Artificial intelligence-based fault diagnosis method for rotating devices

Abstract

The present invention relates to an artificial intelligence-based fault diagnosis method for a rotating device that enables more efficient management, such as maintenance, of a device equipped with a rotating element for purposes such as power transmission, and is characterized by comprising: a data collection step for collecting continuous real-time vibration values ​​generated from one or more vibration sensors attached to the device; a first analysis step for determining whether the total vibration value during a predetermined period selected from the collected real-time vibration values ​​is above a value of interest; a second analysis step for determining whether the same vibration pattern occurs periodically per unit time if the total vibration value is above the value of interest; and an artificial intelligence analysis step for deriving the cause of a fault based on a fault diagnosis learning algorithm learned through deep learning using fault simulation vibration data if the same vibration pattern occurs periodically more than a specified number of times.

Description

Technology Field

[0001] The present invention relates to an artificial intelligence-based fault diagnosis method for a rotating device, which enables more efficient management, such as maintenance, of a device equipped with rotating elements for purposes such as power transmission. Background Technology

[0003] Rotating equipment is being utilized in various industrial sites, including manufacturing plants, shipbuilding, and smart factories. Rotating equipment refers to elements or devices designed to transmit rotational force generated from a power source to where it is needed.

[0004] Generally, rotary equipment responsible for power transmission can be composed of a combination of motors, bearings, shafts, clutches, gears, pulleys, chains, etc. For stable power transmission, each component must be installed accurately in its proper position according to specifications, and it must possess defect-free initial quality. Of course, during the assembly and placement of these components, they must also be installed precisely within the assembly tolerance range in accordance with the design.

[0005] Due to the nature of rotary equipment, which consists of the assembly and combination of multiple components, if installed on a manufacturing line for product production, a motor failure will cause the entire line to shut down. Furthermore, damage to bearings, gears, or shafts will generate excessive noise or vibration; continued operation in such a state will adversely affect the entire rotary equipment and become a new variable that can cause product defects on the manufacturing line.

[0006] As rotating equipment is an essential and important part of almost every industrial sector, it must be operated stably and manufactured and installed to ensure sufficient durability. However, no matter how much care is taken, problems may arise due to various causes, and the countermeasure is to identify the problem as early as possible and repair or replace it promptly.

[0007] When there is a problem with rotating equipment, it exhibits anomalies that differ from normal conditions, and typical signs of abnormality include noise and vibration. Therefore, when managers of rotating equipment confirm that unusual noise or vibration is occurring, they usually conduct inspections or take other measures.

[0008] However, if the abnormal noise can be detected through the human ear or if abnormal vibrations transmitted through the hands or other parts of the body can be felt, it can be assumed that significant damage has already occurred somewhere in the rotating equipment.

[0009] Therefore, recently, technology is being applied to install vibration sensors on rotating equipment to check for abnormalities based on changes in vibration values ​​detected by the sensors. Prior art literature

[0011] Republic of Korea Registered Patent No. 10-2275571 The problem to be solved

[0012] Accordingly, the present invention aims to provide a highly reliable AI-based fault diagnosis method for rotating equipment by incorporating AI technology to detect abnormalities in rotating equipment at an earlier stage and enable corresponding countermeasures.

[0013] In addition, the present invention aims to solve the problem of difficulty in obtaining actual failure data from existing equipment or the lack of failure data itself by generating vibration data simulated for various failure types and utilizing it as training data for an artificial intelligence-based failure diagnosis learning algorithm. Through this, the invention aims to improve the performance of the failure diagnosis model and enable the derivation of more accurate and reliable causes of failure. means of solving the problem

[0015] The artificial intelligence-based fault diagnosis method for a rotating body device according to the present invention for achieving the presented task comprises: a data collection step for collecting continuous real-time vibration values ​​generated from one or more vibration sensors attached to the device; a first analysis step for determining whether the total vibration value during a predetermined period selected from the collected real-time vibration values ​​is above a value of interest; a second analysis step for determining whether the same vibration pattern occurs periodically per unit time when the total vibration value is above the value of interest; and an artificial intelligence analysis step for deriving the cause of the fault based on a fault diagnosis learning algorithm learned through deep learning using fault simulation vibration data when the same vibration pattern occurs periodically more than a specified number of times.

[0016] Preferably, the fault diagnosis learning algorithm incorporates the Global Average Pooling (GAP) technique into the CNN learning algorithm, which is a deep learning method, thereby significantly reducing the size of the learning model of the fault-simulated vibration data so that diagnosis results can be derived more quickly and concisely, thus having features optimized for diagnosing multiple devices.

[0017] Preferably, the fault simulation vibration data is characterized as a hybrid vibration value that combines a normal vibration value measured when the device is in a normal state with a vibration value simulated for each major fault.

[0018] Preferably, the normal vibration value is characterized as a measurement value taken after determining the normal operating state following the initial installation or periodic inspection of the device.

[0019] Preferably, the above normal vibration value is characterized by including the background vibration value of the site where the device is installed.

[0020] Preferably, the vibration values ​​simulated for each major failure are characterized by analyzing vibration characteristics according to the failure patterns of the major components of the device and simulating them based on this analysis.

[0021] Preferably, in the data collection step, the vibration sensor is installed in three axes (Vertical, Horizontal, Axial) at the part where rotational force is transmitted from the driving unit, and is characterized by being able to monitor changes in continuous vibration values ​​transmitted according to the rotational force.

[0022] Preferably, the first analysis step is characterized by determining whether there is an abnormality based on the interest value set for each of the vibration sensors.

[0023] Preferably, the same vibration pattern subject to analysis in the artificial intelligence analysis step is characterized by being converted into acceleration and velocity units, and then utilized in a fault diagnosis learning algorithm after analyzing each preprocessed signal.

[0024] Preferably, the preprocessed signal analysis is such that the signal converted into speed units undergoes BandPass Filter, Window Filter, and Short Time Fourier Transform (STFT) processes, wherein the signal converted into speed units is a low-frequency band of 10 to 2,000 Hz in the frequency domain, and the corresponding band is utilized to analyze the overall vibration energy of the rotating body, and the signal converted into acceleration units undergoes HighPass Filter, Window Filter, and Short Time Fourier Transform (STFT) processes, wherein the signal converted into acceleration units is a high-frequency band of 2,000 Hz or higher in the frequency domain, and the corresponding band easily captures even small changes in high frequency and analyzes them to detect failures in the early stages, thereby having the characteristic of simultaneously analyzing low-frequency and high-frequency regions in the vibration pattern of the device to simultaneously diagnose the overall vibration state of the device and signs of early failure. Effects of the invention

[0026] According to the artificial intelligence-based fault diagnosis method for a rotating device according to the present invention, by predicting and detecting large and small problems that may occur in the rotating device at an earlier stage, it is possible to preemptively replace or repair specific parts, thereby enabling the stable operation of the rotating device.

[0027] Furthermore, according to the present invention, various fault simulation vibration data are generated and continuously trained using a fault diagnosis learning algorithm, thereby enabling more accurate fault diagnosis as time progresses, which has the effect of providing a highly reliable fault diagnosis prediction method. Brief explanation of the drawing

[0029] FIG. 1 is a flowchart of an artificial intelligence-based fault diagnosis method for a rotating body device according to the present invention. Figure 2 is a conceptual diagram of a fault diagnosis learning model. Figure 3 is an example diagram of a processor for a fault diagnosis learning model. Figure 4 is an example diagram of a method for acquiring fault-simulated vibration data to be used in a fault diagnosis learning model. Figure 5 is an example of an algorithm for obtaining fault simulation vibration values ​​(data). Figure 6 is a conceptual diagram of CNN+GAP applied to a fault diagnosis learning algorithm (model). Figure 7 is an example of a comparison of the number of parameters between CNN and CNN+GAP. Figure 8 is a conceptual diagram of the STFT (Short Time Fourier Transform). Specific details for implementing the invention

[0030] Hereinafter, a more detailed explanation of an artificial intelligence-based fault diagnosis method for a rotating body device according to the present invention will be provided with reference to the attached drawings. However, the drawings and specific descriptions provided are intended to illustrate one feasible example according to the technical concept of the present invention, and the technical scope of protection of the present invention is not limited thereto.

[0032] Figure 1 attached is a flowchart of an artificial intelligence-based fault diagnosis method for a rotating body device according to the present invention; Figure 2 is a conceptual diagram of a fault diagnosis learning model; Figure 3 is an example diagram of a processor of a fault diagnosis learning model; Figure 4 is an example diagram of a method for acquiring fault-simulated vibration data to be used in a fault diagnosis learning model; Figure 5 is an example diagram of an algorithm for acquiring fault-simulated vibration values ​​(data); Figure 6 is a conceptual diagram of CNN+GAP applied to a fault diagnosis learning algorithm (model); Figure 7 is an example diagram comparing the number of parameters of CNN and CNN+GAP; and Figure 8 is a conceptual diagram of a Short Time Fourier Transform (STFT).

[0034] As illustrated in FIG. 1, the artificial intelligence-based fault diagnosis method for a rotating body device according to the present invention may include a data collection step (S100), a first analysis step (S200), a second analysis step (S300), and an artificial intelligence analysis step (S400).

[0035] The data collection step (S100) means collecting continuous real-time vibration values ​​generated by a device through one or more vibration sensors attached to a device among the rotating equipment that requires continuous management.

[0036] Here, rotating equipment can be various industrial sites such as ships and factories, and refers to a system equipped with a motor to generate rotational force and then transmit that force to where it is needed.

[0037] Vibration sensors can be distributed among the major components constituting the rotating equipment, and multiple vibration sensors may be installed in a single component. Preferably, for critical components, one vibration sensor is installed in each component along three axes to collect and monitor vibration values ​​transmitted in each axis direction. By installing one vibration sensor in each component along three axes, basic data can be obtained to more accurately identify problems with the component. In other words, since multiple components related to rotation are used in a single component, the cause of failure can be analyzed more accurately through changes in vibration values ​​from each vibration sensor installed along three axes.

[0038] Following the data collection step (S100), the first analysis step (S200) follows, and it is determined whether the overall vibration value during a selected predetermined period among the real-time vibration values ​​collected through the data collection step (S100) is greater than or equal to the level of interest.

[0039] Preferably, in the first analysis step (S200), an overall calculation is performed using the overall vibration data over a predetermined period, and the overall is a representative value indicating the overall vibration level of the machine by summarizing the vibration data measured in a specific frequency range into a single number.

[0040] The types of overall include overall velocity, overall acceleration, and overall envelope.

[0041] Overall Velocity is the most common indicator for evaluating the overall vibration level of a machine and is used to comprehensively represent its vibration status. This value is calculated by integrating the acceleration signal to convert it into velocity, applying a Band-Pass Filter (BPF) to extract specific frequency bands, and finally performing a Root Mean Square (RMS) operation.

[0042] Overall Acceleration is a value representing the magnitude of a machine's vibration acceleration and is used to evaluate vibration levels, particularly those characterized by strong impact or prominent high-frequency components. This value is calculated by applying a band-pass filter (BPF) to the original signal to extract only signals within a specific frequency band, followed by a Root Mean Square (RMS) operation.

[0043] The Overall Envelop is a value that emphasizes high-frequency shock components in a machine's vibration signal. By highlighting small vibration signals, it is utilized for early defect detection in bearings and gears. This value is calculated by applying a High-Pass Filter (HPF) to the acceleration signal to remove low-frequency components, followed by Envelopment Detection, which takes the absolute value of the signal. Subsequently, a Low-Pass Filter (LPF) is applied, and finally, a Root Mean Square (RMS) calculation is performed to determine the Overall Envelop value.

[0045] The vibration sensor attached to the device continuously measures and transmits real-time vibration values ​​when the device is operated, and the real-time vibration values ​​are transmitted to the computational unit of the system to analyze the overall vibration values ​​over a set period.

[0046] The system for the present invention allows an administrator to select real-time vibration values ​​for performing a first analysis step (S200). That is, among the real-time vibration values ​​that are continuously measured and transmitted, a determination is made as to whether the overall vibration values ​​sampled over a predetermined period are greater than or equal to a value of interest.

[0047] In other words, the administrator can set the system so that the first analysis step (S200) is performed in real time, and can set the cycle for performing the first analysis step to be shorter or longer by taking into account the industrial environment, such as 5-minute intervals or 10-minute intervals, and the amount of data for analysis is not fixed, but can be 1 minute or 10 minutes instead of the real-time vibration value for 1 second.

[0048] In addition, in this embodiment, real-time vibration values ​​collected from one device among the rotating equipment are analyzed, but real-time vibration values ​​transmitted from multiple devices simultaneously are not excluded.

[0049] If the total vibration value during a selected predetermined period among the real-time vibration values ​​through the first analysis step (S200) is greater than or equal to the value of interest, it proceeds to the second analysis step (S300).

[0050] Since the possibility of an anomaly is detected in the first analysis step (S200), the second analysis step (S300) determines whether the same vibration pattern occurs periodically per unit time. That is, it determines whether the pattern of the sampling vibration value detected in the first analysis step (S200) occurs periodically per predetermined unit time. Here, the predetermined unit time refers to a time longer than the predetermined period of the sampling vibration value in the first analysis step (S200). For example, if the measurement time of the sampling vibration value is 1 minute, the unit time can be set to 30 minutes.

[0051] The unit time set in the second analysis step (S300) is intended to determine whether the same vibration pattern occurs periodically, so it is desirable to set the unit time as long as possible.

[0052] This further improves the accuracy of fault diagnosis by identifying similar vibration patterns that occur repeatedly within a unit of time in the second analysis stage (S300), thereby classifying signals unrelated to the fault, such as temporary vibration signals that may occur in the field.

[0053] If it is confirmed through the second analysis step (S300) that the same vibration pattern per unit time is periodically generated more than a specified number of times, the next step, the artificial intelligence analysis step (S400), follows.

[0055] In the artificial intelligence analysis stage (S400), the cause of the device's failure is identified, specifically by finding the cause of the failure based on an artificial intelligence learning algorithm for failure diagnosis. This algorithm is intended to enable deep learning to be performed by utilizing failure simulation vibration data to supplement the insufficient failure data in the initial stage.

[0056] In industrial sites, a time-based maintenance method is adopted in which appropriate maintenance cycles and maintenance periods are set for each piece of equipment and maintenance is performed when the scheduled maintenance time approaches. However, this method involves performing maintenance when a certain period arrives even if no breakdown occurs, which presents a practical problem in that it is difficult to obtain failure vibration data.

[0057] In particular, the fault diagnosis learning algorithm proposed in this invention combines the Global Average Pooling (GAP) technique with the Convolutional Neural Network (CNN) learning algorithm, which is one of the deep learning methods, thereby significantly reducing the size of the fault simulation vibration data learning model and lowering the complexity of the model, thereby greatly improving the speed of learning and inference. Specifically, the GAP technique effectively mitigates the problem of overfitting by replacing the existing Fully Connected Layer and maximizes the generalization performance of the learned model. Thanks to this structural efficiency, rapid and accurate results can be obtained even in environments where multiple devices must be diagnosed simultaneously, thereby significantly improving the scalability and practicality of the fault diagnosis system.

[0058] Meanwhile, the fault simulation vibration data used in the fault diagnosis artificial intelligence learning algorithm is based on vibration values ​​measured by the device in a normal state, and the fault simulation vibration data is acquired as shown in Fig. 4. These vibration values ​​reflect the characteristics occurring in the environment where the device is installed. This is to reflect the fact that normal vibration values ​​may vary depending on the environment in which the device is installed. These normal vibration signal values ​​are combined with vibration values ​​simulated according to major fault types to form hybrid vibration values.

[0059] As shown in FIG. 4, in order to derive the hybrid vibration value, the normal vibration value of the device must be obtained, and preferably, the normal vibration value is a measurement value taken when the device is determined to be in a normal operating state after initial installation or periodic inspection.

[0060] Generally, when a device is installed on-site, a test run is conducted to determine whether it has been properly installed, and if this test run is determined to be normal operation, the installation is completed. Therefore, when the device is initially installed, the measured value obtained after the test run, when it is determined to be normal operation, can be set as the normal vibration value. Meanwhile, since periodic inspections are required for the device, actual vibration values ​​are measured at set inspection intervals to determine whether it is in a normal state, and if it is normal, the actual measured vibration value can be set as the normal vibration value.

[0061] Most rotating devices are bound to deteriorate over time, and therefore, the normality of the device will have a specific band. That is, when vibration values ​​are used as the main variable for determining normality, normal vibration values ​​can be defined as A±α.

[0062] Therefore, it is highly likely that the vibration value at the time of the initial installation of the device differs from the vibration value measured during subsequent periodic inspections, and if each measured vibration value falls within the range of A±α, it is recognized as a normal vibration value.

[0063] More preferably, the normal vibration value should include the background vibration value of the site where the device is installed. Here, the background vibration value refers to vibration values ​​generated in the environment or surroundings of the site where the device is installed; the background vibration value is defined as the vibration value measured during the time when the most noise or vibration occurs, taking into account the characteristics of the site where the device is installed. Therefore, it is preferable to measure continuous vibration values ​​over a day or several days at the site where the device is to be installed to identify the point in time when the most severe vibration occurs, and to set the range of the normal vibration value by adding the most severe vibration value generated at the site as the background vibration value to the normal vibration value. Alternatively, it is acceptable to operate the device during the time when vibration is most severe and set the measured value at that time as the normal vibration value.

[0064] Meanwhile, simulated vibration values ​​for major failures can be defined as values ​​simulated based on an analysis of vibration characteristics according to the failure patterns of the device's major components. Preferably, vibration characteristic data according to the failure patterns of major components of a rotating device must be obtained; to this end, the major components of the rotating device are artificially induced into a failure state to measure the vibration values.

[0065] For example, vibration values ​​resulting from major defect types in rotating machinery, such as bent shafts, stator eccentricity, mechanical looseness, mass imbalance, and rotor rod breakage, must be secured. Data on vibration values ​​induced by the failure (defect) status of key components related to the basic operation of the rotating machinery must be obtained.

[0067] In this way, meaningful data can be generated from fault-simulated vibration data by combining normal vibration values ​​with vibration values ​​simulated for major failures. Figure 5 is an example of an algorithm for obtaining fault simulation vibration values ​​(data).

[0068] As shown in Fig. 5, based on actual vibration values ​​measured from equipment or devices, frequency components are calculated, amplitude and phase are calculated, and then signals are reconstructed for each frequency domain.

[0069] Signal reconstruction by frequency domain determines whether the signal falls within the fault signal frequency domain; if the signal falls within the fault signal frequency domain, it is classified and stored as the fault signal amplitude.

[0070] If the signal is a normal signal, it is classified as such, and then a fault signal is generated through signal reconstruction. In other words, the signal can be reconstructed by reflecting the characteristics of a fault signal into a signal classified as normal.

[0071] Through an algorithm such as that shown in Fig. 5, actual normal signals and fault signals are acquired, making it possible to acquire and accumulate fault simulation signals. Preferably, when a signal is classified as a fault signal, the specific part of the actual device or equipment where the problem is located is identified to determine the fault characteristics of each major component in a more detailed and accurate manner.

[0072] For example, when a piece of equipment or device is in a normal state, a normal signal can be acquired and analyzed according to the algorithm of Fig. 5, and then a fault signal can be acquired and analyzed by artificially manipulating a specific part to cause damage. Through this difference before and after, vibration characteristics for each fault according to the major components can be obtained, and by adjusting the degree of change (intensity) at a specific frequency, a wider range of simulated vibration values ​​for major faults can be created.

[0074] The process of acquiring (deriving) fault simulation vibration data (signal) {hybrid fault signal} is as follows.

[0075] First, obtain normal vibration data of the device in a normal environment.

[0076] Second, obtain defect vibration frequency characteristic information with characteristics specific to each defect type.

[0077] Third, the steady-state vibration data in the time domain is converted to the frequency domain using frequency analysis techniques.

[0078] Fourth, a frequency band in which defect vibration frequency characteristics appear in the frequency domain of normal vibration data is extracted and replaced with defect vibration frequency analysis feature information in the frequency domain. At this time, the phase information in the frequency domain follows the phase information of the normal vibration data.

[0079] Fifth, hybrid fault data in the frequency domain, which combines normal vibration data and fault vibration frequency characteristic information, is converted into signal data in the time domain.

[0081] Preferably, the acquisition of failure simulation vibration data enables the acquisition of more diverse and accurate data through machine learning reinforcement learning. Rotating devices include a number of key components, and even within a single component, the failure patterns can vary. For example, in the case of a bearing that supports a shaft at a precise position to ensure stable rotation of a shaft transmitting rotational force, the components include an inner ring, an outer ring, balls or rollers, and seals. A problem may occur in any of these components, and the vibration characteristics induced by each element differ, and the vibration characteristics also vary depending on the degree of defect.

[0082] The simulated vibration values ​​for each major failure are generated based on actual measurements taken when damage occurs to a specific component. Ideally, vibration values ​​in a defective state are acquired for the specific component multiple times and combined with normal vibration values ​​to obtain simulated failure vibration data. Meanwhile, rich simulated failure vibration data can be accumulated by predicting vibration values ​​induced by the damage and severity of major components forming the rotating device through a reinforcement learning model based on a failure diagnosis learning algorithm.

[0083] As fault simulation vibration data is continuously accumulated through reinforcement learning, it generates more accurate prediction values, thereby gradually increasing the reliability of fault diagnosis. Additionally, by securing fault simulation vibration data for each of the major components constituting the rotating device, there is an advantage in being able to identify the cause of the fault more precisely during fault diagnosis.

[0085] A specific example of a fault diagnosis learning algorithm (reinforcement learning model) is explained as follows.

[0086] As shown in Fig. 3, the failure simulation vibration data for a rotating device is a hybrid vibration value that combines normal vibration values ​​with vibration values ​​simulated for major failures, so countless hybrid vibration values ​​can be generated, and when these hybrid vibration values ​​are input into a failure diagnosis learning algorithm, the computing device outputs the learning and evaluation results.

[0087] The hybrid vibration value is generated by converting input data into acceleration and velocity units, performing a Short Time Fourier Transform (STFT) operation as shown in the conceptual diagram of Fig. 8, and then visualizing the generated data to use as input data for an artificial intelligence learning algorithm. That is, the input signal passes through an acceleration unit conversion unit and a velocity unit conversion unit.

[0089] The acceleration component generated while passing through the acceleration unit conversion unit passes through a HighPass Filter and then a Window Filter, followed by feature extraction (region above 2,000 Hz) through a Short Time Fourier Transform (STFT).

[0090] Then, the velocity component generated while passing through the velocity unit conversion unit passes through a BandPass Filter and then a Window Filter, followed by feature extraction (region below 2,000 Hz) through the STFT (Short Time Fourier Transform).

[0091] After undergoing a preprocessing step for such hybrid vibration values, CNN and GAP learning algorithms are executed to finally output the learning and evaluation results.

[0092] As shown in Figure 6, the CNN+GAP learning algorithm is designed to compensate for the shortcomings of CNN, and through this, the input hybrid vibration values ​​are visualized to extract failure simulation vibration data for each major failure.

[0093] In CNN algorithms, training is performed using the Fully Connected Layer of the DNN algorithm after applying filtering techniques to new layers of the Convolutional Layer and Pooling Layer. However, when image-based data is used as input data, the following problems occur in the Fully Connected Layer.

[0094] In other words, when the size of the input data increases, the number of learning parameters also increases rapidly due to the rapid increase in input neurons.

[0095] In addition, in the case of multidimensional input data, there is correlation between adjacent data, but this correlation is lost when serialized to be used as input data for a Fully Connected Layer.

[0096] Furthermore, since the overall topology of the input data cannot be considered, it is vulnerable to variations in the input data; consequently, due to the large amount of training data and numerous parameters, not only does the training computation time increase, but it also becomes difficult to expect high performance.

[0097] To solve these problems, Global Average Pooling (GAP) is applied to replace the Fully Connected Layer of the CNN algorithm as shown in Fig. 6. As shown in Fig. 6, the GAP operation method calculates the arithmetic mean of all features of the same channel and creates a vector with as many elements as the number of channels.

[0098] Therefore, since the channel values ​​are output as a single average value and appear as a one-dimensional vector, it does not matter what size the input is of. Furthermore, since the operation reduces the input to a simple one-dimensional vector, no additional parameters are added, making it advantageous from a learning perspective. In addition, because the number of parameters does not increase to the level of a Fully Connected Layer, it is effective in terms of preventing overfitting. Moreover, the reduction in the number of learning parameters leads to a decrease in computational load and an increase in computational speed. Figure 7 is an example diagram comparing the number of parameters of CNN and CNN-GAP.

[0099] As described above, the fault diagnosis learning algorithm applied in the present invention enables learning by securing a hybrid vibration value, which combines vibration values ​​at the time of defects (abnormalities) of various parts (elements) constituting a rotating device, as highly reliable fault simulation vibration data.

[0100] Through the establishment of a fault diagnosis learning algorithm, the artificial intelligence analysis stage is performed, making it possible to identify the presence of abnormalities in the rotating device and the causes of failure.

[0101] If the same vibration pattern occurs periodically from a rotating device more than a specified number of times, an artificial intelligence analysis step is performed to analyze the vibration pattern.

[0102] After the data collection stage, if the total vibration value during a predetermined period selected through the first analysis stage exceeds a value of interest, the second analysis stage determines whether the same vibration pattern occurs periodically per unit time, and if the same vibration pattern occurs periodically more than a specified number of times, the vibration pattern expected to be problematic is input into the fault diagnosis learning algorithm in the artificial intelligence analysis stage, and after undergoing a predetermined analysis process, the cause of the fault is automatically derived. Industrial applicability

[0104] The present invention is expected to be utilized as a useful tool for the stable maintenance of rotating equipment.

[0105] No content Explanation of the symbols

[0106] S100: Data collection stage S200: 1st analysis stage S300: Secondary analysis stage S400: Artificial Intelligence Analysis Stage

Claims

Claim 1 A data collection step for collecting continuous real-time vibration values ​​generated from one or more vibration sensors attached to a device; a first analysis step for determining whether the total vibration value during a predetermined period selected from the collected real-time vibration values ​​is above a value of interest; a second analysis step for determining whether the same vibration pattern occurs periodically per unit time if the total vibration value is above the value of interest; and an artificial intelligence analysis step for deriving the cause of a failure based on a deep learning-trained fault diagnosis algorithm using fault simulation vibration data if the same vibration pattern occurs periodically more than a specified number of times; wherein, in the artificial intelligence analysis step, the same vibration pattern subject to analysis is classified and converted into acceleration and velocity units, and then utilized in the fault diagnosis learning algorithm after analyzing each preprocessed signal, and wherein the preprocessed signal analysis is performed such that the signal converted into velocity units undergoes BandPass Filter, Window Filter, and Short Time Fourier Transform (STFT) processes, wherein the signal converted into velocity units is a low-frequency band of 10 - 2,000 Hz in the frequency domain, and the corresponding band is a rotating body An artificial intelligence-based fault diagnosis method for a rotating device, characterized by utilizing the overall vibration energy analysis, and ensuring that the signal converted into acceleration units undergoes HighPass Filter, Window Filter, and Short Time Fourier Transform (STFT) processes, wherein the signal converted into acceleration units is a high-frequency band of 2,000 Hz or higher in the frequency domain, and by utilizing the analysis of this band to easily detect even small changes in high frequency and to detect faults in the early stages, thereby simultaneously analyzing low-frequency and high-frequency regions in the vibration pattern of the device to simultaneously diagnose the overall vibration state of the device and signs of early failure. Claim 2 In claim 1, the fault diagnosis learning algorithm incorporates the Global Average Pooling (GAP) technique into the CNN learning algorithm, which is a deep learning method, thereby significantly reducing the size of the learning model of the fault-simulated vibration data so that diagnosis results can be derived more quickly and concisely, thereby providing an artificial intelligence-based fault diagnosis method for a rotating device that has features optimized for diagnosing multiple devices. Claim 3 An artificial intelligence-based fault diagnosis method for a rotating body, characterized in that, in claim 1, the fault simulation vibration data is a hybrid vibration value obtained by combining a normal vibration value measured when the device is in a normal state with a vibration value simulated for each major fault. Claim 4 An artificial intelligence-based fault diagnosis method for a rotating body device, characterized in that, in claim 3, the normal vibration value is a measurement value measured after determining a normal operating state following the initial installation or periodic inspection of the device. Claim 5 An artificial intelligence-based fault diagnosis method for a rotating body device, characterized in that, in claim 4, the normal vibration value includes the background vibration value of the site where the device is installed. Claim 6 An artificial intelligence-based fault diagnosis method for a rotating body, wherein, in claim 3, the simulated vibration values ​​for each major failure are simulated based on an analysis of vibration characteristics according to the failure pattern of the major components of the device. Claim 7 An artificial intelligence-based fault diagnosis method for a rotating body device, characterized in that, in the data collection step, the vibration sensor is installed in three axes (Vertical, Horizontal, Axial) at the part where rotational force is transmitted from the driving unit, and changes in continuous vibration values ​​transmitted according to the rotational force are monitored. Claim 8 An artificial intelligence-based fault diagnosis method for a rotating body device according to claim 1, characterized in that the first analysis step determines whether there is an abnormality based on a value of interest set for each of the vibration sensors. Claim 9 delete Claim 10 delete

Images