization fault mode is defined. The fault modes include: insulation fault (F1); winding fault (F2); cooling system fault (F3); harmonic fault (F4); mechanical fault (F5); connector and wiring fault (F6); protection system fault (F7); overvoltage and overcurrent fault (F8); insulating oil and core fault (F9); and other faults (F10).

[0115] Preferably, preprocessing the historical device sensor data includes: normalizing the historical device sensor data and classifying them according to the divided time periods, wherein the normalization method is:

[0116] Normalized value = (sampling value – minimum value of sensor range) / (maximum value of sensor range – minimum value of sensor range)

[0117] After normalization, a multi-dimensional measurement value vector is obtained.

[0118] Preferably, the normalized sensor data is classified according to the following time periods:

[0119] 1) 10 minutes before the fault occurs; 2) From the beginning of the fault until the system stabilizes; 3) 10 minutes after the system stabilizes after the fault occurs.

[0120] Preferably, preprocessing the historical audio data includes: extracting digital features from the historical audio data and classifying the data according to divided time periods.

[0121] Preferably, the method for extracting digital features from audio data is as follows:

[0122] (1) Losslessly compress the sound recorded by the on-site microphone to obtain a WMA format audio file;

[0123] (2) Extracting the Mel-frequency cepstral coefficients of the audio file as the feature expression vector, wherein: when designing the Mel filter bank, select the following 20 sampling frequency points Mf and convert them into Mel scale, calculate and obtain 20 Mel-frequency cepstral coefficients to form a 20-dimensional audio feature vector and normalize it.

[0124] Mf=[62.5,125,250,500,1000,2000,3000,4000,5120,6340,7780,9510,11700,14470,18100,22880,29290,37930,49630,65535]Hz;

[0125] Similarly, the normalized audio feature vectors are classified according to the following time periods:

[0126] 1) 10 minutes before the fault occurs; 2) From the beginning of the fault until the system stabilizes; 3) 10 minutes after the system stabilizes after the fault occurs.

[0127] Preferably, preprocessing the historical video data includes: extracting digital features from the historical video data and classifying them according to the divided time periods.

[0128] Preferably, the method for extracting digital features of video data is as follows:

[0129] (1) Extract image frames from RTSP or RTMP video streams at uniform time intervals. Each image frame contains 256×256 pixels (each pixel can be represented by a three-dimensional vector [R, G, B]). The most common colors in the power grid operation monitoring screen (such as gray-black, blue-gray, etc.) are divided into 16 color intervals. Next, color features and texture features are extracted from all image frames.

[0130] (2) The color feature extraction of the image is performed using the color aggregation vector method. The method is as follows: 1. The connected region analysis is performed on the 16 color intervals using the fixed threshold method. The root mean square difference between each pixel and the typical RGB value of the color interval is calculated. If the difference is less than the fixed threshold, it is an aggregate pixel corresponding to the color interval. Otherwise, it is a non-aggregate pixel. The following aggregation vector is obtained:

[0131] {Color range N＝1……16 (R,G,B),(number of aggregated pixels, number of non-aggregated pixels)}

[0132] (3) Use the grayscale binary value method to extract texture features from the image. The method is as follows: 1. Use the average method to convert the image into grayscale: gray = (R + G + B) / 3. 2. For the 16 color intervals, select the pixel point closest to the typical RGB value of the color interval, calculate its local binary (LPB) value, form a 16-dimensional vector and normalize it.

[0133] As above, the normalized video feature vectors are classified according to the following time periods:

[0134] 1) 10 minutes before the fault occurs; 2) From the beginning of the fault until the system stabilizes; 3) 10 minutes after the system stabilizes after the fault occurs.

[0135] Preferably, the feature vector obtained after preprocessing is used as input, and the historical data of the fault mode is used as output. According to the divided time periods, a deep belief network (DBN) is used for model training, and a total of 9 trained models are obtained, as shown in the following Table 4:

[0136]

[0137] Table 4

[0138] Preferably, based on the sensor, sound and image information acquired in real time, the above-mentioned feature value extraction method is used to obtain feature vectors, and these features are used as inputs into the trained deep belief network model to obtain 9 fault identification results respectively. If the 9 discrimination results are consistent, a definite conclusion can be obtained. If the discrimination results are inconsistent, the probability of each fault appearing in the discrimination results is obtained through Bayesian reasoning. As shown in Table 5 below:

[0139] <![CDATA[DBN 11 :F3]]> <![CDATA[DBN 12 :F3]]> <![CDATA[DBN 13 :F7]]> <![CDATA[DBN 21 :F3]]> <![CDATA[DBN 22 :F1]]> <![CDATA[DBN 23 :F3]]> <![CDATA[DBN 31 :F7]]> <![CDATA[DBN 32 :F3]]> <![CDATA[DBN 33 :F3]]>

[0140] Table 5

[0141] Preferably, if the several fault identification results are inconsistent, the probability of each fault occurring is obtained by Bayesian reasoning, specifically:

[0142] a. Obtain the probability of occurrence of different defined failure modes in all failure times based on historical data, and obtain the total probability of failure basis and the probability of fault misidentification;

[0143] Preferably, the probability of occurrence of the fault modes F1, F3, and F7 in all fault times can be obtained based on historical data, namely: the basic total probability p(F1), p(F3), p(F7), and the probability of fault misdiagnosis p(F1-), p(F3-), p(F7-);

[0144] b. Set the identification weights of several fault identification results to obtain the fault identification accuracy;

[0145] Preferably, set the recognition weight of each model:

[0146] The accuracy of sensor data recognition> the accuracy of video data recognition> the accuracy of audio data recognition> the accuracy of the time period from when the fault occurs to when the system is stable> the accuracy of fault recognition in the 10-minute period before the fault occurs> the accuracy of the time period after the system is stable after the fault occurs. As shown in Table 6:

[0147]

[0148] Table 6

[0149] Therefore, according to the model identification results this time, the identification accuracy of fault modes F1, F3, and F7 is:

[0150] p(F1+)=1 / 9*75%=8.3%;

[0151] p(F3+)=6 / 9*(90%+95%+70%+65%+85%+75%) / 6=53.3%;

[0152] p(F7+)=2 / 9*(80%+85%) / 2=18.3%;

[0153] c. According to the Bayesian inference formula:

[0154] Reasoning result = fault identification accuracy * fault misidentification probability / total basic fault probability

[0155] The final result is obtained. Preferably, P(F1) = p(F1+)*p(F1-) / p(F1); P(F3) = p(F3+)*p(F3-) / p(F3); P(F7) = p(F7+)*p(F7-) / p(F7).

[0156] Embodiment three:

[0157] A computing device comprises: one or more processors, one or more memories and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing a method for identifying and reasoning faults in operation of an electric power device.

[0158] Embodiment 4:

[0159] A computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by a computing device, the computing device executes a method for identifying and reasoning faults in power equipment operation.

[0160] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0161] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, 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, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0162] The computer instructions stored in the storage medium of the above embodiment are used to enable the computer to execute the lightweight power load forecasting method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0163] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present application (including the claims) is limited to these examples. In line with the concept of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0165] Although the present application has been described in conjunction with specific embodiments of the present application, many replacements, modifications and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may use the embodiments discussed.

Claims

1. A method for identifying and reasoning faults in power equipment operation, characterized in that: The following steps are involved: Obtain historical device sensor data, historical audio data, and historical video data, and obtain digital features after preprocessing them respectively; The digital features are used as input, the input is classified according to the divided time periods, a deep belief network is used for model training, and the confirmed fault type is used as output, and several models are obtained after training; Preprocessing the real-time device sensor data, the real-time audio data and the real-time video data and inputting them into the several models to obtain several fault identification results; If several fault identification results are consistent, the confirmed fault information can be obtained. If several fault identification results are inconsistent, the probability of occurrence of each fault in the fault identification results can be obtained through Bayesian reasoning.

2. A method for identifying and reasoning faults in power equipment operation according to claim 1, characterized in that: Preprocessing the historical device sensor data includes: normalizing the historical device sensor data and classifying it according to the divided time periods, wherein the normalization method is: Normalized value = (sampling value – minimum value of sensor range) / (maximum value of sensor range – minimum value of sensor range) After normalization, a multi-dimensional measurement value vector is obtained.

3. A method for identifying and reasoning an operation fault of an electric power equipment according to claim 1, characterized in that: Preprocessing the historical audio data includes: extracting digital features from the historical audio data and classifying them according to the divided time periods. The digital feature extraction method is as follows: Losslessly compress the sound recorded by the on-site microphone to obtain an audio file; Extract the Mel-frequency cepstral coefficients of the audio file as the feature expression vector.

4. A method for identifying and reasoning an operation fault of an electric power equipment according to claim 1, characterized in that: Preprocessing the historical video data includes: extracting digital features from the historical video data and classifying them according to the divided time periods. The digital feature extraction method is as follows: Extract image frames from the video stream at uniform time intervals, and extract color features and texture features from all image frames; The color aggregation vector method is used to extract color features from the image: the fixed threshold method is used to analyze the connected regions of the color interval, and the root mean square difference between each pixel and the typical RGB value of the color interval is calculated. If it is less than the fixed threshold, it is the aggregation pixel corresponding to the color interval, otherwise it is a non-aggregate pixel, and the following aggregation vector is obtained: {Color range N＝1……16 (R,G,B),(number of aggregated pixels, number of non-aggregated pixels)} The grayscale binary value method is used to extract texture features of the image: the image is converted into grayscale using the averaging method: gray = (R+G+B) / 3; for the divided color interval, the pixel point closest to the typical RGB value of the color interval is selected, and its local binary value is calculated to form a multidimensional vector and normalized.

5. A method for identifying and reasoning an operation fault of an electric power equipment according to claim 1, characterized in that: If the fault identification results are inconsistent, the probability of each fault occurring is obtained by Bayesian reasoning, specifically: According to historical data, the probability of occurrence of different defined failure modes in all failure times is obtained, and the total probability of failure basis and the probability of fault misidentification are obtained; Setting identification weights of several fault identification results to obtain the fault identification accuracy; According to the Bayesian inference formula: The final result is obtained by the inference result = fault identification accuracy * fault misidentification probability / total basic fault probability.

6. A power equipment operation fault identification and reasoning system, characterized in that: include: A data acquisition module and a data processing module, wherein the data acquisition module includes a power grid equipment sensor, a power grid field recorder, and a power grid field monitoring camera; the data processing module includes a power grid field edge computer; Acquire historical device sensor data, historical audio data, and historical video data as well as real-time device sensor data, real-time audio data, and real-time video data in the data acquisition module for use in the data processing module; The data processing module obtains historical device sensor data, historical audio data, and historical video data, and obtains digital features after preprocessing them respectively; The digital features are used as input, the input is classified according to the divided time periods, a deep belief network is used for model training, and the confirmed fault type is used as output, and several models are obtained after training; real-time device sensor data, real-time audio data and real-time video data are pre-processed and input into the several models to obtain several fault identification results; If several fault identification results are consistent, the confirmed fault information can be obtained. If several fault identification results are inconsistent, the probability of occurrence of each fault in the fault identification results can be obtained through Bayesian reasoning.

7. The power equipment operation fault identification and reasoning system according to claim 6 is characterized in that: Preprocessing the historical device sensor data includes: normalizing the historical device sensor data and classifying it according to the divided time periods, wherein the normalization method is: Normalized value = (sampling value – minimum value of sensor range) / (maximum value of sensor range – minimum value of sensor range) After normalization, a multi-dimensional measurement value vector is obtained.

8. The power equipment operation fault identification and reasoning system according to claim 6 is characterized in that: Preprocessing the historical audio data includes: extracting digital features from the historical audio data and classifying them according to the divided time periods. The digital feature extraction method is as follows: Losslessly compress the sound recorded by the on-site microphone to obtain an audio file; Extract the Mel-frequency cepstral coefficients of the audio file as the feature expression vector.

9. The power equipment operation fault identification and reasoning system according to claim 6, characterized in that: Preprocessing the historical video data includes: extracting digital features from the historical video data and classifying them according to the divided time periods. The digital feature extraction method is as follows: Extract image frames from the video stream at uniform time intervals, and extract color features and texture features from all image frames; The color aggregation vector method is used to extract color features from the image: the fixed threshold method is used to analyze the connected regions of the color interval, and the root mean square difference between each pixel and the typical RGB value of the color interval is calculated. If it is less than the fixed threshold, it is the aggregation pixel corresponding to the color interval, otherwise it is a non-aggregate pixel, and the following aggregation vector is obtained: {Color range N＝1……16 (R,G,B),(number of aggregated pixels, number of non-aggregated pixels)} The grayscale binary value method is used to extract texture features of the image: the image is converted into grayscale using the averaging method: gray = (R+G+B) / 3; for the divided color interval, the pixel point closest to the typical RGB value of the color interval is selected, and its local binary value is calculated to form a multidimensional vector and normalized.

10. The power equipment operation fault identification and reasoning system according to claim 6, characterized in that: If the fault identification results are inconsistent, the probability of each fault occurring is obtained by Bayesian reasoning, specifically: According to historical data, the probability of occurrence of different defined failure modes in all failure times is obtained, and the total probability of failure basis and the probability of fault misidentification are obtained; Setting identification weights of several fault identification results to obtain the fault identification accuracy; According to the Bayesian inference formula: The final result is obtained by the inference result = fault identification accuracy * fault misidentification probability / total basic fault probability.

11. A computing device, characterized in that: include: One or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing any of the methods according to claims 1 to 5.

12. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions which, when executed by a computing device, cause the computing device to perform any one of the methods according to claims 1 to 5 .

Images