A fault detection method, device and storage medium for weak features of a photovoltaic power station

Abstract

The application discloses a kind of photovoltaic power station weak feature fault detection method, equipment and storage medium;The method includes extracting photovoltaic power station weak feature fault transient characteristics based on wavelet transform;Weak feature fault transient characteristics including rich information are divided into frequency bands, and the information difference on specific frequency band is amplified;Using the synchronism of measurement information, the same dimension feature expression matrix reflecting the fault characteristics of the entire photovoltaic power station is established;The cloud reference model is trained using the feature expression matrix;The cloud reference model trained is deployed on each photovoltaic power station edge side, and the data scale of edge migration is improved based on the data augmentation method of local sensitive hash, and the cloud reference model is fine-tuned according to the fault data characteristics in the specific photovoltaic power station, to achieve the optimal performance in the specific photovoltaic power station application.It has good detection accuracy, sensitivity and reliability.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant fault detection technology, and in particular to a method, equipment and storage medium for detecting weak characteristics of photovoltaic power plants. Background Technology

[0002] Low-current faults in photovoltaic power plants are difficult to detect effectively using traditional protection methods due to their small fault current and weak characteristics. This results in many fault scenarios going unidentified. Therefore, the first step in effectively diagnosing weak-feature faults is to improve their detection success rate. For this type of weak-feature fault, transient features are the most effective identification features. However, existing methods typically rely on threshold discrimination of a single feature value or comparison of transient feature quantities for detection, failing to fully utilize the rich information in transient features. This makes the detection methods susceptible to failure due to the complexity of the scenario and struggles to balance sensitivity and reliability. Artificial intelligence-based detection methods can obtain deeper features through data learning and can achieve detection without threshold settings. However, they often require large amounts of data for training, and many existing methods are only effective after training with specific photovoltaic power plant data, exhibiting poor topology generalization ability, which greatly limits the application of such methods in practical weak-feature fault detection. Summary of the Invention

[0003] The main technical problem solved by this invention is to provide a fault detection method for photovoltaic power plants with weak features, which solves the problem that existing methods are only effective after being trained with specific photovoltaic power plant data and have poor topology generalization ability.

[0004] To solve the above-mentioned technical problems, one technical solution adopted by the present invention is to provide a fault detection method for weak characteristics of photovoltaic power plants, comprising the following steps:

[0005] Step S1: Extract transient features of weak feature faults in photovoltaic power plants based on wavelet transform; divide the transient features of weak feature faults, which contain rich information, into frequency bands to amplify the information differences in specific frequency bands;

[0006] Step S2: Utilize the synchronicity of measurement information to establish a feature representation matrix of the same dimension that reflects the fault characteristics of the entire photovoltaic power station; use the feature representation matrix to train the cloud benchmark model;

[0007] Step S3: Migrate and deploy the trained cloud benchmark model to the edge of each photovoltaic power station. Based on the locality-sensitive hashing data augmentation method, increase the data scale of the edge migration. Fine-tune the cloud benchmark model according to the fault data characteristics in the specific photovoltaic power station to achieve optimal performance in the specific photovoltaic power station application.

[0008] This application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the method.

[0009] This application also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method described.

[0010] The beneficial effects of this invention are as follows: Based on the extraction of time-frequency transient features of weak-feature fault transient zero-sequence current using wavelet transform, this application proposes a feature representation matrix extraction method based on PCA, highlighting key global features while maintaining the same feature scale. Furthermore, a convolutional neural network cloud benchmark model for mining common knowledge of weak-feature fault transient features is trained in the cloud. Based on this, an edge deployment model is established through transfer learning, achieving efficient weak-feature fault detection for different photovoltaic power plant topologies. During the transfer process, the data level of the target photovoltaic power plant significantly affects the transfer performance. Therefore, a data augmentation method based on locality-sensitive hashing is proposed. This method combines efficient search under edge-cloud collaboration with cloud-based similar data to augment local data, specifically improving the transfer performance of the specific detection model. The proposed method, while addressing the bottlenecks in practical applications of artificial intelligence, effectively detects small-current faults with good detection accuracy, sensitivity, and reliability. Attached Figure Description

[0011] Figure 1 This is a flowchart according to an embodiment of the present invention;

[0012] Figure 2 This is a schematic diagram of the decomposed components and corresponding frequency bands of each layer according to an embodiment of the present invention;

[0013] Figure 3 This is a flowchart of the feature extraction process according to an embodiment of the present invention;

[0014] Figure 4 This is a schematic diagram of the specific structure of a Boltzmann machine according to an embodiment of the present invention;

[0015] Figure 5 This is a schematic diagram illustrating the working principle of RBM according to an embodiment of the present invention;

[0016] Figure 6 This is a schematic diagram of a deep belief network according to an embodiment of the present invention;

[0017] Figure 7 This is a flowchart of intelligent pre-diagnosis according to an embodiment of the present invention. Detailed Implementation

[0018] To facilitate understanding of the present invention, a more detailed description is provided below with reference to the accompanying drawings and specific embodiments. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0019] It should be noted that, unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0020] Figure 1 An embodiment of the fault detection method for weak characteristics of photovoltaic power plants according to the present invention is shown, including:

[0021] Step S1: Extract transient features of weak feature faults in photovoltaic power plants based on wavelet transform; divide the transient features of weak feature faults, which contain rich information, into frequency bands to amplify the information differences in specific frequency bands;

[0022] Step S2: Utilize the synchronicity of measurement information to establish a feature representation matrix of the same dimension that reflects the fault characteristics of the entire photovoltaic power station; use the feature representation matrix to train the cloud benchmark model;

[0023] Step S3: Migrate and deploy the trained cloud benchmark model to the edge of each photovoltaic power station. Based on the locality-sensitive hashing data augmentation method, increase the data scale of the edge migration. Fine-tune the cloud benchmark model according to the fault data characteristics in the specific photovoltaic power station to achieve optimal performance in the specific photovoltaic power station application.

[0024] This application proposes a PCA-based feature representation matrix extraction method to extract time-frequency transient features of weak-feature fault transient zero-sequence current, building upon wavelet transform extraction. This method highlights key global features while maintaining the same feature scale. Furthermore, a convolutional neural network cloud benchmark model for mining common knowledge of weak-feature fault transient features is trained in the cloud. Based on this, an edge-deployment model is established through transfer learning to achieve efficient weak-feature fault detection for different photovoltaic power plant topologies. During the transfer process, the data level of the target photovoltaic power plant significantly affects the transfer performance. Therefore, a data augmentation method based on locality-sensitive hashing is proposed. This method combines efficient search under edge-cloud collaboration with cloud-based similar data to augment local data, specifically improving the transfer performance of the specific detection model. The proposed method effectively detects small-current faults while addressing the bottlenecks in practical applications of artificial intelligence, exhibiting good detection accuracy, sensitivity, and reliability.

[0025] Transient Feature Extraction of Weak-Feature Faults in Photovoltaic Power Plants Based on Wavelet Transform: For low-current faults, effective identification can be achieved based on the transient characteristics of zero-sequence current. However, the transient process is short-lived. Therefore, this application chooses to further extract the transient time-frequency features of weak-feature faults using wavelet transform to enhance the identifiability of fault features. Wavelet transform uses "small wavelets" with variable frequency and duration as basis functions. Through scaling and translation operations, it subdivides the waveform data to be processed into different frequency bands. This avoids the limitations of Fourier transform for unbalanced signals and eliminates the need to add time-domain information by setting fixed panes as in short-time Fourier transform. Therefore, wavelet transform can be used to divide the information-rich transient features of weak-feature faults into frequency bands, amplifying the information differences in specific frequency bands.

[0026] Wavelets are a family of functions that satisfy the condition that their mean is 0 and that are locally defined in both the time and frequency domains. The mother wavelet is defined as... Then there is

[0027] By scaling and shifting the mother wavelet at different scales, a wavelet sequence can be obtained, which can be represented as follows in the continuous case:

[0028] in, For scaling, This represents the translation amount.

[0029] Let the signal to be processed be The continuous wavelet transform is defined as follows:

[0030] As can be seen from the formula, the continuous wavelet transform is used to transform the signal to be processed. The inner product with the wavelet sequence is computationally intensive and contains information redundancy. Therefore, in practical applications, the discrete wavelet transform is usually used, which involves scaling the wavelet sequence. Translational displacement Discretizing the wavelet transform yields the following definition:

[0031] Among them, the scaling scale Perform exponential discretization, let As for translation Perform linear discretization. .

[0032] To ensure the orthogonality of the wavelet decomposition signals and facilitate rapid computation, the orthogonal mother wavelet can be viewed as a set of mirror filters. The signal is decomposed quickly by passing it through these filters. Specifically, the signal is passed through a set of high-pass and low-pass filters corresponding to different frequency bands. The output of the low-pass filter is the approximate coefficient. This corresponds to the low-frequency band; the output after passing through the high-pass filter is the detail coefficient. For the corresponding high-frequency band, the specific formula is as follows:

[0033] in, and These are the decomposition coefficients of the low-pass and high-pass filters in the corresponding layers.

[0034] The sampling precision in this application is: Taking a 5-layer decomposition as an example, the components of each decomposed layer and the corresponding frequency bands of each layer are as follows: Figure 2 As shown.

[0035] After decomposing the input signal to obtain the wavelet coefficients of each layer in the above manner, the discrete wavelet reconstruction coefficients of each frequency band can be obtained through reconstruction, and the transformed characteristic signal can be obtained. All wavelet reconstruction coefficients have the same data length as the original signal and inherit the dimensions of the original signal.

[0036] The choice of mother wavelet has a significant impact on the transformed result; a suitable mother wavelet can effectively improve the discriminativeness of the decomposed features. For the weak-feature fault transient characteristics of photovoltaic power plants, this application study selects... As the mother wavelet used. The (Daubechies) wavelet set is a classic discrete orthogonal wavelet set. This indicates the order of the wavelet. Except for the first order There is no explicit expression for wavelet extrapolation. Its characteristic is that the higher the order, the better the frequency band division effect.

[0037] by Using a mother wavelet, a decomposition layer of 5 was selected to perform wavelet decomposition and reconstruction on typical weak-feature fault transient data. For low-current grounding faults, metallic grounding faults with the same initial phase angle were selected. The transient wavelet reconstruction coefficients among different low-current grounding faults showed high similarity, and the transient wavelet reconstruction coefficients of low-current grounding faults and high-resistance faults exhibited a certain trend of similarity. The anti-parallel DC source type high-resistance fault model fits the actual high-resistance fault transients well and can reflect the transient characteristics and trends of actual high-resistance faults to a certain extent. Simulation using this high-resistance fault model can express the characteristics of actual high-resistance faults to a certain extent, demonstrating practicality and feasibility.

[0038] To integrate fault data from multiple photovoltaic power plants and uncover their common characteristics, it is necessary to enhance and extract key features of each photovoltaic power plant and unify their scale, and jointly train a cloud benchmark diagnostic model to fully utilize the data potential of multiple photovoltaic power plants.

[0039] This application will establish a synchronous fault feature extraction method based on the edge and end side, and establish an applicable cloud benchmark model from the edge to the cloud according to the data characteristics.

[0040] Based on the analysis of the transient processes of weak-feature faults in photovoltaic power plants, it can be concluded that the transient zero-sequence current contains rich fault information. Wavelet transform can effectively extract time-frequency features across different frequency bands, and these features exhibit certain similarities across various weak-feature fault scenarios. To fully extract the weak-feature transient features, this application establishes a cloud benchmark model. This model constructs a feature representation matrix based on data from a micro synchronous phasor measurement device (D-PMU) in the photovoltaic power plant. The D-PMU can collect time-stamped waveforms and phasors of physical quantities such as three-phase voltage, three-phase current, and zero-sequence current in the photovoltaic power plant. The measurement data has a high sampling frequency and accuracy, and is synchronous.

[0041] By utilizing the synchronization of D-PMU measurement information at the edge of each photovoltaic power station, a feature expression matrix with the same dimension that can reflect the fault characteristics of the entire photovoltaic power station is established. First, transient information is further extracted. Wavelet transform is used to extract the reconstruction coefficients of each level from the transient zero-sequence current as transient features. To enhance the representativeness of the state of the entire photovoltaic power station, transient features are fused through procedural feature analysis (PCA) and finally combined to form a feature expression matrix with a unified data scale.

[0042] The process of generating the feature representation matrix can be divided into the following three steps:

[0043] Extracting transient features using discrete wavelet transform;

[0044] PCA is used to fuse and reduce the transient features of all extracted synchronous feature data at the same wavelet decomposition level to the same scale.

[0045] The reduced features are combined into a fixed-scale feature expression matrix that reflects the state of the photovoltaic power station at that time section.

[0046] (1) Extracting wavelet reconstruction coefficient features

[0047] Assume a photovoltaic power station includes For each D-PMU, then for the D-PMU The time window is set to The raw transient measurement data at time can be expressed as Discrete wavelet transform has been systematically explained in the previous section, and it can transform time series... Decomposed into approximate reconstruction coefficients and detail reconstruction coefficients at multiple resolutions, when the number of decomposition layers is hour, The reconstructed detail coefficients can be obtained through discrete wavelet decomposition. and approximation coefficients In discrete wavelet transform, the two most important performance factors are the selection of the mother wavelet and the decomposition derivative. The choice. Mother wavelet. The choice of the number of decomposition levels has been explained above. Considering that interference in photovoltaic power plants may generate a large number of low-frequency harmonics, the low-frequency spectrum should be subdivided as much as possible during selection to ensure reliability and distinguish it from disturbance signals. Meanwhile, according to the Nyquist sampling theorem, the sampling frequency should be at least twice the highest frequency of the signal of interest, and generally should be at least 2.56 times. That is, the lowest frequency band after decomposition should not be lower than the power frequency. The low-frequency features obtained from the excessively high number of decomposition layers (2.56 times higher) are redundant. Therefore, at a D-PMU sampling frequency of... Based on this, select the number of decomposition layers. The lowest frequency band is At this point, for a single D-PMU extracting features... The following (wherein, )

[0048] (2) Feature fusion and unified dimensionality reduction through PCA

[0049] In different photovoltaic power plants, the topology and the number of D-PMUs vary. To standardize the data scale and facilitate subsequent information fusion among multiple distribution networks, this application uses PCA to unify the data structure and enhance data features. PCA is a dimensionality reduction method that best reveals the internal structure of data. It can reconstruct high-dimensional data in a low-dimensional space by establishing a certain linear projection method, minimizing data loss during the dimensionality reduction process. The specific principle of PCA is as follows. Given a... 3D matrix , in order to Reduced to Taking the low-dimensional matrix as the target, the basic formula for PCA dimensionality reduction can be written as:

[0050]

[0051] in, The principal component matrix is ​​constructed by sorting the components by the amount of features retained. The transformation matrix can be obtained through the covariance matrix. To calculate, For matrix A diagonal matrix of eigenvalues ​​arranged in descending order. for eigenvector matrix, Original matrix The average value of each vector in the vector. This is the intermediate matrix. Therefore, we take... The first in the matrix The principal components are then obtained after dimensionality reduction. A low-dimensional matrix.

[0052] In the specific calculation process, singular value decomposition (SVD) can be used to calculate the covariance matrix. Let the eigenvalues ​​and eigenvectors of be such that the eigenvalue matrix is ​​. ,but It can be represented as

[0053] in, Represents a matrix arranged in descending order. Eigenvalues, at this point we have

[0054] Here This indicates the degree to which the dimensionality-reduced data retains the original data.

[0055] According to the PCA algorithm, synchronization features of different dimensions can be reduced to the same dimension, thereby fusing and enhancing the features of the entire photovoltaic power plant while unifying the data scale, and eliminating redundancy and noise caused by multiple measurement information. Let the D-PMU wavelet reconstruction data at the same decomposition level at this time be... ; Through transformation, the previous Each principal component PCs is represented as follows: .

[0056] (3) Form a global feature representation matrix with unified dimensions.

[0057] For the entire photovoltaic power plant, a fixed scale can be formed by combining the reduced-dimensional features. The feature representation matrix reflects the fault status of the photovoltaic power station within a corresponding time window. Parameters The characteristics of a photovoltaic power plant are related to its scale, electrical parameters, and fault transient characteristics. Based on simulation results, when... When set to 3, It can achieve The above means that at least [value] can be retained after dimensionality reduction. The original data characteristics. Time window length. Set to 64 (the number of samples for half a cycle of the D-PMU at 50Hz), the feature representation matrix can be generated in real time by moving the time window.

[0058] The entire feature extraction process is as follows Figure 3 As shown, this feature representation matrix extraction method can effectively extract and enhance transient information from the entire photovoltaic power plant within the corresponding time window, and form a fixed-scale feature representation matrix for unified integration into the cloud benchmark model for training and common feature learning.

[0059] Cloud benchmark model construction; Deep Belief Network (DBN) consists of an unsupervised learning feature extraction part and a supervised learning inverse fine-tuning part. DBN is a graphical model that learns deep features of data, and its main building block is a Restricted Boltzmann Machine (RBM). Each Composed of a two-layer network structure, the first layer, called the visible layer, reads the network data input. The RBM (Representative Business Model) uses a greedy learning algorithm to obtain data features and passes them to the second layer, called the hidden layer, which represents these features. Therefore, by stacking multiple RBMs, a higher-level, lower-dimensional feature representation of the data can be obtained.

[0060] The specific limitations of the Boltzmann machine are as follows: Figure 4 As shown. In a bipartite graph, it consists of two layers of network, one of which is the visible layer. Receive the training data. One layer is a hidden layer. This represents a high-level, low-dimensional representation of data features. The network satisfies the constraint that layers are fully bidirectionally connected to each other, but not connected to each other within layers, greatly improving the parallelism of network training. This model is called RBM.

[0061] In RBM, there is a weight representing the connection strength between any two connected neurons, typically denoted as . Each neuron also has a bias coefficient representing its weight. For neurons in the visible layer, the bias coefficient is generally denoted as... For hidden layer neurons, the bias coefficient is generally denoted as... Thus, the energy of an RBM This can be represented as: .

[0062] In the formula —Number of neurons in the visible layer; ——No. One visible layer neuron, ; ——No. The bias coefficient of each visible layer neuron. ; —Number of neurons in the hidden layer; ——No. One hidden layer neuron, ; ——No. The bias coefficients of each hidden layer neuron. ; 11th The first visible layer neuron and the second The connection weights between neurons in the hidden layer.

[0063] In an RBM, hidden layer neurons Probability of activation by neurons in the visible layer:

[0064]

[0065] In the formula Sigmoid function;

[0066] It is bidirectionally connected, therefore neurons in the visible layer can also be activated by neurons in the hidden layer:

[0067]

[0068] Neurons in the same layer are independent of each other, which is also the "restriction" that limits Boltzmann machines. Therefore, the probability density satisfies independence, and we can obtain:

[0069]

[0070] as well as:

[0071]

[0072] The above describes the basic structure of an RBM and its working principle, as follows: Figure 5 As shown, in the visible layer After inputting the data, through Hidden layers can be obtained And thus obtain the hidden layer Afterwards, through It also allows you to obtain the visible layer.

[0073] By adjusting the parameters, the visible layer obtained from the hidden layer can be... It can be maximally described The resulting hidden layer is another expression of the visible layer, and the result of the hidden layer can be used as a feature of the input data of the visible layer.

[0074] like Figure 6 As shown, a deep belief network consists of multiple Restricted Block Models (RBMs). As mentioned above, the visible layer of the first RBM layer serves as the receiving layer for training data, and its hidden layers serve as the feature outputs of this RBM layer, while also acting as the visible layers of the next RBM layer. These layers are stacked sequentially and combined with a backpropagation (BP) neural network at the top to form the entire deep belief network.

[0075] Deep Belief Network (RBM) model. Model training mainly consists of two steps. First, by stacking Restricted Belief Models (RBMs), deeper low-dimensional feature representations of the data are continuously obtained, ensuring the local optima of parameters in each layer of the model. Second, the Backpropagation (BP) algorithm is used to fine-tune the parameters of the entire model using labeled data to achieve the global optimum of the entire model's parameters.

[0076] A framework for intelligent pre-diagnosis based on deep learning, such as Figure 7 As shown. After feature selection, the selected features are combined with their corresponding health conditions to train the deep learning model. The pre-diagnosis process mainly includes four parts: sample collection, DBN pre-training, parameter fine-tuning, and fault diagnosis.

[0077] When using DBN models for data processing, parameter selection is crucial, as different parameter settings significantly impact DBN training results. Key DBN parameters include the weights and biases of the RBMs that constitute the DBN, DBN network node parameters, network depth, training learning rate, training batch size, and number of iterations. Setting DBN parameters first requires adjusting the connection weights of each layer of the RBMs that make up the DBN. , visible layer bias and hidden layer bias Perform initialization operations. Set connection weights. Perform initialization that conforms to a normal distribution, and initialize the explicit layer bias and the hidden layer bias with an initial value of 0 or a normal random distribution.

[0078] The structure of a Deep Learning Network (DBN) can be divided into two parts: network depth and the number of nodes in each layer. Network depth refers to the number of neural network layers in the DBN, while the number of nodes refers to the number of neurons in each layer. One of the main differences between deep learning and traditional neural networks is its multi-layered network structure. Increasing the number of hidden layers or network depth can reduce the gap in the number of nodes between layers, thereby improving feature extraction capabilities.

[0079] The choice of the number of nodes in each layer of the network has a significant impact on the model's feature extraction capability. The number of neurons in the input layer is determined by the dimensionality of the input data, while the number of neurons in the output layer is determined by the types of labels on the training samples. The number of nodes in the hidden layer of a DBN can be set using methods such as constant-value, increasing-value, decreasing-value, convex, and concave shapes.

[0080] Different learning rates have a significant impact on the learning ability of DBN networks. While an excessively high learning rate can accelerate learning, it can easily lead the network into local optima. Conversely, a very low learning rate, while ensuring relatively stable network training, will drastically increase the number of iterations needed to obtain an optimal network structure, resulting in a significant increase in training time. Therefore, setting an appropriate learning rate is necessary for different network training processes. Based on empirical values, a suitable learning rate can be set to... between.

[0081] The sample input process employs the Gibbs sampling method to generate samples with a multivariate probability distribution. After multiple samplings, a training process simulating the overall population data can be obtained, thus better reflecting the characteristics of the population. The minimum training batch size is the number of categories to ensure that each batch contains samples from all categories. Batch sizes that are too large result in poor simulation performance, while batch sizes that are too small significantly increase computation time. During training, the training batch size can be appropriately selected to ensure excellent sampling training simulation results and reasonable training time.

[0082] From Cloud to Edge: Migration of Edge Deployment Models; After establishing the cloud baseline model, it is necessary to migrate and deploy this model at the edge of each photovoltaic power station to achieve optimal performance in specific distribution network applications. This application proposes a data augmentation method based on locality-sensitive hashing to increase the data scale used for edge migration, and introduces the edge model deployment method from cloud to edge based on this.

[0083] Data augmentation methods based on locality-sensitive hashing; in the process of generating cloud benchmark models, the focus is on the ability to mine common knowledge of weak feature faults. In the cloud-to-edge model migration and deployment, the model needs to be fine-tuned according to the characteristics of fault data in specific photovoltaic power plants to improve the application performance of the model in a targeted manner.

[0084] In this process, the availability of data at specific endpoints will have a certain impact on the model transfer performance. To improve the training or transfer capabilities of small sample data, minor modifications are made to the existing dataset to increase the total amount of data. Common methods in fields such as image recognition include cropping, translation, rotation, and mirroring. However, many data augmentation techniques are meaningless in weak feature fault diagnosis and are difficult to apply directly.

[0085] Therefore, this application proposes a data augmentation algorithm based on locality-sensitive hashing in a cloud-edge architecture, which utilizes the interaction and collaboration between the edge and the cloud to leverage local edge data and cloud big data to increase the scale of local data used for transfer training.

[0086] The amount of usable weak-feature fault data in a single photovoltaic power plant is often very limited, while cloud databases contain feature data from multiple edge devices, potentially including feature data from other sources with high similarity. Based on this potential similarity, searching for feature information similar to the target photovoltaic power plant can effectively expand the feature dataset and improve the transfer efficiency when deploying models at the edge. However, for high-dimensional feature data search, traversal search will bring high time and computational costs, making it impractical. Therefore, a method using Locality Sensitive Hashing (LSH) is proposed to quickly and efficiently search for similar fault feature data samples from the cloud. LSH uses hashing to map large sets of data objects to smaller hash values. Once established, if two objects are close to each other, their hash values ​​are also highly likely to maintain their similarity. Therefore, searching can be performed using hash values ​​to greatly improve search efficiency.

[0087] The main idea behind Locality Sensitive Hashing (LSH) is to map data to a hash table using a hash function. By using the hash table's index, the overall search time complexity can be reduced from linear search. Down to This approach effectively handles the problem of approximating high-dimensional data, significantly improving the efficiency of retrieving similar data from cloud databases based on edge features. Under certain similarity metrics, locality-sensitive hashing (LSH) makes it easier for similar objects to be mapped into the same bucket than dissimilar objects. During data searching, data objects within the same bucket are considered as candidate objects, and the distance between candidate objects and the query object is calculated sequentially. To avoid insufficient similarity between candidate objects hashed into buckets and their nearest neighbors, multiple hash tables can be maintained during the creation process using different hash functions.

[0088] The specific performance of the proposed algorithm depends on the setting of the locality-sensitive hash function. Given a set of hash functions... Euclidean space Mapped to hash coding space In the case of any two points and Random selection A function And analyze The probability of a hash function. A hash function is called a function that satisfies both of the following conditions. for Sensitive (make) ):

[0089] when hour, ;

[0090] when hour, .

[0091] Based on the feature representation matrix extraction method proposed above, the size of each feature representation matrix is... 64. After expanding the data into a 1152-dimensional column vector, this application employs a random threshold hash function, setting it to include... One key and A hash function, that is, randomly selected from the column vector. Each time, the number of keys selected is [number] different key locations. Through this Each corresponding key is normalized to form a corresponding threshold hash function, thereby forming the corresponding... There are several hash tables. In practical applications, increasing the number of hash functions used can improve the reliability of search results, but the average search time will also increase accordingly. When using threshold hashing, the number of keys affects the number of generated buckets, which in turn affects the uniformity of distribution among the buckets. Too few keys may result in too many corresponding elements in some buckets. and The settings and specific data, the number of data points, etc., are all related. For the weak feature fault feature expression matrix generated in this application, the selection... The hash table exhibits a good bucket distribution, averaging approximately 600-800 buckets per hash table during experiments with 3000 sets of data, and demonstrates high search efficiency. Hash table creation can be performed offline, allowing all necessary hash tables to be pre-built in the cloud to store the hash values ​​of all features. When the target photovoltaic power station sends a data augmentation command to the cloud server, the cloud server can quickly search for a certain number of similar feature data using locality-sensitive hashing and send them back to the edge computing platform. Finally, on the edge side corresponding to the target photovoltaic power station, all feature representation matrices, including the original data and the similar data searched and downloaded using the extracted augmentation method, are multiplied by a scaling factor. Further increase the scale of data, The values ​​are set in increments of 0.01, ranging from 0.94 to 1.05. This is achieved by setting a coefficient approximately equal to 1. This can, to a certain extent, simply and effectively enhance the generalization ability for different fault locations and fault impedances.

[0092] Edge deployment model based on transfer learning: Due to the differences between different photovoltaic power plants, directly running the cloud baseline model in a single photovoltaic power plant may not yield good results. Therefore, based on data augmentation completed under edge-cloud collaboration, the cloud baseline model can be transferred from the cloud to the edge through fine-tuning using data from specific edge devices.

[0093] Fine-tuning refers to the process of freezing certain layer parameters in the DBN of the baseline cloud model and retraining other unfrozen layers using data from the target photovoltaic power plant. In DBN, earlier layers typically contain more general features, while later layers provide more specific details for classification. The last fully connected layer can be considered a classifier. The cloud model is trained using multi-source data from a large number of photovoltaic power plants, focusing on mining the common characteristics of data during weak feature faults. Therefore, it can be considered that the convolutional-pooling layer parameters in the baseline cloud DBN model can effectively mine relevant knowledge representations in synchronous global features. Therefore, this application adopts a migration strategy of freezing all hidden layers and fine-tuning the last two fully connected layers during transfer fine-tuning. This strategy effectively inherits the knowledge mining capabilities of the cloud model obtained from a large amount of data while ensuring the efficiency of local deployment.

[0094] Considering the differences in actual operating environment, communication conditions, and data security requirements of various edge photovoltaic power stations, based on the edge-cloud collaborative framework, two different operating modes—online and offline—are presented for the application of the proposed method. The online operation mode is usually the primary choice for photovoltaic power stations. In this mode, data transmission between the edge and cloud is conducted through interactive collaboration, and the model is updated iteratively. For photovoltaic power stations with poor communication conditions or high data confidentiality requirements, the offline operation mode can be adopted. In this mode, only one-way transmission collaboration occurs between the edge and cloud, eliminating the need to upload feature data from the edge to the cloud server, effectively expanding the applicability of the diagnostic method. The specific operating procedures under the two modes are as follows.

[0095] In online operation mode, historical data is first uploaded from the edge to the cloud as raw training data, forming the initial data foundation for the cloud baseline model. After model deployment at the edge, real-time monitoring and diagnosis are performed. When a diagnostic fault occurs, relevant data is stored locally at the edge, tagged accordingly after event verification, and the local model is updated and uploaded to the cloud (regardless of whether the verification result confirms a fault). The corresponding parameters of the cloud baseline model are updated based on this data. After a certain period of time, migration and edge deployment are performed again via the cloud model, iteratively updating the local model to achieve cyclical interactive upgrades between the cloud and edge models.

[0096] In offline operation mode, there is no need to upload historical data to the cloud in the initial stage. The cloud baseline model trained using other edge data is deployed locally on the edge. After deployment, the model is also deployed on the edge for real-time monitoring and diagnosis. When a diagnostic failure occurs, the model is updated on the edge following the same process: storing data locally, verifying and labeling it, and updating the local model. However, the data is no longer uploaded to the cloud; it is only recorded and stored locally. After a certain period of time, the model is migrated again via the cloud model and deployed using the updated local dataset for a one-way model upgrade.

[0097] For edge-deployed models specifically applied at the edge, the diagnostic approach is consistent in both operating modes. When an abnormal state occurs, it is temporarily stored locally. After verification and the addition of accurate label information, the data is stored in the edge database, and the edge model is updated simultaneously. This ensures that the edge model continuously evolves in-situ during operation, improving its localization characteristics. During the update process, regardless of whether the label after verification of the abnormal state data matches the model's judgment, the correct label must be added before updating and propagating to enhance the model's robustness and resistance to disturbances. However, in online operating mode, data is uploaded to the cloud in real time to update the cloud baseline model. Furthermore, during the initial establishment phase of the cloud model, data is collected from the edge's fault database and used to train the model. This edge-to-cloud data upload allows the cloud baseline model to continuously iterate and evolve based on data from multiple edge devices. It also ensures that the baseline cloud model incorporates the feature information of the edge data during its construction and updates, resulting in better performance when migrating and deploying to the corresponding edge model. In offline operating mode, local data does not interact with the cloud; it is only used to assist in deployment when migrating the cloud baseline model after a certain period. At the same time, it is also impossible to augment data based on locality-sensitive hashing, which will result in some loss of diagnostic performance of the edge model.

[0098] This application, building upon wavelet transform for extracting time-frequency transient features of weak-feature fault transient zero-sequence current, proposes a PCA-based feature representation matrix extraction method, highlighting key global features while maintaining the same feature scale. Furthermore, a convolutional neural network cloud benchmark model for mining common knowledge of weak-feature fault transient features is trained in the cloud. Based on this, an edge-deployment model is established through transfer learning, achieving efficient weak-feature fault detection for different photovoltaic power plant topologies. During the transfer process, the data level of the target photovoltaic power plant significantly affects the transfer performance. Therefore, a data augmentation method based on locality-sensitive hashing is proposed. This method combines efficient search under edge-cloud collaboration with cloud-based similar data to augment local data, specifically improving the transfer performance of the specific detection model. The proposed method, while addressing the bottlenecks in practical applications of artificial intelligence, effectively detects small-current faults with good detection accuracy, sensitivity, and reliability.

[0099] The above are merely embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A fault detection method for weak characteristics in photovoltaic power plants, characterized in that, Including the following steps: Step S1: Extract transient features of weak feature faults in photovoltaic power plants based on wavelet transform; divide the transient features of weak feature faults, which contain rich information, into frequency bands to amplify the information differences in specific frequency bands; Step S2: Utilize the synchronicity of measurement information to establish a feature expression matrix with the same dimensions that reflects the fault characteristics of the entire photovoltaic power station; The cloud baseline model is trained using the feature representation matrix; Step S3: Migrate and deploy the trained cloud benchmark model to the edge of each photovoltaic power station. Based on the locality-sensitive hashing data augmentation method, increase the data scale of the edge migration. Fine-tune the cloud benchmark model according to the fault data characteristics in the specific photovoltaic power station to achieve optimal performance in the specific photovoltaic power station application.

2. The fault detection method for weak characteristics of photovoltaic power plants according to claim 1, characterized in that, Wavelets are a family of functions that satisfy the condition that the mean is 0 and that are locally defined in both the time and frequency domains. The mother wavelet is defined as follows: Then there is By scaling and shifting the mother wavelet at different scales, a wavelet sequence can be obtained, which can be represented as follows in the continuous case: in, For scaling, This is the translation amount; Let the signal to be processed be The continuous wavelet transform is defined as follows: Using discrete wavelet transform, the scaling factor is increased. Horizontal displacement Discretize the wavelet transform to obtain the following definition: Among them, the scaling scale Perform exponential discretization, let For translation amount Perform linear discretization. .

3. The fault detection method for weak characteristics of photovoltaic power plants according to claim 2, characterized in that, The orthogonal mother wavelet is treated as a set of mirror filters. The signal is decomposed rapidly by passing it through the filters. The signal is then passed through a set of high-pass and low-pass filters corresponding to different frequency bands. The output of the low-pass filter is the approximation coefficient. This corresponds to the low-frequency band; the output after passing through the high-pass filter is the detail coefficient. Corresponding to the high-frequency band, the specific formula is as follows: in, and These are the decomposition coefficients of the low-pass and high-pass filters in the corresponding layers.

4. The fault detection method for weak characteristics of photovoltaic power plants according to claim 3, characterized in that, The process of generating the feature representation matrix can be divided into the following three steps: (1) Extract transient features through discrete wavelet transform; (2) PCA is used to fuse and reduce the transient features of all extracted synchronous feature data at the same wavelet decomposition level to the same scale; (3) Combine the reduced features into a fixed-scale feature expression matrix that reflects the state of the photovoltaic power station at that time section.

5. The fault detection method for weak characteristics of photovoltaic power plants according to claim 4, characterized in that, The cloud benchmark model includes multiple RMBs. Each RMB has a two-layer network structure. The first layer network is used to read network data input, obtain data features through a greedy learning algorithm, and pass them to the second layer network. The second layer network is used to represent the data features. Between any two connected neurons, there is a weight representing the connection strength, typically denoted as . Each neuron also has a bias coefficient representing its weight. For neurons in the visible layer, the bias coefficient is generally denoted as... For hidden layer neurons, the bias coefficient is generally denoted as... ; The energy of one RBM This can be expressed as: In the formula —Number of neurons in the visible layer; ——No. One visible layer neuron, ; ——No. The bias coefficient of each visible layer neuron. ; —Number of neurons in the hidden layer; ——No. One hidden layer neuron, ; ——No. The bias coefficients of each hidden layer neuron. ; ——No. The first visible layer neuron and the second The connection weights between neurons in the hidden layer; In an RBM, hidden layer neurons Probability of activation by neurons in the visible layer: In the formula Sigmoid function; Neurons in the visible layer are also activated by neurons in the hidden layer: We can obtain: as well as: 。 6. The fault detection method for weak characteristics of photovoltaic power plants according to claim 5, characterized in that, Given a set of hash functions Euclidean space Mapped to hash coding space In the case of any two points and Random selection A function And analyze The probability of a hash function; if both of the following conditions are met, it is called a hash function. for Sensitive (make) ): when hour, ; when hour, .

7. The fault detection method for weak characteristics of photovoltaic power plants according to claim 6, characterized in that, The process involves freezing certain layer parameters of the DBN in the baseline cloud model and then retraining other unfrozen layers using data from the target photovoltaic power station; a migration strategy involves freezing all hidden layers and fine-tuning the last two fully connected layers.

8. The fault detection method for weak characteristics of photovoltaic power plants according to claim 7, characterized in that, The application offers two different operating modes: online and offline. In the online mode, the edge and cloud interact and collaborate to transmit data, and use iterative loops to update the model. In the offline mode, the edge and cloud collaborate only through one-way transmission.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images