Life prediction system and life prediction method

Abstract

This invention provides a life prediction system and method that can predict the lifespan of gearboxes and other components in a manner that reflects actual conditions. [Solution] A life prediction system 1 predicts the life of a mechanism (for example, a speed reducer) that can adjust the rotational speed of a motor by combining gears with different numbers of teeth, comprising: a vibration sensor that acquires vibration data of the mechanism; a monitoring device 20 that monitors the rotational speed of the motor and collects the vibration data in a steady state; and a life prediction device 30 that receives the vibration data from the monitoring device 20 and predicts the life of the mechanism. The life prediction device 30 includes a first prediction unit 33 that uses the vibration data corresponding to stepwise speeds to predict physical quantities related to vibration within a short first period and predicts abnormalities in the speed reducer that will occur within the first period.

Description

Technical Field

[0001] The present invention relates to a life prediction system and a life prediction method for a speed reducer used in a crane or the like.

Background Art

[0002] In the case of a machine using parts with a long lead time (long lead time parts) from the order to the delivery, it is a problem that the machine cannot be used for a long time due to the failure of the long lead time parts. For example, the speed reducer used in a harbor crane falls under long lead time parts, and there is a problem that the crane cannot be used for a long time due to the failure of the speed reducer. In some local ports, there is only one crane, and in that case, the impact is particularly large. Therefore, it is important to predict the life of the machine and perform preventive maintenance.

[0003] Currently, various technologies for predicting the life of machines are being developed (see, for example, Patent Document 1). The system described in Patent Document 1 predicts the life of a speed reducer. It detects the load applied to the bearing with a load sensor and predicts the life by comparing it with a pre-made S-N diagram.

Prior Art Documents

Patent Documents

[0004] [[ID=*]] [[ID=*]]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] The system described in Patent Document 1 has the problem that it can only make long-term predictions based on bearing fatigue failure. In other words, although there are various factors other than bearing fatigue failure that can cause speed reducer failure (for example, gear damage or foreign matter jamming), it is difficult to predict failures based on these factors. Therefore, conventionally, there has been a problem in that it is difficult to predict the lifespan of speed reducers in a way that reflects actual conditions. This problem can also occur in mechanisms other than speed reducers that can adjust the rotational speed of a motor by combining gears with different numbers of teeth (for example, a transmission).

[0006] Therefore, the object of the present invention is to provide a lifespan prediction system and a lifespan prediction method that can perform lifespan predictions that are in line with actual conditions. [Means for solving the problem]

[0007] The present invention relates to a life prediction system for predicting the life of a mechanism capable of adjusting the rotational speed of a motor by a combination of gears with different numbers of teeth, comprising: a vibration sensor that acquires vibration data of the mechanism; a monitoring device that monitors the rotational speed of the motor and collects the vibration data in a steady state; and a life prediction device that receives the vibration data from the monitoring device and predicts the life of the mechanism, wherein the life prediction device comprises: a vibration data storage unit that stores the vibration data for each stepwise speed that the mechanism can change; and a first prediction unit that uses the vibration data corresponding to the stepwise speed to predict physical quantities related to vibration within a short first period and predicts abnormalities of the mechanism that will occur within the first period.

[0008] According to the present invention, the first prediction unit can predict short-term abnormalities such as gear damage and foreign object jamming. Therefore, it is possible to predict the lifespan in a manner that reflects actual conditions. Furthermore, when the vibration sensor is installed in the case, it is easier to "add later" compared to when the load sensor is installed in the bearing, and a dedicated design as a reduction gear is not required, for example. Replacement in the event of failure is also easy. In addition, since vibration data in a steady state is collected and used for prediction, it is possible to improve the accuracy of the prediction.

[0009] The life prediction device may be configured such that the vibration sensor is mounted around the bearing where the rotational speed of the reduction gear shaft is high, and acceleration in the axial, radial, and circumferential directions is measured.

[0010] Since axial vibration acceleration does not exhibit seasonal variations, it is preferable to collect 3-axis acceleration data using a vibration sensor, as in the present invention.

[0011] The life prediction device may be configured to include a state data storage unit that continuously compares predicted physical quantities related to vibration with actually measured physical quantities related to vibration over time and stores the changes in the comparison results over time, and a second prediction unit that predicts the deterioration of the mechanism over time in a second period which is longer than the first period, based on the changes in the comparison results over time.

[0012] According to the present invention, the second prediction unit can predict long-term abnormalities due to deterioration over time (for example, wear and surface fatigue). Therefore, it is possible to predict the lifespan in a way that is more in line with actual conditions.

[0013] Furthermore, it is preferable that the first prediction unit performs frequency analysis on the vibration data and predicts physical quantities related to vibration within the first period using the frequencies of the meshing vibrations of the multiple gears in the mechanism, and information on the frequency bands where there is no superposition and information on the frequency bands where there is superposition among the harmonics. For example, the frequency band in which the three gear vibrations are superimposed (e.g., the 750 Hz band) is used for overall diagnosis, and the respective frequency bands where there is no superposition are used for individual diagnosis.

[0014] According to the present invention, abnormalities in each gear can be predicted with high accuracy.

[0015] Furthermore, the first prediction unit may have a short-term prediction model that predicts physical quantities related to vibration within the first period, and may extract feature quantities using the results of frequency analysis and the autoregressive parameters of the autoregressive model as explanatory variables for the short-term prediction model.

[0016] Furthermore, the monitoring device may monitor the motor rotation speed, determine when a trigger occurs when the motor transitions from a constant speed state to a speed change state, and collect the vibration data immediately before the trigger occurs.

[0017] Furthermore, the second prediction unit may predict the deterioration of the mechanism over time based on the relationship between the prediction error of the first prediction unit and a predetermined threshold value for the prediction error.

[0018] Furthermore, the second prediction unit may use an anomaly score calculated with different prediction parameters than those used in the first prediction unit.

[0019] Furthermore, the first prediction unit may calculate an anomaly degree from the predicted physical quantity related to vibration and the physical quantity related to vibration that was actually measured, and predict an anomaly in the mechanism based on the anomaly degree.

[0020] According to the present invention, it is possible to predict anomalies that are difficult to verify from physical quantities related to vibration. However, it is thought that the degree of anomaly changes as changes occur over time, and the accuracy of short-term anomaly detection decreases. In order to suppress this in operation, it is preferable that the second prediction unit uses an anomaly degree calculated with different prediction parameters than the first prediction unit.

[0021] The first prediction unit may predict anomalies based on a statistical prediction model (for example, principal component analysis) using coefficients from an autoregressive model that predicts vibration acceleration in multiple meshing vibration frequency bands of the gearbox or their harmonic frequency bands.

[0022] According to this invention, false alarms can be suppressed because the model is constructed from various angles. In operation, minimizing false alarms is a crucial point.

[0023] Furthermore, the first prediction unit has a plurality of short-term prediction models that predict physical quantities related to vibration within the first period, and each of the short-term prediction models may predict abnormalities in the mechanism that occur within different first periods.

[0024] Further, the first prediction unit may predict an abnormality from a weighted sum of degrees of abnormality output by a plurality of short-term prediction models.

[0025] Also, the prediction model used in the first prediction unit may use prediction parameters learned based on vibration acceleration data when a certain period has elapsed after oil replacement of the speed reducer.

[0026] Also, the state data accumulation unit creates a histogram regarding the prediction error of the first prediction unit, the second prediction unit obtains a representative value of the prediction error based on the histogram, and based on the relationship between an approximate curve representing the relationship between the passage of time and the representative value and the threshold value, the deterioration over time of the mechanism may be predicted.

[0027] Also, the mechanism may be provided in a crane and may be capable of changing one or both of the lateral travel speed and the hoisting speed stepwise.

[0028] The life prediction method of the present invention is a life prediction method for predicting the life of a mechanism capable of adjusting the rotational speed of a motor by a combination of gears with different numbers of teeth, and includes a monitoring step of monitoring the rotational speed of the motor and collecting vibration data of the mechanism in a steady state, and a life prediction step of receiving the vibration data and predicting the life of the mechanism. In the life prediction step, for each stepwise speed that the mechanism can change, a vibration data accumulation step of accumulating the vibration data, and a first prediction step of predicting a physical quantity related to vibration within a short-term first period and predicting an abnormality of the mechanism occurring within the first period are included.

[0029] According to the present invention, the first prediction step makes it possible to predict short-term abnormalities such as gear damage and foreign object jamming. Therefore, it is possible to predict the lifespan in a manner that reflects actual conditions. Furthermore, when the vibration sensor is installed in the case, it is easier to "add later" compared to when the load sensor is installed in the bearing, and a dedicated design as a reduction gear is not required, for example. Replacement in the event of failure is also easy. In addition, since vibration data in a steady state is collected and used for prediction, it is possible to improve the accuracy of the prediction. [Effects of the Invention]

[0030] According to the present invention, it is possible to predict lifespan in a manner that is in line with actual conditions. [Brief explanation of the drawing]

[0031] [Figure 1] This is a schematic diagram of a life prediction system according to an embodiment of the present invention. [Figure 2] This figure shows the data flow in a life prediction system according to an embodiment of the present invention. [Figure 3] This diagram illustrates the trigger conditions for transferring vibration data to the cloud server. Figure 3(a) is an example of a diagram showing the relationship between the passage of time and motor rotation speed. Figure 3(b) is an enlarged view of Figure 3(a). Figure 3(c) is an example of vibration data acquired by the agent PC. [Figure 4] This diagram illustrates the timing of trigger occurrence. [Figure 5] This is a block diagram relating to life prediction in a life prediction system according to an embodiment of the present invention. [Figure 6] These diagrams illustrate gear vibration. Figure 6(a) is a schematic diagram of meshing gears, Figure 6(b) is an illustrative diagram of vibration data when there is a gear abnormality, and Figure 6(c) is the Fast Fourier Transform of the vibration data shown in Figure 6(b). [Figure 7] This shows the relationship between meshing vibrations and sideband waves using mathematical formulas. [Figure 8] This graph shows the relationship between the frequency and order of the meshing vibration of a gearbox. [Figure 9] There is an illustrative diagram of the short-term prediction model. [Figure 10] This compares the results of vibration acceleration predictions using a short-term prediction model with the measured values. [Figure 11] This is an example of a histogram related to prediction error. [Figure 12] This is an example of an approximation curve that shows the relationship between time progression and the representative value of the prediction error. [Figure 13] This diagram illustrates variations in short-term prediction models. [Figure 14] This is an example of a cepstrum, which is a vibrational acceleration generated in a gearbox. [Figure 15] This is a schematic diagram of an autoencoder. [Figure 16] This is an example of an autoencoder design. [Figure 17] There is a diagram comparing the cepstrum before and after the autoencoder input. [Figure 18] This graph shows the difference (loss) between the input and output of an autoencoder. [Figure 19] This is a schematic diagram of an LSTM. [Figure 20] This is an example of an LSTM design. [Figure 21] This is the result of predicting vibration acceleration using LSTM, which utilizes the cepstrum features of vibration acceleration. [Figure 22] This is an example of the results of detecting anomalies. [Figure 23] This is one example of a method for calculating the degree of abnormality. [Figure 24] This is a block diagram relating to the life prediction of the life prediction system for modified cases. [Figure 25] This diagram illustrates the processing performed by the cloud server in the modified example. [Figure 26] This is an example of a learning flow during an oil change. [Figure 27] This is an example of how the anomaly score is calculated for each model. [Figure 28]This is a diagram illustrating how the VAE model calculates the degree of abnormality from cepstrum. [Figure 29] This figure shows an example of calculating the degree of abnormality using PCA (Principal Component Analysis). [Figure 30] This is an example of the calculation results for long-term abnormality. [Modes for carrying out the invention]

[0032] The life prediction system of the present invention will be described in detail with reference to the drawings. The embodiments and modifications described below are merely illustrative, and each embodiment and modification can be used in appropriate combination.

[0033] <Overall outline structure> As shown in Figure 1, the life prediction system 1 mainly comprises a crane 10, an agent PC 20, and a cloud server 30. Each component of the life prediction system 1 is connected to enable data transmission and reception via a communication network. The cloud server 30 may be a server installed in the facility that manages the crane 10 (for example, an on-premise server connected to the company's own network). In this embodiment, we will describe a case where the lifespan prediction system 1 predicts the lifespan of the speed reducer of the crane 10. However, the lifespan prediction system 1 is not limited to the speed reducer of the crane 10, and it is possible to predict the lifespan of speed reducers of machines other than the crane 10.

[0034] The crane 10 shown in Figure 1 is a machine that uses power to lift loads and transport them horizontally. In this embodiment, a gantry crane installed in a port is assumed. A gantry crane is a large bridge-shaped crane with a structure that can move on rails. Gantry cranes are installed in ports, for example, and are used for unloading containers from container ships or loading containers onto container ships.

[0035] As shown in Figure 1, the crane 10 mainly comprises wheels 11, a frame 12, a trolley 13, and a machine room 14. The "front / back," "up / down," and "left / right" directions of the crane 10 follow the arrows in Figure 1. These directions are defined for the convenience of explanation and do not limit the present invention. The wheels 11 are located at the lower part of the frame 12 and move the crane 10 in the forward and backward direction (movement in the δ direction).

[0036] The frame section 12 is composed of beams arranged in the front-to-back, left-to-right, up-and-down, and diagonal directions. The frame section 12 has a girder 12a and a boom 12b. The girder 12a is integrated with the boom 12b when the boom 12b is fixed horizontally, and allows the trolley 14 to move horizontally (in the β direction). The boom 12b is connected to the end of the girder 12a by a pin with a hinge, and has a structure that allows it to rise and fall (move in the α direction). Rails for the movement of the trolley 13 are laid on the girder 12a and the boom 12b.

[0037] The trolley 13 is a structure that traverses the girder 12a and boom 12b. The trolley 13 has a spreader 13a and an operator's cab 13b. The spreader 13a is a lifting device for lifting containers. The spreader 13a is suspended from the trolley 13 by a wire rope 19 and is structured to move vertically (move in the gamma direction) by winding up and down the wire rope 19. The spreader 13a is equipped with lifting fittings at its four corners and engages with the four corners of the container to lift it. The operator's cab 13b is located in the operator's cab 13b to operate the crane 10.

[0038] The machine room 14 is a room that houses mechanical and electrical equipment such as the hoisting device 14a, the traversing device 14b, and the luffing device 14c. The hoisting device 14a comprises a motor, a brake, a drum, and a reduction gear. A wire rope 19 is wound around the drum. By rotating the drum using the motor and feeding out or winding in the wire rope 19, the spreader 13a moves up and down. The traversing of the trolley 13 and the up and down movement of the spreader 13a can be performed simultaneously. In this embodiment, the reduction gear of the hoisting device 14a is assumed to be the target for predicting the lifespan. Detailed descriptions of the traversing device 14b and the luffing device 14c are omitted.

[0039] The agent PC 20 shown in Figure 1 is a device that controls the flow of data in the life prediction system 1. The agent PC 20 has the function of monitoring the operation of the hoisting device 14a. The agent PC 20 is an example of a "monitoring device". For example, the agent PC 20 monitors the rotational speed of the motor equipped with the hoisting device 14a. Based on the status of the hoisting device 14a, the agent PC 20 transmits the information measured by the crane 10 and necessary for life prediction to the cloud server 30.

[0040] The cloud server 30 shown in Figure 1 is built on the internet and has the function of predicting the lifespan of the speed reducer equipped in the crane 10. The cloud server 30 is an example of a "lifespan prediction device". The agent PC 20 may perform the speed reducer lifespan prediction on behalf of the cloud server 30, or together with the cloud server 30. In that case, the agent PC 20 will have all or part of the functions of the cloud server 30.

[0041] Refer to Figure 2 to explain the main data flow in the lifetime prediction system 1. Figure 2 is a diagram showing the data flow in the lifetime prediction system 1. As shown in Figure 2, the PLC (Programmable Logic Controller) 60 of the crane 10 transmits a command value for the motor rotation speed to the agent PC 20 via the HUB 70. The communication interval for the motor rotation speed command value is, for example, "100 milliseconds".

[0042] As shown in Figure 2, three vibration sensors 51 are attached to the gearbox 50. The vibration sensors 51 detect vibrations of the gearbox 50 and transmit the detection results as vibration data to the agent PC 20. The vibration sensors 51 may be mounted on the case of the gearbox 50, for example, but it is preferable to be close to the bearing. The vibration sensors 51 are, for example, acceleration pickups that sense the acceleration of vibration and convert it into an electrical signal, and in this embodiment, acceleration pickups are assumed to be the vibration sensors 51. Each vibration sensor 51 detects acceleration in three mutually orthogonal axes (x axis, y axis, and z axis). A measurement module 52 is connected to the vibration sensors 51, and the measurement module 52 converts the analog signals generated by the vibration sensors 51 into Ethernet format. The vibration data generated by the vibration sensors 51 and converted by the measurement module 52 is transmitted to the agent PC 20 via the HUB 70.

[0043] Agent PC20 transfers vibration data from the gearbox 50 to the cloud server 30. Agent PC20 monitors the command value of the motor rotation speed and sends vibration data to the cloud server 30 when the motor rotation speed is stable. For example, if Agent PC20 detects that the command value of the motor rotation speed has remained constant for a predetermined time or longer (for example, several seconds or more) and then detects a change in the command value of the rotation speed, it sends vibration data included in the few seconds before the change.

[0044] Refer to Figures 3 and 4 to explain the transmission of vibration data by the agent PC 20. Figure 3 is a diagram illustrating the trigger conditions for transferring vibration data to the cloud server 30. Figure 3(a) is an example of a diagram showing the relationship between the passage of time and the motor rotation speed. Figure 3(b) is an enlarged view of the range indicated by reference numeral 8 in Figure 3(a). Figure 3(c) is an example of vibration data acquired by the agent PC 20.

[0045] The vibration data acquired by agent PC20 (see Figure 3(c)) is created, for example, in binary format. As shown in Figure 3(a), a trigger is set to activate when a constant rotation speed is maintained for a predetermined time (here, "1.8 seconds or more"). For example, "constant rotation speed time ΔT C Let's explain the case where the trigger is applied when the value is ≥1.8 seconds. When the monitoring interval is "100 milliseconds", as shown in Figure 3(b), for example, ΔT is calculated from the change in motor speed. S The trigger time T0 is set to 0.2 seconds prior (see the range indicated by symbol 9). When a trigger occurs, the agent PC 20 transfers two binary vibration data for one vibration sensor 51 to the cloud server 30 (see Figure 3(c)).

[0046] Figure 4 is a diagram illustrating the timing of trigger generation. In the graph shown in Figure 4, the horizontal axis represents elapsed time, and the vertical axis represents the commanded motor speed. The motor speed is positive in the winding direction (the direction that raises the spreader 13a) and negative in the winding direction (the direction that lowers the spreader 13a). In Figure 4, the occurrence of the trigger is indicated by a white arrow. As shown in Figure 4, the trigger occurs when a constant rotational speed is maintained for longer than a predetermined time and the rotational speed changes.

[0047] <Functions related to lifespan prediction> Next, with reference to Figure 5, the functions related to life prediction of the life prediction system 1 will be described. Figure 5 is a block diagram related to life prediction of the life prediction system 1. The functions related to life prediction are realized, for example, by the execution of a program by the CPU. As shown in Figure 5, the agent PC 20 (monitoring device) mainly comprises a monitoring unit 21 and a transmission unit 22.

[0048] The monitoring unit 21 monitors the command value of the motor rotation speed and determines when a trigger has occurred. Specifically, the monitoring unit 21 monitors the motor rotation speed and determines when a trigger has occurred when the motor transitions from a constant speed state to a speed change state. For example, the trigger is determined to occur just before the transition from a steady state to a transient state. The transmitting unit 22 collects vibration data immediately before the trigger and transfers the collected vibration data to the cloud server 30. Other data is deleted. For example, the transmitting unit 22 collects n (n is a natural number, and in this embodiment, 2) binary format vibration data immediately before the trigger and transfers the collected vibration data to the cloud server 30. Note that the transfer of vibration data by the transmitting unit 22 does not necessarily have to be performed in real time.

[0049] As shown in Figure 5, the cloud server 30 (life prediction device) mainly comprises a data reception unit 31, a vibration data storage unit 32, a first prediction unit 33, a data display unit 34, a status data storage unit 35, and a second prediction unit 36.

[0050] The data reception unit 31 receives vibration data from the agent PC 20 for the gearbox 50 (see Figure 2). Specifically, the data reception unit 31 acquires the vibration data immediately preceding the trigger (for example, n binary vibration data immediately preceding the trigger).

[0051] The vibration data storage unit 32 stores vibration data from the reduction gear 50 (see Figure 2) received by the data reception unit 31. The vibration data storage unit 32 stores the vibration data in a storage unit such as an HDD (hard disk drive) or SSD (solid state drive) for data storage.

[0052] The main faulty parts of the reduction gear 50 (see Figure 2) are gears and bearings. In this embodiment, we will focus on the gears as the faulty parts. Meshing vibration and sideband waves are useful indicators for diagnosing gear failures. Figure 6 is a diagram illustrating gear vibration. Figure 6(a) is a schematic diagram of meshing gears, Figure 6(b) is an illustrative diagram of vibration data when there is a gear abnormality, and Figure 6(c) is the result of performing a Fast Fourier Transform (FFT) on the vibration data shown in Figure 6(b) (vibration data when there is a gear abnormality).

[0053] As shown in Figure 6(a), if there is some abnormality in some of the gear teeth, as shown in Figure 6(b), the abnormality in some teeth may be superimposed on the modulation amplitude. Also, as shown in Figure 6(c), when vibration data with abnormal gear teeth is subjected to a Fast Fourier Transform, the modulation amplitude affects the sidebands that appear on both sides of the meshing vibration order component. Figure 7 shows the relationship between meshing vibration and sidebands mathematically.

[0054] The frequency at which meshing vibrations occur is determined by the rotational speed of the shaft and the number of teeth on the gear. Therefore, it is necessary to collect vibration data for the same rotational speed and then evaluate the vibration data for each meshing vibration. The vibration data storage unit 32 stores the vibration data by clustering it according to the stepwise speeds that the reduction gear 50 can change.

[0055] The first prediction unit 33, shown in Figure 5, uses vibration data corresponding to stepwise speeds to predict physical quantities related to vibration (for example, vibration acceleration) within a short first period (for example, within one month), and predicts any abnormalities in the speed reducer 50 that may occur within the first period. The abnormalities in the speed reducer 50 predicted by the first prediction unit 33 include, for example, gear damage or foreign matter jamming.

[0056] The first prediction unit 33 reads vibration data from the vibration data storage unit 32 and performs frequency analysis on the read vibration data. In subsequent processing, it uses information of specific frequencies to predict abnormalities in the speed reducer 50 that will occur in the short term. The frequency band used by the first prediction unit 33 for prediction preferably includes the range of frequencies of the meshing vibrations of the multiple gears of the speed reducer 50 that are not superimposed. The frequency band used by the first prediction unit 33 for prediction may also include both the range of frequencies of the meshing vibrations of the multiple gears of the speed reducer 50 and their harmonics that are not superimposed and the range that is superimposed. For example, when predicting the lifespan of the speed reducer 50 of the crane 10 using multiple prediction models, it is preferable to use the superimposed frequency band for the overall diagnosis using multiple prediction models, and the frequency band that is not superimposed for the individual diagnosis using each prediction model. Here, harmonics are frequency components that are integer multiples of the waveform of the fundamental frequency (in this case, the frequency of the meshing vibration). For example, a frequency component that is twice the frequency is called the second harmonic, and a frequency component that is three times the frequency is called the third harmonic. Specifically, it is best to use both the frequency bands that do not overlap and the frequency bands that do overlap between the frequency of the nth-order (n is a natural number) meshing vibration (frequency of the meshing vibration and the nth harmonic) caused by a certain gear in the reduction gear 50 and the frequency of the nth-order (n is a natural number) meshing vibration (frequency of the meshing vibration and the nth harmonic) caused by other gears.

[0057] Figure 8 is a graph showing the relationship between the frequency and order of the meshing vibration of the gearbox 50. In Figure 8, the horizontal axis represents the frequency of the meshing vibration, and the vertical axis represents the order. Here, we assume that the gearbox 50 has three gears (1st gear M1, 2nd gear M2, and 3rd gear M3). The 1st gear M1 is on the motor side, and the 2nd gear M2 and 3rd gear M3 move away from the motor in that order. Points where the order is "0 (zero)" are plotted by dropping the points from each graph. In the example in Figure 8, for example, the meshing vibration frequency of the 1st gear M1 is set to the "6000Hz band," which is the 8th order meshing vibration; the meshing vibration frequency of the 2nd gear M2 is set to the "1028.57Hz band," which is the 4th order meshing vibration; and the meshing vibration frequency of the 3rd gear M3 is set to the "108.07Hz band," which is the 1st order meshing vibration.

[0058] The first prediction unit 33 has a short-term prediction model that predicts physical quantities related to vibration (for example, vibration acceleration) within the first period. The short-term prediction model is preferably a machine learning model such as an autoregressive model or LSTM (Long Short-Term Memory), and a model using a neural network that takes time series into account is desirable. Before making predictions, the short-term prediction model is trained in advance. The first prediction unit 33 extracts features to be used as explanatory variables for the short-term prediction model from the results of frequency analysis.

[0059] Figure 9 shows an illustrative diagram of a short-term prediction model, assuming a Recurrent Neural Network (RNN), a type of neural network that takes time series into account. Examples of explanatory variables include frequency features of meshing vibration or sideband waves, such as RMS value, peak value, kurtosis, distortion, sideband energy ratio, sideband index, sideband level coefficient, and acoustic features (one example being cepstrum). Other examples of explanatory variables may include lifting load, traverse speed, traverse position, and main hoisting time. At least one, preferably all, of these are used as explanatory variables. On the other hand, examples of dependent variables include vibration acceleration at the frequency of meshing vibration or sideband waves, and the shift amount of the sideband wave frequency peak.

[0060] The RMS (Root Mean Square) value is obtained by squaring the acceleration response for each frequency of a random wave and then taking the square root of the sum of these values. The wave height value is the ratio of the peak value to the RMS value. The peak value is the maximum value within a given interval. Kurtosis is the degree to which an vibration waveform is abrupt, and it represents the degree of sharpness of a signal. Distortion is the degree of distortion around the average value of an oscillation waveform, and represents the degree of asymmetry of the signal.

[0061] The sideband energy ratio is calculated from high-resolution spectral data. The sideband index is the average spectral amplitude from the sidebands of the first GMF. The sideband level factor is the spectral amplitude of the sidebands of the first GMF.

[0062] Acoustic features are numerical representations of the physical characteristics of sound, expressing its physical properties and characteristics. One example of an acoustic feature is the cepstrum, which represents information about the degree of change in different spectral bands.

[0063] Figure 10 compares the predicted vibration acceleration using a short-term prediction model with the measured values. In Figure 10, the horizontal axis represents the data exponent (time), and the vertical axis represents vibration acceleration. The thin solid line graph represents the actual vibration results. The dashed line graph represents the input to the autoregressive model, and the thick solid line graph represents the output (predicted result) of the autoregressive model. In Figure 10, the explanatory variables used in the short-term prediction model were "RMS value, wave height, kurtosis, strain, sideband energy ratio, sideband exponent, sideband level coefficient, and cepstrum." As shown in Figure 10, it is possible to predict the trend of vibration acceleration.

[0064] The first prediction unit 33 displays the results of the vibration prediction for a short-term first period (for example, within one month) on the data display unit 34. The data display unit 34 is, for example, a display device.

[0065] Furthermore, the first prediction unit 33 may calculate an abnormality level from the predicted physical quantity related to vibration (for example, vibration acceleration) and the actually measured physical quantity related to vibration, and predict an abnormality in the reduction gear 50 based on the abnormality level. In this way, it is possible to predict abnormalities that are difficult to verify from the physical quantity related to vibration alone. Specifically, the first prediction unit 33 continuously compares the predicted physical quantity related to vibration and the actually measured physical quantity related to vibration over time. For example, the first prediction unit 33 compares the actual vibration acceleration of the meshing vibration frequency with the prediction result to obtain the RMS value of the error over time. The change in the RMS value of the error over time indicates the state of the equipment. The first prediction unit 33 may display the change in the RMS value of the error over time on the data display unit 34. The first prediction unit 33 also passes the RMS value of the error obtained over time to the state data storage unit 35.

[0066] The state data storage unit 35 shown in Figure 5 stores the time-dependent changes in the comparison results obtained by continuously comparing the predicted physical quantities related to vibration with the physical quantities related to vibration that were actually measured over time. The state data storage unit 35 stores the RMS value of the error obtained over time in a storage unit such as an HDD or SSD for data storage.

[0067] The second prediction unit 36, shown in Figure 5, predicts the time-dependent deterioration of the speed reducer 50 in a second period that is longer than the first period, based on the time-dependent changes in the comparison results stored in the state data storage unit 35. The second prediction unit 36 ​​predicts the time-dependent deterioration of the speed reducer 50 based, for example, on the relationship between the prediction error of the first prediction unit 33 and a predetermined threshold value for the prediction error. The second prediction unit 36 ​​displays the predicted result of time-dependent deterioration over the longer second period (for example, 3 months to 6 months) on the data display unit 34.

[0068] The state data storage unit 35 creates a histogram of the prediction error of the first prediction unit 33, and the second prediction unit 36 ​​may determine representative values ​​of the prediction error (for example, RMS value, mean, variance, kurtosis, skewness, etc.) based on the histogram. Figure 11 is an example of a histogram of the prediction error. In Figure 11, the horizontal axis is the prediction error, and the vertical axis is the frequency. The prediction error is "measured value - predicted value" and can be positive or negative. If the predicted value is large compared to the measured value, it will be negative.

[0069] Furthermore, the second prediction unit 36 ​​predicts the relationship between the elapsed time to date and the representative value of the prediction error (for example, the RMS value) in a longer-term second period. For this prediction, for example, a long-term prediction model is used. An example of a long-term prediction model is a linear regression method, preferably a polynomial regression. This makes it possible to obtain an approximate curve that represents the relationship between the elapsed time and the representative value of the prediction error.

[0070] Figure 12 is an example of an approximation curve showing the relationship between the passage of time and a representative value of the prediction error. In Figure 12, the horizontal axis represents the date, and the vertical axis represents the RMS value of the prediction error. In Figure 12, a threshold value of "1" for the prediction error is set, and it is determined that a failure due to deterioration over time has occurred if this threshold is exceeded. The second prediction unit 36 ​​determines (classifies) whether the second period is normal or abnormal based on the relationship between the approximation curve and the threshold value. The threshold value can be determined from past crane performance values. Alternatively, data on similar cranes and speed reducers may be accumulated, and the threshold value may be determined based on the accumulated data.

[0071] As described above, the life prediction system 1 according to this embodiment can predict short-term abnormalities such as gear damage and foreign matter jamming by the first prediction unit 33. In addition, the second prediction unit 36 ​​can predict long-term abnormalities due to deterioration over time (for example, wear and surface fatigue). Therefore, it is possible to predict the lifespan of the gearbox 50 in a manner that reflects the actual conditions. Furthermore, when the vibration sensor 51 is installed in the case, it is easier to "add later" compared to when the load sensor is installed in the bearing, and a dedicated design for the reduction gear 50 is not required. Also, replacement in case of failure is easy. Furthermore, by collecting vibration data under steady-state conditions and using it for prediction, it is possible to improve the accuracy of the prediction. Furthermore, by using information from frequency bands where there is no superposition among the frequencies of the meshing vibrations of the multiple gears in the reduction gear 50, abnormalities in each gear can be predicted with high accuracy.

[0072] While embodiments of the present invention have been described above, the design can be modified as appropriate, as long as it does not contradict the spirit of the present invention.

[0073] For example, in the embodiment, the reduction gear 50 of the hoisting device 14a was assumed as the target for predicting lifespan, and the case in which the hoisting speed of the reduction gear 50 is changed in steps was described. However, the reduction gear 50 may also be able to change speeds other than hoisting speed (for example, traverse speed). In that case, the reduction gear 50 can change one or both of the traverse speed and hoisting speed in steps, and it is preferable to cluster the vibration data for each speed.

[0074] Furthermore, although the embodiment described a case where one short-term prediction model is used for one clustering, multiple models of different types may be used for one clustering. In that case, for example, safety can be further ensured by performing the evaluation based on the most stringent prediction result. Each short-term prediction model may predict abnormalities in the speed reducer 50 that occur within different first periods.

[0075] Furthermore, the embodiment assumed a case where the lifespan of the reduction gear 50 was to be predicted. However, the present invention can also be applied to mechanisms other than reduction gears that can adjust the rotational speed of a motor by a combination of gears with different numbers of teeth (one example being a transmission). In other words, reduction gears and transmissions are examples of mechanisms that can adjust the rotational speed of a motor by a combination of gears with different numbers of teeth.

[0076] Furthermore, in the embodiment, an RNN was shown as the short-term prediction model of the first prediction unit 33 (see Figure 9). However, other models can be used as the short-term prediction model, for example, a multiple regression model or an LSTM can be used. Note that a neural network with more parameters (more complex) will have higher accuracy and is therefore preferable. An image of using a multiple regression model or an LSTM is shown in Figure 13. Figure 13 is a diagram to explain the variations of the short-term prediction model of the first prediction unit 33.

[0077] The first prediction unit 33 calculates a predicted value of vibration acceleration using, for example, features extracted from time-series data of vibration acceleration. The first prediction unit 33 detects an anomaly in the speed reducer 50 based on the difference between the predicted value and the measured value. Here, we assume that anomaly detection is performed using the cepstrum of vibration acceleration. The cepstrum is obtained by taking the logarithm of the power spectrum obtained by Fourier transforming the waveform, and then performing an inverse Fourier transform on that value. Figure 14 is an example of the cepstrum of vibration acceleration generated in the speed reducer 50. The horizontal axis is cefrenci (the independent variable in the cepstrum graph), and the vertical axis is the cepstrum. Code s1 is the graph of the 1st gear, code s2 is the graph of the 2nd gear, and code s3 is the graph of the 3rd gear. Figure 14 shows the cepstrum from "0 to 10 [ms]".

[0078] Figure 15 is a schematic diagram of an autoencoder used for feature extraction. The autoencoder shown in Figure 15 is a type of neural network. An autoencoder is an algorithm that compresses input data, retains only the important features, and then restores it to its original dimensions. As shown in Figure 15, the autoencoder takes data from the left and outputs it to the right. The circular parts in Figure 15 are called nodes, and the arrows are called edges. The values ​​input to a node are weighted by unique weights at each edge, and then input to the nodes of the next layer. Each node receives input from multiple edges, and the sum of these inputs becomes the final input value of the node.

[0079] Figure 16 shows an example of an autoencoder design. As shown in Figure 16, for example, the encoder layer and decoder layer each consist of two layers, with approximately 220 to 200 nodes. The number of input data points is approximately 250. Then, cepstrum features are extracted from the input data, and the extracted features are used to calculate the predicted vibration acceleration using an LSTM.

[0080] Figure 17 compares the cepstrum before and after inputting data into an autoencoder. In Figure 17, the horizontal axis represents time, and the vertical axis represents the cepstrum. In Figure 17, the input value is shown as a solid line, and the output as a dotted line. However, since the input and output values ​​of the autoencoder overlap (the input and output cepstrums are the same), only the input value (solid line) is visible. As shown in Figure 17, the fact that the input and output cepstrums are the same indicates that the autoencoder is functioning correctly.

[0081] Figure 18 is a graph showing the difference (loss) between the input and output of an autoencoder. Figure 18 compares the loss during training with the loss during validation. In Figure 18, the horizontal axis represents the number of epochs (how many times the training data was repeated for learning), and the vertical axis represents the loss. As shown in Figure 18, repeated training ultimately results in the same loss (the same output value).

[0082] Figure 19 is a schematic diagram of an LSTM. An LSTM is a neural network used for analyzing time series data. Because an LSTM can be intentionally trained to remember past events rather than the immediate present, it can learn long-term dependencies. Figure 20 shows an example of an LSTM design. For example, the LSTM layer is composed of 8 layers and inputs 10 past data points.

[0083] Figure 21 shows the prediction results of vibration acceleration using LSTM with cepstrum features of vibration acceleration. In Figure 21, the horizontal axis is date and time, and the vertical axis is vibration acceleration. The solid line with a circle represents the actual vibration acceleration result (measured value), the dotted line with a triangle represents the prediction result by LSTM on the training data, and the solid line with a square represents the prediction result by LSTM on the test data. Data from "2021 / 4 / 1 to 2022 / 8 / 23" was used as training data, and data from "2022 / 9 / 1 to 2023 / 2 / 17" was used as test data.

[0084] The first prediction unit 33 calculates the degree of anomaly from the difference between the predicted value and the measured value. An example of the formula for calculating the degree of anomaly is shown below.

number

[0085] Figure 22 shows the results of detecting the degree of abnormality using equations (1) and (2). Figure 22 relates to the meshing vibration of the 1st gear at the first frequency. In Figure 22, the horizontal axis represents the date and time, and the vertical axis represents the degree of abnormality. No increase in the degree of abnormality was observed until "February 2023," indicating that there were no signs of abnormality.

[0086] Next, we will explain variations in calculating the anomaly score. It is also possible to calculate the anomaly score by looking at internal parameters rather than looking at the prediction results. Figure 23 shows an example of a method for calculating the anomaly score, illustrating a case where the anomaly score is calculated based on the similarity of regression coefficients. Here, we assume the autoregressive (AR) model shown in equation (3) below. Exogenous variables (crane operating status) are not used.

number

[0087] In this case, the cosine similarity can be calculated using equation (4) shown below. It is thought that the similarity decreases as the oscillation pattern changes. Note that a higher order results in higher sensitivity.

number

[0088] Alternatively, the degree of anomaly can be calculated based on the accuracy of vibration acceleration prediction using the same AR model. Let's consider the case where vibration acceleration is predicted using "n-step ahead forecast". In this case, the evaluation index can be WAPE. For example, the monthly WAPE distribution can be obtained using the following equation (5) to calculate a statistic. Anomalies can be predicted by extrapolating and comparing these statistics.

number

[0089] Furthermore, it is possible to predict extreme values ​​of vibration acceleration using the same AR model and predict anomalies based on changes in the residual distribution (anomaly prediction using predicted residual distribution). In this case, the target of prediction may be a standardized value of the extreme values ​​of vibration acceleration.

[0090] <An example of using multiple predictive models> Next, we will explain an example of predicting the lifespan of the speed reducer 50 installed in the crane 10 using multiple prediction models. Here, we will use four prediction models: the VAE model, the LSTM model, the VAR model, and the AR model.

[0091] Referring to Figure 24, the configuration of the lifetime prediction system 101 when using four prediction models—VAE model, LSTM model, VAR model, and AR model—will be explained. Figure 24 is a block diagram relating to lifetime prediction of the lifetime prediction system 101 according to a modified example. As shown in Figure 24, the lifetime prediction system 101 comprises an agent PC 20 (monitoring device) and a cloud server 130 (lifetime prediction device). The configuration of the cloud server 130 in the lifetime prediction system 101 differs from that of the cloud server 30 (see Figure 5) according to the embodiment.

[0092] The cloud server 130 mainly comprises a data reception unit 31, a vibration data storage unit 32, a first prediction unit 133, a data display unit 34, a state data storage unit 35, and a second prediction unit 36. The first prediction unit 133 performs waveform data analysis processing compared to the first prediction unit 33 according to this embodiment (see Figure 5).

[0093] The first prediction unit 133 extracts features, for example, the results of frequency analysis and the autoregressive parameters of the autoregressive model, to be used as explanatory variables in the short-term prediction model. Furthermore, the first prediction unit 133 predicts anomalies based on a statistical model (for example, principal component analysis) using coefficients from an autoregressive model that predicts the vibration acceleration of multiple meshing vibration frequency bands of the reduction gear 50 or their harmonic frequency bands.

[0094] Figure 25 provides a more detailed description of the processing of the first prediction unit 133. The first prediction unit 133 executes the processes shown in steps A1 to A18 of Figure 25.

[0095] (Step A1) The first prediction unit 133 intermittently receives vibration data from the vibration sensor 51 and transfers the read vibration data to the relevant program (function). The vibration data is, for example, divided into 1.8 seconds as one data point.

[0096] (Step A2) When vibration data is transferred, the first prediction unit 133 inputs the waveform data into the AR model and calculates regression coefficients for each time step. For example, if the sampling rate is "25kHz", then "25,000 × 1.8" time points constitute one data point. For example, a 10-dimensional regression coefficient can be obtained from a model that predicts the next point from the past 10 points of a time-series vibration waveform.

[0097] (Step A3) The first prediction unit 133 calculates the residual for the single-point prediction obtained in step A2, and if the value is greater than or equal to a threshold, it determines that an anomaly occurred in the middle of the waveform and outputs a high-speed anomaly detection alert. To prevent false detections, the alert may be output if the threshold is exceeded for n consecutive points.

[0098] (Step A4) Meanwhile, the first prediction unit 133 performs frequency analysis on each data point of the vibration data using continuous wavelet transform (CWT). From the time-series changes of the wavelet coefficients in a specific frequency band, waveform data with points can be extracted for each data point. A specific frequency band is, for example, the "750Hz" band.

[0099] (Step A5) The first prediction unit 133 extracts waveform features from the waveform data obtained in step A4, for each of the three axes. The three axes are the axial, radial, and circumferential axes of the gears of the reducer 50. The waveform features are, for example, peak value, kurtosis, skewness, and RMS.

[0100] (Step A6) Furthermore, the first prediction unit 133 extracts sideband features for each of the three axes from the waveform data obtained in step A4. Sideband features include, for example, the sideband energy ratio, sideband index, and sideband level factor.

[0101] (Step A7) Meanwhile, the first prediction unit 133 performs frequency analysis on the vibration data for each time step using a short-time FFT. The frequency analysis is performed for each of the three axes, and waveform data of a specific frequency is acquired for each axis. The specific frequency may be the same for all three axes, or it may be a different frequency for each axis.

[0102] (Step A8) The first prediction unit 133 inputs the waveform data of the three axes at specific frequencies for each time step obtained in step A7 into a VAR model with learned parameters to obtain the next prediction point. For example, if the parameters have 10-dimensional values ​​for each axis, the unit predicts the next point for each axis from the past 10 points.

[0103] (Step A9) Furthermore, the first prediction unit 133 obtains the average of one data point for each time unit for the frequency spectra of the three axes obtained in step A7. An FFT is performed on the averaged frequency spectra of the three axes to extract the cepstrum waveform. Only cepstrum waveforms with a cefrenci of less than or equal to a certain value (excluding "0ms") are selected. This certain value is, for example, "10ms".

[0104] (Step A10) The first prediction unit 133 calculates the average for one data point for the 10-dimensional regression coefficients of the three axes obtained in step A2.

[0105] (Step A11) The first prediction unit 133 acquires data for a certain period by repeatedly acquiring data. This certain period is, for example, "10 days".

[0106] (Step A12) The first prediction unit 133 obtains the average value of the waveform features of the three axes obtained in step A5, the sideband features of the three axes obtained in step A6, and the 10-dimensional regression coefficients of the three axes obtained in step A10, at regular intervals. This regular interval is, for example, "1 day".

[0107] (Step A13) The first prediction unit 133 inputs time series data with a series length equal to the period obtained in step A11 into the LSTM, based on the average value for a certain period obtained in step A12, and predicts the average value for the next period. The residual between the obtained predicted value and the actual value is set to an anomaly score of "1".

[0108] (Step A14) The first prediction unit 133 obtains the average value of each of the 10-dimensional parameters of the three axes obtained in step A8 at regular intervals. This regular interval is, for example, "1 day".

[0109] (Step A15) The first prediction unit 133 calculates the principal component quantities using PCA (Principal Component Analysis) on the average value over a certain period obtained in step A14. The ratio of the sum of the principal component quantities from the second principal component onward to the first principal component is defined as an anomaly score of "2".

[0110] (Steps A16, 17) The first prediction unit 133 takes a 3D cepstrum as input to the VAE model trained on the three-axis (3D) cepstrum obtained in step A9 and obtains the residuals from the model's output. The residuals are obtained over a certain period, and the average value is set as the anomaly score "3". This period is, for example, "1 day".

[0111] (Step A18) The first prediction unit 133 calculates the overall anomaly score using a certain formula based on the anomaly score "1" obtained in step A13, the anomaly score "2" obtained in step A15, and the anomaly score "3" obtained in step A17. This formula is, for example, "α1 × anomaly score "1" + α2 × anomaly score "2" + α3 × anomaly score "3", where α1 to α3 can be arbitrarily determined. In other words, the first prediction unit 133 predicts anomalies from the weighted sum of anomaly scores output by multiple short-term prediction models. An alert is issued if the short-term anomaly score exceeds a pre-set threshold.

[0112] (Step A19) The first prediction unit 133 uses a model trained on the initially obtained training data to approximate the anomaly scores obtained daily with a regression model, similar to step A18. This regression model may be a simple linear regression model, a multiple regression model, or a generalized linear model. The resulting approximate model is used to predict long-term anomalies. The second prediction unit 36 ​​may perform predictions using anomaly scores calculated with different prediction parameters than those used by the first prediction unit 133.

[0113] Referring to Figure 26, we will explain how to train a predictive model during an oil change. Figure 26 is an example of the training flow during an oil change. After a statutory crane inspection (step B1) and an oil change (step B2), the crane is operated for a certain period (step B3). During this period of operation (step B3), vibration data at a specified rotational speed is acquired (step B4), and short-term model parameters are automatically trained (step B5). The short-term model is a model for calculating short-term anomaly severity, and is, for example, a VAE model, an LSTM model, or a VAR model.

[0114] In this way, the prediction model used in the first prediction unit 133 will use prediction parameters learned based on vibration acceleration data obtained a certain period after an oil change in the reduction gear 50. When the life prediction system 101 is started up (when the life prediction system 101 is introduced), the learning of the model for calculating the long-term anomaly rate is started (step B6).

[0115] Refer to Figure 27 to see an example of how the anomaly score is calculated for each model. Figure 27 shows an example of how the anomaly score is calculated for each model. In a VAE model, for example, the reconstruction error is calculated from the cepstrum, and the anomaly score is determined based on this reconstruction error. Figure 28 shows an example of how the anomaly score is calculated using a VAE model. In an LSTM model, for example, a weighted absolute error rate is calculated, and the anomaly score is determined based on that weighted absolute error rate.

[0116] In a VAR model, for example, the mean absolute scale error is calculated, and the degree of anomaly is determined based on that mean absolute scale error. In AR models, for example, the degree of anomaly is determined by performing principal component analysis on the autoregression coefficients and calculating their ratio to the first principal component.

[0117] Refer to Figure 29 to explain an example of calculating the anomaly score using PCA (Principal Component Analysis). Figure 29 is a diagram showing an example of calculating the anomaly score using PCA (Principal Component Analysis). Figure 29 shows a total of nine graphs, three in the vertical direction and three in the horizontal direction. The graphs in the left column correspond to the vibration sensor 51 that detects horizontal acceleration. The graphs in the middle column correspond to the vibration sensor 51 that detects axial acceleration. The graphs in the right column correspond to the vibration sensor 51 that detects vertical acceleration.

[0118] The graph in the upper part of Figure 29 shows the results of calculating the anomaly score using VAR. The horizontal axis represents the date, and the vertical axis represents the anomaly score. Here, VAR is used to calculate the 10-dimensional parameters of an autoregression model.

[0119] The graph in the middle of Figure 29 shows the results of principal component analysis of the parameters of an autoregressive model. The horizontal axis represents the first principal component, and the vertical axis represents the second principal component. Here, the parameters have been reduced to two dimensions through dimensionality reduction. Since the second principal component spreads well relative to the first principal component, it is thought that when an anomaly occurs, changes are more likely to be observed in the principal components from the second principal component onward.

[0120] The graph in the lower section of Figure 29 shows the results of determining the anomaly score from the principal component analysis. The horizontal axis represents the date, and the vertical axis represents the anomaly score. Here, the anomaly score is calculated from the ratio of the component amount of the first principal component to the component amounts of the second principal component and subsequent components.

[0121] In the case of the VAR calculation of the anomaly level in the upper panel, the anomaly level is high in April 2021. This is likely because the crane's oil was changed. On the other hand, the anomaly level in the lower graph does not increase even after the oil change, indicating that the judgment was correct.

[0122] Refer to Figure 30 to explain the calculation results of the long-term anomaly. Figure 30 is an example of the calculation results of the long-term anomaly. In Figure 30, the horizontal axis represents the date and time, and the vertical axis represents the daily average value of the anomaly.

[0123] The daily average value of the abnormality shown in Figure 30 is calculated using the following formula (6). Here, α1, α2, and α3 are coefficients.

number

[0124] Figure 30 plots the daily average value of the anomaly score and shows an approximation curve illustrating the relationship between the daily average value of the anomaly score and the passage of time. The approximation curve shows three limits: "upper limit," "average," and "lower limit." Based on the relationship between the approximation curve and the thresholds, normal / abnormal status is determined (classified) for the long-term second period. The thresholds can be determined from past crane performance values. Here, two thresholds are shown: one used for warnings and another for indicating danger. The approximation curves for the "upper limit" and "lower limit" indicate the range of prediction. [Explanation of Symbols]

[0125] 1,101 Life Prediction System 10 Cranes 11 wheels 12 Frame section 12a Garda 12b Boom 13 Trolley 13a Spreader 13b Driver's cab 14 Machine room 14a Hoisting device 14b Traverse device 14c Lifting device 19 Wire rope 20 Agent PCs (monitoring devices) 21 Monitoring Department 22 Transmitter 30,130 Cloud Server (Life Prediction Device) 31 Data Reception Department 32 Vibration data storage unit 33,133 First Prediction Section 34 Data display section 35 Status data storage unit 36 Second Prediction Section 50 reducer 51 Vibration Sensor 52 Measurement Modules 60 PLC 70 HUB

Claims

1. A life prediction system for predicting the lifespan of a mechanism that can adjust the rotational speed of a motor by combining gears with different numbers of teeth, A vibration sensor that acquires vibration data of the aforementioned mechanism, A monitoring device that monitors the rotational speed of the motor and collects the vibration data in a steady state, The device includes a life prediction device that receives vibration data from the monitoring device and predicts the lifespan of the mechanism, The life prediction device is The mechanism includes a vibration data storage unit that stores the vibration data at each stepwise speed that can be changed, The system includes a first prediction unit that uses the vibration data corresponding to the stepwise speed to predict physical quantities related to vibration within a short first period and predicts abnormalities in the mechanism that occur within the first period. A lifespan prediction system characterized by the following features.

2. The life prediction device is The vibration sensor is mounted around the bearing where the rotational speed of the reduction gear shaft is high, and the acceleration in the axial, radial, and circumferential directions is measured. The life prediction system according to feature 1.

3. The life prediction device is A state data storage unit continuously compares the predicted physical quantities related to vibration with the physically measured physical quantities related to vibration over time, and accumulates the changes in the comparison results over time. The system includes a second prediction unit that predicts the deterioration of the mechanism over time in a second period longer than the first period, based on the changes in the comparison results over time. The life prediction system according to feature 1.

4. The first prediction unit performs frequency analysis on the vibration data and uses the frequencies of the meshing vibrations of the multiple gears in the mechanism, as well as information on the frequency bands where there is no superposition and information on the frequency bands where there is superposition, to predict physical quantities related to vibration within the first period. The life prediction system according to feature 1.

5. The first prediction unit has a short-term prediction model that predicts physical quantities related to vibration within the first period, and extracts feature quantities that use the results of frequency analysis and the autoregressive parameters of the autoregressive model as explanatory variables for the short-term prediction model. The life prediction system according to feature 1.

6. The monitoring device monitors the motor rotation speed, determines when a trigger occurs when the motor transitions from a constant speed state to a speed change state, and collects the vibration data immediately before the trigger occurs. The life prediction system according to feature 1.

7. The second prediction unit predicts the deterioration of the mechanism over time based on the relationship between the prediction error of the first prediction unit and a predetermined threshold value for the prediction error. The life prediction system according to claim 3.

8. The second prediction unit uses anomaly scores calculated with different prediction parameters than those of the first prediction unit. The life prediction system according to claim 3.

9. The first prediction unit calculates an anomaly degree from the predicted physical quantity related to vibration and the physical quantity related to vibration that was actually measured, and predicts an anomaly in the mechanism based on the anomaly degree. The life prediction system according to feature 1.

10. The first prediction unit predicts anomalies based on a statistical prediction model using coefficients of an autoregressive model that predicts vibration acceleration in multiple meshing vibration frequency bands or their harmonic frequency bands of the gearbox. The life prediction system according to feature 1.

11. The first prediction unit has a plurality of short-term prediction models that predict physical quantities related to vibrations within the first period, Each of the aforementioned short-term prediction models predicts anomalies in the mechanism that occur within different first periods. The life prediction system according to feature 1.

12. The first prediction unit predicts anomalies from a weighted sum of anomaly scores output by multiple short-term prediction models. The life prediction system according to feature 11.

13. The prediction model used in the first prediction unit uses prediction parameters learned based on vibration acceleration data taken a certain period after the reduction gear oil change. The life prediction system according to feature 1.

14. The state data storage unit creates a histogram relating to the prediction error of the first prediction unit, The second prediction unit obtains a representative value of the prediction error based on the histogram, and predicts the deterioration of the mechanism over time based on the relationship between the approximation curve, which represents the relationship between the time progression and the representative value, and the threshold value. The life prediction system according to feature 7.

15. The aforementioned mechanism is provided by the crane and is capable of gradually changing either the traversing speed or the hoisting speed, or both. The life prediction system according to feature 1.

16. A life prediction method for a mechanism that can adjust the rotational speed of a motor by combining gears with different numbers of teeth, A monitoring step involves monitoring the motor's rotational speed and collecting vibration data of the mechanism in a steady state. The system includes a life prediction step that receives the vibration data and predicts the lifespan of the mechanism, In the life prediction process, A vibration data storage step is performed to store the vibration data at each stepwise speed that the mechanism can change, The system includes a first prediction step, which uses the vibration data corresponding to the stepwise speed to predict physical quantities related to vibration within a short first period and predicts abnormalities in the mechanism that occur within the first period. A method for predicting lifespan, characterized by the following features.

Images