Adaptive detection method for multi-variable dc fault arc based on similarity measure and transfer learning

Abstract

The application discloses a kind of multi-variable DC fault arc adaptive detection method based on similarity measure and transfer learning, to the system output current of different time period applies wavelet packet decomposition and coefficient reconstruction processing after extraction feature, and input to state identification model to carry out multi-period fault arc judgment, to current working condition through machine learning evaluation feature similarity measure value D and model output probability distribution P joint distribution mode, according to this, the extracted feature is distinguished into unidentifiable working condition type data and identifiable working condition type data, using field adaptive strategy to the simplest feature group of unidentifiable working condition type data carries out transfer learning, to update state identification model, to realize real-time accurate detection of DC fault arc under complex and variable working condition in this way.

Description

Technical Field

[0001] This invention belongs to the field of electrical fault detection technology. It involves using similarity measurement and transfer learning to mine the commonalities and differences in the electrical characteristics of fault arcs under different operating conditions (referring to working equipment under different application scenarios and source load types) during the operation of a DC system. In this way, it utilizes an operating condition state (divided into normal, fault arc, and arc-like) identification algorithm that can adapt to changes in operating conditions to achieve real-time and accurate detection of variable DC fault arcs (i.e., DC fault arcs under complex and variable operating conditions). Background Technology

[0002] Traditional arc fault detection methods typically require machine learning with a large training distribution, and the training and testing data must follow an independent and identically distributed (IOD) pattern. However, in reality, electrical equipment operates in harsh environments with complex and variable conditions, resulting in data distribution differences under different conditions. For conditions where fault data is scarce or labeled data is lacking, such as photovoltaic DC power generation systems, long-distance high-voltage DC power transmission, and electric vehicle charging stations, traditional machine learning-based arc fault detection methods are difficult to apply. With the expansion of 5G base stations, data centers, and other infrastructure, the variability of application scenarios and load types makes DC arc fault characteristics more complex and diverse. If the arc fault detection method cannot adaptively determine the variable DC arc faults, the DC arc fault detection device may fail to operate. If an arc fault causes a fire, it can lead to serious loss of life and property. Furthermore, if the problem is misjudged as an arc fault, meaning a normally operating DC system is identified as experiencing an arc fault, the DC arc fault detection device may malfunction due to misjudgment, causing system shutdown, reduced efficiency, and wasted resources. Meanwhile, real fault arc data obtained in industrial environments are generally unlabeled. For traditional fault arc detection methods based on machine learning, unlabeled data cannot be used for model training.

[0003] Given that differences in application scenarios and variations in source load types can lead to changes in fault arcs, and that traditional fault arc detection methods based on machine learning are no longer applicable when there is a lack of effective fault data for equipment in newly built DC systems, the safety protection issues of DC systems under dual-carbon backgrounds are more prominent. Therefore, it is urgent to propose an adaptive detection method for variable DC fault arcs that can extract the essential characteristics of fault arcs during DC system operation and accurately identify the moment when fault arcs occur.

[0004] In the field of fault diagnosis, deep learning significantly outperforms machine learning. It learns from input data layer by layer, from low to high levels, gradually extracting specific features from general surface features. Ultimately, the network establishes a function mapping relationship from input to output, demonstrating powerful data mining capabilities. Compared to machine learning or other deep learning methods, Convolutional Neural Networks (CNNs) have unparalleled advantages in natural language and image classification. CNNs are trained using gradient descent, with forward propagation calculating the network output value and backpropagation adjusting the network weights and biases. Local connectivity and weight sharing are the two most significant characteristics of CNNs. Each neuron in a convolutional layer is locally connected to the feature map of the previous layer, and weights are shared among convolutional kernels. Compared to fully connected networks, CNNs have fewer network connections and weight parameters, lower computational complexity, stronger robustness and fault tolerance, and are easier to train and optimize. However, most research applying CNNs to fault arc detection is based on fault arc data collected in laboratories, which hinders the application of related fault arc detection algorithms in industrial environments. Furthermore, the probability distribution of data obtained in the laboratory differs significantly from that of real data obtained in engineering practice. When intelligent fault diagnosis methods are applied across domains, their performance will drop sharply due to the different data distributions.

[0005] The DC fault arc detection method disclosed in Chinese patent CN114062880A achieves rapid and accurate detection of DC fault arcs under complex multi-source load system operation conditions through adaptive signal processing and CART tree ensemble learning. However, this detection method is limited to application scenarios with fixed system architecture. With the unknown scenarios of topology changes caused by the flexible access of source load devices in the system architecture, the detection of DC fault arcs faces the problem of failure. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive detection method for variable DC fault arcs based on similarity measurement and transfer learning.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] An adaptive detection method for variable DC fault arcs based on similarity measurement and transfer learning, comprising the following steps:

[0009] 1) Perform time-frequency transformation on the acquired system output current, and then extract its features;

[0010] 2) Input the features extracted in step 1 into the deep learning-based state identification model to determine the system operating status by distinguishing between DC fault arcs and arc-like structures; evaluate the joint distribution pattern between the feature similarity metric and the model output probability distribution for the current operating status through machine learning, thereby classifying the features extracted in step 1 as either unidentifiable or identifiable operating status data, and selecting unidentifiable operating status data to update the deep learning-based state identification model through transfer learning, wherein the feature similarity metric is the distance between the features extracted in step 1 and the training features in the deep learning-based state identification model.

[0011] Preferably, step 1 specifically includes the following steps:

[0012] 1.1) The output current of the DC system is sampled point by point at the sampling frequency fs to obtain the current signal x for one analysis period. q The number of sampling points for this analysis period is q. Proceed to step 1.2.

[0013] 1.2) The acquired current signal x q After performing time-frequency transformation, a frequency band reflecting fault arc information is selected, and features (one-dimensional features) are constructed according to the many-to-one mapping form of fault arc information points in the frequency band during the current analysis period.

[0014] Preferably, q is 8000 to 12000, fs is 200 kHz to 1 MHz, and the analysis time period is q / fs.

[0015] Preferably, in step 1.2, the fault arc information point is the point corresponding to the current signal x. q Reconstructed wavelet packet coefficients of consistent length; these wavelet packet coefficients are obtained by using the Rbio3.1 mother wavelet to reconstruct the current signal x. q The wavelet packet coefficients in the frequency band from 31250Hz to 62500Hz in the time-frequency matrix obtained by wavelet packet decomposition are reconstructed.

[0016] Preferably, in step 2, the state identification model based on deep learning adopts a one-dimensional convolutional neural network, so that the input features (specifically the extracted one-dimensional features) pass through two convolutional layers, a batch normalization layer, a pooling layer and a fully connected layer in sequence, and outputs a one-dimensional identification result, thereby judging the arc occurrence of the DC system one by one according to the analysis period.

[0017] Preferably, the initial training parameters of the one-dimensional convolutional neural network include: 15 to 30 iterations, 75 to 125 batch size, 0.0005 to 0.0015 learning rate, 0.7 to 1.0 first momentum decay factor beta1, and 0.8 to 1.0 infinity norm decay factor beta2. The training set used in the initial training of the one-dimensional convolutional neural network consists of features extracted under known operating conditions (training features) and operating condition status labels.

[0018] Preferably, in step 2, the distinction between DC fault arc and arc-like state specifically includes the following steps: After inputting features into the state identification model, if the model outputs an identification result indicating that the system is in an arc-generating state (i.e., an arc occurs) during the current analysis period (e.g., an output of 1), then multi-period fault arc judgment is performed starting from the current analysis period (conversely, if the system is in a normal state without arc occurrence, e.g., an output of 0, then multi-period fault arc judgment is not performed). If the state identification model outputs an identification result indicating that the system is in an arc-generating state at a frequency exceeding a set frequency based on subsequently input features (specifically, one-dimensional features extracted in subsequent consecutive analysis periods), then the system operating condition is determined to be a fault arc (and a fault arc cutoff signal is output to stop system operation). Otherwise (i.e., the state identification model outputs an identification result indicating that the system is in an arc-generating state at a frequency lower than a set frequency based on subsequently input features, or even no further identification result indicating that the system is in an arc-generating state), the system operating condition is determined to be an arc-like state.

[0019] Preferably, the parameters for judging the multi-time period fault arc include: the total number of analysis periods is 8 to 15, and the proportion of the number of arc generation state identifications required to determine a fault arc is ≥80%.

[0020] Preferably, in step 2, the output probability distribution of the model (e.g., the state identification model using a one-dimensional convolutional neural network as described above) is obtained through probability distribution judgment. The probability distribution judgment specifically includes the following steps: starting from the analysis period in which the identification result indicates that the system is in an arc-generating state, determine the proportion of analysis periods in the consecutive m analysis periods starting from this analysis period in which the identification result indicates that the system is in an arc-generating state. Determine whether this proportion meets the probability distribution requirements. If it does, record this proportion as probability P and output probability P as the model output probability distribution result for the consecutive N (N≤400) analysis periods starting from the analysis period (N>m, and N is an integer multiple of m). If it does not meet the requirements, then sequentially judge the identification results in the (m+1)th to 2mth, (2m+1)th to 3mth, ... (N-m+1)th to Nth analysis periods within the above N analysis periods. To determine whether the proportion of analysis periods in which the system is in an arc-generating state meets the probability distribution requirements, if a certain consecutive m analysis periods meet the probability distribution requirements, then the corresponding proportion of the m consecutive analysis periods is denoted as probability P, and probability P is output as the model output probability distribution result within the above N analysis periods. If the proportion of analysis periods in which the system is identified as being in an arc-generating state still does not meet the probability distribution requirements until the last m consecutive analysis periods within the above N analysis periods, then the average of the corresponding proportions of all m analysis periods contained in the above N analysis periods is taken, and this average is denoted as probability P, and probability P is output as the model output probability distribution result within the above N analysis periods (by dividing the N analysis periods into m analysis periods, as many ways as possible to determine probability P are provided, ensuring that the model output probability distribution result can be obtained).

[0021] Preferably, the probability distribution judgment further includes the following steps: during the process of statistically analyzing the proportion of analysis periods in any consecutive m analysis periods where the identification result indicates that the system is in an arc-generating state, if the state identification model outputs an identification result representing that the system is in an arc-generating state in the current analysis period based on the features extracted from a certain analysis period, then multi-objective particle swarm optimization is used to update the parameters of the probability distribution judgment to more accurately distinguish between unidentifiable working condition type data and identifiable working condition type data. The parameters include the upper limit P2 (i.e., the upper limit of the judgment probability P2) and the lower limit P1 (i.e., the lower limit of the judgment probability P1) required by the probability distribution, the initial value of P1 is 10% to 15%, and the initial value of P2 is 50% to 55%.

[0022] Preferably, m is 20 to 40.

[0023] Preferably, in step 2, the similarity measurement (which is unrelated to the system operating condition identification result, i.e., regardless of the identification result, a similarity measurement must be performed to distinguish between identifiable and unidentifiable operating condition data, and only one measurement value can be obtained for multiple analysis periods under one operating condition, rather than obtaining a measurement value for each analysis period separately) is calculated based on the maximum mean difference based on particle swarm optimization. The similarity measurement value (denoted as measurement value D) between the features extracted from the above N analysis periods and the training features in the state identification model based on deep learning is calculated. Based on the principle of universality, the maximum mean difference adopts the Gaussian kernel function, which can fit any distribution to the Hilbert space. The parameter σ of the Gaussian kernel function is 6000 to 7000.

[0024] Preferably, in step 2, the discrimination between unidentifiable operating condition type data and identifiable operating condition type data specifically includes the following steps: inputting the metric value D and the probability P into the support vector machine, and having the support vector machine output the judgment result of identifiable or unidentifiable operating condition type data. That is, within a state identification model update cycle, the features extracted from multiple analysis periods are used to determine whether the current operating condition is identifiable, thereby determining whether the operating condition that has occurred is an unidentifiable operating condition (if it is an unidentifiable operating condition, then the features extracted from the corresponding multiple analysis periods are unidentifiable operating condition type data). Thus, when the operating condition or a similar operating condition is encountered again after the state identification model is updated, the misidentification of the operating condition will no longer occur.

[0025] Preferably, the state identification model is updated after identifying unidentifiable working condition type data. Specifically, the update cycle is when the number C of unidentifiable working condition type data reaches 10 to 20, or the state identification model is updated in a fixed time unit (i.e., update cycle).

[0026] Preferably, in step 2, the transfer learning is based on a domain adaptation strategy, specifically including the following steps: performing pairwise cross-measures on the unidentifiable working condition type data determined within the update cycle, storing the cross-measure results in the distance metric matrix Dist, and then performing threshold judgment (the threshold is set to 0.005~0.015), storing the threshold judgment results in the matrix Dist_new (if it exceeds the threshold, it is marked as 1, otherwise it is marked as 0, that is, the elements in Dist_new are marked as 1 / 0 corresponding to each element in Dist); determining the simplest feature group according to the requirement of covering all unidentifiable working condition type data with the fewest number of unidentifiable working condition type data (that is, selecting as few rows as possible in the Dist_new matrix for OR operation to obtain as many row vectors as possible with 1), and using the domain adaptation method in combination with the simplest feature group to train the state identification model.

[0027] Preferably, training the state identification model in conjunction with the simplest feature set refers to training using different types of unidentifiable working condition data within the simplest feature set in a certain order. Each training session specifically includes the following steps: based on the feature fuzzy sample principle of the maximum effect classification decision boundary, the maximum classifier difference domain adaptive method is selected, that is, by maximizing the difference between two different classifiers (fully connected layers) containing the conditional distribution information of the source domain (training set of a one-dimensional convolutional neural network), the sample features of the target domain (simplest feature set) rejected by the source domain are found, and then the feature extractor (convolutional layer, batch normalization layer, pooling layer) is trained with the goal of minimizing the classifier difference, and features that conform to the conditional distribution of the source domain are regenerated, thereby adapting to the conditional distribution.

[0028] Preferably, the simplest feature group is determined by a compromise based on the number of unidentifiable working condition types of the simplest feature group, the training time, and the detection accuracy requirements of the state identification model.

[0029] An adaptive detection device for variable DC fault arc based on similarity measurement and transfer learning includes a signal analysis module, a multi-time fault arc judgment module, a similarity measurement module, an operating condition judgment module, and a state identification model update module.

[0030] The signal analysis module is used to extract features after performing time-frequency transformation processing on the system output current collected according to the analysis period.

[0031] The multi-period fault arc judgment module is used to input the features extracted by the signal analysis module into the state identification model based on deep learning, thereby determining the system operating state by distinguishing between DC fault arcs and arc-like objects, and to make probability distribution judgments on the output results of the state identification model.

[0032] The similarity measurement module is used to calculate the distance value of the similarity measurement between the features extracted by the signal analysis module and the training features in the deep learning-based state identification model;

[0033] The working condition judgment module is used to evaluate the joint distribution pattern between the feature similarity metric value and the model output probability distribution of the current working condition through machine learning, thereby identifying the features extracted by the signal analysis module as either unidentifiable working condition type data or identifiable working condition type data.

[0034] The state identification model update module is used to select unidentifiable working condition type data and use it to update the deep learning-based state identification model through transfer learning.

[0035] The beneficial effects of this invention are reflected in:

[0036] This invention identifies the collected current signals in real time and updates the state identification model periodically using the identified unidentifiable operating condition data. This makes the fault arc detection method applicable to variable operating condition DC systems where the topology changes due to source load access, or to complex and diverse practical engineering scenarios that differ from laboratory DC system application scenarios. It has high versatility and the ability to identify fault arcs under complex and variable operating conditions. While ensuring the normal and stable operation of the DC system, it solves the problem of malfunction of the DC-side fault arc detection device when the DC system faces transient processes such as power changes and startup.

[0037] The fault arc detection method proposed in this invention can be effectively integrated with existing detection framework devices to achieve adaptive detection of variable DC fault arcs without significantly increasing hardware costs, and has broad market promotion and application prospects.

[0038] Furthermore, by inputting the metric D and probability P into the support vector machine, this invention achieves accurate identification of unidentifiable working condition types and enables transfer learning to autonomously determine the timing of transfer, adaptively performing transfer learning only when a small number of unidentifiable working conditions occur, thus significantly reducing the computational load required for model updates.

[0039] Furthermore, the transfer learning of this invention is based on a domain-adaptive strategy, which does not require state labeling for unidentifiable working conditions in actual engineering scenarios. Moreover, a small amount of unidentifiable working condition data can obtain training results with high recognition accuracy, thereby improving the generalization ability of the state identification model. This enhances the engineering application adaptability of the fault arc detection method and corresponding detection device, and improves the adaptive detection capability of fault arcs under varying working conditions.

[0040] Furthermore, this invention determines the simplest feature group according to the requirement of full coverage of data of unidentifiable working conditions, which can update the state identification model with the least amount of data, reduce the amount of training data and training time, and thus update the state identification model more quickly after obtaining the updated information, reducing the lag in system detection. Attached Figure Description

[0041] Figure 1 This is a flowchart of the adaptive detection method for multivariable DC fault arcs in an embodiment of the present invention.

[0042] Figure 2 This describes the DC fault arc detection process under known operating conditions (the system component is an induction cooker). (The waveform at the top is the current signal, and the waveform at the bottom is the current signal characteristic.)

[0043] Figure 3The accuracy of DC fault arc detection using a one-dimensional convolutional neural network trained with the current signal characteristics of a known operating condition (the system component is an induction cooker) under various different operating conditions (new system components).

[0044] Figure 4 The relationship between the distance value obtained from the similarity measurement of current signal characteristics under identifiable and unidentifiable operating conditions and the accuracy of DC fault arc detection.

[0045] Figure 5 The accuracy of DC fault arc detection was compared between the state identification models before and after training for different strategies of transfer learning; where "old" refers to the induction cooker condition, "new" refers to the television condition, and "other" refers to conditions other than induction cookers and televisions.

[0046] Figure 6 To apply the particle swarm optimization algorithm to solve for the simplest feature set under different feature number requirements; where: (a) the optimization solution for three feature numbers, and (b) the optimization solution for four feature numbers.

[0047] Figure 7 The differences in fault arc detection accuracy (mean accuracy and standard deviation) after updating training using domain adaptive transfer learning (maximum classifier difference domain adaptive strategy) under different working condition sequences and different numbers of working conditions are shown. Among them: (a) the results of three transfer learnings in different sequences, (b) the results of four transfer learnings in different sequences, and (c) the results after changing the working condition sequence 1-8-12-15. Detailed Implementation

[0048] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The embodiments described are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0049] This invention proposes an adaptive detection method for variable DC fault arcs based on similarity measurement and transfer learning. The method collects current signals from a low-voltage DC distribution network system using a current sensor, extracts features from the collected current signals, inputs the extracted features into a trained state identification model, and performs probability distribution judgment on the output of the state identification model. The extracted features and the trained features are then compared using a similarity measurement. A support vector machine is used to determine whether the current operating condition is identifiable. The similarity measurement value and the model's output probability distribution ensure accurate and reliable judgment of whether the operating condition is identifiable. For unidentifiable operating conditions, a threshold is determined after pairwise cross-measurement of the features, and the simplest feature set is determined according to the requirement of full coverage of unidentifiable operating condition features. By utilizing domain-adaptive transfer learning combined with the simplest feature set to update and train the state identification model, the identified unidentifiable operating conditions are expanded into identifiable operating conditions through the model update process (using labeled source domain data and some unlabeled target domain data to train a one-dimensional convolutional neural network model, which can predict the label information corresponding to the unlabeled target domain data after training). This effectively reduces the difference between identifiable and unidentifiable operating condition data, enabling the state identification model to effectively identify the detected unidentifiable operating conditions after updating and training. This achieves the adaptability of the detection process, allowing the state identification model to ensure the normal and stable operation of systems with different DC application scenarios or variable topologies without human intervention in parameter tuning.

[0050] Combination Figure 1 The steps of the adaptive detection method for variable DC fault arcs based on similarity measurement and transfer learning are explained in detail.

[0051] Step one, parameter initialization, includes setting the sampling frequency fs of the current sensor for the current signal output from the low-voltage DC distribution network system to 1MHz, and setting parameters for similarity measurement, probability distribution judgment, support vector machine, and transfer learning. The current signal x is obtained by sampling the current signal output from the low-voltage DC distribution network system point by point using the current sensor. q When the obtained current signal x q Once the number of sampling points q (q∈N+) reaches the requirement of the analysis period (e.g., q=8000), proceed to step two for current signal characteristic analysis;

[0052] Step 2: Collect the current signal x qWavelet packet decomposition is performed based on the Rbio3.1 mother wavelet to obtain the wavelet packet coefficients of each node. That is, the transformed time-frequency matrix is ​​obtained through wavelet packet decomposition based on the Rbio3.1 mother wavelet. Then, wavelet packet coefficients in the 31250Hz to 62500Hz frequency band (since the accuracy of fault arc detection based on Rbio3.1 features decreases in the 31250Hz to 62500Hz frequency band, this band is considered a significant frequency band for fault arc information) are reconstructed to obtain reconstructed wavelet packet coefficients with the same length as the original data (the number of sampling points is q). After feature extraction of the reconstructed wavelet packet coefficients, the process proceeds to steps three and four. Specifically, feature extraction uses a many-to-one mapping method to construct one-dimensional features.

[0053]

[0054] Where, d j,k For the wavelet packet coefficients corresponding to the wavelet packet decomposition tree nodes, Φ j,k Let be the wavelet packet basis function, j be the scale index, k be the position index, and t be the sampling time.

[0055] Step 3 identifies the system's operating status using a one-dimensional convolutional neural network model. This involves inputting the one-dimensional features extracted in Step 2 into the trained one-dimensional convolutional neural network. If the model output is 0 during the current analysis period, the identification result is no arc (determined as normal), and the process returns to Step 1 to collect current signals for the next analysis period. If the model output is 1 during the current analysis period, the identification result is an arc. Furthermore, the system distinguishes between fault arcs and arc-like phenomena by using multi-period fault arc judgment. Specifically, starting from the current analysis period, if at least 8 out of 10 analysis periods are at a high level (i.e., the model output is 1), it is determined to be a fault arc, and a fault arc cutoff signal is output to stop the system. Otherwise, it is determined to be an arc-like phenomenon.

[0056] In step three above, the one-dimensional convolutional neural network is initially trained with the following parameters to form a one-dimensional convolutional neural network model: 20 iterations, 100 batch size, 0.001 learning rate, 0.9 momentum decay factor beta1, and 0.999 infinity norm decay factor beta2. Simultaneously, the first convolutional layer is set to have 1-dimensional input channels, 32-dimensional output channels, a kernel size of 20*1, and a stride of 1. The second convolutional layer has 32-dimensional input channels, 48-dimensional output channels, a kernel size of 20*1, and a stride of 1. Each convolutional layer is followed by a batch normalization layer (Bn layer). The final output data of the convolutional layers is 49*985*1. A pooling layer is then added to reduce the dimensionality of the data. Finally, a fully connected layer is added to transform the dimensionality-reduced data output into a one-dimensional vector, i.e., the output judgment result is 0 / 1. The training set consists of the current signal features of the induction cooker's operating conditions (i.e., training features).

[0057] Step 4: Feature Similarity Measurement and Network Output Probability Distribution

[0058] 4.1 Network Output Probability Distribution

[0059] The probability distribution of the state identification model output is determined by a probability distribution strategy based on multi-objective particle swarm optimization. Specifically, it is determined whether the proportion of high-level signals within m (m=25) analysis periods starting from the analysis period where the output result is 1 meets the probability distribution requirements. If it does, the proportion is denoted as probability P.

[0060]

[0061] Where g is the analysis period in which the model output result is 1 for the first time under a certain working condition, m is the number of analysis periods considered, i is the sequence number of the analysis period, and y is the sequence number of the analysis period. i This is the model output result for the i-th analysis period.

[0062] If not, then iteratively determine whether the proportion of high-level signals in the corresponding analysis time periods divided into m analysis time periods within the N (N=400) analysis time periods meets the probability distribution requirements. If there is an analysis time period range that meets the probability distribution requirements (i.e., a certain m analysis time periods), the corresponding proportion of high-level signals is recorded as probability P. If the proportion of high-level signals in the last m analysis time periods divided within the N analysis time periods still does not meet the probability distribution requirements, take the average of the proportions of high-level signals in each of the m analysis time periods divided within the N analysis time periods, record it as probability P, and go to step five.

[0063] In the above probability distribution strategy, multi-objective particle swarm optimization means that if the model output result corresponding to a certain analysis period is 1, then multi-objective particle swarm optimization is first used to update the upper probability limit P2 and lower probability limit P1 of the probability distribution judgment, and then it is judged whether the probability distribution requirements are met.

[0064] 4.2 Feature Similarity Measurement

[0065] Input the current signal characteristics into the sample data matrix X:

[0066]

[0067] Among them, the current signal characteristics are one-dimensional data with only one variable, q, which is the number of data points;

[0068] The similarity metric input consists of training features and features extracted over N (N=400) analysis periods. The two are then subjected to a maximum mean difference (MMD) similarity metric based on particle swarm optimization to obtain a metric value D.

[0069]

[0070] Where F1 represents the training features, f i Let F1 be each sample, M be the number of training feature samples, F2 be the features extracted in each analysis period, and r be the number of samples. j For each sample in F2; the maximum mean difference similarity measure kernel function k is a Gaussian kernel function, and the Gaussian kernel function parameter σ is 6761.11.

[0071] Step 5: Input the above metric D and probability P into the SVM model. If the SVM model output is 0, mark the current signal features extracted in N analysis periods as identifiable operating condition data; if the output is 1, mark the current signal features extracted in N analysis periods as unidentifiable operating condition data. Then determine whether the state identification model update cycle has been reached. If the update cycle has been reached, proceed to step 6; otherwise, proceed to step 1.

[0072] The SVM model described above was trained using the following method: 50% of the samples calculated under various working conditions (D and probability P) were randomly selected as the training set, and the remaining 50% were used as the test set. A Support Vector Machine (SVM) classification hyperplane was fitted to the training set, and performance was validated on both the training and test sets. The accuracy rate for working condition type identification was 100%, indicating that the SVM was successfully trained and could be used for classification in working condition identification problems.

[0073] Step 6: The 15 batches of unidentifiable working condition data identified by the Support Vector Machine are each classified as a category, and pairwise cross-metrics are performed. The results are stored in the distance metric matrix Dist. A threshold is applied to Dist (threshold set to 0.01). If the threshold is exceeded, it is marked as 1; otherwise, it is marked as 0. The marking results are then stored in the matrix Dist_new. ij In (where i and j refer to the row and column of the matrix, respectively), the simplest feature set is determined according to the requirement of full coverage of unidentifiable working condition data (i.e., according to the principle of covering as many unidentifiable working condition features as possible with the fewest feature combinations). This determination process can be expressed as solving the following optimization problem:

[0074]

[0075] Here, ∨ is the OR operator.

[0076] The above optimization problem can be solved using the Particle Swarm Optimization (PSO) algorithm.

[0077] The state identification model (the aforementioned one-dimensional convolutional neural network model) is updated and trained using a maximum classifier neighborhood adaptation strategy, thereby upgrading the state identification model. The update and training process is as follows: First, the feature generator and classifier are pre-trained using source domain samples to ensure correct classification, and the network is trained to minimize the Softmax cross-entropy loss. Second, the feature generator is fixed, and the cross-entropy loss of the source domain samples is added to the loss function. Third, the two classifiers are fixed, and the features are optimized to ensure that the features affect the two classifiers as consistently as possible.

[0078] See Figure 2 The detection process for DC fault arcs under known operating conditions (system component is an induction cooker) is divided into three stages: Stage I (no fault arc occurs, no fault arc interruption action), Stage II (fault arc occurs), and Stage III (fault arc interruption). Figure 2 As can be seen, during the arc fault stage, abnormal phenomena such as reduced current amplitude and high-frequency harmonic components occur. Furthermore, wavelet packet decomposition effectively extracts features, clearly distinguishing between the normal and arc-generating states, and can effectively interrupt the arc fault. This proves that using wavelet packet decomposition and coefficient reconstruction to extract features from the output current signal of a practical low-voltage DC distribution network system is effective. For example, features formed by a 10-second current flow can be used as training feature samples (M=1250).

[0079] See Figure 3The accuracy results of fault arc detection using a one-dimensional convolutional neural network trained on a known operating condition (system component: induction cooker) under various different operating conditions show that the trained one-dimensional convolutional neural network model can effectively identify fault arcs under known operating conditions, but it cannot effectively identify fault arcs under new system components, such as multi-functional rice cookers and standing fans. Its accuracy fluctuates within a certain range and is less than 1. Although there are some new operating conditions that can be effectively identified (presumably because these conditions have highly similar features to known conditions, allowing for approximate equivalent processing), it still indicates that the detection method needs to have the function of judging whether the detected operating condition is new (especially unidentifiable conditions).

[0080] See Figure 4 By comparing the similarity values ​​D obtained from measuring the similarity between identifiable and unidentifiable operating conditions and the induction cooker operating conditions, it can be seen that relying solely on the distance value of the similarity measurement to determine the accuracy of unidentifiable operating conditions is insufficient. The comparison revealed some instances where the measurement value D could not be fitted or was incorrectly fitted. Therefore, further judgment steps are needed to improve the accuracy of judging different operating conditions.

[0081] See Figure 5By comparing the detection accuracy of the state identification model before and after training with different strategies of transfer learning, it can be seen that the state identification model without transfer learning training has a detection accuracy of 100% for the old working condition (induction cooker), while the detection accuracy for the new working condition ([4] TV) and other working conditions ([1] multi-functional rice cooker; [2] 48V full load; [3] humidifier; [5] standing fan; [6] refrigerator; [7] LED; [8] air conditioner cabinet; [9] air conditioner + induction cooker, bus arcing;

[10] air conditioner + induction cooker, air conditioner branch arcing;

[11] air conditioner + induction cooker, induction cooker branch arcing;

[12] multi-split unit;

[13] energy storage + induction cooker;

[14] induction cooker arcing test air conditioner;

[15] air conditioner arcing test induction cooker) is only 50% and 22.22% respectively, that is, the detection of fault arc is relatively limited. The state identification model was updated using different types of transfer learning strategies. Results showed that the state identification models updated using feature transfer (e.g., Transfer Component Analysis, TCA) and model transfer (e.g., fine-tuning; unfrozen parameter strategy) achieved detection accuracies of 40.48%, 71.43%, and 61.90% respectively on the old operating conditions. This indicates that the updated state identification model forgot the old operating conditions. However, the state identification model updated using the Maximum Classifier Difference Domain Adaptive (MCD_DA) method achieved 100% accuracy. Regarding the detection performance on new operating conditions, the state identification models updated using feature transfer and sample transfer (e.g., Tradaboost strategy) achieved detection accuracies of 25.00% and 75.00% respectively, indicating that the model was not effectively updated. The state identification model updated using MCD_DA still achieved 100% accuracy. Furthermore, the state identification model updated using MCD_DA also showed a certain accuracy in detecting fault arcs (50%) on other operating conditions, outperforming the other four strategies. Therefore, when using transfer learning to iteratively optimize new working conditions, the maximum classifier difference domain adaptive strategy should be selected.

[0082] See Figure 6 a, Figure 6b. In the previous transfer process targeting the target operating condition, the detection performance of other operating conditions improved. This is because there are cases where the feature distributions are relatively similar in the unidentifiable operating condition type data. Based on this, the representation idea of ​​the simplest feature group is proposed. The pairwise cross-similarity measure (MMD) between all unidentifiable operating condition type data is calculated, and combined with threshold judgment and OR operation, the optimization problem in step six is ​​solved using PSO. The smaller the optimization fitness, the more unidentifiable operating condition type data can be represented by the simplest feature group, and vice versa. It can effectively reduce the model update training data and reduce the update training time. Whether it is the optimization of three simplest feature combinations or four simplest feature combinations, different transfer learning times (three times, four times) ultimately yield five feasible solutions. However, the optimization of four simplest feature combinations is better than that of three simplest feature combinations, achieving the highest accuracy in identifying fault arcs in new and old operating conditions.

[0083] See Figure 7 a, Figure 7 b. By examining the differences in fault arc detection accuracy caused by the variations in the types and number of different working conditions (specifically, the number of different unidentifiable working condition types in the simplest feature group) in the training state identification model updated using the maximum classifier difference domain adaptive strategy transfer learning, it can be seen that different types of working conditions used in multiple transfer learning sessions will result in differences in adaptive detection performance. Figure 7 In 3-14-15 shown in a, the three unidentifiable working condition features (3 rows) corresponding to [3],

[14] , and

[15] can be ORed to obtain fourteen 1s, which means that the three features cover the features of fourteen unidentifiable working conditions. Figure 7 In the 1-13-14-15 shown in b, the four unidentifiable operating condition features (4 rows) corresponding to [1], [3],

[14] , and

[15] can be ORed to obtain fifteen 1s, which means that the four features cover the features of fifteen unidentifiable operating conditions. In terms of the final fault arc detection accuracy, the detection accuracy of three transfer learnings is as high as 91.46%, and the accuracy of four transfer learnings is as high as 96.34%. That is, the detection accuracy of the model after four transfer learnings is higher than that of the model after three transfer learnings, but the time and computational load required for four transfer learnings will be more and greater. Therefore, in practical applications, the number of transfer learnings can be dynamically adjusted when the accuracy requirements are different.

[0084] Due to comprehensive Figure 7 a, Figure 7The results of b indicate that when the transfer learning working condition order (specifically, the order in which different unidentifiable working condition types of data in the simplest feature group are used) is 1-8-12-15, the detection accuracy of the finally updated state identification model is the highest. Therefore, the order in which the features corresponding to the working conditions [1], [8],

[12] , and

[15] are used in transfer learning is changed, and the detection accuracy of the corresponding model obtained by updating the training is analyzed. The set usage order is roughly divided into the following different situations: A, B, C, and D. A is the case where [1] and [8] appear before

[12] at the same time. B and C are the influence of the order of [1] and [8] in A (B is [1] first, C is [8] first). D is the case where [1] or [8] appears after

[12] . The results are as follows: Figure 7 As shown in c, it can be seen that when [1] and [8] are before

[12] , the average accuracy is higher than in other cases, and the relationship between [1] and [8] has little impact on the accuracy when [1] and [8] are before

[12] . These results indicate that the order in which features are used for different working conditions in transfer learning can also lead to differences in the effectiveness of fault arc detection.

[0085] In summary, the adaptive detection method for DC fault arcs based on similarity measurement and transfer learning proposed in this invention can accurately, reliably, and rapidly identify the output current state (arc-occurring and arc-free) of a complex and diverse DC system, especially a low-voltage DC distribution network system, by mining the essential characteristics of fault arcs occurring during system operation. This effectively distinguishes between normal, arc-like, and fault arc operating states, and disconnects the system when a fault arc occurs, thus achieving the real-time and accurate performance requirements of a DC fault arc detection device. Furthermore, this invention determines the transfer learning time based on the accurate identification of unidentifiable operating conditions, utilizes the simplest feature set of unidentifiable operating condition type data to comprehensively represent the knowledge to be transferred, and finally applies domain-adaptive methods to improve the generalization ability of the state identification model, thereby achieving adaptive detection of DC fault arcs under complex and variable operating conditions.

Claims

1. An adaptive detection method for variable DC fault arcs based on similarity measurement and transfer learning, characterized in that: The adaptive fault arc detection method includes the following steps: 1) Perform time-frequency transformation on the acquired system output current, and then extract its features; 2) Input the features extracted in step 1) into the state identification model based on deep learning, thereby determining the system operating state by distinguishing between DC fault arcs and arc-like structures; evaluate the joint distribution pattern between the feature similarity metric and the model output probability distribution for the current operating state through machine learning, thereby classifying the features extracted in step 1) as either unidentifiable operating state type data or identifiable operating state type data, and select unidentifiable operating state type data to update the state identification model based on deep learning through transfer learning, wherein the feature similarity metric is the distance value between the features extracted in step 1) and the training features in the state identification model based on deep learning.

2. The adaptive detection method for variable DC fault arcs based on similarity measurement and transfer learning according to claim 1, characterized in that: Step 1) specifically includes the following steps: 1.1) The output current of the DC system is sampled according to the sampling frequency. fs By sampling point by point, a current signal for one analysis period is obtained. x q The number of sampling points during this analysis period is q The q The range is 8000~12000. fs 200kHz~1MHz; 1.2) The acquired current signal x q After time-frequency transformation, a frequency band reflecting fault arc information is selected, and features are constructed according to a many-to-one mapping of fault arc information points in that frequency band during the current analysis period; the fault arc information points are related to the current signal. x q Reconstructed wavelet packet coefficients of consistent length; these wavelet packet coefficients are obtained by using the Rbio3.1 mother wavelet to reconstruct the current signal. x q The wavelet packet coefficients in the frequency band from 31250Hz to 62500Hz in the time-frequency matrix obtained by wavelet packet decomposition are reconstructed.

3. The adaptive detection method for variable DC fault arcs based on similarity measurement and transfer learning according to claim 1, characterized in that: The state identification model uses a one-dimensional convolutional neural network, which sequentially passes the input features through two convolutional layers, a batch normalization layer, a pooling layer, and a fully connected layer, and outputs the identification results, thereby judging the arc occurrence of the DC system one by one according to the analysis period. The initial training parameters of the one-dimensional convolutional neural network include: 15 to 30 iterations, 75 to 125 batch size, 0.0005 to 0.0015 learning rate, 0.7 to 1.0 momentum decay factor beta1, and 0.8 to 1.0 infinity norm decay factor beta2.

4. The adaptive detection method for variable DC fault arcs based on similarity measurement and transfer learning according to claim 1, characterized in that: In step 2), the distinction between DC fault arcs and arc-like phenomena specifically includes the following steps: After inputting features into the state identification model, if the model outputs an identification result indicating that the system is in an arc-generating state during the current analysis period, then multi-period fault arc judgment is performed starting from the current analysis period. If the state identification model outputs an identification result indicating that the system is in an arc-generating state more frequently than a set frequency based on subsequently input features, then the system operating condition is determined to be a fault arc; otherwise, the system operating condition is determined to be an arc-like phenomenon. The parameters for multi-period fault arc judgment include: the total number of analysis periods is 8 to 15, and the proportion of arc-generating state identification times required to determine a fault arc is ≥80%.

5. The adaptive detection method for variable DC fault arcs based on similarity measurement and transfer learning according to claim 1, characterized in that: The model output probability distribution is obtained through probability distribution judgment, which specifically includes the following steps: starting from the analysis period in which the system is identified as being in an arc-generating state, determine the percentage of analysis periods in the m consecutive analysis periods starting from that analysis period that are identified as being in an arc-generating state. If this percentage satisfies the probability distribution requirements, then this percentage is recorded as the probability distribution. P And output the probability P As the probability distribution result of the model output over N consecutive analysis periods starting from the stated analysis period, if the proportion of this quantity does not meet the probability distribution requirements, then the proportion of the remaining m consecutive analysis periods within the N analysis periods that identify the system as being in an arc-generating state is determined sequentially. If any of the m consecutive analysis periods meets the probability distribution requirements, then the corresponding proportion of that m consecutive analysis periods is recorded as the probability distribution. P And output the probability P As the model output probability distribution results over N analysis periods, if the percentage of analysis periods in the last m consecutive analysis periods of the N analysis periods still does not meet the probability distribution requirements, then the average percentage of the corresponding percentages for all m consecutive analysis periods within the N analysis periods is taken, and this average is denoted as the probability distribution. P And output the probability P As the model output probability distribution results over N analysis periods; Where N≤400 and m is 20~40, in the process of statistically analyzing the proportion of analysis periods in which the system is identified as being in an arc-generating state in any consecutive m analysis periods, if the state identification model outputs an identification result representing that the system is in an arc-generating state in the current analysis period based on the features extracted from a certain analysis period, then multi-objective particle swarm optimization is used to update the parameters for probability distribution judgment. The parameters include the upper limit P2 and the lower limit P1 required by the probability distribution, with the initial value of P1 being 10%~15% and the initial value of P2 being 50%~55%.

6. The adaptive detection method for variable DC fault arcs based on similarity measurement and transfer learning according to claim 5, characterized in that: The similarity metric is calculated based on the maximum mean difference of particle swarm optimization, and the similarity metric value between the features extracted over N analysis periods and the training features in the state identification model is denoted as the metric value D. The maximum mean difference is calculated using a Gaussian kernel function with a parameter σ of 6000~7000.

7. The adaptive detection method for variable DC fault arcs based on similarity measurement and transfer learning according to claim 6, characterized in that: In step 2), the distinction between unidentifiable operating condition type data and identifiable operating condition type data specifically includes the following steps: [The text then describes a process involving] measuring values... D With probability P The data is input into a support vector machine (SVM), which then outputs the determination result of whether the data is of a identifiable or unidentifiable operating condition type.

8. The adaptive detection method for variable DC fault arcs based on similarity measurement and transfer learning according to claim 1, characterized in that: The state identification model is updated after identifying unidentifiable working condition type data. Specifically, the update cycle is when the number of unidentifiable working condition type data reaches 10 to 20, or the state identification model has a fixed time length as the update cycle.

9. The adaptive detection method for variable DC fault arcs based on similarity measurement and transfer learning according to claim 8, characterized in that: The transfer learning is based on a domain adaptation strategy and specifically includes the following steps: performing pairwise cross-measures on the unidentifiable working condition type data determined within the update cycle, and storing the cross-measure results in a distance metric matrix. Dist After the threshold is determined, the result is stored in a square matrix. Dist _ new In accordance with the requirement of covering all unidentifiable working condition types with the fewest possible types of data, a minimal feature set is determined. The state identification model is then trained using a domain adaptation method combined with the minimal feature set. Training the state identification model with the minimal feature set refers to training the model using different types of unidentifiable working condition types within the minimal feature set in a certain order. Each training session includes the following steps: Based on the principle of feature fuzzy samples at the maximum effect classification decision boundary, the domain adaptation method with the maximum classifier difference is selected. That is, by maximizing the difference between two different classifiers containing source domain conditional distribution information, target domain sample features rejected by the source domain are found. Then, the feature extractor is trained with the goal of minimizing the classifier difference, and features that conform to the source domain conditional distribution are regenerated, thereby adapting to the conditional distribution.

10. An adaptive detection device for variable DC fault arc based on similarity measurement and transfer learning, comprising a signal analysis module, a multi-time fault arc judgment module, a similarity measurement module, a working condition judgment module, and a state identification model update module; The signal analysis module is used to extract features after performing time-frequency transformation processing on the system output current collected according to the analysis period. The multi-period fault arc judgment module is used to input the features extracted by the signal analysis module into the state identification model based on deep learning, thereby determining the system operating state by distinguishing between DC fault arcs and arc-like objects, and to make probability distribution judgments on the output results of the state identification model. The similarity measurement module is used to calculate the distance value of the similarity measurement between the features extracted by the signal analysis module and the training features in the deep learning-based state identification model; The working condition judgment module is used to evaluate the joint distribution pattern between the feature similarity metric value and the model output probability distribution of the current working condition through machine learning, thereby identifying the features extracted by the signal analysis module as either unidentifiable working condition type data or identifiable working condition type data. The state identification model update module is used to select unidentifiable working condition type data and update the deep learning-based state identification model through transfer learning.

Images