Electric bicycle charging abnormity intelligent identification method and system fusing experience knowledge

Abstract

The invention belongs to the technical field of electric data processing, and discloses an electric bicycle charging abnormity intelligent identification method and system fusing experience knowledge. The method comprises the following steps: reconstructing charging power data of the electric bicycle; secondly, convolutional layers with different time lengths are set in the TCN model to capture features under different time scales, and a cosine annealing learning rate adjustment strategy and a gating cycle unit are added to improve the recognition precision and the training speed; and finally, calling artificial experience knowledge to judge whether a potential abnormal charging condition exists or not for a sample of which the confidence coefficient is lower than a threshold value output by the model. The defects of the TCN model under the abnormal condition can be made up by combining with artificial experience knowledge, the possible misjudgment or missed judgment problem is corrected, and the hidden abnormal information in the data is further mined. According to the method, the detection capability of the TCN deep learning model and the flexibility and accuracy of artificial experience knowledge are fused, and the accuracy and reliability of electric bicycle charging abnormity identification are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of electronic digital data processing, relates to abnormal charging recognition of electric bicycles, and specifically to an intelligent recognition method for abnormal charging of electric bicycles integrating empirical knowledge. Background Art

[0002] Abnormal conditions during the charging process of electric bicycles can easily lead to safety hazards and equipment failures, so it is crucial to identify and prevent abnormal charging. In the prior art, machine learning technology is used to identify abnormal charging, and abnormal charging behavior curves are identified by analyzing charging power data.

[0003] However, in practical applications, machine learning methods are difficult to accurately identify different types of charging anomalies, especially the data imbalance problem caused by the fact that the amount of charging anomaly data is far less than that of normal data. This leads to the model's insufficient recognition ability for minority charging anomaly sample data, seriously affecting the training effect of the machine learning model.

[0004] Some methods use deep neural networks to detect abnormalities in electric loads. For example, the BP neural network algorithm is used as the input layer to receive raw data such as power consumption, voltage, current, power factor, etc. collected by smart meters in different periods of time. The hidden layer uses activation functions to perform nonlinear transformation and feature extraction on the data, mining the complex associations between the data, and the output layer outputs whether the power consumption behavior is abnormal; the CNN algorithm is used to clean and normalize the power consumption data, and then feature extraction is performed to capture local and global features in the power consumption pattern and identify abnormal power loads; RNN is used to process time series data, capture the time series features of load data, better handle the dependencies in long sequence data, and detect abnormal power loads. Although these methods can achieve a high recognition accuracy, it is difficult to make accurate judgments on special situations generated in the actual production process. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention proposes an intelligent identification method and system for electric bicycle charging anomalies that integrates experience knowledge, optimizes and reconstructs highly unbalanced data, improves the time series neural network in combination with gated recurrent units, introduces a cosine annealing learning rate adjustment strategy, and combines it with an artificial experience knowledge recognition mechanism to quickly and accurately identify different types of anomalies in the charging process of electric bicycles, thereby improving the accuracy and reliability of the classification and identification of electric bicycle charging anomalies.

[0006] An intelligent identification method for abnormal charging of electric bicycles integrating experience knowledge comprises the following steps:

[0007] Step 1: Electric bicycle charging power data collection and preprocessing

[0008] The real-time power of the electric bicycle charging pile is collected to generate a charging power sequence, expand the number of charging power sequences, and then divide the charging power sequences into 5 types according to normal, charger abnormality, battery abnormality, other abnormalities and worthless data.

[0009] Step 2: Reconstruction of charging power data of electric bicycles

[0010] Set the uniform length of the charging power sequence to , for charging power sequence exceeding The part with insufficient sequence length is discarded; The charging power sequence is supplemented with 0 elements. The charging power sequence data is truncated or padded to make the length of all sequences consistent.

[0011] Step 3: Build a time series neural network

[0012] A time series neural network is constructed for the classification of electric bicycle charging power data, wherein the time series neural network is a time convolutional neural network TCN, a recurrent neural network RNN ​​or a long short-term memory model LSTM.

[0013] As a preferred method, multi-scale convolution is introduced and combined with gated recurrent units to improve the TCN model:

[0014] s3.1. Construction of multi-scale convolution module

[0015] Four convolution kernels of different sizes are selected to construct a multi-scale convolution module. The multi-scale convolution module is used to replace the dilated causal convolution (DCC) layer in the residual block of the TCN model, so that the network can capture features at different time resolutions and improve recognition accuracy. The convolution layer output of the four convolution kernels , , Input into the concatenation layer to get the output of the multi-scale convolution module .

[0016] s3.2, residual block output

[0017] Input Data After passing through the multi-scale convolution module, weight normalization WN( ), Relu activation Relu( ) and random inactivation Dropout( ), the residual connection is performed with the result of 1x1 convolution conv( ) as the output of the residual block :

[0018] =

[0019] s3.3, Gated recurrent unit construction

[0020] A gated recurrent unit is added between each residual block of the TCN model to The output of the network at the current moment is obtained through the gated recurrent unit and used as the input data of the next residual block.

[0021] Step 4: Cosine annealing learning rate adjustment strategy

[0022] Use the time series neural network constructed in step 3 to detect the charging power sequence of the electric bicycle reconstructed in step 2, introduce the cosine annealing learning rate adjustment strategy, and train the time series neural network:

[0023]

[0024] in, is the current epoch, is the maximum epoch number, and . It is The learning rate for each epoch, and are the minimum and maximum learning rates respectively.

[0025] Step 5: Construct threshold judgment based on artificial experience knowledge

[0026] Calculate the difference between adjacent data of the charging power sequence, save the absolute value sequence and the symbol sequence. Build a threshold judgment model based on artificial experience knowledge, and determine the category according to the characteristics of the charging power sequence:

[0027] s5.1. Determination of worthless data

[0028] If the length of the input charging power sequence is less than the shortest sequence length , it means that the sequence contains too little information and cannot provide sufficient basis for classification, so it is judged as worthless data.

[0029] s5.2, Charger abnormality determination

[0030] Then count the number of data greater than 0 in the absolute value sequence and record it as the number of fluctuations. If the number of fluctuations is greater than 4 times and there is data greater than 5 in the difference sequence, it is determined that the charger is abnormal.

[0031] s5.3、Battery abnormality determination

[0032] Traverse the symbol sequence, if there are continuous After the value decreases more than times, If the value increases, it is considered that the battery is abnormal. .

[0033] s5.4, Normal data determination

[0034] If the absolute values ​​of the differences between adjacent data in the charging power sequence are all less than or equal to 5, and the number of fluctuations is less than 4 times, and there is no battery problem as mentioned above, then it is judged to be normal data.

[0035] s5.5. For other situations that do not meet the above conditions, they are judged as other abnormalities.

[0036] Step 6: Hybrid Model Detection

[0037] Set the confidence threshold threshold, input the charging power sequence of the electric bicycle into the time series neural network trained in step 4 for detection, and for the sequence whose detection result confidence is lower than the threshold, call the threshold judgment function based on artificial experience knowledge in step 5 for artificial experience knowledge judgment, otherwise use the detection result of the time series neural network.

[0038] The electric bicycle charging anomaly intelligent identification system integrating experience knowledge provided by the present invention comprises: a processor, a processor (CPU) and a graphics processing unit (GPU) for executing the computer program / instructions to implement the above-mentioned steps 1 to 6 of the electric bicycle charging anomaly intelligent identification method integrating experience knowledge;

[0039] The memory is used to store data, parameters and computer programs / instructions of the intelligent identification method for electric bicycle charging abnormality integrating empirical knowledge in steps one to six.

[0040] The present invention has the following beneficial effects:

[0041] 1. The data preprocessing, data truncation, filling and other data reconstruction methods of the present invention are adopted to improve the quality and consistency of the electric bicycle charging power input data, and provide a good data basis for subsequent recognition models and artificial experience knowledge judgment.

[0042] 2. Use the cosine annealing learning rate adjustment strategy and gated recurrent unit to improve the recognition speed and accuracy of the temporal convolutional neural network (TCN) model.

[0043] 3. Artificial experience knowledge is introduced, and multi-dimensional judgment rules are designed according to the different characteristics of the data. A threshold judgment function based on artificial experience knowledge is constructed, which can make targeted judgments on different types of electric bicycle charging anomalies and enhance the adaptability of the recognition method and system to different types of anomalies.

[0044] 4. By integrating the respective advantages of deep learning models and artificial experience knowledge, the shortcomings of the two methods are compensated, and the accuracy and reliability of the deep learning model for abnormality recognition are improved. Especially for various complex and diverse charging abnormality data, classification abnormality recognition can be achieved more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of an intelligent identification method for abnormal charging of electric bicycles that integrates empirical knowledge.

[0046] Figure 2 It is a schematic diagram of charging power data reconstruction.

[0047] Figure 3 It is a schematic diagram of the multi-scale convolution module.

[0048] Figure 4 This is a schematic diagram of the improved temporal convolutional neural network (TCN) model structure;

[0049] Figure 5 It is a schematic diagram of the gated recurrent unit structure.

[0050] Figure 6 This is a schematic diagram of a threshold judgment method based on artificial experience knowledge.

[0051] Figure 7 It is a structural diagram of an intelligent identification system for abnormal charging of electric bicycles that integrates experience knowledge. DETAILED DESCRIPTION

[0052] The present invention will be further explained below with reference to the accompanying drawings;

[0053] like Figure 1 As shown, the intelligent identification method for abnormal charging of electric bicycles integrating experience knowledge specifically includes the following steps:

[0054] Step 1: Use sensors to collect the real-time power of electric bicycle charging piles every minute. Arrange the power data collected each time in chronological order to form a charging power data sequence. The collected data set is expanded several times, and the total number of data reaches about 10,000. There are 5 types of charging power, namely normal, charger problem, battery problem, other problems and worthless data, with about 2,000 items in each category.

[0055] Step 2: Set the maximum length of the charging power data sequence to , for each charging power data sequence, if the length is greater than , then discard the excess If the length is less than , then fill it with 0 elements at the end, such as Figure 2All charging power sequence data are truncated and padded to make them of the same length to obtain the training set data.

[0056] Step 3: Improve the temporal convolutional neural network (TCN) model by combining the gated recurrent unit to improve the model recognition accuracy and recognition speed:

[0057] s3.1, such as Figure 3 As shown, select Four convolution kernels form a multi-scale convolution module, replacing the dilated causal convolutional (DCC) layer in the residual block of the TCN model. Convolutional layer output with four convolution kernels , , Splice to get the output of the multi-scale convolution module .

[0058] s3.2, residual block output

[0059] Input Data After passing through the multi-scale convolution module, weight normalization WN( ), Relu activation Relu( ) and random inactivation Dropout( ), the residual connection is performed with the result of 1x1 convolution conv( ) as the output of the residual block :

[0060] =

[0061] s3.3, such as Figure 4 As shown, a gated recurrent unit is added between each convolutional layer to convert the output of the residual block As the input of the gated recurrent unit, the output of the network at the current moment is obtained through the gated recurrent unit as the input data of the next residual block. Figure 5 As shown, the gated recurrent unit directly and the state of the network at the previous moment Add a linear dependency between them to solve the problem of gradient disappearance and gradient explosion, so that the TCN network can better capture long-term dependencies. Represents a time step:

[0062] (1) Set the initial hidden state ,in The dimension is The all-zero vector of the current step input gated recurrent unit network data .

[0063] (2) Calculate the reset gate to determine whether the candidate state at the current moment needs to rely on the network state at the previous moment, and how much it needs to rely on:

[0064]

[0065] In the formula, and are the trainable weight matrices of the network state at the current moment and the previous moment, respectively. is the bias vector. Represents sigmoid activation.

[0066] (3) According to the reset gate The value of determines the candidate hidden state Status of the previous moment Dependence:

[0067]

[0068] In the formula, and is the trainable hidden state weight matrix, is the hyperbolic tangent function, represents element-wise multiplication, is the bias vector of the hidden state.

[0069] (4) Calculate the update gate to control the state of the output at the current moment How much history state to keep in , and how many candidate hidden states to keep at the current moment :

[0070]

[0071] In the formula, and To update the trainable weight matrix of the gate, is the bias vector for the update gate.

[0072] (5) Update the output of the gate Separate and historical status and candidate hidden states Multiply the elements to get the output of the network at the current moment , as the input data of the next residual block:

[0073]

[0074] Step 4: Input the training set data into the improved TCN model in step 3 for detection. During the model training process, the cosine annealing learning rate adjustment strategy is used to improve the model convergence speed and recognition accuracy:

[0075]

[0076] in, is the current epoch, is the maximum epoch number, and . It is The learning rate for each epoch, and are the minimum and maximum learning rates respectively.

[0077] Step 5: Build Figure 6 The threshold judgment model based on artificial experience knowledge shown in FIG. calculates the difference between adjacent data of the charging power sequence, saves the absolute value sequence diffs and the symbol sequence trend, and determines the category according to the characteristics of the charging power sequence:

[0078] s5.1. Determination of worthless data

[0079] If the length of the input charging power sequence is less than 15, it means that the sequence contains too little information and cannot provide sufficient basis for classification, and is judged as worthless data.

[0080] s5.2, Charger abnormality determination

[0081] The number of elements in the absolute value sequence diffs that are not 0 is counted as the fluctuation count fluctuate_count. If there are more than 5 elements in diffs and the fluctuation count fluctuate_count is greater than 4 times, it is determined that the charger of the sample is abnormal.

[0082] s5.3、Battery abnormality determination

[0083] Traverse the symbol sequence trend, if there is a situation where the value decreases for more than 10 consecutive times and then increases for more than 10 consecutive times, it is determined that the sample has a battery abnormality.

[0084] s5.4, Normal data determination

[0085] If the absolute value sequence diffs does not contain elements greater than 5, and the number of fluctuations fluctuate_count is less than 4 times, and there is no battery problem mentioned above, then the sample is determined to be normal data.

[0086] s5.5. For other situations that do not meet the above conditions, they are judged as other abnormalities.

[0087] Step 6: Set the confidence threshold to 0.5. Figure 7As shown, the charging power sequence of the electric bicycle is input into the improved TCN model trained in step 4 for detection, and the recognition results and confidence are stored in a list. For samples whose confidence of the model recognition result is lower than the threshold, the artificial experience knowledge function is called to perform artificial experience knowledge judgment to determine its category; for samples whose confidence is not lower than the threshold, the TCN model recognition result is directly used as the final category judgment.

Claims

1. An intelligent identification method for abnormal charging of electric bicycles that integrates experience knowledge, inputs the classified charging power data of electric bicycles into a time series neural network, and trains the time series neural network to classify the charging power data of electric bicycles into normal, abnormal charger, abnormal battery, other abnormal, and worthless data, characterized by: Input the charging power data of electric bicycles into the trained time series neural network, calculate the confidence of the output results of the time series neural network, and if the confidence is lower than the set threshold, use the threshold judgment model based on artificial experience knowledge for reclassification; The threshold judgment model based on artificial experience knowledge will be less than the shortest sequence length The data is judged as worthless data; For a sequence whose length is not less than the shortest sequence length , calculate the difference between adjacent data, save the absolute value sequence diffs and the symbol sequence trend; traverse the symbol sequence trend, if there is a situation where the value decreases for more than 10 consecutive times and then increases for more than 10 consecutive times, the data is judged to be battery abnormal; count the number of elements that are not 0 in the absolute value sequence diffs as the number of fluctuations, if the number of fluctuations is greater than 4 and there are elements greater than 5 in the absolute value sequence diffs, then the data is judged to be charger abnormal; if there is no element greater than 5 in the absolute value sequence diffs and the number of fluctuations is less than 4, and there is no battery abnormality, then the data is judged to be normal data; otherwise, it is judged to be other abnormalities.

2. The intelligent identification method for charging anomalies of electric bicycles integrating experience knowledge as claimed in claim 1 is characterized by: The real-time power of the electric bicycle charging pile is collected to generate a charging power sequence, expand the data volume, and then divide it into five types according to normal, charger abnormality, battery abnormality, other abnormalities and worthless data; Set the uniform length of the charging power sequence to , for charging power sequence exceeding The part with insufficient sequence length is discarded; The charging power sequences are supplemented with 0 elements to make the lengths of all sequences consistent for training time series neural networks.

3. The intelligent identification method for charging anomalies of electric bicycles integrating experience knowledge as claimed in claim 1 is characterized by: The time series neural network is a time convolutional neural network TCN, a recurrent neural network RNN ​​or a long short-term memory model LSTM.

4. The intelligent identification method for charging anomalies of electric bicycles integrating experience knowledge as claimed in claim 1 is characterized by: The time series neural network is a time convolutional neural network TCN combined with a gated recurrent unit; specifically, the gated recurrent unit is inserted between each residual block of the time convolutional neural network TCN, and the output of the previous residual block is Input into the gated recurrent unit, and use the output of the gated recurrent unit as the input of the next residual block.

5. The intelligent identification method for charging anomalies of electric bicycles integrating experience knowledge as claimed in claim 4 is characterized by: The gated recurrent unit works by directly and the network status at the last moment Add a linear dependency between them to solve the problem of gradient disappearance and gradient explosion. Represents a time step: (1) Set the initial hidden state ,in The dimension is The all-zero vector of the current step input gated recurrent unit network data ; (2) Calculate the reset gate : ； In the formula, and are the trainable weight matrices of the network state at the current moment and the previous moment, respectively. is the bias vector; Represents sigmoid activation; (3) Calculate candidate hidden states : ； In the formula, and is the trainable hidden state weight matrix, is the hyperbolic tangent function, represents element-wise multiplication, is the bias vector of the hidden state; (4) Calculate the update gate : ； In the formula, and To update the trainable weight matrix of the gate, is the bias vector of the update gate; (5) Calculate the output of the network at the current moment , as the input data of the next residual block: 。 6. The intelligent identification method for charging anomalies of electric bicycles integrating experience knowledge as claimed in claim 4 is characterized by: Four convolution kernels of different sizes are selected to construct a multi-scale convolution module. The multi-scale convolution module is used to replace the dilated causal convolution in the residual block, and the convolution layer output of the four convolution kernels is , , Input into the concatenation layer to get the output of the multi-scale convolution module ; Output of residual block for: = ； in, Represents the input data of the residual block, WN( ) represents weight normalization, Relu( ) represents Relu activation, Dropout( ) represents random dropout, and conv( ) represents 1x1 convolution.

7. The intelligent identification method for charging anomalies of electric bicycles integrating experience knowledge as claimed in claim 6 is characterized by: The size of the convolution kernel in the multi-scale convolution module is .

8. The intelligent identification method for charging anomalies of electric bicycles integrating experience knowledge as claimed in claim 1 is characterized by: Introduce the cosine annealing learning rate adjustment strategy in the training process of time series neural network: ； in, is the current training round, is the maximum number of training epochs, and It is The learning rate for each training round, and are the minimum and maximum learning rates respectively.

9. The intelligent identification method for charging anomalies of electric bicycles integrating experience knowledge as claimed in claim 1 is characterized by: Minimum sequence length l min =15, confidence threshold = 0.

5.

10. An intelligent identification system for abnormal charging of electric bicycles integrating experience knowledge is characterized by: include: A processor, a processor and a graphics processor for executing the computer program / instructions to implement the method for intelligently identifying charging anomalies of electric bicycles integrating empirical knowledge as described in any one of claims 1 to 9; A memory for storing data, parameters and computer programs / instructions of the method for intelligently identifying charging anomalies of electric bicycles integrating empirical knowledge as described in any one of claims 1 to 9.

Images