ANOMALIA DETERMINATION DEVICE

Abstract

An anomaly detection device used in a vehicle equipped with a transmission (26) configured to transmit power by rotating a shaft, the anomaly detection device comprising: a memory (46) configured to store map data, which are data defining a map learned by machine learning, wherein the map is a map in which an input variable is a variable indicating time series data of a rotational speed of the shaft, and an output variable is a state variable indicating a state of the shaft, wherein the state variable includes four state variables, the four state variables being a variable indicating that the shaft is in a normal state, a variable indicating that the shaft is in a divergent vibration state, and a variable indicating that the shaft is in a torsional vibration state.and a variable indicating that the shaft is in a bending vibration state; and a processor configured to perform a capture process to acquire the variable indicating the time series data and to use the variable as a value of the input variable, and to perform a determination process to determine, based on values ​​of the four state variables as output variables obtained using the value of the input variable and the characteristic map, whether an anomaly has occurred in the gearbox (26).

Description

Technical field

[0001] The present invention relates to an anomaly detection device that determines whether an anomaly has occurred in a gearbox that transmits power by rotating a shaft. State of the art

[0002] The published Japanese patent application JP 2009 - 063 455 A discloses a detection device that detects an anomaly based on whether the detected value is within the normal range or the abnormal range. Summary of the invention

[0003] An anomaly detection device requires not only the ability to detect whether an anomaly has occurred, but also the ability to detect the phenomenon causing it. For example, if the object under investigation vibrates due to an anomaly, its state is assumed to vary depending on the vibration mode. That is, if the vibration mode can be determined, the state of the object can be ascertained, thus helping to capture the actual state of the phenomenon. With the detection device disclosed in JP 2009-063455A, it is possible to detect whether an anomaly has occurred, but it cannot determine the state of the object under investigation.

[0004] An anomaly detection device according to a first aspect of the present invention is applied to a vehicle equipped with a transmission configured to transmit power by rotating a shaft. The anomaly detection device includes a processor and a memory. The memory is configured to store map data (map data), which is data that defines a map (map) learned by machine learning. The map is a map in which an input variable is a variable that represents time-series data of a rotational speed of the shaft, and an output variable is a state variable that represents a state of the shaft.The processor is set up to perform a capture process of acquiring the variable that displays the time series data and using the variable as a value of the input variable, and a determination process of determining whether an anomaly has occurred in the gearbox, based on a value of the output variable acquired using the value of the input variable and the characteristic map.

[0005] An anomaly can occur in the gearbox, and the power-transmitting shaft can vibrate. If the shaft is in a vibration state, it can be concluded that the time series data of the shaft's rotational speed will exhibit different characteristics than the time series data when the shaft is not in a vibration state. Furthermore, it can be concluded that the characteristics shown by the time series data differ depending on the vibration mode of the shaft.

[0006] According to the configuration described above, when a variable displaying the time-series data of the shaft rotational speed is entered into the characteristic map, a state variable indicating the shaft's condition is output. This allows the shaft's condition to be determined based on the variable displaying the time-series data of the shaft rotational speed. In other words, based on the time-series data of the shaft rotational speed, it is possible to determine the shaft's condition by extracting features that cannot be derived solely from the magnitude of the shaft rotational speed, and it can be determined whether an anomaly has occurred in the gearbox.

[0007] In the foregoing aspect, the processor can be configured to perform a feature size calculation process to calculate a feature size obtained by processing the time series data, wherein the acquisition process can be a process for acquiring the feature size and using the feature size as the value of the input variable, and wherein the feature size calculation process can be a process for dividing values ​​of rotational speed that constitute the time series data into classes according to a size of rotational speed and for calculating an incidence for each of the classes as a feature size.

[0008] In the configuration described above, the rotational speed time series data is processed into a feature metric that exhibits an incidence distribution. This makes it easy to distinguish the feature indicated by the rotational speed time series data when an anomaly has occurred in the shaft from the feature indicated by the rotational speed time series data when no anomaly has occurred. In other words, the accuracy of determining the shaft's condition can be improved.

[0009] In the above aspect, the feature size calculation process can include a process of normalizing the time series data, so that a maximum value of the rotational speed becomes “1” and a minimum value of the rotational speed becomes “0”.

[0010] According to the above configuration, by normalizing the time series data, the fluctuation of the rotational speed can be captured as a feature indicated by the time series data in order to determine the condition of the shaft without being affected by the magnitude of the absolute value of the rotational speed.

[0011] In the above aspect, the feature size can be a time feature size, the feature size calculation process can be an incidence analysis process to calculate the time feature size, the processor can be set up to execute a frequency analysis process to calculate a distribution of frequency components obtained by fast Fourier transform of the time series data as the frequency feature size, and in the acquisition process, the time feature size and the frequency feature size obtained by processing the time series data can be acquired and used as the value of the input variable.

[0012] In the configuration described above, the time-based characteristic parameter, in which the time series data is processed into a parameter representing an incidence distribution, and the frequency-based characteristic parameter, in which the time series data is converted from the time domain to the frequency domain, are calculated and entered into the characteristic map. By entering a combination of differently processed data into the characteristic map, it becomes easier to distinguish the characteristics indicated by the time series data, and therefore the accuracy of the determination is expected to be further improved.

[0013] In the foregoing aspect, the processor may be configured to execute a frequency analysis process to calculate a frequency feature quantity obtained by processing the time series data, wherein the acquisition process may be a process for acquiring the frequency feature quantity and using the frequency feature quantity as the value of the input variable, and the frequency analysis process may be a process for calculating a distribution of frequency components obtained by fast Fourier transform of the time series data as the frequency feature quantity.

[0014] In the configuration described above, the rotational speed time series data is processed into a feature metric converted from the time domain to the frequency domain. Determining the shaft's condition based on the features of the time series data appearing in the frequency domain makes it easy to distinguish the feature indicated by the rotational speed time series data when an anomaly has occurred in the shaft from the feature indicated by the rotational speed time series data when no anomaly has occurred. This means that the accuracy of determining the shaft's condition can be improved.

[0015] According to the above aspect, the frequency analysis process may include a process for calculating a primary frequency based on an average of the rotational speed in the time series data and normalizing the frequency components with the primary frequency as a reference.

[0016] According to the above configuration, by normalizing the frequency components, the state of the shaft can be determined based on the characteristics indicated by the time series data of the rotational speed, without being affected by the intensity of the frequency components.

[0017] In the above aspect, the frequency analysis process can include a process of normalizing the frequency components, so that the intensity of the frequency components at a primary resonant frequency, if the wave mode is a torsional oscillation, becomes "1".

[0018] When the wave's oscillation is subdivided, multiple mode shapes emerge. In the data, when the mode shape is torsional and resonance is generated, the intensity of the frequency components at the resonance frequency is relatively higher than at other frequencies. Therefore, when the data is inputted in the state where torsional oscillation is generated and normalized as described above, the high intensities of the frequency components at the primary, secondary, ... nth order resonance frequencies are highlighted. Conversely, in the data where no torsional oscillation is generated, the difference between the intensity at the resonance frequency and the intensity at other frequencies can be assumed to be small.Therefore, if the data are normalized as described above in the state where torsional vibration is not generated, it can be assumed that the distribution has a high incidence across the entire frequency range.

[0019] This means that, according to the above configuration, the difference in the characteristics that can be extracted from the time-series data of the rotational speeds can be clearly demonstrated, depending on whether the shaft's vibration mode is torsional or not. This facilitates the differentiation of the shaft's condition.

[0020] In the above aspect, the frequency analysis process can include a process of normalizing the frequency components, so that the intensity of the frequency components at a primary resonant frequency, if the wave mode is a bending vibration, becomes “1”.

[0021] When the wave's oscillation is subdivided, several mode shapes emerge. In the data, when the mode shape is bending and resonance is generated, the intensity of the frequency components at the resonance frequency is relatively higher than the intensity at other frequencies. Therefore, when the data is inputted in the state where bending vibration is generated and normalized as described above, the high intensities of the frequency components at the primary, secondary, ... nth order resonance frequencies are highlighted. Conversely, in the data for the state where no bending vibration is generated, the difference between the intensity at the resonance frequency (when bending) and the intensity at other frequencies can be assumed to be small.Therefore, if the data are normalized as described above in the state where bending vibration is not generated, the distribution is assumed to have a high incidence across the entire frequency range.

[0022] This means that, according to the above configuration, the difference in the characteristics that can be extracted from the time-series rotational speed data, depending on whether the shaft's vibration mode is a bending vibration or not, can be clearly demonstrated. This makes it easier to distinguish the shaft's condition.

[0023] In the above aspect, time series data can be calculated based on a signal from a speed sensor that detects the shaft's rotational speed. In the above aspect, time series data can also be calculated based on the vehicle's speed.

[0024] Based on the above, the input variable can include a variable indicating the temperature of the hydraulic oil in the transmission. According to the configuration described above, in addition to the variable displaying the time series data, the variable indicating the hydraulic oil temperature is entered into the characteristic map as an input variable. By inputting different types of variables, it is possible to improve the accuracy of determining whether an anomaly has occurred in the transmission.

[0025] Based on the above, the input variable can include a variable that specifies the magnitude of the torque transmitted by the shaft. According to the above configuration, it is possible to improve the accuracy of determining whether an anomaly has occurred in the gearbox.

[0026] In the above aspect, the input variable can include a variable that indicates a dimension of the shaft based on the shaft specifications. According to the above configuration, it is possible to improve the accuracy of determining whether an anomaly has occurred in the gearbox.

[0027] In the above aspect, the input variable can include a variable indicating the amount of imbalance in the shaft. According to the above configuration, it is possible to improve the accuracy of determining whether an anomaly has occurred in the gearbox.

[0028] In the above aspect, the input variable can include a variable indicating a fit tolerance between the shaft and a bearing on the shaft. According to the above configuration, it is possible to improve the accuracy of determining whether an anomaly has occurred in the gearbox.

[0029] Based on the above, the input variable can include a variable indicating coaxiality between the shaft and a bearing on the shaft. According to the above configuration, it is possible to improve the accuracy of determining whether an anomaly has occurred in the gearbox.

[0030] Based on the above, the input variable can include a variable representing a value detected by a vibration sensor configured to detect vibrations. This configuration can improve the accuracy of determining whether an anomaly has occurred in the transmission.

[0031] Based on the above, the input variable can include a variable representing a value detected by a noise sensor configured to detect sound. According to the above configuration, it is possible to improve the accuracy of determining whether an anomaly has occurred in the transmission.

[0032] Based on the above, the input variable can include a variable indicating the hydraulic pressure in a vehicle's braking system. According to the above configuration, it is possible to improve the accuracy of determining whether an anomaly has occurred in the transmission.

[0033] In the above aspect, the input variable can include a variable indicating the vehicle's mileage. According to the above configuration, it is possible to improve the accuracy of determining whether an anomaly has occurred in the transmission.

[0034] In the above aspect, the state variable can include a variable indicating that the shaft is in a divergent vibration state. When the shaft is in a divergent vibration state, features characteristic of divergent vibration appear in the time series data. According to the above configuration, a state variable indicating that the shaft is in a divergent vibration state can be output if a variable displaying time series data with features characteristic of divergent vibration is entered into the characteristic map. That is, the divergent vibration of the shaft can be detected based on the features displayed by the time series data.

[0035] In the above aspect, the state variable can include a variable indicating that the shaft is in a torsional vibration state. When the shaft is in a torsional vibration state, features characteristic of torsional vibration appear in the time series data. According to the above configuration, a state variable indicating that the shaft is in a torsional vibration state can be output if a variable displaying time series data with features characteristic of torsional vibration is entered into the characteristic map. That is, the shaft's torsional vibration can be detected based on the features displayed by the time series data. This allows the vibration mode to be classified and the shaft's state to be determined.

[0036] In the above aspect, the state variable can include a variable indicating that the shaft is in a bending vibration state. When the shaft is in a bending vibration state, features characteristic of bending vibration appear in the time series data. According to the above configuration, a state variable indicating that the shaft is in a bending vibration state can be output when a variable displaying time series data with features characteristic of bending vibration is entered into the characteristic map. That is, the bending vibration of the shaft can be detected based on the features displayed by the time series data. This allows the vibration mode to be classified and the condition of the shaft to be determined. List of characters

[0037] Features, advantages, and technical and industrial significance of exemplary embodiments of the invention are described below with reference to the accompanying drawings, in which the same symbols denote the same elements, and wherein: Fig. 1 is a diagram showing a configuration of a vehicle and a control unit according to an embodiment; Fig. 2 is a flowchart showing the sequence of a process carried out by the control device according to the embodiment; Sections (a) and (b) in Fig. Three diagrams show the time series data of the rotational speed of a shaft according to the embodiment; Sections (a) and (b) in Fig. Four diagrams are shown, which depict characteristic values ​​obtained by processing the time series data of the rotational speed of the shaft according to the embodiment; Fig. 5 is a diagram that defines output variables according to the embodiment; Fig. 6 is a diagram showing time series data of the rotational speed of the shaft in a torsional vibration state; Fig. 7 is a diagram showing feature sizes obtained by processing the time series data of the rotational speed of the shaft in the torsional vibration state by a process executed by a control device of a modified example; and Fig. Figure 8 is a diagram showing a system configuration according to another modified example. Detailed description of the embodiments

[0038] An embodiment of an anomaly detection device is described below with reference to the drawings. As shown in Fig. As shown in Figure 1, a vehicle VC comprises an internal combustion engine 10, a first motor-generator 22, and a second motor-generator 24. A power-sharing device 20 is mechanically connected to a crankshaft 12 of the internal combustion engine 10. The power-sharing device 20 divides the power of the internal combustion engine 10, the first motor-generator 22, and the second motor-generator 24. The power-sharing device 20 includes a planetary gear set. A carrier C of the planetary gear set is mechanically connected to the crankshaft 12. A sun gear S of the planetary gear set is mechanically connected to a rotational shaft 22a of the first motor-generator 22. A ring gear R of the planetary gear set is mechanically connected to a rotational shaft 24a of the second motor-generator 24. The output voltage of a first inverter 23 is applied to the terminal of the first motor-generator 22.The output voltage of a second inverter 25 is applied to the terminal of the second motor generator 24.

[0039] In addition to the rotary shaft 24a of the second motor-generator 24, drive wheels 30 are mechanically connected via a gearbox 26 to the ring gear R of the power distribution device 20. The gearbox 26 includes a shaft, which is a rotary shaft that transmits power. The shaft is rotatably mounted by bearings.

[0040] A driven shaft 32a of an oil pump 32 is mechanically connected to the support C of the power distribution device 20. The oil pump 32 circulates oil in an oil pan 34 as lubricating oil to the power distribution device 20 and delivers the oil as hydraulic oil to the gearbox 26. The pressure of the hydraulic oil delivered by the oil pump 32 is adjusted by a hydraulic control circuit 26a in the gearbox 26 for use as hydraulic oil.

[0041] A control unit 40 controls the internal combustion engine 10. The control unit 40 controls various operating units of the internal combustion engine 10 to control the torque, exhaust gas component ratio, and the like, which are control variables of the internal combustion engine 10. Furthermore, the control unit 40 controls the first motor-generator 22. The control unit 40 controls the first inverter 23 to control torque, speed, and the like, which are control variables of the first motor-generator 22. Furthermore, the control unit 40 controls the second motor-generator 24. The control unit 40 controls the second inverter 25 to regulate torque, speed, and the like, which are control variables of the second motor-generator 24. Fig. 1 are the signals transmitted by the control unit 40 for the operation of the internal combustion engine 10 and the transmission 26, designated as operating signals MS.

[0042] When the control unit 40 controls the aforementioned control variables, the control unit 40 references the output signal Scr of a crank angle sensor 50, the output signal Sm1 of a first rotation angle sensor 52, which detects the rotation angle of the rotating shaft 22a of the first motor generator 22, and the output signal Sm2 of a second rotation angle sensor 54, which detects the rotation angle of the rotating shaft 24a of the second motor generator 24. Furthermore, the control unit 40 references the oil temperature Toil, i.e., the temperature of the oil detected by an oil temperature sensor 56.

[0043] The control unit 40 comprises a central processing unit (CPU) 42, a read-only memory (ROM) 44, an electrically rewritable non-volatile memory 46, and a peripheral circuit 48. The CPU 42, the ROM 44, the memory 46 (which is an electrically rewritable non-volatile memory), and the peripheral circuit 48 can communicate with each other via a local network 49. The peripheral circuit 48 includes a circuit that generates a clock signal defining the internal operation, a power supply circuit, a reset circuit, and the like. The control unit 40 controls the control variables when the CPU 42 executes the program stored in the ROM 44.

[0044] Part of the processes executed by the control unit 40 is described. The control unit 40 performs a drive torque adjustment process. This process calculates a drive torque setpoint Trq*, which is the target value of the torque to be applied to the drive wheels 30. The drive torque setpoint Trq* is calculated by inputting the actuation value of an accelerator pedal actuator provided in the vehicle VC and increases with increasing actuation value.

[0045] The control unit 40 performs a drive force distribution process. This process sets the target torque Trqe* for the internal combustion engine 10, the target torque Trqm1* for the first motor-generator 22, and the target torque Trqm2* for the second motor-generator 24, based on the drive torque target Trq*. These target torque values ​​Trqe*, Trqm1*, and Trqm2* are generated by the internal combustion engine 10, the first motor-generator 22, and the second motor-generator 24, respectively, so that the torque applied to the drive wheels 30 is adjusted to the drive torque target Trq*.

[0046] The control unit 40 performs a speed calculation process. This process calculates the shaft speed Nshaft, which is the speed of the shaft mounted on the gearbox 26. The shaft speed Nshaft is calculated based on the output signal Sm1. The shaft speed Nshaft is calculated at each predetermined interval. The transmitted shaft speed Nshaft data is stored in the memory 46 as time series data.

[0047] The control unit 40 performs a process to detect an anomaly in the shaft of the gearbox 26. This process is described below. Fig. Figure 2 shows a sequence of the process executed by the control unit 40. The in Fig. The process shown in Figure 2 is implemented when the CPU 42 repeatedly executes the program stored in ROM 44, for example, in a predefined cycle. It should be noted that in the following, the step number of each process step is represented by a number preceded by an "S".

[0048] In the Fig. In the series of processes shown in Figure 2, the CPU 42 first acquires the time series data of the shaft speed Nshaft and the oil temperature Toil (S101). Next, the CPU 42 performs a normalization process to normalize the time series data of the shaft speed Nshaft in the specified period (S102). The CPU 42 further calculates the normalized feature NFv, which is obtained by processing the normalized time series data (S103). The normalized feature NFv is calculated as data representing the incidence distribution of the shaft speed Nshaft in the specified period.

[0049] The processes of S102 and S103 are described with reference to Fig. 3 and Fig. 4 described. Section (a) in Fig. Figure 3 shows a portion of the shaft speed Nshaft time series data when no abnormal vibration has occurred in the shaft, i.e., when the shaft is in a normal state. Section (b) in Fig. Figure 3 shows a portion of the shaft speed time series data when the shaft is in a divergent vibration state, i.e., when an anomaly has occurred in the shaft. Sections (a) and (b) in Fig. 3. It is assumed that the magnitude of the drive torque setpoint Trq* is the same in the period shown.

[0050] As in section (a) in Fig. As shown in Figure 3, if no anomaly has occurred in the shaft, the shaft speed Nshaft changes with a regular cycle and amplitude. On the other hand, as shown in section (b) in Fig. 3 shows that if an anomaly has occurred in the shaft, the shaft speed Nshaft is transferred to produce a different waveform than that shown in section (a) in Fig. 3 to display the waveform shown. In particular section (b) in Fig. Figure 3 shows an example where the shaft is in a divergent state of vibration, such that the amplitude of the shaft speed Nshaft increases over time. As described above, the value of the shaft speed Nshaft can become extremely large or extremely small if an abnormal vibration has occurred in the shaft. If an abnormal vibration has occurred in the shaft, the transmission of the shaft speed Nshaft differs depending on the mode shape of the shaft.

[0051] In the present embodiment, the time series data of the shaft speed Nshaft are provided by a plurality of samples that are continuous in the time series when the shaft speed Nshaft is sampled in a specific period with a constant sampling cycle. The circles in Fig. Figure 3 illustrates the sampled values, which are sampled with a constant sampling cycle. The period for sampling the shaft speed Nshaft is the specified period.

[0052] In process step S102, CPU 42 normalizes the time series data of the shaft speed Nshaft so that the maximum value of the shaft speed Nshaft in the specified period becomes "1" and the minimum value of the shaft speed Nshaft in the specified period becomes "0". The normalization can be performed, for example, using the following mathematical formula (Equation 1). N=(n−nmin) / (nmax−nmin)

[0053] In the preceding mathematical formula (equation 1), “N” is the shaft speed Nshaft after normalization, “n” is the shaft speed Nshaft to be normalized, “nmax” is the maximum value of the shaft speed Nshaft before normalization, and “nmin” is the minimum value of the shaft speed Nshaft before normalization.

[0054] In process step S103, the CPU 42 first divides the range from the minimum to the maximum value of the shaft speed Nshaft in the specified period into five equal ranges and classifies these ranges into five classes. In other words, the classification is performed such that the width of each class is 0.2. The CPU 42 then calculates the number of shaft speed samples Nshaft in each class, i.e., the incidence, as a normalized feature NFv. The five classes are subsequently designated as Bin1 (first class), Bin2 (second class), Bin3 (third class), Bin4 (fourth class), and Bin5 (fifth class), in order from smallest to largest.

[0055] section (a) in Fig. Figure 4 is an example where the process steps of S102 and S103 are applied to the shaft speed time series data when no anomaly has occurred in the shaft. That is, section (a) in Fig. Figure 4 shows an example of a normalized feature size NFv when no anomaly has occurred in the wave. As in section (a) in Fig. As shown in Figure 4, the incidence increases in the order of incidence in the first class Bin1, the incidence in the second class Bin2, and the incidence in the third class Bin3, provided no wave anomaly has occurred. Among the incidences in the five classes, the incidence in the third class Bin3, including the median wave speed Nshaft, is the highest, and the incidence in the fourth class Bin4 and the incidence in the fifth class Bin5 decrease in that order. Furthermore, the incidence in the second class Bin2 and the incidence in the fourth class Bin4 are equal, and the incidence in the first class Bin1 and the incidence in the fifth class Bin5 are equal. That is, section (a) in Fig. Figure 4 shows a symmetrical distribution centered on the third class Bin3.

[0056] Section (b) in Fig. Figure 4 is an example where the process steps of S102 and S103 are applied to the shaft speed time series data when an anomaly has occurred in the shaft. That is, section (b) in Fig. Figure 4 shows an example of a normalized feature size NFv when an anomaly has occurred in the wave. Similar to section (a) in Fig. 4. The incidence is highest in the third class Bin3, and the incidence is lowest in the first class Bin1 and the incidence in the fifth class Bin5. In contrast to section (a) in Fig. In sample 4, where no wave anomaly occurred, the incidence in the third class, Bin3, is much higher compared to the incidences in the other classes. This is because, due to the irregular fluctuations in the shaft speed Nshaft, a small number of values ​​far from the median are sampled. To further illustrate this, consider the incidence distribution of data where a small number of values ​​are sampled far from the median. The range from the minimum to the maximum value of the shaft speed Nshaft in the data is divided into the same number of classes as in section (a) in Fig. Example 4 shows how the time series data of the shaft speed Nshaft, which transmits regularly, is processed. Therefore, the class that includes the form contains more samples than in the one in section (a) in Fig. As shown in example 4. Therefore, if a small number of values ​​far from the median are sampled, the incidence of the class containing the shape will be higher, while the incidence of the other classes will be lower.

[0057] As in sections (a) and (b) in Fig. As shown in Figure 4, the normalized feature NFv, obtained by processing the time-series data of the shaft rotational speed Nshaft, exhibits different features between a case where no anomaly has occurred in the shaft and a case where the shaft is in a vibrational state. When the shaft is in a vibrational state, the features indicated by the normalized feature NFv differ depending on the vibration mode.

[0058] Back to Fig. 2: After calculating the normalized feature NFv, the CPU 42 replaces the normalized feature NFv calculated by the S103 process and the oil temperature Toil in the input variables x(1) to x(6) for the characteristic map, which are defined in the Fig. The memory shown contains 46 stored characteristic map data DM (S104). More precisely, the incidence of the first class Bin1 up to the incidence of the fifth class Bin5 is inserted into the input variables x(1) to x(5). The oil temperature Toil is inserted into the input variable x(6). The mean value of the oil temperature Toil in the specified period can be inserted into the input variable x(6).

[0059] Next, the CPU 42 inserts the values ​​of the input variables x(1) to x(6) into the characteristic map to calculate the values ​​of the output variables y(0) to y(4), which are variables that indicate the state of the shaft (S105).

[0060] In the present embodiment, a functional approximation is illustrated as the characteristic map, and more precisely, a fully connected forward propagation neural network with an intermediate layer is illustrated. Specifically, the values ​​of the nodes in the intermediate layer are determined by inserting each of the "m" values ​​of the input variables x(1) to x(6), into which the values ​​in the process of S105 are inserted, into the activation function f, and by using the input variable x(0) as a bias parameter, which is transformed by the linear characteristic map defined by the coefficient wFjk (j = 1 to m, k = 0 to 6). Furthermore, the values ​​of the output variables y(0), y(1), y(2), and y(3) are determined by inserting each of the values ​​obtained by transforming each of the values ​​of the nodes in the intermediate layer by the linear characteristic map defined by the coefficient wSij (i = 0 to 3) into the activation function g.A hyperbolic tangent or similar function can be used as the activation function f. A softmax function can be used as the activation function g.

[0061] As in Fig. As shown in Figure 5, the output variables y(0), y(1), y(2), and y(3) are state variables that indicate the state of the wave. The output variable y(0) indicates the probability that the wave is not in an abnormal vibration state, i.e., the wave is in a normal state. The output variable y(1) indicates the probability that the wave is in a divergent vibration state. The output variable y(2) indicates the probability that the wave is in a torsional vibration state. The output variable y(3) indicates the probability that the wave is in a bending vibration state.

[0062] In order to Fig. To return to step 2, CPU 42 selects the maximum value ymax among the output variables y(0) to y(3) (S106). Then, CPU 42 determines the state of the wave based on the output variable that is equal to the maximum value ymax of the output variables y(0) to y(3) (S107). Next, CPU 42 executes a memory process to store the result of the wave state determination in memory 46 (S108). For example, if the value of the output variable y(1) is equal to the maximum value ymax, CPU 42 stores in memory 46 that the wave is in a divergent oscillation state. In the process of S108, CPU 42 can also... Fig. The display shown in Figure 1, Display 70, is used to execute a notification process to report the state of the gearbox 26 based on the state of the shaft stored in memory 46. Display 70 is an example of a notification device that reports the state of the gearbox 26. The notification process can be executed, for example, by using a loudspeaker as the notification device and operating the loudspeaker to output an audio signal. When the process of S108 is complete, CPU 42 temporarily terminates the operation shown in Figure 108. Fig. 2. A series of processes is shown.

[0063] The DM map data is a pre-learned model. The DM map data is learned using teacher data, in which the data representing the current state of the shaft are marked as the correct answer for the normalized feature NFv and the oil temperature Toil. The normalized feature NFv in the teacher data is a value calculated using the same processes as S102 and S103, based on time-series shaft speed data Nshaft obtained by driving a prototype vehicle or similar. The DM map data is learned using the teacher data, including data when the shaft is in a normal state, data when the shaft is in a divergent vibration state, data when the shaft is in a torsional vibration state, and data when the shaft is in a bending vibration state.

[0064] The following describes the function and effect of the present embodiment. The CPU 42 determines the condition of the shaft based on the time series data of the shaft speed Nshaft. By referencing the time series data of the shaft speed Nshaft in this way, it is possible to determine whether an anomaly has occurred in the gearbox 26.

[0065] According to the embodiment described above, the following operations and effects can also be obtained. (1) The normalized characteristic NFv, calculated by processing the time series data of the shaft speed Nshaft, is used as an input variable for the characteristic map. The processing of the time series data into data representing the incidence distribution, as in Fig. As shown in Figure 4, it is easy to distinguish the feature indicated by the shaft speed time series data when an anomaly has occurred in the shaft from the feature indicated by the shaft speed time series data when no anomaly has occurred in the shaft. That is, the accuracy of distinguishing the shaft's condition can be improved, and thus the accuracy of determining whether an anomaly has occurred in the gearbox 26 can be improved. (2) When the time series data of the shaft speed Nshaft are processed into the normalized feature NFv, the shaft speed Nshaft is classified into classes. This allows the order of the input variables to be reduced. This makes it possible to reduce the computational effort of the process for determining the shaft anomaly. (3) Since the time series data of the shaft speed Nshaft are normalized to the normalized feature NFv, the fluctuation of the shaft speed Nshaft can be captured as a feature indicated by the time series data in order to determine the condition of the shaft without being affected by the magnitude of the absolute value of the shaft speed Nshaft. (4) In addition to the normalized characteristic variable NFv, the mean oil temperature Toil over the specified period is included in the input variables that are fed to the characteristic map all at once. This allows the values ​​of the state variables to be calculated taking the oil temperature into account. By inputting different types of variables, the accuracy of determining the shaft's condition can be improved. (5) After calculating the values ​​of the output variables y(0) to y(3), the CPU 42 determines the state of the shaft based on the maximum value among the output variables y(0) to y(3) and stores the state of the shaft in memory 46. This allows not only the indication of whether an anomaly has occurred in the shaft, but also the state of the shaft when an anomaly has occurred. In other words, it is possible to determine whether the shaft is in a divergent vibration state, a torsional vibration state, or a bending vibration state. (6) Since the result of the determination of the condition of the shaft is stored in memory 46, when the vehicle VC is taken to the repair shop, measures can be taken in accordance with the cause of the anomaly stored in memory 46. (7) At the initial point in time when an anomaly occurs in the gearbox 26 and the shaft begins to vibrate, it is possible that the gearbox 26 has not yet reached the stage where its performance deteriorates. By determining the condition of the shaft based on the changes occurring in the time-series data of the shaft speed Nshaft, it is possible to detect that the shaft has begun to vibrate abnormally. In other words, the anomaly can be detected before the performance of the gearbox 26 deteriorates. correspondence

[0066] An example of the anomaly detection device is the control unit 40 in Fig. 1. An example of the processor is the CPU 42 and the ROM 44 in Fig. 1. An example of memory is memory 46 in Fig. 1. An example of the characteristic map data is the DM characteristic map data. An example of the data acquisition process is the process of S104 in Fig. 2. An example of the determination process are the processes from S105 to S107 in Fig. 2. An example of the characteristic value calculation process are processes S102 and S103 in Fig. 2. Other embodiments

[0067] The foregoing embodiment can be modified to be implemented as follows. The foregoing embodiment and the modified examples described below can be combined within a technically consistent range. About the value referenced by the control unit

[0068] As in Fig. As shown in Figure 8, the control unit 40 can reference not only the oil temperature detected by the oil temperature sensor 56, but also the values ​​detected by various sensors contained in the vehicle VC. Of the values ​​shown in Fig. The 8 elements shown are the elements that make up the in Fig. The elements shown in 1 correspond to those shown, which for simplicity are designated with the same reference symbols, and their descriptions are omitted.

[0069] For example, the control unit 40 can refer to the wheel speed detected by a wheel speed sensor 57. The control unit 40 can calculate the vehicle speed SPD based on the wheel speed. The control unit 40 can reference the vibration VB detected by a vibration sensor 58 mounted on the vehicle. The control unit 40 can refer to the noise NZ detected by a noise sensor 59 attached to the transmission 26.

[0070] Furthermore, the control unit 40 can also reference the measured value taken by an instrument provided in the vehicle VC, or the state variable that can be detected by another control unit. For example, the control unit 40 can reference the integrated odometer reading OD of the vehicle VC, which is measured by an odometer 60 attached to the vehicle VC. The control unit 40 can also reference the brake pressure PB, which can be obtained from a brake control unit 80. The brake control unit 80 is a control unit for the braking system of the vehicle VC. The brake pressure PB can, for example, be the pressure with which the brake actuator is actuated. If the braking system is a hydraulic braking system, the pressure of the master cylinder can also be referred to as the brake pressure PB. Using time series data

[0071] Although the time series data of the shaft speed Nshaft in Fig. 3 are shown, limited Fig. 3 not the number of samplings of the shaft speed Nshaft in the specified period.

[0072] In the speed calculation process of the preceding embodiment, an example of calculating the shaft speed Nshaft based on the output signal Sm1 is described. The means for calculating the shaft speed Nshaft, which is the rotational speed of the shaft, are not limited to this. For example, the rotational speed of the shaft provided in the gearbox 26 correlates with the vehicle speed SPD. Therefore, the shaft speed Nshaft can also be calculated based on the vehicle speed SPD. It is also possible to use a sensor to detect the shaft speed in order to calculate the shaft speed Nshaft based on the value detected by the sensor. Using the characteristic size as an input variable

[0073] In the process of S102 in the aforementioned embodiment, the time-series data of the shaft speed Nshaft are normalized such that the maximum value of the shaft speed Nshaft in the specified period becomes "1" and the minimum value of the shaft speed Nshaft in the specified period becomes "0". The process for normalizing the time-series data of the shaft speed Nshaft is not limited to this. For example, it is conceivable to normalize the time-series data of the shaft speed Nshaft such that the mean value of the shaft speed Nshaft becomes "0" and the variance becomes "1".

[0074] In the S103 process, as described above, the range from the minimum to the maximum value of the shaft speed Nshaft during the specified period is divided into five equal ranges, and these ranges are classified into five classes. As long as the number of classes corresponds to the input variables, the number of classes set in the S103 process can be changed as needed. In other words, the width of each class can be modified.

[0075] In processes S102 and S103 of the preceding embodiment, the time-series data of the shaft speed Nshaft are normalized, and then the normalized data are processed to calculate the normalized feature NFv, which represents the incidence distribution. The present invention is not limited to this, and it is also possible to process the time-series data of the shaft speed Nshaft to calculate the feature Fv as the feature representing the incidence distribution, and then to calculate the normalized feature NFv based on the feature Fv.

[0076] In the foregoing embodiment, the normalized feature NFv is shown as an example input variable for the characteristic map defined by the characteristic map data DM, but the present invention is not limited thereto. For example, the feature Fv, which represents the incidence distribution data, can be calculated from the time series data of the shaft speed Nshaft, and the feature Fv can be used as an input variable. That is, the normalization process is not essential. Even if the width of the classes in the feature Fv is not adjusted, the feature Fv represents data obtained by extracting features from the time series data of the shaft speed Nshaft. Using the feature Fv as an input variable, the condition of the shaft can be determined.

[0077] In the foregoing embodiment, the data representing the incidence distribution are, by way of example, the feature obtained by processing the time series data of the shaft speed Nshaft, but the present invention is not limited thereto. For example, the distribution of the frequency components obtained by fast Fourier transform of the time series data of the shaft speed Nshaft can be calculated as a feature. By using the feature calculated in this way as an input variable for the characteristic map, the state of the shaft can be determined based on the feature of the time series data occurring in the frequency domain. That is, the state of the shaft can be determined based on the features obtained by frequency analysis of the time series data of the shaft speed Nshaft.The process of converting time series data into the frequency domain using the fast Fourier transform to calculate the feature parameter is referred to as the frequency analysis process. The feature parameter calculated by the frequency analysis process can be called the frequency parameter.

[0078] It is also possible to normalize the distribution of frequency components obtained by the fast Fourier transform of the shaft speed Nshaft time series data and to use the normalized characteristic value as an input variable for the characteristic map. That is, the normalized frequency characteristic value can be calculated during the frequency analysis process. For example, the primary frequency can be calculated from the average of the shaft speed Nshaft over the specified period, and the frequency components can be normalized using the primary frequency as a reference. By normalizing the frequency components, the shaft's condition can be determined based on characteristics obtained through the frequency analysis of the shaft speed Nshaft time series data, without being affected by the intensity of the frequency components.

[0079] If the characteristic value obtained through frequency analysis of the shaft rotational speed time series data Nshaft is used as an input variable, the frequency range can be divided into a multitude of frequency bands, and the mean of the intensities of the frequency components in each frequency band can be used as the intensity in that frequency band. This process is referred to as the reduction process. By performing the reduction process, the elements of the input variables can be reduced when the characteristic value obtained through frequency analysis of the shaft rotational speed time series data Nshaft is used as the input variable. That is, it is possible to reduce the computational load in the process of determining whether the shaft is in a state of vibration.

[0080] The processing content of the frequency analysis process for calculating the frequency characteristic value is not limited to the configuration of the modified example above. An example of another frequency analysis process is given with reference to Fig. 6 and Fig. 7 described. Fig. Figure 6 shows time series data of the shaft speed Nshaft when the shaft is in a torsional vibration state. As in Fig. As shown in Figure 6, the time series data in the torsional vibration state of the wave show a different transmission than in the example of the divergent vibration, which is described in section (b) in Fig. 3 is shown. Fig. Figure 7 shows a normalized feature size in which the time series data of the shaft speed Nshaft are processed by the frequency analysis process when the shaft is in a torsional vibration state.

[0081] In this frequency analysis process, the CPU 42 first calculates the distribution of the frequency components, which is obtained as a feature variable by fast Fourier transforming the time series data of the shaft rotational speed Nshaft. Next, the CPU 42 performs a process to subdivide the frequency range into a multitude of frequency bands, in the same manner as the reduction process. This reduction process can be omitted.

[0082] Furthermore, the CPU 42 performs a process to normalize the frequency components, so that the intensity of the frequency components at the primary resonant frequency becomes "1" when the wave's mode is a torsional oscillation. Values ​​calculated in advance through experimentation or similar methods are stored as the wave's resonant frequency.

[0083] As a result, in the Fig. In example 7, the maximum intensity value at the primary resonant frequency is displayed, and the intensity of the frequency components is normalized to the range from "0" to "1". In data from a torsional vibration state where resonance is generated, the intensity of the frequency components at the resonant frequency is relatively higher than the intensity at other frequencies. Therefore, if the data is input in a torsional vibration state and the frequency analysis process is performed as in the configuration above, as shown in Fig. As shown in Figure 7, the frequency characteristic quantity is calculated, which highlights the high intensities of the frequency components at the resonance frequencies of primary, secondary, ... nth order.

[0084] On the other hand, if the time series data in the state where no torsional vibration is generated are converted to the frequency domain, it can be assumed that the difference between the intensity at the resonant frequency (when the vibration mode is torsional) and the intensity at other frequencies is not large. Therefore, if the data in the state where torsional vibration is not generated are processed by the frequency analysis process, the distribution is assumed to have a high incidence across the entire frequency domain.

[0085] In this way, by calculating the normalized frequency characteristic quantity based on the intensity of the frequency components at the resonance frequency, the difference in the characteristics that can be extracted from the time series data of the shaft rotational speed Nshaft, depending on whether the shaft vibration mode is a torsional vibration or not, can be clearly demonstrated. This facilitates the differentiation of the shaft's state.

[0086] In the modified example above, a process for normalizing the frequency components is presented such that the intensity of the frequency components at the primary resonant frequency becomes "1". Alternatively, the frequency components can be normalized based on their intensity at the secondary resonant frequency or their intensity at the third-order resonant frequency. By normalizing the frequency components so that not only their intensity at the primary resonant frequency but also their intensity at the nth-order resonant frequency becomes "1", it is possible to calculate the normalized feature that highlights the high intensity of the frequency components at the resonant frequency, as in the modified example above.For example, if normalization is performed such that the intensity of the frequency components at the second-order resonance frequency becomes "1", it is to be expected that the intensity of the frequency components at the first-order resonance frequency will be greater than "1". If a value greater than "1" is calculated in this way, the calculated value can be used, or the value greater than "1" can be treated as "1".

[0087] The modified example above illustrates a process for normalizing the frequency components so that their intensity at the primary resonant frequency becomes "1". However, it is not essential that the maximum intensity of the frequency components also becomes "1". The normalized frequency characteristic can be calculated with respect to the intensity of the frequency components at the resonant frequency.

[0088] Furthermore, the preceding modified example presents a process for normalizing the frequency components such that the intensity of the frequency components at the resonance frequency becomes "1" if the wave mode is a torsional vibration. Alternatively, a process for normalizing the frequency components can also be performed such that the intensity of the frequency components at the resonance frequency becomes "1" if the wave mode is a bending vibration. Following this process, unlike the preceding modified example, it is possible, when the time series data obtained when the wave mode is a bending vibration is inputted, to calculate the frequency characteristic that highlights the high intensity of the frequency components of the bending vibration at the primary, secondary, ... nth order resonance frequencies.Therefore, the difference in the characteristics that can be extracted from the shaft rotational speed time series data can be clearly demonstrated depending on whether the shaft vibration mode is a bending vibration or not. This facilitates the differentiation of the shaft's condition.

[0089] Both the feature quantity representing the incidence distribution, such as the normalized feature quantity NFv in the preceding embodiment, and the frequency feature quantity obtained through the frequency analysis process as in the modified example above, can be used as input variables. By inputting a combination of differently processed data into the characteristic map, it becomes easier to distinguish the features indicated by the time series data, and thus the accuracy of the determination is expected to be further improved.

[0090] The preceding modified example shows an example in which the feature obtained by frequency analysis of the shaft speed Nshaft time series data is used as an input variable. The present invention is not limited thereto, and the feature obtained by analyzing the shaft speed Nshaft time series data through the rotational order ratio can also be used as the input variables. About the input variables for the characteristic map

[0091] In the above embodiment, the characteristic value obtained by processing the time series data of the shaft speed Nshaft is used as the input variables for the characteristic map. Alternatively, the time series data of the shaft speed Nshaft can also be used as input variables. For example, the sampled values ​​can be used as input variables.

[0092] In the foregoing embodiment, the mean oil temperature (Toil) is used as an input variable for the characteristic map, but the present invention is not limited thereto. For example, time series data of the oil temperature (Toil) can be used as an input variable.

[0093] It is not strictly necessary to include the oil temperature (Toil) in an input variable for the characteristic map defined by the characteristic map data (DM). The input variables only need to contain the time series data of the shaft speed (Nshaft).

[0094] The oscillation VB can be contained in an input variable for the characteristic map defined by the characteristic map data DM.

[0095] The noise NZ can be contained in an input variable for the map defined by the map data DM.

[0096] The integrated odometer reading (OD) can be included in an input variable for the map defined by the map data (DM). Alternatively, the integrated time during vehicle startup (VC) can be used instead of the integrated odometer reading (OD). With this configuration, the aging of the shaft, bearings, and the like can be taken into account when determining whether an anomaly has occurred in the gearbox 26.

[0097] The brake pressure PB can be included in an input variable for the characteristic map defined by the characteristic map data DM. According to this configuration, vibrations due to the deceleration of the vehicle VC, fluctuations in the shaft speed Nshaft due to the deceleration of the vehicle VC, and the like can be taken into account when determining whether an anomaly has occurred in the transmission 26.

[0098] The input variable for the characteristic map, defined by the characteristic map data DM, can contain design values ​​that specify the dimensions of the individual shaft sections based on the shaft specifications. These design values ​​include the shaft's inner diameter, outer diameter, axial length, and similar parameters.

[0099] For example, the shaft data DS, including the design values ​​that specify the dimensions of each section of the shaft based on the shaft specifications, can be stored in memory 46, as shown in Fig. Figure 8 shows that the CPU 42 can read the values ​​from the wave data DS and replace the values ​​in the input variables.

[0100] In addition to the specification data based on the specifications of the shaft mounted on the gearbox 26, the shaft data DS may contain the following data. Examples of data included in the shaft data DS are actual measured values ​​indicating the dimensions of each part of the shaft, measured after the shaft has been manufactured. Examples of data included in the shaft data DS are the fit tolerance value between the shaft and the bearing, the shaft's unbalance, and the coaxiality of the shaft supported by the bearing. The unbalance and coaxiality are values ​​measured in advance during the manufacturing of the gearbox 26.

[0101] The actual measured values ​​of the shaft can be included in the input variables for the characteristic map defined by the characteristic map data DM.

[0102] The imbalance amount can be included in the input variable for the characteristic map defined by the characteristic map data DM.

[0103] The value of the fit tolerance can be included in the input variables for the characteristic map defined by the characteristic map data DM.

[0104] The value of the coaxiality can be included in the input variables for the characteristic map defined by the characteristic map data DM.

[0105] The magnitude of the torque transmitted by the shaft can be included in the input variables for the characteristic map defined by the characteristic map data DM. Regarding the characteristic map

[0106] The neural network is not limited to a fully connected forward propagation network. For example, a one-dimensional convolutional neural network can also be used. The model learned through machine learning is not limited to neural networks. For example, the state of the wave can be distinguished using classification by a support vector machine.

[0107] In the process of S105, a neural network with an intermediate layer is shown, but the present invention is not limited to this, and the number of intermediate layers can be two or more.

[0108] The characteristic map is not limited to a map with the four output variables y(0), y(1), y(2), and y(3). If there is another state of the wave that can be specified based on the characteristic value obtained by processing the time series data, this wave state can be adopted as a new state variable and used as a state variable in the characteristic map.

[0109] The output variable of the characteristic map can be used to determine whether an anomaly has occurred in the waveform. That is, the probability that an anomaly occurred in the waveform, and the probability that an anomaly did not occur, can be calculated from the output variable provided by the characteristic map.

[0110] A large number of map data can be stored in memory 46 of the vehicle VC, as shown in Fig. Figure 8 shows that, as an example, three map data sets, from the first DM (A1) to the third DM (A3), can be stored in memory 46. The first DM (A1) is a pre-trained model, pre-trained using teacher data based on the normalized feature NFv when the shaft is in a normal state, and teacher data based on the normalized feature NFv when the shaft is in a divergent vibration state. The second DM (A2) is a pre-trained model, pre-trained using teacher data based on the normalized feature NFv when the shaft is in a normal state, and teacher data based on the normalized feature NFv when the shaft is in a torsional vibration state.The third characteristic map data DM (A3) is a pre-trained model that was pre-trained using teacher data based on the normalized feature size NFv when the shaft is in the normal state, and teacher data based on the normalized feature size NFv when the shaft is in a bending vibration state.

[0111] When a large number of map data sets are stored, a type of map data set is selected from the large number of map data sets, and the selected map data set is used to execute the determination process according to a sequence of processes that are described in Fig.Figure 2 of the preceding embodiment is shown. For example, when the first characteristic map data DM (A1) is selected, the probability that the shaft is in a normal state and the probability that the shaft is in a divergent vibration state are calculated as output variables indicating the shaft's state. The CPU 42 determines the shaft's state based on the calculated maximum value of the output variable. For example, the CPU 42 determines that the shaft is in a normal state if the result shows that the probability of the shaft being in a normal state is 80% or higher. Conversely, the CPU 42 determines that the shaft is in a divergent vibration state if the result shows that the probability of the shaft being in a divergent vibration state is 80% or higher.Furthermore, CPU 42 determines that an anomaly, which differs from a divergent oscillation, has occurred if the result shows that both the probability that the wave is in a normal state and the probability that the wave is in a divergent oscillation state are less than 80%.

[0112] By repeatedly selecting the characteristic map data and executing the determination process until the shaft state can be distinguished, it is possible to reduce the processor's computational load per process when the processor executes a series of processes that include the determination process, compared to the case where the shaft state is distinguished by a single type of characteristic map data. As a result, it is possible to suppress an increase in the processor's computational load and a decrease in the accuracy of the determination when determining whether an anomaly has occurred in gearbox 26.

[0113] In the modified example above, three map data points, from the first DM (A1) to the third DM (A3), are represented as a plurality of map data points. Alternatively, two map data points can be used, or four or more map data points can be applied as a plurality of map data points. About the storage process

[0114] In the foregoing embodiment, the memory for storing the determination result is the same as the memory for storing the characteristic map data DM, but the present invention is not limited thereto.

[0115] Instead of executing the storage process to save the result of the output variable's determination, a transmission process can also be executed to send the result to the vehicle manufacturer (VC), the data analysis center, or similar entity. It is also possible to execute both the storage and transmission processes simultaneously. Regarding the processor

[0116] The processor is not limited to a processor comprising the CPU 42 and the ROM 44 that performs software processing. For example, the processor may include a dedicated hardware circuit (e.g., an application-specific integrated circuit (ASIC), etc.) that executes at least some of the software processes performed in the preceding embodiment. That is to say, the processor need only have one of the following configurations (a) to (c): (a) a configuration comprising a processing unit that performs all of the above-mentioned processes according to a program and a program storage unit, such as a ROM, for storing the program; (b) a configuration comprising a processing unit that performs some of the foregoing processes according to a program, a program storage unit, and a dedicated hardware circuit that performs the remaining processes; and (c) a configuration that includes a dedicated hardware circuit that performs all of the above processes.

[0117] The above configurations can include a variety of software processors, including a processing unit and a program memory unit, and a variety of dedicated hardware circuits. To the vehicle

[0118] In the foregoing embodiment, the vehicle VC is shown by way of example with the internal combustion engine 10, the first motor-generator 22 and the second motor-generator 24. As long as the vehicle is equipped with a transmission for power transmission via a shaft, the control device 40 can be used, and the state of the shaft can be determined in the same way as in the foregoing embodiment. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] JP 2009063455 A [0002, 0003]

Claims

[1] An abnormality determining device applied to a vehicle provided with a transmission (26) adapted to transmit power by rotation of a shaft, the abnormality determining device comprising: a memory (46) configured to store map data which is data defining a map learned by machine learning, the map being a map in which an input variable is a variable indicative of time series data of a rotational speed of the shaft, and an output variable is a state variable indicative of a state of the shaft; and a processor configured to execute an acquisition process to acquire the variable indicating the time series data and use the variable as a value of the input variable, and execute a determination process to determine whether an abnormality has occurred in the transmission (26) based on a value of the output variable obtained using the value of the input variable and the map. [2] Anomaly determination device according to claim 1, wherein the processor is configured to execute a feature size calculation process for calculating a feature size obtained by processing the time series data; the acquisition process is a process to acquire the feature quantity and use the feature quantity as the value of the input variable; and the feature size calculation process is a process of dividing values of rotation speed constituting the time series data into classes according to a size of rotation speed and calculating an incidence for each of the classes as the feature size. [3] The abnormality determination apparatus according to claim 2, wherein the feature quantity calculation process includes a process of normalizing the time series data so that a maximum value of the rotational speed becomes "1" and a minimum value of the rotational speed becomes "0". [4] Anomaly determination device according to claim 2 or 3, wherein: the feature variable is a time feature variable; the feature size calculation process is an incidence analysis process to calculate the time feature size; the processor is configured to perform a frequency analysis process to calculate a distribution of frequency components obtained by fast Fourier transform of the time series data as a frequency feature quantity; and in the acquisition process, the time feature size and the frequency feature size obtained by processing the time series data are acquired and used as the value of the input variable. [5] Anomaly determination device according to claim 1, wherein: the processor is configured to perform a frequency analysis process to calculate a frequency feature quantity obtained by processing the time series data; the acquisition process is a process to acquire the frequency feature quantity and use the frequency feature quantity as the value of the input variable; and the frequency analysis process is a process to calculate a distribution of frequency components obtained by fast Fourier transform of the time series data as the frequency feature quantity. [6] The abnormality determination device according to claim 4 or 5, wherein the frequency analysis process includes a process of calculating a primary frequency based on an average value of the rotational speed in the time series data and normalizing the frequency components with the primary frequency as a reference. [7] The anomaly determining device according to claim 4 or 5, wherein the frequency analysis process includes a process for normalizing the frequency components so that an intensity of the frequency components at a primary resonance frequency when a vibration mode of the shaft is a torsional vibration becomes "1". [8] The anomaly determining device according to claim 4 or 5, wherein the frequency analysis process includes a process for normalizing the frequency components so that an intensity of the frequency components at a primary resonance frequency when a vibration mode of the shaft is a bending vibration becomes "1". [9] An abnormality determination device according to any one of claims 1 to 8, wherein the time series data is calculated based on a detection signal of a rotational speed sensor configured to detect the rotational speed of the shaft or based on a vehicle speed of the vehicle. [10] An abnormality determination device according to any one of claims 1 to 9, wherein the input variable includes at least one variable indicative of a temperature of hydraulic oil in the transmission (26), a magnitude of torque transmitted by the shaft, a dimension of the shaft based on specifications of the shaft, an unbalance amount of the shaft, a fitting tolerance between the shaft and a bearing of the shaft, a coaxiality between the shaft and a bearing of the shaft, a value detected by a vibration sensor configured to detect vibrations, a value detected by a noise sensor configured to detect noises, a hydraulic pressure in a braking device of the vehicle, and / or a mileage of the vehicle. [11] An abnormality determination device according to any one of claims 1 to 10, wherein the state variable includes at least one of a variable indicating that the shaft is in a divergent vibration state, a variable indicating that the shaft is in a torsional vibration state, and a variable indicating that the shaft is in a flexural vibration state.

Images