KR20200002665A - Prognostics and health management systems for component of vehicle and methods thereof - Google Patents

Abstract

The present invention relates to a system and method for prognostics and health management for a component of a transport means. The method comprises the steps of: receiving time series sensor data from a module installed in a component of a transport means; updating a parameter of a recurrent neural network for estimating a state of the component or detecting a symptom using the received time series sensor data; and estimating the state of the component of the transport means or detecting the symptom using the recurrent neural network to which the updated parameter of the recurrent neural network is applied.

Description

Fault prediction and soundness management systems and methods for parts of vehicles {PROGNOSTICS AND HEALTH MANAGEMENT SYSTEMS FOR COMPONENT OF VEHICLE AND METHODS THEREOF}

The present invention relates to a failure prediction and health management system for a part of a vehicle.

With the rise of environmental and energy resources, electric vehicles are in the spotlight as the vehicle of the future. Electric vehicles have a merit that there is no exhaust gas and noise is low because a secondary battery (cell) capable of charging and discharging uses a battery formed as a pack as a main power source.

In an electric vehicle, the battery serves as an engine and a fuel tank of a gasoline vehicle, so for the safety of the electric vehicle user, it may be important to check the state of the battery.

In recent years, research to increase the user's convenience while checking the state of the battery more accurately has continued.

There are two main methods for measuring battery state of charge (SOC).

First, there is an open circuit voltage method.

By measuring the open voltage of a battery cell, an OCV-SOC table is created and used to estimate the battery SOC. In order to measure the open circuit voltage, it is impossible to estimate in real time since the vehicle must be stopped and the battery waits for the stabilization state.

Second, there is a current integration method.

The current integration method is a method of estimating the SOC change from the initial SOC value by continuously integrating the amount of current in the battery. Current integration has the disadvantage that the estimation error accumulates if the initial SOC value has an error.

In addition, both methods have a disadvantage in that they do not respond to changes in battery aging over time.

Embodiments of the present invention, by using a regression neural network using a data sequence mixed at least one of the short-term, medium-term and long-term data of the time series sensor data, it is possible to manage the failure prediction and health of the parts of the vehicle, To provide a failure prediction and health management system and method for parts of vehicles.

Embodiments of the present invention provide a failure prediction and health management system and method for parts of a vehicle, which can estimate battery information more accurately and respond to changes due to aging of the battery.

According to an embodiment of the present invention, there is provided a failure prediction and health management method of a vehicle, the method comprising: receiving time series sensor data from a module installed in a component of the vehicle; Updating a parameter of a recurrent neural network using the received time series sensor data to estimate a state of the part or detect an abnormal symptom; And estimating a state of a part of the vehicle or detecting an abnormal symptom by using the regression neural network to which the parameters of the updated regression neural network are applied.

In the updating of the parameter, the parameter of the regression neural network may be updated by using a data sequence in which at least one of the short, medium and long term data of the time series sensor data is mixed.

The regression neural network may include at least one of a state of charge (SOC), a state of health (SOH), a power measurement, a voltage measurement, a current measurement, and a temperature measurement of the component. The included input vector may be input through the input layer.

The regression neural network may be a long short-term memory (LSTM) neural network.

The LSTM neural network may include at least one neural network layer, which is a collection of cells forming an LSTM neural network, and a fully connected neural network having an input of an output of each cell of the neural network layer. .

The LSTM neural network includes one neural network layer including a plurality of cells, and a fully connected neural network having an input of an output of each cell of the one neural network layer, wherein each of the plurality of cells includes the received time series sensor data. The corresponding short-term data may be input from a plurality of short-term data divided in time series.

The LSTM neural network includes a plurality of neural network layers sequentially connected to each other and a fully connected neural network having an output of a neural network layer last connected in the plurality of neural network layers, wherein each of the plurality of cells includes: The received time series sensor data is inputted with corresponding short-term data among a plurality of short-term data classified in chronological order, wherein the plurality of neural network layers are selected from two of the plurality of previous cells included in a neural network layer of a previous order among the plurality of neural network layers. It may include a next order neural network layer including a plurality of next cells that receive the input of the outputs of the one or more previous cells together.

The fully connected neural network may further receive an output of a neural network layer last connected in the plurality of neural network layers and an output of at least one cell in the plurality of neural network layers.

According to another embodiment of the invention, the communication unit for communicating with the module installed in the component of the vehicle; A memory for storing at least one program; And a processor coupled to the communication unit and the memory, wherein the processor receives the time series sensor data from a module installed in a component of a vehicle by executing the at least one program, and uses the received time series sensor data. Update a parameter of a recurrent neural network for estimating the state of the part or detecting an abnormal symptom, and estimate a state of the part of the vehicle or an abnormal symptom using the regression neural network to which the updated regression neural parameter is applied. Failure detection and health management device of the vehicle can be provided for detecting.

The processor may update a parameter of the regression neural network by using a data sequence obtained by mixing at least one data among short, medium and long term data of the time series sensor data.

The regression neural network may include at least one of a state of charge (SOC), a state of health (SOH), a power measurement, a voltage measurement, a current measurement, and a temperature measurement of the component. The included input vector may be input through the input layer.

The regression neural network may be a long short-term memory (LSTM) neural network.

The LSTM neural network may include at least one neural network layer, which is a collection of cells forming an LSTM neural network, and a fully connected neural network having an input of an output of each cell of the neural network layer. .

The LSTM neural network includes one neural network layer including a plurality of cells, and a fully connected neural network having an input of an output of each cell of the one neural network layer, wherein each of the plurality of cells includes the received time series sensor data. The corresponding short-term data may be input from a plurality of short-term data divided in time series.

The LSTM neural network includes a plurality of neural network layers sequentially connected to each other and a fully connected neural network having an output of a neural network layer last connected in the plurality of neural network layers, wherein each of the plurality of cells includes: The received time series sensor data is inputted with corresponding short-term data among a plurality of short-term data classified in chronological order, wherein the plurality of neural network layers are selected from two of the plurality of previous cells included in a neural network layer of a previous order among the plurality of neural network layers. It may include a next order neural network layer including a plurality of next cells that receive the input of the outputs of the one or more previous cells together.

The fully connected neural network may further receive an output of a neural network layer last connected in the plurality of neural network layers and an output of at least one cell in the plurality of neural network layers.

According to another embodiment of the present invention, there is provided a module comprising: a module for collecting and transmitting time series sensor data obtained from a part of a vehicle; And updating a parameter of a recurrent neural network for estimating the state of the part or detecting an abnormal symptom using the time series sensor data received from the module, and transmitting the updated regression neural network parameter to the module. And a cloud server configured to update a parameter of the regression neural network using a data sequence obtained by mixing at least one of short-term, medium-term, and long-term data of time-series sensor data. The failure prediction and soundness management system for the parts of the vehicle, which estimates the state of the parts of the vehicle or detects an abnormal symptom using the above, may be provided.

According to another embodiment of the invention, a non-transitory computer readable storage medium comprising at least one program executable by a processor, wherein the at least one program, when executed by the processor, causes the processor to: Receive time series sensor data from a module installed in a component of the device, update the parameters of a recurrent neural network for state estimation or abnormal symptom detection of the component using the received time series sensor data, and update the regression. A non-transitory computer readable storage medium may be provided that includes instructions for estimating a state of a part of the vehicle or detecting an abnormal symptom using a regression neural network to which parameters of a neural network are applied.

Embodiments of the present invention can manage failure prediction and health of parts of a vehicle by using a regression neural network using a data sequence in which at least one data among short, medium and long term data of time series sensor data is mixed.

With the growing market for electric vehicles due to environmental pollution, accurate condition estimation of Li-ion batteries is a priority for safe vehicle operation. According to embodiments of the present invention, the estimated battery information may be used for the motor output and the vehicle operation, and reliable information may be provided to the driver.

1 is a block diagram illustrating a configuration of a failure prevention and health management system of a vehicle according to embodiments of the present invention.
2 is a block diagram illustrating the configuration of a failure prevention and health management system of a battery mounted to a vehicle according to an embodiment of the present invention.
3 is a block diagram illustrating a configuration of an LSTM neural network for estimating battery information in a failure prediction and health management system according to an exemplary embodiment of the present invention.
4 is a view for explaining the LSTM structure used in an embodiment of the present invention.
5 is a view for explaining a single layer LSTM structure used in an embodiment of the present invention.
6 is a diagram illustrating a graph of a state of charge prediction result during an iterative learning process at full temperature in a tomography layer LSTM used in an embodiment of the present invention.
7 to 9 are diagrams for explaining the multi-layer LSTM structure used in the embodiments of the present invention.
10 is a configuration diagram for explaining the configuration of the failure prevention and health management device of the vehicle according to an embodiment of the present invention.
11 and 12 are flowcharts of data learning and data testing performed in a failure prediction and health management system according to an exemplary embodiment of the present invention.
13 to 16 are diagrams for explaining an excellent charge state prediction result graph during an iterative learning process at each temperature for embodiments of the present invention.
17 is a view for explaining the overall configuration of the IoT-based PHM system according to an embodiment of the present invention.
18 shows a conceptual diagram of an IoT module in an EV's PHM system.
FIG. 19 is a diagram illustrating an extension of DB-LSTM-VAE for fault prediction according to an embodiment of the present invention. FIG.
20 is a diagram for describing a deep ensemble fusion configuration diagram having a hierarchical structure.
FIG. 21 illustrates a method for fine tuning self-generated data and data-based DB-LSTM-VAE collected externally.

As the inventive concept allows for various changes and numerous embodiments, particular embodiments will be illustrated in the drawings and described in detail in the written description.

However, this is not intended to limit the present invention to specific embodiments, it should be understood to include all changes, equivalents, and substitutes included in the spirit and scope of the present invention.

Terms such as first and second may be used to describe various components, but the components should not be limited by the terms. The terms are used only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be referred to as the second component, and similarly, the second component may also be referred to as the first component. The term and / or includes a combination of a plurality of related items or any item of a plurality of related items.

When a component is referred to as being "connected" or "connected" to another component, it may be directly connected to or connected to that other component, but it may be understood that other components may be present in between. Should be. On the other hand, when a component is said to be "directly connected" or "directly connected" to another component, it should be understood that there is no other component in between.

The terminology used herein is for the purpose of describing particular example embodiments only and is not intended to be limiting of the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this application, the terms "comprise" or "have" are intended to indicate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, and one or more other features. It is to be understood that the present invention does not exclude the possibility of the presence or the addition of numbers, steps, operations, components, components, or a combination thereof.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art. Terms such as those defined in the commonly used dictionaries should be construed as having meanings consistent with the meanings in the context of the related art, and shall not be construed in ideal or excessively formal meanings unless expressly defined in this application. Do not.

Hereinafter, with reference to the accompanying drawings, it will be described in detail a preferred embodiment of the present invention. In the following description of the present invention, the same reference numerals are used for the same elements in the drawings and redundant descriptions of the same elements will be omitted.

1 is a block diagram illustrating a configuration of a failure prevention and health management system of a vehicle according to embodiments of the present invention.

As illustrated in FIG. 1, the failure prediction and health management system 10 according to the embodiments of the present invention includes an IoT module 12 and a cloud server 13. However, not all illustrated components are essential components. The failure prediction and health management system may be implemented by more components than the illustrated components, and the failure prediction and health management system may be implemented by fewer components.

Hereinafter, specific configurations and operations of the components of the failure prediction and health management system of FIG. 1 will be described.

Firstly the vehicle 11 consists of individual parts. For example, the vehicle 11 may be a vehicle for transporting a user such as a vehicle, an airplane, a train, and the like, and is not limited to a specific vehicle.

Each part of the vehicle 11 is to be maintained and replaced, in order to ensure the reliability of the vehicle, and transmits information about their status to the IoT module 12 in real time. This information can be accumulated to form big data about the state of the vehicle.

The IoT module 12 is mounted on the vehicle 11. The IoT module 12 may include a sensor connected to each component. Alternatively, the IoT module 12 may be connected to a sensor connected to each component. The IoT module 12 transmits the big data transmitted from each component of the vehicle 11 to the cloud server 13, and transmits an abnormality detection and warning message of the components to the digital cluster of the vehicle based on the real-time data. It can transmit to the user's terminal (eg, smart phone, etc.).

The cloud server 13 may store big data information of each component transmitted from the IoT module 12 via the Internet, and transmit failure prediction information considering a long term to a user terminal (eg, a smartphone) of the user. Based on the big data information stored in the cloud server 13, the deep learning network in the failure prediction and health management system 10 may be periodically learned. The IoT module 12 may update the weight of the deep learning network located in the abnormal indication detection and warning system of the IoT module 12 using the weight of the learned network. Can be.

For example, the IoT module 12 may be implemented as a module installed in a part of the vehicle 11, and may collect and transmit time series sensor data obtained from the part of the vehicle 11 to the cloud server 13. have.

The cloud server 13 uses the time series sensor data received from the IoT module 12 to update a parameter of a recurrent neural network for estimating a state of a part or detecting an abnormal symptom and to update the parameter of the updated regression neural network. While transmitting to the IoT module 12, a parameter of the regression neural network may be updated by using a data sequence in which at least one data among short-term, medium-term and long-term data of the time series sensor data is mixed.

The IoT module 12 may estimate a state of a part of the vehicle 11 or detect an abnormal symptom using the regression neural network to which the parameters of the regression neural network updated by the cloud server 13 are applied.

Hereinafter, referring to FIG. 2, a configuration of managing failure prediction and health of a battery mounted on a vehicle among components of the vehicle 11 will be described as an embodiment. However, embodiments of the present invention are not limited to a specific vehicle or a specific part of the vehicle.

2 is a block diagram illustrating the configuration of a failure prevention and health management system of a battery mounted to a vehicle according to an embodiment of the present invention.

As shown in FIG. 2, the failure prediction and health management system of the battery accurately corrects SOC, SOH, and anomalies from battery information including voltage, current, temperature, SOC output by current integration, or voltage change per hour. Battery state, etc., may be estimated.

Battery failure prediction and health management systems can estimate battery status over the entire battery life from a short, medium, long term, or a combination of battery measurement sequences.

The IoT module 110 may estimate battery information in real time using the learned parameter values, and collect / transmit battery measurement values for building a DB for battery modeling.

The cloud server 120 may receive the battery measurement value through the network to learn the parameter and update the parameter based on the learning result. The cloud server 120 may transmit the updated parameter to the IoT module.

Failure prediction and health management system according to an embodiment of the present invention utilizing the voltage, current, temperature or more of the measured value of the electric vehicle lithium-ion battery, that is, the information of the battery, namely state of charge, remaining life Remaining Useful Life, anomaly, etc.

The failure prediction and health management system according to an exemplary embodiment of the present invention may accurately predict battery information by using a long short-term memory (LSTM) neural network suitable for time series data modeling as a battery information estimation model. The LSTM neural network can be used as an input for short, medium, long term, or a mixture of battery measurement sequences. In addition, the failure prediction and health management system according to an embodiment of the present invention collects / transmits battery measurements for real-time estimation of battery information and DB for battery modeling using parameter values learned in a cloud server. The module 110 may be separated into a cloud server 120 that receives a battery measurement value through a network and transmits parameters to a learning / update and IoT module. Alternatively, the failure prediction and health management system according to an embodiment of the present invention may perform a process of collecting battery measurement values and estimating battery information through a system installed in a vehicle.

As shown in FIG. 2, the battery 101 is a battery to be subjected to battery information estimation. The battery 101 may be configured in a package or cell unit.

The battery management system 102 is a battery management system for battery management. The battery management system 102 may transmit voltage, current, and temperature measurement values of the battery to the IoT module in the vehicle.

The IoT module 110 is an IoT module mounted in a vehicle. The IoT module 110 may receive the battery measurement from the battery management system 102 and transmit it to the cloud server 120. The IoT module 110 may serve as a real-time estimation of battery information.

The preprocessor 111 may perform preprocessing on the battery measurement value received from the BMS 102. The pretreatment process may include normalization, absolute temperature conversion of degrees Celsius, and the like.

The communication unit 112 may transmit the battery measurement value received from the BMS 102 to the cloud server 120, and receive the updated parameter value from the cloud server 120. The information transmitted to the cloud server 120 may include not only battery measurement values but also driver information. The cloud server 120 may use various communication technologies such as a cellular network and Wi-Fi.

The estimator 113 may estimate battery information in real time using an LSTM neural network using previously learned parameters.

The cloud server 120 is a server connected to the IoT module 110 through a communication network or the Internet. The cloud server 120 may receive battery measurements and perform new model training for the deep learning network.

The communication unit 121 may receive the battery measurement value transmitted from the IoT module 110 and transmit the updated parameter value to the IoT module 110. In this case, various communication technologies such as a cellular network and a Wi-Fi may be used.

The data storage unit 122 may store data received by the communication unit 121.

The learner 123 may update the parameters of the LSTM neural network by using the data received from the communicator 121. The updated parameter may be stored in the parameter storage unit 124. The parameter storage unit 124 may serve to store the parameter updated by the learning unit 123.

The cloud server 120 may transmit a parameter through the communication unit 121 when a parameter update is required in the IoT module 110.

3 is a block diagram illustrating a configuration of an LSTM neural network for estimating battery information in a failure prediction and health management system according to an exemplary embodiment of the present invention.

In the failure prediction and health management system according to an exemplary embodiment of the present invention, the LSTM neural network for estimating battery information may include LSTM neural network layers such as an input layer, a hidden layer, and an output layer.

The input vector 201 is an input vector of the LSTM neural network and may include a battery voltage (V), a current (I), or a temperature (T) measurement value at each time. The dimension of the vector input to one LSTM cell may be determined according to the length of the time series. The length of the time series can be determined by the amount of information collected. In addition, the length of the input vector of each cell may be the same or different.

The cell 202 refers to a cell forming an LSTM neural network. Each cell may output the characteristics of time series data in a vector form. The gate and state constituting the cell may be defined by Equation 1 as follows.

Each cell does not have a fixed structure as shown in [Equation 1] above, and may be transformed into a more advanced form such as a bidirectional LSTM cell.

The neural network layer 203 is a layer which is a collection of cells constituting the LSTM neural network. If necessary, the LSTM neural network may be formed of at least one layer. That is, the plurality of layers 203 may include a single layer or two or more layers. Each cell receives the cell output of the previous layer as an input. In this case, the input may receive one or more cell outputs existing in the previous layer as inputs. Cells are layered so that at least one of short-term, medium-term and long-term data according to the length of the input time series data can be used or combined to be used for battery estimation.

Fully connected neural network 204 means a fully connected neural network having an LSTM neural network output as an input. This input may include the output of each layer as well as the output of the final layer of the LSTM neural network. Through this, at least one prediction result among short, medium, and long term prediction results of time series data may be reflected in the final output. Each node has a weight, and these values are determined through iterative learning. It has one or more layers, and each layer has one or more nodes.

The estimated value 205 means a battery information estimated value that is an output of a fully connected neural network 204 having an LSTM neural network output as an input. The battery information includes a state of charge (SOC), a remaining useful life (RUL), anomaly, and the like.

As such, the LSTM neural network may include at least one neural network layer 203, which is a collection of cells 202 that make up the LSTM neural network, and a fully connected neural network 204 having an input of the output of each cell of the neural network layer 203. Can be.

4 is a view for explaining the LSTM structure used in an embodiment of the present invention.

IoT-based Prognostics and Health Management (PHM) for an electric truck according to an embodiment of the present invention may perform failure prediction and health management by detecting abnormal symptoms based on time series data obtained from various parts of the vehicle.

To this end, the failure prediction and health management system according to an embodiment of the present invention may use a Deep Bidirectional LSTM Variational Auto Encoder (DB-LSTM-VAE) based on Long Short Term Memory (LSTM). . The failure prediction and health management system according to an embodiment of the present invention may use a multilayer structure considering at least one of short, medium and long term based on LSTM.

In an embodiment of the present invention, the failure prediction and health management system may apply a multi-layered LSTM framework to predict the state of charge (SoC) of a battery in order to more accurately predict the remaining useful life (RUL) of the battery. .

Hereinafter, the basic structure and operating principle of the LSTM used in the failure prediction and health management system will be described in an embodiment of the present invention.

Long-Short Term Memory module (LSTM) is developed from simple Recurrent Neural Networks (RNN) to have cell state and gate, and the existing slope value disappears by using gate and cell state. And long-term dependence.

The cell state, the core of the LSTM, penetrates all connected within the LSTM, manages the memory of the LSTM, utilizes gates to manage information that changes in the cell state, and the gates can add or remove cell state information.

The LSTM has three gates, each of which can be:

First, the forgetting gate may serve as a first gate of the LSTM to determine how long the cell state is maintained or reflected. Through the operation of the previous LSTM output and the current input in the sigmoid layer, the output of the forgetting gate at the value between 0 and 1 can be determined as the cell state, so that the cell state can be determined.

Second, the input gate determines that the input data is reflected in the cell state, and the input gate has two layers, and the first layer may determine to reflect the cell state as a sigmoid layer. The second layer is the Tanh layer, which can be used to adjust the values that will fit into the cell state. The output from the two layers is multiplied and reflected in the cell state, which can be completed from the past state to the current state.

Third, the output gate may be an output value of the cell state considering the flow up to now by multiplying the input value by the filter value in the sigmoid layer and the updated cell state calculated in the Tanh layer.

LSTM takes into account the current output using previously stored values through three gate and cell states and can be used to process musical, language, etc. in the form of time series data.

5 is a view for explaining a single layer LSTM structure used in an embodiment of the present invention.

Failure prediction and soundness management system according to an embodiment of the present invention can use a combination of various LSTM for prediction considering at least one of the short-term, medium-term, and long-term. LSTM model used in the embodiment of the present invention can be shown to be excellent in basically reflecting long-term and short-term memory in the data input over time.

Failure prediction and health management system according to an embodiment of the present invention can predict the battery SoC.

6 is a diagram illustrating a graph of a state of charge prediction result during an iterative learning process at full temperature in a tomography layer LSTM used in an embodiment of the present invention.

When the single layer is used, the difference between the predicted value and the actual value of the battery charge state is 1.46%, and the result is shown in FIG. 6. An embodiment of the present invention may use a multiple layer LSTM structure in which a single layer structure is changed to improve this.

7 to 9 are diagrams for explaining the multi-layer LSTM structure used in the embodiments of the present invention.

Among the experimental results for various structures of the embodiments of the present invention, the LSTM framework of the multi-layer structure shown in FIGS. 7 to 9 shows the best performance.

The structure and features of the LSTM framework with multi-layer structure will be described.

Embodiments of the present invention may use a multi-layer LSTM model to increase the predictability of at least one of short-term, medium-term, and long-term data instead of just one neural network. Predictive battery charge forecasting requires forecasting based on short-term time series data generation, and designing a model that can predict mid to long term over time may be necessary.

As shown in FIG. 7, the multiple layer LSTM structure may consist of one layer and four LSTM cells. The LSTM cell of the first layer can receive and output a total of 50 time series data (five time series data, 10 time series data, 15 time series data, and 20 time series data) from four LSTM cells. That is, the length of the input sequence may be different for each LSTM cell. The multi-layer LSTM structure according to the embodiments of the present invention may predict the final SoC (battery charge state) by inputting data of the last column of the matrix output from the first layer into a fully connected neural network. As described above, the multilayer LSTM neural network includes one neural network layer, which is a first floor including a plurality of cells, and a fully connected neural network having an input of an output of each cell of one neural network layer, and each of the plurality of cells. The received time series sensor data may receive corresponding short term data among a plurality of short term data divided in chronological order.

As shown in FIG. 8, the multiple layer LSTM structure may be composed of three layers and seven LSTM cells. The LSTM cell of the first layer can receive and output five (20 total) time series data from four LSTM cells. The LSTM cell of the second layer consists of two, and the first output can be combined to output each of the two inputs. The third layer consists of one LSTM cell, and the output of the second layer can be combined and output. That is, the viewpoint may be different for each LSTM layer. The multi-layer LSTM structure according to the embodiments of the present invention may predict the final SoC (battery charge state) by inputting data of the last column of the matrix output from the third layer into a fully connected neural network. As such, a multi-layer LSTM neural network includes a plurality of neural network layers (e.g., a first layer, a second layer, a third layer, etc.) connected in sequence and including a plurality of cells, and an output of a neural network layer last connected in the plurality of neural network layers as an input. And a fully connected neural network. Each of the plurality of cells may receive the short-term data corresponding to the received time-series sensor data among the plurality of short-term data in which the time-series sensor data are divided in time series. The plurality of neural network layers may include a next order neural network layer including a plurality of next cells which are inputted by combining outputs of two or more previous cells among a plurality of previous cells included in a previous order neural network layer among the plurality of neural network layers. have.

As shown in FIG. 9, the multiple layer LSTM structure may be composed of three layers and seven LSTM cells. The LSTM cell of the first layer can receive and output five (20 total) time series data from four LSTM cells. The LSTM cell of the second layer consists of two, and the first output can be combined to output each of the two inputs. The third layer consists of one LSTM cell, and the output of the second layer can be combined and output. The multi-layer LSTM structure according to the embodiments of the present invention is completely connected by combining data of the last column of the matrix output from the first layer, data of the last column of the matrix output from the second layer, and data of the last column of the matrix output from the third layer. (Fully Connected) Neural network can be input to predict the final SoC (battery charge state). As such, the fully connected neural network may further receive an output of the neural network layer last connected in the plurality of neural network layers and an output of at least one cell in the plurality of neural network layers.

When using the LSTM model of multi-layer structure than the single-layer LSTM, the accuracy of the battery charging state is about 2-3% higher.

For detailed experimental results, see the following section. The data used in the experiment is as follows.

The battery data used in the experiment was conducted based on data of experiments on changes in four operating environments of the battery charge and discharge provided by the Center for Advanced Life Cycle Engineering (CALCE) of Marland University (https: // calceumdedu /). .

The battery data is provided as the data tested under four operating conditions. The four operating environments are as follows.

-BJDST: Beijing Dynamic Stress Test

-DST: Dynamic Stress Test

FUDS: Federal Urban Driving Schedule

-US06: Highway Driving Schedule

The battery data used for training is 162,081 and the battery data used for testing is 25,932.

The information of the battery used in the data is shown in the following [Table 1].

10 is a configuration diagram for explaining the configuration of the failure prevention and health management device of the vehicle according to an embodiment of the present invention.

As illustrated in FIG. 10, the failure prediction and health management apparatus 300 according to the embodiments of the present invention includes a communication unit 310, a memory 320, and a processor 330. However, not all illustrated components are essential components. The failure prediction and health management system may be implemented by more components than the illustrated components, and the failure prediction and health management system may be implemented by fewer components.

Hereinafter, specific configurations and operations of the components of the failure prediction and health management apparatus of FIG. 10 will be described.

The communication unit 310 communicates with a module installed in the component of the vehicle. Here, the module may be implemented as an IoT module.

The memory 320 stores at least one program. The memory 320 may store time series sensor data received from the module. The memory 320 may store the parameters of the regression neural network.

The processor 330 is connected to the communication unit 310 and the memory 320.

By executing at least one program, the processor 330 receives time series sensor data from a module installed in a component of a vehicle, and uses the received time series sensor data to calculate a state of a part or a regression neural network for detecting an abnormal symptom ( Recurrent Neural Network) parameters are updated, and the regressive neural network to which the updated regressive neural network parameters are applied is used to estimate the state of parts of the vehicle or to detect abnormal symptoms.

According to embodiments, the processor 330 may update a parameter of the regression neural network by using a data sequence obtained by mixing at least one of short, medium, and long term data of the time series sensor data.

According to embodiments, the regression neural network may include a state of charge (SOC), a state of health (SOH), a power measurement, a voltage measurement, a current measurement, and a temperature measurement of a component. An input vector including at least one may be received through the input layer.

According to embodiments, the regression neural network may be a long short-term memory (LSTM) neural network.

According to embodiments, the LSTM neural network includes at least one neural network layer, which is a collection of cells constituting the LSTM neural network, and a fully connected neural network having an input of an output of each cell of the neural network layer. It may include.

According to embodiments, the LSTM neural network may include one neural network layer including a plurality of cells and a fully connected neural network having an input of an output of each cell of one neural network layer. Each of the plurality of cells may receive short-term data corresponding to the plurality of short-term data in which the received time-series sensor data are classified in chronological order.

According to embodiments, the LSTM neural network may include a plurality of neural network layers including a plurality of cells and sequentially connected, and a fully connected neural network having an output of a neural network layer last connected in the plurality of neural network layers. Each of the plurality of cells may receive short-term data corresponding to the plurality of short-term data in which the received time-series sensor data are classified in chronological order. The plurality of neural network layers may include a next order neural network layer including a plurality of next cells that receive inputs of two or more previous cells from among a plurality of previous cells included in the previous order neural network layers among the plurality of neural network layers. have.

According to embodiments, the fully connected neural network may further receive an output of a neural network layer last connected in the plurality of neural network layers and an output of at least one cell in the plurality of neural network layers.

11 and 12 are flowcharts of data learning and data testing performed in a failure prediction and health management system according to an exemplary embodiment of the present invention.

Let's look at the experiment for performance evaluation.

Learning and testing of the framework for performance evaluation may proceed in the same manner as in FIGS. 10 and 11.

The model for predicting the state of charge of a battery using a multilayer LSTM is classified into a process of learning data and a data test process of verifying the state of predicting the state of charge of a battery.

The learning process illustrated in FIG. 11 will be described.

In step S101, the failure prediction and health management system reads a file containing battery data and then converts at least one of current, voltage, and temperature into one data.

In step S102, the failure prediction and health management system declares and initializes the variables to be used.

In step S103, the failure prediction and health management system designates an epoch to one.

In step S104, the failure prediction and health management system learns for a predetermined number of times (e.g., 10,000 times).

In step S105, the failure prediction and health management system compares the loss function values representing the difference between the predicted value and the actual value during the learning process.

In step S106, the failure prediction and health management system stores a learning state of a value having a low loss function value.

In step S107, the failure prediction and health management system checks whether there is the lowest validation error.

In step S108, the failure prediction and health management system stores the lowest validation error state.

In step S109, the failure prediction and soundness management system checks whether the number of times of learning execution in the learning process is a multiple of a predetermined number (for example, 500).

In step S110, the failure prediction and health management system stores the learning state when the number of learning trials is a multiple of a preset number (eg, 500).

In step S111, the failure prediction and soundness management system checks whether the number of times of learning execution in the learning process is a multiple of a predetermined number (eg, 10,000).

In step S112, the failure prediction and health management system adds 1 to the epoch and performs step S104, if the number of learning trials is not a multiple of a preset number (e.g., 10,000).

Thereafter, the failure prediction and health management system ends when the predetermined number of times (eg, 10,000 times) learning is completed.

A test process illustrated in FIG. 12 will be described.

In step S201, the failure prediction and health management system reads the test data from the battery data and then converts the current, voltage, and temperature into one data.

In step S202, the failure prediction and health management system reads a stored state of a value having a low loss function value.

In step S203, the failure prediction and health management system stores the result value after the test proceeds in a state where the read loss function value is stored at a low value.

In step S204, the failure prediction and health management system reads a learning state of 500 multiples.

In step S205, the failure prediction and health management system stores the result value after the test proceeds.

In step S206, the failure prediction and health management system checks whether it is in the last test state.

The failure prediction and health management system terminates when testing at all 500 multiples of learning is complete.

The experiment conducted as above was performed on the assumption that the loss function value was excellent, but the prediction result was not excellent in the test. When the experiment was performed on the embodiments of the present invention, the test when the loss function value having the excellent learning state was the lowest. Was not better than the lowest error among the learning states of multiples of 500.

13 to 16 are diagrams for explaining an excellent charge state prediction result graph during an iterative learning process at each temperature for embodiments of the present invention.

The graphs shown in FIGS. 13 to 16 express the difference between the battery remaining charge value and the battery charge state predicted value at 0 ° C., 25 ° C., 45 ° C., and all temperatures, and the difference between each value is the mean square root error ( RMSE: Root Mean Square Error

FIG. 13 shows a graph of excellent state of charge prediction results during an iterative learning process at 0 ° C.

FIG. 13 is a graph comparing the predicted result and the actual value at the excellent prediction state at 0 ° C. during the test. The difference between the actual value and the predicted value when the learning was 10,000 was 1.49%, and the largest error. The difference represents 0.29%.

Figure 14 shows a graph of the excellent state of charge prediction results during the iterative learning process at 25 ℃.

FIG. 14 is a graph comparing the predicted result and the actual value at the excellent prediction condition at 25 ° C. during the test. When the learning is 6,500, the difference between the actual value and the predicted value is 0.75% and the largest error. The difference represents 0.08%.

Figure 15 shows a graph of the excellent state of charge prediction results during the iterative learning process at 45 ℃.

FIG. 15 is a graph comparing the predicted result and the actual value at the excellent prediction state at 45 ° C. during the test. The difference between the actual value and the predicted value when the learning is 8,000 is 0.5%, and the largest error The difference represents 0.05%.

16 shows a graph of the result of excellent charge state prediction during the iterative learning process at all temperatures.

FIG. 16 is a graph comparing the predicted result and the real value in the state of good prediction at all temperatures of 0 ° C., 25 ° C. and 45 ° C., and the comparison of the real value and the predicted value when the learning is 6,500 is 1.26%. The largest error difference is 0.47%.

At 0 ° C, the lowest difference between the battery charge value and the predicted value represents a difference of 1.49% of the RMSE as tested on 10,000 trained data, and the lowest difference between the battery charge and the predicted value at 25 ° C is 6,500 trained data. The lowest difference between RMSE 0.77% and the battery charge and predicted value at 25 ° C was tested on 8,000 trained data, showing a RMSE 0.5% difference.

The lowest difference between the battery state of charge value and the predicted value at all temperatures of 0 ℃, 25 ℃ and 45 ℃ was the smallest difference tested in the model trained 6,500 times, representing a difference of RMSE 126%.

Table 2 shows the comparison of SoC estimation errors in six studies.

The experimental results in Table 2 are “Long Short-Term Memory-Networks for Accurate State of Charge Estimation of Li-ion Batteries (Ephrem Chemali, Student Member, IEEE, Phillip J Kollmeyer, Member, IEEE, Matthias Preindl, Member, IEEE) , Ryan Ahmed, Member, IEEE, and Ali Emadi, Fellow, IEEE) ”, the error measurements are different from each other in comparison with the numerical values compared with each other, but it is difficult to make a simple comparison. Experimental results according to an embodiment of the present invention. RMSE error of 0.7% at 25 ° C is lower than 0.75% RMSE of 25 ° C of the embodiment of the present invention, but the above study did not reflect the state of charge of the battery. Therefore, the result of the embodiment of the present invention reflecting the charge and discharge in the experiment is close to the actual battery operating environment, showed excellent results, and also can be predicted in various temperature conditions by performing the experiment at a temperature difference of 0 ~ 45 ℃ It can have

Looking at the implementation of the IoT module, for the IoT-based PHM for the electric truck according to an embodiment of the present invention, the IoT module that can be mounted on the vehicle to perform the estimation of the state of the vehicle parts through the sensor data in the vehicle. This module must be able to support CAN Bus to communicate with the VCU in the vehicle to receive the sensor's measurements, and through the local gateway to transfer the vehicle's data to the Cloud Server. ) And broadband connection.

To this end, in an embodiment of the present invention, various necessary performances such as sufficient calculation performance for PHM, CAN communication capable of communicating with the VCU, and wireless communication with a cloud server may be satisfied.

17 is a view for explaining the overall configuration of the IoT-based PHM system according to an embodiment of the present invention.

IoT based failure prediction and health management system according to embodiments of the present invention efficiently detects each component through IoT-based real-time abnormality detection and long-term failure prediction of the status of each component mounted on the vehicle (eg electric truck) By maintaining and replacing, you can ensure the reliability of the electric truck you are developing. Here, embodiments of the present invention relates to a deep learning-based abnormal symptom detection and long-term failure prediction system that combines a model-based and data-driven approach to ensure reliability. This means that the information about the state of each part of the vehicle generated in real time can be processed more effectively and efficiently by utilizing the deep learning which is emerging as a big data with a large amount of data.

The PHM system that implements the IoT-based fault prediction technology proposed through the embodiments of the present invention is composed of each component of the vehicle, an IoT module mounted on the vehicle, and a cloud server as shown in FIG. 17. Same as

Each part of the vehicle is to be maintained and repaired to ensure the reliability of the vehicle, and transmits information about their status to the IoT module in real time. This information is accumulated to form big data about the state of the vehicle.

The IoT module installed in the vehicle transmits big data transmitted from each component of the vehicle to the cloud server, and transmits abnormal sign detection and warning messages of the components to the digital cluster of the vehicle and the user's smartphone based on real-time data. do.

The cloud server may store big data information of each component transmitted from the IoT module through the Internet and transmit failure prediction information considering the long term to the user's smartphone. Based on the big data information stored in the cloud server, the deep learning network in the failure prediction system may be periodically learned. The weight of the deep learning network located in the abnormal symptom detection and warning system of the IoT module may be updated using the weight of the learned network.

Details on the development of the above-mentioned components are as follows.

The embodiment of the present invention may include an IoT-based intelligent failure prediction technology that can determine the vehicle inspection time by analyzing the driving pattern and the power load pattern of the vehicle.

In an embodiment of the present invention, an IoT module for collecting sensor data for measuring various parts states and operation / environmental conditions of an electric truck may be included.

Embodiments of the present invention include component-wise, subsystem-wise, and system-wise anomalies based on collected data, including point anomalies, collective anomalies, and contextual anomalies. Real time detection and warning system may be included.

In an embodiment of the present invention, a system for analyzing a failure mode / mechanism by connecting a fault indication result and a fault tree analysis (FTA) may be included.

Based on the results, there is an integrated PHM system that enables component-wise, subsystem- and system-wise fault prediction and remaining useful lifetime (RUL) prediction. May be included.

In an embodiment of the present invention, the selection of parts and subsystems to which the PHM is applied may be based on Critical to Quality (CTQ) and Critical to Customer (CTC).

In an embodiment of the present invention, a deep learning network with a new structure as a core engine for implementing PHM functions through anomaly indication and fault prediction including point anomalies, collective anomalies, and contextual anomalies Deep Bidirectional Long-Short Term Memory Variational Auto Encoder Network (DB-LSTM-VAE).

According to an embodiment of the present invention, vehicle driving records and simulation-based data are collected through an IoT module and a cloud server connected to the Internet, and online fine tuning and performance improvement of a Deep Learning Network-based PHM system may be included.

In an embodiment of the present invention, anomaly detection, failure mode / mechanism analysis, failure prediction, and RUL prediction of component-wise, subsystem-wise, and system-wise integrated PHM systems are provided. Simulation and real data based performance evaluation.

Bidirectional Long-Short Term Memory (BLSTM) has been successfully modeled for changes in temporal dynamics of input sequences, and successfully extracts relevant features under variable period conditions. can do. These extracted features can be used for accurate temporal modeling and prediction. Because bidirectional LSTM structures learn both forward and reverse sequential dependencies, Bidirectional LSTM structures can outperform LSTM structures in prediction accuracy.

In order to extract features having various levels of locality / abstractions / temporal-period, in the embodiment of the present invention, the framework includes a DB-LSTM Decoder for restoring and predicting future data. Bidirectional Long Short Term Memory (DB-LSTM) Variational Auto Encoder can be implemented. Variational Auto Encoders (VAE) are well suited for non-supervised learning to extract high-abstract features from raw images. In an embodiment of the present invention, this VAE can be extended by combining with DB-LSTM as follows.

Data accumulated in one cycle is created as an input image.

Dynamic characteristics of the cross-section time-series of the extracted features are modeled using DB-LSTM with convolutional layer.

Anomaly can be divided into Point, Collective and Contextual Anomaly. These anomalies are detected based on statistical characteristics obtained from features extracted from learning of LSTM-VAE and Latent Space Association VAE (LSA-VAE) frameworks with varying time windows.

LSA-VAE and DB-LSTM-VAE are combined to detect signs of contextual abnormality.

Make predictions for long cycles using DB-LSTM-VAE.

In an embodiment of the present invention, abnormal indication detection for the battery PHM may be included.

An anomaly is defined as an abnormal operation in which a part does not meet its normal characteristics, and a failure is not necessarily a failure, but all failures are caused by an anomaly.

Anomalies can be divided into three types: point, collective and contextual anomalies.

Point anomaly sign means that there is a variable that is out of the normal operation range at the present moment.

Collective anomaly means that the pattern of variables over a period of time is similar to the anomaly pattern.

Contextual anomaly means that the pattern of variables in a given context condition is similar to the anomaly pattern.

Because anomalies are unexpected and not frequent, there is a small amount of prior data on all anomalies, and detection of anomalies under these conditions is a challenging task. The statistical characteristics of the feature extracted through all the learning are stored, and the stored statistical characteristic is used to detect the feature of the abnormality extracted from the time interval representing the abnormality.

Abnormal Signs You can define different types of abnormal signs depending on the depth of the feature. For example, abnormal signs detected in the shallow layer correspond to Point abnormal signs, while abnormal signs detected in the Deep Layer correspond to Collective abnormal signs. LSA-VAE is used to create a high-level feature space under conditions of environment and operation, and uses the statistical characteristics of features extracted based on this environmental condition to detect Contextual anomalies.

An anomaly detection algorithm for lithium-ion batteries will be developed and applied.

Let's look at the IoT module implementation.

A computational platform is needed to store the computational load and sensor data needed to process the sensor data. These platforms not only need to be able to receive sensor measurements via the CAN Bus, but also the VCU and Cloud Servers using local area networks (LANs) and broadband connections through local gateways. Each must be able to communicate with each other. To this end, an embodiment of the present invention intends to use an IoT module, which is a platform having sufficient computational performance, which can collect, store, process, and transmit data to a VCU and a cloud server. IoT-based PHMs for electric trucks not only provide predictive capabilities for IoT-based driver assistance, but also provide user services for electric truck companies.

18 shows a conceptual diagram of an IoT module in an EV's PHM system.

We will be experimenting with the development toolkit. In this experiment, we plan to apply artificial sensor data to the toolkit to evaluate the performance of the system. The right figure below is a conceptual diagram of IoT module experiment, and the table below is a comparison of various systems that can be used as IoT modules.

Let's look at the integration of IoT modules with cloud servers.

The IoT module and cloud server will integrate the PHM framework available and will conduct experiments based on simulation data. At this time, the IoT module collects the measured sensor data and performs anomaly detection in real time. The IoT module also transmits the collected data to the cloud server for failure mode / mechanism analysis, failure prediction, data logging, and fine tuning of the DB-LSTM-VAE via PHM.

Let's look at the development of algorithms for fault prediction.

DB-LSTM-VAE was developed to model the characteristics of measurements aimed at extracting features for the detection of anomalies, and DB-LSTM- was developed to predict failure prediction criteria such as State Of Health (SOH). Expand your VAE network. To do this, re-learn the decoder of DB-LSTM, and perform fault deposit by reconfiguring future intervals rather than reconfiguring the input intervals as shown below.

FIG. 19 is a diagram illustrating an extension of DB-LSTM-VAE for fault prediction according to an embodiment of the present invention. FIG.

Let's look at completing a PHM system for a battery.

Since the internal state variables of a lithium-ion battery cannot be measured with an external sensor, the CBM / PHM of a lithium-ino battery is not an easy task. In order to perform CBM / PHM, values of State of Charge (SOC), State of Health (SOH), Remaining Useful Life (RUL), and Time of Failure (ToF) must be estimated. Although recent Battery Management Systems (BMS) are able to estimate SOC and SOH, it is still difficult to accurately predict RUL.

In an embodiment of the present invention, the DB-LSTM-VAE is used to model the dynamic characteristics of the battery PHM over time and to predict future changes in the variables of the battery PHM, particularly SOH. This method has a variable structure that can predict various potential spaces, and calculate the predicted mean and variance. Probability estimation of Remaining Useful Life (RUL) is performed using predicted future changes based on Expert-defined failure threshold. The table below shows the input and output of the PHM for the proposed lithium-ion battery.

Table 3 shows the input and output of PHM for lithium-ion batteries.

Let's look at the PHMs for chargers / motors 1 and 2.

The DB-LSTM-VAE is trained to detect abnormal signs of PHM for chargers and motors, analyze failure modes / mechanisms, and predict failures. The inputs and outputs are as follows.

Table 4 shows the input and output of the PHM for the charger / mutter 1, 2.

Due to the complexity of the electric truck system, it is very difficult to model the electric truck system from the raw data of the individual parts, which means that the PHM of the electric truck system has a very low model dependency on the individual parts' PHM. To this end, an embodiment of the present invention represents a system in a hierarchical structure with sub-systems defined as Critical to Quality (CTQ) and Critical to Customer (CTC), and features for groups of parts and feature pools of sub-systems. Deep Ensemble Fusion is applied to fuse the pool. The feature pool of this subsystem is then combined with the feature pool of other subsystems to form the PHM feature pool of the electric truck.

20 is a diagram for describing a deep ensemble fusion configuration diagram having a hierarchical structure.

Proceed with learning about PHM of the following subsystems.

Electrical subsystem: battery and charger

Motor Subsystem: Motor 1 and Motor 2

Let's look at the completion of a PHM subsystem for an electric truck.

In an embodiment of the present invention, component-based PHM is performed through DB-LSTM_VAE learning on an air conditioner, an air brake, a power take-off (PTO), and an electric power steering (EPS) component. And the PHM of the subsystem may be included through learning about the PHM fusion of the control subsystem.

In addition, a new Attention / Affordance mechanism for deep ensemble fusion to enhance the PHM performance of subsystems and electric trucks by fusing only the necessary features of the subsystem and the electric truck system. This may be included.

In an embodiment of the present invention, a PHM for an air conditioner / brake / PTO / EPS may be included.

DB-LSTM-VAE is trained to detect abnormal signs of PHM for air conditioner / brake / PTO / EPS, failure mode / mechanism analysis, and predictive failure, where input and output are as follows.

[Table 5] shows the input / output of PHM for air conditioner / brake / PTO / EPS.

In an embodiment of the present invention, a PHM for a control subsystem may be included.

Apply Attention / Affordance-based Deep Ensemble Fusion technique to fuse air conditioning / brake / PTO / EPS with the control subsystem and cross-section time-series features for the PHM of the control subsystem Create a cross-section time-series feature pool.

In an embodiment of the present invention, an algorithm for fine tuning of self-generated data and externally collected data-based DB-LSTM-VAE may be included.

As mentioned earlier, the presence of anomalies does not necessarily cause a breakdown, but an anomaly means a new characteristic that was not seen during learning. In order to improve the performance of the framework developed in the embodiment of the present invention, a real-time learning function may be used. Through real-time learning, new measurements are defined and used for self-learning. Moreover, the DB-LSTM-VAE can be built on a cloud server because the cloud-based server records the PHM data obtained from the operation of the electric truck / automatic offline learning / regular updates. Data from the various data sources mentioned above is constantly updated and accumulated, and the parameters of the DB-LSTM-VAE network can be enhanced through gradual fine tuning and the network performance can be improved.

FIG. 21 illustrates a method for fine tuning self-generated data and data-based DB-LSTM-VAE collected externally.

In the embodiment of the present invention, the overall PHM system for an electric truck, implements fine-tuning of the cloud-based DB-LSTM-VAE, and can evaluate the performance of the PHM developed in the actual test-bed.

The failure prediction and health management method for the parts of the vehicle according to the embodiments of the present invention described above may be implemented as computer readable codes on a computer readable recording medium. Failure prediction and health management method for the parts of the vehicle according to the embodiments of the present invention may be implemented in the form of program instructions that can be executed by various computer means may be recorded on a computer-readable recording medium.

A non-transitory computer readable storage medium comprising at least one program executable by a processor, wherein the at least one program, when executed by the processor, causes the processor to: time-series sensor data from a terminal installed on a component of a vehicle; Receive a, update the parameters of the recurrent neural network for the estimation of the state of the component or the detection of abnormal symptoms using the received time series sensor data, and using the regression neural network to which the parameters of the updated regression neural network is applied A non-transitory computer readable storage medium can be provided that includes instructions to estimate a state of a part of the vehicle or to detect an abnormal symptom.

Computer-readable recording media include all kinds of recording media having data stored thereon that can be decrypted by a computer system. For example, there may be a read only memory (ROM), a random access memory (RAM), a magnetic tape, a magnetic disk, a flash memory, an optical data storage device, and the like. The computer readable recording medium can also be distributed over computer systems connected over a computer network, stored and executed as readable code in a distributed fashion.

As described above with reference to the drawings and examples, it does not mean that the scope of protection of the present invention is limited by the above drawings or embodiments, and those skilled in the art are skilled in the art It will be understood that various modifications and variations can be made in the present invention without departing from the spirit and scope.

Specifically, the described features may be implemented within digital electronic circuitry, or computer hardware, firmware, or combinations thereof. The features may be executed in a computer program product implemented in storage in a machine readable storage device, for example, for execution by a programmable processor. And features may be performed by a programmable processor executing a program of instructions to perform functions of the described embodiments by operating on input data and generating output. The described features include at least one programmable processor, at least one input device, and at least one output coupled to receive data and directives from a data storage system, and to transmit data and directives to a data storage system. It can be executed in one or more computer programs that can be executed on a programmable system comprising a device. A computer program includes a set of directives that can be used directly or indirectly within a computer to perform a particular action on a given result. A computer program is written in any form of programming language, including compiled or interpreted languages, and included as a module, element, subroutine, or other unit suitable for use in another computer environment, or as a standalone program. Can be used in any form.

Suitable processors for the execution of a program of instructions include, for example, both general purpose and special purpose microprocessors, and one of a single processor or multiple processors of another kind of computer. Computer program instructions and data storage devices suitable for implementing the described features are, for example, magnetic memory such as semiconductor memory devices, internal hard disks and removable disks such as EPROM, EEPROM, and flash memory devices. Devices, magneto-optical disks and all forms of non-volatile memory including CD-ROM and DVD-ROM disks. The processor and memory may be integrated in application-specific integrated circuits (ASICs) or added by ASICs.

Although the present invention described above has been described based on a series of functional blocks, the present invention is not limited to the above-described embodiments and the accompanying drawings, and various substitutions, modifications, and changes without departing from the technical spirit of the present invention. It will be apparent to one of ordinary skill in the art that this is possible.

Combinations of the above-described embodiments are not limited to the above-described embodiments, and various types of combinations as well as the above-described embodiments may be provided according to implementation and / or need.

In the above-described embodiments, the methods are described based on a flowchart as a series of steps or blocks, but the present invention is not limited to the order of steps, and any steps may occur in a different order or at the same time than the other steps described above. have. Also, one of ordinary skill in the art appreciates that the steps shown in the flowcharts are not exclusive, that other steps may be included, or that one or more steps in the flowcharts may be deleted without affecting the scope of the present invention. I can understand.

The foregoing embodiments include examples of various aspects. While not all possible combinations may be described to represent the various aspects, one of ordinary skill in the art will recognize that other combinations are possible. Accordingly, the invention is intended to embrace all other replacements, modifications and variations that fall within the scope of the following claims.

Although described above with reference to the drawings and embodiments, it does not mean that the scope of protection of the present invention is limited by the above drawings or embodiments, and those skilled in the art to the spirit of the present invention described in the claims It will be understood that various modifications and variations can be made in the present invention without departing from the scope of the invention.

10: Failure Prediction and Health Management System
11: means of transport
12: IoT Module
13: cloud server
101: battery
102: BMS
110: IoT module
111: preprocessing unit
112: communication unit
113; Estimator
120: cloud server
121: communication unit
122: data storage
123: learning unit
124: parameter storage unit
300: failure prediction and health management device
310: communication unit
320: memory
330: processor

Claims (18)

In the failure prevention and soundness management method of the transportation means,
Receiving time series sensor data from a module installed in the component of the vehicle;
Updating a parameter of a recurrent neural network using the received time series sensor data to estimate a state of the part or detect an abnormal symptom; And
Estimating a state of a part of the vehicle or detecting an abnormal symptom using the regression neural network to which the parameters of the updated regression neural network are applied. The method of claim 1,
Updating the parameter,
And updating a parameter of a regression neural network using a data sequence in which at least one of the short-term, medium-term and long-term data of the time series sensor data is mixed. The method of claim 1,
The regression neural network,
An input vector including at least one of a state of charge (SOC), a state of health (SOH), a power measurement, a voltage measurement, a current measurement, and a temperature measurement of the component Means of failure prevention and health management of vehicles received through the input layer. The method of claim 1,
The regression neural network,
Long Short-Term Memory (LSTM) Neural network, characterized in that the failure failure prevention and health management method of the vehicle. The method of claim 4, wherein
The LSTM neural network,
Predicting failure of a vehicle, comprising at least one neural network layer, which is a collection of cells constituting an LSTM neural network, and a fully connected neural network having as its input the output of each cell of the neural network layer; and How to manage health. The method of claim 4, wherein
The LSTM neural network includes one neural network layer including a plurality of cells and a fully connected neural network having an input of an output of each cell of the one neural network layer,
Each of the plurality of cells receives the short-term data corresponding to the received time-series sensor data from the plurality of short-term data, which are divided in chronological order. The method of claim 4, wherein
The LSTM neural network includes a plurality of neural network layers sequentially connected to each other and a plurality of neural network layers connected to each other, and a fully connected neural network having an output of a neural network layer last connected to the plurality of neural network layers.
Each of the plurality of cells receives short-term data corresponding to a plurality of short-term data in which the received time-series sensor data are divided in time-series order,
The plurality of neural network layers includes a next order neural network layer including a plurality of next cells which are inputted by combining outputs of two or more previous cells among a plurality of previous cells included in a previous order neural network layer among the plurality of neural network layers. Failure prevention and soundness management method of the means of transportation. The method of claim 6,
The fully connected neural network,
And an output of a neural network layer last connected in the plurality of neural network layers and an output of at least one cell in the plurality of neural network layers. A communication unit for communicating with a module installed in a component of the vehicle;
A memory for storing at least one program; And
A processor connected to the communication unit and the memory,
The processor executes the at least one program,
Receive time series sensor data from modules installed in parts of vehicles,
Using the received time series sensor data to update a parameter of a recurrent neural network for estimating the state of the part or detecting an abnormal symptom,
A device for predicting failure and soundness of a vehicle using the regression neural network to which the parameters of the updated regression neural network are applied, to estimate a state of a part of the vehicle or to detect an abnormal symptom. The method of claim 9,
The processor,
And updating the parameters of the regression neural network using a data sequence obtained by mixing at least one of short-term, medium-term and long-term data of the time series sensor data. The method of claim 9,
The regression neural network,
An input vector including at least one of a state of charge (SOC), a state of health (SOH), a power measurement, a voltage measurement, a current measurement, and a temperature measurement of the component Failure prediction and soundness management device of the vehicle, received through the input layer. The method of claim 9,
The regression neural network,
Long short-term memory (LSTM) Neural network, characterized in that the failure prevention and health management device of the vehicle. The method of claim 12,
The LSTM neural network,
Predicting failure of a vehicle, comprising at least one neural network layer, which is a collection of cells constituting an LSTM neural network, and a fully connected neural network having as its input the output of each cell of the neural network layer; and Health Management Device. The method of claim 12,
The LSTM neural network includes one neural network layer including a plurality of cells and a fully connected neural network having an input of an output of each cell of the one neural network layer,
Each of the plurality of cells receives the short-term data corresponding to the plurality of short-term data in which the received time-series sensor data are divided in time-series order. The method of claim 12,
The LSTM neural network includes a plurality of neural network layers sequentially connected to each other and a plurality of neural network layers connected to each other, and a fully connected neural network having an output of a neural network layer last connected to the plurality of neural network layers.
Each of the plurality of cells receives short-term data corresponding to a plurality of short-term data in which the received time-series sensor data are divided in time-series order,
The plurality of neural network layers includes a next order neural network layer including a plurality of next cells which are inputted by combining outputs of two or more previous cells among a plurality of previous cells included in a previous order neural network layer among the plurality of neural network layers. Failure prevention and soundness management device of the vehicle. The method of claim 15,
The fully connected neural network,
And an output of a neural network layer last connected in the plurality of neural network layers and an output of at least one cell in the plurality of neural network layers. A module for collecting and transmitting time series sensor data obtained from parts of the vehicle; And
Using the time series sensor data received from the module to update the parameters of the recurrent neural network for estimating the state of the part or detecting abnormal symptoms, and transmits the parameters of the updated regression neural network to the module, wherein the time series It includes a cloud server for updating the parameters of the regression neural network using a data sequence of at least one of the short-term, medium-term and long-term data of the sensor data,
Wherein the module estimates the state of a part of the vehicle or detects an abnormal symptom using the regression neural network to which the parameters of the updated regression network are applied. A non-transitory computer readable storage medium comprising at least one program executable by a processor, wherein the at least one program, when executed by the processor, causes the processor to:
Receive time series sensor data from modules installed in parts of vehicles,
Using the received time series sensor data to update a parameter of a recurrent neural network for estimating the state of the part or detecting an abnormal symptom,
And instructions for estimating a state of a part of the vehicle or detecting an abnormal symptom using the regression neural network to which the parameters of the updated regression neural network are applied.

Images