A substation battery pack abnormal state evaluation method based on multi-modal features

Abstract

The application discloses a kind of based on multimodal feature's abnormal state evaluation method of transformer substation battery pack, the method includes the following steps: obtaining the current, temperature, humidity and vibration frequency of battery in historical data;Formalization fault prediction task and construct fault detection model, import pre-processing data to feature extraction module, obtain local feature;Import local feature to time series feature extraction module, obtain time series feature;Import time series feature to global feature extraction module;Obtain global feature;Import local feature, time series feature and global feature to feature fusion module, obtain multimodal feature;Import multimodal feature to prediction module, obtain predicted fault type.The application utilizes multiple detection indexes of transformer substation battery pack, sequentially extracts local feature, time series feature and global feature, and carries out multimodal data fusion to feature, which helps to improve the data anomaly capturing ability of model and improve the abnormal state recognition accuracy of transformer substation battery pack.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of storage batteries, and particularly relates to a substation storage battery pack abnormal state evaluation method based on multi-modal features. BACKGROUND

[0002] The direct current operating power supply system is one of the indispensable secondary devices in power plants and substations. As the core equipment of the direct current operating power supply system, the storage battery provides protection power in time when the station AC fails, and maintains the reliable operation of the protection equipment of the power station and the distribution station. The high reliability, long cycle service life and durability of the storage battery become the focus of research and development. Based on the high reliability of the storage battery, the project develops a mobile portable battery which is small in size, light in weight, does not need external power supply, is simple in wiring and convenient to install. In addition, during the long-term operation of the storage battery, the storage battery is in a floating charge state, and it is difficult to judge the performance of the storage battery, such as capacity reduction, internal resistance increase, and virtual connection of the battery pole. At present, the only way to monitor the performance of the storage battery is to do a qualitative discharge test on the storage battery. Generally, it is required to do it once a year according to the relevant regulations and standards. This method is not real-time enough and cannot find hidden troubles in time. It also needs personnel to maintain on site, and the overall efficiency is low, the cost is high, and the workload is large. In the prior art, an online monitoring system is embedded in the portable power supply, which can monitor the voltage, resistance and other parameters of the storage battery, control the charging and discharging of the battery pack, and estimate the state of the battery, so as to realize high-precision monitoring and control of the high-voltage battery pack.

[0003] The Chinese patent with the patent publication number CN107150602B provides a BMS battery management system for detecting the battery system, but there is a problem of not being able to fully utilize the detection data and a technical problem of not being able to effectively predict the fault of the storage battery in real time. SUMMARY

[0004] In view of the technical problems existing in the existing trunk whitening machine, a substation storage battery pack abnormal state evaluation method based on multi-modal features is provided, which comprises the following steps:

[0005] Step S1: obtaining the historical operation data of voltage, current, temperature, humidity and vibration frequency of the storage battery pack through the storage battery management system, and intercepting and quantitatively labeling the data according to the actual abnormal state type and its severity;

[0006] Step S2: pre-processing the data by using abnormal filtering, format correction, data splicing and data standardization, obtaining array format pre-processed data, enhancing the data with abnormal state by downsampling, flipping and twisting, and constructing a training sample data set;

[0007] Step S3: formalize the abnormal state evaluation task and build an intelligent evaluation model based on a deep learning network structure, which includes a local feature extraction module, a time series feature extraction module, a global feature extraction module, a feature fusion module, and an identification output module; the preprocessed data is sequentially imported into the feature extraction module, the time series feature extraction module, and the global feature extraction module, respectively, to obtain local features, time series features, and global features, and the obtained local features, time series features, and global features are multi-scale fused to obtain multi-modal features, which are intelligently identified by the identification output module to obtain evaluation indicators of each abnormal state type;

[0008] Step S4: build a model loss function using quantile loss, randomly select 30 groups of training sample set data in batches, import them into the intelligent evaluation model, and train the intelligent evaluation model in S3 by minimizing the loss function as the target using the Adam optimization algorithm to obtain the parameter-optimized intelligent evaluation model.

[0009] Step S5: obtain the on-site battery pack voltage, current, temperature, humidity, and vibration frequency data, which are preprocessed in S2 and then imported into the optimized intelligent evaluation model to output the battery pack state type prediction result.

[0010] Further, the abnormal state types are battery leakage, internal short circuit, overcharging, undercharging, installation failure, abnormal environmental temperature, and abnormal environmental humidity, and include combinations of different abnormal state types, and the severity of each abnormal state type is divided into three levels: mild, severe, and critical.

[0011] Further, in S3, the formalized abnormal state evaluation task and the construction of the intelligent evaluation model based on the deep learning network structure are as follows:

[0012] Given a set of preprocessed data , wherein, , respectively, represent the voltage, current, temperature, humidity, and vibration frequency of the battery pack, and the corresponding abnormal state type and severity are diagnosed by the intelligent evaluation model, represented as:

[0013] ;

[0014] , wherein, represents the first state with the highest severity, represents the second state with the next highest severity, represents the intelligent evaluation model.

[0015] Further, the local feature extraction module in S3 is a combination of ResNeSt blocks and Feature refinement blocks.

[0016] Further, the time sequence feature extraction module in S3 first performs a maximum pooling operation on the output F1 of the local feature extraction module to obtain Then, the time sequence dependent features are extracted through a long short-term memory network, specifically as follows:

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] ;

[0023] wherein, represents a candidate cell state, represents an updated cell state, represents a hidden state, , , and each represent a preset weight matrix, , , and each represent a preset weight parameter, represents a SigMoid activation function, represents a tanh activation function, represents a hidden state, i.e., a time sequence feature , represents a hidden state at the previous time.

[0024] Further, the global feature extraction module in S3 processes the time sequence features in the following manner through a multi-head self-attention mechanism:

[0025] First, the input time sequence features are transformed through three different linear transformations to obtain a query matrix Q, a key matrix K, and a value matrix V; then, the attention weight of each head is calculated using a scaled dot-product attention; the above process is processed in parallel to obtain global features :

[0026] ;

[0027] ;

[0028] ;

[0029] MultiHead Concat ;

[0030] ;

[0031] in, It is a learnable weight matrix; It is the dimension of the key vector; and It is a learnable weight matrix.

[0032] Furthermore, multimodal features are obtained in S3. The specific operation is as follows:

[0033] ;

[0034] Furthermore, the identification output module described in S3 is a fully connected layer that uses a probability density function to calculate the probability of each abnormal state type, sorts the corresponding abnormal state types according to their magnitude, and outputs the first two types and their probabilities.

[0035] The positive and progressive effects of this invention are as follows:

[0036] 1) This invention improves the ability to capture detection data and enhances the generalization ability of the model by extracting multiple detection indicators of the battery module and extracting local features, time-series features and global features in sequence, enabling fault type analysis at multiple scales.

[0037] 2) This invention makes full use of historical testing data and can promptly obtain two fault types predicted by probability, which simplifies the procedure for battery maintenance and inspection and improves efficiency. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the steps of a substation battery pack abnormal state identification method based on multimodal features according to the present invention. Detailed Implementation

[0039] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0040] Reference Figure 1 A method for assessing abnormal states of substation battery banks based on multimodal features, in one example, the method includes the following steps:

[0041] Step S1: Obtain the historical operation data of the voltage, current, temperature, humidity and vibration frequency of the battery pack through the battery management system, and intercept and quantitatively label the data according to the actual abnormal state type and its severity;

[0042] Step S2: Preprocess the data by using abnormal filtering, format correction, data splicing and data standardization to obtain array format preprocessed data, enhance the data with abnormal state by downsampling, flipping and twisting, and construct a training sample data set;

[0043] Step S3: Formalize the abnormal state evaluation task and construct an intelligent evaluation model based on a deep learning network structure, which includes a local feature extraction module, a time series feature extraction module, a global feature extraction module, a feature fusion module and an identification output module; the preprocessed data is sequentially imported into the feature extraction module, the time series feature extraction module and the global feature extraction module, respectively, to obtain local features, time series features and global features, and the obtained local features, time series features and global features are multi-scale fused to obtain multi-modal features, which are intelligently identified by the identification output module to obtain evaluation indicators of each abnormal state type;

[0044] Step S4: Construct a model loss function using quantile loss, randomly select 30 groups of training sample set data in batches, import them into the intelligent evaluation model, and train the intelligent evaluation model in S3 by using the Adam optimization algorithm to minimize the loss function as the target to obtain the parameter-optimized intelligent evaluation model;

[0045] Step S5: Obtain the voltage, current, temperature, humidity and vibration frequency data of the on-site battery pack, import them into the optimized intelligent evaluation model after preprocessing in S2, and output the battery pack state type prediction result.

[0046] Further, in an example, the abnormal state types are battery leakage, internal short circuit, overcharging, undercharging, installation failure, abnormal environment temperature and abnormal environment humidity, and include combinations between different abnormal state types, and the severity of each abnormal state type is divided into three levels: mild, severe and critical.

[0047] Further, in an example, the formalized abnormal state evaluation task in S3 and the construction of the intelligent evaluation model based on the deep learning network structure are specifically:

[0048] Given a set of preprocessed data wherein, respectively represent the voltage, current, temperature, humidity and vibration frequency of the battery pack, and the corresponding abnormal state type and severity are diagnosed by the intelligent evaluation model, which is represented as:

[0049] ;

[0050] wherein, represents a first state with the highest severity, represents a second state with the second highest severity, represents an intelligent evaluation model.

[0051] Further, in an example, the local feature extraction module in S3 is a combination of ResNeSt block and Feature refinement block.

[0052] Further, in an example, the time sequence feature extraction module in S3 first performs a max-pooling operation on the output F1 of the local feature extraction module to obtain , and then extracts time sequence dependent features through a long short-term memory network, specifically as follows:

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] ;

[0058] ;

[0059] wherein, represents a candidate cell state, represents an updated cell state, represents a hidden state, , , and all represent preset weight matrices, , , and all represent preset weight parameters, represents a SigMoid activation function, represents a tanh activation function, represents a hidden state, i.e., time sequence features , represents a hidden state at the previous time.

[0060] Further, in an example, the global feature extraction module in S3 processes the time sequence features through a multi-head self-attention mechanism in the following manner:

[0061] First, the input time sequence features X are processed through three different linear transformations to obtain query matrix Q, key matrix K and value matrix V. Then, the attention weight of each head is calculated using the scaled dot-product attention, and the above process is processed in parallel to obtain the global feature :

[0062] ;

[0063] ;

[0064] ;

[0065] MultiHead Concat ;

[0066] ;

[0067] wherein, is a learnable weight matrix; is the dimension of the key vector; and are learnable weight matrices.

[0068] Further, in an example, the specific operation of obtaining the multi-modal feature in S3 is as follows:

[0069] ;

[0070] Further, in an example, the identification output module in S3 is a fully connected layer that uses a probability density function to calculate the probability of each abnormal state type, sorts the corresponding abnormal state types according to their size, and takes the first two types and their probabilities as the output.

[0071] The above embodiments of the application are described in detail in conjunction with the accompanying drawings, and those of ordinary skill in the art can make various changes to the application based on the above description. Therefore, some details in the embodiments should not be construed as limiting the application, and the scope of protection of the application will be defined by the appended claims.

Claims

1. A substation battery bank abnormal state evaluation method based on multi-modal features, characterized in that, The method comprises the following steps: Step S1: Obtain the historical operation data of the voltage, current, temperature, humidity and vibration frequency of the battery pack through the battery management system, and intercept and quantitatively label the data according to the actual abnormal state type and its severity; Step S2: Preprocess the data by using abnormal filtering, format correction, data splicing and data standardization, obtain array format preprocessed data, enhance the data with abnormal state by downsampling, flipping and twisting, and construct a training sample data set; Step S3: Formulate the abnormal state evaluation task and construct an intelligent evaluation model based on a deep learning network structure, which includes a local feature extraction module, a time sequence feature extraction module, a global feature extraction module, a feature fusion module and an identification output module; the preprocessed data is sequentially imported into the feature extraction module, the time sequence feature extraction module and the global feature extraction module, respectively, to obtain local features, time sequence features and global features, and the obtained local features, time sequence features and global features are multi-scale fused to obtain multi-modal features, which are intelligently identified by the identification output module to obtain evaluation indexes of each abnormal state type; Step S4: Construct a model loss function using quantile loss, randomly select 30 groups of training sample set data in batches, import them into the intelligent evaluation model, and train the intelligent evaluation model in S3 by using the Adam optimization algorithm to minimize the loss function as the target, to obtain the parameter-optimized intelligent evaluation model; Step S5: Obtain the voltage, current, temperature, humidity and vibration frequency data of the on-site battery pack, import them into the optimized intelligent evaluation model after preprocessing in S2, and output the battery pack state type prediction result.

2. The method according to claim 1, wherein, The abnormal state types are battery liquid leakage, internal short circuit, overcharging, undercharging, installation failure, abnormal environmental temperature and abnormal environmental humidity, and include combinations of different abnormal state types, and the severity of each abnormal state type is divided into three degrees: mild, severe and critical.

3. The method of claim 1, wherein the method further comprises: The formalized abnormal state evaluation task in S3 and the construction of the intelligent evaluation model based on the deep learning network structure are as follows: Given a set of pre-processed data wherein, respectively represent the voltage, current, temperature, humidity and vibration frequency of the battery pack, diagnose their corresponding abnormal state types and severity through the intelligent evaluation model, represented as: ； wherein, represents a first state of the highest severity, represents a second state of the next highest severity, represents an intelligent assessment model.

4. The method of claim 1, wherein the method further comprises: The local feature extraction module in S3 is a combination of ResNeSt block and Feature refinement block.

5. The method of claim 1, wherein the method further comprises: The time sequence feature extraction module in S3 first performs a maximum pooling operation on the output F1 of the local feature extraction module to obtain And then extracts time sequence dependent features through a long short-term memory network, specifically as follows: ； ； ； ； ； ； wherein, denotes a candidate cell state, denotes an updated cell state, denotes a hidden state, , , and each denote a preset weight matrix, , , and each denote a preset weight parameter, denotes a SigMoid activation function, denotes a tanh activation function, denotes a hidden state, i.e. a temporal feature , denotes a hidden state of a previous time instant.

6. The method of claim 1, wherein the method is based on multi-modal features of the battery bank. The way of processing time sequence features by the global feature extraction module in S3 through the multi-head self-attention mechanism is as follows: First, the input time sequence features are processed by a convolutional neural network to obtain the query matrix Q, the key matrix K, and the value matrix V. The query matrix Q, the key matrix K, and the value matrix V are obtained through three different linear transformations; then, the attention weight of each head is calculated by using the scaled dot-product attention; and the above process is processed in parallel to obtain the global feature : ； ； ； MultiHead Concat ; ； wherein, is a learnable weight matrix; is the dimension of the key vector; and is a learnable weight matrix.

7. The method of claim 1, wherein the method further comprises: S3 acquires multi-modal features The specific operation is as follows: 。 8. The method of claim 1, wherein the method further comprises: The identification output module in S3 is a full connection layer that uses a probability density function to calculate the probability of each abnormal state type, and sorts the corresponding abnormal state types according to their size, taking the first two types and their probabilities as the output.

Images