TWI718762B - Smart wearable system and operation method thereof - Google Patents

Abstract

A smart wearable system and operation method thereof are disclosed. The smart wearable system comprises a wearable device; a plurality of sensors disposed on the wearable device for obtaining a plurality of sensing data during performing a motion by a plurality of fingers; a control device for generating appropriate control command by comparing the sensing data with a self-organizing learning, training and evolution data of an ANM system; and a driving device disposed on the wearable device for driving an fake finger to perform a corresponding motion according to the control command from the control device.

Description

Smart wearable system and its operating method

This creation is a wearable device, and especially relates to a smart wearable system with autonomous learning ability and its operation method.

Assistive devices refer to tools, equipment, or products that have been designed or modified to maintain or improve the functions of the user, so that the user can be more independent, convenient and safe in daily life, work or study. For example, a prosthesis is an artificial prosthesis that is specially designed and assembled to make up for the limbs of the amputee or the incomplete limb defect by using the existing advanced engineering techniques and methods, also known as "prosthetics" (or "prosthetics") ). Its main function is to replace some of the functions of the limbs, so that the amputee can restore a certain degree of self-care and work ability.

However, the general prosthesis faces many challenges in operation. For example, the user cannot precisely control the strength of the prosthetic arm to grasp the object, nor can it perform finer (delicate) hand movements.

In view of this, the purpose of this creation is to provide a smart wearable system. Among them, the smart wearable system includes at least: a wearable device with at least one artificial sense finger; a plurality of sensing devices are provided on the wearable device to sense when the user's multiple fingers perform an action A plurality of sensed data; a control device for comparing the sensed data with one of the autonomous learning training and evolution data of the ANM system to generate a control command corresponding to the action; and a driving device on the wearable device , Used to drive the artificial meaning finger of the wearable device to perform a corresponding action according to the control command of the control device.

Wherein, the sensing device includes a plurality of curvature sensors, which are arranged on the wearable device to capture a plurality of curvature data when the fingers perform the aforementioned actions.

Wherein, the sensing device includes at least one muscle sensor for capturing at least one muscle data when a first limb performs the aforementioned action.

Wherein, the sensing device includes a plurality of pressure sensors, which are arranged on the wearable device to capture a plurality of pressure data when the fingers perform the aforementioned actions.

Wherein, the finger is located on the first limb, and the wearable device is worn on the first limb.

Among them, the action performed by the finger is the same as the corresponding action performed by the artificial finger of the wearable device.

Among them, the fingers are located on the first limb, and the wearable device is worn on the second limb

Among them, the corresponding action performed by the artificial sense finger is complementary to the action performed by the finger, and the time sequence of the corresponding action corresponds to the action performed by the finger.

Among them, the ANM system has information processing neurons and control neurons, and the ANM system inputs the sensing data generated by the fingers of a plurality of data collectors to perform the action. The evolutionary learning of information processing neurons and the evolutionary learning of control neurons are used to obtain autonomous learning training and evolution data.

This creation also provides an operation method of a smart wearable system, which at least includes: sensing a plurality of sensing data when a plurality of fingers of the user perform an action; comparing these sensing data with the autonomous learning training and training of the ANM system The evolution data is used to generate a control command corresponding to the action; and at least one artificial finger of the wearable device is driven to perform the corresponding action according to the control command of the control device.

The smart wearable system and its operation method of this creation have the following advantages:

(1) Different from other traditional neural networks, the main neurons of the ANM system created by this author integrate different time signals generated by different neurons into a series of signals at different times to control the functions of other neurons .

(2) This creation can assist rehabilitation patients to understand their hand finger movements and output analysis.

(3) This creation can help patients who have lost their fingers to make a bionic prosthesis, analyze the remaining finger pressure through a smart wearable system, and then scan and analyze the patient's hand, and then model the patient's finger status, and then use the 3D column The prosthesis needed by the patient is printed, and finally driven by the servo motor. It is hoped that the patient can resume the movement like normal fingers.

10: Wearable device

11: Artificial meaning finger

13: Wire

14: Artificial knuckle

20: Sensing device

22: Curvature sensor

24: Pressure sensor

26: Muscle Sensor

40: control device

50: drive device

Figure 1 is a schematic diagram of the use of the created smart wearable system, in which Figure 1 (A) is the palm surface, and Figure 1 (B) is the back of the hand, and the user lacks two fingers (dotted line).

Figure 2 is a schematic diagram of the use of the created smart wearable system when collecting public data.

Figure 3 is a schematic block diagram of the circuit of the created smart wearable system.

Figure 4 is the architecture diagram of the core components of the molecular-like nervous system.

Figure 5 shows the relationship between the detailed architecture of the molecular-like nervous system and its input/output.

Figure 6 is a diagram showing the molecular structure of an information processing neuron.

FIG. 7 is a schematic diagram showing the movement path of the signal.

Figure 8 shows the evolutionary learning diagram of information processing neurons.

Figure 9 shows the evolutionary learning diagram of the control neuron.

Figure 10 is a diagram showing that the learning of the molecular-like nervous system is similar to that of people using the fingers of the hand.

In order to understand the technical features, content and advantages of this creation and its achievable effects, this creation is combined with the diagrams, and detailed descriptions are given in the form of embodiments as follows, and the diagrams used therein have only To illustrate and supplement the manual, it may not be the true proportions and precise configuration after the implementation of this creation. Therefore, the proportion and configuration relationship of the attached drawings should not be interpreted or limited to the scope of rights of the creation in actual implementation. In addition, in order to facilitate understanding, the same elements in the following embodiments are denoted by the same symbols. In addition, the size ratios of the components shown in the drawings are only to facilitate the explanation of the components and their structures, and are not intended to be limiting.

In addition, the terms used in the entire specification and the scope of the patent application, unless otherwise noted In addition, it usually has the usual meaning of each word used in this field, in the content disclosed here, and in the special content. Some terms used to describe this creation will be discussed below or elsewhere in this specification to provide those skilled in the art with additional guidance on the description of this creation.

Secondly, if the terms "include", "include", "have", "contain", etc. are used in this article, they are all open terms, which means including but not limited to.

The purpose of this creation is to make a simple (low cost and light weight) robotic arm wearable device with fingers and grip pressure, and it can be developed as an artificial prosthesis. This creation uses the Artificial Neuro Molecular (ANM) system to learn how people use both hands to control the finger and grip pressure of the hand, and develops robotic prosthetic limbs to further assist in poor hand use. On the patient, to strengthen the use of hand activities in daily life.

In short, after the wearable device provided by this creation has been measured and calibrated, it can first collect data/data of different finger movements, and then use the ANM system to learn how to cooperate with the collected data/data, so that the ANM system Learning is similar to how people use the fingers of their hands to control. The data collected above is selected from one or more of the group consisting of finger flexion, muscle induction, and finger pressure.

Please refer to the illustration. This creation is a smart wearable system, which at least includes a wearable device 10, a plurality of sensing devices 20, an ANM system, a control device 40, and a driving device 50. These sensing devices 20 are provided on the wearable device 10 to sense a plurality of sensing data collected when the wearable device 10 is worn by the user's multiple fingers (ie fingers) to perform a certain action . Such an action may be, for example, a predetermined action, such as holding a cup or holding a knife. Then, the control device 40 is used to compare the sensed data with the autonomous learning training and evolution data of the ANM system to generate a control command (appropriate control command) corresponding to the action. For example, the control device 40 may be electrically connected to the sensing device 20 via the wire 13 to receive the sensing data.

Among them, the characteristic of the ANM system is that it can perform autonomous learning training and evolution, and generate autonomous learning training and evolution data. For example, the ANM system is built-in to the control device 40 In addition, the ANM system, for example, inputs the sensor data of the data collector (the public) via the control device 40, so as to perform autonomous learning training and evolution to generate autonomous learning training and evolution data. Alternatively, the ANM system is built on a cloud system (not shown), for example, and the control device 40 is wired or wirelessly connected to the cloud system, so as to input the sensing data of the data collector (the public) to the ANM system for autonomous learning Training and evolution, and self-learning training and evolution data are generated. In detail, the ANM system collects in advance the sensory data and other data when the fingers perform the above-mentioned predetermined actions, so as to carry out the autonomous learning training and the evolutionary learning process, that is, the autonomous learning training and the establishment of the autonomous learning training and the establishment of the control of the fingers of the hand. Evolution data. Among them, when the smart wearable system of the present invention collects the sensing data of the public, the wearable device 10 may not have the meaning finger, that is, it does not have the driving device 50 and the artificial knuckle 14, and only has the sensing device 20.

One of the characteristics of this creation is that when the ANM system receives the sensing data of the data collector (the public), the ANM system can generate autonomous learning training and evolution data through the process of autonomous learning training and evolution. In addition, the control device 40 can compare the sensed data of the user (patient) with the autonomous learning training and evolution data provided by the ANM system to generate control commands. The driving device 50 is electrically connected to the control device 40, so the driving device 50 can drive at least one artificial finger 11 of the wearable device 10 to perform corresponding actions according to the control command of the control device 40. However, in another embodiment, the ANM system can also receive sensing data of an action performed by the user (patient), and generate an output data after self-learning training and evolution. Then, the control device 40 compares the output data with the self-learning training and evolution data to generate a control command corresponding to the action.

The aforementioned sensing device 20 is selected from one or more of the group consisting of the curvature sensor 22, the pressure sensor 24, and the muscle sensor 26. Taking the wearable device 10 worn on upper limbs such as fingers or palms as an example, the wearable device 10 may be, for example, a glove, and optionally, for example, has at least one artificial finger 11. The number of the aforementioned curvature sensors 22 may be multiple, for example, and are provided in the wearable device 10, for example. The outside of the finger near the joint is used to sense the curvature information of the user's finger when performing the above action. The number of pressure sensors 24 can be multiple, for example, and are arranged on the distal joints, proximal joints, and the inner side of the metacarpophalangeal joints of the fingers (from fingertips to the bottom of the fingers) of the glove-type wearable device for collection and use. The pressure information/data of the distal joint, proximal joint, and metacarpophalangeal joint of the user’s fingers during the above-mentioned actions. The number of muscle sensors 26 may be at least one, for example, to capture muscle data of the limbs when the user's fingers perform the above-mentioned actions.

In one embodiment, the finger is located on the first limb, and the wearable device 10 is also worn on the first limb. For example, the first limb is the right hand, and the fingers are the fingers of the right hand. Taking a user who lacks the middle finger and ring finger of the right hand as an example, the wearable device 10 is worn on the right hand of the user. The wearable device 10 is preferably a glove structure, and preferably has a piercing hole for the user's original The fingers of the pierced into it, and the curvature sensor and the pressure sensor are arranged on the glove structure to sense the data of the glove structure (the user’s original fingers and/or prosthesis) during exercise. The sense finger on the wearable device 10 is used to replace the missing middle finger and ring finger, and the sense finger 11 is preferably provided on the wearable device 10, and the sense finger 11 has at least a plurality of human labors corresponding to the missing fingers. The knuckles 14 and the driving mechanism 50 connected between the artificial knuckles, such as a motor or a servo motor. Therefore, when the user uses the original three fingers (index finger, thumb and little finger) of the right hand to perform actions, such as holding a cup, the sensing device can sense the sensing signals of the three fingers and arm muscles, and ANM The system inputs these sensing signals and performs autonomous learning training and evolution process to provide at least one output signal, wherein the output signal has at least one output data, so that the control device 40 can compare the output signal and the autonomous learning training and evolution data, thereby Generate control instructions to make the driving device 50 drive the two artificial sense fingers on the wearable device 10 to perform the same actions as the original fingers (three fingers), that is, the two artificial sense fingers can be combined with the other three fingers. Hold the cup with your fingers. In this embodiment, the sensing data may be curvature data and muscle data, for example. However, this creation is not limited to this. For example, in this embodiment, the sensing data can also be curvature data, pressure data, etc. And muscle data. The implementation of the meaning finger 11 of this creation can adopt the structure and design of the meaning finger of the prior art, and the meaning finger 11 has at least a plurality of artificial knuckles 14 corresponding to the above-mentioned missing fingers and a drive connected between the artificial knuckles. The mechanism 50 is, for example, a motor or a servo motor. However, the structure of Yizhi 11 in this creation is not limited to the above examples, and Yizhi 11 in this creation can adopt any existing meaning-finger structure on the market.

In another embodiment, the finger is located on the first limb, such as the right hand, and the wearable device 10 is worn on at least the second limb, such as the left hand. Take fruit peeling as an example. For example, holding the fruit in the left hand and holding the knife in the right hand usually requires complementary movements of the left and right hands and coordinated timing. Therefore, in this other embodiment, when the fingers of the user's right hand perform actions such as holding a knife, the sensing device can sense the sensing data of the right finger and arm muscles, and the ANM system inputs these senses. The measured data is used to provide at least one output signal, wherein the output signal has at least one output data, so that the control device 40 can compare the output signal with the self-learning training and evolution data to generate a control command, so that the driving device 50 drives the left hand worn The artificial prosthetic finger of the wearable device 10 performs an action complementary to that of the right finger, and the time series of the artificial prosthetic finger of the left hand holding an apple corresponds to the right hand holding a knife, that is, when the right finger holds the knife to cut the fruit When the time, the artificial finger of the left hand can hold the fruit tightly.

In the above-mentioned another embodiment, the first limb and the second limb are preferably worn with the wearable device 10 respectively. Among them, depending on actual needs, one or both of the wearable devices 10 worn by the first limb and the second limb may have one or more meanings.

This creation also provides the operation method of the above-mentioned smart wearable system, which at least includes: sensing a plurality of sensing data when a plurality of fingers of the user perform an action; comparing the sensing data with the autonomous learning training of the ANM system The evolution data is used to generate a control command corresponding to the action; and at least one artificial finger of the wearable device is driven to perform a corresponding action according to the control command of the control device.

One of the characteristics of this creation is to use the ANM system to create autonomous learning training and evolution data. The ANM system has information processing neurons and control neurons, and the ANM system inputs multiple finger wears of multiple data collectors (the public). The sensing data generated when the wearable device 10 is performing the above-mentioned actions is used to perform the evolution learning of information processing neurons and the evolution learning of control neurons, so as to obtain autonomous learning training and evolution data.

In detail, the ANM system is mainly motivated by the information processing system of the human brain to capture the close structure/function relationship of biological systems and implement it on a digital computer (program) to obtain a biological body Highly adaptable ability. It has multiple competitive learning networks, and each learning network learns from each other the adjustment of actions according to the information sent from the sensor. The most important thing is that it not only includes the information processing between neurons, it also extends to the information processing of the molecular formula within the neurons. The basic structure of the system is to put the characteristics that are conducive to self-learning (that is, the characteristics of gradual transformation of structure/function) into the structure of the system, especially the interior of neurons. When the structure of the system changes slightly, the functions (functions) shown by the structure of the entire system also change slightly. Taking advantage of the characteristics of this gradual transformation of the structure, it can move forward in the direction of achieving the assigned tasks through an autonomous learning method.

The main component of the ANM system is a central processing subsystem composed of a group of neurons, which can be applied to different problem areas through different output interfaces. The biggest difference between the ANM system and other traditional neural networks is that its main neurons integrate different time signals generated by different neurons into a series of signals at different times to control the functions of other neurons. . If you think of the central processing subsystem as a hierarchical nested network, where each neuron itself is a network with signal integration, this type of neuron can be called an information processing neuron. The ANM system connects these neurons into a population of local networks in a certain way. The structure of the neurons in the local area network is very similar to each other, that is, only in the Some places are slightly different. Therefore, for the same input data, on the one hand, because the structure of each subnet is very similar to each other (only slightly different), on the other hand, because the system has the characteristics of gradually changing the structure/function, the two This factor makes each subnet have a similar output. Using this feature, you can evaluate the performance of each local area network for a certain input, and then select several better-performing local networks and copy them to the lower-performing local networks. Through the above learning process, it is possible to educate (train) these local area networks to achieve certain tasks (or functions).

The ANM system is composed of two main types of neurons. The first type of neuron is called the "information processing neuron", which has the ability to integrate different time signals generated by different neurons and generate a series of different time signals to control other neurons. The second type of neuron is called "control neuron". It has a memory function and can select (combine) a group of "information processing neurons" or "control neurons" to perform a specific function. Figure 4 is the architecture diagram of the core components of the molecular-like nervous system. Figure 5 shows the relationship between the detailed architecture of the molecular-like nervous system and its input/output.

The general neural network neuron processing method is to multiply the input value by its relative weight value and add it, and then go through the activation function to determine whether its neuron produces excitation behavior. Basically, this is a conversion process of a mathematical function. The main learning core of the ANM system is the information processing neuron. This neuron processing method converts a series of input signals into signals on the neuron's cytoskeleton to integrate spatiotemporal signals from different places into a series of output control signals. .

The main learning core of the ANM system is the information processing neuron, which is based on the assumption that the information processing in the system occurs on the cytoskeleton within the neuron. Here, this creation uses a two-dimensional spatial cellular automata (Cellular Automata, CA) to simulate the information processing process on the cytoskeleton. Figure 6 is a diagram showing the molecular structure of an information processing neuron, in which each grid represents the basic unit of the cytoskeleton, represented here by C1, C2, and C3. This is assuming that there are three different elements in the neuron responsible For signal transmission, each element has different transmission characteristics. For example, the transmission speed of the C1 element is the slowest, but the signal transmission energy is the strongest; the transmission speed and energy of the C2 element are both medium; the signal transmission energy of the C3 element is the weakest, But the transmission speed is the fastest.

In the cytoskeleton, each component unit may be a location for signal input and output. The input point is called the import enzyme, and the output point is called the read enzyme. The read enzyme is responsible for receiving the signal transmitted from outside the cell to the cell membrane. And turn it into a signal of molecular structure. When a certain combination of signals arrives and the sum of the signal energy reaches a certain level, the neuron will initiate the trigger when a certain combination of signals arrives. However, this mode has some limitations, that is, the read-in enzyme can be configured on any element, but the read-out enzyme can only be configured in the C1 element, which is based on the triggering of neurons caused by a certain element.

The same elements can transmit signals to each other. When the signal from outside the cell is transmitted to the cell membrane to cause the read-in enzyme to start, the read-in enzyme also turns on the element at the same position as it, and the turned-on element affects the neighboring same element again. Therefore, the flow of signals on a certain cytoskeleton is formed. When the reading enzyme at position (2, 2) receives the signal, the C2 element is activated, generating a signal that moves along the C2 element, moving from (2, 2) to (8). ,2). In order to form a signal flow in a single direction in the process, the element that is turned on will enter a very short deceleration period after the signal is transmitted. During the deactivation period, the element cannot be activated again, and it must wait until the end of the deflection period to ensure that One-way delivery.

Different elements can also transmit signals to each other, because there is a kind of connected protein (Microtubule Associated Protein, MAP) on the cytoskeleton that specifically connects different elements. When a signal on a certain element encounters a connected protein, it will pass through This protein is transferred to a different element at the other end, causing a change in the energy level of the element at the other end. The change in energy may form a new signal to flow between the elements. For example, when the input enzyme of (4, 3) receives the signal, the C1 element is activated, and then flows along C1 to (8, 3), because (8, 3) has two connecting proteins connected respectively (7, 2) C2 And C3 of (8, 4), so it will affect the energy of these two positions. Since the energy of C1 is greater than C2 and C3, new transmission signals will be generated on the elements C2 and C3 separately. If the signal is generated by (7, If the C2 of 2) is transmitted to (8, 3) C1, it may not be able to generate a new transmission signal on C1. This is because of the different energy transmission intensity between different elements. In the ANM system, in order to make the cytoskeleton have the function of integrating different time and space signals, it is assumed that different elements have different degrees of energy influence (as shown in Table 1). The relationship of energy determines how easy it is for the elements to be activated ( The higher the energy level, the easier it is to be activated), it also determines the signal movement time between elements and the signal combination required for neuron triggering.

As shown in Figure 6, although the signal transmitted from C2 of (7, 2) to C1 of (8, 3) does not generate a new transmission signal, because the energy intensity has not reached S, it increases the energy of C1 to I. This means that it is close to the activated state. If a signal comes in from C3 of (8, 4) within a short period of time, it is likely to increase its strength to S and activate the signal movement of the element. On the contrary, if there is no When the signal comes in, the energy will diminish and disappear with time. In addition, it can be seen from Table 1 that the energy transfer intensity between the same elements is all S, that is, C1 vs. C1, C2 vs. C2, and C3 vs. C3. This is because even if there are differences between C1, C2 and C3 The energy intensity of, but the same element has a special relationship between each other and can easily activate each other, so this creation gives it the energy intensity of S, which can also explain why the same element can be used for signal transmission.

As mentioned above, when a certain combination of signals arrives at the reading enzyme, the neuron will trigger because the cytoskeleton integrates the signals at different times and locations. For example, in Figure 6 (8, 3), C1 has Read the enzyme. At the same time, there are two connecting proteins connecting C2 and C3. Therefore, there are three different signals that can be assembled here to trigger neuron triggers. However, these three signals have different activation times and transmission speeds, C2 signal It may be activated by (2, 2) or (3, 2), the C3 signal may also be activated by (2, 4) or (4, 4), and the moving speeds of C1, C2, and C3 are also different. Therefore, the neuron’s The trigger time depends on how the cytoskeleton in the neuron integrates and processes these messages. The time when the user's behavioral action determination is triggered comes from the time from reading in to the reading out of the periosteum in the process of perturbation, and the fitness function is an evaluation of the best score for the user's behavioral actions. When the cell periosteum is in operation, it will randomly arrange the basic components of C1, C2, C3, etc. After evaluation, the cell periosteum in the sub-network with the best score is selected, and its pattern is copied to the next cell periosteum. And add a little variation.

The Wrap-Around Fashion of the cytoskeleton is expressed in a two-dimensional space in the ANM system. Its arrangement is to adopt a wrap-around connection, that is, to connect the corresponding two sides together, so that the signal is in the cytoskeleton There will be no boundary restriction when moving inside. As far as the signal movement path is concerned, it is a circular path. Figure 7 shows a schematic diagram of the signal movement path. For example, when (3, 3) in Figure 7 (left picture) After the reading enzyme on C3 receives the signal, the signal will move along (2, 3), (1, 3), (8, 3), and finally stop at (7, 3). (1,3) and stop, because the previous square of (1,3) is (8,3), this is a circular path. In the same way, when the reading enzyme of (5, 2) C2 in Figure 7 (right) receives the signal, the signal will follow (4, 1), (3, 8), (2, 7), (1). , 6) to move, and finally stop at (8, 5). In other words, as far as each basic unit is concerned, it can move in eight directions, and each direction can form a circular path.

The ANM system is a multi-layered architecture, and its evolutionary learning (or genetic algorithm) allows it to occur at five levels. The first to fourth layers belong to the learning of the "information processing neuron" layer, and the fifth layer belongs to the learning of the "control neuron" layer). The current operating mode of the system is to allow only one level of evolutionary learning within a certain period of time, while other levels of learning are turned off. When a fixed period of study is over, the system will turn off this level of study and turn on other levels of study.

Each "information processing neuron" has the ability to integrate different temporal and spatial information, which is similar to the information processing on the cytoskeleton in the actual neuron. We can imagine that the cytoskeleton itself is a small neuron network. When the signals from outside the cell are transmitted to the cytoskeleton, it will produce a certain flow of signals, and when these signals are in a certain combination on the cytoskeleton It may cause the neuron to emit when it is assembled somewhere. This creation is based on a two-dimensional cell state machine to simulate the information processing and signal flow on the actual neuron cytoskeleton. There are four parameters that can be changed by "information processing neuron": "reading enzyme", "reading enzyme", "basic constituent molecule", and "connectivity protein". Their functions are described as follows: "reading enzyme" It mainly responds to the place where the signal is integrated on the cytoskeleton. In other words, when different signals are assembled in the place where the "reading enzyme" is assembled, the neuron emits. Therefore, its existence determines the firing behavior of "information processing neurons".

"Reading enzyme" is the main place for signal input from "information processing neuron". It is responsible for converting the signal transmitted to the cell membrane of the neuron into the signal on the cytoskeleton. Similarly, its existence determines the type of input signal of the "information processing neuron". .

The main function of "connective protein" is to connect different "basic constituent molecules" on the cytoskeleton, and is responsible for transmitting signals between different "basic constituent molecules". Its existence will control the mutual influence of different types of constituent molecules and the flow of signals.

The function of "basic constituent molecules" is mainly to transmit signals: set different types of "basic constituent molecules" to form transmission signals of different properties (that is, different transmission speeds and different intensities of influence). Changing these "basic components" will affect the type of signal.

The molecular-like nervous system is a multi-layered structure. The learning method of its calculations is mainly to use the evolutionary chemistry proposed by Darwin. Using the "mutation-selection" method, learning can be divided into two major stages. The first stage is the information processing neuron. Evolutionary learning method (Figure 8). The second stage is to control the evolutionary learning method of neurons (Figure 9). The learning steps can be divided into three types:

1. Evaluation: Evaluate the performance of each subnet and learning objective, and select the best subnet.

2. Copy: Copy the selected best subnet cytoskeleton distribution to other subnets.

3. Variation: The cytoskeleton distribution in the subnet is similar, but with some changes.

Generally speaking, computer programs are not suitable for evolutionary learning, because slight changes to a program may lead to an inoperable program. This is because its adaptability curve is a mountain and valley full of distinct heights. The feasible path between each mountain is steep. This high and low structure is very unsuitable for evolutionary learning, because evolutionary learning may stagnate. This creation will use the adjective "multidimensional bypass" (multidimensional bypass) to describe it. Intuitively, there is a bit more space, which increases the chance of saddle points. This intuition is described in a more formal way, that is, the relationship between complexity and stability in an arbitrary dynamic body. The theory The basis is that when the number of constituent elements increases and the relationship between their interaction and adjustment increases, the chances of saddle points will also increase. On the other hand, repetitive and weak interactions play an important role. These factors using evolutionary learning will be placed in neurons and even the network structure of the entire ANM system.

The basic structure of the ANM system is to incorporate the characteristics of evolutionary learning, that is, the characteristics of gradual transformation of the structure, into the structure of the system, so that it will have the function of changing the structure of the organism. The characteristic of this gradual transformation structure is that when the input information changes slightly, its output information also changes slightly. In figure recognition, the characteristics of this gradual transformation structure are used in conjunction with the evolutionary search method to extract the characteristics of each figure to distinguish different figures. On the other hand, when the input figure changes slightly (such as a certain degree of In case of interference), the ANM system has characteristics similar to the gradual transformation of the structure of the organism, and its output information may not have any changes (that is, it can still recognize these slightly changed patterns).

When no time limit is given to the system, it is expected that the system can progress through self-organizing learning to achieve the task delivered. Basically, the structure of the ANM system has the characteristic of "evolution friendliness", which makes it have a gradual transformation of the structure/function of a living organism to provide self-organizing learning, that is, when the structure of the system produces a little The function of the whole system structure also changes slightly. Autonomous learning is a feature of the ANM system, and the premise of this function is that the system must have the characteristics of a gradual transformation of the structure of a living organism. Most of the actual effects of the nervous system must be fine-tuned to be suitable for a certain problem, for example, to recognize a certain set of graphics or Chinese characters. However, when this graphic or Chinese character group changes, the system needs to be fine-tuned again, and these fine-tuned actions are in the general neural learning system. It may be completed by the designer of the system. On the contrary, the ANM system emphasizes that the system performs fine-tuning actions through autonomous learning.

The following is an in-depth description of the adaptability of the ANM system. This creation will use "computational adaptability" to describe it, briefly as follows: Permanently progressive learning: refers to a learning space In an environment where the system exhibits the ability to continuously learn, that is to say, assuming that the difficulty of the problems faced by the system is gradually increased, and the system is given considerable learning time, the system can perform continuous learning; this includes when a certain problem When the degree of difficulty increases to a considerable extent (that is, when the system cannot completely solve the problem under limited computer resources), the system can still show continuous and progressive learning during the learning process.

Mobile problem domain (changing problem domain): Refers to the ability of a system to adapt to changes in the environment (permanent changes in the environment) in a modest way of changing itself. This includes changing the problem domain used in training (or learning). This is the ability of the so-called system of this author to cope with a moving (changing) problem.

Interference fault tolerance: refers to the ability of the system to continue to operate in an environment subject to interference. Unlike the above-mentioned theme (mobile problem domain), this type of environmental change is temporary (not permanent). Therefore, fault tolerance and system credibility are closely related, but not exactly the same. The reason for the instability of a system may be the interference generated within it. However, there is no doubt that fault tolerance is a type of system generalization capability.

The architecture of the ANM system itself has two representations: "internal representation" and "external representation". "Internal expression" refers to the relationship between the internal structure/function of the system, and "external expression" refers to the actual reaction of the system to the environment. In other words, "internal representation" refers to a certain input data, the system "information office The firing situation of the rational neuron, and the “external expression” is the actual firing situation of the “reactor neuron”, that is, the behavior of the system that truly reflects the environment. As mentioned earlier, "information processing neurons" are divided into four groups (for convenience, we will call them N, E, S, W), that is, 32 "information processing neurons" are divided into 8 " N" group of neurons, 8 "S" group of neurons, 8 "E" group of neurons, and 8 "W" group of neurons. The following two examples illustrate the relationship between "internal expression" and "external expression". The first example is that the "internal expression" has changed, but the "external expression" has no change; the second example is that the "internal expression" has changed, but the "external expression" has also changed. The first example is assuming that for a certain input data, if the first "information processing neuron" emitted by the system is a neuron in the "W" group, the second is a neuron in the "N" group Yuan, the rest are "E", "S", "N", "S" in order, and its "internal expression" is (WNESNS). Now suppose that when the internal structure of the system changes, the firing behavior of the system's "information processing neurons" is "W", "N", "E", "E", "S", "N", "S" in order. ", the changed "internal expression" is WNEESNS. As far as "internal expression" is concerned, due to changes in the internal structure of the system, the system's internal expression of a certain input data is changed from WNESNS to WNEESNS (the slashed boldface represents the difference between these two expressions). As far as the "internal representation" is concerned, the system output is only slightly different (because only one "information processing neuron" emits differently). However, as far as the "external representation" is concerned, if the second one generates a trigger " If the time between the neurons in the E group and the first neuron that triggers the "E" group is very short, the "reactor neurons" controlled by the "E" group are still in the "triggered state" Or during the "anti-period", the second signal that generates the neuron that triggers the "E" group will be ignored, so the "external expression" (external behavior) displayed by the system is unchanged. The second example is a continuation of the first example. Now suppose that the internal structure of the system changes again, resulting in the system’s internal representation of a certain input data. The "internal representation" is changed from WNEESNS to WNEWSNS (the slashed bold words represent the two Different expressions). on In terms of "internal expression", the output of the system is only slightly different (because only one "information processing neuron" emits different behaviors). However, in terms of "external expression", it is changed from behavior WNESNS to behavior WNEWSNS, this behavior may cause a 180-degree change in system output. The above two expressions play a very important role in the learning of the ANM system, because, on the one hand, it allows the internal structure to change gradually, but the behavior that appears on the outside does not change at all. This is "redundancy." One of the "surplus" characteristics is also one of the conditions for the gradual transformation of the system. On the other hand, it also allows interaction with the external environment in a greatly changed manner. Through these two alternate learning methods, complete the assigned tasks.

In an implementation aspect, the technology of this creation can be divided into three stages. The first stage is to provide a wearable device for the bending of the fingers of the right hand and the muscle sensing of the right forearm, and perform actual measurement and calibration of the device, and then perform different movements of the fingers Finally, the ANM system is used to learn how to cooperate with the data obtained through the right hand finger flexion and the right forearm muscle sensing device, so that the ANM system learns similar to the control of people using the fingers of the hand (Figure 10). In detail, in terms of the actual measurement and calibration of curvature and pressure, this creation is to stick or set sensing devices (such as curvature sensor and pressure sensor) on the glove, and use the control device (Arduino control By means of sensing data and capturing data, the curvature is adjusted according to the maximum and minimum values of each sensor, so that the curvature conforms to the degree of freedom of the human finger. Subsequently, this creation collects the gesture curvature and pressure data of different subjects (data collectors), and analyzes the difference in finger curvature and pressure in different actions of the same subject.

The second stage is to complete the wearable device for left hand pressure and left forearm muscle sensing, and perform actual measurement and calibration of the device, and then collect data on different objects held by the hand. Finally, the ANM system is used to learn left hand sensing The information obtained by the device allows the ANM system to learn the control similar to what people use to hold with their hands.

The third stage of this creation is applied to assist people (or patients with hand and finger inconvenience). This third stage of the two-handed wearable device (the left hand is responsible for left finger pressure and left forearm muscle sensing, while the right hand is responsible for right finger bending and right forearm muscle sensing pressure) are used together to complete the functional complementation of a certain action. It is the so-called "artificial prosthesis." Therefore, this creation can assist patients whose left and right hands are incomplete or whose movements remain unchanged to complete a specific action.

In short, the smart wearable system of this creation combines the sensing of the completion of a certain action with both hands (for example, the action of using both hands to cut fruit, the left hand is holding the fruit, and the right hand is holding the fruit knife), That is, while performing a certain action, collecting the time series obtained by the left hand pressure and the left forearm muscle sensing device, and collecting the time series data of the right finger bending and right forearm muscle sensing. Then use the ANM system to learn how to match the time series data obtained from the left and right hand sensing devices. In this creation, for example, a 3D printer can be used to print the left and right hands of the prosthesis, which is the so-called "artificial prosthesis". Assuming that the user loses a certain finger function, for example, due to a certain inherent defect, so that part of the function is lost, the ANM system of this creation can learn and adapt to this defect under this limitation, and show normal Operational conditions.

This creation uses the ANM system to learn the values retrieved from the sensors, that is, the molecular-like nervous system learns to make controls similar to the hand movements of the subject. In other words, the output result of the ANM system for each action should be as close as possible to the value captured from the sensor. The following description first explains the interpretation method (combining the system output and the sensed value) of the output result of the ANM system for each sensing data, and then explains the input method of the ANM system.

In the output part, each finger bending control in this creation is controlled by a "reactor-type neuron". In terms of four finger bending controls in each hand, the entire system has four "reactor-type nerves". yuan". In the learning process, when the "reactor neuron" produces the first firing When, it represents the activation of the control of the finger. When the "reactor neuron" produces a second firing, it represents the cessation of the control of the finger. Therefore, the difference between the first two firing times of a certain "reactor neuron" represents the start and stop time of the finger bending. Since the rotation speed of the motor (that is, the driving device) is fixed, the time the motor rotates can determine the degree of bending of the finger. In addition, when the "reactor neuron" fires for the first time, it represents the time to start the bending of the finger. Each "reactor neuron" is controlled by 8 "information processing neurons". In other words, the firing behavior of these 8 "information processing neurons" determines the control of the "reactor neuron" Launch time. According to the current practice, the firing time of the first two "information processing neurons" of the 8 "information processing neurons", indirectly, it represents the activation time and the degree of bending of the control finger. For the control of four fingers, it needs to be controlled by 4 "reactor neurons" (indirectly, it means 32 "information processing neurons"). In order to produce the above-mentioned harmonious bending behavior of the four fingers, the 32 "information processing neurons" of each subnet must also learn the best control time (that is, they must compete with each other to find the best time to control the movement of this finger ), in other words, it must coordinate the trigger time of 32 "information processing neurons" to produce the ANM system's four hand fingers to bend to control the rhythmic movement.

In the input part, for each test subject, each experimental action is encoded into a 64-bit data, that is, each input data is a series of "...*.**.. "Symbol, it corresponds to a series of "receptor neuron", each bit (* or .) corresponds to a "receptor neuron" (* means that the neuron is activated, and that the neuron is in Suppressed state) (Figure 10).

This creation uses the molecular-like nervous system to learn the values retrieved from the left-hand pressure sensor, which means that the molecular-like nervous system learns to control the pressure similar to the fingers of the subject’s hand. In other words, the output result produced by the ANM system for each action should be as close as possible to the one captured from the sensor To the value. The following description first explains the interpretation method of the output result of the ANM system for each sensing data, and then explains the input method of the ANM system.

In the output part, considering that there are four finger pressure controls in each hand, each finger pressure control in this creation is controlled by a "reactor neuron". In the learning process, when the "reactor neuron" produces the first firing, it represents the activation of the finger pressure control, and when the "reactor neuron" produces the second firing, it represents Stop the finger pressure control. Therefore, the difference between the first two firing times of a certain "reactor neuron" represents the start and stop time of the finger pressure (interpreted as the pressure of the finger). In addition, when the "reactor neuron" emits for the first time, it represents the time to activate the finger pressure. In the input part, for each experimental action, it is encoded into a 64-bit data. This method is the same as the former, so I won’t repeat it.

This creation uses Micro Electro Mechanical Systems (MEMS) to make a pair of left and right hand prostheses with different functions (ie, bionic manipulator fingers). In terms of prosthetic development, this creation uses a 3D printer to produce (print) an InMoov robotic finger similar to an open source hardware, and then the time series data collected by the sensor, and through the algorithm ANM system, Learn to control the intelligent prosthesis to produce different manual behaviors, and further learn to use the different functions of the left and right hands to complete a specific manual function, so as to assist people to use the prosthesis to make movements that are similar to humans.

The smart wearable system and its operation method of this creation have the following advantages:

(1) Different from other traditional neural networks, the main neurons of the ANM system created by this author integrate different time signals generated by different neurons into a series of signals at different times to control the functions of other neurons .

(2) This creation can assist rehabilitation patients to understand their hand finger movements and output analysis.

(3) This creation can help patients who have lost their fingers to make a bionic prosthesis, analyze the remaining finger pressure through a smart wearable system, and then scan and analyze the patient's hand, and then model the patient's finger status, and then use the 3D column The prosthesis needed by the patient is printed, and finally driven by the servo motor. It is hoped that the patient can resume the movement like normal fingers.

10: Wearable device

11: Artificial meaning finger

13: Wire

14: Artificial knuckle

20: Sensing device

22: Curvature sensor

24: Pressure sensor

26: Muscle Sensor

40: control device

50: drive device

Claims (6)

A smart wearable system includes at least: a wearable device; a plurality of sensing devices are provided on the wearable device for sensing when a user's fingers perform an action with a plurality of time series A plurality of sensed data having the time series; a control device that compares the sensed data with one of the autonomous learning training and evolution data of an ANM system to generate a control command corresponding to the action; and a drive The device is installed on the wearable device, and is used to drive an artificial sense finger of the wearable device to perform a corresponding action according to the control command of the control device, wherein the corresponding action performed by the artificial sense finger is complementary The actions performed on the fingers, the fingers are located on a first limb, the wearable device is worn on the first limb or a second limb, and the plural time series of the corresponding actions are The time series corresponding to the action. For example, the smart wearable system described in claim 1, wherein the sensing devices include a plurality of curvature sensors, which are arranged on the wearable device to capture the fingers when performing the action The plural curvature data. For the smart wearable system described in item 1 or item 2 of the scope of patent application, the sensing devices include at least one muscle sensor for capturing at least one muscle data when the first limb performs the action. For example, the smart wearable system described in item 3 of the scope of patent application, wherein the sensing devices include a plurality of pressure sensors, which are arranged on the wearable device to capture the movement of the fingers when performing the action Multiple pressure data. Such as the smart wearable system described in item 1 of the scope of patent application, wherein the ANM system has information processing neurons and control neurons, and the ANM system inputs multiple fingers of multiple data collectors to perform the action. The generated sensing data is used for the evolution learning of the information processing neuron and the evolution learning of the control neuron, so as to obtain the autonomous learning training and evolution data. An operation method of a smart wearable system at least includes: sensing a plurality of sensed data having a plurality of time series when a plurality of fingers of a user perform an action having the plurality of time series; Test data and one of the autonomous learning training and evolution data of an ANM system to generate a control command corresponding to the action; and drive at least one artificial finger of a wearable device to perform a corresponding action according to the control command of the control device , Wherein the corresponding action performed by the artificial finger is complementary to the action performed by the fingers, the fingers are located on a first limb, and the wearable device is worn on the first limb or the first limb On the two limbs, and the plural time series of the corresponding action correspond to the time series of the action.

Images