An arc fault detection model construction method based on expert experience constraint and performance iterative promotion

Abstract

The application discloses a kind of arc fault detection model construction methods based on expert experience constraint and performance iterative promotion, belong to low-voltage distribution network fault protection field, solve the problem of insufficient generalization ability of existing pure data-driven model.The method first collects current signal to extract time domain, frequency domain characteristics, then based on the expert experience constraint feature threshold interval of electrical physical mechanism, construct initial white box detection model;Subsequently, using the two-stage iterative optimization mechanism of reducing false alarm first, then reducing false negative, combined with the hierarchical progressive correction sub-process of direct ancestor node to adjust the model, while limiting the upper limit of node splitting times to prevent overfitting.The model constructed by the application has high detection accuracy, low false alarm rate and strong interpretability, and there is no error in the generalization test of unknown load, which significantly improves the reliability and practicality of series arc fault detection technology in complex scenarios.

Description

Technical Field

[0001] This invention relates to the field of fault protection technology for low-voltage distribution networks, and in particular to a method for constructing an arc fault detection model based on expert experience constraints and performance iteration improvement. Background Technology

[0002] Arc faults in low-voltage power distribution systems, especially series arc faults, are one of the main causes of electrical fires. Their detection is extremely difficult, primarily because the fault point is connected in series with the load, and traditional protection features such as significant overcurrent or residual current are lacking when the fault occurs. Furthermore, the current waveforms before and after the fault are subject to severe aliasing due to the nonlinear effects of various load types, making it difficult to reliably distinguish between faulty and non-faulty states using only a single or a few features.

[0003] To address this challenge, existing arc fault detection methods are mainly divided into two categories based on their model construction methods. One category is multi-feature fusion methods based on manually set thresholds. This method relies on expert subjective experience to extract features and construct criteria through logical combinations. While it has some interpretability, it struggles to uncover deep feature relationships and has limited generalization ability. The other category is multi-feature fusion methods based on machine learning. This type of method utilizes data-driven approaches, automatically mining or constructing features and their relationships through algorithms, and has become the mainstream research approach. Based on model interpretability, machine learning models can be further divided into black-box models and white-box models. While black-box models may achieve high training accuracy, their decision-making process lacks physical meaning support, their internal logic is invisible, and adjustments and optimizations are difficult. White-box models, on the other hand, are characterized by transparent structure and clear decision-making paths. Their judgment rules have clear physical meaning support, thus possessing high interpretability. While black-box models may achieve high training accuracy, their decision-making process lacks physical meaning and is difficult to adjust internally. On the other hand, existing white-box model applications, such as using principal component analysis combined with decision trees or integrating multi-period features with counter strategies, improve interpretability, but their model construction is still essentially in the realm of pure data-driven approaches.

[0004] The existing methods mentioned above, whether relying on manual experience to set thresholds or on machine learning for data-driven approaches, all share a common limitation: the model construction process is overly dependent on labeled sample data. This makes feature selection and threshold setting entirely dominated by data distribution, potentially deviating from expert experience based on electrophysical mechanisms. Such models, purely based on historical data fitting, often exhibit insufficient generalization ability and decreased adaptability when faced with unknown load types or complex and variable working scenarios not covered by the training data, hindering the reliable application and promotion of arc fault detection technology in engineering practice. Therefore, how to construct a detection method that possesses both data-driven learning capabilities and effectively integrates expert prior knowledge such as electrophysical mechanisms, thereby significantly improving the model's adaptability to unknown scenarios, has become a key technical problem urgently needing to be solved in this field. Summary of the Invention

[0005] To address the shortcomings of existing data-driven arc fault detection models in terms of insufficient generalization ability and poor adaptability to unknown loads, this invention provides an arc fault detection method that integrates expert experience constraints with performance iteration improvement. By constructing a detection architecture based on white-box models such as decision trees, the aim is to build an arc fault detection model that combines high accuracy, low false alarm rate, strong interpretability, and good generalization ability.

[0006] To achieve the above objectives, the technical solution of this invention is as follows: A method for constructing an arc fault detection model based on expert experience constraints and performance iterative improvement is proposed, comprising: building an arc fault experimental platform, collecting fault and non-fault current signals, extracting time-domain and frequency-domain features to construct a feature set, and dividing the set into a training set and a test set; based on the training set, defining an effective threshold range for each feature dimension according to expert experience; generating a candidate splitting threshold set based on the effective threshold range, selecting the optimal splitting point from the candidate splitting threshold set, and recursively constructing an initial white-box detection model; performing a two-stage iterative optimization on the initial white-box detection model, and outputting the detection model.

[0007] Optionally, generating the candidate split threshold set includes: for each feature dimension, sorting the values ​​of all samples in the training set on the feature in ascending order to obtain an ordered sequence; generating a full candidate threshold set on the feature based on the midpoint of the values ​​of adjacent samples in the ordered sequence; and mapping the full candidate threshold set to an effective threshold interval defined by expert experience to form a candidate split threshold set, such that each candidate value in the set falls within the effective threshold interval.

[0008] Optionally, selecting the optimal split point from the candidate split threshold set includes: traversing the candidate split threshold set corresponding to each feature dimension for each node to be split in the initial white-box detection model; calculating the Gini coefficient of the child nodes obtained after splitting with the threshold for each candidate split threshold; and selecting the corresponding feature and threshold as the optimal split rule for the current node based on minimizing the Gini coefficient.

[0009] Optionally, the initial white-box detection model is subjected to a two-stage iterative optimization, including: calculating the false positive rate of the current model on the training set and determining whether it meets a preset hard constraint threshold for the false positive rate; if the false positive rate does not meet the hard constraint threshold, then the false positive rate reduction stage is entered; if the requirements are still not met after adjustment, the number of model node splits is increased, the initial model is reconstructed and iterated until the false positive rate meets the constraint; if the false positive rate meets the hard constraint threshold, then the false negative rate reduction stage is entered; in the false negative rate reduction stage, if the model adjustment causes the false positive rate to exceed the hard constraint threshold or the accuracy cannot be further optimized, then the iteration is stopped.

[0010] Optionally, if the false alarm rate does not meet the hard constraint threshold, the process enters the false alarm rate reduction stage: Locate the leaf node with the highest false alarm risk in the model, and call the hierarchical progressive correction process of the direct ancestor node to adjust the model structure, then recalculate the false alarm rate; if the adjustment still fails to meet the requirements, increase the number of model node splits, rebuild the initial model, and iterate until the false alarm rate meets the constraints, including: for all leaf nodes in the decision tree model whose prediction results are faulty; for each leaf node, calculate the leaf node with the highest false alarm risk index as the worst leaf node to be optimized; call the hierarchical progressive correction subprocess of the direct ancestor node to correct the decision path where the worst leaf node is located; re-evaluate the false alarm rate of the corrected model; if it meets the hard constraint threshold, the process moves to the false alarm rate reduction stage; otherwise, check whether the correction attempt for all high-risk leaf nodes has been completed; if so, and the false alarm rate still does not meet the standard, increase the number of splits and rebuild the initial decision tree to start a new round of iterative optimization.

[0011] Optionally, if the false positive rate meets the hard constraint threshold, the process enters the false negative rate reduction stage: Locating the leaf node with the highest false negative risk in the model, and invoking the hierarchical progressive correction process of the direct ancestor node to adjust the model structure, ensuring that the false positive rate never exceeds the hard constraint threshold during the adjustment process includes: starting from the worst leaf node to be optimized, backtracking along the decision path to its parent node; at the parent node level, generating a candidate splitting threshold set based on the effective threshold range of each feature, and constructing a correction search space; for each candidate splitting scheme in the constructed search space, reconstructing and evaluating its subtree performance in parallel, and selecting the optimal scheme according to the current optimization stage; if the optimal scheme is better than the original scheme, the correction is completed; otherwise, backtracking continues; repeating the step-by-step backtracking until the root node; if there is still no better scheme at the root node, it is determined that the current path cannot be optimized, and this process ends.

[0012] Optionally, if the false alarm rate meets the constraint threshold, the process proceeds to the false alarm rate reduction stage: The stage involves locating the leaf node with the highest false alarm risk in the model and invoking the hierarchical progressive correction process of the direct ancestor node to adjust the model structure. During the adjustment process, it is ensured that the false alarm rate never exceeds the hard constraint threshold. This includes: for all leaf nodes in the decision tree model whose prediction results are normal; for each leaf node, calculating the leaf node with the highest false alarm risk index as the worst leaf node to be optimized; invoking the hierarchical progressive correction subprocess of the direct ancestor node to correct the decision path where the worst leaf node is located; and evaluating the corrected model to obtain the final optimized model.

[0013] Optionally, for each candidate splitting scheme in the search space, the performance of its subtree is reconstructed and evaluated in parallel; the selection of the optimal scheme according to the current optimization stage can be implemented using a parallel computing architecture, simultaneously evaluating all splitting schemes in the corrected search space.

[0014] Optionally, in the two-stage iterative optimization process, an upper limit for the number of node splits is preset to prevent model overfitting, and a false alarm rate hard constraint threshold is set as the core criterion for iterative transformation and termination.

[0015] Optionally, in the hierarchical progressive correction process of the direct ancestor node, after updating the splitting rules of the parent node or the higher-level ancestor node, the subtree to which it belongs needs to be reconstructed; the splitting criterion for subtree reconstruction is still based on minimizing the Gini coefficient, and the splitting threshold of each node after reconstruction must still meet the effective threshold range defined by the expert experience.

[0016] Beneficial effects: The arc fault detection model construction method of this invention, based on expert experience constraints and performance iteration improvement, effectively limits the model search space by introducing expert experience-constrained feature threshold ranges based on electrophysical mechanisms. This avoids overfitting and physical uninterpretability problems caused by purely data-driven approaches, thereby significantly improving the model's generalization ability to unknown loads and complex scenarios. By adopting a two-stage iterative optimization mechanism that first reduces false alarms and then reduces false negatives, the model structure can be gradually optimized while ensuring a low false alarm rate—a core engineering indicator—continuously improving the overall detection accuracy and achieving both detection reliability and accuracy. It effectively balances accuracy and performance; through the hierarchical progressive correction subprocess of direct ancestor nodes, it can achieve fine-grained and interpretable adjustments to the model structure without increasing model complexity, enhancing its adaptability to new loads and unknown operating conditions; by setting an upper limit on the number of node splits and combining it with a hard constraint threshold on the false alarm rate as the core of iterative control, it can effectively prevent model overfitting, ensure stable convergence of the optimization process, and finally construct an arc fault detection model with high accuracy, low false alarm rate, strong interpretability and good generalization performance, significantly improving the reliability and practicality of this technology in complex real-world scenarios of low-voltage distribution networks.

[0017] To make the above-mentioned features and advantages of the invention more apparent and understandable, specific embodiments are described below, and detailed descriptions are provided in conjunction with the accompanying drawings. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for constructing an arc fault detection model.

[0019] Figure 2 This is a flowchart of step S3 in a method for constructing an arc fault detection model.

[0020] Figure 3 This is a flowchart of step S32 in a method for constructing an arc fault detection model.

[0021] Figure 4 This is a flowchart of step S324 in a method for constructing an arc fault detection model.

[0022] Figure 5 This is a flowchart of step S33 in a method for constructing an arc fault detection model.

[0023] Figure 6 This is a schematic diagram of the initial decision tree model topology.

[0024] Figure 7 A schematic diagram of the final optimized decision tree model topology.

[0025] Figure 8 This is a comparison chart of the confusion matrices for model performance. Among them, Figure 8Figure (a) in the figure is a quantitative evidence of the model performance based on the confusion matrix of the training set. Figure 8 Figure (b) in the figure is a quantitative evidence of the model performance of the confusion matrix of the test set. Detailed Implementation

[0026] To make the objectives and technical solutions of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0027] In one embodiment, see Figure 1 , Figure 1 This is a flowchart of an arc fault detection model construction method based on expert experience constraints and performance iteration improvement provided in one embodiment of the present invention. The present application provides an arc fault detection model construction method, including the following steps: steps S1 to S5.

[0028] Step S1: Build an arc fault experimental platform, collect fault and non-fault current signals, extract time-domain and frequency-domain features to construct a feature set, and divide the training set and test set.

[0029] Step S2: Based on the training set, define an effective threshold range for each feature dimension according to expert experience; based on the effective threshold range, generate a candidate splitting threshold set, select the optimal splitting point from the candidate splitting threshold set, and recursively construct the initial white-box detection model.

[0030] Step S3: Perform two-stage iterative optimization on the initial white-box detection model and output the detection model.

[0031] This application presents a method for constructing an arc fault detection model based on expert experience constraints and iterative performance improvement. By introducing expert experience to impose physical constraints on the effective threshold range of feature dimensions, it effectively limits the model search space and avoids overfitting and physical uninterpretability caused by purely data-driven approaches. Through a two-stage iterative optimization mechanism that first reduces false positives and then false negatives, it continuously improves the overall accuracy of the model while ensuring a low false positive rate, a core engineering indicator. Furthermore, through a hierarchical progressive correction sub-process at the direct ancestor node level, it enables refined and interpretable adjustments to the model structure, enhancing its adaptability to new loads and unknown scenarios. Ultimately, the constructed model possesses high detection accuracy, low false positive rate, strong interpretability, and good generalization performance, significantly improving the reliability and practicality of arc fault detection technology in complex real-world scenarios.

[0032] In step S1, please refer to Figure 1 An arc fault experimental platform was built to collect fault and non-fault current signals, extract time-domain and frequency-domain features to construct a feature set, and divide it into training set and test set.

[0033] Specifically, an arc fault experimental platform was built, and various load types with representative electrical characteristics were selected to collect normal current signals and fault current signals under each load type.

[0034] As an example, the training load type can be selected from household appliances with representative electrical characteristics, including induction cookers, vacuum cleaners, dimmers, refrigerators, electric kettles, and air conditioners.

[0035] Furthermore, current signals can be acquired in half-wave units. Based on the acquired current signal waveform, multiple features in the time and frequency domains can be extracted. Based on the extracted features, a multi-dimensional feature set of arc faults can be constructed. The feature set can be divided into a training set and a test set according to a preset ratio for subsequent model construction and verification.

[0036] As an example, multiple features in the time and frequency domains may include characteristic energy kurtosis, high-frequency spectrum amplitude sum, characteristic energy sum, odd-even harmonic ratio, characteristic energy entropy, and cosine similarity.

[0037] In step S2, please refer to Figure 1 Based on the training set, for each feature dimension, an effective threshold range is defined according to expert experience; based on the effective threshold range, a candidate splitting threshold set is generated, and the optimal splitting point is selected from the candidate splitting threshold set to recursively construct the initial white-box detection model.

[0038] As an example, the white-box detection models include decision trees, linear regression, logistic regression, etc. This embodiment will use decision trees as an example for detailed explanation.

[0039] Specifically, the specific expression for the training set is:

[0040]

[0041] in, for 3D feature vector R , For category labels, i This refers to the index number of the sample in the training set.

[0042] Furthermore, for any feature dimension The effective threshold range defined by expert experience is: ,in, L This is the lower limit of the effective threshold for the feature. U This represents the upper limit of the effective threshold for the feature.

[0043] Furthermore, an initial decision tree is constructed, targeting the features. j , training set D The values ​​of all samples on this feature are sorted in ascending order to obtain an ordered sequence. Among them, ordered sequence The expression is as follows:

[0044]

[0045] in, For the training set n The sample at the th j The specific values ​​that can be taken on the feature dimension.

[0046] Furthermore, based on the midpoint of adjacent sample values, a complete set of unconstrained candidate thresholds for this feature is generated. S， Among them, the full set of candidate thresholds S The expression is as follows:

[0047]

[0048] in, Represents a set S The first in n One candidate threshold, Representation of features j In the ordered sequence of the first n The feature values ​​of each sample Representation of features j In the ordered sequence of the first n+1 The feature values ​​of each sample.

[0049] Furthermore, the full candidate set S Mapped to expert-defined effective threshold range This forms a set of effective candidate thresholds that are physically interpretable. This ensures that the search space for the model's splitting threshold conforms to physical laws. The effective candidate threshold set... The expression is as follows:

[0050]

[0051] in, S Represents the full candidate set. L This represents the lower limit of the effective threshold for a feature as defined by expert experience. U This represents the upper limit of the effective threshold for a feature as defined by expert experience. T Represents the complete set S A single candidate threshold in the data.

[0052] Furthermore, when constructing each node to be split in the decision tree, the candidate values ​​in the set of effective candidate thresholds corresponding to each feature dimension are traversed to form multiple candidate splitting schemes; for each candidate splitting scheme, the Gini coefficient of the child nodes obtained after splitting according to the scheme is calculated; based on minimizing the Gini coefficient, the corresponding features and thresholds are selected from the multiple candidate splitting schemes and determined as the optimal splitting rule for the current node.

[0053] Furthermore, the above node splitting process is recursively executed until the preset decision tree stopping growth condition is met, thereby completing the construction of the initial decision tree model; wherein, the splitting threshold of each node in the model is derived from the physical experience constraint interval, so that the initial decision tree model has initial generalization ability and interpretability based on physical experience constraints.

[0054] In step S3, please refer to Figure 1 In step S3, the initial white-box detection model is subjected to two-stage iterative optimization to output the detection model.

[0055] For example, please refer to Figure 2 , Figure 2 This is a flowchart of step S3 in an arc fault detection model construction method based on expert experience constraints and performance iteration improvement provided in an embodiment of the present invention. Step S3 may include the following steps: steps S31 to S34.

[0056] Step S31: Calculate the false alarm rate of the current model on the training set and determine whether it meets the preset hard constraint threshold for false alarm rate.

[0057] Step S32: If the false alarm rate does not meet the hard constraint threshold for false alarm rate, then proceed to the false alarm rate reduction stage.

[0058] Step S33: If the false alarm rate meets the hard constraint threshold for false alarm rate, then proceed to the stage of reducing the false alarm rate.

[0059] Step S34: During the false negative rate reduction phase, if the model adjustment causes the false positive rate to exceed the hard constraint threshold of the false positive rate or the accuracy cannot be further optimized, then the iteration stops.

[0060] Specifically, in step S31, the number of node splits can be set to N, and the upper limit of the number of node splits can be limited to N. max =10, you can set a hard constraint threshold for false positive rate a=0.2, calculate the false positive rate FPR of the initial model on the training set, and determine whether it meets the preset hard constraint threshold a for false positive rate.

[0061] Further, in step S32, if the current false alarm rate FPRIf the false alarm rate is subject to a hard constraint threshold of a%, the model is deemed not to meet the standard, and the system enters an iterative process to reduce false alarms.

[0062] For example, please refer to Figure 3 , Figure 3 This is a flowchart of step S32 in an arc fault detection model construction method based on expert experience constraints and performance iteration improvement provided in an embodiment of the present invention. Step S32 may include the following steps: steps S321 to S324.

[0063] Step S321: For all leaf nodes in the decision tree model whose prediction result is faulty.

[0064] Step S322: For each leaf node, calculate the leaf node with the highest false alarm risk index as the worst leaf node to be optimized.

[0065] Step S323: Call the hierarchical progressive correction sub-process of the direct ancestor node to correct the decision path where the worst leaf node is located.

[0066] Step S324: Re-evaluate the false alarm rate of the corrected model. If the false alarm rate meets the hard constraint threshold, proceed to the false alarm rate reduction stage; otherwise, check whether the correction attempts for all high-risk leaf nodes have been completed. If so, and the false alarm rate is still not up to standard, increase the number of splits and rebuild the initial decision tree to start a new round of iterative optimization.

[0067] Specifically, in step S322, the false alarm risk index R FP The expression is as follows:

[0068]

[0069] in, This represents the number of samples in the leaf nodes whose true class label is normal but which were mistakenly identified as faulty. This represents the total number of samples contained in a leaf node.

[0070] Furthermore, in step S324, if the false alarm rate (FPR) still does not meet the target, then the number of splits is increased incrementally in an iterative manner of N=N+1. If N≤N max =10, then rebuild the initial decision tree and start a new round of iterative optimization; if N>N max If the number of splits reaches 10, the incrementing of the number of splits will stop, and the optimal model structure will be determined for the current sample set.

[0071] Specifically, please refer to Figure 4 , Figure 4This is a flowchart of step S324 in an arc fault detection model construction method based on expert experience constraints and performance iteration improvement provided in an embodiment of the present invention. The hierarchical progressive correction of the direct ancestor node may include the following steps: steps S3241 to S3245.

[0072] Step S3241: Starting from the worst leaf node to be optimized, backtrack along the decision path to its parent node.

[0073] Step S3242: At the parent node level, a set of candidate splitting thresholds is generated based on the effective threshold range of each feature, and a modified search space is constructed.

[0074] Step S3243: For each candidate splitting scheme in the constructed search space, reconstruct and evaluate its subtree performance in parallel, and select the optimal scheme according to the current optimization stage.

[0075] Step S3244: If the optimal solution is better than the original solution, complete the correction; otherwise, continue backtracking.

[0076] Step S3245: Repeat the backtracking process until the root node is reached; if there is still no better solution at the root node, it is determined that the current path cannot be optimized and the process ends.

[0077] Specifically, in step S3241, the worst leaf node to be optimized is input, and the process is to backtrack along the decision path to its parent node, starting from the worst leaf node.

[0078] Further, in step S3242, at the parent node level, a set of candidate splitting thresholds is generated based on the effective threshold range of each feature, and a correction search space is constructed; based on the set of effective candidate thresholds, a search space for node correction is constructed, the search space containing all possible splitting schemes that replace the splitting threshold of the parent node with other candidate values ​​in the set of effective candidate thresholds other than the original threshold.

[0079] Furthermore, in step S3243, for each candidate splitting scheme in the modified search space, the corresponding subtree is independently reconstructed, and the performance indicators of the new subtree under each candidate scheme are evaluated synchronously using parallel computing. If the current stage is the false positive reduction stage, the scheme that minimizes the false positive rate (FPR) of the new subtree on the training set is selected. If the current stage is the false negative reduction stage, the scheme that maximizes the accuracy of the new subtree is selected, provided that the false positive rate (FPR) does not exceed the hard constraint threshold 'a' of the false positive rate.

[0080] Further, in step S3244, the selected optimal splitting scheme is compared with the original splitting scheme: if the performance of the optimal splitting scheme is better than the original scheme, the splitting rule of the parent node is updated with the optimal scheme, and the corresponding subtree is reconstructed accordingly to update the model, completing this correction and exiting the process; otherwise, it is determined that effective optimization cannot be achieved at the parent node level, and backtracking continues to the next higher level node, and the same correction operation as the parent node is performed on the grandparent node, including obtaining splitting features and a set of effective candidate thresholds, constructing a correction search space, evaluating each candidate splitting scheme in parallel and selecting the optimal scheme according to the current optimization stage, and comparing and deciding whether to update the splitting rule.

[0081] Furthermore, in step S3245, the process is repeated step by step upwards until the root node is reached; if a splitting scheme that satisfies the optimization objective is still not obtained after backtracking to the root node, it is determined that further optimization cannot be achieved by adjusting the current decision path, and the hierarchical progressive correction process of the direct ancestor node is terminated.

[0082] As an example, during the hierarchical progressive correction of the direct ancestor node, the update of the splitting rules of higher-level nodes such as the grandfather node, great-grandfather node, and even the root node will lead to the reconstruction of some or all of their subordinate subtrees. The splitting criterion for subtree reconstruction is still based on minimizing the Gini coefficient, and the splitting threshold after reconstruction still falls within the effective threshold range defined by expert experience.

[0083] Furthermore, in step S33, if the false alarm rate meets the hard constraint threshold for the false alarm rate, then the false alarm rate reduction stage begins.

[0084] For example, please refer to Figure 5 , Figure 5 This is a flowchart of step S33 in an arc fault detection model construction method based on expert experience constraints and performance iteration improvement provided in an embodiment of the present invention. Step S33 may include the following steps: steps S331 to S333.

[0085] Step S331: For all leaf nodes in the decision tree model whose prediction results are normal.

[0086] Step S332: For each leaf node, calculate the leaf node with the highest false negative risk index as the worst leaf node to be optimized.

[0087] Step S333: Call the hierarchical progressive correction sub-process of the direct ancestor node to correct the decision path where the worst leaf node is located.

[0088] Step S334: Evaluate the corrected model and output the final optimized detection model.

[0089] Specifically, in step S332, the underreporting risk index RFN The parameter expression is as follows:

[0090]

[0091] in, N FN This indicates the number of samples in the leaf node whose true class label is faulty but were mistakenly classified as normal. N total This indicates the total number of samples contained in the leaf node.

[0092] Furthermore, in step S34, during the false negative reduction phase, an evaluation is required after each correction. If a correction causes the false positive rate (FPR) to exceed the hard constraint threshold 'a', or if backtracking correction of the worst-performing node fails to improve the overall model accuracy, then the entire iterative optimization process is stopped.

[0093] In another embodiment, six typical household appliances were selected as test loads: induction cooker, vacuum cleaner, dimmer, refrigerator, electric kettle, and air conditioner. Six features were extracted: characteristic energy kurtosis, high-frequency spectrum amplitude sum, characteristic energy sum, odd-even harmonic ratio, characteristic energy entropy, and cosine similarity. Using half-wave as the unit, 100 sets of normal / fault waveforms were collected for each load to construct a multi-dimensional feature set for arc faults.

[0094] Furthermore, based on expert experience, the splitting threshold range for each feature dimension is constrained as follows: feature energy kurtosis [30, 120], high-frequency spectrum amplitude sum [6, 40], feature energy sum [10, 500], odd-even harmonic ratio [4, 20], feature energy entropy [3.5, 6.5], and cosine similarity [0.5, 0.95].

[0095] As an example, within the effective threshold range corresponding to each feature, the optimal split point of each node is calculated based on the Gini coefficient minimization criterion, thereby constructing an initial decision tree model, the topology of which is as follows: Figure 6 As shown, Figure 6 This is a schematic diagram of the initial decision tree model topology in an arc fault detection model construction method based on expert experience constraints and performance iteration improvement provided in one embodiment of the present invention. Figure 6 It intuitively displays the complete topology of the initial white-box detection model built based on expert experience constraints, clearly presenting the splitting features, splitting thresholds, and node hierarchical relationships of each level of the decision tree. The splitting thresholds of all nodes strictly fall within the effective threshold range of the 6 types of features defined by the experts.

[0096] Furthermore, the constructed initial decision tree model achieved an accuracy of 95.52% and a false positive rate (FPR) of 1.03%. Through iterative improvement, the model underwent two effective update iterations. Each iteration focused on a specific node, calculating and selecting the optimal splitting threshold based on its valid candidate set, thereby progressively optimizing the model's performance. Ultimately, a highly robust decision tree model was obtained, with the following topology: Figure 7 As shown, Figure 7 This is a schematic diagram of the final optimized decision tree model topology in an arc fault detection model construction method based on expert experience constraints and performance iteration improvement provided in one embodiment of the present invention. Figure 7 This demonstrates the converged final decision tree model topology after two effective two-stage iterative optimizations and hierarchical progressive corrections to the direct ancestor nodes. Compared to... Figure 6 The initial model underwent fine-tuning of the splitting characteristics and splitting thresholds of some core nodes, resulting in a topology structure that better aligns with the electrical and physical laws of arc fault detection. The changes in the topology structure in the figure intuitively demonstrate the improvement effect of the two-stage iterative optimization—first reducing false alarms and then reducing false negatives—on the model structure.

[0097] Furthermore, the final model achieved an accuracy of 97.6% and a false positive rate of 0% on the training set; and an accuracy of 97.08% and a false positive rate of 0% on the test set. The corresponding confusion matrix is ​​shown below. Figure 8 As shown, Figure 8 This is a comparison diagram of the model performance confusion matrix in an arc fault detection model construction method based on expert experience constraints and performance iteration improvement provided in one embodiment of the present invention. In the diagram, (a) is the confusion matrix of the training set and (b) is the confusion matrix of the test set. Figure 8 The quantitative evidence chart for the final model performance clearly shows the model's classification results for faulty and non-faulty samples in the training and test sets. The quantitative results show that the confusion matrix directly verifies that the final model has no false positives and very few false negatives in both the training and test sets, achieving high accuracy rates of 97.6% and 97.08%, respectively. This data supports the core engineering metrics of low false alarm rate and high detection accuracy in this invention and provides a basic performance reference for subsequent tests on the generalization ability under unknown loads.

[0098] Furthermore, the model was applied to four types of unknown loads—fluorescent lamps, computers, and electric drills—to test its generalization ability. The experimental results show that the model can accurately distinguish between faulty and non-faulty samples under each load without generating false alarms. This confirms that the arc fault detection model constructed in this invention, based on expert experience constraints and performance iteration improvement, has good practicality and robustness.

[0099] In summary, the arc fault detection model construction method of this invention, based on expert experience constraints and performance iterative improvement, effectively limits the model search space by introducing expert experience to impose physical meaning constraints on feature thresholds, avoiding overfitting and physical uninterpretability problems caused by purely data-driven approaches, thereby improving the model's generalization ability to unknown loads. Through a two-stage iterative optimization mechanism that first reduces false alarms and then reduces false negatives, the overall accuracy of the model can be continuously improved while ensuring a low false alarm rate—a core engineering indicator—thus balancing detection reliability and accuracy. The hierarchical progressive correction sub-process of direct ancestor nodes enables refined and interpretable adjustments to the model structure, enhancing its adaptability to new loads and unknown scenarios without increasing model complexity. Ultimately, the constructed model possesses high detection accuracy, low false alarm rate, strong interpretability, and good generalization performance, significantly improving the reliability and practicality of arc fault detection technology in complex real-world scenarios.

[0100] Although the present invention has been disclosed above by way of embodiments, it is not intended to limit the present invention. Anyone skilled in the art can make some modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for constructing an arc fault detection model based on expert experience constraints and performance iterative improvement, characterized in that, include: An arc fault experimental platform was built to collect fault and non-fault current signals, extract time-domain and frequency-domain features to construct a feature set, and divide it into training set and test set. Based on the training set, an effective threshold range is defined for each feature dimension according to expert experience; Based on the effective threshold range, a candidate splitting threshold set is generated, and the optimal splitting point is selected from the candidate splitting threshold set to recursively construct the initial white-box detection model. Based on the false alarm rate of the initial white-box detection model on the training set, the initial white-box detection model is optimized in a two-stage iterative process to output the detection model. The initial white-box detection model is a decision tree, and the two-stage process includes a false positive rate reduction stage and a false negative rate reduction stage. The false alarm rate reduction stage includes: locating the leaf node with the highest false alarm risk in the model, and calling the hierarchical progressive correction process of the direct ancestor node to adjust the model structure, and then recalculating the false alarm rate; if the adjustment still cannot meet the requirements, the number of model node splits is increased, the initial model is reconstructed and iterated until the false alarm rate meets the constraints. The false alarm rate reduction stage includes: locating the leaf node with the highest false alarm risk in the model, and calling the hierarchical progressive correction process of the direct ancestor node to adjust the model structure, ensuring that the false alarm rate never exceeds the hard constraint threshold of the false alarm rate during the adjustment process.

2. The method for constructing an arc fault detection model based on expert experience constraints and performance iterative improvement as described in claim 1, characterized in that, The process of generating a candidate split threshold set includes: for each feature dimension, sorting the values ​​of all samples in the training set on the feature in ascending order to obtain an ordered sequence; generating a full candidate threshold set on the feature based on the midpoint of the values ​​of adjacent samples in the ordered sequence; and mapping the full candidate threshold set to an effective threshold interval defined by expert experience to form a candidate split threshold set, such that each candidate value in the set falls within the effective threshold interval.

3. The method for constructing an arc fault detection model based on expert experience constraints and performance iterative improvement as described in claim 1, characterized in that, Selecting the optimal split point from the candidate split threshold set includes: for each node to be split in the initial white-box detection model, traversing the candidate split threshold set corresponding to each feature dimension; for each candidate split threshold, calculating the Gini coefficient of the child nodes obtained after splitting with the threshold; and selecting the corresponding feature and threshold as the optimal split rule for the current node based on minimizing the Gini coefficient.

4. The method for constructing an arc fault detection model based on expert experience constraints and performance iterative improvement as described in claim 1, characterized in that, The initial white-box detection model is optimized through a two-stage iterative process, including: Calculate the false alarm rate of the current model on the training set and determine whether it meets the preset hard constraint threshold for false alarm rate. If the false alarm rate does not meet the hard constraint threshold for false alarm rate, then the false alarm rate reduction stage begins. If the false alarm rate meets the hard constraint threshold for the false alarm rate, then the false alarm rate reduction stage begins. During the false negative rate reduction phase, if model adjustments cause the false positive rate to exceed the hard constraint threshold or the accuracy cannot be further optimized, then the correction is stopped. Specifically, entering the false alarm rate reduction phase includes: For all leaf nodes in the decision tree model whose predictions are faulty; For each leaf node, the leaf node with the highest false alarm risk index is calculated as the worst leaf node to be optimized. The decision path containing the worst leaf node is corrected by calling the hierarchical progressive correction sub-process of the direct ancestor node; Reassess the false alarm rate of the corrected model. If the false alarm rate meets the hard constraint threshold, proceed to the false alarm rate reduction stage. Otherwise, check whether the correction attempts for all high-risk leaf nodes have been completed. If so, and the false alarm rate is still not up to standard, increase the number of splits and rebuild the initial decision tree to start a new round of iterative optimization. Entering the stage of reducing the false negative rate specifically includes: For all leaf nodes in the decision tree model whose prediction results are normal; For each leaf node, the leaf node with the highest false negative risk index is calculated as the worst leaf node to be optimized; The decision path containing the worst leaf node is corrected by calling the hierarchical progressive correction sub-process of the direct ancestor node; The revised model is evaluated to obtain the final optimized model.

5. The method for constructing an arc fault detection model based on expert experience constraints and performance iterative improvement as described in claim 4, characterized in that, The hierarchical progressive correction sub-process of the direct ancestor node is invoked to correct the decision path containing the worst leaf node, including: Starting from the worst leaf node to be optimized, backtrack along the decision path to its parent node; At the parent node level, a set of candidate splitting thresholds is generated based on the effective threshold range of each feature, and a modified search space is constructed. For each candidate splitting scheme in the constructed search space, its subtree performance is reconstructed and evaluated in parallel, and the optimal scheme is selected based on the current optimization stage. If the optimal solution is better than the original solution, complete the correction; otherwise, continue backtracking. Repeat the backtracking process level by level until the root node is reached; if there is still no better solution at the root node, it is determined that the current path cannot be optimized and the process ends.

6. The method for constructing an arc fault detection model based on expert experience constraints and performance iterative improvement as described in claim 5, characterized in that, For each candidate splitting scheme in the search space, its subtree performance is reconstructed and evaluated in parallel; the optimal scheme is selected based on the current optimization stage and implemented using a parallel computing architecture, while all splitting schemes in the modified search space are evaluated simultaneously.

7. The method for constructing an arc fault detection model based on expert experience constraints and performance iterative improvement as described in claim 5, characterized in that, In the two-stage iterative optimization process, an upper limit is preset for the number of node splits to prevent model overfitting, and a false alarm rate hard constraint threshold is set as the core criterion for iterative transformation and termination.

8. The method for constructing an arc fault detection model based on expert experience constraints and performance iterative improvement as described in claim 5, characterized in that, In the hierarchical progressive correction sub-process of the direct ancestor node, after updating the splitting rules of the parent node or the ancestor node at a higher level, the subtree to which it belongs needs to be reconstructed; the splitting criterion for subtree reconstruction is still based on minimizing the Gini coefficient, and the splitting threshold of each node after reconstruction must still meet the effective threshold range defined by the expert experience.

Images