Engine status diagnosis method and diagnosis modeling method

Abstract

The present disclosure relates to a method for diagnosing an engine state and a diagnostic modeling method thereof. A method for diagnosing an engine state may include: measuring the vibration of an engine, wherein the assembly of the engine is completed in an automated end-of-line (EOL) process of engine production; and preliminarily diagnosing the vibration of the engine during the EOL process using a primary deep learning classification model, wherein the vibration signals of the engine are classified according to features obtained by learning using multiple algorithms based on big data about multiple vibration signals measured at three or more locations on the engine, wherein the assembly state of the engine in the EOL process is classified as a normal state or an abnormal state by preliminarily diagnosing the vibration of the engine.

Description

Technical Field

[0001] Exemplary embodiments of the present disclosure relate to a method of diagnosing an abnormal state of an engine using artificial intelligence (AI) based on deep learning and a method of configuring a diagnosis model. Background Art

[0002] The human brain is composed of multiple nerve cells called neurons. Each of the neurons is connected to hundreds to thousands of other neurons through synapses called junctions. Each of the neurons receives electrical and chemical signals from other neurons connected to it and aggregates these signals in the cell body. When the aggregation value is greater than a threshold (i.e., a neuron's inherent limit), the neuron is activated to send outputs to adjacent neurons through axons. Information exchange between neurons is performed simultaneously, and the function of this information exchange is improved through learning.

[0003] In terms of the technical structure in the relevant field, "artificial intelligence (AI)" is the highest concept. AI is a technology that will enable computers and robots to imitate the human brain and neural network and think and act like humans one day in the future.

[0004] Research is being conducted on control systems based on learning related to AI even in vehicles. However, AI has been applied to vehicles only as a technology combining speaker recognition (speech recognition) with mobile information technology (IT).

[0005] That is, examples of AI applications to transportation include navigation and audio manipulation through voice recognition and application manipulation through smartphone interlocking.

[0006] Meanwhile, vehicles are finished products composed of tens of thousands of parts, and even when some parts fail, it is not easy to identify the problem and accurately determine which part failed.

[0007] Therefore, when deep learning-based AI is used for component fault diagnosis, faulty components can be identified and repaired more accurately and quickly.

[0008] The present disclosure relates to a method for diagnosing an abnormal state of an engine by constructing a diagnosis model for diagnosing the abnormal state of the engine, the engine being a core component of a vehicle.

[0009] Conventional diagnosis of engine abnormalities can be performed only through a simple status check under end-of-line (EOL) conditions, which is the last operation that can be performed during the engine production process. Figure 1 As shown, when the final state is checked during the engine production process, normality and abnormality of the engine are diagnosed only by comparing the levels of frequency so that synthetic diagnosis is not performed for the abnormal state of the engine.

[0010] Conventionally, as shown in the accompanying drawings, a reference value compared with the vibration decibel noise level and the kurtosis level is set by a specific frequency filter using test data in an abnormal state and then only the reference value is compared with the vibration decibel noise level and the kurtosis level. Therefore, when a complex abnormal state occurs, the reference value cannot be set.

[0011] The engine should be classified from normal to abnormal during the engine production process. When the abnormality is not classified and then discovered after the vehicle is assembled, the engine should be separated from the vehicle.

[0012] Therefore, man-hours (M / H) and costs are consumed in the above process, and when a vehicle assembled with an abnormal engine is delivered to a consumer, not only is the repair cost consumed but also the complete image of the vehicle is damaged.

[0013] The contents described in the above background art are intended to help understanding the background of the present disclosure and may include contents previously known to those skilled in the art to which the present disclosure belongs. Summary of the invention

[0014] Embodiments of the present disclosure relate to a method for diagnosing an abnormal state of a vehicle engine using deep learning-based artificial intelligence (AI).

[0015] Other purposes and advantages of the present disclosure can be understood through the following description and become apparent with reference to the embodiments of the present disclosure. In addition, it is obvious to those skilled in the art that the objects and advantages of the present disclosure can be achieved through the claimed methods and combinations thereof.

[0016] According to one embodiment of the present disclosure, a method for diagnosing an engine status includes: measuring the vibration of the engine, wherein assembly of the engine is completed in an automated end-of-line (EOL) process of engine production; and preliminarily diagnosing the vibration of the engine during the EOL process using a primary deep learning classification model, wherein the vibration signals of the engine are classified according to features obtained by learning using multiple algorithms based on big data about multiple vibration signals measured at three or more locations on the engine, wherein the assembly status of the engine in the EOL process is classified as a normal state or an abnormal state by preliminarily diagnosing the vibration of the engine.

[0017] Measuring vibrations of the engine in the EOL process may include measuring vibrations at the same locations as measurement locations of the plurality of vibration signals of the preliminary deep learning classification model.

[0018] The big data about the plurality of vibration signals used to construct the primary deep learning classification model may include data for normal signals according to acceleration or constant speed driving conditions and various types of abnormal vibration signals.

[0019] Each vibration signal of the plurality of vibration signals may include data over an entire frequency range.

[0020] More specifically, a primary deep learning classification model is constructed by sequentially learning multiple vibration signals using a recursive neural network (RNN) (many-to-one) based classification algorithm, a convolutional neural network (CNN) + attention mechanism algorithm, and a DenseNet algorithm.

[0021] When the engine in the EOL process is classified as abnormal in the preliminary diagnosis, the method of diagnosing the engine status may also include performing a secondary diagnosis of the vibration of the engine in the EOL process using a secondary deep learning diagnostic model, in which the diagnostic reference setting value is determined according to the abnormal situation based on big data about multiple vibration signals.

[0022] The vibration of the engine in the EOL process is remeasured on the repaired engine according to the abnormality diagnosed in the secondary diagnosis.

[0023] The method of diagnosing the engine status may further include storing the diagnosis results obtained in the primary diagnosis and the secondary diagnosis in the central server; and including the data stored in the central server and the vibration data for the engine measured separately in the big data.

[0024] More specifically, a secondary deep learning diagnostic model can be constructed by learning multiple vibration signals using frame energy, pitch, zero crossing, precise diagnosis algorithm based on kurtosis signal processing, binary classification, and integrated deep neural network (DNN) / result selection algorithm.

[0025] According to another embodiment of the present disclosure, a diagnostic modeling method for an engine state includes: collecting big data about vibration signals of an engine; measuring multiple vibration signals at three or more locations on an engine in an automated end-of-line (EOL) process of engine production; analyzing data based on the big data and multiple vibration signals and extracting features for supervised learning; constructing a primary deep learning classification model, in which the vibration signals of the engine are classified based on the big data according to features obtained by learning using multiple algorithms; and constructing a secondary deep learning diagnostic model, in which a diagnostic reference setting value is determined based on the big data according to abnormal conditions classified using the primary deep learning classification model.

[0026] Each of the classification accuracy of the primary deep learning classification model and the diagnostic accuracy of the secondary deep learning diagnosis model may be 99.9% or above.

[0027] When the classification accuracy of the primary deep learning classification model and the diagnostic accuracy of the secondary deep learning diagnosis model were less than 99.9%, the data were reanalyzed and features for supervised learning were extracted.

[0028] The big data about the plurality of vibration signals used to construct the primary deep learning classification model may include data for normal signals according to acceleration or constant speed driving conditions and various types of abnormal vibration signals.

[0029] Each vibration signal of the plurality of vibration signals may include data across a frequency range.

[0030] More specifically, a primary deep learning classification model is constructed by sequentially learning multiple vibration signals using a recursive neural network (RNN) (many-to-one) based classification algorithm, a convolutional neural network (CNN) + attention mechanism algorithm, and a DenseNet algorithm.

[0031] A secondary deep learning diagnostic model is constructed by learning multiple vibration signals using frame energy, pitch, zero crossing, an accurate diagnosis algorithm based on kurtosis signal processing, binary classification, and an integrated deep neural network (DNN) / result selection algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a diagram illustrating a conventional method of diagnosing an abnormal state of an engine.

[0033] Figure 2 and Figure 3 is a diagram illustrating a method for diagnosing an engine state and a method for diagnosing the engine state according to the present disclosure.

[0034] Figure 4 is a diagram illustrating a plurality of vibration signal measurement systems.

[0035] Figure 5A , Figure 5B , Fig. 6A as well as Figure 6B is a graph showing vibration data of an engine in normal and abnormal states.

[0036] Fig. 7A , Figure 7B as well as Figure 7C is a diagram showing a classification method according to a preliminary diagnosis.

[0037] Fig. 8A and Figure 8B is a diagram illustrating a secondary diagnosis algorithm. DETAILED DESCRIPTION

[0038] Reference should be made to the accompanying drawings which illustrate exemplary embodiments of the present disclosure and to the description therein for a fuller understanding of the present disclosure and its operating advantages, as well as objectives attained by practicing the present disclosure.

[0039] In the description of the exemplary embodiments of the present disclosure, known techniques or repeated descriptions that make the main points of the present disclosure unnecessarily obscure may be reduced or omitted.

[0040] Figure 2 and Figure 3 is a diagram illustrating a method for diagnosing an engine state and a method for diagnosing the engine state according to the present disclosure.

[0041] In the following, reference will be made to Figure 2 and Figure 3 A method for diagnosing an engine state and a method for diagnosing the engine state according to the present disclosure are described.

[0042] According to the present disclosure, a method is provided for constructing a primary deep learning model and a secondary deep learning model using multiple vibration signals of an engine based on big data about various types of abnormal status conditions, and diagnosing the engine status according to the constructed primary deep learning model and the secondary deep learning model.

[0043] First, in order to construct a deep learning model, big data on vibration signals of an engine are collected ( S11 ), and a plurality of vibration signals are measured at three or more locations on the engine ( S12 ).

[0044] Collection of big data on vibration signals is performed by collecting data including vibration data result information to be described below, which is measured for evaluation of a production line ( S11 ).

[0045] In addition, through Figure 4 The illustrated multiple vibration signal measurement system in the automated end-of-line (EOL) process measures multiple vibration signals of an engine ( S12 ).

[0046] Whether assembly defects occur in the engine completed in assembly and delivery is determined by cold room evaluation as an engine EOL process. Multiple vibration signals of the engine are measured by a multiple vibration signal measurement system.

[0047] Since the position and characteristics of vibration generation are different according to the abnormal type of the engine, in order to ensure high EOL diagnosis accuracy, the drawings show that multiple vibration signal measurements are performed so as to measure vibration signals at multiple positions on the engine. Therefore, the diagnosis accuracy can be improved. Then, data of normal state and abnormal state are collected according to multiple situations.

[0048] It is necessary to measure multiple vibration signals of the engine to collect synchronous data conditions and fifteen or more test data required for the learning model and should reflect the deviation characteristics of the engine.

[0049] Then, based on the big data collected in operation S11 and the multiple vibration signals measured in operation S12, the data is analyzed and features for supervised learning are extracted (S13), a primary deep learning classification model is constructed (S14), and then a secondary deep learning diagnostic model is constructed (S15).

[0050] The classification accuracy in operation S14 is required to be 99.9% or higher. When the classification accuracy is less than 99.9%, data analysis and feature extraction are repeated in operation S13.

[0051] In addition, due to the modeling result in operation S15, the classification accuracy is required to be 99.9% or more. When the classification accuracy is less than 99.9%, data analysis and feature extraction are repeated in operation S13.

[0052] The feature extraction of the plurality of vibration signals in operation S13 extracts representative features based on the plurality of channel signals. FIG. 5A to FIG. 6B As shown in , in normal and abnormal states under acceleration conditions and constant speed conditions, features and pattern changes in the entire frequency range are extracted instead of analyzing frequencies within a specific range.

[0053] That is, features of vibration data regarding normal and abnormal states are extracted according to frequency and operating conditions (S13) so as to construct a primary deep learning classification model (S14) by vibration data collected according to normal and abnormal situations.

[0054] Therefore, by applying the extracted feature data in a composite manner 7A to 7C The algorithm constructs a primary deep learning classification model (S14) and preliminarily classifies the abnormal state of the engine to be described below according to the constructed primary deep learning classification model (S23).

[0055] Fig. 7A A classification algorithm based on a recurrent neural network (RNN) (many-to-one) is shown, in which main information about each frame is stored and defined as features, and n total feature information is extracted, defined, and stored as features of multiple vibration signals in normal / abnormal situations.

[0056] Figure 7B The convolutional neural network (CNN) + attention mechanism algorithm is shown, in which the shared Fig. 7A The input value of each frame unit is obtained and the information of the transduction (characteristic) density of each frame unit is stored, thereby extracting high-probability features about each vibration signal.

[0057] also, Figure 7C A DenseNet algorithm is shown, in which DenseNet blocks are applied to form a high-precision model at a layer close to an input signal, and unlike existing DNNs, the input signal shares information with all layers to improve the classification model.

[0058] As described above, gradually applying the three-level composite algorithm to multiple vibration input signals (big data for supervised learning) makes it possible to improve the classification accuracy of the deep learning model.

[0059] Next, by applying Fig. 8A and Figure 8B The algorithm is used to construct a secondary deep learning diagnosis model (S15).

[0060] Fig. 8A A first stage diagnostic algorithm is shown that applies frame energy, pitch, zero crossings, and an exact diagnostic algorithm based on kurtosis signal processing.

[0061] Zero crossing is a technique for classifying the energy of input signals (vibration) using changes in positive and negative values, i.e., using the sign and magnitude changes of the value obtained by multiplying the previous frame data value by the current frame data value to classify the features. In this case, the features of the parts where the sign changes and the magnitude changes significantly are the main information.

[0062] The kurtosis is characterized in that the kurtosis value has a large value due to a large deviation between the input data and the reference data, and has a small value due to a small deviation therebetween. In this case, when the vibration data in the normal state is regarded as the reference data, the feature having a large kurtosis value when the vibration data in the abnormal state is input becomes the main information.

[0063] In addition, the principal component analysis (PCA) algorithm is an unsupervised learning technique and uses a probabilistic technique to set the diagnostic segmentation axis. The PCA algorithm is an algorithm characterized in that the classification (detailed diagnosis) segmentation axis is optimized with respect to the fault type (case 1, case 2, ... case n) using the changes in the absolute value of the frame energy and pitch and the characteristics of zero crossing and kurtosis.

[0064] Next, Figure 8B A two-stage diagnostic algorithm is shown, and a binary classification and an integrated deep neural network (DNN) / result selection algorithm are applied to the two-stage diagnostic algorithm.

[0065] Binary classification is an algorithmic structure that shows high performance when adding hidden layers to the deep learning structure for accurate diagnosis.

[0066] The integrated DNN / result selection algorithm uses multiple input vibration signals to achieve more accurate diagnosis, which is a feature of the present disclosure. The integrated DNN / result selection algorithm is an algorithm that learns multi-channel characteristics, distinguishes important channel information from other channel information, and assigns high weights to important channel information, thereby improving diagnostic accuracy.

[0067] According to the present disclosure, an abnormal state of an engine can be diagnosed by applying the deep learning model constructed through the above process.

[0068] First, refer back Figure 3 , in order to diagnose the engine, the vibration of the cold chamber is measured in the automatic EOL process of engine production (S21), and the data in the evaluation mode (constant speed / acceleration) is measured in the automatic EOL process of engine production (S22).

[0069] It is preferable that the measurement in operation S21 is performed to measure vibration at the same position where the collection of big data for learning the model in operation S11 is performed, and the measurement in operation S22 is performed for the purpose of keeping the production time constant.

[0070] Therefore, vibration data according to the driving condition of the engine to be diagnosed is preliminarily diagnosed by applying the deep learning classification model constructed in operation S14 (S23).

[0071] Compared with the feature vectors classified into four to five cases according to the abnormal cases, the assembly state of the engine is classified into a normal state or an abnormal state according to the results of the initial classification model.

[0072] At the same time, when it is determined that the engine is in an abnormal state, a secondary diagnosis is performed by applying the deep learning diagnostic model constructed in operation S15 (S25), and a secondary diagnosis is performed on the engine to determine whether it is bad (NG) according to the abnormal situation by using the secondary diagnostic model that sets the diagnostic reference setting value according to the category, and a corresponding situation in the abnormal situation is derived so that the engine can be moved to check the line corresponding to the derived situation and then repaired (S26).

[0073] Furthermore, in order to re-inspect the engine after repairing the abnormal state of the engine, the engine is re-input into the production line in operation S21 so that diagnosis can be performed on the engine.

[0074] The raw data and vibration data of the diagnosis results in operations S24 and S26 are transmitted to the central server and stored therein (S27).

[0075] Furthermore, after collecting vibration data (result information), the vibration data (result information) measured for the purpose of evaluating the production line may be included in big data ( S28 ) so that the learning model may be updated and thus the diagnostic accuracy may be improved.

[0076] Alternatively, the engine may be diagnosed for an abnormal state of the engine based on a primary deep learning classification model to be classified into normal and abnormal states only, and the engine may be selectively diagnosed for an abnormal state of the engine based on a secondary deep learning classification model.

[0077] As described above, according to the present disclosure, a deep learning model for classifying and diagnosing abnormal conditions of an engine is constructed based on big data about driving conditions and multiple vibration signals according to a certain frequency, so that it can accurately diagnose abnormal conditions of the engine before assembling the produced engine.

[0078] According to an exemplary embodiment of the present disclosure, a deep learning algorithm can be configured with multiple levels based on big data about various situations of abnormal state conditions so that abnormal situations can be classified and a normal state or an abnormal state can be accurately determined. That is, in addition to the level comparison, a technique for comparing the transduced energy flow with a feature image and an algorithm for comparing features in the entire frequency range are configured so that the abnormal state of the engine in the production stage can be diagnosed.

[0079] Although the present disclosure has been described with reference to the accompanying drawings, it is not limited to the exemplary embodiments described, and it is obvious to those skilled in the art that substitutions and modifications may be made without departing from the spirit and scope of the present disclosure as defined in the appended claims. Therefore, it should be noted that such substitutions and modifications are within the scope of the claims of the present disclosure, and the scope of the present disclosure should be interpreted based on the appended claims.

Claims

1. A method for diagnosing an engine status, comprising: Measuring vibrations of an engine, wherein assembly of the engine is completed in an automated end-of-line process of engine production; and preliminarily diagnosing the vibration of the engine during the end-of-line process using a primary deep learning classification model, in which the vibration signals of the engine are classified according to features obtained by learning using a plurality of algorithms based on big data regarding a plurality of vibration signals measured at three or more locations on the engine, wherein the assembly state of the engine in the off-line process is classified as a normal state or an abnormal state by performing a preliminary diagnosis on the vibration of the engine; The primary deep learning classification model is constructed by sequentially learning the multiple vibration signals using a classification algorithm based on a recursive neural network, a convolutional neural network + attention mechanism algorithm, and a DenseNet algorithm. The method further comprises: When the engine in the offline process is classified as abnormal in the preliminary diagnosis, a secondary diagnosis is performed on the vibration of the engine in the offline process using a secondary deep learning diagnosis model, in which a diagnosis reference setting value is determined according to each abnormal situation based on big data about the plurality of vibration signals, The vibration of the engine in the offline process is secondary diagnosed by deriving the corresponding abnormal situation in the abnormal situation through the secondary deep learning diagnosis model.

2. The method according to claim 1, wherein: Measuring vibrations of the engine in the end-of-line process includes measuring vibrations at the same locations as measurement locations of the plurality of vibration signals of the preliminary deep learning classification model.

3. The method according to claim 1, wherein: The big data about the plurality of vibration signals used to construct the primary deep learning classification model includes data about normal signals according to acceleration or constant speed driving conditions and a plurality of types of abnormal vibration signals.

4. The method according to claim 3, wherein: Each vibration signal of the plurality of vibration signals includes data over an entire frequency range.

5. The method according to claim 1, wherein: The vibration of the engine in the off-line process is measured again for the repaired engine according to the abnormality diagnosed in the secondary diagnosis.

6. The method according to claim 1, further comprising: storing the diagnosis results obtained in the primary diagnosis and the secondary diagnosis in a central server; as well as The data stored in the central server and the separately measured vibration data for the engine are included in the big data.

7. The method according to claim 1, wherein: The secondary deep learning diagnostic model is constructed by learning the multiple vibration signals using frame energy, fundamental pitch, zero crossing, precise diagnosis algorithm based on kurtosis signal processing, binary classification, and integrated deep neural network / result selection algorithm.

Images