TWI612828B - System and method for detecting falling down in a predetermined living space - Google Patents

Abstract

A home space fall detection system includes a transmitter, a receiver and an information processor. The transmitter is disposed in a preset space, and the transmitter transmits a plurality of wireless network packets, each of the wireless network packets including a signal strength indicator information. The receiver is disposed in the preset space and receives each of the wireless network packets. The information processor is in telecommunication connection with the receiver. The information processor includes an attribute extraction unit and an automatic encoder-like neural network. The invention also discloses a home space fall detection method. With the above system and method, the present invention uses the signal strength indicator information to perform the fall detection work without the need for personnel to carry special equipment, and has the advantages of convenience and privacy.

Description

Home space fall detection system and method thereof

The invention relates to a home space fall detection system and a method thereof, in particular to a home space falling down using a signal strength indicator information to perform a fall detection work without requiring a person to carry a special device, which has the convenience and the advantages of taking privacy into consideration. Detection system and method.

The advancement of medical technology has contributed to the growth of human life, following the advanced countries of Europe, America and Japan. Taiwan has gradually entered an aging society. Falling at home is the most common accident that threatens the elderly living alone.

A fall is a change in posture from a standing or sitting posture that suddenly changes the body to tilt or lie flat. According to US estimates, in 2000, the elderly fell for a total of $19 billion in medical costs; this number grew to 300 in 2010. One hundred million U.S. dollars. If the elderly fall is not immediately discovered, it will seriously threaten the lives of the elderly. The advancement of medical standards has prolonged the life span of human beings, but the changes in the social environment have also produced many elderly people living alone. How to use technology to provide comfortable, safe and privacy-protected home living spaces for these elders is an important administrative goal for many developed countries.

Most of the detection techniques for known fall events must use special equipment, such as accelerometers and gyroscopes on smart phones. This technology uses App to read sensor data in data. When the value exceeds a threshold, the fall subsequent processing is started.

The Republic of China Patent Publication No. I391880 discloses a wearable motion sensing device and a method thereof, which include a wristwatch and a sensing component, the wristwatch and the sensing component are each provided with a three-axis acceleration sensor for measurement. Body posture change and active acceleration value, when the wearer's activity acceleration exceeds the fall threshold, and the posture is detected as a lying posture, it is judged to be a fall, and an alarm can be issued and an emergency ambulance can be notified, for example, the wearer can operate the device by mistake. Cancel the alarm, and when the amount of activity detected in the body and wrist during the non-sleep time is lower than the activity threshold and continues for more than a certain period of time, the activity is too low, and the wearer cancels the alarm if there is an action response, no action response In case of warning and notification of emergency ambulance, if the wearer is misjudged, the device can also be operated to cancel the alarm.

The Republic of China Patent Bulletin No. M449320 discloses a fall detection and warning system for smart mobile phones, mainly comprising a portable smart phone and a sensing module for detecting a fall in a smart mobile phone. . Among them, the sensing module placed on the smart mobile phone is a three-axis acceleration sensor and a gyroscope, and the smart mobile phone is carried by an elderly person living alone or a caretaker. When the user accidentally falls, the sensor can be activated by the sensor module to send an emergency message to the relevant medical staff or contact person for immediate treatment by the medical staff or contact person, thereby improving home care. Safety to prevent physical injury caused by delay in emergency time.

The above patents require the user to carry the equipment for a long time to be effective, and the above patent cannot be implemented if the elderly get up at night and forget to carry it.

The Republic of China Patent Publication No. I275045 discloses a fall event detection method that utilizes three characteristic values of a person's fall time interval, centroid position, and vertical projection height in three different motion modes to detect an observer in a fall time interval. The rate of change of the centroid position and the rate of change of the vertical projection can be notified to the remote family through the monitoring device when the fall event is detected, and the medical damage can be immediately applied to the smart home monitoring.

The above patents use photographic equipment and image processing technology. Although there is no need for personnel to wear equipment, the above patents require specific photographic equipment, and images must be analyzed under a certain optical environment. In addition, long-term image tracking also infringes on the elderly. The privacy of the technology makes the promotion of technology difficult.

Therefore, how to design a home space fall detection system that does not require people to wear equipment, convenience and privacy, has become a common expectation of related equipment manufacturers and developers.

In view of the fact that the fall detection device of the prior art requires personnel to wear equipment and the lack of privacy infringement, the inventor of the present invention actively develops, in order to improve the above-mentioned shortcomings, and after continuous trial and effort, finally develops the present. invention.

The first object of the present invention is to provide a home space fall detection system that does not require a person to wear equipment, and which is convenient and takes privacy into consideration.

A second object of the present invention is to provide a home space fall detection method that does not require a person to wear equipment, and which is convenient and takes privacy into account.

In order to achieve the above object, the home space fall detection system of the present invention comprises a transmitter, a receiver and an information processor.

The transmitter is disposed in a preset space, and the transmitter transmits a plurality of wireless network packets, each of the wireless network packets including a signal strength indicator information.

The receiver is disposed in the preset space and receives each of the wireless network packets.

The information processor is in telecommunication connection with the receiver, the information processor includes an attribute extraction unit and an automatic encoder-like neural network, and the automatic encoder-like neural network includes a complex de-duplication automatic encoder-like neural network layer. And a classification-like neural network layer, each of the de-encoded automatic encoder-like neural network layers includes a plurality of neurons, each of which is connected with the neurons of the previous layer and the next layer, and the neurons of the same layer are not connected. And each of the neurons has a link weight.

The receiver transmits a plurality of the wireless network packets to the information processor at a first time. During the first time, a tester generates a plurality of fall events and a plurality of no fall events in the preset space. The receiver transmits a plurality of the wireless network packets to the information processor at a second time.

The attribute extracting unit calculates a plurality of first derived data of the signal strength indicator sequence of the first time, and the attribute extracting unit calculates a second derived data of the signal strength indicator sequence of the second time.

The automatic encoder-based neural network calculates a complex first event calculation value according to the first derived data and each of the connection weights, where the first event calculation value includes a fall probability in the preset space in the first time period. And the value of the first event is calculated by the automatic encoder-based neural network according to the first event calculation value and the first time event first event actual value of the first time, so that the first event calculation value And the actual value of the first event is within a certain error range, and the actual value of the first event is an actual value of the event corresponding to the first event calculated value in the preset space in the first time, and the actual value of the first event includes The automatic encoder-based neural network calculates a second event calculation value according to the second derived data and the adjusted connection weights, and the second event calculation value includes a fall event actual value and a fall-free actual value. A fall probability value and a no fall probability value, and determining whether a fall event occurs in the preset space in the second time by using the fall probability value and the no fall rate value.

In order to achieve the above object, the home space fall detection method of the present invention comprises the steps of:

Step A: providing a transmitter, a receiver, and an information processor, the transmitter is disposed in a preset space, the receiver is disposed in the preset space, the information processor is electrically connected to the receiver, and the The information processor includes an attribute extraction unit and an automatic encoder-like neural network. The automatic encoder-like neural network is connected to the attribute extraction unit, and the automatic encoder-like neural network includes a complex de-duplication automatic encoder. a neural network layer and a classification neural network layer, each of the de-encoded automatic encoder-like neural network layers includes a plurality of neurons, each of which is connected with a neuron of the previous layer and the next layer, and the same layer of neurons They are not connected, and each of the neurons has a link weight.

Step B: The transmitter transmits a plurality of wireless network packets, and each of the wireless network packets includes a signal strength indicator information.

Step C: The receiver receives each of the wireless network packets.

Step D: providing a tester to generate a plurality of fall events and a plurality of no fall events in the preset space in a first time, the receiver transmitting a plurality of the wireless network packets to the information processing at the first time Device.

Step E: The attribute extraction unit calculates a plurality of first derived data of the signal strength indicator sequence of the first time.

Step F: The automatic encoder-like neural network calculates a complex first event calculation value according to the first derived data and each of the connection weights, where the first event calculation value includes the first time in the preset space. A fall probability value and a no fall probability value.

Step G: The automatic encoder-like neural network adjusts each of the connection weights according to the first event calculation value and the first time event first event actual value of the first time, and the first event actual value is the first time pair The first event should be calculated as the actual value of the event in the preset space, the first event actual value including a fall actual value and a no fall actual value.

Step H: determining whether the difference between the first event calculation value and the actual value of the first event is within an error percentage. If the difference is within the error percentage, performing the next step, if the difference In addition to the error percentage, step F to step G is performed.

Step I: The receiver transmits a plurality of the wireless network packets to the information processor at a second time.

Step J: The attribute extraction unit calculates a second derived data of the signal strength indicator sequence of the second time.

Step K: The automatic encoder-like neural network calculates a second event calculation value according to the second derived data and the adjusted connection weights, and the second event calculation value includes a fall probability value and a fall-free value. The probability value is used to determine whether a fall event occurs in the preset space in the second time by using the fall probability value and the no fall probability value.

With the above system and method, the present invention uses the signal strength indicator information to perform the fall detection work without the need for personnel to carry special equipment, and has the advantages of convenience and privacy.

The preferred embodiments of the present invention are described in detail below with reference to the drawings.

Referring to FIG. 1 to FIG. 2, the home space fall detection system (1) of the present invention comprises a transmitter (10), a receiver (11) and an information processor (12).

The transmitter (10) is disposed in a predetermined space, the transmitter (10) transmits a plurality of wireless network packets (2), and each of the wireless network packets (2) includes a signal strength indicator information.

The receiver (11) is disposed in the preset space and receives each of the wireless network packets (2).

The information processor (12) is telecommunicationally connected to the receiver (11). The information processor (12) includes an attribute extraction unit (120) and an automatic encoder-like neural network (121). The automatic encoder class The neural network (121) includes a complex de-internal automatic encoder-like neural network layer and a classification-like neural network layer, and each of the de-encoded automatic encoder-like neural network layers includes a plurality of neurons (1210), each of which The element (1210) is connected to the neurons of the previous layer and the next layer (1210), and the neurons of the same layer (1210) are not connected, and each of the neurons (1210) has a link weight (W).

The receiver (11) transmits a plurality of the wireless network packets (2) to the information processor (12) at a first time, and a tester generates a plurality of the preset spaces in the first time. In the event of a fall and a plurality of no-fall events, the receiver (11) transmits a plurality of the wireless network packets (2) to the information processor (12) at a second time.

The attribute extraction unit (120) calculates a plurality of first derived data of the signal strength indicator sequence at the first time, and the attribute extraction unit (120) calculates a second derivative of the signal strength indicator sequence at the second time. data.

The automatic encoder-like neural network (121) calculates a plurality of first event calculation values according to the first derived data and each of the connection weights, where the first event calculation value includes the first time in the preset space. A fall probability value and a no fall probability value.

The automatic encoder-like neural network (121) adjusts each of the connection weights according to the first event calculation value and the first first event actual value of the first time, so that the first event calculation value and the first The actual value of the event is within a certain range, and the actual value of the first event is the actual value of the event corresponding to the calculated value of the first event in the preset space in the first time, and the actual value of the first event includes a fall actual value. And the actual value of no fall.

In a preferred embodiment of the present invention, the actual value of the fall of the first event actual value is 1, and the actual value of the no fall is 0, indicating that a fall event occurs, in another preferred embodiment of the present invention. The actual value of the fall of the first event is 0, and the actual value of the no fall is 1, indicating that no fall event occurred.

The automatic encoder-like neural network (121) calculates a second event calculation value according to the second derived data and the adjusted connection weights, and the second event calculation value includes a fall probability value and a fall-free value. The probability value is used to determine whether a fall event occurs in the preset space in the second time by using the fall probability value and the no fall probability value.

In a preferred embodiment of the present invention, the fall probability value of the second event calculation value is greater than the no-fall probability value, and it is determined that the fall event occurs in the preset space in the second time, in another In a preferred embodiment, the fall probability value of the second event calculation value is less than the no-fall probability value, and it is determined that the fall event does not occur in the preset space in the second time.

In a preferred embodiment of the invention, the transmitter (10) is a WiFi sharer. The receiver (11) is a wireless device that can receive WiFi signals, including a single-chip system such as a Raspberry Pi.

In another preferred embodiment of the present invention, the first derived data includes an average value, a standard deviation, an instantaneous frequency, and an instantaneous phase, and the second derived data includes an average value, a standard deviation, an instantaneous frequency, and an instantaneous phase.

In another preferred embodiment of the present invention, the attribute extraction unit (120) uses an Empirical Mode Decomposition (EMD) to decompose the signal strength indicator sequence into a plurality of Intrinsic Mode Functions (Instrinsic Mode Function, IMF), and using Hilbert-Huang Transform to calculate the instantaneous frequency and instantaneous phase of the complex essential mode function, and then obtain the instantaneous frequency of the first derived data and the second derived data. And the instantaneous phase.

Please refer to FIG. 3 together, the home space fall detection method (3) of the present invention includes the following steps:

Step 300: providing a transmitter (10), a receiver (11), and an information processor (12), the transmitter (10) is disposed in a preset space, and the receiver (11) is set in the preset Space, the information processor (12) is telecommunicationally connected to the receiver (11), and the information processor (12) includes an attribute extraction unit (120) and an automatic encoder-like neural network (121), the automatic The encoder-like neural network (121) is telecommunicationally connected to the attribute extraction unit (120), and the automatic encoder-like neural network (121) includes a complex de-noising automatic encoder-like neural network layer and a classification-like neural network. The layer, each of the de-encoded autoencoder-like neural network layers includes a plurality of neurons (1210), each of which is connected to the neurons of the previous layer and the next layer (1210), and the neurons of the same layer (1210) There is no connection between them, and each of the neurons (1210) has a connection weight (W).

Step 301: The transmitter (10) transmits a plurality of wireless network packets (2), and each of the wireless network packets (2) includes a signal strength indicator information.

Step 302: The receiver (11) receives each of the wireless network packets (2).

Step 303: Providing a tester to generate a plurality of fall events and a plurality of no fall events in the preset space in a first time, and the receiver (11) encapsulates the plurality of wireless network packets at the first time (2) Transfer to the information processor (12).

Step 304: The attribute extraction unit (120) calculates a plurality of first derived data of the signal strength indicator sequence of the first time.

Step 305: The auto-encoder-like neural network (121) calculates a complex first event calculation value according to the first derivation data and each of the connection weights, where the first event calculation value includes the pre-time in the first time Set a fall probability value for the space and a drop-free probability value.

Step 306: The auto-encoder-like neural network (121) adjusts each of the connection weights according to the first event calculation value and the first-time complex first event actual value, and the first event actual value is the first The actual value of the event in the preset space is calculated corresponding to the first event in a time, and the actual value of the first event includes a fall actual value and a fall-free actual value.

Step 307: Determine whether the difference between the first event calculation value and the actual value of the first event is within an error percentage. If the difference is within the error percentage, perform the next step, if the difference In addition to the error percentage, step 305 to step 306 are performed.

Step 308: The receiver (11) transmits a plurality of the wireless network packets (2) to the information processor (12) at a second time.

Step 309: The attribute extraction unit (120) calculates a second derived data of the signal strength indicator sequence of the second time.

Step 310: The auto-encoder-like neural network (121) calculates a second event calculation value according to the second derivation data and each of the adjusted connection weights, where the second event calculation value includes a fall probability value and A fall-free probability value is used to determine whether a fall event occurs in the preset space in the second time by using the fall probability value and the no-fall probability value.

In a preferred embodiment of the present invention, the transmitter (10) in the step 300 is a WiFi sharer. The receiver (11) in this step 300 is a wireless device that can receive WiFi signals, including a single-chip system such as a Raspberry Pi.

In another preferred embodiment of the present invention, the first derived data in the step 304 includes an average value, a standard deviation, an instantaneous frequency, and an instantaneous phase. The second derived data in the step 309 includes an average value and a standard. Difference, instantaneous frequency and instantaneous phase.

In another preferred embodiment of the present invention, the attribute extraction unit (120) uses an Empirical Mode Decomposition (EMD) to decompose the signal strength indicator sequence into a plurality of Intrinsic Mode Functions (Instrinsic Mode Function, IMF), and using Hilbert-Huang Transform to calculate the instantaneous frequency and instantaneous phase of the complex essential mode function, and then obtain the instantaneous frequency of the first derived data and the second derived data. And the instantaneous phase.

The invention utilizes a wireless network environment to be characterized by an instantaneous change characteristic of a signal strength indicator caused by a fall event in a home space, and is supplemented by a machine learning algorithm to train a fall detection mode.

In order to expand the detection range, the transmitter (10) and the receiver (11) can be installed at a diagonal position of the preset space; if the single receiver (11) cannot cover the entire preset space, it can be installed. The second or third receiver (11), all of the receivers (11) receive the respective wireless network packets (2) of the transmitter (10). The wireless network packet (2) sent by the transmitter (10) can be calculated by the receiver (11) to calculate the signal strength indicator information of the wireless network packet (2), however, the fall action is a time The consecutive events cannot be judged by the signal strength indicator information of the single wireless network packet (2). Therefore, the receiver (11) must receive a continuous sequence of signal strength indicators before analyzing.

The attribute extraction unit (120) extracts the fluctuation amplitude and phase characteristics calculated from the Hilbert-Yellow transformation from the received sequence of continuous signal strength indicators. Finally, the trained automatic encoder-like neural network (121) is used to make a fall event judgment based on the instantaneous change characteristics of the collected signal strength indicators.

The Hilbert-Yellow conversion calculation consists of two steps:

(1) Empirical Mode Decomposition (EMD) decomposes the signal strength indicator information sequence into the sum of the Intrinsic Mode Function (IMF);

(2) Calculate the instantaneous amplitude and phase of each essential mode function using a Hilbert-Huang transform.

Suppose the receiver (11) receives a sequence of signal strength indicators x ₁ , x ₂ , ..., x _n , x _n = x(t _n ), n = 1, 2, ..., N, t _n is a time series The empirical mode decomposition decomposes the signal strength indicator sequence into a sum of a plurality of essential modal functions and a residual function.

As shown in equation (1), c _j (t) is the jth intrinsic mode function, and the residual function r(t) is a very simple function of the wave characteristic.

(1)

(1)

The so-called intrinsic mode function means that the number of local extremes is the same as or different from the number of zero crossings in the time range considered, and the regional maximum is formed. The average of the lower envelope formed by the envelope and the region minima is zero. The empirical mode decomposition method for obtaining the intrinsic mode function is achieved by the well-known sifting step. The intrinsic modal function with smaller foot size captures the characteristics of the volatility in the original sequence, and the larger the foot, the more slowly volatility. The Hilbert-Yellow transformation of equation (2) can then be used for each essential mode function.

(2)

(2)

The combined intrinsic mode function and its Hilbert-yellow conversion can obtain an analytic function, so it can be expressed as instantaneous amplitude a _j (t) (instantaneous amplitude) and instantaneous phase ψ _j (t) (instantaneous phase), such as equation ( 3) shown:

(3)

(3)

，瞬間頻率則定義為瞬間相位的微分

。 among them

Instantaneous frequency is defined as the differentiation of instantaneous phase

.

Each intrinsic mode function produces a sequence of instantaneous amplitudes a _j (t _n ), n = 1, ..., N and an instantaneous phase sequence ψ _j (t _n ), n = 1, 2, ..., N.

Assuming that the transmitter (10) transmits the packet at a rate of 0.1 kHz, a sequence of signal strength indices of length N = 100 can be generated in one second. Such a signal strength indicator sequence can generate log ₂ (N) – 1 ≒ 6 intrinsic mode functions. Therefore, the Hilbert-Yellow transform extracts the signal strength index sequence fluctuation feature with a total of 600 instantaneous amplitudes a _j (t _n ), n=1,...,100, j=1,...,6 and 600 instantaneous phases ψ _j (t _n ), n=1,2,...,100, j=1,2 ,...,6.

It is known from equation (3) that the 1200 data completely describe the fluctuation characteristics of the original signal strength index sequence. Therefore, the instantaneous variation characteristics of the extracted signal strength indicators have a total of 1202 dimensions:

(μ, σ, a ₁ (t ₁ )), ..., a ₆ (t ₁₀₀ ), ψ ₁ (t ₁ ), ..., ψ ₆ (t ₁₀₀ )).

Where μ is the average of the sequence of signal strength indicators and σ is the standard deviation. Faced with such high-dimensional attributes, special machine learning algorithms must be used to learn the relationship between attributes and fall events.

Referring to FIG. 2, the automatic encoder-like neural network (121) is effective in processing the correspondence of such high-dimensional data. The present invention uses the internal stacking of the automatic encoder-like neural network (121). The Stacked Denoising Autoencoder (SdA) preprocesses the high-dimensional data, and the classification neural network of the above-mentioned layer is used to establish the correspondence between the instantaneous variation characteristics of the signal strength indicator and the fall event. The stacked de-duplication autoencoder is an unsupervised learning mechanism in which multi-layer auto-encoders are stacked together. As shown in FIG. 2, a de-internal automatic encoder is used. The first layer of the encoder is an input layer (I), and the input value is an instantaneous variation characteristic of the signal strength indicator in the above 1202 dimension, and the upper layer is a hidden coding layer ( H), the number of neurons (1210) of the hidden layer (H) may be moderately decreased from the number of neurons (1210) of the input layer (I), and the uppermost layer is the decoded output layer (O), The purpose of the autoencoder is to use symmetric and trained link weights (W) so that the more the output layer (O) can replicate the value of the input layer (I). When training the link weight (W), part of the noise can be imported to the input layer (I) to achieve the purpose of the de-embedded auto-encoder.

The de-duplication auto-encoder can be stacked together to form a stacked de-duplication auto-encoder to achieve the purpose of extracting different levels of abstract features of the original input data, and the stacked de-duplication auto-encoder adopts a layered processing method during training. Avoid the difficulties of traditional multi-level neural network training. That is to say, after the first automatic encoder is trained, the calculation result of the hidden layer (H) is taken as an input, and the second de-duplication automatic encoder is trained, so that a stacking and de-duplication automatic coding can be obtained. Finally, a conventional classification-like neural network is added to it to make a category (ie, a fall event) judgment. When such a deep neural network is used for training learning, the initial link weight (W) of the next few layers of the automatic encoder can be set to the weight of the previously trained auto-encoder, so that faster convergence and generalization can be achieved. A more effective neural network.

The wireless network signal will change when it is disturbed by the human body containing 70% of water. The fall event will form a rapid spatial environment change, and the degree of interference with wireless signal transmission will be more severe. Therefore, if you can try to establish the mode between the fall event and the change of the signal strength indicator based on the Received Signal Strength Indicator (RSI), you can follow the mode to do the subsequent automatic fall detection. The use of this technique must first overcome the feature extraction problem that causes the signal strength indicator to change instantaneously due to a fall event.

The wireless network signal is transmitted in multi-path mode in space. Even if there is no person or object moving, the signal strength index is not completely static, and it will change in a small range with time. Once there is a person moving, especially an environmental change such as a fall, the signal strength indicator received by the receiver will produce a more intense change. Accordingly, by using the characteristics of the instantaneous change of the signal strength indicator to reverse the personnel movement event, the purpose of automatically detecting the fall event can be achieved.

The motion detection technology based on the change of signal strength index is mostly based on the average and standard deviation of the sequence. However, these two attributes cannot fully describe the fluctuation characteristics of the signal strength indicator sequence, for example, any adjustment. The order of the elements of the sequence does not change its mean and standard deviation, but it can change its fluctuations. Therefore, the present invention uses the Hilbert-Huang transform to extract the fluctuation characteristics of the signal intensity index sequence to compensate for the lack of the mean and standard deviation. Hilbert-Huang is one of the important tools for analyzing time series. It has the characteristics of data adaptability, can analyze non-steady-state, non-linear time series, and can perform time-frequency-amplitude three-dimensional refined analysis.

With the above structure, the present invention uses the signal strength indicator information to perform the fall detection work, which not only does not require the user to carry any special equipment, nor does it perform image tracking on the user, but has the advantages of convenience and privacy. Moreover, its structural form is not easily reached by those skilled in the art, and it is novel and progressive.

Through the above detailed description, it can fully demonstrate that the object and effect of the present invention are both progressive in implementation, highly industrially usable, and are new inventions not previously seen on the market, and fully comply with the invention patent requirements. , 提出 apply in accordance with the law. The above is only the preferred embodiment of the present invention, and is not intended to limit the scope of the invention. All changes and modifications made in accordance with the scope of the invention shall fall within the scope covered by the patent of the invention. I would like to ask your review committee to give a clear explanation and pray for it.

(1) ‧‧‧Home Space Fall Detection System
(10)‧‧‧transmitters
(11)‧‧‧ Receiver
(12)‧‧‧Information Processor
(120)‧‧‧Attribute extraction unit
(121)‧‧‧Automatic encoder-like neural networks
(1210) ‧‧‧ neurons
(2) ‧‧‧Wireless network packets
(3) ‧ ‧ home space fall detection method
(W) ‧ ‧ link weights
(I) ‧‧‧Input layer
(O)‧‧‧ Output layer
(H)‧‧‧Hidden Layer Steps 300~Step 318

1 is a schematic diagram of a home space fall detection system of the present invention; FIG. 2 is a schematic diagram of an automatic encoder-like neural network of the present invention; and FIG. 3 is a flow chart of a method for detecting a home space fall detection according to the present invention.

(1) ‧‧‧Home Space Fall Detection System

(10)‧‧‧transmitters

(11)‧‧‧ Receiver

(12)‧‧‧Information Processor

(120)‧‧‧Attribute extraction unit

(121)‧‧‧Automatic encoder-like neural networks

(2) ‧‧‧Wireless network packets

Claims (8)

A home space fall detection system includes: a transmitter disposed in a predetermined space, the transmitter transmitting a plurality of wireless network packets, each of the wireless network packets including a signal strength indicator information; a receiver disposed on the The preset space receives each of the wireless network packets; and an information processor is electrically connected to the receiver, the information processor includes an attribute extraction unit and an automatic encoder-like neural network, the automatic encoder class The neural network includes a complex de-encoded automatic encoder-like neural network layer and a classification-like neural network layer, each of the de-encoded automatic encoder-like neural network layers including a plurality of neurons, each of the neurons and the previous layer and the lower a layer of neurons is connected, the same layer of neurons are not connected, and each of the neurons has a link weight; wherein the receiver transmits a plurality of the wireless network packets to the information at a first time a processor, in the first time, a tester generates a plurality of fall events and a plurality of no fall events in the preset space, and the receiver will plural in a second time Transmitting the wireless network packet to the information processor; the attribute extracting unit calculates a plurality of first derived data of the signal strength indicator sequence at the first time, and the attribute extracting unit calculates the signal strength of the second time a second derived data of the indicator sequence; the automatic encoder-like neural network calculates a complex first event calculation value according to the first derived data and each of the connection weights, and the first event calculation value includes the first time Incorporating a fall probability value and a fall-free probability value of the preset space, the auto-encoder-like neural network adjusts each of the association weights according to the first event calculation value and the first-time complex first event actual value of the first time So that the first event calculation value and the first event actual value are within a certain error range, and the first event actual value is the event corresponding to the first event calculation value in the preset space in the first time. a value, the first event actual value includes a fall actual value and a no fall actual value, and the automatic encoder-like neural network is based on the second derived data and Calculating a second event calculation value, the second event calculation value includes a fall probability value and a no-fall probability value, and determining the second by using the fall probability value and the no-fall probability value Whether a fall event occurs in the preset space during the time. The home space fall detection system of claim 1, wherein the transmitter is a wireless device capable of receiving a WiFi signal. The home space fall detection system according to claim 1, wherein the first derived data includes an average value, a standard deviation, an instantaneous frequency, and an instantaneous phase, and the second derived data includes an average value, a standard deviation, and an instantaneous frequency. And the instantaneous phase. For example, the home space fall detection system described in claim 3, wherein the attribute extraction unit uses the Empirical Mode Decomposition (EMD) to decompose the signal strength index sequence into a plurality of essential mode functions (Instrinsic Mode Function (IMF), and using the Hilbert-Huang Transform to calculate the instantaneous frequency and instantaneous phase of the complex modal function, and then obtain the first derived data and the second derived data. Instantaneous frequency and instantaneous phase. A home space fall detection method includes the following steps: Step A: providing a transmitter, a receiver, and an information processor, the transmitter is disposed in a preset space, and the receiver is disposed in the preset space, the information is The processor is in telecommunication connection with the receiver, and the information processor includes an attribute extraction unit and an automatic encoder-like neural network, and the automatic encoder-like neural network is connected to the attribute extraction unit by a telecommunication device. The neural network includes a complex de-encoded automatic encoder-like neural network layer and a classification-like neural network layer, each of the de-encoded automatic encoder-like neural network layers including a plurality of neurons, each of the neurons and the previous layer and the lower One layer of neurons is connected, the same layer of neurons are not connected, and each of the neurons has a link weight; Step B: The transmitter transmits a plurality of wireless network packets, each of the wireless network packets including a signal Strength indicator information; Step C: The receiver receives each of the wireless network packets; Step D: provides a tester in the first time in the first Setting a plurality of fall events and a plurality of no fall events in the space, the receiver transmits a plurality of the wireless network packets to the information processor at the first time; Step E: the attribute extracting unit calculates the first time a plurality of first derived data of the signal strength indicator sequence; Step F: the automatic encoder-like neural network calculates a complex first event calculation value according to the first derived data and each of the connection weights, the first event The calculated value includes a fall probability value and a no fall probability value in the preset space in the first time period; Step G: the automatic encoder-like neural network calculates the value according to the first event and the plural number of the first time The first event actual value adjusts each of the connection weights, and the first event actual value is an event actual value corresponding to the first event calculation value in the preset space in the first time, and the first event actual value includes a fall actual Value and a no fall actual value; Step H: determining whether the difference between the first event calculated value and the actual value of the first events is an error Within the fractional ratio, if the difference is within the error percentage, the next step is performed, and if the difference is outside the error percentage, then the step F is performed to the step G; Step I: the receiver is in the second Transmitting a plurality of the wireless network packets to the information processor; Step J: the attribute extraction unit calculates a second derived data of the signal strength indicator sequence at the second time; and step K: the automatic encoder The neural network calculates a second event calculation value according to the second derived data and the adjusted connection weights, and the second event calculation value includes a fall probability value and a no fall rate value, and the fall probability is utilized. The value and the no-fall probability value determine whether a fall event occurs in the preset space during the second time. The method for detecting a home space fall as described in claim 5, wherein the transmitter in the step A is a wireless device capable of receiving a WiFi signal. The home space fall detection method according to claim 5, wherein the first derived data in the step E includes an average value, a standard deviation, an instantaneous frequency, and an instantaneous phase, and the second export in the step J The data includes the mean, standard deviation, instantaneous frequency, and instantaneous phase. For example, in the home space fall detection method described in claim 7, wherein the attribute extraction unit uses the Empirical Mode Decomposition (EMD) to decompose the signal strength index sequence into a plurality of essential mode functions (Instrinsic Mode Function (IMF), and using the Hilbert-Huang Transform to calculate the instantaneous frequency and instantaneous phase of the complex modal function, and then obtain the first derived data and the second derived data. Instantaneous frequency and instantaneous phase.