control device

Abstract

The present application provides a control device, more appropriately reduces noise of a sensor signal, and suppresses performance reduction of control based on the sensor signal. The control device as one example of the present disclosure has: a noise reduction processing section that acquires a sensor signal containing noise, and reduces the noise contained in the sensor signal based on a recurrent neural network, wherein the sensor signal is a signal based on an output from a sensor that detects time series data, and the recurrent neural network is trained in a manner of learning a correspondence between a first signal and a second signal, wherein the first signal is a signal corresponding to the sensor signal containing the noise, and the second signal represents the first signal from which the noise is removed; and a control processing section that controls an actuator based on an output from the noise reduction processing section.

Description

Technical Field

[0001] This disclosure relates to control devices. Background Technology

[0002] Previously, techniques were known for reducing high-frequency noise in sensor signals by using low-pass filters when performing control based on sensor signals, wherein the sensor signals are based on the output of a sensor that detects time-series data.

[0003] Patent Document 1: Japanese Patent Application Publication No. 2019-135120

[0004] However, in conventional techniques as described above, low-frequency noise cannot be completely reduced, and phase delay occurs in the high-frequency band. In such cases, the performance of control based on sensor signals is degraded. Summary of the Invention

[0005] Therefore, one of the objectives of this disclosure is to provide a control device that can more appropriately reduce the noise of sensor signals and suppress the performance degradation of control based on sensor signals.

[0006] As an example of the control device disclosed herein, it includes: a noise reduction processing unit that acquires a sensor signal containing noise and reduces the noise contained in the sensor signal based on a recurrent neural network, wherein the sensor signal is a signal output from a sensor that detects time-series data, and the recurrent neural network is trained in a manner that learns a correspondence between a first signal and a second signal, wherein the first signal is a noisy signal corresponding to the sensor signal, and the second signal is a signal representing the first signal after the noise has been removed; and a control processing unit that controls an actuator based on the output from the noise reduction processing unit.

[0007] According to the control device described above, based on a properly trained recurrent neural network, the noise of the sensor signal can be reduced more effectively, and the performance degradation of the control based on the sensor signal can be suppressed.

[0008] In the aforementioned control device, the sensor includes a state quantity sensor that detects time-series data related to the vehicle's state quantities. With this structure, it is possible to suppress performance degradation in control systems that utilize time-series data related to the vehicle's state quantities.

[0009] In this configuration, the state quantity sensor includes a displacement sensor that detects time-series data relating to the vehicle's vertical displacement, which serves as a state quantity for the vehicle. The actuator includes a suspension actuator that controls the vehicle's suspension. With this configuration, performance degradation of the suspension control utilizing time-series data relating to the vehicle's vertical displacement can be suppressed.

[0010] Furthermore, the aforementioned control device also includes a signal processing unit disposed between the sensor and the noise reduction processing unit. This signal processing unit performs signal processing on the output from the sensor, while the noise reduction processing unit acquires the output from the signal processing unit as a sensor signal and at least reduces the noise contained in the sensor signal generated during the sensor's detection of time-series data, as well as the noise generated during signal processing. With this structure, it is possible to at least reduce the noise generated during the sensor's detection of time-series data and the noise generated during signal processing by the signal processing unit, improve the accuracy of the signal processing result, and suppress the performance degradation of control based on the signal processing result.

[0011] In this case, the signal processing unit performs differential or integral processing as signal processing. This structure improves the accuracy of the results of differential or integral processing and suppresses performance degradation in control based on the signal processing results.

[0012] Furthermore, the aforementioned control device also includes a signal processing unit disposed between the noise reduction processing unit and the control processing unit. This signal processing unit performs signal processing on the noise-reduced sensor signal output from the noise reduction processing unit. The noise reduction processing unit acquires the output from the sensor as a sensor signal and at least reduces the noise contained in the sensor signal generated when the sensor detects time-series data. The control processing unit controls the actuator based on the output from the signal processing unit corresponding to the output from the noise reduction processing unit. With this structure, by reducing the noise generated at least when the sensor detects time-series data, the accuracy of the signal processing result can be improved, and performance degradation of the control based on the signal processing result can be suppressed.

[0013] Furthermore, in the aforementioned control device, the recurrent neural network is constructed from a Seq2Seq (Sequence to Sequence) model based on LSTM (Long Short-Term Memory). This structure allows the recurrent neural network to be configured in a form suitable for reducing noise in time series data. Attached Figure Description

[0014] Figure 1 This is an exemplary and schematic block diagram illustrating the structure of the control device in an embodiment.

[0015] Figure 2 This is an exemplary and schematic block diagram illustrating the structure of the noise reduction network in the implementation method.

[0016] Figure 3 This is an exemplary and schematic block diagram illustrating an example of the structure of the encoder section of the noise reduction network in an embodiment.

[0017] Figure 4 This is an exemplary and schematic block diagram illustrating an example of the structure of the decoder section of the noise reduction network in an embodiment.

[0018] Figure 5 This is an exemplary and schematic diagram illustrating one example of the noise reduction results of a technique based on a comparative example.

[0019] Figure 6 This is an exemplary and schematic diagram illustrating an example of the noise reduction result of a technique based on an implementation method.

[0020] Figure 7 This is an exemplary and schematic block diagram illustrating the structure of a control device as a variation of an embodiment.

[0021] Explanation of reference numerals in the attached figures

[0022] 30… Sensors (state sensors, displacement sensors), 50… Actuators (suspension actuators), 100, 700… Control devices, 110… Signal processing unit, 120, 720… Noise reduction processing unit, 121, 721… Noise reduction network, 130… Control processing unit. Detailed Implementation

[0023] Hereinafter, embodiments and modifications of the present disclosure will be described based on the accompanying drawings. The structures of the embodiments and modifications described below, as well as the effects and functions brought about by such structures, are merely examples and are not limited to the following description.

[0024] <Implementation Method>

[0025] Figure 1 This is an exemplary and schematic block diagram illustrating the structure of the control device 100 in an embodiment.

[0026] like Figure 1 As shown, the control device 100 in this embodiment is configured to control the actuator 50 based on the output from the sensor 30. The control device 100 is configured as, for example, a microcomputer equipped with hardware resources such as a processor and memory.

[0027] Furthermore, as an example, the following description will illustrate an example where the sensor 30 and actuator 50 are mounted in the vehicle, i.e., the control device 100 is mounted in the vehicle's microcomputer, i.e., the ECU (Electronic Control Unit). However, the technology of this embodiment can also be applied to general control beyond vehicle control.

[0028] Furthermore, as an example, the following description will focus on an example where sensor 30 is configured as a state quantity sensor that detects time-series data related to the state quantity of the vehicle, and more specifically, a displacement sensor (vehicle height sensor) that detects time-series data related to the vertical displacement of the vehicle, and actuator 50 is configured as a suspension actuator that controls the vehicle's suspension. However, in implementations, the combination of sensor 30 and actuator 50 can be any way, as long as sensor 30 and actuator 50 correspond to each other.

[0029] In the past, it was known that when performing control based on sensor signals, a technique was used to reduce high-frequency noise contained in the sensor signals by using a low-pass filter, wherein the sensor signals were the output signals of a sensor 30 as described above, which was based on the detection time series data.

[0030] However, in the aforementioned conventional techniques, there are instances where low-frequency noise cannot be completely reduced, and phase delay occurs in the high-frequency band (examples are described later). Figure 5 In this case, the performance of control based on sensor signals may be reduced.

[0031] Therefore, by configuring the control device 100 in such a way, the implementation achieves a more appropriate reduction in the noise of the sensor signal and suppresses the performance degradation of the control based on the sensor signal.

[0032] More specifically, the control device 100 includes a signal processing unit 110, a noise reduction processing unit 120, and a control processing unit 130. For example, it can read a computer program stored in memory and execute the result of the computer program as a processor of the control device 100, which is configured as a microcomputer; that is, these structures are implemented through a combination of hardware and software. However, in some embodiments, at least some or all of these structures may be implemented using only hardware such as dedicated circuitry.

[0033] The signal processing unit 110 performs signal processing on the output from the sensor 30. The signal processing is, for example, differential processing. Thus, when the sensor 30 is configured as a vehicle height sensor, the signal processing unit 110 can differentiate the detection result of the vehicle height sensor and output a sensor signal representing the rate of change of the suspension travel.

[0034] The sensor signal output from the signal processing unit 110 may contain various types of noise, which can be classified as EMI (Electromagnetic Interference) and EMS (Electromagnetic Susceptibility). For example, the sensor signal may contain noise generated by electromagnetic or mechanical factors due to external or internal environmental influences when the sensor 30 detects time-series data, as well as noise generated due to signal processing by the signal processing unit 110.

[0035] The noise reduction processing unit 120 reduces the noise contained in the sensor signal. More specifically, the noise reduction processing unit 120 uses a noise reduction network 121 to reduce the noise contained in the sensor signal output from the signal processing unit 110.

[0036] As follows Figures 2-4 As shown, the noise reduction network 121 is configured as a recurrent neural network (RNN), which is trained in a manner that learns the correspondence between a first signal containing noise corresponding to a sensor signal and a second signal representing the first signal after the noise has been removed.

[0037] Figure 2 This is an exemplary and schematic block diagram illustrating the structure of the noise reduction network 121 in the embodiment.

[0038] like Figure 2 As shown, the noise reduction network 121 in this embodiment is configured as a Seq2Seq (Sequence to Sequence) model. More specifically, the noise reduction network 121 includes: an encoder unit 121a, which accepts input of a sensor signal containing noise and performs encoding processing; and a decoder unit 121b, which performs decoding processing based on the encoder result of the encoder unit 121a and outputs a sensor signal without noise.

[0039] For example, in Figure 2 In the example shown, as input to encoder unit 121a, data x1, x2, x3, x4, x5, ... represented by black dots are shown drawn at predetermined time intervals. Although the black dots representing these data x1, x2, x3, x4, x5, ... are drawn approximately along a sinusoidal dotted line L200, they are not drawn to be completely aligned with the dotted line L200, thus corresponding to a sensor signal containing noise.

[0040] On the other hand, Figure 2In the example shown, as the output from the decoder unit 121b, data y1, y2, y3, y4, y5, ... represented by black dots are shown, drawn at predetermined time intervals. Since the black dots representing these data y1, y2, y3, y4, y5, ... are drawn in accordance with the dotted line L200 in the shape of a sine wave, they correspond to a sensor signal that does not contain noise.

[0041] The noise reduction network 121 of the implementation is pre-trained by machine learning to accept input time series data representing the output of the sensor signal containing noise at each moment, such as the data x1, x2, x3, x4, x5, ... described above, and output time series data representing the output of the sensor signal without noise at each moment, such as the data y1, y2, y3, y4, y5, ... described above.

[0042] In addition, although Figure 2 The example shown illustrates a structure where the sensor signal, free of noise, has a sinusoidal waveform; however, this is merely an example. In implementations, structures using sensor signals with waveforms other than sine waves, rather than those with sinusoidal waveforms, are also considered.

[0043] The following is for reference Figure 3 as well as Figure 4 The specific structures of the encoder section 121a and decoder section 121b of the noise reduction network 121 in the embodiment will be described below. However, the following... Figure 3 as well as Figure 4 The structure shown is just one example.

[0044] Figure 3 This is an exemplary and schematic block diagram showing an example of the structure of the encoder section 121a of the noise reduction network 121 in the embodiment.

[0045] like Figure 3 As shown, the encoder unit 121a of the embodiment is constructed based on LSTM (Long Short-Term Memory). That is, multiple (e.g., N) LSTM modules B are used. 11 B 12 ...B 1N This constitutes the encoder section 121a. Each LSTM module B 11 B 12 ...B 1N The structure is a general structure with input gates, output gates, and forget gates.

[0046] LSTM Module B 11It accepts input data x1 and passes data h1, representing the corresponding output, and data c1, representing the storage unit, to the next LSTM module B. 12 LSTM module B 12 Subsequent modules operate in the same manner, with the Nth LSTM module B... 1N Accept data x N The input, and the data h representing the corresponding output. N , and represent the data c of the storage unit N It is handed over to the outside of the encoder section 121a (i.e., the decoder section 121b).

[0047] Figure 4 This is an exemplary and schematic block diagram illustrating an example of the structure of the decoder section 121b of the noise reduction network 121 in an embodiment.

[0048] like Figure 4 As shown, the decoder unit 121b in this embodiment is the same as the encoder unit 121a described above, and is also based on LSTM. That is, the decoder unit 121b also uses multiple (e.g., N) LSTM modules B. 21 B 22 ...B 2N To construct. Each LSTM module B 21 B 22 ...B 2N The structure is the same as the LSTM module B mentioned above. 11 B 12 ...B 1N It is the same as the general structure that has an input gate, an output gate, and a forget gate.

[0049] However, the decoder unit 121b differs from the encoder unit 121a described above. The decoder unit 121b includes: a connection to the LSTM module B 21 B 22 ...B 2N The corresponding multiple (e.g., N) transformation layers L1, L2, ... L N These transition layers L1, L2, ... L N The symbols representing the source from LSTM module B will be displayed respectively. 21 B 22 ...B 2N The output data H1, H2, ..., H N Converted into data y1, y2, ... y corresponding to the sensor signals N .

[0050] LSTM Module B 21 Receive data h from encoder section 121a N and c NThe input is passed to the transformation layer L1, and the data H1 representing the corresponding output is passed to the storage unit, along with the data C1 representing the storage unit, to the next LSTM module B. 22 LSTM modules B22 and later operate in the same way. The Nth LSTM module B... 2N The data h representing the output corresponding to the input. N Handover to conversion layer L N .

[0051] In this way, the noise reduction network 121 of the implementation is constructed based on a recurrent neural network, which is constructed by an LSTM-based Seq2Seq model. More specifically, in the implementation, the LSTM module B is executed in a manner that learns the correspondence between a first signal corresponding to a noisy sensor signal and a second signal representing the first signal after the noise has been removed, using a recurrent neural network as described above. 11 B 12 ...B 1N LSTM module B 21 B 22 ...B 2N Transition layers L1, L2, ... L N The training results, with the set weights and biases, constitute the noise reduction network 121.

[0052] Furthermore, in one implementation, for example, a second signal for teacher data used for learning is generated by performing a filtering process based on a forward-backward filtering method on the first signal. According to this method, a second signal that reduces noise and avoids phase delay can be generated from the first signal.

[0053] Return to Figure 1 The control processing unit 130 calculates a command value for controlling the actuator 50 based on the output from the noise reduction processing unit 120. More specifically, when the actuator 50 is configured as a suspension actuator, the control processing unit 130 can calculate the command value based on the sensor signal output from the noise reduction processing unit 120, which suppresses phase delay and reduces noise, wherein the command value is used to achieve a high level of both vehicle ride comfort and driving stability.

[0054] Here, the noise reduction effect (results) of the technology based on the implementation method will be briefly explained while comparing it with a comparative example. As an example, the following explanation will be based on the premise that the object to which the noise reduction technology is applied is a sensor signal representing the rate of change of suspension travel.

[0055] Figure 5This is an exemplary and schematic diagram illustrating an example of the noise reduction result of the technique based on the comparative example. In the technique of the comparative example, noise (in the high-frequency band) is reduced from the sensor signal by a simple low-pass filter, using the same concept as the prior art described above.

[0056] Figure 5 (A) shows an example of a low-frequency waveform in a sensor signal whose noise has been reduced using the technique described in the comparative example. Figure 5 In the example shown in (A), the dashed line L510 corresponds to the measured value of the rate of change of suspension travel, and the solid line L511 corresponds to the estimated value of the rate of change of suspension travel obtained by differentiating the detection result of the vehicle height sensor. Comparing the dashed line L510 and the solid line L511, it can be seen that in the technology of the comparative example, the noise of the estimated value is not sufficiently reduced.

[0057] in addition, Figure 5 (B) shows an example of a high-frequency waveform of a sensor signal whose noise has been reduced using the technique described in the comparative example. Figure 5 In the example shown in (B), the dashed line L520 corresponds to the measured value of the rate of change of suspension travel, and the solid line L521 corresponds to the estimated value of the rate of change of suspension travel obtained by differentiating the detection result of the vehicle height sensor. Comparing the dashed line L520 and the solid line L521, it can be seen that in the comparative example's technology, although the noise of the estimated value is reduced, the phase of the estimated value is delayed relative to the phase of the measured value.

[0058] In the comparative example technology, since low-frequency noise cannot be completely reduced and phase delay occurs in the high-frequency band, it can be said that using the estimated values ​​for control is prone to adverse effects. However, since it is difficult to actually mount the sensors used to obtain the measured values ​​on the vehicle, it is desirable to improve the accuracy of the estimated values ​​obtained from the detection results of the vehicle height sensor.

[0059] Therefore, according to the technology of the implementation method, as follows: Figure 6 As shown, this can improve the accuracy of the estimated values ​​obtained from the detection results based on the vehicle height sensor.

[0060] Figure 6 This is an exemplary and schematic diagram illustrating an example of the noise reduction result of the technology based on the implementation method. In the technology of the implementation method, as described above, noise is reduced from the sensor signal using a noise reduction network 121 that has been pre-trained by machine learning.

[0061] Figure 6 (A) shows an example of a waveform in the low-frequency band of a sensor signal whose noise has been reduced by the techniques employed in this embodiment. Figure 6In the example shown in (A), the dashed line L610 corresponds to the measured value of the rate of change of suspension travel, and the solid line L611 corresponds to the estimated value of the rate of change of suspension travel obtained by differentiating the detection result of the vehicle height sensor. Comparing the dashed line L610 and the solid line L611, it can be seen that in the technology of the implementation, since the estimated value is approximately consistent with the measured value, the noise of the estimated value is sufficiently reduced.

[0062] in addition, Figure 6 (B) shows an example of a high-frequency waveform of a sensor signal whose noise has been reduced by the techniques employed in this implementation. Figure 6 In the example shown in (B), the dashed line L620 corresponds to the measured value of the rate of change of suspension travel, and the solid line L621 corresponds to the estimated value of the rate of change of suspension travel obtained by differentiating the detection result of the vehicle height sensor. Comparing the dashed line L620 and the solid line L621, it can be seen that in the technology of this implementation, since the estimated value is approximately consistent with the measured value, the noise of the estimated value is sufficiently reduced, and the aforementioned issues are not observed. Figure 5 The phase delay is as shown in (B).

[0063] In this way, according to the technology of the embodiment, unlike the technology of the comparative example, any noise in both the low-frequency and high-frequency bands can be sufficiently reduced without phase delay. Therefore, according to the technology of the embodiment, since the accuracy of the estimated value obtained by the vehicle height sensor can be improved, the performance degradation caused by noise can be suppressed.

[0064] As described above, the control device 100 of the embodiment includes a noise reduction processing unit 120 and a control processing unit 130.

[0065] The noise reduction processing unit 120 acquires a noisy sensor signal and reduces the noise in the sensor signal based on the noise reduction network 121. The sensor signal is the output signal from the sensor 30 that detects time-series data. The noise reduction network 121 is a recurrent neural network trained to learn the correspondence between a first signal and a second signal. The first signal is the noisy signal corresponding to the sensor signal, and the second signal represents the first signal after the noise has been removed. The control processing unit 130 controls the actuator 50 based on the output from the noise reduction processing unit 120.

[0066] Based on the structure described above, and using a properly trained noise reduction network 121, the noise in the sensor signal can be reduced more effectively (see reference). Figure 6 This helps to suppress the performance degradation of sensor-based control.

[0067] In this embodiment, sensor 30 includes a state quantity sensor that detects time-series data related to the state quantity of the vehicle. With this structure, performance degradation of control utilizing time-series data related to the vehicle's state quantity can be suppressed.

[0068] More specifically, in this embodiment, the aforementioned state quantity sensor includes a displacement sensor (vehicle height sensor) that detects time-series data related to the vertical displacement of the vehicle, which serves as a state quantity of the vehicle. The actuator 50 includes a suspension actuator that controls the vehicle's suspension. With this structure, it is possible to suppress performance degradation of suspension control utilizing time-series data related to the vertical displacement of the vehicle.

[0069] Furthermore, the control device 100 in this embodiment also includes a signal processing unit 110. The signal processing unit 110 is disposed between the sensor 30 and the noise reduction processing unit 120, and performs signal processing on the output from the sensor 30. The noise reduction processing unit 120 acquires the output from the signal processing unit 110 as a sensor signal, and at least reduces the noise contained in the sensor signal generated when the sensor 20 detects time-series data, as well as the noise generated by the signal processing of the signal processing unit 110. With this structure, it is possible to at least reduce the noise generated when the sensor 20 detects time-series data and the noise generated by the signal processing of the signal processing unit 110, improve the accuracy of the signal processing result, and suppress the performance degradation of control based on the signal processing result.

[0070] Furthermore, in this embodiment, the signal processing unit 110 performs differential processing as signal processing. With this structure, the accuracy of the differential processing result can be improved, and performance degradation of control based on the signal processing result can be suppressed.

[0071] Furthermore, in this implementation, the noise reduction network 121 is constructed using an LSTM-based Seq2Seq model. Based on this structure, the noise reduction network 121 can be configured in a form suitable for reducing noise in time series data.

[0072] <Variation Example>

[0073] Furthermore, in the above embodiments, a structure is illustrated where sensor 30 is a vehicle height sensor and actuator 50 is a suspension actuator. However, the technology of this disclosure can be applied to any structure as long as it utilizes the detection results of sensors to control the actuator. That is, the technology of this disclosure can of course be applied to structures that utilize the detection results of vehicle-mounted sensors other than vehicle height sensors to control vehicle-mounted actuators other than suspension actuators, and it can also be applied to structures that utilize the detection results of general sensors from fields other than vehicles to control general actuators.

[0074] Furthermore, the above embodiments primarily illustrate structures for reducing noise generated by differential processing. However, the technology of this disclosure can also be applied to structures for reducing noise generated by integral processing. Therefore, the technology of this disclosure can, for example, be applied to structures for reducing noise generated by the conversion processing of various state quantities accompanying differential or integral processing, such as the conversion processing between displacement, velocity, acceleration, and jerk, and the conversion processing between angle, angular velocity, angular acceleration, and angular jerk.

[0075] Furthermore, in the above embodiment, a recurrent neural network for noise reduction is constructed using an LSTM-based Seq2Seq model. However, as a variation, it is also considered to construct the recurrent neural network for noise reduction using GRU (Gated Reccurent Unit) or Bi-directional RNN. Additionally, as a variation, it is also considered to further implement the structure of the above embodiment by adding attention functionality, multi-layering, bi-directionality, and adding Skip Connections.

[0076] Furthermore, in the above embodiment, a structure is shown that performs noise reduction after signal processing. However, as a variation, the following is also considered. Figure 7 The structure shown performs signal processing after noise reduction.

[0077] Figure 7 This is an exemplary and schematic block diagram illustrating the structure of the control device 700, which is a modified example of the implementation.

[0078] like Figure 7 As shown, in this modified example, the control device 700 includes a noise reduction processing unit 720 between the sensor 30 and the signal processing unit 110. This noise reduction processing unit 720 uses a noise reduction network 721 to reduce the noise of the sensor signal output from the sensor 30. In this modified example, the signal processing unit 110 performs signal processing on the output from the noise reduction processing unit 720, and the control processing unit 130 controls the actuator 50 based on the output from the signal processing unit 110.

[0079] That is, in Figure 7In the illustrated variation, the signal processing unit 110 is disposed between the noise reduction processing unit 720 and the control processing unit 130, and performs signal processing on the noise-reduced sensor signal output from the noise reduction processing unit 720. The noise reduction processing unit 720 acquires the output from the sensor 30 as a sensor signal and at least reduces the noise contained in the sensor signal generated when the sensor 30 detects time-series data. The control processing unit 130 controls the actuator 50 based on the output from the signal processing unit 110 corresponding to the output from the noise reduction processing unit 720.

[0080] In addition, Figure 7 In the variant shown, the noise reduction network 721 is the same as the noise reduction network 121 in the above-described embodiment, and is configured as a recurrent neural network. The noise reduction network 721 is pre-trained based on appropriately prepared teacher data, etc., to reduce the noise generated when the sensor 30 detects time-series data, that is, to appropriately reduce the noise contained in the output from the sensor 30.

[0081] according to Figure 7 The modified example shown can improve the accuracy of the signal processing results by at least reducing the noise generated when the sensor 30 detects time series data, and suppress the performance degradation of the control based on the signal processing results.

[0082] The embodiments and modifications of this disclosure have been described above, but these embodiments and modifications are merely examples and are not intended to limit the scope of the invention. The new embodiments and modifications described above can be implemented in various ways, and various omissions, substitutions, and changes can be made without departing from the spirit of the invention. The embodiments and modifications described above are included within the scope and spirit of the invention, and are included within the scope of the invention as described in the claims and its equivalents.

Claims

1. A control device comprising: The noise reduction processing unit acquires a sensor signal containing noise and reduces the noise contained in the sensor signal based on a recurrent neural network, wherein... The aforementioned sensor signal is based on the output signal from a sensor detecting time-series data. The aforementioned recurrent neural network is trained in a manner that learns the correspondence between a first signal and a second signal, wherein the first signal is a signal containing the aforementioned noise corresponding to the aforementioned sensor signal, and the aforementioned second signal represents the aforementioned first signal after the aforementioned noise has been removed; and The control processing unit controls the actuator based on the output from the aforementioned noise reduction processing unit. The phase delay of the estimated second signal suppression value of the first signal after the aforementioned noise has been removed, relative to the measured phase value, represents the phase delay of the second signal suppression estimate. The aforementioned actuator includes a suspension actuator that controls the vehicle's suspension.

2. The control device according to claim 1, wherein, The aforementioned sensors include state quantity sensors, which detect the aforementioned time-series data related to the state quantity of the vehicle.

3. The control device according to claim 2, wherein, The aforementioned state quantity sensor includes a displacement sensor, which detects the aforementioned time series data related to the vertical displacement of the vehicle, which is the aforementioned state quantity of the vehicle.

4. The control device according to any one of claims 1 to 3, wherein, It also includes a signal processing unit, which is disposed between the sensor and the noise reduction processing unit, and performs signal processing on the output from the sensor. The noise reduction processing unit acquires the output from the signal processing unit as the sensor signal, and at least reduces the noise contained in the sensor signal generated when the sensor detects the time series data and the noise generated by the signal processing unit.

5. The control device according to claim 4, wherein, The aforementioned signal processing unit performs differential or integral processing as part of the aforementioned signal processing.

6. The control device according to any one of claims 1 to 3, wherein, It also includes a signal processing unit, which is disposed between the noise reduction processing unit and the control processing unit, and performs signal processing on the sensor signal output from the noise reduction processing unit after the noise has been reduced. The noise reduction processing unit acquires the output from the sensor as the sensor signal, and at least reduces the noise contained in the sensor signal generated when the sensor detects the time series data. The control processing unit controls the actuator based on the output from the signal processing unit, which corresponds to the output from the noise reduction processing unit.

7. The control device according to any one of claims 1 to 3, wherein, The aforementioned recurrent neural network is composed of a Seq2Seq (Sequence to Sequence) model based on LSTM (Long Short-Term Memory).

8. The control device according to claim 4, wherein, The aforementioned recurrent neural network is composed of a Seq2Seq (Sequence to Sequence) model based on LSTM (Long Short-Term Memory).

9. The control device according to claim 5, wherein, The aforementioned recurrent neural network is composed of a Seq2Seq (Sequence to Sequence) model based on LSTM (Long Short-Term Memory).

10. The control device according to claim 6, wherein, The aforementioned recurrent neural network is composed of a Seq2Seq (Sequence to Sequence) model based on LSTM (Long Short-Term Memory).

Images