Charging pile fault diagnosis method based on fusion of gradient boosting tree and multi-layer perceptron

Abstract

The present invention discloses a charging pile fault diagnosis method based on the fusion of gradient boosting tree and multi-layer perceptron, and relates to the technical field of charging pile fault diagnosis. In order to solve the common problem of feature extraction in charging pile fault diagnosis and the problem that complex artificial intelligence models are prone to overfitting, the present invention proposes a new method based on the fusion of gradient boosting tree GBDT and multi-layer perceptron MLP. According to the theory of integrated learning, multiple independent gradient boosting tree GBDT models are established, multiple fully connected GBDT layers are formed, and finally a multi-layer perceptron MLP model is connected to perform feature learning and classification. This fusion scheme avoids the difficulty of manual feature extraction, and while reducing the dependence on the performance and complexity of a single model, its fusion performance is improved and more stable. Experimental results on public data sets show that the scheme proposed by the present invention is superior to typical independent machine learning schemes.

Description

Technical Field

[0001] The present invention relates to the technical field of charging pile fault diagnosis, and in particular to a charging pile fault diagnosis method based on the fusion of a gradient boosting tree and a multi-layer perceptron. Background Art

[0002] The world today faces increasingly severe energy and environmental crises. The development of electric vehicles not only effectively alleviates the non-renewable nature of traditional fossil fuels but also serves as an effective means of environmental governance. Charging stations, as energy recharge sources, are crucial supporting infrastructure for electric vehicles, and charging piles are core equipment for both charging stations and charging service networks. my country's electric vehicle market is projected to reach 15 million vehicles by 2030, significantly increasing demand for charging piles. Charging piles are exposed to the elements for extended periods, causing component degradation and ultimately leading to failure, significantly inconvenient for electric vehicle charging. Intelligent detection of charging pile failures is crucial for timely identification of faulty charging piles, enabling maintenance and maintaining reliable operation.

[0003] Huang Caijuan's "Fault Diagnosis and Troubleshooting of Electric Vehicle AC Charging Piles" and Li Miao's "Fault Diagnosis and Detection of Electric Vehicle Charging Piles" discussed common fault types and solutions for charging piles;

[0004] Lin Yue et al.'s "Charging Pile Fault Diagnosis Based on AP-HMM Hybrid Model" collects voltage data from charging piles, combines the deterministic similarity propagation (AP) clustering method to quickly and accurately extract fault features, and the powerful fault classification ability of the statistical hidden Markov model (HMM), to construct an AP-HMM hybrid charging pile fault diagnosis model, realizing the diagnosis of charging pile relay failure. Lin Yue et al.'s "Charging Pile Fault Diagnosis Research Based on Hidden Birth and Death Process Model" proposes to establish a hidden birth and death process model for charging pile relay failure to achieve fault diagnosis.

[0005] Li Yifei et al.'s "Application of Supervised Machine Learning Algorithms in Charging Facility Status Assessment" analyzes the causes of charging pile failures and uses remote signaling data, charging capacity, charging time, average power, and other information as input to construct a decision tree diagnostic model that can identify five charging pile health states, ranging from normal to abnormal.

[0006] Cai Songhuan's "Research on Charging Pile Data Mining and Charging Capacity Prediction Algorithms" attempts to implement charging pile fault diagnosis using a variety of tree-based classifiers, such as random forest and XGBoost;

[0007] Zhao Xiang et al.'s "A DC Charging Pile Fault Diagnosis Method Based on Deep Neural Network" proposed a deep learning-based fault diagnosis scheme using the operating voltage data of the charging pile as input, achieving automatic diagnosis of DC charging pile fault conditions;

[0008] Wang Qunfei et al.'s "Research on Fault Diagnosis of V2G Charging Pile Based on Random Forest Algorithm" uses wavelet packet analysis to extract characteristic information from fault current signals and uses random forest to construct a fault diagnostic device suitable for DC charging pile switch modules;

[0009] Zhou Rui et al.'s "Design of the Operating Status and Fault Diagnosis System of Automobile Charging Pile" and Gao Dexin et al.'s "Design of Mobile Monitoring and Fault Diagnosis System for Electric Vehicle Charging Pile" proposed software and hardware design and development solutions for charging pile fault diagnosis systems.

[0010] To describe the problem of charging pile fault detection, the present invention uses the vector represents a set of physical quantities collected from the charging pile, such as the total harmonic distortion of voltage, total harmonic distortion of current, electronic lock drive signal, etc. The task is to classify the charging piles into two categories based on the vector x: normal and faulty. The category labels can be represented by the variable y∈{0,1}. Therefore, the charging pile fault diagnosis model can be expressed as

[0011] y=Det(x) (1) The mapping Det(·) is unknown and will be obtained in a data-driven way.

[0012] In order to construct the mapping Det(·), a set of charging pile data can usually be measured Based on this, the general machine learning processing mode can be used for this classification problem, that is, feature extraction and classifier formation. However, each input x i The data may be from different charging piles, so it is difficult to establish statistics to observe the changes in the status of the same charging pile at different times. At the same time, the data difference between normal charging piles and faulty charging piles is very small, and the statistical distribution of their data overlaps seriously, making it difficult to distinguish. Moreover, the input quantity x i The number of rows L is usually small, making it difficult to extract features through dimensionality reduction. In addition, the mapping Det(·) is nonlinear. These aspects significantly increase the difficulty of the problem, making the performance of traditional solutions unsatisfactory.

[0013] The performance of traditional machine learning solutions relies heavily on feature extraction. Manual feature extraction has always been a challenge, especially for charging pile fault diagnosis. To circumvent this problem, designing deep neural networks can be considered. Neural networks can automatically extract features. By stacking neural network layers with nonlinear processors, deep feature representations can be learned. This allows for joint optimization of feature extraction and classifiers within an end-to-end framework. Neural networks can be considered as fully automatic optimization tools, achieving good performance on large datasets. However, due to their numerous parameters, neural networks are difficult to optimize and are prone to local optima and overfitting. Furthermore, the lack of rotational invariance for tabular data significantly degrades the performance of the neural network. This is particularly problematic for charging pile fault diagnosis.

[0014] For tabular data, tree-based classifiers offer advantages over deep neural networks and can also be used to build nonlinear models. However, tree-based classifiers generally require manual feature extraction, and as the tree depth increases, they are prone to overfitting and computational complexity increases.

[0015] Based on the above analysis, due to the characteristics of charging pile data, manual feature extraction for fault diagnosis is relatively difficult. Complex artificial intelligence models are prone to overfitting. Fully automated solutions, such as neural networks, are prone to local optimality and difficult to train. Therefore, research is necessary to address the common feature extraction challenges in charging pile fault diagnosis and the overfitting problem that complex artificial intelligence models are prone to. Summary of the Invention

[0016] In view of the defects of the existing technology, the present invention proposes a charging pile fault diagnosis method based on the fusion of gradient boosting tree and multi-layer perceptron. The method collects the charging pile data set to be tested, constructs a feature vector, and inputs the feature vector into the fusion model to output the charging pile fault diagnosis result.

[0017] The fusion model includes N layers of learners, where N≥2,

[0018] The first N-1 layers of learners are composed of multiple independent gradient boosting tree GBDT models. The output of the previous layer of learners is used as the input of the next layer of learners. The gradient boosting tree GBDT model in the first layer of learners takes the feature vector of the charging pile as input and outputs the charging pile fault prediction. The gradient boosting tree GBDT model in the next layer integrates the output of the gradient boosting tree GBDT model in the previous layer to output a new charging pile fault prediction.

[0019] The Nth layer learner is a multi-layer perceptron (MLP) model with two hidden layers, which is used to receive the output of the gradient boosting tree (GBDT) model in the previous layer learner and output the charging pile fault diagnosis results.

[0020] Furthermore, the charging pile data set includes the K1K2 drive signal of the charging pile, the electronic lock drive signal, the charging pile emergency stop signal, the charging pile access control signal, and the total harmonic distortion of the charging pile voltage and current.

[0021] Furthermore, the fusion model can be composed of three layers of learners, the first layer of learners includes 5 gradient boosting tree GBDT models, the second layer of learners includes 3 gradient boosting tree GBDT models, and the third layer of learners includes a multi-layer perceptron MLP model with two hidden layers.

[0022] Furthermore, in the fusion model, the feature vector is sampled multiple times to provide different input data for each gradient boosting tree GBDT model in the first layer learner, and the charging pile fault prediction values ​​output by all gradient boosting tree GBDT models in the previous layer learner are stacked to provide different input data for each gradient boosting tree GBDT model in the next layer learner.

[0023] Furthermore, the hyperparameter settings of each gradient boosting tree GBDT model in the same layer of learners are different.

[0024] Furthermore, each gradient boosting tree (GBDT) model in the fusion model is trained using a Softmax objective function and an exact greedy algorithm.

[0025] Furthermore, the number of neurons in the input layer of the multilayer perceptron MLP model in the N-th layer learner is equal to the number of gradient boosting tree GBDT models in the N-1-th layer learner, and one neuron in the input layer of the multilayer perceptron MLP model receives a charging pile fault prediction value output by a gradient boosting tree GBDT model in the N-1-th layer learner.

[0026] Furthermore, the output layer of the multi-layer perceptron MLP model in the N-th layer learner contains two neurons, corresponding to the normal and fault states of the charging pile respectively.

[0027] Furthermore, a Softmax layer is connected after the multi-layer perceptron (MLP) model in the Nth layer learner to output the confidence score of the fault.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] First, based on ensemble learning theory, we establish multiple independent gradient boosting tree (GBDT) models, build multiple fully connected GBDT layers, and finally connect them to a multi-layer perceptron (MLP) model for feature learning and classification. This fusion solution avoids the difficulty of manual feature extraction, while reducing dependence on the performance and complexity of individual models. It also improves and stabilizes fusion performance, resulting in more accurate prediction results.

[0030] Second, the performance of each base learner involved in the fusion is high, and each algorithm meets the diversity requirements. Multi-source information fusion can give full play to the attention information of different models on features and make full use of feature information to improve fusion.

[0031] Thirdly, the present invention reduces the complexity of a single model by setting a reasonable sampling data base, thus greatly saving computing resources;

[0032] Fourthly, compared with traditional charging pile fault diagnosis methods, the gradient boosting tree GBDT and multi-layer perceptron MLP fusion model adopted in the present invention can be applied to diagnosis and evaluation in other power electronics fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 It is a diagnostic model that combines the gradient boosting tree GBDT and the multi-layer perceptron MLP. DETAILED DESCRIPTION

[0035] The technical solutions of the present invention will be described clearly and completely below with reference to the accompanying drawings. It is apparent that the embodiments described are only a portion of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0036] Those skilled in the art should be aware that the following specific embodiments or implementations are a series of optimized configurations listed in the present invention to further explain the specific content of the invention, and these configurations can be combined or used in conjunction with each other, unless the present invention explicitly states that some or a specific embodiment or implementation cannot be combined or used in conjunction with other embodiments or implementations. At the same time, the following specific embodiments or implementations are only intended to be optimized configurations and are not to be understood as limiting the scope of protection of the present invention.

[0037] The specific embodiments of the present invention are described below with reference to the accompanying drawings (tables).

[0038] Considering the difficulties of traditional machine learning solutions in solving the problem of charging pile status detection, the present invention proposes a fusion classification model of gradient boosting tree and multi-layer perceptron.

[0039] The present invention proposes a new method that integrates gradient boosting decision tree (GBDT) and multi-layer perceptron (MLP) to solve the problem of charging pile fault diagnosis. The scheme is based on ensemble learning theory, and multiple fully connected GBDT layers are formed and connected with MLP to perform feature learning and classification. The fusion of gradient boosting tree GBDT and multi-layer perceptron MLP can effectively avoid their shortcomings, combine their advantages, and improve the accuracy of charging pile fault diagnosis. Moreover, the scheme avoids the difficulty of manual feature extraction, reduces the probability of network overfitting, and ensures that the fault diagnosis performance is improved and more stable. Experimental results on public data sets show that the scheme proposed by the present invention is superior to typical independent machine learning schemes.

[0040] Figure 1 It is a diagnostic model that combines the gradient boosting tree GBDT and the multi-layer perceptron MLP.

[0041] The present invention proposes a charging pile fault diagnosis method based on the fusion of gradient boosting tree and multi-layer perceptron. Specifically, the method comprises the following steps: collecting a data set of the charging pile to be tested, constructing a feature vector, inputting the feature vector into a fusion model, and outputting the charging pile fault diagnosis result.

[0042] The fusion model includes N layers of learners, where N≥2,

[0043] The first N-1 layers of learners are composed of multiple independent gradient boosting tree GBDT models. The output of the previous layer of learners is used as the input of the next layer of learners. The gradient boosting tree GBDT model in the first layer of learners takes the feature vector of the charging pile as input and outputs the charging pile fault prediction. The gradient boosting tree GBDT model in the next layer integrates the output of the gradient boosting tree GBDT model in the previous layer to output a new charging pile fault prediction.

[0044] The Nth layer learner is a multi-layer perceptron (MLP) model with two hidden layers, which is used to receive the output of the gradient boosting tree (GBDT) model in the previous layer learner and output the charging pile fault diagnosis results.

[0045] The gradient boosting tree GBDT model is a polynomial regression or classification model. In order to achieve model fitting, multiple classification and regression trees (CART) are established, and the parameters of each CART are optimized by minimizing the loss function. The advantage of the gradient boosting tree GBDT model is that it can recognize nonlinearly separable data and flexibly process complex tabular data. The multi-layer perceptron MLP model is a fully connected neural network. It organizes neurons in layers, and the network layers are connected in a one-way, non-loop, fully connected manner. The network weight parameters are optimized by minimizing the loss function. It has strong feature learning capabilities and can be combined with support vector machines (SVM) to achieve classification, but it also has the common problems of neural networks. The present invention combines the two under the framework of integrated learning, which can effectively avoid the shortcomings of the two and combine the advantages of the two to better realize the fault diagnosis of charging piles.

[0046] Specifically, based on the stacking framework and theory of ensemble learning, multiple independent gradient boosting tree (GBDT) models were established. GBDT models that received the same input were then placed on the same layer, forming multiple GBDT layers. The output of the previous GBDT layer served as the input for the next GBDT layer, similar to a fully connected neural network. The first GBDT layer used the charging pile data vector as input to predict charging pile failures. The GBDTs in the subsequent layers obtained new predictions by integrating the outputs of the previous GBDT layer. Following these GBDT layers, a multi-layer perceptron (MLP) model was added to receive the outputs of the previous GBDT layer for feature learning and classification.

[0047] Ensemble learning theory dictates that to achieve a superior fusion model, each base learner should be distinct, providing independent classification errors and exhibiting superior performance. To achieve diverse base learners, the hyperparameters of the same GBDT layer are set differently, and bootstrap sampling is used to generate a different training set for each GBDT model. The first layer of GBDT models is designed to be more complex than the second layer, ensuring high classification performance while minimizing the risk of overfitting.

[0048] The final GBDT layer is followed by an MLP. The MLP, which operates on a completely different principle from the GBDT, ensures diversity in the base learners. Furthermore, the MLP's fusion capabilities effectively integrate the outputs of the GBDT layer, driving improved performance for the entire model. The number of hidden layers and neurons in each hidden layer of the MLP can be adjusted based on model performance, but the MLP's output layer must contain two neurons, one each outputting a score representing the normal and fault categories. To provide a confidence score for the fault, the MLP is followed by a Softmax layer to normalize the output data.

[0049] The proposed fusion model demonstrates that by fully connecting multiple layers of GBDTs, the output of the previous GBDT layer serves as the input feature for the next GBDT layer, avoiding the difficulty of manual feature extraction and gradually improving the performance of the classification model. Model fusion utilizes the Stacking integration framework and bootstrap sampling to integrate the advantages of a diverse set of GBDT models while fully leveraging the performance strengths of each model. The use of GBDT integration also reduces the difficulty of optimizing the hyperparameters of each GBDT. Finally, the output of the last GBDT layer is integrated using an MLP. Due to the expressive power of the MLP, it achieves better integration results than traditional integration schemes such as weighted summation and voting.

[0050] In this scheme, the number of gradient boosting tree GBDT models in each layer of learners and the number of hidden layers and the number of neurons in each hidden layer of the multilayer perceptron MLP model can be adjusted according to the performance of the model, but the number of neurons in the input layer of the multilayer perceptron MLP model is equal to the number of gradient boosting tree GBDT models in the second layer of learners. Each neuron in the input layer of the multilayer perceptron MLP model receives the charging pile fault prediction value output by a gradient boosting tree GBDT model in the second layer of learners; the output layer of the multilayer perceptron MLP model must contain two neurons, which output 0 and 1, corresponding to the normal and faulty states of the charging pile, respectively; in order to give a confidence score for the fault, a Softmax layer is added after the MLP to normalize the output data.

[0051] In this embodiment, the fusion model is composed of three layers of learners. The first layer of learners includes 5 gradient boosting tree GBDT models, the second layer of learners includes 3 gradient boosting tree GBDT models, and the third layer of learners includes a multi-layer perceptron MLP model with two hidden layers. The charging pile data vector is input into the first GBDT layer, which is composed of 5 GBDT models. The second GBDT layer is composed of 3 GBDT models, and the 5 predicted values ​​output by the first layer of GBDT are stacked to form a feature vector as its input. The hyperparameters of GBDT include: the number of subtrees, the maximum depth of the tree, the minimum loss reduction required to split the internal nodes, and the minimum number of samples required for the subtrees. They are respectively controlled by the variables given in Table 1, which also shows the hyperparameter settings of each GBDT.

[0052] The MLP is a fully connected neural network with two hidden layers. The input layer has three neurons and receives the output of the final GBDT layer. The first hidden layer consists of six neurons with a ReLU (rectified linear unit) nonlinear activation, and the second hidden layer has 16 neurons with a ReLU activation. The output layer consists of a fully connected layer with two neurons and no activation, and a softmax layer. Table 1 summarizes the MLP hyperparameter settings.

[0053] Table 1 Hyperparameter configuration

[0054]

[0055] Among them: n_estimators is the number of subtrees, max_depth is the maximum depth of the tree, min_sam_split is the minimum loss reduction required to split an internal node, and min_sam_leaf is the minimum number of samples required for a subtree.

[0056] according to Figure 1 The fusion diagnosis model of GBDT and MLP in the present invention is a network containing three-stage learners. The fusion model is trained by the following method:

[0057] Step S1: Obtain a charging pile dataset, including feature vectors and labels, and divide it into a training set and a test set. The charging pile feature vector consists of the charging pile's K1K2 drive signal, electronic lock drive signal, charging pile emergency stop signal, charging pile access control signal, and the total harmonic distortion of the charging pile voltage and current;

[0058] Step S2: Train the first-layer gradient boosting tree GBDT model: Use the training set obtained in step S1 to train each gradient boosting tree GBDT model in the first layer;

[0059] Step S3: Train the second-layer gradient boosting tree GBDT model: The prediction values ​​given by each gradient boosting tree GBDT model in the first layer are combined into a new feature vector, a new training set is constructed, and together with the corresponding sample labels, each gradient boosting tree GBDT model in the second layer is trained;

[0060] Step S4: training the multi-layer perceptron (MLP) model: the predicted values ​​given by the gradient boosting tree (GBDT) model of the second layer are combined into a new feature vector, and the multi-layer perceptron (MLP) model is trained;

[0061] Step S5: Model evaluation: Input the test set into the fusion model to evaluate the fusion model.

[0062] In the fusion model, the feature vector is sampled multiple times to provide different input data for each gradient boosting tree (GBDT) model in the first layer of learners, and the charging pile fault prediction values ​​output by all gradient boosting tree (GBDT) models in the previous layer of learners are stacked to provide different input data for each gradient boosting tree (GBDT) model in the next layer of learners.

[0063] The number of neurons in the input layer of the multilayer perceptron MLP model is equal to the number of gradient boosting tree GBDT models in the second-layer learner. One neuron in the input layer of the multilayer perceptron MLP model receives the charging pile fault prediction value output by a gradient boosting tree GBDT model in the second-layer learner.

[0064] Specifically, we first establish training and test datasets for charging pile fault assessment. In this step, we collect the required voltage, current, and other relevant data to form the charging pile input vector and corresponding labels. All data is then divided into two groups according to a certain ratio, serving as the training and test datasets.

[0065] Next, the five Gradient Boosted Tree (GBDT) models in the first layer are trained. Bootstrap sampling, a common approach in bagging ensemble learning, is used to construct five different training sets. These are then used to train the five Gradient Boosted Tree (GBDT) models, optimizing model parameters so that their outputs approximate the true sample labels. Therefore, the first GBDT layer is also called the bagging layer. Each Gradient Boosted Tree (GBDT) model in the first and second layers is trained using a softmax objective function and an exact greedy algorithm, with different hyperparameter settings for the GBDT models within the same layer.

[0066] Next, the three Gradient Boosted Tree (GBDT) models in the second layer are trained using a stacking framework. For each charging station input vector in the original training set, the five Gradient Boosted Tree (GBDT) models in the first layer generate five predictions, which are then combined into a new feature vector. Combined with the sample labels, a new training set is constructed. The three Gradient Boosted Tree (GBDT) models in the second layer are trained using cross-validation. Therefore, the second GBDT layer is also called a stacking layer.

[0067] Finally, the third-level learner, the MLP, is trained using the stacking framework. Each charging station vector in the original training set is fed into a two-stage GBDT network. The three predicted values ​​output are stacked into a new feature vector, which is then fed into the MLP along with the corresponding sample label. The MLP is trained using a binary cross-entropy loss and a stochastic gradient descent optimizer.

[0068] To verify the performance of the present invention, the proposed scheme was trained and tested on a public charging pile dataset and compared with typical machine learning algorithms.

[0069] 1. Charging pile dataset

[0070] The present invention uses the Baidu Dianshi new energy automobile charging pile dataset. Each charging pile data contains a feature vector and a fault label. The charging pile feature vector is composed of the charging pile's K1K2 drive signal, electronic lock drive signal, charging pile emergency stop signal, charging pile access control signal, and the total harmonic distortion of the charging pile voltage and current. By deploying gateways or concentrator devices at several charging stations, the charging pile data in the charging stations are fused and collected, and the data is transmitted to the background data server using parallel communication, resulting in a total of 122,144 data. The dataset is divided into a training set and a test set in a ratio of 7:3, resulting in 85,500 training data and 36,644 test data.

[0071] 2. Model implementation and training

[0072] The proposed model was implemented using the Python Scikit-learn package on a computer equipped with an AMD Ryzen 9 5950X 16-core CPU and 32GB of RAM. The charging station fault data was normalized using the z-score before model training. Next, the features and labels of the training set were input into the first-layer GBDT model using bagging sampling. Each GBDT in the first layer was sampled with 10,000 samples, approximately 1 / 7 of the dataset. The second-layer GBDT was trained using 5-fold cross-validation. The training dataset was divided into five equal-sized groups. Four groups were used for model training, and the remaining group was used for model evaluation. To train the third-layer learner, the MLP, a binary cross-entropy loss with a regularization coefficient of 0.08 was used. The momentum value of the stochastic gradient descent optimizer was fixed at 0.9, and the batch size was set to 200. The learning rate was initialized to 0.01 and decayed during training using an inverse-proportional learning rate with a parameter of 0.5 (exponential decay).

[0073] The performance of the detection scheme is measured by multiple objective indicators: accuracy, recall, precision and F1 score, and compared with traditional GBDT and MLP to show the performance of the present invention.

[0074] 3. Comparison and analysis of results

[0075] The proposed method was tested on a charging pile fault detection test set. Table 2 lists the accuracy, recall, precision, and F1 score of the proposed method, along with the results of traditional GBDT and MLP for comparison. In the comparative experiment, the MLP had 6 hidden layers, each with 100 neurons. The GBDT had 150 subtrees, with a maximum tree depth of 8. It can be seen that the parameter complexity of these two comparison models is much greater than that of the model constructed by the present invention.

[0076] Table 2 shows that although the traditional MLP model is complex, it can only obtain an F1 score of 89%, and using a higher depth and larger capacity GBDT alone can only achieve an accuracy of approximately 98%. The four indicators of the model of the present invention have all reached close to 99%, which is better than the previous two single models, indicating that the text method has achieved high accuracy in all indicators, can effectively detect the fault of the charging pile, and ensure the stable and safe operation of the charging pile. This high-precision fusion method is produced by the combined effect of two aspects. First, the performance of each base learner participating in the fusion is relatively high. Secondly, the various algorithms participating in the fusion meet the diversity requirements, and the multi-source information fusion can give full play to the attention information of different models to the features, thereby making full use of the feature information to improve the fusion performance. In addition, the present invention has made many optimizations in complexity. For example, when training the first layer of GBDT of the model, only 10,000 data are sampled, which greatly saves computing resources.

[0077] Table 2 Quantitative performance of different models on the test set

[0078]

[0079]

[0080] We further conducted ablation experiments to analyze the role of each layer in the proposed model. To this end, we implemented the following variations of the complete model: 1) Bagging layer model: five GBDT models with bootstrap sampling; 2) Bagging layer + stacking layer model: three GBDTs added in a fully connected fashion to the bagging layer model; and 3) Complete model: a bagging layer + stacking layer model followed by an MLP.

[0081] As can be seen, among these variations, the model with the bagging layer performs the worst. Although the F1 score increases with the complexity of the GBDT model, their performance is still below 98%. When the three GBDTs in the second layer are added one by one, the performance is improved, achieving F1 scores of 97.68%, 98.92%, and 98.99%, respectively. When all components are used, the full model shows the best performance, with an F1 score of 99.3%. That is, by adding the MLP, our full model can surpass the highest performance of the fusion using GBDT in terms of F1 score. These results show that all components provide performance gains and contribute to the overall performance.

[0082] Table 3 F1 scores of multiple variants of the fusion model

[0083]

[0084] This paper proposes a new method that integrates GBDT and MLP to solve the problem of charging pile fault diagnosis. Based on ensemble learning theory, this solution constructs multiple fully connected GBDT layers and connects them with MLP to perform feature learning and classification. The fusion of GBDT and MLP can effectively avoid their shortcomings, combine their advantages, and improve the accuracy of charging pile fault diagnosis. Moreover, this solution avoids the difficulties of manual feature extraction, reduces the probability of network overfitting, and ensures that the fault diagnosis performance is improved and more stable. Experimental results on public datasets show that the solution proposed by this invention is superior to typical independent machine learning solutions.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A charging pile fault diagnosis method based on the fusion of gradient boosting tree and multi-layer perceptron is characterized by: Collect the charging pile data set to be tested, construct it into a feature vector, input the feature vector into the fusion model, and output the charging pile fault diagnosis result. The fusion model includes N layers of learners, where N≥2, The first N-1 layers of learners are composed of multiple independent gradient boosting tree GBDT models, and the Nth layer of learners is a multi-layer perceptron MLP model with multiple hidden layers; The output of the previous layer of learners serves as the input of the next layer of learners. The gradient boosting tree GBDT model of the first layer of learners takes the feature vector of the charging pile as input and outputs the charging pile fault prediction. The gradient boosting tree GBDT model of the next layer integrates the output of the gradient boosting tree GBDT model of the previous layer to output a new charging pile fault prediction. The multi-layer perceptron MLP model of the Nth layer of learners receives the output of the gradient boosting tree GBDT model of the N-1th layer of learners and outputs the charging pile fault diagnosis result. The fusion model consists of three layers of learners, the first layer of learners includes five gradient boosting tree GBDT models, the second layer of learners includes three gradient boosting tree GBDT models, and the third layer of learners includes a multi-layer perceptron MLP model with two hidden layers; In the fusion model, the feature vector is sampled multiple times to provide different input data for each gradient boosting tree (GBDT) model in the first layer of learners, and the charging pile fault prediction values ​​output by all gradient boosting tree (GBDT) models in the previous layer of learners are stacked to provide different input data for each gradient boosting tree (GBDT) model in the next layer of learners.

2. The charging pile fault diagnosis method based on the fusion of gradient boosting tree and multi-layer perceptron according to claim 1 is characterized in that: The charging pile data set includes the charging pile's K1K2 drive signal, electronic lock drive signal, charging pile emergency stop signal, charging pile access control signal, and total harmonic distortion of the charging pile voltage and current.

3. The charging pile fault diagnosis method based on the fusion of gradient boosting tree and multi-layer perceptron according to claim 1 is characterized in that: The hyperparameter settings of each gradient boosting tree GBDT model in the same layer of learners are different.

4. The charging pile fault diagnosis method based on the fusion of gradient boosting tree and multi-layer perceptron according to claim 1 is characterized in that: Each gradient boosting tree (GBDT) model in the fusion model is trained using a Softmax objective function and an exact greedy algorithm.

5. The charging pile fault diagnosis method based on the fusion of gradient boosting tree and multi-layer perceptron according to claim 1 is characterized in that: The number of neurons in the input layer of the multilayer perceptron MLP model in the N-th layer learner is equal to the number of gradient boosting tree GBDT models in the N-1-th layer learner. One neuron in the input layer of the multilayer perceptron MLP model receives the charging pile fault prediction value output by a gradient boosting tree GBDT model in the N-1-th layer learner.

6. The charging pile fault diagnosis method based on the fusion of gradient boosting tree and multi-layer perceptron according to claim 1 is characterized in that: The output layer of the multi-layer perceptron MLP model in the Nth layer learner contains two neurons, corresponding to the normal and fault states of the charging pile respectively.

7. The charging pile fault diagnosis method based on the fusion of gradient boosting tree and multi-layer perceptron according to claim 1 is characterized in that: A Softmax layer is connected after the multi-layer perceptron MLP model of the Nth layer learner to output the confidence score of the fault.

Images