Method for evaluating health of dc converter based on pulse neural network
By improving the network structure and feature extraction method of spiking neural networks, and combining wavelet transform and compressed attention mechanism, a lightweight model is constructed, which solves the problems of computational accuracy and adaptability of spiking neural networks in resource-constrained environments, and realizes efficient health assessment and detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANTONG UNIV
- Filing Date
- 2026-02-26
- Publication Date
- 2026-06-02
Smart Images
Figure CN122133726A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to a health assessment method for DC-DC converters based on pulse neural networks. Background Technology
[0002] While spiking neural networks (SNNs) have demonstrated certain advantages in resource-constrained embedded and edge devices due to their low power consumption and asynchronous parallel processing capabilities, existing SNNs still face a series of technical challenges. First, because the feature representation of SNNs primarily relies on discrete time-stamped sequences, their computational accuracy is often lower than that of traditional continuous-value neural networks. Second, limited by network structure and information representation capabilities, their ability to model high-dimensional complex features is insufficient, resulting in limited feature representation capabilities. Third, in complex task scenarios (such as high-precision classification, predictive maintenance, or dynamic scene recognition), their adaptability is poor, making it difficult to meet practical application requirements.
[0003] In summary, this invention aims to propose a novel network structure that improves the feature representation capability and computational accuracy of spiking neural networks while also considering low power consumption, high efficiency, and adaptability to complex tasks. This further enhances the overall performance of the model, overcomes the main bottlenecks of existing technologies, and meets the needs of practical applications. Summary of the Invention
[0004] The purpose of this invention is to at least partially solve the above-mentioned technical problems and to provide a health assessment method for DC-DC converters based on pulse neural networks.
[0005] Building upon the foregoing discussion, and addressing the combined demands for low power consumption, high efficiency, and accurate computation in resource-constrained environments, there is an urgent need to develop a novel network architecture. This architecture should overcome the shortcomings of existing spiking neural networks in terms of accuracy, feature representation, and task adaptability, while fully leveraging their advantages in low power consumption and sparse computation. An ideal solution should possess the following characteristics: firstly, by improving the network structure or introducing new computational modules, the feature representation capability should be enhanced, thereby improving the network's adaptability to complex tasks; secondly, the network computation path should be optimized to reduce power consumption and resource consumption, further improving the model's practical deployment performance in embedded and edge devices.
[0006] According to a first aspect of the present invention, a method for assessing the health of a DC-DC converter based on a spiking neural network is provided, comprising:
[0007] By integrating the information from each channel in the feature matrix, this module performs axial compression and enhancement operations on the features, proposing a compressed attention mechanism. First, this mechanism designs a compression strategy based on feature matrix compression operations, compressing the input features in both horizontal and vertical directions (horizontal and vertical compression needs to be specified in the claims), reducing computational complexity while ensuring the retention of key information from the input features across different dimensions. Then, features are extracted from the original data using a weight matrix, improving the network's feature discrimination capability.
[0008] To further enhance the relevance of feature information, the compression enhancement module also strengthens the sensitivity of compressed axial attention to specific locations by constructing learned parameters. This method, through the introduction of positional embedding, enables the feature matrix to more accurately describe the positional information of features across different extended dimensions. Simultaneously, this module utilizes scaling operations to process different input signals, improving its ability to capture local fine-grained features.
[0009] The health assessment method for a DC-DC converter based on an improved pulse neural network according to an embodiment of the present invention has at least one of the following advantages:
[0010] (1) This invention combines wavelet transform technology to perform feature extraction and noise reduction processing by sampling the voltage signal of DC-DC converter in multiple time sequences, thereby generating a high-quality image dataset, which provides a solid data foundation for subsequent model training and health assessment.
[0011] (2) A lightweight spiking neural network based on the Ghost-Spiking module is proposed, which significantly reduces the computational complexity and energy consumption of the model, making it particularly suitable for the real-time deployment requirements of edge devices. At the same time, by fusing the compressed axial attention mechanism and the detail enhancement module, the feature capture capability and the expressive power of the model are further improved.
[0012] (3) By introducing an optimized spiking neuron model, this invention abandons the high energy consumption design of traditional networks and achieves efficient processing with low power consumption, enabling accurate detection of different health states of power converters in resource-constrained environments.
[0013] The "Ghost-Spiking module," a modeling and lightweight module based on the information transmission and dynamic mechanisms between neurons, significantly reduces the computational cost of feature extraction by employing phantom convolutions, thus achieving network lightweighting.
[0014] (4) Based on (3), the number of feature channels in the unit blocks is gradually increased to construct an efficient network structure suitable for DC-DC converter health assessment. To further improve the robustness and detection accuracy of the model in complex environments, this invention introduces a hardware-friendly compression-enhanced axial attention mechanism in some pulse unit blocks, thereby achieving a balance between network depth and detection accuracy. Experimental results show that the method described in this invention can maintain high detection accuracy while reducing computational resource consumption, is suitable for edge detection devices, and shows broad application prospects in the engineering application of power supply equipment health assessment. Attached Figure Description
[0015] These and / or other aspects and advantages of the present invention will become apparent and readily understood from the following description of preferred embodiments taken in conjunction with the accompanying drawings, in which:
[0016] Figure 1 A flowchart illustrating a health assessment method for a DC-DC converter based on an improved pulsating neural network according to an embodiment of the present invention.
[0017] Figure 2 for Figure 1 A schematic diagram of the information transmission principle of a spiking neural network in a computer.
[0018] Figure 3 for Figure 1 A schematic diagram of axial compression enhancing attention;
[0019] Figure 4 for Figure 1 The structure diagram of the improved spiking neural network. Detailed Implementation
[0020] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings. In this specification, the same or similar reference numerals indicate the same or similar components. The following description of the embodiments of the present invention with reference to the accompanying drawings is intended to explain the overall inventive concept of the present invention and should not be construed as a limitation thereof.
[0021] See Figure 1 The diagram shows a flowchart of a method for assessing the health of a DC-DC converter based on a pulsating neural network according to an embodiment of the present invention.
[0022] Specifically, the health assessment method described in this invention is particularly suitable for resource-constrained edge devices. Traditional health assessment methods for DC-DC converters based on convolutional neural networks suffer from high computational complexity and difficulty in deployment.
[0023] To address this technical problem, the inventors of this invention have innovatively designed a health assessment method for DC-DC converters based on spiking neural networks from the perspective of neural network structure optimization.
[0024] The health assessment method includes the following steps:
[0025] Step S1: Collect voltage ripple signals of the DC-DC converter circuit under normal conditions and different health conditions to obtain the original sample dataset.
[0026] The main purpose of step S1 is to collect fault data, such as voltage ripple data, and then obtain raw sample datasets under, for example, 10 different health conditions (state stages), to provide a basis for subsequent data processing.
[0027] Specifically, for example, an oscilloscope is used to acquire the voltage ripple signal in step S1 to obtain a raw sample dataset covering, for example, 10 different state stages.
[0028] Specifically, the process of acquiring the voltage ripple signal of the DC-DC converter circuit under normal conditions and different health conditions in step S1 includes:
[0029] S11. Use an oscilloscope to acquire various voltage ripple signals of the DC-DC converter circuit under normal conditions and different health conditions; for example, use an oscilloscope to perform the acquisition action in step S11. The different health conditions include various possible health conditions, such as normal operating conditions and various fault conditions.
[0030] S12. Divide the voltage ripple signal into a predetermined number of different health state stages to generate an original sample dataset covering each state. This original sample dataset provides a foundation for subsequent data analysis and model training. The predetermined number can be 3, 5, 10, or more. In this example, the predetermined number is selected as 10.
[0031] Step S2: Use wavelet transform to preprocess the collected original sample dataset to generate a processed image dataset and divide it into a training dataset and a test dataset.
[0032] Specifically, the image dataset is divided into training and testing datasets in a ratio of, for example, 1:4. This is understandable; other ratios such as 1:2 or 1:5 can also be used.
[0033] Step S2 involves data preprocessing of the collected raw sample dataset, specifically including the following steps:
[0034] S21. Wavelet transform is used to preprocess the original sample dataset.
[0035] Specifically, wavelet transform can be used to extract features or reduce noise in the data, ensuring that the feature information of the data can be effectively preserved in subsequent analysis, while reducing the impact of noise on model performance.
[0036] S22. After preprocessing the data using the wavelet transform method, the processed image dataset is generated.
[0037] Specifically, the preprocessed data is converted into image data that can be processed by a computer. The resulting image dataset can be directly used for subsequent deep learning model training or validation.
[0038] S23. Divide the image dataset into a training dataset and a test dataset according to a predetermined ratio.
[0039] To ensure the scientific and reasonable training and testing process of the model, the generated image dataset is divided into two parts according to a certain ratio. The training dataset accounts for 80% (used for training the deep learning model), and the testing dataset accounts for 20% (used for model performance evaluation). This 1:4 division method can avoid data leakage problems and provide reliable verification of model performance. As mentioned above, the predetermined ratio can be adjusted, meaning other suitable ratios can be selected.
[0040] Step S3: Combine the Ghost convolution module with the spiking neural network model to establish a lightweight spiking neural network model. Integrate the lightweight spiking neural network model with phantom convolution and pulse transmission models, and add a compressed axial attention mechanism.
[0041] The spiking neural network model can be lightweight. The compressed axial attention mechanism is mobile device friendly.
[0042] Step S3 includes the following steps:
[0043] S31: Combine the Ghost convolution module with the spiking neural network model using the Ghost-Spiking module to build a lightweight spiking neural network model.
[0044] This invention proposes a lightweight network building block specifically designed for health assessment of DC-DC converters, effectively improving the computational efficiency and processing power of neural networks in resource-constrained environments. The Ghost-Spiking module combines a Ghost convolutional module with a spiking neural network model, forming a module with low power consumption and high efficiency, making it particularly suitable for embedded devices and edge computing applications.
[0045] The core idea is to use the convolutional module in the Ghost-Spiking module to replace the traditional convolutional layer, so as to significantly reduce the computational burden and energy consumption of the model. Furthermore, by integrating the spiking neuron model in the spiking neural network model, it acquires the dynamic characteristics of biological neural networks, enabling efficient feature extraction and real-time response capabilities with low power consumption.
[0046] First, a standard convolutional layer is used as the initial feature extractor to extract basic low-level features from the input image.
[0047] Secondly, the Ghost-Spiking modules are stacked progressively in multiple stages, and the number of feature channels is increased in each stage.
[0048] Specifically, in the initial stages, lightweight spiking neural network models use smaller convolutional kernels and fewer channels to maintain their lightweight nature. As the lightweight spiking neural network model becomes deeper, the number of feature channels is gradually increased, allowing it to more effectively capture information from complex patterns.
[0049] This design reduces computational complexity while improving the model's learning ability and stability, effectively maintaining a balance between lightweight design and accuracy.
[0050] After stacking all Ghost-Spiking modules, global average pooling is used to map high-dimensional features to low-dimensional feature vectors, and then a fully connected layer is used to complete the classification task.
[0051] S32: Modeling the information transmission and dynamic mechanisms between neurons;
[0052] To achieve more efficient information transmission, a lightweight spiking neural network model was adopted as the basic framework. The information transmission and dynamic mechanisms between neurons were modeled, which can be represented as:
[0053]
[0054] and V rest These represent membrane potential and resting potential, respectively. Indicates the input current. Indicates leakage resistance; when When the value reaches or exceeds the threshold, the neuron emits a pulse, and the membrane potential... It will then be reset to the reset potential.
[0055] Suppose that the postneuron receives spike signals from K preneurons, and the learnable weights of each preneuron are w1, w2, ..., w n , This is the action potential input from the i-th preneuron. The neuronal dynamics mechanism of SNN can be modeled as follows:
[0056]
[0057] Among them, the output pulse The following conditions must be met:
[0058]
[0059] SNNs, by simulating the dynamics of biological neurons, abandon the information transmission methods of traditional neural networks. They can typically achieve low power consumption in resource-constrained environments and have the advantages of efficient processing and real-time response.
[0060] The above SNN neuron dynamics model is as follows: Figure 2 The diagram illustrates a typical information transmission process in a spiking neural network. On the left side of the figure are the input pulse signals emitted by the preneurons. Each signal is transmitted to the postneurons via synaptic connections, each with a corresponding synaptic weight to adjust the influence of the input signal. All weighted inputs are accumulated at the postneurons, generating changes in membrane potential. Once the membrane potential reaches a threshold, an output pulse signal is triggered, thus achieving information transmission and response.
[0061] S33: Integrates a lightweight spiking neural network model with phantom convolution and pulse propagation models;
[0062] To address the resource constraints unique to embedded applications, a lightweight spiking neural network model is organically integrated with phantom convolution and pulse propagation models. The specific implementation method is as follows:
[0063] First, a lightweight phantom convolution module is introduced to handle the pulse sequence features transmitted in spiking neural networks. In spiking neural networks, input data is represented in pulse form, and feature extraction needs to consider both time-series information and computational efficiency. Therefore, the phantom convolution module is modified to adapt to the feature extraction requirements of the pulse transmission model. Specifically, the initial feature map is generated through a small number of standard convolutions, and then additional "ghost" feature maps are generated through a series of simple linear transformation operations (such as pointwise convolution or depthwise convolution). The generation of these "ghost" feature maps is based on the temporal correlation of the pulse sequence input, ensuring that the temporal dimension information is preserved while maintaining low computational cost.
[0064] Secondly, during the fusion process, the feature maps generated by phantom convolution are combined with the dynamic membrane potential accumulation process of the pulse transmission model. The characteristics of spiking neural networks lie in the sparsity and temporal dependence of the input. After fusion, the feature maps generated by phantom convolution are used to jointly model the pulse sequence spatially and temporally. Simultaneously, by dynamically adjusting the synaptic weights in the phantom convolution, it can adapt to pulse inputs at different time steps, further improving the accuracy and efficiency of feature extraction.
[0065] Finally, this design optimizes computational efficiency and energy consumption. On one hand, phantom convolution significantly reduces the computational cost of convolution, lowering the model's complexity; on the other hand, the inherent sparsity of spiking neural networks further reduces unnecessary computations. Combining these two aspects, the model can generate high-quality feature representations with lower resource consumption, thus meeting the needs of embedded environments for efficient, low-power models.
[0066] This fusion design achieves complementary advantages of lightweight spiking neural networks and phantom convolution modules in feature extraction and time series modeling, making the model more practical in resource-constrained embedded applications.
[0067] S34: Added compression-enhanced axial attention mechanism.
[0068] When processing data on mobile devices, traditional attention mechanisms often underperform due to limited computing resources. To address this, a compressed and enhanced axial attention module is introduced in the Ghost-Spiking module. This mechanism applies attention in both the height and width directions, effectively improving feature capture capabilities while maintaining a balance between lightweight design and performance.
[0069] To improve computational efficiency and effectively integrate global information, this invention relates to the aforementioned compressed axial attention mechanism, the specific process of which is as follows.
[0070] First, a feature matrix-based compression operation is designed to compress the input features horizontally and vertically. This method reduces computational complexity while preserving key information of the input features across different dimensions. This is achieved through a specific weight matrix. From raw data Extracting feature vectors , , ,in , , The compression process is achieved using the following formula:
[0071]
[0072] By repeating the above process, the model extracts and saves global features in the horizontal and vertical directions. The compression operations on feature vector q in the horizontal and vertical directions are also repeated on feature vectors k and v, thus ultimately obtaining...
[0073]
[0074] Subsequently, the feature matrix is updated using the following attention mechanism:
[0075]
[0076] This method reduces computational complexity while fully preserving feature relevance. To further enhance the expressive power of compressed axial attention, this invention proposes introducing positional encoding information. By adding learned positional embedding vectors to the compressed axial direction, the model can capture the spatial positional information of axial features. Specifically, the features are enhanced using the following formula:
[0077]
[0078] Here, the position vector *r* is a learning parameter used to enhance the sensitivity of axial attention to positional information. This method introduces position embedding. This allows q(h) and k(h) to know their positions within the extrusion axial features, where these position embeddings are derived from learnable parameters. This is obtained through linear interpolation. Here, L is a constant that defines the dimension of the position embedding. Similarly, we can obtain... Used to process q(v) and k(v).
[0079] To enhance the perception of local details while preserving global information, embodiments of the present invention further introduce a detail enhancement module.
[0080] This detail enhancement module uses multi-scale convolutional units (such as 3×3 convolutional kernels) to extract local features and combines them with a linear activation function for feature mapping. Finally, the features output by the detail enhancement module are fused into the result of the compressed axial attention module, achieving a balance between global and local information.
[0081] like Figure 3 As shown, the proposed compressed axial attention module has a horizontal axial feature compression path on the left side based on a multi-head attention mechanism, where the input features are used to extract feature vectors in the horizontal direction. , , After the vector is processed, separate attention calculations and pointwise convolution operations are performed. The right side shows the convolution path along the vertical axis, using depthwise separable convolution and 1×1 convolution to extract and compress features. Finally, the two paths are fused to enhance the model's ability to focus on spatial axial features and improve the overall feature representation effect.
[0082] This method reduces computational complexity overall while improving the expressive power of the model, and can be widely applied to intelligent devices in resource-constrained environments.
[0083] Step S4: Train the improved lightweight spiking neural network model and verify the model's detection performance.
[0084] Specifically, the divided training dataset is input into the lightweight spiking neural network model for training, and the optimal model is deployed on the edge device. The test dataset is input into the optimal model to conduct a health assessment experiment, obtain the health detection results of the DC-DC converter under different health conditions, and output the assessment results.
[0085] like Figure 4 As shown, the proposed lightweight spiking neural network architecture begins with an input layer. Continuous signals are first transformed into discrete pulse sequences via a pulse convolutional encoder, which then enters the core Ghost-Spiking module. This module consists of phantom convolutional layers, pulse normalization layers, spiking neuron modeling layers, max-pooling layers, and an axial compression-enhanced attention mechanism, used for joint modeling and enhanced extraction of spatial and temporal features. Subsequently, the features are fused and classified through flattening layers and fully connected layers, ultimately outputting a health assessment result.
[0086] The specific process of step S4 is as follows:
[0087] Deploy the trained, optimal model to edge devices to ensure its real-time performance in resource-constrained environments. Verify successful deployment by testing its runtime and resource consumption on edge devices.
[0088] The preprocessed test dataset is input into the optimal model deployed on an edge device to conduct a health assessment experiment. The model classifies the health status of the converter (e.g., normal, slightly degraded, severely degraded) based on the test data.
[0089] The experimental results are compared with those of various algorithms, as shown in the table below:
[0090] Table 1 Comparison of experimental results
[0091]
[0092] As can be seen, without increasing the computational load, the improved model achieved a significant improvement in accuracy, and the number of parameters and computational load were also better than the original model, demonstrating higher overall detection accuracy.
[0093] While some embodiments of the present general inventive concept have been shown and described, those skilled in the art will understand that changes may be made to these embodiments without departing from the principles and spirit of the present general inventive concept, the scope of which is defined by the claims and their equivalents.
Claims
1. A health assessment method for a DC-DC converter based on a spiking neural network, wherein, The health assessment methods include: Step S1. Acquire voltage ripple signals of the DC-DC converter circuit under normal conditions and different health conditions to obtain the original sample dataset; Step S2. Use wavelet transform to preprocess the collected original sample dataset to generate a processed image dataset and divide it into a training dataset and a test dataset; Step S3. Combine the Ghost convolution module with the spiking neural network model to build a lightweight spiking neural network model, fuse the lightweight spiking neural network model with phantom convolution and pulse transmission model, and add a compressed axial attention mechanism; Step S4. Train the lightweight spiking neural network model from step S3 and verify the model's detection performance.
2. The health assessment method according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11. Use an oscilloscope to acquire various voltage ripple signals of the DC-DC converter circuit under normal conditions and different health conditions; S12. Divide the voltage ripple signal into a predetermined number of different health state stages to generate an original sample dataset covering each state.
3. The health assessment method according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21. Wavelet transform method is used to preprocess the original sample dataset; S22. After preprocessing the data using the wavelet transform method, a processed image dataset is generated; S23. Divide the image dataset into a training dataset and a test dataset according to a predetermined ratio.
4. The health assessment method according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. Combine the Ghost convolution module with the spiking neural network model using the Ghost-Spiking module to build a lightweight spiking neural network model; S32. Model the information transmission and dynamic mechanisms between neurons; S33. Integrate the lightweight spiking neural network model with phantom convolution and pulse propagation models; S34. Add a compression-enhanced axial attention mechanism to the Ghost-Spiking module.
5. The health assessment method according to claim 4, characterized in that, Step S31 specifically involves: A standard convolutional layer is used as the initial feature extractor to extract basic low-level features from the input image; The Ghost-Spiking modules are stacked progressively in multiple stages, and the number of feature channels is increased in each stage. After stacking all Ghost-Spiking modules, global average pooling is used to map high-dimensional features to low-dimensional feature vectors, and then a fully connected layer is used to complete the classification task.
6. The health assessment method according to claim 4, characterized in that, Step S32 specifically involves: A lightweight spiking neural network model was adopted as the basic framework to model the information transmission and dynamic mechanisms between neurons, which is represented as follows: and These represent membrane potential and resting potential, respectively. Indicates the input current. Indicates leakage resistance; when When the value reaches or exceeds the threshold, the neuron emits a pulse, and the membrane potential... It will then be reset to the reset potential; Suppose that the postneuron receives spike signals from K preneurons, and the learnable weights of each preneuron are w1, w2, ..., w n , If the action potential input comes from the i-th preneuron, then the neuronal dynamics mechanism of the SNN is modeled as follows: Among them, the output pulse The following conditions must be met: 。 7. The health assessment method according to claim 4, characterized in that, The fusion of the lightweight spiking neural network model with the phantom convolution and pulse propagation models in step S33 includes: First, the input signal is pulse-coded to convert the continuous signal into a discrete pulse sequence, making it suitable for the processing method of the spiking neural network. Then, the feature extraction process of the spiking neural network is optimized by using a phantom convolution module. An initial feature map is generated by a small number of standard convolutions, and an additional "ghost" feature map is generated by a series of linear transformations. A time-step feature update mechanism was designed, enabling the phantom convolution module to adaptively extract key features of the pulse signal according to the dynamic changes of the time series, and retain time information in the feature representation; Finally, the spatial features generated by phantom convolution are deeply fused with the temporal features in the pulse sequence through the membrane potential integration mechanism of the spiking neural network.
8. The health assessment method according to claim 4, characterized in that, Step S34 specifically involves: The input image features are compressed in both horizontal and vertical directions; Through the weight matrix From raw data Extracting feature vectors , , ,in , , The compression process is achieved using the following formula: ； The compression operations in the horizontal and vertical directions of feature vector q are also repeated on feature vectors k and v, thus finally obtaining... ； Subsequently, the feature matrix is updated using the following attention mechanism: 。 9. The health assessment method according to claim 8, characterized in that, By adding a learned position embedding vector along the compressed axis, the model captures the spatial location information of the axial features and enhances the features using the following formula: Where the position vector r is the learning parameter, , L is a constant.
10. The health assessment method according to claim 1, characterized in that, Step S4 specifically involves: The partitioned training dataset is input into the designed lightweight spiking neural network model for training, and the optimal model is deployed on edge devices. Input the test dataset into the trained model, conduct a health assessment experiment, obtain the health detection results of the transformer under different states, and output the assessment results.