Intelligent detection method for direct-current series arc fault in more-electric aircraft scene

Abstract

The invention provides an intelligent detection method for a direct-current series arc fault in a more-electric aircraft scene, and relates to the field of electrical parameter detection, and the method comprises the following steps: collecting direct-current series current data in a working process of a direct-current power supply system; inputting the direct-current series current data into an arc fault detection model to obtain a direct-current series arc fault probability and a direct-current series arc fault type; the arc fault detection model comprises four layers of time convolution networks TCN and one layer of relevance vector machine RVM which are connected in sequence; in the arc fault detection model before training, the number of channels of four layers of TCN is doubled layer by layer; in the trained arc fault detection model, pruning processing is carried out on all TCN channels, so that the parameter quantity of the TCN is reduced to 30% of the parameter quantity before training; and comparing the DC series arc fault probability with an alarm threshold, and when the DC series arc fault probability is greater than the alarm threshold, determining that the DC power supply system will have a corresponding type of DC series arc fault.

Description

Technical Field

[0001] The present invention relates to the field of electrical parameter detection, and in particular to an intelligent detection method for DC series arc faults in a multi-electric aircraft scenario. Background Art

[0002] High-voltage DC power supply technology for more-electric aircraft has advanced from tens of kilovolt-amperes to megavolt-amperes, and future capacity and voltage levels are expected to increase. However, the increased voltage and complex operating environment pose challenges to insulation performance, particularly the risk of DC arc faults. Series arcs, in particular, are difficult to detect due to their short duration, low current intensity, and irregular patterns.

[0003] Under high-noise conditions, series arc faults in the DC power supply system of multi-electric aircraft are instantaneous, random, and have high impedance characteristics. The current signal is often interfered with by the complex electromagnetic environment, and the resulting current fluctuations are often masked by the load noise, resulting in the true fault characteristics being masked in the time and frequency domains and insufficient extraction.

[0004] However, the existing time domain threshold detection method is easily affected by load fluctuations and cannot distinguish between high-frequency noise and real arc characteristics; the frequency domain feature analysis method relies on fixed frequency band division, consumes a lot of computing resources and has poor real-time performance.

[0005] Existing machine learning methods lack the ability to model the distribution differences between different mission scenarios, as the manifestation of arc faults in actual MEL aircraft applications varies depending on factors such as load, temperature, and flight altitude. This can lead to overfitting and severely limit the model's generalization. Single deep learning models have large parameters, high inference latency, and lack probabilistic confidence assessments of their outputs. Single probabilistic models are also weak at extracting time series features and require reliance on pre-processed feature engineering.

[0006] And considering the limitations of the actual deployment platform, most fault detection systems need to be deployed on resource-constrained embedded edge devices, which have strict constraints on inference latency, energy consumption and model size. Existing large models are difficult to meet the requirements of both real-time and reliability.

[0007] Therefore, the existing threshold method, frequency domain analysis, independent deep learning model or probability model all have limitations, and it is difficult to balance detection accuracy, response speed and deployment feasibility.

[0008] The existing intelligent detection methods for DC series arc faults in the scenario of multi-electric aircraft have problems such as difficulty in extracting low signal-to-noise ratio features under high noise interference, insufficient model generalization ability under complex working conditions, insufficient real-time performance on the edge device side, and lack of confidence. Summary of the Invention

[0009] In order to solve the technical problems existing in the above-mentioned prior art, the purpose of the present invention is to provide an intelligent detection method for DC series arc faults in a multi-electric aircraft scenario, which can improve the feature resolution capability in a noisy environment, the model generalization capability under complex working conditions, the real-time performance of the edge device end, and the confidence level.

[0010] To achieve the above-mentioned object, the present invention provides an intelligent detection method for DC series arc faults in a multi-electric aircraft scenario, the method being as follows:

[0011] Collect the DC series current data during the operation of the DC power supply system of the more electric aircraft;

[0012] Inputting the DC series current data into a pre-trained arc fault detection model to obtain a DC series arc fault probability and a corresponding DC series arc fault type;

[0013] The arc fault detection model includes a sequentially connected 4-layer temporal convolutional network TCN and a 1-layer relevance vector machine RVM;

[0014] In the pre-training arc fault detection model, the number of channels in the four-layer TCN doubled layer by layer. In the post-training arc fault detection model, all TCN channels were pruned, reducing the number of TCN parameters to 30% of the pre-training level.

[0015] The DC series arc fault probability is compared with an alarm threshold, and when the DC series arc fault probability is greater than the alarm threshold, it is determined that a DC series arc fault of a corresponding type will occur in the DC power supply system.

[0016] According to a technical solution of the present invention, it also includes:

[0017] generating a fault handling instruction when it is determined that a DC series arc fault will occur in the DC power supply system;

[0018] Based on the fault handling instruction, the DC power supply system is disconnected.

[0019] According to a technical solution of the present invention, the width of the convolution kernel in the 4-layer TCN is 5;

[0020] The expansion factors of the 1st to 4th TCN layers are 1, 2, 4, and 8, respectively;

[0021] In the arc fault detection model before training, the number of channels of the 1st to 4th layer TCN are 64, 128, 256 and 512 respectively.

[0022] According to a technical solution of the present invention, the training process of the arc fault detection model is as follows:

[0023] Simulate DC series arc faults and collect arc fault metadata;

[0024] Initialize global parameters and deploy them into the arc fault detection model;

[0025] The global parameters include TCN model parameters and RVM hyperparameters;

[0026] Construct multiple meta-learning training tasks. In each meta-learning training task, use 5 to 10 arc fault metadata as the support set and 15 arc fault metadata as the query set.

[0027] Perform each meta-learning training task using the corresponding support set and query set to obtain updated global parameters;

[0028] The updated global parameters are back-propagated to update the original global parameters in the arc fault detection model until a stopping condition is reached, thereby finally obtaining a trained arc fault detection model.

[0029] According to a technical solution of the present invention, the meta-learning is MAML meta-learning;

[0030] The internal learning rate of the training task is 0.001;

[0031] The external learning rate of the training task is 0.0001;

[0032] The stopping condition is reaching a preset number of training rounds, which is 500 to 1000;

[0033] The optimizer used in the training task execution process is the Adam optimizer.

[0034] According to a technical solution of the present invention, it also includes:

[0035] In the trained arc fault detection model, the TCN parameters are quantized and the weights in the TCN are converted into 8-bit fixed-point numbers.

[0036] According to a technical solution of the present invention, arc fault metadata includes current data corresponding to arcs of different durations at different voltage levels for different types of loads;

[0037] The types of loads include resistive loads, capacitive loads and inductive loads;

[0038] The voltage levels include 270V and 540V;

[0039] The duration is 0.2 ms to 5 ms.

[0040] According to a technical solution of the present invention, the DC series arc fault probability is compared with the alarm threshold, and the specific process is as follows;

[0041] Based on the DC series current data under the current working conditions, the corresponding alarm threshold is matched with the rules in the alarm threshold rule library;

[0042] The alarm threshold rule base is pre-built based on historical DC series current data;

[0043] The rule is the correspondence between historical DC series current data and corresponding alarm thresholds under different working conditions;

[0044] The DC series arc fault probability is compared with the corresponding alarm threshold.

[0045] The present invention also provides an electronic device, comprising: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory, so that the electronic device performs the above-mentioned intelligent detection method for DC series arc faults in the multi-electric aircraft scenario.

[0046] The present invention also provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the above-mentioned intelligent detection method for DC series arc faults in the multi-electric aircraft scenario is implemented.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The intelligent detection method for DC series arc faults in the multi-electric aircraft scenario proposed in this paper combines the efficient feature extraction capability of TCN in multi-scale sequence modeling with the advantages of RVM in sparse classification and confidence output. The specific details are as follows:

[0049] 1. The introduction of TCN, using its dilated causal convolution structure and multi-level convolution kernel, can capture the long-term temporal dependencies of current signals and adaptively extract multi-scale features without manually presetting frequency bands or thresholds, thereby improving feature resolution in noisy environments.

[0050] 2. Incorporating relevance vector machines (RVMs) and leveraging their sparse Bayesian learning properties, this approach reduces backend classification computational overhead and enables end-to-end low-latency inference. Furthermore, an RVM probabilistic classifier is embedded in the TCN backend, mapping high-dimensional features into confidence levels for fault probabilities. This output, rather than a single label, avoids false triggering of protective devices due to occasional interference.

[0051] 3. By extracting multi-scale temporal features using TCN and filtering sparsity using RVM, this model significantly enhances its sensitivity to weak arc signatures, achieving an average 7.2% improvement in recognition accuracy across a test set containing varying voltage levels and load disturbances. Combined with an end-to-end optimized architecture, the entire model's inference latency is kept below 5 milliseconds, meeting the real-time protection requirements of DC circuit breakers.

[0052] 4. By introducing channel pruning and parameter quantization strategies into the network structure design, the model's overall parameter size is compressed to 30% of the original while maintaining performance, and converted into 8-bit fixed-point representation, making it suitable for embedded platforms with limited computing resources such as the ARM Cortex-M7, achieving efficient and low-power edge deployment.

[0053] In summary, the present invention solves the problem of collaborative optimization of feature robustness, real-time performance, and decision confidence in DC arc detection through the coupling design of deep features and probabilistic reasoning. This enables the present invention to have significant technical advantages and practical value in DC scenarios with high noise and high reliability requirements, such as multi-electric aircraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0055] Figure 1 A schematic diagram illustrating a principle diagram of an intelligent detection method for DC series arc faults in a more-electric aircraft scenario according to one embodiment of the present invention;

[0056] Figure 2 Schematically showing the structure of a signal conditioning circuit in an intelligent detection method for DC series arc faults in a multi-electric aircraft scenario according to one embodiment of the present invention;

[0057] Figure 3 A schematic diagram illustrating a system structure of an intelligent detection method for DC series arc faults in a multi-electric aircraft scenario according to an embodiment of the present invention;

[0058] Figure 4 A schematic diagram illustrating a structure of a connection between a TCN and an SVM in an intelligent detection method for DC series arc faults in a multi-electric aircraft scenario according to an embodiment of the present invention;

[0059] Among them, TCN is the 4th layer TCN. DETAILED DESCRIPTION

[0060] The description of the embodiments in this specification should be combined with the corresponding drawings, which should be considered a complete part of this specification. In the drawings, the shapes and thicknesses of the embodiments may be exaggerated and indicated for simplicity or convenience. Furthermore, the various structural components in the drawings will be described separately. It is worth noting that components not shown in the drawings or not described in words are known to those of ordinary skill in the art.

[0061] The description of the embodiments herein and any references to directions and orientations are for ease of description only and are not to be construed as limiting the scope of the present invention. The following description of the preferred embodiments may involve combinations of features, which may exist independently or in combination. The present invention is not specifically limited to the preferred embodiments. The scope of the present invention is defined by the claims.

[0062] like Figures 1 to 4 As shown, the present invention provides an intelligent detection method for DC series arc faults in a multi-electric aircraft scenario, and the method is as follows:

[0063] Collect the DC series current data during the operation of the DC power supply system of the more electric aircraft;

[0064] Inputting DC series current data into a pre-trained arc fault detection model to obtain the DC series arc fault probability and the corresponding DC series arc fault type;

[0065] The arc fault detection model consists of a sequentially connected 4-layer temporal convolutional network (TCN) and a 1-layer relevance vector machine (RVM);

[0066] In the pre-training arc fault detection model, the number of channels in the four-layer TCN doubled layer by layer. In the post-training arc fault detection model, all TCN channels were pruned, reducing the number of TCN parameters to 30% of the pre-training level.

[0067] The DC series arc fault probability is compared with the alarm threshold. When the DC series arc fault probability is greater than the alarm threshold, it is determined that a DC series arc fault of a corresponding type will occur in the DC power supply system.

[0068] In this embodiment, a high-precision, low-latency, and strongly generalized DC series arc fault detection system is constructed, which is a hybrid model (TCN-RVM) that integrates TCN (Temporal Convolutional Networks) and RVM (Relevance Vector Machine). Through the multi-scale time series feature extraction of TCN, it effectively separates noise and fault features, enhances the ability to resist noise interference, and improves the stability and reliability in the complex working environment of multi-electric aircraft, ensuring that the detection system can still accurately perform signal processing and fault detection tasks under high noise background; the lightweight TCN and sparse RVM joint architecture achieves a model inference time of less than 5ms, meeting the action requirements of the circuit breaker and achieving millisecond-level response; RVM outputs fault probability confidence, supports dynamic threshold adjustment, reduces false alarm rate, and realizes probabilistic decision-making.

[0069] like Figure 1 As shown, the system for implementing the intelligent detection method for DC series arc faults in the multi-electric aircraft scenario of this embodiment includes two parts: a hardware acquisition layer and an algorithm processing layer.

[0070] The hardware acquisition layer includes a signal conditioning circuit and a high-speed analog-to-digital conversion (ADC) module. The signal conditioning circuit converts the current in the DC power supply system of the multi-electric aircraft into a voltage signal, which is then fed into the high-speed ADC module for measurement. The high-speed ADC module has a sampling rate of 500 kSps and an accuracy of 16 bits.

[0071] like Figure 2 As shown in the figure, in the signal conditioning circuit, a 1mΩ chip alloy sampling resistor is selected for arc current sampling. Its maximum power is 3W and the maximum sampling current it can support is 54A, which is sufficient to meet the maximum power requirement of the system. In the subsequent signal amplification circuit, a non-inverting amplification configuration is adopted, and the amplification factor is set to 100 times. This can effectively amplify the weak sampling signal, thereby meeting the sampling accuracy requirements of the subsequent analog-to-digital converter (ADC). Figure 2 As shown in the figure, R1 is the current sampling resistor, while R2 serves as the load resistor, whose value can be flexibly adjusted according to the actual application scenario. R3, R4, R5, and R6 correspond to the key parameter resistors in the non-inverting amplifier circuit. C1 and C2 serve as filter capacitors to suppress high-frequency noise interference. In addition, the amplifier is powered by a stable 5V power supply, and its output is connected to the ADC to ensure accurate signal digitization.

[0072] The algorithm processing layer consists of three parts: TCN feature extraction part, RVM probability classification part and threshold decision module.

[0073] Among them, the TCN feature extraction module is a lightweight time convolutional network used for multi-scale time series feature learning. The RVM probability classification module is a sparse classifier based on the Bayesian framework, which outputs DC series arc faults. The threshold decision module provides an alarm threshold and compares the DC series arc fault probability and the alarm threshold to ultimately determine whether a DC series arc fault has occurred.

[0074] The TCN feature extraction module mainly consists of three parts: causal convolution, dilated convolution and residual connection.

[0075] Causal convolution limits the access scope of the convolution kernel, ensuring that the output of the current time step depends only on the current and previous inputs, thus preventing future information leakage. The mathematical form is Equation (1), which ensures that the model makes judgments based only on historical current signals during real-time detection.

[0076]

[0077] Among them, K is the width of the convolution kernel, ω k is the kth convolution kernel weight, x t-k For historical input.

[0078] Dilated convolution inserts a "hole" into the standard convolution to expand the receptive field without increasing the number of parameters. The dilation factor d controls the hole spacing, and the dilation factor of the lth layer is usually d = 2 l-1 , the output is calculated as Equation (2). In this implementation, a 4-layer TCN is used with an expansion factor sequence of [1, 2, 4, 8], so that the top-layer receptive field covers multiple time steps, which is sufficient to capture the transient and continuous characteristics of arc faults.

[0079]

[0080] In each layer, the TCN module consists of dilated convolution, weight normalization (WeightNorm), ReLU activation function and residual skip connection. The output is formula (3), which is used to alleviate the gradient vanishing problem of deep networks, accelerate model convergence, and retain the low-level features of the original signal.

[0081] Output=ReLU(Conv(x))+x (3)

[0082] The network architecture of this embodiment accepts a current signal sequence (DC series current data) with a length of T = 1024 (corresponding to a 500kHz sampling rate and a 2.048ms time window). The convolution kernel width K = 5, the dilation factor increases layer by layer (d = 1, 2, 4, 8), and the number of channels doubles layer by layer (64 → 128 → 256 → 512) to enhance feature expression.

[0083] Weight normalization (WeightNorm) is used to adapt to online streaming data, and nonlinearity is introduced with the ReLU activation function. To optimize the model, a channel pruning strategy is employed. After training, the L1 norm of each channel is calculated to filter out low-importance channels, reducing the number of parameters to 30% of the original TCN. Fine-tuning then restores accuracy (loss of <1%). Furthermore, parameter quantization converts 32-bit floating-point weights to 8-bit fixed-point numbers, reducing the model size by 75%, meeting the memory requirements of embedded devices.

[0084] During feature extraction, the current signal sequence is preprocessed using Z-score normalization. In multi-scale feature extraction, the first TCN layer (d = 1) captures local high-frequency fluctuations and microsecond arc pulses; the second TCN layer (d = 2) extracts millisecond-level current trends; and the third and fourth TCN layers (d = 4, 8) model second-level load states. Finally, residual connections are used to fuse shallow and deep features, outputting a 128-dimensional feature vector.

[0085] The RVM module is mainly constructed by four steps: input feature space construction, kernel function selection and parameter optimization, probabilistic classification output, sparsity implementation and related vector screening.

[0086] Construction of the input feature space: The input source is the 128-dimensional high-dimensional feature vector output by the above-mentioned TCN, which is used as the input x of the RVM. The TCN features are further standardized (using the Z-score) to eliminate dimensional differences and improve the stability of the classifier.

[0087] Kernel function selection and parameter optimization: The kernel function type adopts radial basis kernel function, which is expressed as (4), where σ is the kernel width parameter.

[0088]

[0089] In the process of automatic parameter optimization, this implementation is based on the Bayesian inference framework and assumes that the weight parameter w follows the prior distribution p(w|α), where α represents a hyperparameter. The Type-II ML method is used to maximize the marginal likelihood function p(t|α,σ 2 ), iteratively update the hyperparameters α and σ until convergence. During the optimization process, most of α is induced by sparsity. i Will tend to infinity, making the corresponding weight w i approaches zero. Eventually, only a small number of i The samples of are retained as correlation vectors.

[0090] Then, the probabilistic classification output is performed. For the input feature x, the RVM outputs the posterior probability that it belongs to the fault category as formula (5), where φ(x) is the feature vector after kernel function mapping, and w is the sparse weight vector.

[0091]

[0092] Based on the actual operating conditions of a multi-electric aircraft, a probability threshold T can be set. When the posterior probability P(y=1|x) exceeds the probability threshold T, the system triggers an alarm; otherwise, the system is deemed to be in a normal state. This probability threshold-based judgment mechanism allows the system to flexibly adjust its sensitivity based on actual operating conditions, avoiding missed or misjudgment.

[0093] Sparsity implementation and related vector screening: The related vector is the corresponding hyperparameter α i is a finite value sample, whose weight w i Non-zero, plays a decisive role in classification decisions. The screening process first initializes all samples to α i The same value is used. α is iteratively updated by maximizing the second-class maximum likelihood method to automatically eliminate redundant samples. Finally, about 5% of the samples are retained as relevant vectors to form a sparse model.

[0094] Through the construction of the above modules, a joint training framework is finally formed. The optimization goal is to integrate the TCN feature extraction and RVM classification tasks into the same loss function, which is defined as Equation (6).

[0095]

[0096] Among them, the first term is the negative logarithm of the marginal likelihood of RVM, the second term is the TCN weight regularization term, and λ is a hyperparameter.

[0097] like Figure 3 As shown in Figure 2, during the backpropagation process, the DC series current data is extracted through the TCN and then input into the RVM to calculate the probability output. After calculating the loss function, the gradient is backpropagated to update the parameters of the TCN and RVM, thus achieving joint optimization.

[0098] Furthermore, this implementation utilizes a dynamic feature adaptation mechanism, weighting feature importance to allow the RVM to feed back classification errors to the TCN module during training. This process leverages gradients to adjust the channel weights of the TCN convolution kernels, thereby enhancing the ability to extract fault-sensitive features. Furthermore, to prevent overfitting, this implementation introduces a 10% dropout layer at the end of the TCN to improve feature generalization and prevent the RVM from over-relying on specific patterns in the training set.

[0099] The intelligent detection method for DC series arc faults in a multi-electric aircraft scenario also includes:

[0100] generating a fault handling instruction when it is determined that a DC series arc fault will occur in the DC power supply system;

[0101] Based on the fault handling instructions, the DC power supply system is disconnected.

[0102] In this embodiment, the system for implementing the intelligent detection method of DC series arc faults in the multi-electric aircraft scenario of this embodiment further includes an embedded execution layer, which includes a protection execution mechanism and an edge computing unit.

[0103] Among them, the protection execution mechanism is linked with the DC circuit breaker. When it is determined that a DC series arc fault will occur in the DC power supply system, a fault processing instruction is generated to trigger the fault isolation action.

[0104] The edge computing unit is a single-chip microcontroller based on the ARM Cortex-M7 architecture, which is used to carry the arc fault detection model, and can perform DC arc fault identification and processing on multi-electric aircraft.

[0105] In the intelligent detection method of DC series arc faults in the multi-electric aircraft scenario, the convolution kernel width in the four-layer TCN is 5;

[0106] The expansion factors of the 1st to 4th TCN layers are 1, 2, 4, and 8, respectively;

[0107] In the arc fault detection model before training, the number of channels of the 1st to 4th layer TCN are 64, 128, 256 and 512 respectively.

[0108] In this embodiment, a lightweight TCN module is constructed by optimizing TCN and convolution kernel width.

[0109] In the intelligent detection method for DC series arc faults in the multi-electric aircraft scenario, the arc fault detection model training process is as follows:

[0110] Simulate DC series arc faults and collect arc fault metadata;

[0111] Initialize global parameters and deploy them into the arc fault detection model;

[0112] Global parameters include TCN model parameters and RVM hyperparameters;

[0113] Construct multiple meta-learning training tasks. In each meta-learning training task, use 5 to 10 arc fault metadata as the support set and 15 arc fault metadata as the query set.

[0114] Perform each meta-learning training task using the corresponding support set and query set to obtain updated global parameters;

[0115] The updated global parameters are back-propagated to update the original global parameters in the arc fault detection model until the stopping condition is reached, and finally a trained arc fault detection model is obtained.

[0116] In this embodiment, the algorithm processing layer also includes a meta-learning module.

[0117] During the training task construction phase, diverse fault simulation scenarios are introduced, including current data (arc fault metadata) corresponding to arcs of different durations and different types of loads at different voltage levels.

[0118] On this basis, the total number of training tasks in meta-learning training is set to 20. Each training task contains a support set consisting of 5 to 10 samples for local arc fault detection model update, and a query set consisting of 15 samples for global gradient optimization evaluation. In this way, shared structures across training tasks can be extracted from multiple small-sample training tasks, and the arc fault detection model can be quickly migrated to unknown working conditions.

[0119] The meta-learning training task is performed through the meta-learning module as follows:

[0120] Initialization: Global parameters θ (including TCN weights and RVM hyperparameters).

[0121] Inner loop update: For each task T i , perform gradient descent on the support set.

[0122] θ' i =θ-α▽ θ L support (θ)

[0123] Outer loop update: calculate meta-loss on the query set, update,

[0124] θ←θ-β▽ θ Σ i L query (θ' i )

[0125] Iteration: until the meta-epoch number of meta-learning training rounds is completed.

[0126] This implementation introduces the MAML (Model-Agnostic Meta-Learning) mechanism, which optimizes parameter initialization through a shared structure between tasks during the training phase. This allows the arc fault detection model to quickly adapt to unknown working conditions by relying on only a small number of new samples, ultimately achieving a highly robust, highly portable, and highly real-time intelligent fault identification solution for the end user.

[0127] Leveraging the MAML meta-learning mechanism, this implementation enables rapid fine-tuning of the model's local parameters under completely unseen operating conditions using only 5–10 samples, significantly accelerating the transfer of arc fault detection models to new tasks. Experiments demonstrate that this approach improves adaptability by approximately three times compared to existing transfer learning methods, making it particularly suitable for new load scenarios with limited arc fault samples.

[0128] The TCN-RVM joint training framework of this embodiment realizes end-to-end automated learning by synchronously optimizing the feature extraction and classification modules through gradient backpropagation. It utilizes the sparsity of RVM and the parameter sharing mechanism of TCN to reduce the model storage and computing resource requirements, making it embeddable in edge devices to meet the needs of real-time detection on site.

[0129] During the training process of the arc fault detection model, meta-learning is MAML meta-learning;

[0130] The internal learning rate of the training task is 0.001;

[0131] The external learning rate for the training task is 0.0001;

[0132] The stopping condition is to reach the preset number of training rounds, which is 500 to 1000;

[0133] The optimizer used during the training task execution is the Adam optimizer.

[0134] In this implementation, the parameter configuration table for meta-learning is as follows:

[0135]

[0136] The intelligent detection method for DC series arc faults in the multi-electric aircraft scenario also includes:

[0137] In the trained arc fault detection model, the TCN parameters are quantized and the weights in the TCN are converted into 8-bit fixed-point numbers.

[0138] In the intelligent detection method for DC series arc faults in the multi-electric aircraft scenario, the arc fault metadata includes the current data corresponding to arcs of different durations at different voltage levels for different types of loads;

[0139] The types of loads include resistive loads, capacitive loads and inductive loads;

[0140] Voltage levels include 270V and 540V;

[0141] The duration is 0.2ms to 5ms.

[0142] In the intelligent detection method of DC series arc faults in the multi-electric aircraft scenario, the DC series arc fault probability is compared with the alarm threshold. The specific process is as follows;

[0143] Based on the DC series current data under the current working conditions, the corresponding alarm threshold is matched with the rules in the alarm threshold rule library;

[0144] The alarm threshold rule base is pre-built based on historical DC series current data;

[0145] The rule is the correspondence between historical DC series current data and corresponding alarm thresholds under different working conditions;

[0146] The DC series arc fault probability is compared with the corresponding alarm threshold.

[0147] In this implementation, the threshold decision module adaptively adjusts the alarm threshold based on historical data and operating conditions. Specifically, a dynamic probability threshold T can be set based on the actual operating conditions of the multi-electric aircraft. When the posterior probability P(y=1|x) exceeds probability threshold T, the system triggers a DC series arc fault alarm; otherwise, the system is deemed to be in a normal state.

[0148] This embodiment is based on a dynamic probability threshold judgment mechanism, which enables the system to flexibly adjust the sensitivity according to the on-site working conditions to avoid missed judgments and misjudgments.

[0149] Compared with the existing hard decision method, the RVM module outputs the posterior probability that each sample belongs to the fault category. Combined with the dynamically adjustable threshold decision mechanism, the system can flexibly adjust the sensitivity according to the operating conditions, effectively suppressing false alarms and missed alarms, and enhancing the confidence and interpretability of the judgment results.

[0150] According to one aspect of the present invention, an electronic device is provided, comprising: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is operating, the processor executes the one or more computer programs stored in the memory, so that the electronic device performs the intelligent detection method for DC series arc faults in a multi-electric aircraft scenario as described in the above technical solution.

[0151] According to one aspect of the present invention, a computer-readable storage medium is provided for storing computer instructions. When the computer instructions are executed by a processor, the intelligent detection method for DC series arc faults in a multi-electric aircraft scenario as described in the above technical solution is implemented.

[0152] Computer-readable storage media may include any medium capable of storing or transmitting information. Examples of computer-readable storage media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, and the like. The code segments may be downloaded via a computer network such as the Internet, an intranet, and the like.

[0153] This embodiment provides an intelligent detection method for DC series arc faults in a multi-electric aircraft scenario, relating to the field of electrical parameter detection. The method comprises: collecting DC series current data during operation of a DC power supply system; inputting the DC series current data into an arc fault detection model to obtain a DC series arc fault probability and a DC series arc fault type; the arc fault detection model comprises a sequentially connected four-layer temporal convolutional network (TCN) and a one-layer relevance vector machine (RVM); in the pre-trained arc fault detection model, the number of channels in the four-layer TCN is doubled layer by layer; in the post-trained arc fault detection model, all TCN channels are pruned to reduce the number of TCN parameters to 30% of the pre-training value; and comparing the DC series arc fault probability with an alarm threshold. When the DC series arc fault probability is greater than the alarm threshold, it is determined that a DC series arc fault of the corresponding type will occur in the DC power supply system.

[0154] Furthermore, it should be noted that the present invention may be provided as a method, apparatus, or computer program product. Thus, embodiments of the present invention may take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention may take the form of a computer program product embodied on one or more computer-usable storage media containing computer-usable program code.

[0155] The embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the process in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0156] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device, so that a series of operation steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for implementing the process in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0157] It should also be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or terminal device comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or terminal device comprising the element.

[0158] Finally, it should be noted that the above is a preferred embodiment of the present invention. It should be noted that although the preferred embodiment of the present invention has been described, it is clear that those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles of the present invention. Such improvements and modifications should also be considered as within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including the preferred embodiment and all changes and modifications that fall within the scope of the embodiments of the present invention.

Claims

1. An intelligent detection method for DC series arc faults in a multi-electric aircraft scenario, characterized in that: The method is as follows: Collect the DC series current data during the operation of the DC power supply system of the more electric aircraft; Inputting the DC series current data into a pre-trained arc fault detection model to obtain a DC series arc fault probability and a corresponding DC series arc fault type; The arc fault detection model includes a sequentially connected 4-layer temporal convolutional network TCN and a 1-layer relevance vector machine RVM; In the pre-training arc fault detection model, the number of channels in the four-layer TCN doubled layer by layer. In the post-training arc fault detection model, all TCN channels were pruned, reducing the number of TCN parameters to 30% of the pre-training level. The DC series arc fault probability is compared with an alarm threshold, and when the DC series arc fault probability is greater than the alarm threshold, it is determined that a DC series arc fault of a corresponding type will occur in the DC power supply system.

2. The intelligent detection method for DC series arc faults in a multi-electric aircraft scenario according to claim 1 is characterized in that: Also includes: generating a fault handling instruction when it is determined that a DC series arc fault will occur in the DC power supply system; Based on the fault handling instruction, the DC power supply system is disconnected.

3. The intelligent detection method for DC series arc faults in a multi-electric aircraft scenario according to claim 1 or 2, characterized in that: The convolution kernel width in the 4-layer TCN is 5; The expansion factors of the 1st to 4th TCN layers are 1, 2, 4, and 8, respectively; In the arc fault detection model before training, the number of channels of the 1st to 4th layer TCN are 64, 128, 256 and 512 respectively.

4. The intelligent detection method for DC series arc faults in a multi-electric aircraft scenario according to claim 3 is characterized in that: The training process of the arc fault detection model is as follows: Simulate DC series arc faults and collect arc fault metadata; Initialize global parameters and deploy them into the arc fault detection model; The global parameters include TCN model parameters and RVM hyperparameters; Construct multiple meta-learning training tasks. In each meta-learning training task, use 5 to 10 arc fault metadata as the support set and 15 arc fault metadata as the query set. Perform each meta-learning training task using the corresponding support set and query set to obtain updated global parameters; The updated global parameters are back-propagated to update the original global parameters in the arc fault detection model until a stopping condition is reached, thereby finally obtaining a trained arc fault detection model.

5. The intelligent detection method for DC series arc faults in a multi-electric aircraft scenario according to claim 4 is characterized in that: The meta-learning is MAML meta-learning; The internal learning rate of the training task is 0.001; The external learning rate of the training task is 0.0001; The stopping condition is reaching a preset number of training rounds, which is 500 to 1000; The optimizer used in the training task execution process is the Adam optimizer.

6. The intelligent detection method for DC series arc faults in a multi-electric aircraft scenario according to claim 4 or 5, characterized in that: Also includes: In the trained arc fault detection model, the TCN parameters are quantized and the weights in the TCN are converted into 8-bit fixed-point numbers.

7. The intelligent detection method for DC series arc faults in a multi-electric aircraft scenario according to claim 6 is characterized in that: Arc fault metadata includes current data corresponding to arcs of different durations at different voltage levels for different types of loads; The types of loads include resistive loads, capacitive loads and inductive loads; The voltage levels include 270V and 540V; The duration is 0.2 ms to 5 ms.

8. The intelligent detection method for DC series arc faults in a multi-electric aircraft scenario according to any one of claims 1, 2, 4, 5 and 7, characterized in that: Comparing the DC series arc fault probability with the alarm threshold, the specific process is as follows; Based on the DC series current data under the current working conditions, the corresponding alarm threshold is matched with the rules in the alarm threshold rule library; The alarm threshold rule base is pre-built based on historical DC series current data; The rule is the correspondence between historical DC series current data and corresponding alarm thresholds under different working conditions; The DC series arc fault probability is compared with the corresponding alarm threshold.

9. An electronic device, characterized in that: include: One or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to enable the electronic device to perform the intelligent detection method for DC series arc faults in a multi-electric aircraft scenario as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that Used to store computer instructions, which, when executed by a processor, implement the intelligent detection method for DC series arc faults in a multi-electric aircraft scenario according to any one of claims 1 to 8.

Images