A solar insecticidal lamp internet of things fault diagnosis method based on multi-model fusion and dynamic migration learning
Patent Information
- Application Number
- CN202610851737.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-12
- Publication Date
- 2026-10-23
AI Technical Summary
[0005]针对现有技术中存在的不足，本发明提供了一种基于多模型融合与动态迁移学习的太阳能杀虫灯物联网故障诊断方法，解决了跨场景下标签稀缺和领域漂移问题，提升了故障诊断系统的泛化能力、准确性和鲁棒性
[0034]本发明通过DSCNN、BiLSTM和Informer的三分支多模型并行融合架构，能够全面捕捉SIL-IoTs传感数据中的空间、局部时序和长时全局依赖特征，显著提升了故障特征的表征能力，在多种迁移场景下均达到96%以上的诊断准确率。本发明提出的DTM-FlexMatch半监督策略，能够根据类别学习进度动态调整伪标签筛选阈值，有效缓解了工业数据中常见的长尾分布问题，仅需5%-20%的标注数据即可实现高精度诊断，大幅降低了人工标注成本；结合MK-MMD领域自适应技术，有效对齐了源域与目标域的高维特征分布，解决了因设备、环境差异导致的领域漂移问题，使得在一个场景下训练的模型能够可靠地迁移到其他不同场景，显著提升了系统的通用性。本发明的模型设计兼顾了轻量化需求，在树莓派等资源受限的边缘计算设备上验证了其运行效率（内存占用仅1.13 MB），满足实时诊断要求，为大规模、分布式太阳能杀虫灯物联网的智能化运维提供了可靠的技术支撑。
Smart Images
Figure CN122964805A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of IoT intelligent operation and maintenance and fault diagnosis technology, and in particular relates to a fault diagnosis method for solar insecticidal lamp IoT based on multi-model fusion and dynamic transfer learning. It is especially suitable for high-precision and robust fault diagnosis of solar insecticidal lamp IoT nodes under conditions of scarce tags and cross-scenario deployment. Background Technology
[0002] Solar Insecticidal Lamp Internet of Things (SIL-IoTs) are widely deployed in outdoor environments as crucial sensing and execution nodes in modern smart agriculture for pest control and ecological monitoring. However, their long-term exposure to harsh natural environments (such as temperature and humidity fluctuations and electromagnetic interference) makes their hardware components (such as insect-attracting lamps, high-voltage power grids, and solar panels) highly susceptible to various failures, such as open circuits and abnormal current / voltage. These failures can lead to significant deviations in measurement data, meaning that the failure of a single node can affect the measurement integrity and data validity of the entire distributed sensor network, ultimately resulting in inaccurate pest and disease predictions and ineffective agricultural management decisions.
[0003] Existing technologies have proposed several fault diagnosis solutions for such systems. For example, methods based on lightweight one-dimensional convolutional neural networks can detect sensor anomalies; schemes based on binary sliding windows can assist nodes in self-detection. However, most of these methods are limited to a single node or specific hardware configurations, ignoring the distribution differences between data collected from different regions and devices, i.e., the domain drift problem. When a model trained in one environment is directly applied to other nodes, its diagnostic performance drops sharply, severely limiting the system's generalization ability and scalability.
[0004] Transfer learning is considered an effective paradigm for solving the aforementioned domain drift problem; among them, semi-supervised transfer learning and domain adaptation techniques are widely used. Semi-supervised methods extract information from unlabeled data through consistency regularization and pseudo-labeling techniques, but existing methods are prone to generating biased pseudo-labels when dealing with imbalanced data. Domain adaptation methods narrow the domain gap by aligning the feature distributions of different domains, but when used alone, they struggle to capture the complex spatiotemporal features of SIL-IoTs sensing data. Therefore, how to construct a fault diagnosis method that can effectively utilize a small amount of labeled data and a large amount of unlabeled data, overcome cross-domain data distribution differences, and possess both high accuracy and lightweight characteristics is a pressing technical challenge in this field. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a fault diagnosis method for solar-powered insecticidal lamps based on multi-model fusion and dynamic transfer learning in the Internet of Things. This method solves the problems of tag scarcity and domain drift across different scenarios, and improves the generalization ability, accuracy, and robustness of the fault diagnosis system.
[0006] The present invention achieves the above-mentioned technical objectives through the following technical means.
[0007] A fault diagnosis method for solar-powered insecticidal lamps based on multi-model fusion and dynamic transfer learning in the Internet of Things includes the following process:
[0008] Step 1: Collect various feature data by using multiple types of sensors deployed on IoT nodes of solar-powered insecticidal lamps in the source and target domains; for source domain data, label all samples with faults; for target domain data, label a small number of samples or leave them unlabeled.
[0009] Step 2: Construct a three-branch parallel multi-model fusion feature extraction network, which includes a DSCNN branch, a BiLSTM branch, and an Informer branch. The three branches process the input source domain data in parallel and extract multi-scale spatial features, bidirectional temporal dependency features, and long-time series global features, respectively. The feature vectors output by the three branches are concatenated and fused to obtain fused features, which are then input into a classifier composed of fully connected layers.
[0010] Step 3: Build and apply a transfer learning framework based on dynamic thresholding and domain adaptation, which is responsible for implementing cross-domain transfer learning and semi-supervised learning at the feature level. The transfer learning framework includes a DTM-based FlexMatch semi-supervised learning module and an MK-MMD-based domain adaptation module.
[0011] Step 4: Deploy the multi-model fusion feature extraction network and classifier processed in Step 3 onto the edge computing device of the solar insecticidal lamp IoT node. This device collects sensor data in real time, performs fault diagnosis, and sends the diagnosis results to the user terminal for alarm via wireless network.
[0012] Furthermore, in step 2, the input to the DSCNN branch is time-series data. , Represents the real number field. Indicates the number of input feature channels. To represent the sequence length, first, channel-wise convolution is performed, using an independent k×1 convolution kernel to extract spatial features for each input channel. Then, pointwise convolution is performed, using a 1×1 convolution kernel to fuse information between channels, finally outputting a feature vector. ; This indicates the dimension of the feature vector output by the DSCNN branch. This indicates the length of the feature vector output by the DSCNN branch.
[0013] Furthermore, in step 2, the BiLSTM branch input sequence , , , These represent the feature vectors of the input sequence at each time step, respectively, and the forward hidden state is calculated through a forward LSTM layer. The backward hidden state is calculated through the backward LSTM layer. The hidden state at each time step is ultimately: The hidden state at the last time step is taken as the output feature vector. , This represents the dimension of the feature vector output by the BiLSTM branch. This represents the feature length after the bidirectional hidden layers are concatenated, i.e., the length of the feature vector output by the BiLSTM branch.
[0014] Furthermore, in step 2, the Informer branch input sequence After positional encoding, the data is fed into the encoder. The encoder's probabilistic sparse self-attention mechanism evaluates the sparsity of the query matrix, filtering out important queries and reducing computational complexity. After passing through multiple attention layers and a feedforward network, adaptive average pooling is used to obtain the output feature vector. , This indicates the dimension of the feature vector output by the Informer branch. This indicates the length of the feature vector output by the Informer branch.
[0015] Furthermore, in step 2, the feature vectors output from the three branches are concatenated and fused to obtain the fused feature. : ; This represents a vector concatenation operation. The feature vector output by the DSCNN branch. The feature vector output by the BiLSTM branch. The feature vector output by the Informer branch.
[0016] Further, step 3 includes:
[0017] Step 3.1: Load the multi-model fusion feature extraction network and its parameters that have been trained using labeled samples from the source domain;
[0018] Step 3.2: Model warm-up: Train the initial classifier using the fusion features of labeled samples from the source domain and their true labels. The initial classifier is the classifier composed of fully connected layers described in Step 2, which enables it to acquire basic fault identification capabilities.
[0019] Step 3.3: Model fine-tuning: Use the network parameters of the multi-model fusion feature extraction network trained with labeled samples in the source domain to initialize the unlabeled or poorly labeled data in the target domain, and perform collaborative training through a semi-supervised learning module and a domain adaptation module to align the features of the source domain with the features of the target domain.
[0020] Step 3.4: Combine supervised loss, semi-supervised loss, and MK-MMD loss to calculate the total loss function, and perform end-to-end joint optimization training; total loss function for:
[0021]
[0022] in, It is the supervised cross-entropy loss of labeled data in the source domain. It is the unsupervised cross-entropy loss of the target domain pseudo-label data. It is a balancing weight. It depends on the training round. Decreasing weighting coefficients; This is the loss of MK-MMD distance;
[0023] Step 3.5: Input the fused features output by the multi-model fusion feature extraction network after joint optimization training into the classifier. The classifier uses a fully connected layer to perform linear mapping and non-linear activation on the fused features, and makes predictions through the Softmax function and saves the model.
[0024] Furthermore, step 3 is specifically as follows:
[0025] Step 3.3.1: In the semi-supervised learning module, maintain a counting matrix. Before recording Categories in round iteration The number of samples selected as pseudo-labels; calculated for each category using the following formula. In the current round Learning status:
[0026]
[0027] Then calculate the dynamic filtering threshold. :
[0028]
[0029] in, It is a preset global threshold. It is a mapping function. Indicates the total number of categories; only when the sample matches the category The predicted probability exceeds Only then will the sample and its pseudo-label be included in the training set;
[0030] Step 3.3.2: In the domain adaptation module, the source domain feature set... and target domain feature set Through multi-core mapping Projected onto the regenerating nucleus Hilbert space Calculate the MK-MMD distance:
[0031]
[0032] in, and These are the feature vectors of the source domain and the target domain, respectively. and This represents the number of samples.
[0033] The present invention has the following beneficial effects:
[0034] This invention employs a three-branch, multi-model parallel fusion architecture combining DSCNN, BiLSTM, and Informer to comprehensively capture spatial, local temporal, and long-term global dependency features in SIL-IoTs sensor data, significantly improving the representation capability of fault features and achieving diagnostic accuracy exceeding 96% in various transfer scenarios. The proposed DTM-FlexMatch semi-supervised strategy dynamically adjusts the pseudo-label selection threshold based on the class learning progress, effectively mitigating the long-tail distribution problem common in industrial data. High-precision diagnosis requires only 5%-20% of labeled data, significantly reducing manual annotation costs. Combined with MK-MMD domain adaptation technology, it effectively aligns the high-dimensional feature distributions of the source and target domains, solving the domain drift problem caused by differences in equipment and environment. This allows models trained in one scenario to reliably transfer to other different scenarios, significantly improving the system's versatility. The model design of this invention takes into account the need for lightweight design and has verified its operating efficiency (memory usage of only 1.13 MB) on resource-constrained edge computing devices such as Raspberry Pi. It meets the requirements for real-time diagnosis and provides reliable technical support for the intelligent operation and maintenance of large-scale, distributed solar insecticidal lamp IoT. Attached Figure Description
[0035] Figure 1 Flowchart of a fault diagnosis method for solar-powered insecticidal lamps based on multi-model fusion and dynamic transfer learning in the Internet of Things.
[0036] Figure 2This is a schematic diagram of a multi-model fusion feature extraction network structure;
[0037] Figure 3 Flowchart of dynamic threshold mechanism and semi-supervised learning;
[0038] Figure 4 This is a schematic diagram of a domain adaptive module based on the maximum mean difference of multiple cores.
[0039] In the diagram: Batch represents the batch size of samples, which helps prevent training overload caused by excessively large parameters; , , , This represents the feature vector of the input sequence at each time step; This is the weight matrix of the first layer of the feedforward LSTM. This is the weight matrix of the first layer of the backward LSTM. Combining the two can effectively extract the bidirectional dynamic features of the original time series data. This is the weight matrix of the second layer of the feedforward LSTM. This is the weight matrix of the second-layer backward LSTM. The combination of the two can effectively extract the dynamic features of time series data. , , , , , , , The output features of BiLSTM are represented by: Linear layer, which performs feature fusion, extraction, and mapping; Dropout is the neuron dropout rate, which can prevent overfitting and improve generalization ability; ReLU is the activation function; and Softmax function is used to predict the output probability. , , To improve the model's global feature extraction capabilities, the Backbone is a three-branch model, i.e., a multi-model fusion feature extraction network; the Bottleneck represents the features output by the model. For classification loss, calculate the output probability. Detailed Implementation
[0040] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited thereto.
[0041] Reference Figure 1 As shown, the fault diagnosis method for solar-powered insecticidal lamps based on multi-model fusion and dynamic transfer learning according to the present invention includes the following process:
[0042] Step 1: Data Acquisition and Preprocessing;
[0043] By deploying various types of sensors on IoT nodes of solar-powered insecticidal lamps in different regions (divided into source and target domains), a variety of characteristic data are collected, including but not limited to light intensity, solar panel voltage, solar panel current, insect-attracting lamp current, insect-attracting lamp voltage, metal mesh current, metal mesh voltage, air temperature, humidity, etc. The collected data is transmitted to cloud servers or local servers for storage and processing via 4G / 5G networks.
[0044] For the source domain data, all samples are labeled with faults. The fault label space contains 9 categories: F0 (no fault), F1 (open circuit of insect-attracting lamp), F2 (open circuit of metal mesh), F3 (open circuit of solar panel), F4 (mismatch between light intensity and solar panel current), F5 (abnormal current of metal mesh), F6 (abnormal voltage of metal mesh), F7 (abnormal current of insect-attracting lamp), and F8 (abnormal voltage of insect-attracting lamp), forming a labeled dataset.
[0045] For the target domain data, perform minimal or no labeling.
[0046] To increase the number of samples, an overlap sampling technique is used, which divides the original long-term series data into multiple samples using a sliding window. The labeled dataset (i.e., the source domain dataset) is divided into training, validation, and test sets according to a certain ratio.
[0047] Step 2: Construct a three-branch parallel multi-model fusion feature extraction network (DSCNN-BiLSTM-Informer) for feature extraction;
[0048] like Figure 2 As shown, the multi-model fusion feature extraction network includes a Deep Separable Convolutional Neural Network (DSCNN) branch, a Bidirectional Long Short-Term Memory (BiLSTM) branch, and an Informer branch. The three branches process the input data in parallel, extracting multi-scale spatial features, bidirectional temporal dependency features, and long-term global features, respectively. Finally, the feature vectors output by the three branches are concatenated and fused to obtain a joint feature representation.
[0049] It should be noted that the input data for the multi-model fusion feature extraction network in this step comes from the labeled samples in the source domain (including training, validation, and test sets) from step 1. During model training, the network can simultaneously receive labeled samples from the source domain (including training, validation, and test sets) and unlabeled (or a small number of labeled) samples from the target domain; during inference, it receives samples collected in real time from the target domain. The inputs to the three parallel branches (DSCNN, BiLSTM, and Informer) are all the same original time-series data, meaning each branch receives the same input sequence. Features are extracted separately using different structures.
[0050] The DSCNN branch employs depthwise separable convolution, decomposing standard convolution into channel-wise convolution and pointwise convolution; the input to the DSCNN branch is temporal data. , Represents the real number field. Indicates the number of input feature channels. To represent the sequence length, first, channel-wise convolution is performed, using an independent k×1 convolution kernel to extract spatial features for each input channel. Then, pointwise convolution is performed, using a 1×1 convolution kernel to fuse information between channels, finally outputting a feature vector. ;in, This indicates the dimension of the feature vector output by the DSCNN branch. This indicates the length of the feature vector output by the DSCNN branch;
[0051] The BiLSTM branch extracts bidirectional contextual information of the sequence through two hidden layers, forward and backward; the BiLSTM branch input sequence , , , These represent the feature vectors of the input sequence at each time step, respectively, and the forward hidden state is calculated through a forward LSTM layer. The backward hidden state is calculated through the backward LSTM layer. The hidden state at each time step is ultimately: The hidden state at the last time step is taken as the output feature vector. ,in, This represents the dimension of the feature vector output by the BiLSTM branch. This represents the feature length after concatenation of the bidirectional hidden layers, i.e., the length of the feature vector output by the BiLSTM branch.
[0052] The Informer branch reduces computational complexity through a probabilistic sparse attention mechanism; the Informer branch input sequence After positional encoding, the data is fed into the encoder. The encoder's probabilistic sparse self-attention mechanism evaluates the sparsity of the query matrix, filtering out important queries and reducing computational complexity from... Reduce to After passing through multiple attention layers and a feedforward network, the output feature vector is obtained through adaptive average pooling. ,in, This indicates the dimension of the feature vector output by the Informer branch. This indicates the length of the feature vector output by the Informer branch.
[0053] Then, the feature vectors output from the three branches are concatenated and fused to obtain the fused feature. : ;in, The feature vector output by the DSCNN branch. The feature vector output by the BiLSTM branch. The feature vector output by the Informer branch. This indicates a vector concatenation operation.
[0054] Finally, the fusion features will be used. The input is fed into a classifier consisting of fully connected layers.
[0055] Step 3: Construct and apply a transfer learning framework based on dynamic thresholding and domain adaptation, responsible for implementing cross-domain transfer learning and semi-supervised learning at the feature level; such as Figure 3 , 4 As shown, the transfer learning framework includes a FlexMatch semi-supervised learning module based on the Dynamic Thresholding Mechanism (DTM) and a domain adaptation module based on the Multi-kernel Maximum Mean Discrepancy (MK-MMD).
[0056] The specific processing procedure is as follows:
[0057] Step 3.1: Load the source domain model and weights, that is, load the multi-model fusion feature extraction network and network parameters that have been trained using labeled samples from the source domain;
[0058] Step 3.2: Model warm-up: Train the initial classifier using the fusion features of labeled samples from the source domain and their true labels to train the initial classifier (i.e., the classifier composed of fully connected layers as described in Step 2) to obtain basic fault identification capabilities.
[0059] Step 3.3: Model fine-tuning: Use the network parameters of the multi-model fusion feature extraction network trained with labeled samples in the source domain to initialize the unlabeled (or poorly labeled) data in the target domain, and perform collaborative training through semi-supervised and domain adaptation modules (i.e., semi-supervised learning module and domain adaptation module) to align the features of the source domain with the features of the target domain.
[0060] Step 3.3.1: In the semi-supervised learning module, maintain a counting matrix. Before recording Categories in round iteration The number of samples selected as pseudo-labels; calculated for each category using the following formula. In the current round Learning status:
[0061]
[0062] Then calculate the dynamic filtering threshold. :
[0063]
[0064] in, It is a preset global threshold. It is a mapping function. Indicates the total number of categories. Only when the sample matches the category... The predicted probability exceeds Only when the specified time is reached will the sample and its pseudo-label be included in the training set. This mechanism effectively mitigates the pseudo-label bias problem under long-tailed distributions.
[0065] Step 3.3.2: In the domain adaptation module, the source domain feature set... and target domain feature set Through multi-core mapping Projected onto the regenerating nucleus Hilbert space Calculate the MK-MMD distance:
[0066]
[0067] in, and These are the feature vectors of the source domain and the target domain, respectively. and The distance represents the number of samples. This distance quantifies the difference between the feature distributions of the two domains; minimizing this distance forces the model to learn domain-independent feature representations, thus overcoming the domain drift problem.
[0068] The core function of the MK-MMD distance is to quantify the degree of difference between the feature distributions of the source and target domains. The smaller the distance, the closer the features of the two domains are statistically, meaning that the features learned by the model are less sensitive to changes in the target domain. By incorporating this distance as a loss term into the overall optimization objective, the model can be guided to actively reduce domain differences during training, thereby enabling the classifier trained on the source domain to reliably generalize to the target domain.
[0069] Step 3.4: Combine supervised loss, semi-supervised loss and MK-MMD loss to calculate the total loss function and perform end-to-end joint optimization training.
[0070] The total loss function for:
[0071]
[0072] in, It is the supervised cross-entropy loss of labeled data in the source domain. It is the unsupervised cross-entropy loss of the target domain pseudo-label data. It is a balancing weight. It depends on the training round. Decreasing weighting coefficients. The loss is the MK-MMD distance defined in step 3. By minimizing this distance in each round of training, the model gradually eliminates the feature distribution shift caused by device differences and environmental changes, thereby improving the robustness of cross-scenario fault diagnosis.
[0073] Step 3.5: Input the fused features output by the multi-model fusion feature extraction network after joint optimization training into the classifier. The classifier uses a fully connected layer to perform linear mapping and non-linear activation on the fused features, and makes predictions through the Softmax function and saves the model.
[0074] Step 4: Edge Deployment and Real-time Diagnostics;
[0075] The trained model is quantized and converted to a new format, then deployed on an edge computing device within the IoT node of the solar-powered insecticidal lamp. This device collects sensor data in real time, runs the model to perform fault diagnosis, and sends the diagnostic results to the user terminal via a wireless network for alarm purposes.
[0076] The embodiments described above are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention shall fall within the protection scope of the present invention.
Claims
1. A fault diagnosis method for solar-powered insecticidal lamps based on multi-model fusion and dynamic transfer learning in the Internet of Things (IoT), characterized in that, The process includes the following: Step 1: Collect various feature data by using multiple types of sensors deployed on IoT nodes of solar-powered insecticidal lamps in the source and target domains; for source domain data, label all samples with faults; for target domain data, label a small number of samples or leave them unlabeled. Step 2: Construct a three-branch parallel multi-model fusion feature extraction network, which includes a DSCNN branch, a BiLSTM branch, and an Informer branch. The three branches process the input source domain data in parallel and extract multi-scale spatial features, bidirectional temporal dependency features, and long-time series global features, respectively. The feature vectors output by the three branches are concatenated and fused to obtain fused features, which are then input into a classifier composed of fully connected layers. Step 3: Build and apply a transfer learning framework based on dynamic thresholding and domain adaptation, which is responsible for implementing cross-domain transfer learning and semi-supervised learning at the feature level. The transfer learning framework includes a DTM-based FlexMatch semi-supervised learning module and an MK-MMD-based domain adaptation module. Step 4: Deploy the multi-model fusion feature extraction network and classifier processed in Step 3 onto the edge computing device of the solar insecticidal lamp IoT node. This device collects sensor data in real time, performs fault diagnosis, and sends the diagnosis results to the user terminal for alarm via wireless network.
2. The method for fault diagnosis of solar-powered insecticidal lamps based on multi-model fusion and dynamic transfer learning according to claim 1, characterized in that, In step 2, the input to the DSCNN branch is time-series data. , Represents the real number field. Indicates the number of input feature channels. To represent the sequence length, first, channel-wise convolution is performed, using an independent k×1 convolution kernel to extract spatial features for each input channel. Then, pointwise convolution is performed, using a 1×1 convolution kernel to fuse information between channels, finally outputting a feature vector. ; This indicates the dimension of the feature vector output by the DSCNN branch. This indicates the length of the feature vector output by the DSCNN branch.
3. The method for fault diagnosis of solar-powered insecticidal lamps based on multi-model fusion and dynamic transfer learning according to claim 1, characterized in that, In step 2, the BiLSTM branch input sequence , , , These represent the feature vectors of the input sequence at each time step, respectively, and the forward hidden state is calculated through a forward LSTM layer. The backward hidden state is calculated through the backward LSTM layer. The hidden state at each time step is ultimately: The hidden state at the last time step is taken as the output feature vector. , This represents the dimension of the feature vector output by the BiLSTM branch. This represents the feature length after the bidirectional hidden layers are concatenated, i.e., the length of the feature vector output by the BiLSTM branch.
4. The method for fault diagnosis of solar-powered insecticidal lamps based on multi-model fusion and dynamic transfer learning according to claim 1, characterized in that, In step 2, the Informer branch input sequence After positional encoding, the data is fed into the encoder. The encoder's probabilistic sparse self-attention mechanism evaluates the sparsity of the query matrix, filtering out important queries and reducing computational complexity. After passing through multiple attention layers and a feedforward network, adaptive average pooling is used to obtain the output feature vector. , This represents the dimension of the feature vector output by the Informer branch. This indicates the length of the feature vector output by the Informer branch.
5. The method for fault diagnosis of solar-powered insecticidal lamps based on multi-model fusion and dynamic transfer learning according to claim 1, characterized in that, In step 2, the feature vectors output from the three branches are concatenated and fused to obtain the fused feature vectors. : ; This represents a vector concatenation operation. The feature vector output by the DSCNN branch. The feature vector output by the BiLSTM branch. The feature vector output by the Informer branch.
6. The method for fault diagnosis of solar-powered insecticidal lamps based on multi-model fusion and dynamic transfer learning according to claim 1, characterized in that, Step 3 includes: Step 3.1: Load the multi-model fusion feature extraction network and its parameters that have been trained using labeled samples from the source domain; Step 3.2: Model warm-up: Train the initial classifier using the fusion features of labeled samples from the source domain and their true labels. The initial classifier is the classifier composed of fully connected layers described in Step 2, which enables it to acquire basic fault identification capabilities. Step 3.3: Model fine-tuning: Use the network parameters of the multi-model fusion feature extraction network trained with labeled samples in the source domain to initialize the unlabeled or poorly labeled data in the target domain, and perform collaborative training through a semi-supervised learning module and a domain adaptation module to align the features of the source domain with the features of the target domain. Step 3.4: Combine supervised loss, semi-supervised loss, and MK-MMD loss to calculate the total loss function, and perform end-to-end joint optimization training; total loss function for: in, It is the supervised cross-entropy loss of labeled data in the source domain. It is the unsupervised cross-entropy loss of the target domain pseudo-label data. It is a balancing weight. It depends on the training round. Decreasing weighting coefficients; This is the loss of MK-MMD distance; Step 3.5: Input the fused features output by the multi-model fusion feature extraction network after joint optimization training into the classifier. The classifier uses a fully connected layer to perform linear mapping and non-linear activation on the fused features, and makes predictions through the Softmax function and saves the model.
7. The method for fault diagnosis of solar-powered insecticidal lamps based on multi-model fusion and dynamic transfer learning according to claim 3, characterized in that, Step 3 is as follows: Step 3.3.1: In the semi-supervised learning module, maintain a counting matrix. Before recording Categories in round iteration The number of samples selected as pseudo-labels; Calculate each category using the following formula In the current round Learning status: Then calculate the dynamic filtering threshold. : in, It is a preset global threshold. It is a mapping function. Indicates the total number of categories; only when the sample matches the category The predicted probability exceeds Only then will the sample and its pseudo-label be included in the training set; Step 3.3.2: In the domain adaptation module, the source domain feature set... and target domain feature set Through multi-core mapping Projected onto the regenerating nucleus Hilbert space Calculate the MK-MMD distance: in, and These are the feature vectors of the source domain and the target domain, respectively. and This represents the number of samples.