A method for power equipment fault analysis based on multi-source data fusion

Abstract

The present invention discloses a method for analyzing power equipment faults based on multi-source data fusion, the method comprising: data collection and storage; data processing; fault diagnosis; integrating multi-source data fusion with a deep learning fault diagnosis algorithm, and optimizing the execution efficiency of the algorithm. The present invention uses multi-source data fusion and deep learning to monitor and warn of power equipment status and faults in real time, and provides corresponding maintenance suggestions and measures according to the type and degree of the fault. It can be applied to the operation and maintenance of power systems, as well as other intelligent management systems that require power equipment information. Its high efficiency and intelligence can greatly improve the operating efficiency and safety of power equipment, and reduce maintenance costs and risks.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment failure, and in particular relates to a power equipment failure analysis method based on multi-source data fusion. Background Art

[0002] Power equipment is an important part of the power system, and its operating status directly affects the quality and stability of power supply. However, due to the complexity and diversity of power equipment, as well as changes and interference in the external environment, power equipment may have various faults and abnormalities. If these faults are not discovered and handled in time, it may cause equipment damage, downtime, or even accidents. Therefore, effective fault diagnosis of power equipment is the key to ensuring the normal operation of the power system. At present, traditional power equipment fault diagnosis methods mainly rely on manual detection and experience judgment, lacking automation and intelligence. These methods are not only time-consuming and labor-intensive, but also easily affected by human factors, resulting in inaccurate or untimely diagnosis results. Summary of the invention

[0003] In order to address the shortcomings of the prior art, the present invention aims to provide a method for power equipment fault analysis based on multi-source data fusion. The method is applied to the operation and maintenance of power systems, as well as other intelligent management systems that require power equipment information. Its high efficiency and intelligence will greatly improve the operating efficiency and safety of power equipment, thereby effectively reducing maintenance costs and risks.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] The present invention proposes a method for analyzing power equipment faults based on multi-source data fusion, comprising the following steps:

[0006] S1. Data collection and storage: Use relevant collection devices to collect relevant data of power equipment including different types and operating status of power equipment, and annotate the power equipment status and fault type for each data, which will be used as training sets and test sets;

[0007] S2. Data processing: normalize the linear function and establish a Bayesian model;

[0008] S3. Fault diagnosis: S3-1. Use deep learning to extract data features, classify data, and predict data for real-time monitoring of power equipment status and fault warning;

[0009] S3-2. Use image recognition to analyze images of the surface of power equipment, identify the structure and components of the power equipment, and detect visual faults such as cracks, corrosion and deformation;

[0010] S3-3. Obtain information similar to or related to the current fault from historical cases through the expert knowledge base, and provide corresponding maintenance suggestions and measures;

[0011] S4. Integrate the multi-source data fusion method with the deep learning fault diagnosis algorithm and optimize the execution efficiency of the algorithm.

[0012] Furthermore, the collection device for collecting the relevant data of the power equipment in step S1 includes sensors and cameras installed on the power equipment; the collected relevant data includes different types and operating states of the power equipment.

[0013] Furthermore, the linear function normalized values ​​in the data processing of step S2 include:

[0014] Normalization is to unify the values ​​into the same range. Linear function normalization transforms the feature data linearly and scales the data geometrically to the range of 0~1. The calculation formula is as follows:

[0015] ;

[0016] Indicates that the collected voltage, current, and temperature data are normalized; Represents feature data, and Represent the maximum and minimum values ​​of the characteristic data respectively; the zero mean normalization method is used to standardize and normalize the noise points of the data during the experiment, and the data is mapped to , The normalized calculation description for the distribution of is:

[0017] ;

[0018] In the formula, is the assumed mean, is the standard deviation; It means standardizing and normalizing the data noise of voltage, current, temperature, etc.

[0019] The role of normalization and the iterative process of optimization training include two types of numerical features, and the features directly represented by numbers are distinguished from categorical features that can only be represented by 0 and 1; voltage The value range is 0~10, current The value range is 0~1; under the same learning rate of the gradient descent method, due to The value interval ratio The interval is large, so Iteration update speed is faster than Large, it takes a long time to find the optimal solution; , Normalized, characteristic voltage and current data , The update rates become closer and the optimization algorithm is used to find the optimal solution more easily.

[0020] Furthermore, the establishment of the Bayesian model in step S2 data processing includes:

[0021] The probability formula of the Bayesian model is:

[0022] ;

[0023] in, are disjoint events, B means events that occur under certain conditions; It is an inevitable event The union of; is the “prior probability” of B, is the class conditional probability of sample δ for class labeled B, is the “evidence” factor used for normalization, is the class conditional probability of sample A for class labeled B;

[0024] Through the event The derivation of conditional probability obtains the event The posterior probability is:

[0025] ;

[0026] in, It is A speculative event, It is known that Sample events; is sample B for class labeled A i The class conditional probability of is sample B for class labeled A i The class conditional probability of is sample B for class labeled A j The class conditional probability of and Yes A i and A j The “prior probability” of ; m represents the number of data.

[0027] Furthermore, the data feature extraction in step S3-1 includes: obtaining new attributes through the relationship between different data to change the original feature space; specifically:

[0028] First, the correlation between features is eliminated, and then the most important features among the irrelevant ones are selected, and the existing features are transformed into other mutually perpendicular and irrelevant spaces, and the features with the largest distribution dispersion, that is, the smallest projection distance, are selected as important features.

[0029] Data classification includes: establishing rules based on known samples in several categories and assigning labels to the categories of unknown samples. If the number of types of divisions exceeds two, the number of divisions corresponds to the number of classifiers, and the data is expressed in the form of feature vectors. In the process of target detection, the selection of classifiers includes , and neural network; bounding box regression is to find the mapping relationship and make the position and size of the candidate box generated by the candidate area step as close as possible to the position and size of the real standard box, which is used to determine the candidate box category and obtain the position of the candidate box.

[0030] After data feature extraction and classification, each feature value information is predicted through a deep learning algorithm, which includes a dendritic neural network DDNet; the algorithm formula is:

[0031] ;

[0032] in, is the weight parameter matrix, is the vector representation of the input value, is the predicted output value, is the Haddam product operation.

[0033] In step S3-2, the image of the surface of the power equipment is analyzed by regressing the image frame, and finally the result is optimized to obtain the optimal output.

[0034] Furthermore, the multi-source data fusion method in step S4 includes:

[0035] S4-1. Calculate the confidence distance matrix of multi-source sensor data. The matrix represents the relative confidence of each sensor data. Assuming that the sensor data obeys the normal distribution, the confidence distance is expressed as:

[0036] ;

[0037] in, It is The sensor and The relative confidence of each sensor, are the estimated model parameters; x i and x j Represents the i-th and j-th fusion information values; Indicates the value of each sensor; represents the derivative with respect to r;

[0038] S4-2. Measure the confidence and select a distance threshold to update the confidence distance matrix and obtain the correlation matrix. The threshold is selected as 0 or 1. If the distance is higher than the threshold, the correlation is reduced, and if the distance is lower than the threshold, the correlation matrix is ​​increased. The confidence distance matrix Its expression is:

[0039] ;

[0040] Among them, u represents the value of relative confidence;

[0041] S4-3. Filter multi-sensor data through correlation matrix to obtain fusion information;

[0042] S4-4. Obtain the estimated value of the fusion model parameters through the function to complete the model data inference. The parameter estimation formula is expressed as:

[0043] ;

[0044] in, are the model parameter estimates, and is the given initial value of the parameter, represents the value of the ith item, It is A fusion information value.

[0045] The present invention has the following beneficial effects:

[0046] The present invention realizes fault diagnosis of power equipment through intelligent analysis of multi-source data fusion and deep learning. It collects various data generated during the operation of power equipment, such as voltage, current, temperature, vibration, sound, etc., and uses deep learning to extract features, classify and predict the data, so as to realize real-time monitoring of the status of power equipment and fault warning, thereby significantly improving the operating efficiency and safety of power equipment and reducing maintenance costs and risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a diagram of the multi-element fusion analysis system of the power equipment of the present invention;

[0048] Figure 2 is a data-level fusion graph of the present invention;

[0049] Figure 3 This is the DDNet network structure diagram of the present invention. DETAILED DESCRIPTION

[0050] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not intended to limit the scope of use of the present invention.

[0051] like Figure 1-3 As shown, the present invention proposes a method for analyzing power equipment faults based on multi-source data fusion, comprising the following steps:

[0052] S1. Data collection and storage: Utilize various sensors and cameras installed on the power equipment and program corresponding data collection programs to collect a series of data containing different types and operating conditions of power equipment, and annotate each data with the power equipment status and fault type so as to serve as training sets and test sets.

[0053] S2. Data processing: normalize the linear function numerically and establish a Bayesian model. Through data processing, we can achieve preprocessing, cleaning, normalization, dimensionality reduction and other operations on different types of data.

[0054] In order to more effectively mine the effective information in the data, it is necessary to provide concise and clean data for the data analysis and mining models. However, the original model data obtained in practical applications usually have many problems and cannot be put into the training of the actual model. At this time, the data needs to be preprocessed. The preprocessing of power equipment time series data can be divided into the following steps: data cleaning, data segmentation, and data standardization. The results of data preprocessing are sent to the model for training, and then the output layer of the model provides classification or prediction results. The LSTM network is trained using the stochastic gradient descent method, and the trained LSTM network provides prediction results. This is the detailed LSTM model construction step. In order to improve the accuracy of the preprocessing experiment, eliminate the dimensional influence between data features, and manually remove the noise points of the data, data cleaning should remove the noise of the data. Data cleaning eliminates the influence of noise on the features, improves the quality of the data, and is conducive to the stability of the model results. Data cleaning can be considered from the following perspectives: unreasonable values, missing values, and dirty data such as duplicate records.

[0055] Among them, the linear function normalized values ​​include:

[0056] Normalization is to unify the values ​​into the same range. There are two normalization methods: linear function normalization and zero mean normalization. Linear function normalization transforms the feature data linearly and scales the data geometrically to the range of [0,1]. The calculation formula is as follows:

[0057] ;

[0058] It means that the collected voltage, current and temperature data are normalized, which is beneficial for data storage and zero mean normalization; Represents feature data, and Represent the maximum and minimum values ​​of the characteristic data respectively. The more widely used zero-mean normalization method is used to normalize and normalize the data noise during the experiment, and the data is mapped to , The normalized calculation is described as:

[0059] ;

[0060] In the formula, is the characteristic data, is the assumed mean, is the standard deviation; It means standardizing and normalizing the data noise such as voltage, current, and temperature.

[0061] The role of normalization and the iterative process of optimization training include two types of numerical features, and the features directly represented by numbers are distinguished from categorical features that can only be represented by 0 and 1; the feature data voltage The value range is 0~10, current The value range is 0~1; under the same learning rate of the gradient descent method, due to The value interval ratio The interval is large, so Iteration update speed is faster than Large, it takes a long time to find the optimal solution; , Normalized, characteristic data voltage , Current The update rates become closer and the optimization algorithm is used to find the optimal solution more easily.

[0062] The Bayesian method is a more traditional method in probability theory. It provides a method for calculating the probability of a hypothesis. It is calculated by the prior probability of a given hypothesis, the probability of observing different data under a given hypothesis, and the observed data itself. The prior information of the unknown parameters obtained is comprehensively analyzed with the provided sample information, and then the posterior information is obtained according to the Bayesian formula. The unknown parameters are derived through the posterior information. In Bayesian reasoning, all related variables are considered to be random variables and the influence of random values ​​is not considered, while the probability of the data is expressed as the degree of ignorance of the related variables under given information. As a traditional statistical method, the Bayesian method is characterized by making full use of existing information such as sample information and overall information, and establishing statistical reasoning on a reasonable posterior distribution. This can greatly reduce the statistical error caused by the lack of sample size, and can estimate and reason about unknown samples.

[0063] In Bayesian reasoning, the probability formula of the Bayesian model is:

[0064] ;

[0065] in, are disjoint events, B means events that occur under certain conditions; It is an inevitable event The union of; is the “prior probability” of B, is the class conditional probability of sample δ for class labeled B, is the “evidence” factor used for normalization, is the class conditional probability of sample A for class label B.

[0066] Through the event The derivation of conditional probability obtains the event The posterior probability is:

[0067] ;

[0068] in, It is A speculative event, It is known that Sample events; is sample B for class labeled A i The class conditional probability of is sample B for class labeled A i The class conditional probability of is sample B for class labeled A j The class conditional probability of and Yes A i and A j The “prior probability” of ; m represents the number of data.

[0069] S3. Fault diagnosis: S3-1. Use deep learning to extract data features, classify data, and predict data for real-time monitoring of power equipment status and fault warning;

[0070] Data feature extraction includes: obtaining new attributes through the relationship between different data to change the original feature space. Specifically:

[0071] First, eliminate the correlation between features. Then select the most important features among the irrelevant ones, transform the existing features into other mutually perpendicular (irrelevant) spaces, and select the most important features with the largest distribution dispersion, that is, the smallest projection distance. Commonly used feature extraction methods designed manually include SIFT, HOG, Haar-like, and LBP.

[0072] Data classification includes: establishing rules based on known samples in several categories and assigning labels to the categories of unknown samples. If the number of types of divisions exceeds two, the number of divisions corresponds to the number of classifiers, and the data is expressed in the form of feature vectors. In the process of target detection, the selection of classifiers includes , and neural network; bounding box regression is to find the mapping relationship and make the position and size of the candidate box generated by the candidate area step as close as possible to the position and size of the real standard box, which is used to determine the candidate box category and obtain the position of the candidate box.

[0073] After data feature extraction and classification, each feature value information is predicted through a deep learning algorithm. The deep learning algorithm includes and Neural network. As a new classification algorithm, the dendritic neural network DDNet has demonstrated its advantages in terms of speed and accuracy on conventional public data sets (such as MNIST).

[0074] DDNet can be understood as a logic extraction tool that extracts the relationship between data, analyzes to obtain effective information and outputs results. Studies have shown that the DDNet module can build a multi-layer network. The more layers there are, the more accurate the information that can be extracted, but the larger the parameters, the slower the model speed. Therefore, finding a suitable number of module links that takes into account both speed and accuracy needs to be verified through actual testing. The formula of the algorithm is:

[0075] ;

[0076] in, is the weight parameter matrix, Z is the vector expression of the input value, is the predicted output value, is the Haddam product operation (element-wise multiplication). The DDNet network structure is as follows Figure 3 As shown. DDNet network input data x1...x n After that, through W1...W n The operation of the first layer parameters generates new neuron data, which is then combined with the original input data x1...x n After the Haddam product operation is completed, the algorithm for each layer is the same as above. After the nth layer operation is completed, the network inputs the predicted value Y.

[0077] The BP (Back Propagation) neural network algorithm is the most important algorithm in deep learning of neural networks. It is an indispensable foundation for deep learning network models. In classification problems, the output can be represented by a unique hot code to represent the classification result. Figure 3 The training results of the BP neural network are shown as three hidden layers with 4, 4, and 4 neurons in each layer. The results can be seen from the visualized model training convergence function Loss value and fluctuation mark. After 1000 times of model training, the Loss value is 0.1, and the convergence of the Loss value is poor. The mean square error is 1.1 and the accuracy is 96%.

[0078] The DDNet neural network algorithm is used as the data fusion method of the present invention. In order to verify its advantages and disadvantages with the BP neural network, three hidden layers are also set. The number of neurons in each layer is consistent with the BP neural network. The final loss value is 0.009, the mean square error is 0.26, and the accuracy rate reaches 98%. It can be seen from the experiment that the DDNet algorithm can predict the feature value more accurately. The comparison of the training prediction data of the BP neural network and the DDNet model is shown in Table 1.

[0079] Table 1 Comparison of training prediction data between BP neural network and DDNet model

[0080] BP Neural Network DDNet Number of model neurons 4，4，4 4，4，4 Model training loss value 0.1 0.009 Model training mean square error 1.1 0.26 Model prediction accuracy 96% 98%

[0081] By comparing with other machine learning algorithms such as BP neural network, it is concluded that DDNet neural network has better performance in data fusion.

[0082] The present invention uses continuous data frames during the operation of power equipment and introduces time series information to construct a prediction model for power equipment. Even if some data frames are abnormal, the system can still accurately estimate the state of the power equipment. At the same time, the prediction model can be used to predict the future state of the power equipment and achieve early warning of faults, thereby avoiding or reducing the losses caused by faults.

[0083] S3-2. Use image recognition to analyze images of the surface of power equipment, identify the structure and components of the power equipment, and detect whether there are visual faults such as cracks, corrosion and deformation.

[0084] Generally speaking, target detection is divided into region proposal, feature representation and region classification. In order to successfully complete the detection of the required target, it can be achieved by regressing the image border, and finally optimizing the result to obtain the optimal output. This method has been widely used in both early recognition algorithms and current target detection algorithms.

[0085] Candidate region Since target detection not only needs to classify all targets in the input image and output the label of the target category probability. It is also necessary to determine the location of the target in the image, mark the target with a bounding box and output it. In order to be able to determine the location of the target in the image, the current location of some sub-regions must be selected first. The next step is to select a suitable calculation algorithm in the selected sub-region location. The final calculated target position is the maximum category probability output value. In the subsequent algorithm work, non-maximum values ​​will be used to suppress and correct the candidate regions representing the preliminary estimated target location, so as to obtain relatively accurate information on the target location.

[0086] S3-3. Obtain information similar to or related to the current fault from historical cases through the expert knowledge base, and provide corresponding maintenance suggestions and measures.

[0087] By using the expert knowledge base, information similar to or related to the current fault is obtained from historical cases, such as the cause of the fault, probability of occurrence, scope of impact, maintenance method, etc. Combining the expert knowledge base with the results of deep learning analysis can provide more targeted and operational fault diagnosis results. By comprehensively applying the above innovations, the method provided by the present invention can create a more efficient, intelligent and predictive power equipment fault diagnosis system, providing more reliable data support and decision-making basis for power equipment operation and maintenance.

[0088] S4. Integrate the multi-source data fusion method with the deep learning fault diagnosis algorithm, and optimize the execution efficiency of the algorithm to meet real-time requirements.

[0089] Use multi-source data fusion and deep learning to monitor and warn of power equipment status and faults in real time. Provide corresponding maintenance suggestions and measures based on the type and severity of the fault. This power equipment fault diagnosis system can be applied to the operation and maintenance of power systems, as well as other intelligent management systems that require power equipment information. Its high efficiency and intelligence will greatly improve the operating efficiency and safety of power equipment, and reduce maintenance costs and risks. Data quality and integrity may affect the accuracy of fault diagnosis, and the training and testing of deep learning models may require a large amount of data and computing resources.

[0090] Since the Bayesian model is insensitive to the number of samples and can make full use of the existing information, the Bayesian reasoning method is often used in data-level fusion and decision-level fusion in multi-source data fusion. In data-level fusion, the Bayesian method shows its superiority in the ability to make full use of data information and reduce data information redundancy. In multi-source signal fusion, each piece of data can be dimensionally mapped, and new data is obtained by fusion by mapping the shortest distance to the selected fitting position in another space. Then the conditional probability of the distance judgment result of each piece of data and the dimensional space is expressed as the likelihood function of the data point, and the concept of the judgment result output of multiple sensors is expressed as the likelihood function of data-level fusion. Generally speaking, the multi-source data fusion method under Bayesian reasoning is mainly divided into four steps.

[0091] S4-1. Calculate the confidence distance matrix of multi-source sensor data. The matrix represents the relative confidence of each sensor data. Assuming that the sensor data obeys the normal distribution, the confidence distance is expressed as:

[0092] ;

[0093] in, It is The sensor and The relative confidence of each sensor, are the estimated model parameters; x i and x j Represents the i-th and j-th fusion information values; Indicates the value of each sensor. represents the derivative with respect to r.

[0094] S4-2. Measure the confidence and select a distance threshold to update the confidence distance matrix and obtain the correlation matrix. The threshold is selected as 0 or 1. If the distance is higher than the threshold, the correlation is reduced, and if the distance is lower than the threshold, the correlation matrix is ​​increased. The confidence distance matrix Its expression is:

[0095] ;

[0096] Among them, u represents the value of relative confidence;

[0097] S4-3. Filter multi-sensor data through correlation matrix to obtain fusion information;

[0098] S4-4. Obtain the estimated value of the fusion model parameters through the function to complete the model data inference. The parameter estimation formula is expressed as:

[0099] ;

[0100] in, are the model parameter estimates, and is the given initial value of the parameter, represents the value of the ith item, It is A fusion information value.

[0101] The Bayesian method has fast calculation speed and can effectively reduce the calculation complexity of the model by using the log-likelihood function. The Bayesian method can make full use of the input data. The most important feature of data-level fusion is the huge amount of information and high information redundancy. The Bayesian method can overcome these problems well. Therefore, the Bayesian method has high applicability in data-level fusion.

[0102] Voltage and current signals can be represented in the time domain, which shows the variation of the signal over time. This can be a periodic waveform, such as a sine wave, or in non-ideal cases may contain noise and other variations. Frequency domain representation of voltage and current: By converting the signal to the frequency domain, the frequency distribution of the signal can be obtained. This includes information about the various frequency components, which helps detect equipment anomalies.

[0103] In the power equipment fault analysis system, the key to multi-source data fusion is to integrate different types of data into a unified framework and ensure that they can complement and verify each other. First, timestamp alignment is performed to obtain information from multiple data at the same time point. Then, deep learning technology is used to extract features from each data, find key fault information from complex raw data, and classify and predict the fused data.

[0104] Traditional data analysis methods usually include statistical methods, model fitting, and rule-based algorithms. Traditional methods are usually based on the linear assumption that the relationship between data is linear. For complex and nonlinear data structures, such methods may not accurately capture the complex patterns and relationships in the data. Manual feature engineering is usually required, that is, manually selecting and designing features for analysis. When faced with large-scale, high-dimensional data, manual feature engineering may become cumbersome and incomplete, resulting in the omission of some important information. And it is more sensitive to noise and outliers, which may lead to a decrease in the generalization performance of the model, especially in datasets with diversity and complexity in the real world. Deep learning models consist of multiple layers, each of which can automatically learn abstract features of the data. Through multi-level feature learning, deep learning models can discover complex hierarchical structures in the data, thereby better capturing the nonlinear relationship of the data, and have strong adaptive expression capabilities, and can learn and generalize large-scale, high-dimensional data. This makes deep learning superior in dealing with complex scenarios and large-scale data sets.

[0105] The present invention introduces multi-source data fusion: in addition to using traditional electrical parameter data, such as voltage, current, etc., other types of data, such as temperature, vibration, sound, etc., can also be introduced. In the power equipment fault analysis system, the key to multi-source data fusion is to integrate different types of data into a unified framework and ensure that they can complement and verify each other. First, timestamp alignment is performed to obtain information from multiple data at the same time point. Then use deep learning technology to extract features from each data, find key fault information from complex original data, and classify and predict the fused data. These data can reflect the operating status of power equipment from different angles, and can complement and verify each other. Multi-source data is integrated to obtain more comprehensive and accurate power equipment information, thereby improving the efficiency and reliability of fault diagnosis.

[0106] The present invention uses deep learning. Traditional data analysis methods may not work well for complex and nonlinear data. Traditional methods are usually based on linear assumptions, that is, they assume that the relationship between data is linear. For complex and nonlinear data structures, such methods may not accurately capture the complex patterns and relationships in the data. The deep learning model consists of multiple levels, and each layer can automatically learn the abstract features of the data. Through multi-level feature learning, the deep learning model can discover the complex hierarchical structure in the data, thereby better capturing the nonlinear relationship of the data, and has a strong adaptive expression ability, and can learn and generalize large-scale, high-dimensional data. This makes deep learning superior in processing complex scenes and large-scale data sets. Using deep learning based on neural networks, features can be extracted from data more effectively, and can be trained and tested on large-scale data sets. This method can learn deeper data laws and better handle noise and anomalies in complex scenes.

[0107] The present invention designs an adaptive mechanism that enables the system to optimize and adjust itself according to real-time analysis results and user feedback. By using technologies such as reinforcement learning, the system can continuously learn and adapt to various types of power equipment and operating conditions, thereby continuously improving the accuracy and stability of fault diagnosis. Obtain environmental information: Use temperature and humidity sensors, light sensors and other equipment to record the environmental conditions of the power equipment, such as temperature, humidity, light and other information. Combining environmental information with the results of power equipment status analysis can enhance the system's understanding and judgment of the cause of the fault, and more accurately diagnose the type and degree of the fault. Introduce image recognition: Through image recognition, such as target detection, semantic segmentation and other methods, analyze the image of the surface of the power equipment, identify the structure and components of the power equipment, and detect whether there are visual faults such as cracks, corrosion, and deformation. This helps to improve the comprehensiveness and intuitiveness of fault diagnosis.

[0108] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. A method for analyzing power equipment faults based on multi-source data fusion, characterized in that: The following steps are involved: S1. Data collection and storage: Use relevant collection devices to collect relevant data of power equipment including different types and operating status of power equipment, and annotate the power equipment status and fault type for each data, which will be used as training sets and test sets; S2. Data processing: normalize the linear function and establish a Bayesian model; S3. Fault diagnosis: S3-1. Use deep learning to extract data features, classify data, and predict data for real-time monitoring of power equipment status and fault warning; S3-2. Use image recognition to analyze images of the surface of power equipment, identify the structure and components of the power equipment, and detect visual faults such as cracks, corrosion and deformation; S3-3. Obtain information similar to or related to the current fault from historical cases through the expert knowledge base, and provide corresponding maintenance suggestions and measures; S4. Integrate the multi-source data fusion method with the deep learning fault diagnosis algorithm and optimize the execution efficiency of the algorithm; Among them, the linear function normalized value in step S2 data processing includes: Normalization is to unify the values ​​into the same range. Linear function normalization transforms the feature data linearly and scales the data proportionally to the range of 0 to 1. The calculation formula is as follows: X norm Indicates that the collected voltage, current, and temperature data are normalized; X represents characteristic data, X max and X min Represent the maximum and minimum values ​​of the characteristic data respectively; the zero-mean normalization method is used to standardize and normalize the noise points of the data during the experiment, and the data is mapped to the distribution of μ = 0, σ = 1. The normalized calculation is described as: Z=(X-μ) / σ Where μ is the assumed mean, σ is the standard deviation; Z represents the standardization and normalization of the voltage, current, and temperature data noise; The role of normalization and the iterative process of optimization training include two types of numerical features, and the features directly represented by numbers are distinguished from the categorical features that can only be represented by 0 and 1; the value range of the feature data voltage X1 is 0 to 10, and the value range of the current X2 is 0 to 1; under the same learning rate of the gradient descent method, since the value interval of X1 is larger than that of X2, the update speed of X1 iteration is faster than that of X2, and it takes a long time to find the optimal solution; normalizing X1 and X2, the update speed of the feature data voltage X1 and current X2 becomes closer, and the optimization algorithm is used to find the optimal solution more easily; The establishment of the Bayesian model in step S2 data processing includes: The probability formula of the Bayesian model is: P(B)=P(B|δ)=∑P(B|A)P(A) Among them, A is a disjoint event, B represents an event that occurs under the condition of δ; δ is the union of the necessary events A; P(B) is the "prior probability" of B, P(B|δ) is the class conditional probability of sample δ for class labeled B, P(A) is the "evidence" factor used for normalization, and P(B|A) is the class conditional probability of sample A for class labeled B; By deducing the conditional probability of event B, the posterior probability of event A is obtained as: Among them, A i is the i-th inferred event, A j is the jth known sample event; P(A i |B) is sample B for class labeled A i The class conditional probability, P(B|A i ) is sample B for class labeled A i The class conditional probability, P(B|A j ) is sample B for class labeled A j The class conditional probability, P(A i ) and P(A j ) is A i and A j The "prior probability" of; m represents the number of data; After data feature extraction and data classification processing in step S3-1, each feature value information is used to complete data prediction through a deep learning algorithm, and the deep learning algorithm includes a dendritic neural network DDNet; the algorithm formula is: Y=(W·Z)·Z Where W is the weight parameter matrix, Z is the vector representation of the input value, Y is the predicted output value, and W·Z is the Haddam product operation; The multi-source data fusion method in step S4 includes: S4-1. Calculate the confidence distance matrix of multi-source sensor data. The matrix represents the relative confidence of each sensor data. Assuming that the sensor data obeys the normal distribution, the confidence distance is expressed as: Among them, d xy is the relative confidence between the xth sensor and the yth sensor, δ i are the estimated model parameters; x i and x j represents the i-th and j-th fusion information values; r represents the value of each sensor; dr represents the derivative of r; S4-2. Measure the confidence and select a distance threshold to update the confidence distance matrix and obtain the correlation matrix. The threshold is selected as 0 or 1. If the distance is higher than the threshold, the correlation is reduced, and if the distance is lower than the threshold, the correlation matrix is ​​increased. D is the confidence distance matrix, and its expression is: Among them, u represents the value of relative confidence; S4-3. Filter multi-sensor data through correlation matrix to obtain fusion information; S4-4. Obtain the estimated value of the fusion model parameters through the function to complete the model data inference. The parameter estimation formula is expressed as: Among them, θ is the estimated value of the model parameter, δ0 and θ0 are the given initial values ​​of the parameters, δ i represents the value of the ith item, x i is the i-th fusion information value; The data feature extraction in step S3-1 includes: obtaining new attributes through the relationship between different data to change the original feature space; specifically: First, eliminate the correlation between features, then select the most important features among the irrelevant ones, and transform the existing features into other mutually perpendicular and irrelevant spaces, and select the important features with the largest distribution dispersion, that is, the smallest projection distance; The data classification in step S3-1 includes: establishing rules based on known samples in several categories, and assigning a model with labels corresponding to the categories of unknown samples. If the number of type divisions exceeds two, the number of divisions corresponds to the number of classifiers, and the data is expressed in the form of feature vectors; in the process of target detection, the selection of classifiers includes Adaboost, SVM and neural network; bounding box regression is to find a mapping relationship, and make the position and size of the candidate box generated in the candidate area step as close as possible to the position and size of the real standard box, so as to determine the category of the candidate box and obtain the position of the candidate box.

2. The method for analyzing power equipment failure based on multi-source data fusion according to claim 1 is characterized in that: The collection device for collecting the relevant data of the power equipment in step S1 includes various sensors and cameras installed on the power equipment; the collected relevant data includes different types and operating states of the power equipment.

3. The method for analyzing power equipment faults based on multi-source data fusion according to claim 1 is characterized in that: In step S3-2, the image of the surface of the power equipment is analyzed by regressing the image frame, and finally the result is optimized to obtain the optimal output.

Images