Magnet temperature estimation device

Abstract

A magnet temperature estimation device is provided, which obtains parameters related to the rotation of a motor (2) measured at fixed intervals and calculates the moving average value of the parameters for each fixed interval. The calculated moving average value is input into a learning model that has been trained to output the temperature of the magnets on the rotor (7) of the motor (2) when the moving average value of the parameters related to the rotation of the motor (2) is input, and the estimated value of the magnet temperature output from the learning model is obtained. Next, the obtained estimated value of the magnet temperature is output.

Description

Technical Field

[0001] This invention relates to a magnet temperature estimation device. Background Technology

[0002] In a rotary electric motor consisting of a rotor and a stator equipped with permanent magnets, the magnetic force of the permanent magnets decreases when the temperature of the permanent magnets increases, causing the electric motor to fail to output the intended output. Therefore, it is necessary to detect the temperature of the permanent magnets. However, in commercially available vehicles, it is difficult to detect the temperature of the permanent magnets mounted on the rotor. Therefore, a magnet temperature estimation device for estimating the temperature of the permanent magnets mounted on the rotor is known (see, for example, Japanese Patent Application Publication No. 2014-93867).

[0003] In this magnet temperature estimation device, the temperature of the rotor, i.e. the temperature of the magnet, is estimated by calculating the heat dissipation from the rotor to the oil per unit time based on the temperature difference between the estimated rotor temperature and the temperature of the oil flowing around the rotor, and the difference between the heat dissipation from the rotor to the oil per unit time based on the temperature rise of the oil. The calculated temperature rise of the rotor per unit time is accumulated to estimate the temperature of the rotor, i.e. the temperature of the magnet. Summary of the Invention

[0004] However, when inferring the magnet temperature by accumulating the temperature rise of the rotor per unit time, there is a problem: during the accumulation of the temperature rise per unit time, errors accumulate, and the inferred magnet temperature often deviates significantly from the actual magnet temperature.

[0005] Therefore, according to the present invention, a magnet temperature estimation device is provided, comprising:

[0006] The parameter acquisition unit acquires parameters related to the rotation of the target motor, measured at fixed intervals.

[0007] The calculation unit calculates the moving average of the parameters over each fixed interval.

[0008] The temperature acquisition unit inputs a moving average calculated by the calculation unit into a learning model that has been trained to output the temperature of the magnets mounted on the rotor of the motor based on a moving average of parameters related to the motor's rotation, and obtains a predicted value of the magnet temperature output from the learning model; and

[0009] The output section outputs the estimated value of the magnet temperature obtained by the temperature acquisition section.

[0010] According to the present invention, the temperature of the magnet can be accurately predicted. Attached Figure Description

[0011] Figure 1 This is an overall diagram of the magnet temperature estimation device.

[0012] Figure 2 This is a timeline diagram used to illustrate the method of predicting magnet temperature.

[0013] Figure 3 This is a diagram used to illustrate one-dimensional convolution processing.

[0014] Figure 4A as well as Figure 4B This is a diagram used to illustrate one-dimensional convolution processing.

[0015] Figure 5 This is a diagram illustrating the construction of a one-dimensional convolutional neural network.

[0016] Figure 6 This is a diagram used to illustrate one-dimensional convolution processing.

[0017] Figure 7 This is a diagram used to illustrate one-dimensional convolution processing.

[0018] Figure 8A as well as Figure 8B This is a diagram used to illustrate one-dimensional convolution processing.

[0019] Figure 9 This is a timeline diagram used to illustrate the method of predicting magnet temperature.

[0020] Figure 10A as well as Figure 10B These are graphs showing a list of the obtained parameters and a list of moving averages, respectively.

[0021] Figure 11A as well as Figure 11B These are graphs illustrating simple moving average processing and exponential smoothing moving average processing, respectively.

[0022] Figure 12 This is a flowchart used to estimate the temperature of a magnet.

[0023] Figure 13 This is a functional structure diagram of the present invention.

[0024] Figure 14 This is a graph showing the relationship between the estimation error of magnet temperature and the frequency distribution.

[0025] Figure 15 It is a graph showing the relationship between the calculation time interval and the vehicle's driving load.

[0026] Figure 16 This is a flowchart used to estimate the temperature of a magnet.

[0027] Figure 17A as well as Figure 17BThese are diagrams illustrating the relationship between magnet temperature and the upper limit of drive current, as well as a flowchart for executing drive current control.

[0028] Figure 18A as well as Figure 18B These are diagrams illustrating the relationship between magnet temperature and the drive current of the oil pump drive motor, as well as a flowchart for executing the drive control of the oil pump. Detailed Implementation

[0029] In reference Figure 1 In this designation, 1 represents the housing of the transaxle of the hybrid vehicle, 2 represents the motor for driving the vehicle, 3 represents the oil pump, and 4 represents the oil cooler. The motor 2 includes a rotating shaft 6 rotatably supported by bearings 5, a rotor 7 fixed to the rotating shaft 6, and a stator 8 surrounding the rotor 7. In an embodiment of the invention, a permanent magnet (not shown) is mounted on the rotor 7. In this case, the permanent magnet may be embedded within the rotor 7 or fixed to the outer peripheral surface of the rotor 7. On the other hand, stator coils (not shown) are arranged within the stator 8. When a drive current is supplied to the stator coils, the rotor 7 rotates due to the interaction between the stator coils and the magnetic field of the permanent magnets on the rotor 7. Furthermore, the motor 2 is used not only as a driving force source for the vehicle but also as a generator. The supply control of the drive current to the stator coils is performed by an inverter 9, which is controlled by an electronic control unit 20.

[0030] Engine oil is supplied to the drive axle housing 1 to lubricate and cool the motor 2 and reduction gear mechanism. The engine oil accumulated at the bottom of the drive axle housing 1 is pumped by the oil pump 3 into the oil cooler 4. In the oil cooler 4, the oil, cooled by heat exchange with the engine coolant, is then pumped back into the drive axle housing 1. The oil pump 3 can be driven by the engine or by an oil pump-driven motor. Figure 1 In the embodiment shown, the oil pump 3 is driven by the engine.

[0031] like Figure 1 As shown, the electronic control unit 20 is composed of a digital computer and includes a ROM (Read-Only Memory) 22, a RAM (Random Access Memory) 23, a CPU (Microprocessor) 24, an input port 25, and an output port 26 interconnected via a bidirectional bus 21. A temperature sensor 10 for measuring the temperature of the stator coils is installed on the stator 8, and the output signal of the temperature sensor 10 is input to the input port 25 via a corresponding AD converter 27. A speed sensor 11 for measuring the rotational speed of the motor 2 is installed on the rotating shaft 6 of the motor 2, and the output signal of the speed sensor 11 is input to the input port 25 via a corresponding AD converter 27.

[0032] Additionally, a temperature sensor 12 is installed at the bottom of the drive axle housing 1 to measure the temperature of the oil stored within the housing 1 of the oil drive axle. The output signal of the temperature sensor 12 is input to the input port 25 via a corresponding AD converter 27. Furthermore, a speed sensor 13 is installed on the oil pump 3 to measure its rotational speed. The output signal of the speed sensor 13 is input to the input port 25 via a corresponding AD converter 27. Moreover, when the oil pump 3 is driven by the engine, a speed sensor that detects engine speed can also be used as the speed sensor 13.

[0033] On the other hand, signals representing the drive current of motor 2, the drive voltage of motor 2, and the inverter frequency are input from inverter 9 to input port 25 via corresponding A / D converters 27. In this case, the drive torque of motor 2 is calculated within electronic control unit 20 based on the drive current and drive voltage of motor 2.

[0034] Furthermore, the magnetic force of a permanent magnet decreases as the temperature increases. Therefore, when the temperature of the permanent magnet installed in the rotor 7 increases, the motor 7 will not output the intended output. Thus, it is necessary to detect the temperature of the permanent magnet. However, in commercially available vehicles, it is difficult to detect the temperature of the permanent magnet installed in the rotor 7, so it is necessary to infer the temperature of the permanent magnet. Therefore, the results of repeated studies on the temperature of the permanent magnet have clarified that the temperature of the permanent magnet installed in the rotor 7 is related to parameters related to the rotation of the motor 7, and that the temperature of the permanent magnet installed in the rotor 7 is significantly affected by the time-varying changes of these parameters from the past to the present.

[0035] In this case, regarding the temperature of the permanent magnet installed on the rotor 7, it is clarified that among the parameters related to the rotation of the motor, the temperature of the stator coil, the speed of the motor 2, the temperature of the engine oil, and the speed of the oil pump 3 are particularly strongly correlated. The time-varying changes of these stator coil temperatures, motor 2 speeds, engine oil temperatures, and oil pump 3 speeds from the past to the present have a significant impact on the temperature of the permanent magnet installed on the rotor 7. Figure 2 The diagram shows an example of the time-varying temperature Ts of these stator coils, the rotational speed Rm of motor 2, the oil temperature Ti, and the rotational speed Rp of oil pump 3. Figure 2 The solid line in the figure represents the actual time variation of the temperature Tr of the permanent magnet installed on rotor 7.

[0036] In addition, Figure 2In this configuration, the temperature Ts of the stator coil is measured by temperature sensor 10, the rotational speed Rm of motor 2 is measured by speed sensor 11, the temperature Ti of the engine oil is measured by temperature sensor 12, and the rotational speed Rp of oil pump 3 is measured by speed sensor 13. Alternatively, a telemetry device capable of wirelessly transmitting temperature information detected using a thermistor can be used to detect the actual temperature Tr of the permanent magnet mounted on rotor 7. In this case, the telemetry device is embedded in rotor 7 to detect the actual temperature Tr of the permanent magnet. Furthermore, the actual temperature Tr of the permanent magnet mounted on rotor 7 can also be obtained through simulation.

[0037] Furthermore, as mentioned above, it is clear that the time variations of the stator coil temperature Ts, the motor 2 rotational speed Rm, the oil temperature Ti, and the oil pump 3 rotational speed Rp from the past to the present have a significant impact on the temperature of the permanent magnet mounted on the rotor 7. Therefore, in the embodiments of the present invention, based on the time variations of the stator coil temperature Ts, the motor 2 rotational speed Rm, the oil temperature Ti, and the oil pump 3 rotational speed Rp from the past to the present, a one-dimensional convolutional neural network is used to infer the current temperature Tr of the permanent magnet mounted on the rotor 7. Furthermore, in Figure 2 In the figure, the dashed line represents the time variation of the estimated value of the temperature Tr of the permanent magnet installed on rotor 7.

[0038] Furthermore, in an embodiment of the present invention, measurements of the stator coil temperature Ts, the motor 2 rotational speed Rm, the oil temperature Ti, and the oil pump 3 rotational speed Rp are obtained every second. In this case, in an embodiment of the present invention, based on the time series data of the stator coil temperature Ts, motor 2 rotational speed Rm, oil temperature Ti, and oil pump 3 rotational speed Rp obtained every second during the 10-second period from time t-9 to time t, a one-dimensional convolutional neural network is used to calculate the estimated value TTr of the current temperature Tr of the permanent magnet mounted on the rotor 7.

[0039] Therefore, next, with Figure 2 The situation shown is an example; please refer to... Figures 3 to 5 This section outlines the inference method using a one-dimensional convolutional neural network. Figure 3 The upper portion shows a table summarizing the measured values ​​a, b, c, and d for each second during the 10-second period from time t-9 to time t. Figure 2 In the cases shown, these measured values ​​a, b, c, and d represent the measured values ​​of the stator coil temperature Ts, the motor speed Rm, the oil temperature Ti, and the oil pump speed Rp.

[0040] On the other hand, Figure 3The central section shows the filters applied to these measurements a, b, c, and d. In these filters, four elements wa are assigned to each measurement value a, b, c, and d. (1)1 ~wa (1)4 wb (1)1 ~wb (1)4 wc (1)1 ~wc (1)4 and wd (1)1 ~wd (1)4 That is, in this filter, the filter size is 4 and the number of channels is 4. In this case, initially at time t-6 ( Figure 3 ), calculated in Figure 3 The sum of the products of the measured values ​​a, b, c, and d from time t-9 to time t-6, enclosed by dashed lines in the upper part, and the corresponding values ​​of the filter elements. That is, for the time interval... Figure 3 In the upper part, the measured values ​​a, b, c, and d from time t-9 to time t-6, enclosed by dashed lines, are multiplied by the corresponding element values ​​of the filter, and the sum of the 16 multiplication results is calculated (a t-9 ·wa (1)1 +a t-8 ·wa (1)2 +……+d t-7 ·wd (1)3 +d t-6 ·wd (1)4 The value of the sum of the products calculated becomes... Figure 3 The output value z at time t-6, enclosed by a dashed line in the lower part of the graph. (1)t-6 .

[0041] Next, in Figure 3 At time t-5, calculate Figure 3 The sum of the products of the measured values ​​a, b, c, d from time t-8 to time t-5 enclosed by a single-dotted line in the upper part of the graph, and the corresponding values ​​of the filter elements (a t-8 ·wa (1)1 +a t-7 ·wa (1)2 +……+d t-6 ·wd (1)3 +d t-5 ·wd (1)4 The value of the sum of the products calculated becomes... Figure 3 The output value z at time t-5 enclosed by a single-dot dash in the lower part. (1)t-5 That is, targeting in Figure 3 The detection values ​​a, b, c, and d from time t-9 to time t-6, enclosed by dashed lines in the upper part of the image, are convolved by a filter to obtain the output value z. (1)t-6 In response to Figure 3The measured values ​​a, b, c, and d from time t-8 to time t-5, enclosed by a single-dot dash in the upper part, are convolved through a filter to obtain the output value z. (1)t-5 .

[0042] In this way, for each measured value a, b, c, d, the filter is moved little by little while performing filter-based convolution to calculate the output value (z) from time t-6 to time t. (1)t-6 ...z (1)t On the other hand, in the example of this one-dimensional convolutional neural network, such as Figure 4A As shown, 19 filters with a filter size of 4 and a channel count of 4 are also used. Regarding each of these 19 filters, for... Figure 3 The measured values ​​a, b, c, and d shown are used to perform filter-based convolution while gradually moving the filter, thus achieving the desired result. Figure 4B The figure shows the calculation of the output values ​​(z) from time t-6 to time t for 19 time periods. (2)t-6 ...z (2)t ...(z) (20)t-6 ...z (20)t That is, calculate the output values ​​(z) for all 20 times from time t-6 to time t. (1)t-6 ...z (1)t ...(z) (20)t-6 ...z (20)t Using the output values ​​from time t-6 to time t, a one-dimensional convolutional neural network is trained. The trained one-dimensional convolutional neural network is then used to calculate the predicted value TTr of the current temperature Tr of the permanent magnet mounted on rotor 7. Furthermore, in this case, the filter's movement can be set to an arbitrary amount, i.e., an arbitrary step size.

[0043] Figure 5 This illustrates the construction of a one-dimensional convolutional neural network. (The following is a convolutional neural network structure, which is not directly related to the previous sentence.) Figure 3 When the measured values ​​a, b, c, and d shown are input into a one-dimensional convolutional neural network, in Figure 5 In the convolutional layer, these measured values ​​a, b, c, and d are subjected to convolution processing based on each filter, and the following calculations are performed: Figure 3 as well as Figure 4B The output values ​​(z) for all 20 times from t-6 to t are shown. (1)t-6 ...z (1)t ...(z) (20)t-6 ...z (20)tThese output values ​​are multiplied by an activation function such as a sigmoid function, and the resulting output values ​​are input to the nodes of the fully connected layer. Furthermore, in this case, a pooling layer can be placed before the fully connected layer. The outputs from the nodes of the fully connected layer are input to the nodes of the output layer, from which the predicted value TTr of the current temperature Tr of the permanent magnet is output.

[0044] However, in response to such Figure 3 When performing convolution operations based on filters on the measured values ​​a, b, c, d for each second during the 10-second period from time t-9 to time t, as shown in the overview table in the upper part, a large storage capacity is required due to the large amount of data to be stored, and a large number of weights to be learned by the one-dimensional convolutional neural network. Therefore, it takes time to calculate the predicted value TTr of the current temperature Tr of the permanent magnet. Therefore, in the embodiment of the present invention, in order to reduce the data storage capacity and the amount of weights to be learned so that the time changes of the measured values ​​a, b, c, d from the past to the present do not impair the effect on the predicted value TTr of the current temperature Tr of the permanent magnet, the moving average of the measured values ​​a, b, c, d is calculated, and convolution processing based on filters is performed on the calculated moving average of the measured values ​​a, b, c, d.

[0045] Therefore, the following reference Figures 6 to 8B To illustrate one-dimensional convolution processing using moving averages. (In the initial reference...) Figure 6 At that time, Figure 6 The upper part shows the relationship with Figure 3 The upper portion shows a list of measured values ​​a, b, c, and d for each second during the 10-second period from time t-9 to time t. In an embodiment of the invention, for... Figure 6 The measured values ​​a, b, c, and d shown in the upper part are processed using a moving average. Figure 6 In the example shown, at time t-5, with respect to the measured value a, the calculation is performed... Figure 6 The measurements taken during the past 5 seconds from time t-9 to time t-5, enclosed by dashed lines in the upper part (a) t-9 a t-8 a t-7 a t-6 a t-5 The moving average of the measured value (a) t-9 a t-8 a t-7 a t-6 a t-5 The moving average of ) becomes Figure 6The lower part of the table lists the time t-5, with ma enclosed in dashed lines. t-5 Additionally, at time t-4, regarding the measured value a, the calculation is performed... Figure 6 The measurements taken in the upper part of the image, enclosed by a single-dot dash, during the past 5 seconds from time t-8 to time t-4 (a) t-8 a t-7 a t-6 a t-5 a t-4 The moving average of the measured value (a) t-8 a t-7 a t-6 a t-5 a t-4 The moving average of ) becomes Figure 6 The lower part of the table lists the time t-4, with ma enclosed in a single-dotted line. t-4 .

[0046] Similarly, calculate the moving averages ma from time t-3 to time t. t-3 ma t-2 ma t-1 ma t Additionally, such as Figure 6 As shown in the table below, for the measured value b, the moving averages mb from time t-5 to time t are also calculated in the same way. t-5 mb t-4 mb t-3 mb t-2 mb t-1 mb t Similarly, for the measured value c, the moving average mc from time t-5 to time t is also calculated. t-5 ,mc t-4 ,mc t-3 ,mc t-2 ,mc t-1 ,mc t Similarly, for the measured value d, the moving average md from time t-5 to time t is also calculated. t-5 md t-4 md t-3 md t-2 md t-1 md t For all measured values ​​a, b, c, and d from time t-9 to time t, calculate as follows: Figure 6 The lower part of the table shows the moving average ma t-5 ma t-4 ...md t-1 md t Then, for these moving averages ma t-5 mat-4 ...md t-1 md t Perform one-dimensional convolution processing.

[0047] Next, refer to Figures 7 to 8B This indicates that the moving average ma t-5 ma t-4 ...md t-1 md t One-dimensional convolution processing. Furthermore, for this moving average ma... t-5 ma t-4 ...md t-1 md t One-dimensional convolution processing is also performed by referring to... Figures 3 to 4B The same method is used for processing one-dimensional convolutions. (Refer to...) Figure 7 At that time, Figure 7 The upper part shows the relationship with Figure 6 The list shown in the lower part is the same as the list shown below. On the other hand, in Figure 7 The central part shows the moving average ma t-5 ma t-4 ...md t-1 md t The applied filter. This filter is also related to... Figure 3 Similarly, the filter shown in the central part has a filter size of 4 and a channel count of 4.

[0048] In this case, also initially in Figure 7 At time t-2, calculate Figure 7 The moving averages ma from time t-5 to time t-2, enclosed by dashed lines in the upper part of the graph. t-5 ma t-4 ...md t-3 md t-2 The sum of the values ​​of the corresponding elements of the filter. That is, for the filter... Figure 7 The moving averages ma from time t-5 to time t-2, enclosed by dashed lines in the upper part of the graph. t-5 ma t-4 ...md t-3 md t-2 Multiply by the values ​​of the corresponding elements of the filter, and calculate the sum of the 16 multiplication results (ma). t-5 ·wa (1)1 +ma t-4 ·wa (1)2 +……+md t-3 ·wd (1)3 +md t-2 ·wd (1)4 The value of the sum of the products calculated becomes... Figure 7 The output value z at time t-2, enclosed by a dashed line in the lower part of the graph. (1)t-2 .

[0049] Next, in Figure 7 At time t-1, calculate Figure 7 The moving averages ma from time t-4 to time t-1 enclosed by a single-dot dash in the upper part of the graph. t-4 ma t-3 ...md t-2 md t-1 The sum of the products of the values ​​of the corresponding elements of the filter (ma) t-4 ·wa (1)1 +ma t-3 ·wa (1)2 +……+md t-2 ·wd (1)3 +md t-1 ·wd (1)4 The value of the sum of the products calculated becomes... Figure 7 The output value z at time t-1 enclosed by a single-dot dash in the lower part (1)t-1 That is, targeting in Figure 7 The moving averages ma from time t-5 to time t-2, enclosed by dashed lines in the upper part of the graph. t-5 ma t-4 ...md t-3 md t-2 The result obtained by convolution processing through a filter is the output value z. (1)t-2 In response to Figure 7 The moving averages ma from time t-4 to time t-1 enclosed by a single-dot dash in the upper part of the graph. t-4 ma t-3 ...md t-2 md t-1 The result obtained by convolution through the filter is the output value z. (1)t-1 .

[0050] This applies to each moving average ma t-5 ma t-4 ...md t-1 md t While gradually moving the filter, filter-based convolution is performed to calculate the output value (z) from time t-2 to time t. (1)t-2 ...z (1)t On the other hand, in the example of this one-dimensional convolutional neural network, it is also as follows: Figure 8A The diagram shows 19 filters with a filter size of 4 and a channel count of 4. Regarding each of these 19 filters... Figure 7 The moving averages ma shownt-5 ma t-4 ...md t-1 md t While gradually moving the filter, filter-based convolution is performed, thus achieving... Figure 8B As shown, the output values ​​(z) from time t-2 to time t are calculated for 19 time intervals. (2)t-2 ...z (2)t ...(z) (20)t-2 ...z (20)t That is, calculate the output values ​​(z) for all 20 times from time t-2 to time t. (1)t-2 ...z (1)t ...(z) (20)t-2 ...z (20)t Using the output values ​​from time t-2 to time t, a one-dimensional convolutional neural network is trained. Using the trained one-dimensional convolutional neural network, the predicted value TTr of the current temperature Tr of the permanent magnet on rotor 7 is calculated. Furthermore, in this case, the filter's movement can be set to an arbitrary amount, i.e., an arbitrary step size.

[0051] exist Figure 9 In, it is shown Figure 2 The following is an example of the time variation of the moving average values ​​of the stator coil temperature Ts, motor 2 speed Rm, oil temperature Ti, and oil pump 3 speed Rp when the moving average values ​​of the stator coil temperature Ts, motor 2 speed Rm, oil temperature Ti, and oil pump 3 speed Rp are shown, along with the actual time variation of the permanent magnet temperature Tr of the rotor 7 (solid line) and the estimated time variation of the permanent magnet temperature Tr of the rotor 7 (dashed line).

[0052] However, in Figure 9 In the example shown, as described above, the output values ​​from time t-2 to time t are used to learn the one-dimensional convolutional neural network. Therefore, the learning period of the one-dimensional convolutional neural network becomes... Figure 9 The range is represented by S in the text. In contrast, in... Figure 2 In the example shown, as described above, the output values ​​from time t-6 to time t are used to learn the one-dimensional convolutional neural network. Therefore, the learning period of the one-dimensional convolutional neural network becomes... Figure 2 The range is represented by S. In this case, regardless of Figure 9 The situation shown is still Figure 2 As shown, the amount of information used in the learning of a one-dimensional convolutional neural network is roughly the same. Therefore, when targeting a moving average ma... t-5 ma t-4 ...md t-1 mdt The case where convolution processing has been performed, i.e. Figure 9 In the case shown, compared to Figure 2 The scenario shown demonstrates how to shorten the learning period of a one-dimensional convolutional neural network while maintaining roughly the same amount of information used in the learning process. As a result, the amount of data retained can be reduced, thus decreasing storage capacity.

[0053] Next, refer to Figure 10A as well as Figure 10B This illustrates the learning process of a one-dimensional convolutional neural network. For example, the learning process of this one-dimensional convolutional neural network is... Figure 1 This is performed within the electronic control unit 20 shown. During the learning of the one-dimensional convolutional neural network, the driving load of the vehicle is initially varied. At this time, for each unit of time, for example, the temperature Ts of the stator coil, the speed Rm of the motor 2, the temperature Ti of the oil, the speed Rp of the oil pump 3, and the actual temperature Tr of the permanent magnet of the rotor 7 are measured every 1 second and stored in the RAM 23 of the electronic control unit 20. Figure 10A The stator coil temperature Ts, motor speed Rm, oil temperature Ti, oil pump speed Rp, and the actual temperature Tr of the permanent magnets on rotor 7 at various times t1, t2, t3, t4, t5, t6, t7, t8... stored in RAM 23 of electronic control unit 20 are displayed. Furthermore, Figure 10A The time t1 in the data represents the time at which the storage of the stator coil temperature Ts, etc., begins.

[0054] Within the electronic control unit 20, according to the RAM 23 stored in the electronic control unit 20 Figure 10A The moving average is calculated from the measured values ​​shown, and the calculated moving average is stored in the RAM23 of the electronic control unit 20. Figure 10B A table showing the moving averages stored in the RAM 23 of the electronic control unit 20 is provided. Furthermore, the following example illustrates the calculation of the moving average of measurements over the past 4 seconds. Figure 11A as well as Figure 11B To explain Figure 10B The moving averages are shown in the overview table.

[0055] Figure 11A This illustrates the case where a moving average obtained through a simple moving average (SMA) is used as the moving average. In this case, as... Figure 11A As shown, Figure 10B The moving average ma4 at time t4 is Figure 10A The simple average of the measured values ​​a1, a2, a3, and a4 from time t1 to time t4. Figure 10BThe moving average mb4 at time t4 is Figure 10A The simple average of the measured values ​​b1, b2, b3, and b4 from time t1 to time t4. Figure 10B The remaining moving averages mc4 and md4 at time t4 are also the same.

[0056] In addition, such as Figure 11A As shown, Figure 10B The moving average ma5 at time t5 is Figure 10A The simple average of the measured values ​​a2, a3, a4, and a5 from time t2 to time t5. Figure 10B The moving average mb5 at time t5 is Figure 10A The simple average of the measured values ​​b2, b3, b4, and b5 from time t2 to time t5. Figure 10B The remaining moving averages mc5 and md5 at time t5 are also the same. Furthermore, at... Figure 11A The table below shows the general formula for the Simple Moving Average (SMA).

[0057] on the other hand, Figure 11B This illustrates the use of a moving average obtained through an exponentially smoothed moving average (EMA) as a moving average. The EMA is calculated by using the previous exponentially smoothed moving average (EMA) instead of past measurements, doubling the weight of the current measurements, and then averaging them. Regarding this exponentially smoothed moving average (EMA), it is calculated using each measurement without calculating the previous exponentially smoothed moving average (EMA).

[0058] For example, in Figure 10B At times t1, t2, and t3, the exponential moving average (EMA) is not calculated, therefore... Figure 10B At time t4, the measured values ​​are used to calculate the exponentially smoothed moving average (EMA). That is, as... Figure 11B As shown, Figure 10B The exponential smoothed moving average (ma4) at time t4 becomes the... Figure 10A The value is obtained by dividing the sum of the measured values ​​a1, a2, and a3 from time t1 to time t3 and twice the measured value a4 at time t4 by the number of terms in the numerator (=5). Figure 10B The exponentially smoothed moving average mb4 at time t4 becomes the time to... Figure 10A The value is obtained by summing the measured values ​​b1, b2, and b3 from time t1 to time t3 and twice the measured value b4 at time t4, divided by the number of terms in the numerator (=5). Regarding Figure 10B The remaining exponentially smoothed moving averages mc4 and md4 at time t4 are also the same.

[0059] On the other hand, such as Figure 11B As shown, in Figure 10B In the calculation of the exponential smoothed moving average ma5 at time t5, instead of Figure 10A The measured values ​​a2, a3, and a4 from time t2 to time t4 are respectively used with the previous exponentially smoothed moving average ma4. The exponentially smoothed moving average ma5 is obtained by summing the three previous exponentially smoothed moving averages ma4 and twice the measured value a5 at time t5, and dividing by the number of terms in the numerator (=5). Figure 10B The exponentially smoothed moving average mb5 at time t5 is obtained by dividing the sum of the three previous exponentially smoothed moving averages mb4 and twice the measured value b5 at time t5 by the number of terms in the numerator (=5). Regarding Figure 10B The remaining moving averages mc5 and md5 at time t5 are also the same. Furthermore, at... Figure 11B The table below shows the general formula for the exponentially smoothed moving average (EMA).

[0060] According to the RAM 23 stored in the electronic control unit 20 Figure 10B The table showing the moving averages uses Figure 5 The one-dimensional convolutional neural network shown is used to learn the weights of the one-dimensional convolutional neural network. That is, Figure 10B This shows the training dataset used for learning the weights of a one-dimensional convolutional neural network. When learning the weights of the one-dimensional convolutional neural network, the moving averages ma4...ma9, mb4...mb9, mc4...mc9, and md4...md9 from time t4 to time t9 are initially input into... Figure 5 The one-dimensional convolutional neural network shown performs filter-based convolution processing on the moving averages ma4...ma9, mb4...mb9, mc4...mc9, and md4...md9 from the input time t4 to t9, while gradually moving the filter, thereby calculating the output value (z) from time t7 to t9. (1)7 z (1)8 z (1)9 ...(z) (20)7 z (20)8 z (20)9 ).

[0061] The output values ​​from time t7 to time t9 are multiplied by activation functions such as the sigmoid function, and the resulting output values ​​are input to the nodes of the fully connected layer. The outputs from the nodes of the fully connected layer are input to the nodes of the output layer, which output a value representing the temperature of the permanent magnet. To minimize the difference between this output value and the teacher's data (i.e., the actual temperature Tr9 of the permanent magnet at time t9), backpropagation is used to learn the values ​​of each element of the filter and the weights of the fully connected layer, i.e., the weights of the one-dimensional convolutional neural network. After learning the weights of the one-dimensional convolutional neural network using the output values ​​from time t7 to time t9, the weights from time t8 to time t9 are then used... 10 The weights of a one-dimensional convolutional neural network are learned to produce the output values. This continues until the desired output value is achieved. Figure 10B The weights of the one-dimensional convolutional neural network are learned up to the last moment in the list.

[0062] After the weights of the one-dimensional convolutional neural network are learned, the learned one-dimensional convolutional neural network is stored in the RAM23 of the electronic control unit 20, and the learned one-dimensional convolutional neural network is used to perform the temperature Tr estimation processing of the permanent magnet installed on the rotor 7.

[0063] Figure 12 This demonstrates a routine for predicting magnet temperature within the electronic control unit 20 during vehicle operation using a trained one-dimensional convolutional neural network. Furthermore, this routine is executed using an interrupt every 1 second.

[0064] In reference Figure 12 In the initial step 40, the stator coil temperature Ts measured by temperature sensor 10, the motor speed Rm measured by speed sensor 11, the oil temperature Ti measured by temperature sensor 12, and the oil pump speed Rp measured by speed sensor 13 are obtained and stored in RAM 23. Next, in step 41, it is determined whether a fixed time X1 has elapsed since the start of the magnet temperature estimation routine, i.e., whether the number of measurements required for calculating the moving average has been obtained. In the case of calculating the moving average based on four measurements, this fixed time X1 is set to four seconds.

[0065] In step 41, if it is determined that no fixed time X1 has elapsed since the start of the routine for estimating the magnet temperature, the processing loop ends. Conversely, if in step 41 it is determined that a fixed time X1 has elapsed since the start of the routine for estimating the magnet temperature, the process proceeds to step 42 to calculate the moving average, and then proceeds to step 43 to store the calculated moving average in RAM 23. Furthermore, if a moving average obtained through a simple moving average (SMA) is used as the moving average, the process is calculated in step 42... Figure 11A The moving average is shown at time t4. Additionally, in the subsequent interrupt routine 1 second later, the value at... Figure 11A The moving average shown at time t5.

[0066] On the other hand, when using a moving average obtained through exponentially smoothed moving average (EMA) as the moving average, in step 42, the calculation is performed on... Figure 11B The moving average is shown at time t4. Additionally, in the interrupt routine after 1 second, the moving average is calculated at... Figure 11B The moving average shown at time t5 is then calculated and compared with the value at time t5. Figure 11B The moving average plot shown at time t5 is the same moving average.

[0067] Next, in step 44, it is determined whether a fixed time X2 has elapsed since the start of the magnet temperature prediction routine, i.e., whether it is a time when the magnet temperature can be predicted using a one-dimensional convolutional neural network. In step 44, if it is determined that no fixed time X2 has elapsed since the start of the magnet temperature prediction routine, the processing loop ends. Conversely, if it is determined that a fixed time X2 has elapsed since the start of the magnet temperature prediction routine, step 45 is initiated, where the learned one-dimensional convolutional neural network is used to calculate the predicted magnet temperature value. That is, the calculated moving average is input into the learned one-dimensional convolutional neural network. When the calculated moving average is input into the learned one-dimensional convolutional neural network, the predicted magnet temperature value is output from the learned one-dimensional convolutional neural network (step 46).

[0068] Furthermore, during the learning of weights in a one-dimensional convolutional neural network, Figure 10B In the process, for the moving averages ma4..., mb4..., mc4..., md4... after time t4, filter-based convolution is performed while the filter is moved little by little, thereby initially calculating the output values ​​(z) from time t7 to time t9. (1)7 z (1)8 z (1)9 ...(z) (20)7 z (20)8 z (20)9Even when using a trained one-dimensional convolutional neural network to infer the temperature of a magnet, in the case of... Figure 10B In the process, for the moving averages ma4..., mb4..., mc4..., md4... after time t4, filter-based convolution is performed while the filter is moved little by little, thereby initially calculating the output value (z) from time t7 to time t9. (1)7 z (1)8 z (1)9 ...(z) (20)7 z (20)8 z (20)9 Therefore, in this case, the magnet temperature can be estimated at time t9, and thus the fixed time X2 in step 44 becomes 9 seconds. If the magnet temperature estimation process begins, the estimated magnet temperature value is then output every second.

[0069] Thus, the magnet temperature estimation device involved in this invention is as follows: Figure 13 The invention, as shown in the structural diagram, comprises: a parameter acquisition unit 50, which acquires parameters related to the rotation of the target motor 2 measured at fixed intervals; a calculation unit 51, which calculates the moving average value of these parameters for each fixed interval; a temperature acquisition unit 52, which, by inputting the moving average value calculated by the calculation unit 51 into a learning model that outputs the temperature of the magnet on the rotor 7 of the motor 2 when the moving average value of the parameters related to the rotation of the motor 2 is input, acquires a predicted value of the magnet temperature output from the learning model; and an output unit 53, which outputs the predicted value of the magnet temperature acquired by the temperature acquisition unit 52.

[0070] In this case, in the embodiment of the present invention, the learning model described above is composed of a one-dimensional convolutional neural network. Furthermore, in the embodiment of the present invention, the parameters related to the rotation of the motor 2 include the stator coil temperature Ts, the motor speed Rm, the oil temperature Ti, and the oil pump speed Rp. Additionally, in the embodiment of the present invention, a moving average obtained by exponential smoothing moving average (EMA) can be used as the moving average. In this case, the calculation unit 51 calculates the exponential smoothing moving average as the moving average.

[0071] Furthermore, it was clarified that the temperature of the permanent magnet installed on the rotor 7 is also related to the torque of the motor 2, the motor current flowing through the stator coils of the motor 2, and the inverter frequency associated with the rotational speed of the motor 2. The time-varying parameters of the motor 2's torque, motor current, and inverter frequency also affect the temperature of the permanent magnet installed on the rotor 7. Therefore, the parameters related to the rotation of the motor 2 include not only the stator coil temperature Ts, motor speed Rm, oil temperature Ti, and oil pump speed Rp, but also the motor torque, motor current, and inverter frequency.

[0072] Figure 14 This illustrates the relationship between the estimated error (estimated value - actual value) of the magnet temperature and the frequency distribution when driving a vehicle while randomly varying its driving state. From... Figure 14 It can be seen that when a moving average was applied to each measured value (after moving average processing), compared to applying a moving average to each measured value as shown below, the result is significantly better. Figures 3 to 4B As shown, without moving average processing (before moving average processing), the estimation error of magnet temperature is smaller.

[0073] In addition, Figure 12 In the illustrated routine for estimating magnet temperature, as described above, if the magnet temperature estimation process begins, then the estimated magnet temperature value is calculated every second, and the estimated magnet temperature value is output every second. However, when the vehicle's drive load is low, the rate of change of magnet temperature slows down, reducing the necessity to obtain the estimated magnet temperature value at shorter intervals. Therefore, when the vehicle's drive load is low, it is preferable to reduce the frequency of the calculation process for estimating the magnet temperature value to reduce power consumption. Figure 15 as well as Figure 16 Other embodiments are shown, illustrating the frequency of computational processing to reduce the estimated value of magnet temperature when the vehicle's drive load is low.

[0074] In this other embodiment, such as Figure 15 As shown, when the vehicle's drive load is less than the set load PK, the calculation time interval Δt for the estimated magnet temperature increases. In this case, Figure 15 In the example shown, when the vehicle's drive load is less than the set load PK, the calculation time interval Δt for the predicted magnet temperature increases. As the vehicle's drive load further decreases, the calculation time interval Δt for the predicted magnet temperature further increases. Specifically, when the vehicle's drive load is greater than the set load PK, the calculation time interval Δt for the predicted magnet temperature becomes 1 second; when the vehicle's drive load is less than the set load PK, the calculation time interval Δt for the predicted magnet temperature becomes 5 seconds; and as the vehicle's drive load further decreases, the calculation time interval Δt for the predicted magnet temperature becomes 10 seconds.

[0075] Figure 16 A predictive routine for determining the magnet temperature is shown for performing this other embodiment. Furthermore, this routine is also executed via a 1-second interrupt. Figure 16 Steps 60 to 64 of the routine for estimating the magnet temperature shown are... Figure 12 Steps 40 to 44 of the routine for estimating the magnet temperature shown are the same.

[0076] That is, in reference Figure 16 In the initial step 60, the stator coil temperature Ts measured by temperature sensor 10, the motor speed Rm measured by speed sensor 11, the oil temperature Ti measured by temperature sensor 12, and the oil pump speed Rp measured by speed sensor 13 are obtained and stored in RAM 23. Next, in step 61, it is determined whether a fixed time X1, for example 4 seconds, has elapsed since the start of the routine for estimating the magnet temperature.

[0077] In step 61, if it is determined that no fixed time X1 has elapsed since the start of the routine for estimating the magnet temperature, the processing loop ends. Conversely, in step 61, if it is determined that a fixed time X1 has elapsed since the start of the routine for estimating the magnet temperature, the process proceeds to step 62, where a moving average is calculated, and then proceeds to step 63, where the calculated moving average is stored in RAM 23. Furthermore, if a moving average obtained through a simple moving average (SMA) is used as the moving average, the process is calculated in step 62... Figure 11A The moving average shown at time t4 is calculated in the next interrupt routine 1 second later. Figure 11A The moving average shown at time t5. On the other hand, if the moving average obtained by exponential smoothing moving average (EMA) is used as the moving average, the moving average is calculated in step 62 at time t5. Figure 11B The moving average shown at time t4 is calculated in the interrupt routine 1 second later. Figure 11B The moving average shown at time t5.

[0078] Next, in step 64, it is determined whether a fixed time X2, for example 9 seconds, has elapsed since the start of the magnet temperature prediction routine. If, in step 64, it is determined that no fixed time X2 has elapsed since the start of the magnet temperature prediction routine, the processing loop ends. Conversely, if it is determined that a fixed time X2 has elapsed since the start of the magnet temperature prediction routine, the process proceeds to step 65, where it is determined whether the vehicle's drive load is greater than a set load PK. If it is determined that the vehicle's drive load is greater than the set load PK, the process jumps to step 68, where the learned one-dimensional convolutional neural network is used to calculate the predicted magnet temperature value. That is, the calculated moving average is input into the learned one-dimensional convolutional neural network. When the calculated moving average is input into the learned one-dimensional convolutional neural network, the predicted magnet temperature value is output from the learned one-dimensional convolutional neural network (step 69). At this time, the process continues... Figure 12 Similarly, the magnet temperature estimation routine shown outputs the estimated magnet temperature value every second after the magnet temperature estimation process begins.

[0079] On the other hand, in step 65, if it is determined that the vehicle's drive load is not greater than the set load PK, the process proceeds to step 66, according to... Figure 15 The relationship shown is used to calculate the estimated magnet temperature value corresponding to the vehicle's drive load over a time interval Δt. Next, in step 67, it is determined whether the estimated magnet temperature value calculation time interval Δt has elapsed. In step 67, if it is determined that the estimated magnet temperature value calculation time interval Δt has not elapsed, the processing loop ends. Alternatively, in step 67, if it is determined that the estimated magnet temperature value calculation time interval Δt has elapsed, the process proceeds to step 68, where the learned one-dimensional convolutional neural network is used to calculate the estimated magnet temperature value. Therefore, at this time, the estimated magnet temperature value is calculated over a time interval Δt corresponding to the vehicle's drive load, for example, every 5 seconds or every 10 seconds, and the estimated magnet temperature value is output every 5 seconds or every 10 seconds.

[0080] That is, in this other embodiment, the temperature acquisition unit 52 ( Figure 13 The interval for obtaining the estimated magnet temperature varies depending on the vehicle's drive load, making the interval for obtaining the estimated magnet temperature longer when the vehicle's drive load is low than when the vehicle's drive load is high.

[0081] Figure 17A as well as Figure 17B This illustrates drive current control of motor 2 using a predicted value of magnet temperature. In this drive current control of motor 2, as... Figure 17A As shown, the higher the magnet temperature, the higher the maximum value I of the drive current of motor 2.max The lower it is. Figure 17B This diagram illustrates the drive current control routine for the motor 2, performed within the electronic control unit 20 during vehicle operation. Furthermore, this routine is executed via interruptions at fixed intervals. (Refer to...) Figure 17B In this process, the required drive load of the vehicle is initially calculated in step 70. Next, in step 71, the drive current I of motor 2 is calculated based on the required drive load of the vehicle. Then, in step 72, the drive current I of motor 2 is calculated based on the estimated value of the magnet temperature. Figure 17A The relationship shown is used to calculate the maximum value I of the drive current of motor 2. max Next, in step 73, it is determined whether the calculated drive current I of motor 2 is greater than the maximum value I. max The calculated drive current I of motor 2 is greater than the maximum value I. max When proceeding to step 74, the drive current I of motor 2 is set to its maximum value I. max .

[0082] Figure 18A as well as Figure 18B The diagram illustrates the drive control of the oil pump 3 using a predicted value of the magnet temperature. In this case, an oil pump driven by an oil pump drive motor is used as the oil pump 3, such as... Figure 18A As shown, the higher the magnet temperature, the higher the drive current I of the oil pump drive motor. p The larger it increases, the greater the driving current I of the oil pump drive motor. p When the size increases, the cooling effect of the oil in the oil cooler 4 is enhanced, so the temperature of the oil decreases, and the temperature of the magnet decreases as well. Figure 18B The diagram shows the drive control routine for the oil pump 3, performed within the electronic control unit 20 during vehicle operation. Furthermore, this routine is executed via interruptions at fixed intervals. (See reference...) Figure 18B At that time, firstly in step 80 according to Figure 18A The relationship shown is based on the estimated value of the magnet temperature to determine the drive current I of the computer oil pump drive motor. p Next, in step 81, the drive current of the oil pump drive motor is made to be the calculated drive current I. p The oil pump 3 is driven and controlled in this manner.

Claims

1. A magnet temperature estimation device, comprising: The parameter acquisition unit acquires parameters related to the rotation of the target motor, including stator coil temperature, motor speed, oil temperature, and oil pump speed, measured at fixed intervals. The calculation unit calculates the moving average of the parameter for each fixed interval; The temperature acquisition unit inputs the moving average value calculated by the calculation unit into a learning model that has been learned to output the temperature of the magnet installed on the rotor of the target motor in a way that outputs the moving average value of the parameter being input, and obtains the estimated value of the magnet temperature output from the learning model. as well as The output unit outputs the estimated value of the magnet temperature obtained by the temperature acquisition unit. The learning model is a one-dimensional convolutional neural network. It performs convolution processing based on the moving average values ​​of stator coil temperature, motor speed, oil temperature, and oil pump speed, while moving a filter of the same size as the number of these parameters little by little. The output value is the sum of the product of the values ​​of each parameter and the filter value. The one-dimensional convolutional neural network is then used to learn.

2. The magnet temperature estimation device according to claim 1, wherein, In addition to stator coil temperature, motor speed, oil temperature, and oil pump speed, the parameters also include motor torque, motor current, and inverter frequency.

3. The magnet temperature estimation device according to claim 1, wherein, The calculation unit calculates an exponentially smoothed moving average as the moving average.

4. The magnet temperature estimation device according to claim 1, wherein, The interval for obtaining the estimated value of the magnet temperature obtained by the temperature acquisition unit varies according to the vehicle's drive load, so that the interval for obtaining the estimated value of the magnet temperature when the vehicle's drive load is low is longer than the interval for obtaining the estimated value of the magnet temperature when the vehicle's drive load is high.

Images