Gait training system, control method, and control program

Abstract

To provide a gait training system, a control method, and a control program that enable effective training to be performed.SOLUTION: A gait training system 1 includes: a load distribution sensor having a plurality of sensors arranged in a two-dimensional array for detecting the distribution of a load applied from a user; and an inference unit having a CNN for executing convolution arithmetic processing with two-dimensional data based on the output of the sensor as an input to estimate a value of the load received from the user.SELECTED DRAWING: Figure 6

Description

[Technical field]

[0001] The present invention relates to a walking training system, a control method, and a control program. [Background technology]

[0002] Patent Document 1 discloses a pressure distribution detection device. This pressure distribution detection device includes a plurality of drive coils and a plurality of detection coils. It also includes spacer means for maintaining a constant distance between the drive coils and the detection coils. The drive coil is provided with a cushioning material and an electromagnetic shielding body. The detection coil is provided with a cushioning material and an electromagnetic coupling variable body. The electromagnetic coupling variable body is a sheet-shaped magnetic material sheet. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2007-218892 A Summary of the Invention [Problem to be solved by the invention]

[0004] When detecting pressure distribution, pressure (load) is detected from the amount of deformation (deformation stress) of the viscoelastic body. If the viscoelastic body is formed into a sheet, the viscoelastic body will also deform in the surrounding area of ​​the point where the load is actually applied. Since the load is also detected in the surrounding area where no load is applied, it is difficult to detect the load with high accuracy.

[0005] In addition, in a walking training device using a treadmill or the like, the timing in the walking cycle is estimated according to the load distribution. Then, the drive control is performed according to the timing in the walking cycle. Therefore, by detecting the load distribution with high accuracy, more appropriate drive control can be performed.

[0006] The present invention has been made in consideration of the above background, and has an object to provide a walking training system, a control method, and a control program that can appropriately calculate a load value. [Means for solving the problem]

[0007] The walking training system according to this embodiment includes a load distribution sensor having a plurality of sensors arranged in a two-dimensional array to detect the distribution of the load applied by the user, and an inference device having a CNN that performs convolution processing using two-dimensional data based on the sensor output of the sensor as input, and that estimates the load value applied by the user.

[0008] The walking training system may further include a robot leg attached to the user's leg, and a drive unit that drives the robot leg based on the load value.

[0009] The walking training system may further include a treadmill having a belt on which the user stands, and the sensor may detect a load applied via the belt.

[0010] In the above walking training system, the inference unit may be a machine learning model that receives an input of the moving speed of the belt.

[0011] The walking training system may be configured to trim sensor outputs from all sensors included in the load distribution sensor to generate the two-dimensional data.

[0012] The walking training system may further include a pressure conversion unit that converts the sensor output into a pressure value, and the inference unit may receive as input two-dimensional data made up of the pressure values.

[0013] In the above walking training system, the inference device may be constructed by supervised machine learning using a load value detected by a load cell of an insole sensor provided in the sole of a shoe as training data.

[0014] In the walking training system, the load distribution sensor may include a viscoelastic sheet, and the load may be detected according to an amount of deformation of the viscoelastic sheet.

[0015] The control method of the walking training system in this embodiment is provided to detect the distribution of the load applied by the user, and acquires sensor output from a load distribution sensor having a plurality of sensors arranged in a two-dimensional array, and inputs two-dimensional data based on the sensor output to an inference device having a CNN that performs convolution calculation processing, thereby estimating the load value applied by the user.

[0016] In the above-mentioned method for controlling a walking training system, the walking training system may include a robot leg attached to the leg of the user, and the robot leg may be controlled based on the load value.

[0017] In the above-described method for controlling a walking training system, the walking training system may further include a treadmill having a belt on which the user stands, and the sensor may detect a load applied via the belt.

[0018] In the above-mentioned method for controlling a walking training system, the inference unit may be a machine learning model that receives an input of the moving speed of the belt.

[0019] In the above-mentioned method for controlling a walking training system, in order to generate the two-dimensional data, sensor outputs from all sensors included in the load distribution sensor may be trimmed.

[0020] In the above-mentioned method for controlling a walking training system, the sensor output may be converted into a pressure value, and the inference unit may receive as input two-dimensional data constituted by the pressure values.

[0021] In the above-mentioned control method for a walking training system, the inference device may be constructed by supervised machine learning using a load value detected by a load cell of an insole sensor provided in the sole of a shoe as teaching data.

[0022] In the above-described control method for a walking training system, the load distribution sensor may include a viscoelastic sheet, and the load may be detected according to an amount of deformation of the viscoelastic sheet.

[0023] The program of this embodiment is a control program for causing a computer to execute a control method for controlling a walking training system, the control method being for detecting the distribution of load applied by a user, acquiring sensor output from a load distribution sensor having a plurality of sensors arranged in a two-dimensional array, inputting two-dimensional data based on the sensor output to an inference device having a CNN that performs convolution processing, and estimating the load value applied by the user. Effect of the Invention

[0024] According to the present disclosure, it is possible to provide a walking training system, a control method, and a control program capable of appropriately calculating a load value. [Brief description of the drawings]

[0025] [Figure 1] FIG. 1 is a schematic perspective view of a walking training system. [Diagram 2] FIG. 2 is a side cross-sectional view showing the configuration of the treadmill. [Diagram 3] FIG. 2 is a top view showing a schematic configuration of a load distribution sensor and a belt. [Figure 4] 5A to 5C are cross-sectional views illustrating the deformation of a viscoelastic sheet in the load distribution sensor. [Diagram 5] This is a two-dimensional map showing the output values ​​of the sensor. [Figure 6] FIG. 2 is a functional block diagram showing the main configuration of a control system of the system. [Figure 7] FIG. 2 is a diagram for explaining a learning model used in the inference device. [Figure 8] FIG. 2 is a diagram showing a schematic arrangement of a load cell which is an insole sensor. [Figure 9]1 is a graph showing the change over time in load value during normal walking. [Figure 10] 13 is a graph showing the change over time in load value during heel walking. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0026] The present invention will be described below through embodiments of the invention, but the invention according to the claims is not limited to the following embodiments. In addition, not all of the configurations described in the embodiments are necessarily essential as means for solving the problems. For clarity of explanation, the following description and drawings are omitted and simplified as appropriate. In each drawing, the same elements are given the same reference numerals, and repeated explanations are omitted as necessary.

[0027] (System Configuration) 1 is an overall conceptual diagram showing an example of the configuration of a rehabilitation support system according to an embodiment of the present invention. The rehabilitation support system (walking training system 1) according to this embodiment is mainly composed of a walking training device 100 and a leg brace 120.

[0028] The walking training device 100 is a specific example of a rehabilitation support device that supports the rehabilitation of a trainee (user) 900. The walking training device 100 is a device that allows the trainee 900, who is a hemiplegic patient suffering from paralysis in one leg, to perform walking training under the guidance of a training staff member 901. Here, the training staff member 901 can be a therapist (physiotherapist) or a doctor, and can also be called a training instructor, training assistant, training assistant, etc., since the training staff member supports the trainee's training by providing guidance or assistance.

[0029] The walking training device 100 mainly comprises a control panel 133 attached to a frame 130 forming the overall skeleton, and a treadmill 131 on which the trainee 900 walks. Furthermore, a leg brace 120 is attached to the affected leg of the trainee 900, which is the paralyzed leg. In FIG. 1, the leg brace 120 is attached to the right leg of the trainee 900. The leg brace 120 is a robotic leg equipped with an actuator and the like. For example, the leg brace 120 has a joint drive mechanism that assists the movement of the knee joint of the trainee 900.

[0030] The frame 130 is erected on a treadmill 131 that is installed on the floor. The treadmill 131 rotates a ring-shaped belt 132 by a motor (not shown). The treadmill 131 is a device that encourages the trainee 900 to walk, and the trainee 900 who is undergoing walking training gets on the belt 132 and tries to walk in accordance with the movement of the belt 132. Note that the training staff 901 can stand on the belt 132 behind the trainee 900 and walk together as shown in FIG. 1, for example, but it is usually preferable for the training staff 901 to be in a position that makes it easy to assist the trainee 900, such as standing astride the belt 132. The treadmill 131 has a load distribution sensor, as will be described later.

[0031] The frame 130 supports a control panel 133 and a training monitor 138. The control panel 133 houses an overall control unit 210 that controls the motors and sensors. The training monitor 138 is, for example, a liquid crystal panel, and displays the progress of training and the like to the trainee 900. The frame 130 also supports a front pulling unit 135 near the front of the upper part of the trainee 900's head, a harness pulling unit 112 near the upper part of the head, and a rear pulling unit 137 near the rear of the upper part of the head. The frame 130 also includes a handrail 130a for the trainee 900 to grasp.

[0032] The handrails 130a are arranged on both the left and right sides of the trainee 900. Each handrail 130a is arranged in a direction parallel to the walking direction of the trainee 900. The handrails 130a are adjustable in vertical and horizontal positions. In other words, the handrails 130a can include a mechanism for changing their height and width. Furthermore, the handrails 130a can be configured so that their inclination angle can be changed, for example, by adjusting the height to be different between the front side and the rear side in the walking direction. For example, the handrails 130a can be given an inclination angle that gradually increases along the walking direction.

[0033] The handrail 130a is also provided with a handrail sensor 218 that detects the load received from the trainee 900. For example, the handrail sensor 218 can be a resistance change detection type load detection sheet in which electrodes are arranged in a matrix. The handrail sensor 218 can also be a six-axis sensor that combines a three-axis acceleration sensor (x, y, z) and a three-axis gyro sensor (roll, pitch, yaw). However, the type and installation position of the handrail sensor 218 are not important.

[0034] The camera 140 functions as an imaging unit for observing the entire body of the trainee 900. The camera 140 is installed near the training monitor 138 so as to face the trainee. The camera 140 takes still images and videos of the trainee 900 during training. The camera 140 includes a lens and imaging element set that provides an angle of view sufficient to capture the entire body of the trainee 900. The imaging element is, for example, a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor, and converts an optical image formed on an imaging surface into an image signal.

[0035] The coordinated action of the front pulling section 135 and the rear pulling section 137 offsets the load of the leg brace 120 so that it does not become a burden on the affected leg, and further assists the swinging motion of the affected leg according to the set degree.

[0036] One end of the front wire 134 is connected to a winding mechanism of the front pulling unit 135, and the other end is connected to the leg brace 120. The winding mechanism of the front pulling unit 135 turns on / off a motor (not shown) to wind or unwind the front wire 134 in response to the movement of the affected leg. Similarly, one end of the rear wire 136 is connected to a winding mechanism of the rear pulling unit 137, and the other end is connected to the leg brace 120. The winding mechanism of the rear pulling unit 137 turns on / off a motor (not shown) to wind or unwind the rear wire 136 in response to the movement of the affected leg. Such coordinated operation of the front pulling unit 135 and the rear pulling unit 137 offsets the load of the leg brace 120 so that the load does not become a burden on the affected leg, and further assists the swinging motion of the affected leg according to the set degree.

[0037] The front wire 134 and the front pulling unit 135 constitute a first pulling means for pulling the legs of the trainee 900 upward and forward. The rear wire 136 and the rear pulling unit 137 constitute a second pulling means for pulling the legs of the trainee 900 upward and backward. The front pulling unit 135 and the rear pulling unit 137 pull the front wire 134 and the rear wire 136 with a pulling force according to the walking phase, as described below. Also, a pulling force operation pattern may be set according to the walking phase.

[0038] For example, the training staff 901, as an operator, sets a high assist level for a trainee with severe paralysis. When the assist level is set high, the front pulling unit 135 winds up the front wire 134 with a relatively large force in accordance with the timing of the swing of the affected leg. As the training progresses and assistance is no longer necessary, the training staff 901 sets the assist level to the minimum. When the assist level is set to the minimum, the front pulling unit 135 winds up the front wire 134 with just enough force to cancel the weight of the leg orthosis 120 in accordance with the timing of the swing of the affected leg.

[0039] The walking training device 100 is equipped with a fall prevention harness device as a safety device, the main components of which are an orthosis 110, a harness wire 111, and a harness tensioning part 112. The orthosis 110 is a belt that is wrapped around the abdomen of the trainee 900 and is fixed to the waist by, for example, a hook and loop fastener. The orthosis 110 is equipped with a connecting hook 110a that connects one end of the harness wire 111, which is a hanging device, and can also be called a hanger belt. The trainee 900 wears the orthosis 110 so that the connecting hook 110a is located at the rear of the trainee.

[0040] One end of the harness wire 111 is connected to a connecting hook 110a of the harness 110, and the other end is connected to a winding mechanism of the harness tensioning unit 112. The winding mechanism of the harness tensioning unit 112 winds and unwinds the harness wire 111 by turning on and off a motor (not shown). With this configuration, when the trainee 900 is about to fall, the fall prevention harness device winds up the harness wire 111 according to an instruction from the overall control unit 210 that detects the movement, and the harness 110 supports the upper body of the trainee 900, preventing the trainee 900 from falling.

[0041] The orthosis 110 includes a posture sensor 217 for detecting the posture of the trainee 900. The posture sensor 217 is, for example, a combination of a gyro sensor and an acceleration sensor, and outputs the inclination angle of the abdomen on which the orthosis 110 is worn with respect to the direction of gravity.

[0042] The management monitor 139 is attached to the frame 130, and is a display input device mainly used for monitoring and operation by the training staff 901. The management monitor 139 is, for example, a liquid crystal panel, and a touch panel is provided on its surface. The management monitor 139 displays various menu items related to training settings, various parameter values ​​during training, training results, etc. In addition, an emergency stop button 232 is provided near the management monitor 139. When the training staff 901 presses the emergency stop button 232, the walking training device 100 is brought to an emergency stop.

[0043] The overall control unit 210 generates rehabilitation data that may include setting parameters related to training settings, various data related to leg movements output from the leg brace 120 as training results, etc. This rehabilitation data may include data indicating the training staff 901 or their years of experience and level of proficiency, data indicating the symptoms, walking ability, recovery level, etc. of the trainee 900, various data output from sensors provided outside the leg brace 120, etc.

[0044] At least a part of the control of the overall control unit 210 may be executed by a computer program. For example, the overall control unit 210 includes a memory for storing the program, a processor for executing the program, and the like.

[0045] The overall control unit 210 can control actuators such as motors based on the detection results of the load distribution sensor. For example, the overall control unit 210 determines the phase (timing) of a walking cycle based on the change in load detected by the load distribution sensor. The overall control unit 210 generates a control signal for controlling the actuator based on the determination result. The joint drive mechanism of the leg brace 120 performs an assisting operation of the knee joint based on the control signal. In this way, an appropriate assisting operation can be performed, and effective training can be achieved. The overall control unit 210 may also control the front tension unit 135 and the rear tension unit 137 according to the phase of the walking cycle. The control based on the detection results of the load distribution sensor will be described later.

[0046] Next, the configuration of the load distribution sensor provided on the treadmill 131 will be described with reference to Figures 2 and 3. Figure 2 is a side cross-sectional view showing the configuration of the treadmill 131 and the load distribution sensor 152. Figure 3 is a top view showing a schematic arrangement of the load distribution sensor 152 on the treadmill 131.

[0047] The treadmill 131 includes at least a ring-shaped belt (moving plate) 132, a pulley 153, and a motor (not shown). The belt 132 is an endless belt and is formed in a ring shape in a side view. Two pulleys 153 are arranged inside the ring-shaped belt 132. The two pulleys 153 are arranged at a distance in the front-rear direction. The rotation axis of the pulley 153 is parallel to the left-right direction. Therefore, when the motor rotates the pulley 153, the belt 132 is driven in the front-rear direction. When the trainee 900 or the training staff 901 sets the moving speed, the motor rotates the pulley 153 so that the belt 132 moves at the moving speed.

[0048] Furthermore, a load distribution sensor 152 is arranged on the inside of the ring-shaped belt 132 (below the belt 132 on which the trainee 900 rides) so as not to be linked to the belt 132. The load distribution sensor 152 is arranged on the inside of the ring of the endless belt. The load distribution sensor 152 is composed of a plurality of sensors 155. The plurality of sensors 155 are arranged in a matrix on the underside of the belt 132 that supports the soles of the feet of the trainee 900 during walking training. By using the plurality of sensors 155, the load distribution sensor 152 can detect the distribution of surface pressure (load distribution) received from the feet FT of the trainee 900 riding on the belt 132.

[0049] 3, the load distribution sensor 152 has a plurality of sensors 155 arranged in a matrix. The load distribution sensor 152 has a sheet shape parallel to a horizontal plane. The sensor 155 may be of an electromagnetic induction type, an electrical resistance type, a capacitance type, or a load cell type. Of course, the type of the sensor 155 is not particularly limited.

[0050] For example, the load distribution sensor 152 is assumed to have m (m is an integer of 2 or more) sensors 155 arranged in the left-right direction and n (n is an integer of 2 or more) sensors 155 arranged in the front-rear direction. In Fig. 3, the XY coordinates of the sensor 155 at the left front corner are (1,1), and the XY coordinates of the sensor 155 at the left rear corner are (1,n). The XY coordinates of the sensor 155 at the right front corner are (m,1), and the XY coordinates of the sensor 155 at the right rear corner are (m,n). Each sensor 155 has the same size. In this manner, an XY address (XY coordinates) is assigned to each sensor 155.

[0051] The sensors 155 are arranged at equal intervals in the front-rear and left-right directions. The sensing area of ​​the load distribution sensor 152 is therefore rectangular. The output from the load distribution sensor 152 is therefore m×n pieces of two-dimensional map data.

[0052] One sensor 155 is also referred to as a cell. The load distribution sensor 152 detects pressure (load) for each cell. The load distribution sensor 152 detects the distribution of the load received from the trainee 900 by sequentially reading out the outputs of the multiple sensors 155. Each sensor 155 detects a two-dimensional distribution of the pressure received from the foot FT via the belt 132. The m×n pieces of two-dimensional data are also referred to as one frame of data.

[0053] The overall control unit 210 estimates a load value from the foot FT based on the detection result of the load distribution sensor 152. The load value here indicates the total load received from the entire foot FT. For example, the output values ​​detected by the multiple sensors 155 can be converted into pressure using a conversion map. Then, the pressure can be converted into a load value by integrating it. The overall control unit 210 makes a decision regarding the walking phase based on the load value. For example, the overall control unit 210 decides that it is the timing to transition from the stance leg to the swing leg.

[0054] Load distribution sensor 152 is a sheet-like sensor and has a viscoelastic body. For example, as shown in Fig. 4, load distribution sensor 152 has substrate 1511, viscoelastic sheet 1512, and conductive sheet 1513. Note that Fig. 4 is a side cross-sectional view that typically shows the cross-sectional configuration of load distribution sensor 152. Here, load distribution sensor 152 will be described as an electromagnetic induction type sensor. Also, treadmill 131, belt 132, etc. are omitted in Fig. 4.

[0055] Load distribution sensor 152 includes substrate 1511, viscoelastic sheet 1512, and conductive sheet 1513. Substrate 1511, viscoelastic sheet 1512, and conductive sheet 1513 are arranged in this order from the bottom side. Viscoelastic sheet 1512 is provided between conductive sheet 1513 and substrate 1511 of load distribution sensor 152. Viscoelastic sheet 1512 is made of a resin material such as urethane foam. The thickness direction of viscoelastic sheet 1512 is the up-down direction.

[0056] The substrate 1511 has a coil pattern having a plurality of coils. The coil pattern has a coil provided for each cell. The plurality of coils are arranged in an array on the substrate 1511. The conductive sheet 1513 is, for example, an aluminum sheet, and is formed over the entire load distribution sensor 152 so as to cover the plurality of coils. The distance between the conductive sheet 1513 and the substrate 1511 changes depending on the load. In other words, the greater the load, the more the viscoelastic sheet 1512 is crushed and becomes thinner. When the viscoelastic sheet 1512 deforms, the electrical characteristics of the substrate 1511 change.

[0057] The load distribution sensor 152 can determine the amount of deformation by measuring electrical characteristics such as electromagnetic induction for each cell. The load distribution sensor 152 converts the amount of deformation of the viscoelastic sheet 1512 in the thickness direction into a load. The amount of deformation can be converted into a load for each cell. In this way, the electrical characteristics change according to the amount of deformation of the viscoelastic sheet 1512, so that each sensor 155 can detect the load. Of course, the load distribution sensor 152 is not limited to the electromagnetic induction type, and a load distribution sensor of a capacitance type, an electrical resistance type, or the like may also be used.

[0058] Hereinafter, the area of ​​the viscoelastic sheet 1512 where the load is actually applied is referred to as the application area A3. The application area A3 is the area directly above which the foot FT is placed. The application area A3 corresponds to the size of the foot FT. The area around the application area A3 is referred to as the peripheral area A2. The area outside the peripheral area A2 is referred to as the outer area A1. The outer area A1 is an area where the viscoelastic sheet 1512 is not deformed. Each of the application area A3, the outer area A1, and the peripheral area A2 includes multiple sensors 155.

[0059] When a load is applied to the application area A3, the viscoelastic sheet 1512 also deforms in the peripheral area A2. That is, the distance between the substrate 1511 and the conductive sheet 1513 is smaller in the peripheral area A2 than in the outer area A1. Although the peripheral area A2 is not receiving a load, the viscoelastic sheet 1512 deforms. In this way, a load is also detected in the peripheral area A2 of the application area A3, which is receiving an actual load (hereinafter, this is also referred to as the roll-in phenomenon). Furthermore, in the peripheral area A2, the closer it is to the application area A3, the greater the amount of deformation.

[0060] Fig. 5 is a map that shows the sensor output values ​​of the sensors 155. In Fig. 5, the sensor output of each sensor 155 is shown as a numerical value. In other words, the greater the amount of deformation, the greater the sensor output value. The output value of the sensor 155 is converted into, for example, an 8-bit or 16-bit digital value by an A / D (Analog to Digital) converter or the like.

[0061] In the outer area A1 where the viscoelastic sheet 1512 is not deformed, the output is 0. In the application area A3, the sensor output value changes depending on the load. In the peripheral area A2, the sensor output value changes depending on the load of the adjacent application area A3.

[0062] In this way, the sensor 155 detects pressure in an area larger than the actual ground contact size of the foot FT. The viscoelastic sheet 1512 deforms and detects pressure (load) in a number of cells greater than the number of cells actually located directly under the foot FT. In other words, it is not possible to distinguish between cells that are crushed and cells that are not crushed based on the output value of each sensor 155 alone.

[0063] Therefore, there is a risk that the load value cannot be detected with high accuracy. If the load value detection accuracy is low, there is a risk that the drive timing of the leg orthosis 120 will deviate from the timing of the walking cycle. Therefore, in this embodiment, the following process is performed to estimate the load value with high accuracy. The process according to this embodiment will be described below.

[0064] 6 is a functional block diagram for explaining the control according to this embodiment. The overall control unit 210 includes a sensor output acquisition unit 171, an inference unit 172, a determination unit 175, and a drive control unit 176. The joint drive unit 221 is provided in the leg brace 120. Specifically, the joint drive unit 221 includes a motor and a rotation mechanism for assisting the knee joint movement of the trainee 900.

[0065] The treadmill driving unit 131a drives the treadmill 131. The treadmill driving unit 131a has a motor or the like for rotating the pulley 153. The treadmill 131 operates so that the belt 132 moves at a moving speed set by the training staff 901 or the like.

[0066] The treadmill 131 is provided with a load distribution sensor 152. The load distribution sensor 152 outputs a detection signal (sensor output) indicating the detection result to the overall control unit 210. The sensor output may be digital data or analog data. In other words, the load distribution sensor 152 may have an A / D converter that converts the analog detection signal into digital data. Alternatively, the sensor output acquisition unit 171 may have the A / D converter.

[0067] As shown in Fig. 5, the sensor output includes an output value for each cell. The sensor output acquisition unit 171 acquires the sensor output from the load distribution sensor 152 for each cell. That is, data indicating the sensor output is associated with a cell address and an output value. The sensor output acquisition unit 171 may acquire the sensor output data in the read order of the load distribution sensor 152. The sensor outputs from all the cells constitute one frame of data. The load distribution sensor 152 reads out data for each line or each cell.

[0068] The inference unit 172 infers the load value based on the sensor output. The inference unit 172 infers the load value using a learning model 178 generated by machine learning. The learning model 178 receives output values ​​of cells arranged in an array as input. Specifically, the inference unit 172 inputs two-dimensional data as shown in FIG. 5 to the learning model 178. The inference unit 172 calculates a load value 173 indicating the load received from the foot FT, and outputs the load value 173 to the determination unit 175.

[0069] The determination unit 175 performs a determination regarding the phase of the walking cycle based on the estimated load value 173. The determination unit 175 compares the load value 173 with a preset threshold value. Then, the overall control unit 210 determines the timing when the load value 173 falls below the threshold value as the switching timing for transitioning from the standing leg to the swing leg. The determination unit 175 performs the determination in order to detect the timing for removing weight from the affected leg.

[0070] The drive control unit 176 controls the joint driving unit 221 according to the determination result by the determination unit 175. For example, the drive control unit 176 outputs a control signal to the joint driving unit 221 so that the joint driving unit 221 operates at the switching timing. Then, an actuator of the joint driving unit 221 operates according to the control signal. The joint driving unit 221 can bend the knee joint at the switching timing, that is, at the timing of lifting weight from the affected leg.

[0071] In this way, the joint driving unit 221 can apply an assist force in accordance with the walking phase. Therefore, the trainee 900 can perform effective walking training. Note that the control target controlled in accordance with the judgment result is not limited to the joint driving unit 221. For example, the drive control unit 176 may control the front pulling unit 135 and the rear pulling unit 137 so that the pulling forces of the front wires 134 and the rear wires 136 operate in accordance with the walking phase. This allows the front pulling unit 135 and the rear pulling unit 137 to assist in swinging forward at the timing of unloading.

[0072] The following describes the learning model 178 used by the inference unit 172. Fig. 7 is a conceptual diagram for explaining an example of the learning model 178. The learning model 178 is an inference model generated by machine learning.

[0073] The learning model 178 is a DL (Deep Learning) model having multiple layers. As described above, the input data of the learning model 178 is two-dimensional data. Therefore, the input data is shown as a two-dimensional grayscale image.

[0074] The learning model 178 includes a convolutional neural network (CNN) that performs convolution processing on two-dimensional data. The CNN includes a convolution layer, a pooling layer, and the like. A fully connected network (FCN) including a fully connected layer is provided downstream of the CNN. Of course, the learning model 178 may use other networks.

[0075] CNN performs convolution processing on two-dimensional data including the pressure value of each cell. The convolutional neural network CNN has a weight called a two-dimensional kernel, which allows weighting processing that takes adjacent cells into account. The convolutional neural network extracts features according to the two-dimensional data. This improves the estimation accuracy of the load value 173. The weights of the CNN can be calculated by machine learning.

[0076] In this way, the load distribution sensor 152 has a plurality of sensors 155 arranged in a two-dimensional array, and detects the load distribution applied by the trainee 900. The inference unit 172 receives two-dimensional data based on the sensor output of the sensor 155. The inference unit 172 has a CNN that performs convolution calculation processing, and estimates the load value received from the user. Since the output value of each sensor 155 is multi-bit data, the input data of the inference unit 172 becomes two-dimensional data similar to image data. In other words, the sensor output becomes similar to image data in which the output of one sensor 155 becomes pixel data for one pixel. As with image recognition and other processing, the CNN can perform convolution processing. Since a CNN for image processing can be used, versatility can be improved.

[0077] This allows the inference unit 172 to calculate the load value with high accuracy. The inference unit 172 does not need to distinguish between the actually crushed applied area A3 and the surrounding area A2 that is deformed due to the current entrapment state. This allows the walking phase to be properly analyzed, enabling effective training.

[0078] In this embodiment, the load distribution sensor 152 is disposed under the belt 132. The sensor 155 detects the load (pressure) applied through the belt 132. Even in such a case, the inference unit 172 can estimate the load value with high accuracy. That is, each parameter such as the weight of the CNN is learned assuming the case where the belt 132 moves. That is, the weighting parameters are optimized in the state where the belt 132 is moving. This can improve the estimation accuracy of the load value 173. By using a machine learning model having a CNN as the inference unit 172, the load value can be calculated with higher estimation accuracy. Therefore, the overall control unit 210 can appropriately analyze the walking phase, and the trainee 900 can perform effective training.

[0079] Furthermore, based on the load value estimated by the inference unit 172, the drive control unit 176 controls the leg orthosis 120, which is the robot leg. For example, the drive control unit 176 outputs a control signal for controlling a joint drive unit 221 for a knee joint or the like. The drive control unit 176 can drive the leg orthosis 120 in accordance with the walking phase. Since the leg orthosis 120 can assist the walking motion with an appropriate assist force, the trainee 900 can perform effective training.

[0080] The degree of the entrapment phenomenon may change depending on the moving speed of the belt 132. Therefore, the moving speed of the belt 132 may be input to the inference device 172. That is, the two-dimensional data and the moving speed become input data of the learning model 178. For example, the learning model 178 may use different parameters depending on the moving speed of the belt 132. Specifically, the activation function, the filter, the weight, etc. may be switched. Alternatively, the inference device 172 may switch between the learning model 178 for high speed and the learning model 178 for low speed. When generating the learning model 178 by machine learning, the learning device may perform machine learning for each moving speed of the belt 132. In other words, the learning device uses the data of the moving speed as learning data by associating it with the two-dimensional data of the sensor output.

[0081] The sensor output acquisition unit 171 may trim the sensor output from all the sensors 155 included in the load distribution sensor 152. Then, the sensor output acquisition unit 171 may input the trimmed two-dimensional data to the inference unit 172. For example, the sensor output acquisition unit 171 may have a trimming size set in advance. Here, the trimming size can be (k×l) rectangular areas. k is greater than 1 and less than m, and l is greater than 1 and less than n. The sensor output acquisition unit 171 cuts out k×l two-dimensional data with the center of gravity of the load distribution of the load distribution sensor 152 as the center. That is, the sensor output becomes (k×l) two-dimensional map data. In this way, the amount of data processing can be reduced, and therefore the processing can be performed at high speed. The inference unit 172 can calculate the load value for each frame.

[0082] The sensor output acquisition unit 171 may function as a pressure conversion unit that converts the sensor output into a pressure value. In this case, the inference unit 172 uses two-dimensional map data composed of pressure values ​​as input data. Specifically, the load distribution sensor 152 A / D converts the output values ​​(voltage values) output from each sensor 155 into digital data. Then, the sensor output acquisition unit 171 converts the sensor output, which is digital data, into a pressure value using a conversion formula or a conversion table. The load distribution sensor 152 has a conversion formula or a conversion table that converts a measured value of an electrical characteristic into a pressure value.

[0083] The conversion table or the conversion formula may indicate a linear conversion result or a non-linear conversion result. For example, the conversion formula may be a quadratic or higher polynomial. The conversion formula or the conversion table for converting the sensor output value into the pressure value may be different for each sensor 155, or may be common to a plurality of sensors 155. For example, the sensor output acquisition unit 171 may use a conversion formula calibrated for each sensor 155. This can improve the estimation accuracy.

[0084] Furthermore, supervised machine learning with teacher data or semi-supervised machine learning can be used as machine learning for generating the learning model 178. In this case, the load value detected by the load cell of the insole sensor provided in the sole of the shoe can be used as the teacher data (correct answer label).

[0085] For example, as shown in FIG. 8, the trainee 900 puts on the shoes SH with the load cell 180 placed on the sole of the shoes SH. As a result, the load cell 180 is placed between the foot FT and the shoes SH. The load cell 180 is placed on the sole of the foot FT. The load cell 180 is an insole-type load sensor or load distribution sensor. The load cell 180 can directly detect the load received by the foot FT. That is, the inference device 172 is constructed by supervised machine learning using the load value detected by the load cell 180 as teaching data. This makes it possible to construct a learning model with higher estimation accuracy.

[0086] 9 and 10 are graphs showing the measurement results of the load cell 180, which is an insole sensor, and the inference results of the inference unit 172. In Figs. 9 and 10, the measurement value of the load cell 180 is shown as the sole load. Fig. 9 shows the inference results and measurement values ​​during normal walking. Fig. 10 shows the inference results and measurement values ​​during heel walking. Figs. 9 and 10 are also graphs showing the change in load value during one step.

[0087] 9 and 10, the horizontal axis indicates the number of frames, i.e., time. The vertical axis indicates the estimated value and the actual measured value of the load value [N]. The estimated value by the inference unit 172 is close to the actual measured value by the load cell 180. By using the inference unit 172, the load value can be estimated with high accuracy, and therefore the trainee 900 can perform effective training.

[0088] The overall control unit 210 is not limited to being a single physical device. In other words, the sensor output acquisition unit 171, the inference unit 172, the determination unit 175, the learning model 178, and the drive control unit 176 may be distributed and arranged in multiple devices. In this case, data may be transmitted and received between the devices via a network. For example, the learning model 178 may be mounted on a server or the like so that it can be used by multiple walking training devices 100.

[0089] In addition, a part or all of the above-mentioned control method can be realized by a computer program. The program can be stored using various types of non-transitory computer readable media and supplied to the computer. The non-transitory computer readable media includes various types of tangible storage media. The non-transitory computer readable media includes, for example, magnetic recording media, magneto-optical recording media, CD-ROM (Read Only Memory), CD-R, CD-R / W, and semiconductor memory. The magnetic recording media are, for example, flexible disks, magnetic tapes, hard disk drives, etc. The magneto-optical recording media are, for example, magneto-optical disks, etc. The semiconductor memory is, for example, mask ROM, PROM (Programmable ROM), EPROM (Erasable PROM), flash ROM, RAM (Random Access Memory), etc. In addition, the program may be supplied to the computer by various types of transitory computer readable media. Examples of the transitory computer readable media include electric signals, optical signals, and electromagnetic waves. The temporary computer-readable medium can supply the program to the computer via a wired communication path such as an electric wire or an optical fiber, or via a wireless communication path. [Explanation of symbols]

[0090] 1. Walking training system 100 Walking training device 120 Leg braces 131 Treadmill 132 Belt 152 Load distribution sensor 153 Pulley 155 Sensors 171 Sensor output acquisition unit 172 Reasoner 173 Load Value 175 Judgment section 176 Drive control unit 178 Learning Model 210 Overall Control Unit 221 Joint drive unit 900 Trainee

Claims

1. A load distribution sensor having a plurality of sensors arranged in a two-dimensional array for detecting a distribution of a load applied by a user; A walking training system comprising: a CNN that performs convolution processing using two-dimensional data based on the sensor output of the sensor as input; and an inference device that estimates the load value received from the user.

2. A robot leg attached to the user's leg; The walking training system according to claim 1 , further comprising: a drive unit that drives the robot leg based on the load value.

3. a treadmill having a belt on which a user stands; 3. The walking training system according to claim 1, wherein the sensor detects a load applied via the belt.

4. The walking training system according to claim 3 , wherein the inference device is a machine learning model that receives the moving speed of the belt as an input.

5. 3. The gait training system according to claim 1, wherein sensor outputs from all sensors included in the load distribution sensor are trimmed to generate the two-dimensional data.

6. A pressure conversion unit that converts the sensor output into a pressure value, 3. The walking training system according to claim 1, wherein the inference unit receives as input two-dimensional data composed of the pressure values.

7. 3. The walking training system according to claim 1, wherein the inference device is constructed using supervised machine learning using a load value detected by a load cell of an insole sensor attached to the sole of a shoe as training data.

8. the load distribution sensor comprises a viscoelastic sheet, The walking training system according to claim 1 or 2, wherein a load is detected according to an amount of deformation of the viscoelastic sheet.

9. acquiring a sensor output from a load distribution sensor provided for detecting a distribution of a load applied by a user and having a plurality of sensors arranged in a two-dimensional array; A control method for a walking training system in which two-dimensional data based on the sensor output is input to an inference device having a CNN that performs convolution calculation processing, and the load value received from the user is estimated.

10. The walking training system includes a robotic leg attached to a leg of the user, The method for controlling a walking training system according to claim 9 , further comprising the step of driving the robot legs based on the load values.

11. The gait training system further comprises a treadmill having a belt on which a user stands; The method for controlling a walking training system according to claim 9 or 10, wherein the sensor detects a load applied via the belt.

12. The method for controlling a walking training system according to claim 11 , wherein the inference device is a machine learning model that receives the moving speed of the belt as an input.

13. 11. The method for controlling a walking training system according to claim 9, further comprising the step of trimming sensor outputs from all sensors included in the load distribution sensor to generate the two-dimensional data.

14. converting the sensor output into a pressure value; 11. The control method for a walking training system according to claim 9, wherein the inference unit receives two-dimensional data constituted by the pressure values ​​as an input.

15. The control method for a walking training system according to claim 9 or 10, wherein the inference device is constructed using supervised machine learning using a load value detected by a load cell of an insole sensor provided in the sole of a shoe as training data.

16. the load distribution sensor comprises a viscoelastic sheet, The method for controlling a walking training system according to claim 9 or 10, wherein a load is detected according to an amount of deformation of the viscoelastic sheet.

17. A control program for causing a computer to execute a control method for controlling a walking training system, The control method includes: acquiring a sensor output from a load distribution sensor provided for detecting a distribution of a load applied by a user and having a plurality of sensors arranged in a two-dimensional array; A control program for inputting two-dimensional data based on the sensor output to an inference device having a CNN that performs convolution calculation processing, and estimating the load value received from the user.

Images