JPH07244065A - Motion evaluation method - Google Patents

Abstract

PURPOSE:To evaluate diversified motion of an object based on the increase of acceleration by inputting the increase of acceleration of the moving object detected into an input layer of a hierarchical type neural network. CONSTITUTION:When an object 0 makes any motion, a sampler 2 samples acceleration and increase of acceleration outputs of an increase of acceleration sensor 1 at each unit time. A data obtained is set with a delay element block 3 on an input layer of a neural network 4 as increase of acceleration patten sequentially. Then, the results are outputted from an output layer as such of motion evaluation after via an intermediate layer of the network 4. In the network 4, a plurality of processing units corresponding to a neurone are connected in a network and operated parallel giving or taking signals mutually to accomplish an information processing with the network as a whole. Furthermore, a desired algorithm is acquired while a learning is repeated even when the algorithm is unclear at first. This enables the evaluation of diversified motion of the object efficiently without the preparation of corresponding algorithms.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a motion evaluation method for evaluating the motion of an object by using a jerk that differentiates acceleration.

[0002]

2. Description of the Related Art As a physical quantity representing the motion of an object, position,
Speed and acceleration can be mentioned. There is a sensor that detects each physical quantity. By the way, a person visually detects a position and a speed, and what is perceptually detected as a bodily sensation is an acceleration and a physical quantity higher than it. For example, when riding in a moving body such as an automobile, a railway vehicle, an aircraft, or an elevator, the user is consciously aware of the state change at the time of acceleration or deceleration.

Development of an acceleration sensor for automobiles (Nissan Technical Report No. 1
It is known from literatures such as No. 23, Sho 62-12) that humans are sensitive to jerk higher than acceleration. This fact shows that speed is deeply involved in the feeling of comfort and discomfort felt by humans. Therefore, in order to evaluate vehicles such as automobiles, railway vehicles, aircraft, and elevators that involve motion, not only this acceleration but also the jerk information that differentiates the acceleration is detected,
It is necessary to evaluate using that information. Conversely speaking, if humans limit their movements to linear movements, acceleration,
Since it is not possible to detect physical quantities above the jerk,
If acceleration and jerk information can be detected, it can be said that the amount of information is sufficient for evaluating a vehicle.

For the above-mentioned uses, Japanese Patent Laid-Open No. 31375/1993
In Japanese Patent Publication No. 3, a jerk sensor capable of simultaneously detecting acceleration and jerk and a method of motion control / motion evaluation using the same are proposed. In this motion evaluation method, the acceleration and jerk detected at a plurality of detection points are evaluated by the magnitude of a value obtained by multiplying an arbitrary weighting constant.

[0005]

The above-mentioned prior art is effective in the sense that the strength of the vibration component is evaluated, for example, for an object that vibrates steadily. There is no guideline for the evaluation method for movements that have a weak periodicity in changes in acceleration and jerk during exercise. Further, if the object to be exercised is changed, the setting of the weighting constant and the algorithm of the evaluation function need to be changed for each object to be exercised, which is complicated. On the other hand, assuming that a person feels comfortable and uncomfortable when riding on an automobile, railway vehicle, aircraft, or elevator, acceleration, jerk is input, and results evaluated by algorithms such as sensuality and sensitivity, The result that it is comfortable or uncomfortable is output. Human evaluation has a personality, and if a human changes during a series of motion evaluations, subsequent evaluations will not be consistent, unlike previous evaluations. In such a case, the exercise evaluation may be restarted from the beginning. Even for the same person, the evaluation may change due to changes in physical condition. Therefore, if an algorithm that simulates human sensory and sensibilities can be formed in motion evaluation, motion evaluation that excludes personality is performed in the same manner as a human evaluator such as a test driver together with the jerk detection means. It should be possible. Furthermore, according to the motion evaluation method as described above, abnormal motion can be detected from the evaluation contents in the same manner as a human evaluator.

An object of the present invention is to provide a motion evaluation method for evaluating the motions of various objects based on jerk.

It is an object of the present invention to provide a motion evaluation method for evaluating the motion of an object in the same way as a human sense.

An object of the present invention is to provide a motion evaluation method for evaluating the motion of an object that changes irregularly.

An object of the present invention is to provide a motion evaluation method for evaluating the motion of an object that changes with periodicity.

An object of the present invention is to provide an abnormal motion detecting method for detecting an abnormal motion of an object.

[0011]

The above object is to detect a jerk which is obtained by differentiating at least the acceleration of a moving object, and to input the detected jerk to an input layer of a hierarchical neural network and to pass through an intermediate layer. Then, it is achieved by outputting the motion evaluation result from the output layer.

The above object is to detect the jerk which is obtained by differentiating at least the acceleration of a moving object for each unit time, delay the jerk detected for each unit time, and to sequentially input to the input layer of the hierarchical neural network. It is achieved by inputting as a jerk pattern and outputting the motion evaluation result from the output layer via the intermediate layer.

The above object is to detect the jerk which is obtained by differentiating at least the acceleration of a moving object for each unit time, frequency-analyze the detected jerk for each unit time, and to apply the frequency to the input layer of the hierarchical neural network. This is achieved by inputting as a spectrum pattern and outputting the motion evaluation result from the output layer via the intermediate layer.

The above-mentioned object is an evaluation obtained by another evaluation means for the combined weights of the input layer and the intermediate layer and the intermediate layer and the output layer of the neural network with respect to the jerk during various movements of the object. Learning the neural network by detecting the jerks of various movements of the object and inputting the detected jerks to the neural network so as to reduce the difference between the ideal output and the output from the output layer. It is achieved by

The above-mentioned object is an ideal output which is an evaluation of the combined weights of the input layer and the intermediate layer and between the intermediate layer and the output layer of the neural network by human sensitivities to the jerk during various movements of the object. Achieved by learning the neural network by detecting the jerks of various movements of an object, inputting the detected jerks to the neural network, and changing them so as to reduce the difference from the output from the output layer. To be done.

The above-mentioned object is an ideal output which is an evaluation of the combined weight of the input layer and the intermediate layer of the neural network and the combined weight of the intermediate layer and the output layer by human sensitivity to the jerk during various movements of the object. By detecting the jerks of various movements of the object, inputting the detected jerks to the neural network, and changing them so as to reduce the difference from the output from the output layer, and learning the neural network in advance. To be achieved.

An ideal output, which is an evaluation of the combined weights of the input layer and the intermediate layer and between the intermediate layer and the output layer of the neural network, which is an evaluation based on human sensitivity to the additional velocity of the motion of the object,
It is desirable to learn the neural network at any time by changing it so that the difference from the output of the neural network becomes small at the time of evaluating the motion of the object.

When the motion of the object cannot be evaluated, it is desirable to generate an abnormal motion signal as unlearned motion.

It is desirable to detect acceleration at the same time as the jerk of the moving object.

It is desirable that the jerk is detected by a jerk sensor.

[0021]

According to the above arrangement, the jerk of a moving object is detected and the neural network is used to evaluate the motion, thereby eliminating the need for various algorithms for evaluating the motion of the object.

The jerk of an object that changes irregularly is detected at arbitrary time intervals, and the detected jerks are delayed and input to the neural network in order, so that they are time-sequentially spaced at equal intervals. Acceleration patterns can be added.

A frequency spectrum pattern can be formed by detecting the jerk of an object whose movement changes with periodicity and analyzing the frequency.

By preliminarily learning the neural network using the human sense as an ideal output (teaching signal),
An algorithm simulating the human sense can be formed, and the motion of an object can be evaluated in the same way as the human sense.

By learning the neural network as needed when evaluating the motion of an object, it is possible to evaluate the motion of a new object.

When the motion of the object cannot be evaluated, an abnormal motion signal is generated as an unlearned motion, so that it is possible to detect an abnormal motion that was not assumed at the time of design.

By detecting the acceleration as well as the acceleration of the moving object, it is possible to perform a motion evaluation that cannot be achieved only by detecting the acceleration.

By detecting the jerk by the jerk sensor, the combination of the acceleration sensor and the electric circuit for differentiating the detected value can detect even the frequency region where the electric circuit has a delayed phase and does not function.

[0029]

Embodiments of the present invention will be described below with reference to the drawings.

First, the basic structure of this embodiment will be described.

FIG. 1 is an explanatory view for explaining the motion evaluation method of the embodiment of the present invention and the conventional sensory evaluation in comparison.

In the past, when performing a sensory evaluation of a vehicle such as an elevator or a vehicle, a tester had a sensory or empirical evaluation of the information actually felt in the vehicle. At this time, the amount of information that can be felt is only the acceleration and jerk applied to the tester. Therefore, if this acceleration and jerk are detected simultaneously, it is necessary and sufficient as physical information for evaluation.After that, if a means (emulator) that mimics the process of linking this physical quantity to evaluation can be realized, it will be considered as a human sensitive evaluation. A similar evaluation should be possible.

In this embodiment, a jerk sensor for detecting acceleration and jerk is used, and a neural network is used as an emulator. A neural network is a neural network in which multiple processing units corresponding to nerve cells (neurons) are connected in a network like a neural network of the brain, and many processing units are moved in parallel while exchanging signals with each other. It is a mechanism that processes information over the entire network, and is an engineering model of the information processing mechanism performed in the brains of living things such as humans. Further, the neural network spontaneously changes the strength of the coupling between neurons, that is, the characteristics of the network, and it becomes possible to gradually output desired output data from the input sample data. . That is, a desired algorithm can be acquired during repeated learning even if the initially clear algorithm is unknown.

The learning method of the neural network for algorithm acquisition in this embodiment uses supervised learning, and the input test pattern is acceleration and jerk information detected at the same time when a human sensory evaluation is performed.
As the teacher signal, a human sensory evaluation is given. As the learning is repeated, the human senses and experiences will be reflected in the neural network. Therefore, the motion evaluation method can be configured by using such a learned neural network and jerk sensor. Furthermore, by changing the exercises to be learned, it is possible to evaluate any exercise.

Next, the device configuration of this embodiment will be described.

FIG. 2 is an explanatory diagram showing the apparatus configuration of the embodiment of the present invention.

As shown in the figure, the device is a moving object 0
Is fixed to, and a jerk sensor 1 that detects the acceleration and jerk that accompany the movement of the object 0 at any time, and the acceleration output and jerk output of the jerk sensor 1 are sampled at arbitrary equal time intervals. Sampler 2 and delay element group 3 for delaying sampled acceleration output and jerk output
The neural network 4 which inputs the acceleration and jerk patterns from the delay element group 3 and outputs the motion evaluation pattern from the output layer, and each label which represents the motion evaluation index with respect to the output of the neural network 4 are stored. The evaluation label storage means 5 is used.

Next, the motion evaluation method of this embodiment will be described.

FIG. 3 is an explanatory view for explaining the functions of the sampler and the delay element group of FIG.

It is assumed that the object 0 is now moving arbitrarily and the acceleration / acceleration time history output as shown in FIG. The sampler 2 samples the acceleration and jerk output of the jerk sensor 1 every unit time from the start of exercise. The sampled acceleration and jerk data are sequentially set by the delay element group 3 in the input layer of the neural network 4 as an input acceleration and jerk pattern. In this way, time series acceleration,
The jerk information can be patterned.

FIG. 4 is an explanatory diagram for explaining a general neuron model.

FIG. 5 is an explanatory diagram for explaining an input conversion function called a general sigmoid function.

The neural network 4 is formed by connecting neuron models 400 whose concept is shown in FIG. The jth neuron model 400 is
The value obtained by multiplying the input signal Xi from the i-th neuron by the connection weight Wij corresponding to the connection strength with that neuron (Wi
The value obtained by adding the bias value θj to the value obtained by adding j · Xi) to all the neurons expressed by Equation 1 is held as the internal potential Xj.

[0044]

[Equation 1]

[0045]

[Equation 2]

Then, based on the input conversion function of FIG.
According to the above, one output signal Yj is output. The output signal Yj output from the j-th neuron model 400 is
It becomes the input Xj of the connected k-th neuron model. Alternatively, the output signal may be used as input for feedback.

FIG. 6 is an explanatory diagram for explaining the modes of the neural network according to the embodiment of the present invention.

As shown in the figure, the neural network 4 has two modes. One is a processing mode that outputs a motion evaluation pattern according to the input acceleration and jerk patterns, and the other is a learning mode that optimizes the neural network 4 itself by learning according to the algorithm shown below. is there.

FIG. 7 is an explanatory diagram for explaining the neural network learning of the embodiment of the present invention. The learning of the neural network is specifically optimization of the connection weight between the neuron models. In this embodiment, a learning method with a teacher is shown. First, as an input pattern, acceleration and jerk information generated when the object 0 moves is used as a sample pattern, and a plurality of sets of ideal evaluation output data corresponding to the sample acceleration and jerk pattern are prepared. The evaluation output may be a result evaluated by a human being actually boarding the vehicle, or when the object 0 makes the motion of the sample jerk pattern described above, it is determined by something other than jerk information. It may have been done. The sample acceleration and jerk pattern are input to the neural network 4, and the evaluation output thereof is obtained. Here, if the evaluation output data and the ideal evaluation output data of the sample acceleration and jerk pattern match each other, at that time, all the connection weights in the path transmitting the signal in the neural network 4 are calculated. Increase the value. On the contrary, if they do not match, the values of all the coupling weights in the path transmitting the signal in the neural network 4 at that time are reduced. When this is repeated using various sample acceleration and jerk data, the neural network 4 gradually becomes
The desired output can be obtained for any input.
Finally, a desired motion evaluation output can be obtained for an arbitrary acceleration and jerk pattern generated by the motion of the object 0.

When the neural network 4 is in the learned state as described above, the learning mode for learning is unnecessary and the motion can be evaluated only in the processing mode for processing the input data. Therefore, when the evaluation target and the movement are limited, the neural network 4 may be a device that has been learned and configured only in the processing mode. Also, if you want to change the evaluation target or exercise, activate the learning mode at any time even after learning is completed,
By learning new sample acceleration and jerk pattern and ideal evaluation output data, it is possible to create a network that is compatible with new movements (acceleration and jerk pattern) at any time.

FIG. 8 is an explanatory view for explaining the method of evaluating the riding comfort when the elevator is rising according to the embodiment of the present invention.

The jerk sensor 11 is attached to the car 10 of the elevator, and can detect vertical acceleration and jerk.

At the same time when the ascent starts, the sampler 12 starts sampling. The sampled acceleration and jerk information is input to the input layer 141 through the delay element group 13, and the intermediate layer 142 is subjected to the conversion process of the equations 1 and 2.
It is output from the output layer 43 via the. Reference numeral 15 is an evaluation label storage means, which stores each label representing the evaluation index of the motion for each neuron model of the output layer 143.

FIG. 9 is an explanatory view for explaining an example of evaluating the riding comfort when the elevator is rising according to the embodiment of the present invention.

As shown in the figure, four-level exercise evaluation labels of "good", "satisfactory", "unsatisfactory", and "impermissible" are attached. The rating for movement is determined by the output of the output layer 143 for each label. The size of the fill in the square in the figure reflects the output value of each labeled neuron model of the output layer 143, with 1.0 being a fill and 0.0 being a blank. Now, in case 1, the output of 1.0 is obtained for the neuron model with the label “good”, and it is close to 0.0 for the other neuron models with the label. Therefore, it is possible to evaluate that the motion of the elevator that gives an output value that is Case 1 is a “good” motion. In case 2, the output of 0.5 is obtained for the neuron model labeled as “good”, which is “satisfied”.
Even for the neuron model with the label of 0.5, 0.5 is obtained, and it can be seen that the motion is such that the evaluation is divided and the value is intermediate between these two evaluations. In case 3, the output of 0.6 is obtained for the neuron model with'unsatisfied 'label, 0.6 is output for the neuron model with'unsatisfied' label, and the neuron model with'satisfied 'label becomes On the other hand, 0.5 was obtained, and as a result, it can be seen that the evaluation of "dissatisfaction" is a valid exercise. In this way, unlike the conventional digital evaluation method, it is possible to perform an evaluation closer to the human sense by performing a fuzzy evaluation. Further, the ride comfort of the elevator can be improved by inputting these signals to the elevator motion controller.

FIG. 10 is an explanatory diagram for explaining a method for recognizing vehicle motion according to the embodiment of the present invention.

The jerk sensors 211 and 212 are used for the vehicle 2
It is installed on the floor of 0, the front-back direction (x axis),
The lateral acceleration (y-axis) and jerk can be detected. Now, at the same time when the running is started, the sampler 22 starts sampling. The sampled acceleration and jerk information is used for delay element groups 231, 232, 233, 234.
Via the input layer 241 of the neural network 24
Is output to the output layer 243 via the intermediate layer 242. Reference numeral 25 denotes an evaluation label storage unit, which stores each label for recognizing the motion state for each neuron model of the output layer 243. In this embodiment, 'straight'and'turning (cor)
nering), 'accelerate', 'decelerat'
e) ',' spinning ',' tire lock (Tyre lo
ck) 'is attached to a label for recognizing six types of motion states. Recognition of motion is determined by the output of output layer 243 for each label.

FIG. 11 is an explanatory diagram for explaining an example of recognizing vehicle motion according to the embodiment of the present invention.

Similar to FIG. 9, the size of the filled area in the figure reflects the output value of each labeled neuron model of the output layer 243, with 1.0 being filled and 0.0 being blank. is there. In case 1, the output of 1.0 is obtained for the neuron model with the label “straight ahead”, and 0.0 for the other neuron models with labels.
Is. Therefore, it is possible to recognize that the motion of the vehicle 20 that can obtain the output value that is Case 1 is the “straight ahead” motion. In case 2, the output of 0.5 is obtained for the neuron model labeled'straight ahead 'and 0.5 for the neuron model labeled'turning'. It is understood that the movement is such that the cornering has a large turning radius (close to a straight line). In case 3, a 0.5 output is obtained for the neuron model labeled'turning ', 0.6 for the neuron model labeled'acceleration', and for the neuron model labeled'deceleration '. 0.6 is obtained, and the motion of the vehicle in this case is not in a state where the acceleration and jerk are input to the neuron model, and a bias is input to the neuron model, and neurons are connected to the bias. Has a large output when is in a strong state,
When equal outputs of labels with contradictory physical quantities are obtained, it can be seen that it is appropriate to recognize that the acceleration is neither constant acceleration nor deceleration, and is a constant circular motion of constant jerk motion. In this way, unlike the conventional digital evaluation method, it is possible to perform an evaluation closer to the human sense by performing a fuzzy evaluation.

FIG. 12 is an explanatory diagram showing the acceleration and jerk when a general vehicle spin occurs. FIG. 13 is an explanatory diagram showing acceleration and jerk when a tire lock occurs in a general vehicle.

When the behavior of the vehicle 20 changes, the centrifugal force in the lateral direction of the vehicle sharply decreases during turning, and the decelerating force in the longitudinal direction of the vehicle decreases rapidly during deceleration. By learning such characteristic acceleration and jerk patterns, it is possible to easily detect and recognize changes in the behavior of the vehicle, and by inputting these signals to the vehicle motion controller, It is possible to improve exercise performance.

Next, an evaluation method for an object that is steadily vibrating will be described.

FIG. 14 is an explanatory view showing the device construction of another embodiment of the present invention.

The motion evaluation apparatus shown in this figure is provided with a jerk sensor 31 which is fixed to the object 30 and detects the acceleration and jerk which accompany the movement of the object 30 at any time, and the acceleration output of the jerk sensor 31. A sampler 32 that samples the jerk output at arbitrary equal time intervals, a frequency analysis device 33 that frequency-analyzes the output from the sampler 32, an acceleration and jerk pattern are input, and a motion evaluation pattern is output from the output layer. The neural network 34 and an evaluation label storage means 35 in which each label representing the evaluation index of the motion with respect to the output of the neural network 34 is stored.

FIG. 15 is an explanatory view for explaining the functions of the sampler and the frequency analyzer of FIG.

It is assumed that the object 30 is now making an arbitrary motion as shown in the figure, and the acceleration and jerk time history output as shown in FIG. 3 is output from the jerk sensor 31. The sampler 32 samples the acceleration and jerk output of the jerk sensor every unit time from the start of exercise. The sampled acceleration and jerk data are subjected to frequency analysis such as Fourier transform by the frequency analysis device 33 and converted into a power spectrum as shown in the figure. Then, the power spectrum for each frequency is input to the input layer of the neural network 34 as an acceleration and jerk spectrum pattern. In this way, time series acceleration,
The jerk information can be converted into a spectrum pattern. The neural network 34 has a power spectrum of acceleration and jerk information generated as the input object 30 moves as a sample pattern, and a plurality of sets of ideal evaluation output data corresponding to the sample acceleration and jerk spectrum pattern. When the neural network 34 is prepared and repeatedly learned, the neural network 34 gradually obtains a desired output with respect to an arbitrary input. And finally, as in the previously described embodiment, the object 30
The desired motion evaluation output can be obtained for any acceleration and jerk pattern generated by the motion.

The examples described above are all cases in which evaluation is possible, but cases in which evaluation is impossible are also conceivable. For example, as in case 4 shown in FIG. 9, the highest "good" and the lowest "impermissible" are simultaneously output. These cases show cases in which the motion of the object to be evaluated is unlearned. When applied to human evaluation, this is considered to be similar to the feeling when first riding on an object with a distinctly different movement pattern from the one with experience of conventional tower riding. Even in this case,
Using the learning mode of the neural network,
By learning the motion of a new object without changing the algorithm, it becomes possible to evaluate the motion of a new object in addition to the conventional evaluation target.

Further, it is also possible to detect abnormal motion by making reverse use of non-evaluation.

FIG. 16 is an explanatory diagram showing the structure of an abnormal motion detecting device according to another embodiment of the present invention.

The abnormal movement includes, for example, an abnormal vibration of an elevator caused by a child jumping and bouncing in a car, and a phenomenon that a brake does not work due to a failure. The abnormal motion detecting device is fixed to the object 40, and detects the acceleration and jerk that occur with the movement of the object 40 at any time, and the acceleration output and jerk output of the jerk sensor 41 are arbitrary. Of the sampler 42 for sampling at equal time intervals of, the delay element group 43 for delaying the output of the sampler 42, and the neural network for inputting the acceleration and jerk pattern output from the delay element group 43 and outputting the motion evaluation pattern from the output layer Network 44, evaluation pattern storage means 46 for storing an evaluation pattern used as an ideal output during learning, neural network 4
Each label representing the motion evaluation index for the output of 4 is stored, and the evaluation inability is detected from the evaluation pattern used as the ideal output at the time of learning stored in the evaluation pattern storage means 46 and the output of the neural network 44,
It is composed of an un-evaluable detection device 45 that outputs an abnormal motion signal. In this embodiment, the neural network 44 is preliminarily given various exercise modes that occur during normal operation and evaluations of the exercises to complete learning. That is, the evaluation is output as learned during normal operation. On the other hand, if an abnormal motion is input to the neural network 44, the evaluation cannot be performed. By detecting this state from the output layer 343 by the non-evaluable detection device 45 and generating an abnormal motion signal, the abnormal motion can be self-determined. Further, it is also possible to detect a change in exercise performance due to a change over time.

In the present embodiment, a jerk sensor for detecting the jerk and acceleration of an object is used, and a process of linking the sensory evaluation from the physical sensation obtained when a human gets on an object is realized by a neural network. The method of realizing the same evaluation as the human sensory evaluation and the method of detecting the abnormal motion of the object from these evaluations were also described. In the present embodiment, the jerk sensor is used to detect the acceleration and jerk of the object, but the invention is not limited to this, and various acceleration sensors and differentiating circuits that are practically used according to the purpose are used. You may use. Also, position, speed,
Image information and the like may be captured and input to a neural network to further enhance the functionality.

Further, the evaluation output by the motion evaluation method of this embodiment may be input to the motion controller to improve the motion performance. Further, the motion control controller may be configured by a neural network common to the motion evaluation device, and the learning may be performed so as to be optimum based on the evaluation by the motion evaluation method.

As described above, according to this embodiment, since the neural network is preliminarily trained by using the evaluation based on the human sense, the motion of the vehicle can be evaluated in the same manner as the human sense. It is also possible to perform evaluation in extreme conditions that cannot be performed by humans, for example, in a test flight in which G is large and severe, and it is possible to obtain a highly reproducible and consistent evaluation by eliminating personality.

[0074]

According to the present invention, the jerk of a moving object is detected and the motion is evaluated by a neural network, so that the motions of various objects can be efficiently evaluated without preparing respective algorithms. .

By preliminarily learning the neural network using the evaluation based on the human sense as an ideal output, an algorithm simulating the human sense can be formed and the motion of the object can be evaluated in the same manner as the human sense.

By detecting the jerk of a moving object, delaying it to form a jerk pattern and inputting it to a neural network, it is possible to efficiently evaluate the movement of an object that changes irregularly.

By detecting the jerk of a moving object, performing frequency analysis to form a frequency spectrum pattern and inputting it to a neural network, the motion of an object whose movement changes periodically can be evaluated efficiently. be able to.

When the motion of the object cannot be evaluated, an abnormal motion signal is generated as an unlearned motion, so that it is possible to detect an abnormal motion that was not assumed at the time of design.

By detecting the acceleration as well as the acceleration of the moving object, it is possible to perform a motion evaluation that cannot be achieved only by detecting the acceleration.

[Brief description of drawings]

FIG. 1 is an explanatory diagram illustrating a motion evaluation method according to an embodiment of the present invention and a conventional sensory evaluation in comparison.

FIG. 2 is an explanatory diagram showing a device configuration of an embodiment of the present invention.

FIG. 3 is an explanatory diagram illustrating functions of a sampler and a delay element group in FIG.

FIG. 4 is an explanatory diagram illustrating a general neuron model.

FIG. 5 is an explanatory diagram illustrating an input conversion function called a general sigmoid function.

FIG. 6 is an explanatory diagram illustrating modes of the neural network according to the embodiment of this invention.

FIG. 7 is an explanatory diagram illustrating neural network learning according to the embodiment of this invention.

FIG. 8 is an explanatory diagram illustrating a method of evaluating a riding comfort when an elevator is rising according to the embodiment of this invention.

FIG. 9 is an explanatory diagram illustrating an example of evaluating the riding comfort when the elevator is rising according to the embodiment of this invention.

FIG. 10 is an explanatory diagram illustrating a vehicle motion recognition method according to the embodiment of this invention.

FIG. 11 is an explanatory diagram illustrating an example of recognition of vehicle motion according to the embodiment of this invention.

FIG. 12 is an explanatory diagram showing acceleration and jerk when a general vehicle spin occurs.

FIG. 13 is an explanatory diagram showing acceleration and jerk when a tire lock occurs in a general vehicle.

FIG. 14 is an explanatory diagram showing a device configuration of another embodiment of the present invention.

FIG. 15 is an explanatory diagram illustrating functions of the sampler and the frequency analysis device in FIG.

FIG. 16 is an explanatory diagram showing a configuration of an abnormal motion detection device according to another embodiment of the present invention.

[Explanation of symbols]

 0 object 1 jerk sensor 2 sampler 3 delay element group 4 neural network 5 evaluation label storage means 10 elevator 11 jerk sensor 12 sampler 13 delay element group 14 neural network 15 evaluation label storage means 20 vehicle 24 neural network 25 evaluation label Storage means 30 Object 31 jerk sensor 32 sampler 33 frequency analysis device 34 neural network 35 evaluation label storage means 40 object 41 jerk sensor 42 sampler 43 delay element group 44 neural network 45 non-evaluable detection device 46 evaluation pattern storage means 141 Input layer 142 Intermediate layer 143 Output layer 211 jerk sensor 212 jerk sensor 221 sampler 222 sampler 223 sampler 224 sampler 231 slow Element group 232 delay element group 233 delay element group 234 delay element group 241 input layer 242 intermediate layer 243 outputs layer 400 neuron model 441 input layer 442 intermediate layer 443 outputs layer

Claims (10)

[Claims] 1. A motion evaluation result obtained by detecting a jerk that is obtained by differentiating at least acceleration of a moving object, inputting the detected jerk to an input layer of a hierarchical neural network, and from an output layer via an intermediate layer. Is output. 2. A jerk that is obtained by differentiating at least the acceleration of a moving object is detected for each unit time, and the jerk that is detected for each unit time is delayed to be added to an input layer of a hierarchical neural network in order. A motion evaluation method characterized by inputting as a velocity pattern and outputting a motion evaluation result from an output layer via an intermediate layer. 3. A frequency spectrum pattern to an input layer of a hierarchical neural network by detecting a jerk which is obtained by differentiating at least acceleration of a moving object for each unit time and performing frequency analysis on the detected jerk for each unit time. , And outputs the motion evaluation result from the output layer via the intermediate layer. 4. An ideal output, which is an evaluation obtained by other evaluation means for the combined weights of the input layer and the intermediate layer and the intermediate layer and the output layer of the neural network with respect to the jerk during various movements of the object. And to detect the jerks of various movements of the object, input the detected jerks to the neural network, and change them so as to reduce the difference from the output from the output layer to learn the neural network. The exercise evaluation method according to any one of claims 1 to 3. 5. An ideal output, which is an evaluation based on human sensitivities with respect to jerks during various movements of the object, of an input layer and an intermediate layer of the neural network, and a combined weight of the intermediate layer and an output layer, and an ideal output of the object. It is characterized in that the jerk of various movements is detected, the detected jerk is input to the neural network, and the neural network is learned by changing so as to reduce the difference from the output from the output layer. The exercise evaluation method according to any one of claims 1 to 3. 6. The ideal output, which is an evaluation of the combined weight of the input layer and the intermediate layer of the neural network, and the combined weight of the intermediate layer and the output layer, based on human sensitivity to the jerk during various movements of the object, and the A feature is that the jerk of various movements is detected, the detected jerk is input to the neural network, and the neural network is learned in advance by changing so as to reduce the difference from the output from the output layer. The exercise evaluation method according to any one of claims 1 to 3. 7. An ideal output, which is an evaluation based on human sensitivity to the jerk of the motion of the object, and a combined load of the input layer and the intermediate layer and the intermediate layer and the output layer of the neural network, and at the time of evaluating the motion of the object. 4. The motion evaluation method according to claim 1, wherein the neural network is learned as needed by changing it so that a difference from the output of the neural network becomes small. 8. When the motion of the object cannot be evaluated,
4. The exercise evaluation method according to claim 1, wherein an abnormal exercise signal is generated as an unlearned exercise. 9. The acceleration according to claim 1 and the acceleration of the moving object are detected at the same time.
The exercise evaluation method according to claim 1. 10. The motion evaluation method according to claim 1, wherein the jerk is detected by a jerk sensor.