Degradation level prediction method and degradation level prediction system

Abstract

A degradation level of a noise source is predicted by performing deep learning using frequency data indicative of vibration of the noise source during a certain period and degradation data indicative of degradation of the noise source during a predetermined period to determine a state estimation model. During a monitoring period, a first vibration signal is measured at a noise point. The first vibration signal is converted into a frequency domain, and the frequency domain converted first vibration signal is input into a state estimation model to estimate time series degradation data of degradation of the noise source during the monitoring period. A degradation prediction model is determined by deep learning using the time series degradation data. Degradation data indicative of predicted degradation of the noise source during a prediction period is predicted via the degradation prediction model. And the residual service life of the noise source can be calculated.

Description

Technical Field

[0001] The present disclosure provides a degradation level prediction method and a degradation level prediction system. Background Art

[0002] In recent years, the comfort of vehicle drivers has been recognized as important, and the noise and vibration experienced by the driver may play an important role in the marketability of the vehicle. In a vehicle, various noises and / or vibrations may be generated.

[0003] In particular, when a vehicle drive system deteriorates, the degradation may cause noise and / or vibration and / or changes in noise and / or vibration. It is useful to determine (eg, detect) to address the degradation (eg, replace, update and / or repair parts). Summary of the invention

[0004] The following summary briefly summarizes certain features. The summary is not an extensive overview and is not intended to identify key or critical elements.

[0005] Systems, devices, and methods for degradation level prediction are described. A method may include performing deep learning to determine a state estimation model by the following steps: using multiple frequency data items indicating multiple vibrations of a noise source during a predetermined period and multiple degradation data items indicating corresponding degradation degrees of the noise source during the predetermined period; measuring multiple first vibration signals from the noise source at a noise point during a monitoring period; estimating time series degradation data indicating the degree of degradation of the noise source during the monitoring period by inputting multiple first frequency data items obtained by converting the multiple first vibration signals into the frequency domain into the state estimation model; determining a degradation prediction model by performing deep learning using the time series degradation data; predicting degradation data indicating the degree of degradation of the noise source at at least one target time point within a prediction period later than the monitoring period based on the degradation prediction model; and calculating the remaining useful life of the noise source based on the time series degradation data and the degradation data at the target time point.

[0006] In addition, or alternatively, the degradation level prediction system may include: an input unit configured to acquire a vibration signal from a noise point configured to measure noise from a noise source; an estimation model learning unit configured to determine a state estimation model by performing deep learning using a plurality of frequency data items indicating a plurality of vibrations of the noise point during a predetermined period and a plurality of degradation data items indicating a degree of degradation of the noise source during the predetermined period; and a controller configured to measure a plurality of first vibration signals at the noise point during a monitoring period, estimate time series degradation data indicating a degree of degradation of the noise source during the monitoring period by inputting the plurality of first frequency data items into the state estimation model, wherein the first frequency data items are obtained by converting the plurality of first vibration signals into a frequency domain, predict degradation data indicating a degree of degradation of the noise source at at least one target time point within a prediction period later than the monitoring period by performing deep learning using the time series degradation data, and calculate the remaining useful life of the noise source based on the time series degradation data and the predicted degradation data at the at least one target time point.

[0007] These and other features and advantages are described in greater detail below. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 is a block diagram schematically illustrating a degradation level prediction system according to an example.

[0009] Figure 2 is a diagram showing one example of a state estimation model in the degradation level prediction system according to this example.

[0010] Figure 3 It is shown Figure 2 An example of the output of the state estimation model is shown in the figure.

[0011] Figure 4 is a diagram showing one example of a degradation prediction model in the degradation level prediction system according to this example.

[0012] Figure 5 It is shown Figure 4 A diagram of an example of multiple LSTM cells shown in .

[0013] Figure 6 It is shown Figure 1 The detailed configuration of the controller is shown in the diagram.

[0014] Figure 7 is a diagram showing one example of a graph illustrating data generated by the degradation level prediction system according to this example.

[0015] Figure 8 is a flow chart of a degradation level prediction method according to an example. DETAILED DESCRIPTION

[0016] Hereinafter, the examples in this specification will be described in detail with reference to the accompanying drawings. The same or similar constituent parts are represented by the same or similar reference numerals, and will not be described repeatedly. As used herein, the suffix "module" and / or the suffix "unit" of the constituent elements are given and used interchangeably only for the convenience of compiling this specification, and the suffixes themselves do not have meanings or functions that can be distinguished from each other. In addition, when describing the examples disclosed in this specification, if it is determined that its specific description may confuse the main points of the examples disclosed in this specification, the detailed description of the known technology will be omitted. In addition, the drawings are provided to help easily understand the examples disclosed in this specification, and the technical spirit disclosed in this specification is not limited by the drawings, and it should be understood that the present disclosure includes all modifications, equivalents and substitutes contained in the spirit and technical scope of the present disclosure.

[0017] Terms including ordinal numbers, such as first and second, are used to describe various constituent elements, but these constituent elements are not limited by these terms. These terms are used only to distinguish one constituent element from another constituent element.

[0018] In the present application, it should be understood that the terms "include", "including" and "having" are intended to specify the existence of the characteristics, numbers, steps, operations, constituent elements and parts or their combinations described in the specification, and do not exclude the possibility of pre-existence or addition of one or more other characteristics, numbers, steps, operations, constituent elements, parts or their combinations.

[0019] In a component for controlling other components under specific control conditions according to one example, a program implemented as a command set of a control algorithm required to control other components may be installed. The control component may generate output data by processing input data and / or stored data according to a program (e.g., a stored and / or installed program). The control component may include a non-volatile memory for storing the program and a memory for storing data. Instructions for implementing the program may be stored in a non-transitory computer-readable storage medium (e.g., in a non-volatile memory).

[0020] Disclosed herein are methods, apparatus and / or systems for determining and / or predicting future degradation of one or more components of a vehicle and / or for determining and / or predicting the remaining useful life of a vehicle drive system.

[0021] Figure 1 is a block diagram schematically illustrating a degradation level prediction system according to an example.

[0022] The degradation level prediction system 10 according to an example of the present disclosure may include a computer system for implementing various examples to be described in this specification. Specifically, the degradation level prediction system 10 may include at least one processor and a memory, and the memory may contain computer-readable instructions, and at least one processor may be configured to execute the instructions stored in the memory, so that various examples to be described in this specification may be implemented.

[0023] A "processor" may refer to a device for processing arithmetic operations, logical operations, determination operations, etc. to provide at least one function. The processor may be implemented in the form of hardware, software, or a combination of hardware and software. For example, the processor may be implemented in the form of software such as tasks, classes, subroutines, processes, objects, execution threads, and programs that can be executed in a predetermined area of ​​a memory, or in the form of hardware such as a field programmable gate array (FPGA) and an application specific integrated circuit (ASIC), or in the form of a combination of software and hardware. The processor may be contained on a computer-readable storage medium, and / or may be distributed on multiple computers.

[0024] For example, the degradation level prediction system 10 may be implemented in the form of a supercomputer, a server, a mainframe, a workstation, a personal computer, a laptop, a tablet computer, an embedded computer, a wearable computer, etc. The degradation level prediction system 10 may include an output device and / or component configured to output one or more results (e.g., an indication and / or notification of the degradation level and / or life of a vehicle component), as discussed herein. However, the scope of the present disclosure is not limited thereto, and the degradation level prediction system 10 may be any computing device capable of implementing the various functions / methods described in this specification.

[0025] The degradation level prediction system 10 may be implemented in the form of a single computing device, or may be implemented in the form of a plurality of computing devices.

[0026] If the vehicle is driven, noise may be generated in various parts. At least one noise generating point in the vehicle (hereinafter referred to as "noise source") may indicate various points in the vehicle where noise may be generated, such as parts of the vehicle drive system, and for example, the noise source may be a motor reducer, brake, tire, etc. of the vehicle.

[0027] Hereinafter, for convenience of explanation, a point located within a predetermined range (eg, within a predetermined distance) from a position of a part corresponding to a noise source (eg, a motor reducer of a vehicle, etc.) will be referred to as a “noise point”.

[0028] In this specification, the degradation level prediction system 10 is described as predicting the degradation level of a vehicle, but the present disclosure is not limited thereto. The degradation level prediction system 10 may be used to predict the degradation level of any other motor equipment.

[0029] The degradation level prediction system 10 may acquire vibration signals of noise points at n time points (where n is a natural number greater than or equal to 2), estimate the degradation level of the noise point (and / or noise source) from the data indicated by the vibration signal of the noise point based on deep learning, and / or predict the degradation level of the noise source at a target time point in the future based on time series data indicating the degradation level of the noise source from the past to the present. In the following, the current time point may refer to the latest time point among the n time points indicated by the vibration signal of the noise point acquired by the degradation level prediction system 10, and the target time point may include at least one time point within a predetermined prediction period after the current time point for which the degradation degree of the noise point (and / or noise source) needs to be predicted.

[0030] Deep learning can be implemented via various models, such as a convolutional neural network (CNN) model, a multi-layer perceptron (MLP) model, a deep neural network regression model, a long short-term memory (LSTM) model, etc. The convolutional neural network model can extract feature maps from data, and / or simplify (reduce) feature maps, and / or convert the dimensions of the simplified feature maps into one-dimensional vectors through a flattening layer in order to perform a learning process.

[0031] In addition, the degradation level prediction system 10 can estimate whether the degradation level of the noise source (e.g., motor reducer) is normal or abnormal based on the vibration signal associated with the noise point in the deep neural network classifier, and / or determine what the degradation level is if the degradation level is abnormal.

[0032] For example, the degradation level prediction system 10 may determine the degradation level of the motor-reducer of the vehicle at a future target time point based on data indicating vibration of noise generated from the motor-reducer of the vehicle.

[0033] refer to Figure 1 , the degradation level prediction system 10 may include an input unit 100, an estimation model learning unit 200, a prediction model learning unit 300, a controller 400, and / or a storage unit 500. The input unit 100, the estimation model learning unit 200, the prediction model learning unit 300, the controller 400, and / or the storage unit 500 may be implemented as one or more processors and a memory storing instructions, which, when executed by the one or more processors, cause the fault diagnosis system 10 to perform one or more operations described herein.

[0034] At least one sensor installed at a noise point in the vehicle can measure a vibration signal (e.g., at the noise point). For example, an acceleration sensor located at a noise point in the vehicle can generate three-axis data indicating a vibration signal associated with an x-axis, a y-axis, and a z-axis.

[0035] The input unit 100 may acquire, from at least one sensor, a vibration signal measured at a noise point corresponding to a noise point generating a vibration signal (e.g., configured to detect vibration of the noise point). Specifically, the input unit 100 may acquire, from at least one sensor, a vibration signal within a predetermined range from a noise source corresponding to the noise point. The vibration signal acquired by the input unit 100 may be measured and / or acquired in real time.

[0036] The vibration signal measured at the noise point may be time series data. For example, the vibration signal measured at the noise point may include data indicating the vibration measured for each unit time during a predetermined monitoring period from time point t_0 to time point t_n later than time point t_0. Time point t_0 may be the initial time point at which at least one sensor starts to collect the vibration signal, and time point t_n may be the current time point. In the following, for ease of explanation, time point t_m (where m is a natural number greater than 1) may be a time point that is (m-1) times later than time point t_1 in unit time. For example, if the unit time is 1 hour, time point t_5 may be a time point that is 4 hours later than time point t_1.

[0037] The input unit 100 may receive (e.g., from an external device) data indicating information about the state of a noise point corresponding to a time point at which one or more sensors measure a vibration signal. Here, the fault information may include data indicating a fault (degradation) type in the vehicle, the presence or absence of a fault, the degree of the fault, etc.

[0038] For example, the input unit 100 may acquire (eg, from an external device) data indicating fault information (eg, at and / or on a motor reducer in a vehicle) at a first time point at which a first vibration signal is measured.

[0039] However, the present disclosure is not limited thereto. The storage unit 500 may contain (e.g., store) vibration signals collected from noise points and / or data indicating information about the state of the noise points when each vibration signal is measured. For example, the storage unit 500 may contain a vibration signal indicating noise at a motor reducer of a vehicle of model A as a noise source in model A and / or data indicating fault information about a motor transmission of the vehicle when the vibration signal is measured.

[0040] Hereinafter, for convenience of explanation, it is assumed that the controller 400 acquires a vibration signal and data indicating corresponding fault information from the input unit 100 .

[0041] The controller 400 may receive a vibration signal acquired by the input unit 100 .

[0042] The controller 400 may convert one or more (eg, each) of the vibration signals collected from the noise points into data in the frequency domain. The way in which the controller 400 converts the vibration signal into data in the frequency domain may be implemented by fast Fourier transform (FFT).

[0043] Hereinafter, for convenience of explanation, data in the frequency domain into which the controller 400 converts the vibration signal at the noise point will be referred to as noise point data. The controller 400 may store the noise point data corresponding to the vibration signal at the noise point acquired in real time in the storage unit 500.

[0044] Hereinafter, noise point data may be associated with data indicating information about the state of the noise point when the noise point data is measured. For example, data indicating fault information of a motor reducer in a vehicle at a first time point may be associated with noise point data indicating vibration at the first time point.

[0045] The estimation model learning unit 200 may acquire the noise point data from the storage unit 500. The noise point data acquired by the estimation model learning unit 200 may be data labeled with state information of the noise point corresponding to the measurement time. Some data labeled with state information of the noise point corresponding to the time point at which the vibration signal is measured at the noise point may be a learning data set, and other data may be a test data set.

[0046] The estimation model learning unit 200 can generate a state estimation model 2000 based on the noise point data marked with the state information of the noise point, and / or can train the state estimation model through deep learning. The state estimation model 2000 can be implemented in the form of a CNN model, an MLP model, a deep neural network regression model, etc.

[0047] The estimation model learning unit 200 may use the learning data set to train the state estimation model 2000. The estimation model learning unit may use the test data set to test the state estimation model 2000. The state estimation model 2000 may be implemented via a regression model for predicting the life of a noise source corresponding to a noise point. The state estimation model 2000 may use the learning data set to learn a method of predicting state information of a noise point from data indicating vibration of the noise point through deep learning.

[0048] The estimation model learning unit 200 can provide the state estimation model 2000 with frequency data indicating the vibration measured from the noise point at a first time point, and / or derive an error between the predicted data output from the state estimation model 2000 and the degree of degradation of the noise source at the first time point, and / or modify the state estimation model 2000 based on the derived error.

[0049] The state estimation model 2000 may be a transfer function configured to output state information of a noise point based on frequency data indicating vibration of the noise point as input. The state estimation model 2000 may perform learning so as to be configured to output state information of a noise point based on frequency data indicating vibration of the noise point as input.

[0050] Hereinafter, for ease of explanation, the learned and / or tested state estimation model 2000 will be referred to as the learned state estimation model 2000 .

[0051] The estimation model learning unit 200 may provide the learned state estimation model 2000 to the controller 400 .

[0052] The controller 400 may control the operation of the input unit 100, the estimation model learning unit 200, the prediction model learning unit 300, and / or the storage unit 500. The controller 400 may be implemented in the form of hardware, or in the form of software, or in the form of a combination of hardware and software. The controller 400 may be implemented in the form of a processor (e.g., a microprocessor).

[0053] The controller 400 can control the entire process of learning and testing of the estimation model learning unit 200. The controller 400 can use the learned state estimation model 2000 to perform control of the detection process based on data indicating vibration of noise points, data indicating the degree of degradation of the noise source when measuring the vibration signal.

[0054] The controller 400 can estimate data indicating the degree of degradation of the noise source via the learned state estimation model 2000 and based on data indicating the vibration measured at the noise point at an arbitrary time point. Therefore, if the learned state estimation model 2000 receives frequency data indicating the vibration measured for each of the n unit times falling within a predetermined monitoring period (e.g., from time point t_0 to time point t_n, hereinafter referred to as "time series frequency data"), it can estimate data indicating the degree of degradation of the noise source corresponding to the n unit times falling within the monitoring period from time point t_0 to time point t_n (hereinafter referred to as "time series degradation data"). Here, n can be the number of unit times. For example, if the predetermined period is 10 hours and each unit time is 1 hour, each of the time series frequency data and the time series degradation data can include data items corresponding to 10 time points.

[0055] The controller 400 may input the time series frequency data to the learned state estimation model 2000, and store the time series degradation data estimated by the state estimation model 2000 in the storage unit 500. The storage unit 500 may store the time series degradation data in real time.

[0056] For example, the storage unit 500 may store the degradation degree of the noise source estimated by the state estimation model 2000 at the time point t_1, and store the degradation degree of the noise source estimated by the state estimation model 2000 at the time point t_2 later than the time point t_1.

[0057] The storage unit 500 may include a storage medium of at least one type of storage medium of a flash memory type, a hard disk type, a micro type, a card type (e.g., a secure digital card (SD card) or an eXtream digital card), etc., and a storage medium of at least one type of storage medium of a random access memory (RAM) type, a static RAM (SRAM) type, a read-only memory (ROM) type, a programmable ROM (PROM) type, an electrically erasable PROM (EEPROM) type, a magnetic RAM (MRAM) type, a magnetic disk type, and an optical disk type.

[0058] The controller 400 may predict the degradation degree of the noise source at a target time point later than the time point of measuring the vibration signal based on the data indicating the estimated degradation degree of the noise source. The controller 400 may predict the degradation degree of the noise source at the target time point using the degradation prediction model 3000 generated by the prediction model learning unit 300.

[0059] The degradation prediction model 3000 may be implemented in the form of an LSTM model or the like.

[0060] The prediction model learning unit 300 may acquire the time series degradation data from the storage unit 500. The prediction model learning unit 300 may use the time series degradation data to train the degradation prediction model 3000. The degradation prediction model 3000 may include a plurality of LSTM blocks, and each of the plurality of LSTM blocks may include a plurality of LSTM units. The degradation prediction model 3000 may learn the characteristics of the data (history) indicated by the time series degradation data over time.

[0061] The degradation prediction model 3000 can input time series degradation data corresponding to n time points falling within a monitoring period from time point t_0 to time point t_n into a unit corresponding to a time point in a plurality of LSTM units belonging to one of a plurality of LSTM blocks through deep learning.

[0062] The degradation prediction model 3000 can learn the degradation history of the noise source from the time series degradation data corresponding to n time points falling within the monitoring period from time point t_0 to time point t_n through multiple LSTM units belonging to each of the multiple LSTM blocks, and predict the degradation degree of the noise source corresponding to the target time point later than time point t_n based on the degradation history.

[0063] According to time series data corresponding to a plurality of continuous time points (e.g., a plurality of time points falling within a monitoring period from time point t_0 to time point t_n), the degradation prediction model 3000 may predict data indicating the degree of degradation of a noise source at a time point later than the plurality of time points (e.g., a target time point later than time point t_n which is the last time point in the monitoring period). The degradation prediction model 3000 may be a transfer function that outputs the degree of degradation of a noise source corresponding to a target time point later than time point t_n at which the time series degradation data is acquired. The degradation prediction model 3000 may use the time series degradation data to learn how to predict the degree of degradation of a noise source corresponding to each target time point falling within a prediction period later than time point t_n by deep learning.

[0064] The prediction model learning unit 300 may provide the learned degradation prediction model 3000 to the controller 400 .

[0065] The controller 400 may control the overall learning process of the prediction model learning unit 300. The controller 400 may perform control in the process of detecting data indicating the degree of degradation of a noise source at a time point later than the time point indicated by the time series degradation data from the time series degradation data using the learned degradation prediction model 3000. The controller 400 may predict data indicating the degree of degradation of a noise source at a target time point of the noise point through the learned degradation prediction model 3000.

[0066] The controller 400 can input data indicating the degree of degradation of the noise source corresponding to the unit time belonging to the monitoring period from time point t_0 to time point t_n into the learned degradation prediction model 3000, and derive the degree of degradation of the noise source corresponding to the target time point later than time point t_n, and store the derived data in the storage unit 500.

[0067] The storage unit 500 may store the vibration signal of the noise source acquired by the input unit 100, the learned state estimation model 2000, and the learned degradation prediction model 3000. The storage unit 500 may store time series degradation data and data indicating the degree of degradation of the noise source at the target time point predicted by the state estimation model 2000.

[0068] Figure 2 is a diagram showing one example of a state estimation model in the degradation level prediction system according to this example.

[0069] refer to Figure 2 , the state estimation model 2000 may include a plurality of FC layers 211 to 214 and a plurality of layers 2113 to 2115 .

[0070] The estimation model learning unit 200 can train the state estimation model 2000 by inputting the time series frequency data 201 to the state estimation model 2000 .

[0071] The time series frequency data 201 may include a plurality of noise point frequency data items corresponding to the vibration signals actually measured at the noise points, and a tag data item indicating the state information of the noise source when measuring the vibration signal, and the tag data items may be associated with the corresponding noise point frequency data items, respectively. Each of the plurality of noise point frequency data items may include data about a two-dimensional 3*m matrix corresponding to three axes (x-axis, y-axis, and z-axis) (where m is a natural number greater than 1). For example, the time series frequency data 201 may be a two-dimensional 3*2048 matrix.

[0072] The first FC layer 211 may include a plurality of layers 2101 to 2103. The plurality of layers 2101 to 2103 may have one weight and one bias. Data corresponding to one axis in the time series frequency data 201 may be input to each of the plurality of layers 2101 to 2103.

[0073] The following description will be made under the assumption that the time series frequency data 201 having a two-dimensional size of 3*2048 is input to the first FC layer 211. A one-dimensional 2048-bit vector indicating the first row corresponding to the x-axis in the time series frequency data 201 may be input to the layer 2101. A one-dimensional 2048-bit vector indicating the second row corresponding to the y-axis in the time series frequency data 201 may be input to the layer 2102. A one-dimensional 2048-bit vector indicating the third row corresponding to the z-axis in the time series frequency data 201 may be input to the layer 2103.

[0074] Each of the plurality of layers 2101 to 2103 may have a predetermined number of nodes and perform matrix calculations on input data and output data corresponding to the number of nodes. For example, each of the plurality of layers 2101 to 2103 may have 6400 nodes, and layer 2101 may output a one-dimensional 6400-bit vector associated with the x-axis, layer 2102 may output a one-dimensional 6400-bit vector associated with the y-axis, and layer 2103 may output a one-dimensional 6400-bit vector associated with the z-axis.

[0075] The second FC layer 212 may include a plurality of layers 2104 to 2106. The plurality of layers 2104 to 2106 may have one weight and one bias. The output of the corresponding layer in the plurality of layers 2201 to 2103 may be input to each of the plurality of layers 2104 to 2106. For example, the output of layer 2101 may be input to layer 2104. The output of layer 2102 may be input to layer 2105. The output of layer 2103 may be input to layer 2106.

[0076] Each of the plurality of layers 2104 to 2106 may have a predetermined number of nodes and perform matrix calculations on input data and output data corresponding to the number of nodes. For example, each of the plurality of layers 2104 to 2106 may have 3200 nodes, and layer 2104 may output a one-dimensional 3200-bit vector associated with the x-axis, layer 2105 may output a one-dimensional 3200-bit vector associated with the y-axis, and layer 2106 may output a one-dimensional 3200-bit vector associated with the z-axis.

[0077] The third FC layer 213 may include a plurality of layers 2107 to 2109. The plurality of layers 2107 to 2109 may have one weight and one bias. The output of the corresponding layer in the plurality of layers 2104 to 2106 may be input to each of the plurality of layers 2107 to 2109. For example, the output of layer 2104 may be input to layer 2107. The output of layer 2105 may be input to layer 2108. The output of layer 2106 may be input to layer 2109.

[0078] Each of the plurality of layers 2107 to 2109 may have a predetermined number of nodes and perform matrix calculations on input data and output data corresponding to the number of nodes. For example, each of the plurality of layers 2107 to 2109 may have 1600 nodes, and layer 2107 may output a one-dimensional 1600-bit vector associated with the x-axis, layer 2108 may output a one-dimensional 1600-bit vector associated with the y-axis, and layer 2109 may output a one-dimensional 1600-bit vector associated with the z-axis.

[0079] The fourth FC layer 214 may include a plurality of layers 2110 to 2112. The plurality of layers 2110 to 2112 may have one weight and one bias. The output of the corresponding layer in the plurality of layers 2107 to 2109 may be input to each of the plurality of layers 2110 to 2112. For example, the output of layer 2107 may be input to layer 2110. The output of layer 2108 may be input to layer 2111. The output of layer 2109 may be input to layer 2112.

[0080] Each of the plurality of layers 2110 to 2112 may have a predetermined number of nodes and perform matrix calculations on input data and output data corresponding to the number of nodes. For example, each of the plurality of layers 2110 to 2112 may have 400 nodes, and layer 2110 may output a one-dimensional 400-bit vector associated with the x-axis, layer 2111 may output a one-dimensional 400-bit vector associated with the y-axis, and layer 2112 may output a one-dimensional 400-bit vector associated with the z-axis.

[0081] The estimation model learning unit 200 may normalize the outputs of the plurality of layers 2110 to 2112, flatten the normalized outputs into one-dimensional data, and input the one-dimensional data to the layer 2113. For example, a one-dimensional 1200-bit vector obtained by flattening a one-dimensional 400-bit vector output from the layer 2110, a one-dimensional 400-bit vector output from the layer 2111, and a one-dimensional 400-bit vector output from the layer 2112 may be input to the layer 2113.

[0082] Each of the plurality of layers 2113 to 2115 may have a predetermined number of nodes and perform matrix calculations on input data and output data corresponding to the number of nodes. For example, layer 2113 may include 256 nodes and output a one-dimensional 256-bit vector, layer 2114 may include 256 nodes and output a one-dimensional 256-bit vector, and layer 2115 may include 1 node and input a one-dimensional 1-bit vector.

[0083] The data output from the layer 2115 may indicate the degree of failure of the noise source. For example, if the data output from the layer 2115 is 0, it may indicate that the motor reducer is normal, and if the data output from the layer 2115 falls within a range greater than 0 and equal to or less than 0.04, it may indicate that the motor reducer is at a first degradation level, and if the data output from the layer 2115 falls within a range greater than 0.04 and equal to or less than 0.28, it may indicate that the motor reducer is at a second degradation level, and if the data output from the layer 2115 falls within a range greater than 0.28 and equal to or less than 1, it may indicate that the motor reducer is at a third degradation level. In this case, the first degradation level indicates a lower degree of degradation compared to the second degradation level, and the second degradation level indicates a lower degree of degradation compared to the third degradation level.

[0084] If a first data item among a plurality of noise point frequency data items is input to the state estimation model 2000, the estimation model learning unit 200 may compare the predicted data output from the layer 2115 with the labeled data item corresponding to the first data item in the learning data set. The output of the layer 2115 may be the output of the state estimation model 2000.

[0085] The estimation model learning unit 200 may derive an error between the predicted data and the labeled data item output from the state estimation model 2000 , and output the state estimation model 2000 based on the derived error.

[0086] For example, the estimation model learning unit 200 may train the plurality of FC layers 211 to 214 and the plurality of layers 2113 to 2115 so that a loss function value indicating a difference between predicted data output from the layer 2115 and labeled data is minimized.

[0087] Furthermore, the estimation model learning unit 200 may test the state estimation model 2000 by inputting a test data set including the time series frequency data 201 into the state estimation model 2000 .

[0088] The estimation model learning unit 200 may predict the degree of failure of the noise source indicated by each of the plurality of noise point frequency data items through the state estimation model 2000 .

[0089] The estimation model learning unit 200 may provide the state estimation model 2000 generated by deep learning to the controller 400. The storage unit 500 may store the learned state estimation model 2000.

[0090] In this way, the controller 400 can provide the learning data set and the test data set to the estimation model learning unit 200 and receive the state estimation model 2000 from the estimation model learning unit 200.

[0091] The controller 400 may receive the time series frequency data to the learned state estimation model 2000 and acquire the time series degradation data corresponding to the time series frequency data.

[0092] If the learned state estimation model 2000 receives noise point data, it can estimate the degradation degree of the noise source corresponding to the noise point data. The storage unit 500 can store the degradation degree of the noise source output from the state estimation model 2000.

[0093] Furthermore, if the learned state estimation model 2000 receives time series frequency data, it may estimate time series degradation data corresponding to the time series frequency data. The storage unit 500 may store the time series degradation data output from the state estimation model 2000.

[0094] Figure 3 It is shown Figure 2 An example of the output of the state estimation model is shown in the figure.

[0095] refer to Figure 3 , the state estimation model 2000 may output the degradation degree of the noise source, and in this case, the degradation degree of the noise source may be one of a first degradation level, a second degradation level, and a third degradation level.

[0096] exist Figure 2 In the example of , the state estimation model 2000 may output a value indicating the degradation level of the motor reducer. If the data output from the state estimation model 2000 is 0, it may indicate that the motor reducer is normal, and if the data output from the state estimation model 2000 falls within a range greater than 0 and equal to or less than 0.1, it may indicate that the motor reducer is at a first degradation level, and if the data output from the state estimation model 2000 falls within a range greater than 0.1 and equal to or less than 0.6, it may indicate that the motor reducer is at a second degradation level, and if the data output from the state estimation model 2000 falls within a range greater than 0.6 and equal to or less than 1, it may indicate that the motor reducer is at a third degradation level. In this case, the first degradation level indicates a lower degree of degradation compared to the second degradation level, and the second degradation level indicates a lower degree of degradation compared to the third degradation level. In other words, the threshold value of the first degradation level may be 0.1, the threshold value of the second degradation level may be 0.6, and the threshold value of the third degradation level may be 0.9.

[0097] Reference again Figure 3 , most of the values ​​output from the state estimation model 2000 that has received the noise point data indicating that the motor reducer is normal are values ​​close to 0.

[0098] Most of the values ​​output from the state estimation model 2000 that has received the noise point data indicating that the motor reducer is at the first degradation level (degrad lev 1) are values ​​close to 0.1.

[0099] Most of the values ​​output from the state estimation model 2000 that has received the noise point data (degrad lev 2) indicating that the motor reducer is at the second degradation level are values ​​close to 0.6.

[0100] Most of the values ​​output from the state estimation model 2000 that has received the noise point data indicating that the motor reducer is at the third degradation level (degrad lev 3) are values ​​close to 0.9.

[0101] refer to Figure 3 , the value output from the state estimation model 2000 has a tendency similar to the degradation degree of the noise source corresponding to the noise point data input to the state estimation model 2000. Therefore, the state estimation model 2000 can be considered to be able to estimate the degradation degree of the motor reducer (noise source) with remarkable accuracy.

[0102] Figure 4 is a diagram showing one example of a degradation prediction model in the degradation level prediction system according to this example.

[0103] refer to Figure 4 , the degradation prediction model 3000 may include multiple LSTM blocks 311 and 312 , a flattening layer 313 , and a FC layer 314 .

[0104] The prediction model learning unit 300 may provide the time series degradation data to the degradation prediction model 3000, and with respect to the LSTM type degradation prediction model 3000, based on data corresponding to a first period (e.g., from time point t_0 to time point t_4) in the time series degradation data, predict degradation data indicating a degradation degree of a noise source at a second time point (e.g., time point t_5) later than a first period (e.g., from time point t_0 to time point t_4) corresponding to a plurality of consecutive first time points. The prediction model learning unit 300 may modify the degradation prediction model 3000 based on an error between data corresponding to the second time point (e.g., time point t_5) in the time series degradation data and degradation data predicted as a degradation degree of the noise source at the second time point (e.g., time point t_5).

[0105] exist Figure 4 , it is shown that the number of LSTM blocks 311 and 312 is two; however, this is only for convenience of explanation, and the present disclosure is not limited thereto.

[0106] The following description will be based on the following assumptions: Figure 4In the degradation prediction model 3000 shown, the epoch is 1000, the batch size is 64, the optimizer is Adam (Ir=0.001), and the loss is the mean absolute error.

[0107] The time series degradation data x corresponding to n time points belonging to the monitoring period from time point t_0 to time point t_n 1 to x n It can be used as input data 301 of the degradation prediction model 3000. Time series degradation data x 1 to x n can respectively indicate the degree of degradation of the noise source at a time point. For example, the data x 1 It can indicate the degree of degradation of the noise source at time point t_1, while the data x 2 The degradation degree of the noise source at the time point t_2 can be indicated.

[0108] The plurality of LSTM blocks 311 and 312 may include a plurality of LSTM units 3101 to 3110. Specifically, the LSTM block 311 may include a plurality of LSTM units 3101 to 3105, and the LSTM block 312 may include a plurality of LSTM units 3106 to 3110. In the plurality of LSTM units 3101 to 3110, the data dimension size (h) value of the LSTM unit state may be 64.

[0109] In the following, Figure 4 In the description of Figure 4 , it is shown that the number of LSTM units 3101 to 3110 included in each of the plurality of LSTM blocks 311 and 312 is 5 corresponding to n=5; however, the present disclosure is not limited thereto.

[0110] LSTM block 311 may receive time series degradation data x indicating 1 to x n The time point data corresponding to the time point data in the time series (for example, data x t , where t is a natural number equal to or greater than 1 and equal to or less than n). The plurality of LSTM units 3101 to 3105 may refer to the same unit having one parameter, and their outputs differ according to one or more input data.

[0111] In the following, reference will be made to Figure 5 An example of the configuration of each of the plurality of LSTM units 3101 to 3110 is described.

[0112] Figure 5 It is shown Figure 4A diagram of an example of multiple LSTM cells shown in .

[0113] refer to Figure 5 Each of the multiple LSTM units 3101 to 3110 can receive the long-term state c of the previous unit at time point t. t-1 and the short-term state h of the previous unit t-1 The corresponding input data x t , and output the long-term state c of the current unit t , the short-term state of the current unit h t And the output data y t Here, the long-term state and the short-term state can indicate the input data x t Each of the plurality of LSTM units 3101 to 3110 may determine a weight and a bias corresponding to each of the plurality of parameters so as to be able to output a value that is comparable to the input data x based on the long-term state at the previous time point and the short-term state at the previous time point. t The output data, long-term status, and short-term status corresponding to a time point later than the indicated time point.

[0114] In the following, reference will be made to Figure 4 and Figure 5 To describe the operation of multiple LSTM units 3101 to 3110.

[0115] In the following, for the sake of convenience, it is assumed that the input data corresponding to the plurality of LSTM units 3101 to 3105 is the time series degradation data x 1 to x n , and the long-term state c corresponding to the plurality of BSTM units 310-1 to 3105 t is the long-term state c1 t , and the short-term states h corresponding to the plurality of LSTM units 3101 to 3105 t It is a short-term state h1 t , and the output data y corresponding to the plurality of LSTM units 3101 to 3105 t is the output data y1 t Furthermore, it is assumed that the input data corresponding to the plurality of LSTM units 3106 to 3110 is the output data y1 of the plurality of LSTM units 3101 to 3105 t To y1 n , and the long-term state c corresponding to the plurality of LSTM units 3106 to 3110 t is the long-term state c2 t , and the short-term states h corresponding to the plurality of LSTM units 3106 to 3110 t It is a short-term state h2t , and the output data y corresponding to the plurality of LSTM units 3106 to 3110 t is the output data y2 t .

[0116] LSTM unit 3101 can receive time series degradation data x corresponding to time point t_1 1 , the long-term state of the previous unit c1 0 and the short-term state h1 of the previous unit 0 , and output the long-term state c1 of the current unit 1 、The short-term state of the current unit h1 1 And output data y1 1 . The long-term state c1 of the previous unit can be pre-set 0 and the short-term state h1 of the previous unit 0 Determined as initial information.

[0117] LSTM unit 3102 may receive time series degradation data x corresponding to time point t_2 2 , the long-term state of the previous unit c1 1 and the short-term state h1 of the previous unit 1 , and output the long-term state c1 of the current unit 2 、The short-term state of the current unit h1 2 And output data y1 2 .

[0118] LSTM unit 3103 may receive time series degradation data x corresponding to time point t_3 3 , the long-term state of the previous unit c1 2 and the short-term state h1 of the previous unit 2 , and output the long-term state c1 of the current unit 3 、The short-term state of the current unit h1 3 And output data y1 3 .

[0119] LSTM unit 3104 may receive time series degradation data x corresponding to time point t_4 4 , the long-term state of the previous unit c1 3 and the short-term state h1 of the previous unit 3 , and output the long-term state c1 of the current unit 4 、The short-term state of the current unit h1 4 And output data y1 4 .

[0120] LSTM unit 3105 may receive time series degradation data x corresponding to time point t_5 5, the long-term state of the previous unit c1 4 and the short-term state h1 of the previous unit 4 , and output the long-term state c1 of the current unit 5 、The short-term state of the current unit h1 5 And output data y1 5 .

[0121] The plurality of LSTM units 3101 to 3105 can learn how to respectively degrade the time series data x 1 to x n To determine the short-term state h1 of the unit at each time point 1 To h1 5 、Long-term state c1 1 To c1 5 And output data y1 1 To y1 5 Output data y1 1 To y1 5 Based on the input data x t 、Long-term state c1 1 To c1 5 and short-term status h1 1 To h1 5 The weights and biases are determined.

[0122] The output data of the LSTM block 311 may be input to the LSTM block 312. The plurality of LSTM units 3106 to 3110 may refer to the same unit having one parameter, and their outputs differ according to one or more input data.

[0123] LSTM unit 3106 can receive the output data y1 of LSTM unit 3101 corresponding to time point t_1 1 , the long-term state of the previous unit c2 0 and the short-term state h2 of the previous unit 0 , and output the long-term state c2 of the current unit 1 , the short-term state of the current unit h2 1 And output data y2 1 . The long-term state c2 of the previous unit can be pre-set 0 and the short-term state h2 of the previous unit 0 Determined as initial information.

[0124] LSTM unit 3107 can receive the output data y1 of LSTM unit 3102 corresponding to time point t_2 2 , the long-term state of the previous unit c2 1 and the short-term state h2 of the previous unit 1 , and output the long-term state c2 of the current unit2 , the short-term state of the current unit h2 2 And output data y2 2 .

[0125] LSTM unit 3108 may receive output data y1 of LSTM unit 3103 corresponding to time point t_3 3 , the long-term state of the previous unit c2 2 and the short-term state h2 of the previous unit 2 , and output the long-term state c2 of the current unit 3 , the short-term state of the current unit h2 3 And output data y2 3 .

[0126] LSTM unit 3109 may receive output data y1 of LSTM unit 3104 corresponding to time point t_4 4 , the long-term state of the previous unit c2 3 and the short-term state h2 of the previous unit 3 , and output the long-term state c2 of the current unit 4 , the short-term state of the current unit h2 4 And output data y2 4 .

[0127] LSTM unit 3110 may receive output data y1 corresponding to LSTM unit 3105 at time point t_5 5 , the long-term state of the previous unit c2 4 and the short-term state h2 of the previous unit 4 , and output the long-term state c2 of the current unit 5 , the short-term state of the current unit h2 5 And output data y2 5 .

[0128] The plurality of LSTM units 3106 to 3110 can learn how to respectively output data y1 based on the plurality of LSTM units 3101 to 3105. 1 To y1 5 To determine the short-term state h2 of the unit at each time point 1 To h2 5 、Long-term state c2 1 To c2 5 And output data y2 1 To y2 5 Output data y2 1 To y2 5 Based on the input data y1 1 To y1 5 、Long-term state c2 1 To c25 and short-term state h2 1 To h2 5 The weights and biases are determined.

[0129] refer to Figure 5 Each of the plurality of LSTM units 3101 to 3110 may include an input gate, a forget gate, and an output gate, and includes a forget gate layer f t , input gate layer i t , input hyperbolic tangent layer g t and output gate layer o t .

[0130] Forget gate layer f t can indicate based on the input data x t The data output is obtained by the weights and biases, and the short-term state h of the previous unit t-1 is input to the sigmoid layer σ. The forget gate layer f t The weights and bias values ​​may be determined via learning (eg, by learning).

[0131] Input gate layer i t It can be indicated based on the input data x t and the short-term state h of the previous unit that has been input to the sigmoid layer σ t-1 The data output of the value obtained by applying the weights and biases. Input gate layer i t The weights and / or bias values ​​may be determined via learning (eg, by learning).

[0132] Input hyperbolic tangent layer g t It can be indicated based on the input data x t and the short-term state h of the previous unit that has been input to the hyperbolic tangent layer tanh t-1 The data output of the values ​​obtained by applying the weights and biases. Input to the hyperbolic tangent layer g t The weights and / or bias values ​​may be determined via learning (eg, by learning).

[0133] Output gate layer o t It can be indicated based on the input data x t and the short-term state h of the previous unit that has been input to the sigmoid layer σ t-1 The data output of the value obtained by applying the weights and biases. Output gate layer o t The weights and bias values ​​may be determined via learning (eg, by learning).

[0134] The input gate can output the long-term state c of the previous unit t-1 With the forget gate layer (f t ) values.

[0135] The forget gate can transform the input gate layer (i t ) value and the input hyperbolic tangent layer (g t ) value, and / or the output of the input gate (the long-term state c of the previous unit) t-1 With the forget gate layer (f t ) value) is added to the product to update the long-term state c of the current unit t .

[0136] The output gate can be implemented by passing the long-term state c of the current unit based on the input to the hyperbolic tangent layer tanh t Output data y t And the output gate layer (o t ) value to update the short-term state h of the current unit t-1 .

[0137] Reference again Figure 4 , the flattening layer 313 can flatten the output data y2 outputted from the plurality of LSTM units 3106 to 3110 1 To y2 n The planarization layer 313 can receive the output data y2 1 To y2 n And generate a one-dimensional vector. For example, the flattening layer 313 can receive the output data y2 1 To y2 5 , and generates a one-dimensional 5-bit vector. The flattening layer 313 can generate output data by flattening n data into a one-dimensional vector through dimensional transformation while keeping the size of the input data the same.

[0138] The FC layer 314 may be completely connected to the flattening layer 313. The output data indicates the degree of degradation of the noise source corresponding to each of the plurality of time points t_0 to t_n and the target time point later than the time point t_n. The FC layer 314 may receive the output of the flattening layer 313 as input data, and output a one-dimensional vector by performing an operation on the input data. The output of the FC layer 314 may be the output of the degradation prediction model 3000.

[0139] For example, the FC layer 314 may receive a one-dimensional n-bit vector as input and generate a one-dimensional n-bit vector by performing an operation of Wx+b. Here, x, W, and b may indicate input, weight, and bias, respectively. The FC layer 314 may determine weight and bias values ​​by learning.

[0140] The degradation prediction model 3000 may generate output data 302 including time series degradation data corresponding to a predetermined monitoring period indicated by the input data 301 and time series degradation data corresponding to a predetermined prediction period later than the predetermined monitoring period and predicted based on the input data 301 .

[0141] The controller 400 may include the time series degradation data x 1 to x n The degradation data corresponding to the output data at at least one time point in the degradation prediction model 3000 and the output data output from the degradation prediction model 3000 at the previous time point (for example, data x n+1 ) is input to the degradation prediction model 3000, thereby generating output data 302 indicating the degree of degradation of the noise source corresponding to a time point later than the time point indicated by the first input data 301. The controller 400 may repeatedly collect output data from the degradation prediction model 3000 by delaying the time point of the input data 301 by one unit time until the time point indicated by the output data 302 reaches a final time point belonging to a predetermined prediction period.

[0142] For example, if the input data 301 is time series degradation data x corresponding to the period from time point t_0 to time point t_5 1 to x 5 , the prediction data 302 output from the degradation prediction model 3000 may include the degradation degree of the noise source at time point t_6. If the input data 301 is the time series degradation data x corresponding to the period from time point t_2 to time point t_6 2 to x 6 , the prediction data 302 output from the degradation prediction model 3000 may include the degradation degree of the noise source at time point t_7. If the predetermined prediction period is the period from time point t_6 to time point t_10, the controller 400 may input the time series degradation data x corresponding to the period from time point t_4 to time point t_9 to the degradation prediction model 3000. 4 to x 9 , and repeatedly collect the output data from the degradation prediction model 3000 until the degradation degree of the noise source at the time point t_10 is predicted.

[0143] The controller 400 can output time series data (hereinafter referred to as "time series prediction data") indicating the degree of degradation of the noise source corresponding to a predetermined monitoring period (the period from time point t_0 to time point t_n) and a predetermined prediction period (the period from time point t_(n+1) to time point t_(n+k) where k is a natural number greater than 2) from the degradation prediction model 3000.

[0144] The time series prediction data may include data indicating the degree of degradation of the noise source corresponding to the period from time point t_0 to time point t_(n+k). For example, the flattening layer 313 and the FC layer 314 flatten the one-dimensional n-bit vector corresponding to the predetermined monitoring period (the period from time point t_0 to time point t_n) and the one-dimensional k-bit vector corresponding to the predetermined prediction period (the period from time point t_(n+1) to time point t_(n+k)), thereby generating the time series prediction data in the form of a one-dimensional (n+k) bit vector.

[0145] The storage unit 500 may store time series prediction data predicted from the state estimation model 2000 .

[0146] As described above, in one example, the time series degradation data output from the state estimation model 2000 may be used as learning data for learning the degradation prediction model 3000. However, the example is not limited thereto, and in some examples, the storage unit 500 may store data indicating the degradation level of the noise source in association with a plurality of time points belonging to a predetermined period, and the degradation prediction model 3000 may learn how to predict data at a later time point based on the characteristics of the data stored in the storage unit 500 over time, the data indicating the degradation level of the noise source associated with the plurality of time points.

[0147] Figure 6 It is shown Figure 1 The detailed configuration of the controller is shown in the diagram.

[0148] refer to Figure 6 , the controller 400 may include a state estimation unit 410 , a degradation level prediction unit 420 and / or a remaining service life calculation unit 430 .

[0149] The state estimation unit 410 may generate and / or train the state estimation model 2000 as described above, and / or estimate the state information of the noise source according to the learned state estimation model 2000 .

[0150] The degradation level prediction unit 420 may generate and / or train the degradation prediction model 3000 as described above, and predict the degree of degradation of the noise source corresponding to the target time point according to the learned degradation prediction model 3000 .

[0151] In the following, reference will be made to Figure 7 The state information of the noise source estimated by the state estimation unit 410 and the degree of degradation of the noise source predicted by the degradation level prediction unit 420 are described.

[0152] Figure 7 is a diagram showing one example of a graph illustrating data generated by the degradation level prediction system according to this example.

[0153] refer to Figure 7 , shows the remaining useful life (RUL) corresponding to the time series degradation data (Degradation Data) estimated by the state estimation unit 410 based on the number of cycles of the vehicle corresponding to the time (Time) and / or the degradation degree (Predicted Data) of the noise source predicted by the degradation level prediction unit 420.

[0154] The following will refer to Figure 7 Description is made assuming that the input data is data corresponding to a period from the 0th cycle to the 1000th cycle.

[0155] refer to Figure 7 As time goes by, the remaining useful life (RUL) corresponding to the time series degradation data (Degradation Data) estimated by the state estimation unit 410 in association with each time point from the 0th cycle to the 1000th cycle and / or the remaining useful life (RUL) corresponding to the degradation degree of the noise source predicted by the degradation level prediction unit 420 in association with the 1000th cycle to the 1500th cycle tends to decrease.

[0156] In addition, the time series degradation data (Degradation Data) and / or the degradation degree of the noise source (Predicted Data) may have a trend similar to the actual data (True value) indicating the actual failure of the noise source. Therefore, in this example, it can be considered that it is possible to estimate and predict the degradation level of the noise source using the state estimation model 2000 and the degradation prediction model 3000 with significant accuracy.

[0157] Reference again Figure 6 The remaining service life calculation unit 430 may calculate the remaining service life at a reference time point belonging to the target time point based on the state information of the noise source estimated by the state estimation unit 410 and the degradation degree of the noise source predicted by the degradation level prediction unit 420.

[0158] The remaining service life calculation unit 430 may perform curve fitting on the time series degradation data corresponding to the monitoring period from time point t_0 to time point t_n and the data indicating the degree of degradation of the noise source later than time point t_n, thereby generating a trend graph. The remaining service life calculation unit 430 may generate a trend graph corresponding to the time series prediction data by estimating a series of linear relationships of the data based on the time series prediction data. The manner in which the remaining service life calculation unit 430 performs curve fitting may be implemented by various algorithms.

[0159] The remaining service life calculation unit 430 can calculate the remaining service life at the current time point based on the trend graph generated by curve fitting and the degree of degradation of the noise source at the current time point. The remaining service life calculation unit 430 can derive the degradation level of the noise source corresponding to the reference time point from the trend graph. Here, the reference time point may be one of at least one time point belonging to the target time point. The degree of degradation of the noise source derived from the trend graph may be different from the degree of degradation of the noise source at the reference time point in the time series prediction data. This is because the trend graph is a curve graph generated by the time series prediction data predicted by the degradation prediction model 3000 through curve fitting.

[0160] The remaining service life calculation unit 430 can derive the remaining service life of the noise source at the current time point based on the degradation level of the noise source at the current time point derived from the trend graph and the degradation level of the noise source at the reference time point. The remaining service life calculation unit 430 can calculate the remaining service life in a manner that can be implemented through various algorithms.

[0161] For example, when it is assumed that the time point t=200 is the current time point, the remaining service life calculation unit 430 can derive the degradation level of the motor reducer at the time point t=400 from the trend chart, and can calculate the remaining service life of the motor reducer at the current time point based on the degradation level of the motor reducer at the current time point and the degradation level of the motor reducer at the time point t=400.

[0162] Figure 8 is a flow chart of a degradation level prediction method according to an example.

[0163] The degradation level prediction system 10 has been described, and will not be redundantly described below. A degradation level prediction method to be described below may be a method performed by at least one processor included in the degradation level prediction system 10.

[0164] refer to Figure 8 , the sensor can collect the vibration signal from the noise point (S100).

[0165] The sensor may collect the vibration signal at the noise point, and the input unit 100 may acquire the collected vibration signal from the sensor.

[0166] The controller 400 may convert the vibration signal into data in the frequency domain ( S200 ).

[0167] The estimation model learning unit 200 may generate and train the state estimation model 2000 under the control of the controller 400 to estimate the degradation level through deep learning ( S300 ).

[0168] The estimation model learning unit 200 may determine the state estimation model 2000 based on learning data in which noise point data indicating a vibration signal at a first point in time is labeled with a degree of degradation of a noise source at the first point in time.

[0169] The state estimation unit 410 may estimate time series degradation data indicating the degree of degradation of the noise source corresponding to the unit time based on the learned state estimation model 2000 during the monitoring period from the time point t_0 to the time point t_n, and store the estimated data in the storage unit 500 (S400). The state estimation unit 410 may provide the time series frequency data corresponding to the monitoring period from the time point t_0 to the time point t_n to the state estimation model 2000, and acquire the time series degradation data output from the state estimation model 2000.

[0170] The prediction model learning unit 300 may generate and train a degradation prediction model 3000 through deep learning under the control of the controller 400 to predict a degradation state of a noise source at a target time point later than a current time point ( S500 ).

[0171] The prediction model learning unit 300 may determine the degradation prediction model 3000 based on the time series degradation data output from the state estimation model 2000 .

[0172] The degradation level prediction unit 420 may predict the degree of degradation of the noise source corresponding to the target time point later than the time point t_n based on the learned degradation prediction model 3000 and store the predicted data in the storage unit 500 ( S600 ).

[0173] The degradation level prediction unit 420 may provide the time series degradation data to the degradation prediction model 3000, and derive the time series prediction data output from the degradation prediction model 3000. The degradation prediction model 3000 may flatten the time series degradation data and the degree of degradation of the noise source at a target time point belonging to a prediction period later than the monitoring period indicated by the time series degradation data, thereby generating the time series prediction data.

[0174] The remaining service life calculation unit 430 may calculate the remaining service life of the noise source at the current time point based on the state information of the noise source estimated by the state estimation unit 410 and the degree of degradation of the noise source predicted by the degradation level prediction unit 420 ( S700 ).

[0175] The remaining service life calculation unit 430 can generate a trend graph corresponding to the time series prediction data according to the time series prediction data. The remaining service life calculation unit 430 can calculate the remaining service life at the current time point based on the trend graph generated by curve fitting and the degradation degree of the noise source at the current time point.

[0176] The remaining service life calculation unit 430 may derive the degradation level corresponding to the reference time point from the trend graph. The remaining service life calculation unit 430 may derive the remaining service life of the noise source at the current time point based on the degradation level of the noise source at the current time point derived from the trend graph and the degradation level of the noise source at the reference time point.

[0177] The present disclosure attempts to provide a degradation level prediction method and a degradation level prediction system, which are capable of measuring noise detected from a noise point in a vehicle, estimating the degradation level of the noise source based on the measured noise, and predicting the degradation level at a future time point.

[0178] A method according to a feature of the present disclosure is a degradation level prediction method, which is executed by at least one processor on a noise point that generates noise, and includes a step of determining a state estimation model by performing deep learning using a plurality of frequency data items indicating a plurality of vibrations of the noise point during a predetermined period and a plurality of degradation data items indicating a degree of degradation of the noise point (and / or a noise source) during the predetermined period, a step of measuring a plurality of first vibration signals at the noise point during a monitoring period, a step of estimating time series degradation data indicating a degree of degradation of the noise point (and / or a noise source) during the monitoring period by inputting in a frequency domain a plurality of first frequency data items obtained by converting a plurality of first vibration signals into a state estimation model, a step of determining a degradation prediction model by performing deep learning using the time series degradation data, a step of predicting degradation data indicating a degree of degradation of the noise point (and / or a noise source) at at least one target time point belonging to a prediction period later than the monitoring period from the degradation prediction model, and a step of calculating a remaining useful life of the noise source based on the time series degradation data and the degradation data at the target time point.

[0179] The step of determining the state estimation model may include: the step of providing a frequency data item indicating vibration measured from a noise point at a first time point as input to the state estimation model, the step of predicting a first degradation data item indicating the degree of degradation of the noise point (and / or noise source) based on the frequency data item at the first time point, and the step of modifying the state estimation model based on an error between the data item indicating the degree of degradation of the noise point (and / or noise source) at the first time point and the first degradation data item.

[0180] The step of determining a degradation prediction model may include: a step of providing time series degradation data as input to the degradation prediction model, a step of predicting a first degradation data item indicating a degree of degradation of a noise point (and / or a noise source) at a second time point later than a first time period corresponding to a plurality of consecutive first time points in a long short-term memory (LSTM) manner based on data corresponding to the first time period in the time series degradation data, and a step of modifying the degradation prediction model based on an error between the data corresponding to the second time point in the time series degradation data and the first degradation data item.

[0181] The step of predicting degradation data indicating the degree of degradation of noise points (and / or noise sources) may include: the step of generating time series prediction data by flattening the time series degradation data and the data indicating the degree of degradation of the noise points (or noise sources) corresponding to the prediction period.

[0182] The step of calculating the remaining useful life of the noise source may include: the step of generating a trend graph corresponding to the time series prediction data by performing curve fitting on the time series prediction data, the step of deriving a first degradation level of the noise source corresponding to a reference time point among the target time points from the trend graph, and the step of calculating the remaining useful life of the noise source at the current time point based on the degradation level of the noise source at the current time point and the first degradation level of the noise source.

[0183] A degradation level prediction system according to another feature of the present disclosure includes: an input unit that acquires a vibration signal from a noise point that generates noise, an estimation model learning unit that determines a state estimation model by performing deep learning using a plurality of frequency data items indicating a plurality of vibrations of the noise point during a predetermined period and a plurality of degradation data items indicating a degree of degradation of the noise point (and / or noise source) during the predetermined period, and a controller that measures a plurality of first vibration signals at the noise point during a monitoring period, and estimates time series degradation data indicating a degree of degradation of the noise point (and / or noise source) during the monitoring period by inputting a plurality of first frequency data items in a frequency domain obtained by converting the plurality of first vibration signals into the state estimation model, and predicts degradation data indicating a degree of degradation of the noise point (and / or noise source) at least one target time point belonging to a prediction period later than the monitoring period according to a degradation prediction model that performs deep learning using the time series degradation data, and calculates a remaining useful life of the noise source based on the time series degradation data and the degradation data at the target time point.

[0184] The estimation model learning unit can provide a frequency data item indicating the vibration measured from the noise point at a first time point as input to the state estimation model, and predict a first degradation data item indicating the degree of degradation of the noise point (and / or the noise source) based on the frequency data item at the first time point, and modify the state estimation model based on the error between the data item indicating the degree of degradation of the noise point (and / or the noise source) at the first time point and the first degradation data item.

[0185] The degradation level prediction system may also include a prediction model learning unit, which provides the time series degradation data as input to a long short-term memory (LSTM) type degradation prediction model, and predicts a first degradation data item indicating the degree of degradation of a noise point (and / or noise source) at a second time point later than the first time period corresponding to multiple consecutive first time points based on data corresponding to the first time period in the time series degradation data, and modifies the degradation prediction model based on an error between the data corresponding to the second time point in the time series degradation data and the first degradation data item.

[0186] The degradation prediction model may generate time series prediction data by flattening the time series degradation data and data indicating the degree of degradation of noise points (and / or noise sources) corresponding to the prediction period.

[0187] The controller can generate a trend graph corresponding to the time series prediction data by performing curve fitting on the time series prediction data, and derive a first degradation level of the noise source corresponding to a reference time point among the target time points from the trend graph, and calculate the remaining service life of the noise source at the current time point based on the degradation level of the noise source at the current time point and the first degradation level of the noise source.

[0188] According to the present disclosure, a vibration signal of a noise point may be used to determine an abnormality in a vehicle drive system.

[0189] According to the present disclosure, a deep learning model can be used to more accurately estimate abnormalities in a vehicle drive system based on vibrations at noise points.

[0190] According to the present disclosure, real-time data is used in the vehicle driving system in actual operation, so that the degradation state of the vehicle parts reflecting the current state of the vehicle can be estimated, and the future degradation time point can be predicted or the remaining service life at a certain time point in the future can be estimated.

[0191] According to the present disclosure, degradation in a vehicle drive system can be accurately determined so that the degradation can be addressed. Therefore, the time and cost required for vehicle maintenance can be reduced.

[0192] According to the present disclosure, as a result, the driving comfort of the vehicle driver can be improved, and the satisfaction of the vehicle user can be improved.

[0193] While the present disclosure has been described in conjunction with what are presently considered to be practical examples, it should be understood that the present disclosure is not limited to the disclosed examples. Rather, it is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A method comprising: The state estimation model is determined by performing deep learning using the following data items: a plurality of frequency data items indicating a plurality of vibrations of a noise source during a predetermined period, and a plurality of degradation data items indicating corresponding degrees of degradation of the noise source during the predetermined period; measuring a plurality of first vibration signals from the noise source at the noise points during a monitoring period; estimating time series degradation data indicating a degree of degradation of the noise source during the monitoring period by inputting a plurality of first frequency data items into the state estimation model, the plurality of first frequency data items being obtained by converting the plurality of first vibration signals into a frequency domain; determining a degradation prediction model by performing deep learning using the time series degradation data; predicting, based on the degradation prediction model, degradation data indicating a degree of degradation of the noise source at at least one target time point within a prediction period later than the monitoring period; as well as The remaining useful life of the noise source is calculated based on the time series degradation data and the degradation data at the target time point.

2. The method according to claim 1, wherein: Determining the state estimation model includes: providing as input to the state estimation model a frequency data item indicative of vibration measured at the noise point at a first point in time; predicting a first degradation data item indicating a first degree of degradation of the noise source based on the frequency data item at the first time point; and The state estimation model is modified based on an error between a data item indicating the degree of degradation of the noise source at the first time point and the first degradation data item.

3. The method according to claim 1, wherein: Determining the degradation prediction model includes: providing the time series degradation data as input to the degradation prediction model; predicting, via the degradation prediction model, a first degradation data item based on data corresponding to a first time period in the time series degradation data, the first degradation data item indicating a second degradation degree of the noise source at a second time point later than the first time period corresponding to a plurality of consecutive first time points; and The degradation prediction model is modified according to an error between data corresponding to the second time point in the time series degradation data and the first degradation data item.

4. The method according to claim 3, wherein: The degradation prediction model includes a long short-term memory (LSTM) model.

5. The method according to claim 1, wherein: The predicting the degradation data indicating the degree of degradation of the noise source includes generating time series prediction data by flattening the time series degradation data and the data indicating the degree of degradation of the noise source corresponding to the prediction period.

6. The method according to claim 5, wherein: Calculation of the remaining useful life of the noise source includes: generating a trend graph corresponding to the time series forecast data by performing curve fitting on the time series forecast data; deriving, from the trend graph, a first degradation level of the noise source corresponding to a reference time point among the at least one target time point; and Based on the degradation level of the noise source at the current time point and the first degradation level of the noise source, a remaining service life of the noise source at the current time point is determined.

7. The method according to claim 1, wherein: The noise source includes one or more of a motor reducer, a brake or a tire.

8. A degradation level prediction system, comprising: an input unit configured to acquire a vibration signal from a noise point, wherein the noise point is configured to measure noise from a noise source; An estimation model learning unit configured to determine a state estimation model by performing deep learning using the following data items: a plurality of frequency data items indicating a plurality of vibrations of the noise point during a predetermined period, and a plurality of degradation data items indicating a degree of degradation of the noise source during the predetermined period; as well as A controller, the controller being configured to: measuring a plurality of first vibration signals at the noise points during a monitoring period, estimating time series degradation data indicating a degree of degradation of the noise source during the monitoring period by inputting a plurality of first frequency data items into the state estimation model, wherein the plurality of first vibration data items are obtained by converting the plurality of first vibration signals into a frequency domain, predicting degradation data indicating the degree of degradation of the noise source at at least one target time point within a prediction period later than the monitoring period via a degradation prediction model that performs deep learning using the time series degradation data, and The remaining useful life of the noise source is calculated based on the time series degradation data and the predicted degradation data at the at least one target time point.

9. The degradation level prediction system according to claim 8, wherein: The estimation model learning unit is configured as follows: providing as input to the state estimation model a frequency data item indicative of a vibration measured at the noise point at a first point in time, predicting a first degradation data item indicating a degree of degradation of the noise source from the frequency data item at the first time point, and The state estimation model is modified according to an error between a data item indicating a degree of degradation of the noise source at the first time point and the first degradation data item.

10. The degradation level prediction system according to claim 8, further comprising: A prediction model learning unit, wherein the prediction model learning unit is configured to: providing the time series degradation data as input to the degradation prediction model, predicting a second degradation data item indicating a degradation degree of the noise source at a second time point later than the first time period based on data corresponding to a first time period in the time series degradation data, wherein the first time period corresponds to a plurality of consecutive first time points, and The degradation prediction model is modified according to an error between a portion of the time-series degradation data corresponding to the second time point and the second degradation data item.

11. The degradation level prediction system according to claim 10, wherein: The degradation prediction model includes a long short-term memory (LSTM) model.

12. The degradation level prediction system according to claim 8, wherein: The degradation prediction model is configured to generate time series prediction data by flattening the time series degradation data and data indicating the degree of degradation of the noise source corresponding to the prediction period.

13. The degradation level prediction system according to claim 12, wherein: The controller is configured to: generating a trend graph corresponding to the time series forecast data by performing curve fitting on the time series forecast data, deriving a first degradation level of the noise source corresponding to a reference time point among the at least one target time point based on the trend graph, and The remaining service life of the noise source at the current time point is calculated based on the degradation level of the noise source at the current time point and the first degradation level of the noise source.

14. The degradation level prediction system according to claim 8, wherein: The noise source includes one or more of a motor reducer, a brake or a tire.

Images