A Sleep Posture Recognition Pillow Based on Acoustic Characteristics and a Recognition Method

Abstract

This invention discloses a sleep posture recognition pillow and method based on acoustic characteristics. A slit is cut into the bottom of the latex pillow, and a microphone, speaker, and chip are all located at this slit. The microphone and speaker are respectively positioned at the front and middle of the pillow. The two ends of the chip are connected to the microphone and a data processing device, respectively. The speaker and data processing device are electrically connected. The method includes collecting sound signals from different sleep positions and postures using the microphone; preprocessing the converted electrical signals using the data processing device; extracting recognition feature values ​​to construct a pose training dataset; establishing a neural network model; training the model using the pose training dataset; obtaining the trained neural network model; and recognizing sleep postures and positions based on the trained neural network model. This invention has a simple structure, low equipment cost, meets privacy protection requirements, and is easy to use.

Description

Technical Field

[0001] This invention relates to an acoustic smart sleep pillow in the field of sleep monitoring, and in particular to a sleep posture recognition pillow and recognition method based on acoustic characteristics. Background Technology

[0002] Sleep plays a vital role in daily life, and sleep posture reflects sleep quality and health status. In the field of sleep monitoring, sleep posture recognition is an important research area. Sleep posture recognition helps maintain cervical spine health and is crucial for preventing pressure sores and obstructive sleep apnea. In terms of clinical needs, patient-specific sleep posture recognition can help caregivers more effectively adjust patient posture, providing medical value in a range of areas.

[0003] Based on how they capture human body states, non-invasive sleep posture monitoring methods can be divided into wearable and non-wearable types. Wearable methods involve placing sensors on the body, such as the wrists, ankles, and chest. Non-wearable methods include the use of cameras, pressure sensors, infrared sensors, millimeter-wave radar, dense flexible sensor arrays, printed electrodes, and other sensors built into the bed and pillow.

[0004] However, existing wearable sleep posture recognition systems may interfere with users' sleep, cause discomfort, and are expensive; camera-based sleep posture recognition systems also raise privacy concerns; and optical sensors or cameras are susceptible to interference from ambient light. There is a lack of a non-wearable sleep posture recognition system that maintains high accuracy, is cost-effective, does not interfere with sleep, and protects user privacy. Summary of the Invention

[0005] In order to solve the problems existing in the background art, the purpose of this invention is to design a sleep posture recognition pillow and recognition method based on acoustic characteristics, which can monitor the user's head sleep posture through a built-in microphone and speaker.

[0006] The technical solution of the present invention is as follows:

[0007] I. A sleep posture recognition pillow based on acoustic characteristics:

[0008] The sleep posture recognition pillow includes a latex pillow, a built-in microphone array for receiving sound signals, a built-in speaker array for playing sound, a chip, and a data processing device. The bottom of the latex pillow has a cutout, and the built-in microphone array, built-in speaker array, and chip are all located at the cutout at the bottom of the latex pillow. The built-in microphone array and built-in speaker array are respectively located at the front and middle of the latex pillow. The input and output terminals of the chip are respectively connected to the microphone of the built-in microphone array and the data processing device. The speakers of the built-in speaker array are electrically connected to the data processing device.

[0009] The built-in microphone array is mainly formed by two sets of built-in microphones arranged at intervals along the length of the latex pillow. Each set of built-in microphones is mainly formed by two microphones arranged at intervals along the length of the latex pillow. The built-in speaker array is mainly formed by two speakers arranged at intervals along the length of the latex pillow. The speaker is located in the middle of the two microphones in the built-in microphone set.

[0010] The built-in speaker array outputs sound signals from the speakers, and the microphone receives the sound signals output from the speakers. The received sound signals are then transmitted to the chip, which converts the input sound signals into electrical signals and transmits them to the data processing device. The data processing device processes and identifies the input electrical signals and finally outputs the human body's sleeping posture and sleeping position.

[0011] The outer surface of the latex pillow is made of knitted fabric, and the inner core of the latex pillow is made of latex and slow-rebound memory foam.

[0012] The sleep posture recognition pillow is used to identify the sleep posture and position of a person during sleep, including supine and lateral sleeping postures.

[0013] II. A sleep posture recognition method based on acoustic characteristics, comprising the following steps:

[0014] Step S1: Turn on the speaker. First, four sound signals are collected from the subject in different sleep positions and postures through four microphones. The chip converts the four sound signals into four electrical signals and transmits them to the data processing device.

[0015] Step S2: Then, the electrical signal is preprocessed using a data processing device to extract the recognition feature values ​​under different sleep positions and sleep postures, thereby constructing a pose training dataset.

[0016] Step S3: Establish a neural network model for sleep posture and position recognition. Use the pose training dataset from Step S2 to train the model. After training, obtain the trained neural network model.

[0017] Step S4: Identify human sleeping posture and position

[0018] When the human body is asleep, the newly acquired actual sound signals are obtained, and the sleep posture and position are identified according to the pre-trained neural network model. The actual sleep position and sleep posture of the human body are then determined and recorded.

[0019] Step S2 specifically involves:

[0020] Step S2.1: Use the four electrical signals output by the chip as the initial electrical signals of the four channels respectively, and use the data processing device to normalize the initial electrical signals of the four channels respectively to obtain four continuous electrical signals.

[0021] Step S2.2: Use s windows to perform sliding window processing on the four continuous electrical signals. The j-th continuous electrical signal is truncated into s time series P by the window. jn Each window contains four time series P jn For each time series P jn After performing a Fast Fourier Transform, the corresponding frequency domain sequence Q is obtained. jm A time series P jn and the corresponding frequency domain sequence Q jm Form a set of discrete sequence groups P jn -Q jm This ensures that each window contains four discrete sequence groups, time series P jn and frequency domain sequence Q jm Specifically, it is expressed as follows:

[0022] P jn =p j1 ,p j2 ,...,p ji ,...,p jn

[0023] Q jm =q j1 ,q j2 ,...,q ji ,...,q jm

[0024] Where, p ji and q ji They are time series P jn and frequency domain sequence Q jm The i-th element in the sequence, where the index i represents the ordinal number of the element and the index n represents the time series P. jn The total number of elements in the sequence, where the subscript m represents the frequency domain sequence Q. jm The total number of elements in the middle;

[0025] Step S2.3: Extract time-domain and frequency-domain features from the four discrete sequence groups in each window to obtain recognition feature values, and then construct a pose training dataset.

[0026] The specific method for obtaining a continuous electrical signal using normalization processing in step S2.1 is as follows:

[0027] The amplitude x of the continuous electrical signal is obtained by processing it according to the following formula. norm :

[0028]

[0029] Where x is the signal amplitude of the initial electrical signal, x min and x max These are the minimum and maximum amplitudes of the initial electrical signal, respectively.

[0030] Step S2.3 specifically includes:

[0031] Step S2.3.1: Extract the recognition feature values ​​from the four discrete sequence groups of each window. The recognition feature values ​​of each discrete sequence group include the maximum value (max), minimum value (min), and average value. Variance D, Median M, Energy E, Entropy S, Frequency Domain Mean The center frequency f, frequency domain variance YD, and the identification feature values ​​of each discrete sequence group are obtained by processing them in the following way:

[0032] Max is the time series P jn p ji The maximum value; Min is the time series P jn p ji The minimum value of P; M is the time series P jn p ji The minimum value;

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039]

[0040] Step S2.3.2: First, set the window to a 4×4 matrix. The element corr(j,k) in the j-th row and k-th column of the 4×4 matrix is ​​obtained as follows:

[0041]

[0042] Where j, k = 1, 2, 3, 4;

[0043] Then, K-means clustering is performed on the 4×4 matrix to obtain the class ID of each window;

[0044] Step S2.3.3: Construct a pose training dataset, which includes all the recognition feature values ​​and category IDs in s windows, and use the pose training dataset as the input data for the neural network model.

[0045] The topology of the neural network model in step S3 is as follows:

[0046] The identification feature value and category ID are connected to the input end of the input layer, the output end of the input layer is connected to the input end of the hidden module, and the output end of the hidden module is connected to the output layer; the hidden module is mainly composed of five hidden layers connected in sequence.

[0047] The acoustic properties of latex pillows are affected by the magnitude and location of applied pressure. Different sleeping positions and postures apply varying degrees of pressure to different parts of the latex pillow, causing it to deform and affecting its acoustic properties. This results in varying acoustic impedances encountered by sound as it travels from the speaker to the microphone, thus altering the characteristics of the received sound signal.

[0048] The sleep posture recognition pillow uses the difference in sound signals received by the built-in microphone as one of the standards for determining sleep position and sleep posture. Environmental noise outside the latex pillow and the sound emitted by the latex being squeezed will not affect the sleep posture recognition results.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. The sleep posture recognition pillow of the present invention utilizes the acoustic properties of latex material to recognize sleep posture, which has the advantage of causing less harm to the user.

[0051] 2. The sleep posture recognition pillow of the present invention is a non-wearable sleep posture recognition system. The microphone and speaker are installed inside the latex pillow and will not disturb the user's sleep.

[0052] 3. The sleep posture recognition pillow of the present invention has a simple structure, low equipment cost, meets the needs of privacy protection, and has the advantages of being easy to use.

[0053] 4. The latex pillow of the present invention has a good sound absorption coefficient. The ambient noise outside the latex pillow has almost no effect on the sound signal received by the built-in microphone, that is, the result of sleep posture recognition is less affected by ambient noise.

[0054] 5. The sleep posture recognition pillow of the present invention uses a deep learning algorithm to classify and recognize the user's sleep posture, and can maintain high accuracy in detecting sleep position and posture. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the sleep posture recognition pillow of the present invention;

[0056] Figure 2 This is a schematic diagram of the sleep position recognition area of ​​the sleep posture recognition pillow of the present invention;

[0057] Figure 3 This is a flowchart illustrating the sleep posture recognition process of the sleep posture recognition pillow of the present invention.

[0058] Figure 4 A schematic diagram showing a user lying supine on the sleep posture recognition pillow of this invention;

[0059] Figure 5 A schematic diagram of the user lying on their side on the invention of the sleep posture recognition pillow;

[0060] Figure 6 This is a diagram of the neural network structure for sleep posture recognition in the sleep posture recognition pillow of the present invention.

[0061] In the picture: 1-Latex pillow, 2-Microphone, 3-Speaker, 4-Chip. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] like Figure 1As shown, the sleep posture recognition pillow in a specific implementation includes a latex pillow 1, a built-in microphone array for receiving sound signals, a built-in speaker array for playing sound, a chip 4, and an external data processing device. The bottom of the latex pillow 1 has a cutout, and the built-in microphone array, the built-in speaker array, and the chip 4 are all located at the cutout at the bottom of the latex pillow 1. The built-in microphone array and the built-in speaker array are respectively located at the front and middle of the latex pillow 1. The input end and the output end of the chip 4 are respectively connected to the built-in microphone array and the data processing device. The built-in speaker array is electrically connected to the data processing device.

[0064] The front direction of latex pillow 1 corresponds to the direction the head faces when the person is lying flat, and the back direction of latex pillow 1 corresponds to the direction the feet face when the person is lying flat. The left-right direction of latex pillow 1 is perpendicular to the front-back direction of latex pillow 1.

[0065] like Figure 4 and Figure 5 As shown, the built-in microphone array is mainly formed by two sets of built-in microphones arranged at intervals along the length direction (i.e., left and right direction) of the latex pillow 1. Each set of built-in microphones is mainly formed by two microphones 2 arranged at intervals along the length direction of the latex pillow 1. That is, the built-in microphone array is mainly formed by four microphones 2 arranged at intervals along the length direction of the latex pillow 1. The built-in speaker array is mainly formed by two speakers 3 arranged at intervals along the length direction of the latex pillow 1. The speaker 3 is located in the middle of the two microphones 2 in the built-in microphone set.

[0066] The speaker 3 in the built-in speaker array outputs a sound signal, the microphone 2 receives the sound signal output by the speaker 3, and then transmits the received sound signal to the chip 4. The chip 4 converts the input sound signal into an electrical signal and transmits it to the data processing device. The data processing device processes and identifies the input electrical signal and finally outputs the human body's sleeping posture and sleeping position.

[0067] The outer surface of the latex pillow 1 is made of knitted fabric, and the inner core of the latex pillow 1 is made of latex and slow rebound memory foam.

[0068] Sleep posture recognition pillows are used to identify a person's sleeping posture and position during sleep. Sleep postures include two common postures: supine and side-lying.

[0069] Sleep posture recognition pillows can identify sleep positions, such as... Figure 2 As shown, the four microphones 2 divide the sleep posture recognition pillow into five position areas along its length, from left to right: area 1, area 2, area 3, area 4 and area 5. Each microphone 2 is located at the boundary between two adjacent areas, and the two speakers 3 are located in area 2 and area 4 respectively.

[0070] The latex pillow 1 has a cutout at the bottom, creating a cross-section inside for securing the speaker 3 and microphone 2. Microphone 2 is specifically a four-channel high-sensitivity USB microphone used to receive audio signals. The four built-in microphones divide the length of the latex pillow 1 into five sections. The speaker 3 is used to play soothing music that promotes sleep.

[0071] The data processing device allows selection of the sound played by speaker 3. USB microphone 2 transmits the received sound signal to the data processing device, where it is processed. The data processing device employs deep learning algorithms to identify and classify sleep position and sleep posture.

[0072] The latex pillow 1 used in this invention measures 720mm × 410mm × 115mm. Its outer surface material is knitted fabric, and its inner layer consists of latex and slow-rebound memory foam. The latex content is over 80%, and the overall net weight is 1500g. A 70mm slit with a 250mm depth is located at the bottom of the pillow, forming a cross-section inside. Microphone 2, built-in speaker 3, and chip 4 are all fixed here. The four built-in microphones divide the pillow's length into five sections, each 144mm long. Two built-in speakers 3 are located between the two microphones. Tape marking the boundaries of the five sections is affixed to the part of the pillow near the neck. When the user lies on a section of the pillow, their head and neck must be positioned between two boundaries.

[0073] In this invention, the acoustic properties of latex are a crucial factor in sleep position and posture recognition. The acoustic properties of latex material (including transmission loss and sound absorption coefficient) are related to the average pore size of the latex material. The pore structure of latex foam is the sound transmission medium; therefore, the smaller the average pore size, the stronger the resistance to sound signal transmission and the higher the transmission loss. When a user's head lies on a latex pillow, different head positions and postures apply varying degrees of pressure to different parts of the pillow, causing deformation and altering the average pore size of the local latex material. This results in varying resistance to sound transmission from the built-in speaker to the built-in microphone, thus changing the characteristics of the sound signal received by the built-in microphone. Simultaneously, due to the latex pillow's excellent sound absorption coefficient, ambient noise outside the pillow has almost no impact on the sound signal received by the built-in microphone, meaning it does not affect the sleep posture recognition results. However, when the user's head rests on the latex pillow, the latex is compressed and emits a slight sound. Therefore, this invention utilizes the differences in sound signals captured by the four built-in microphones as one of the criteria for determining sleep position and posture.

[0074] When using the sleep posture recognition pillow for sleep posture recognition, the recognizable sleep positions include 5 zones, divided equally from left to right by 4 built-in microphones, with a length of 144mm. Recognizable sleep postures include supine and lateral positions. Different sleep positions and postures have varying effects on sound propagation within the latex pillow. These effects are received and processed by the built-in microphones, and each combination of sleep position and posture generates a four-channel sound signal time sequence. The USB microphone array transmits this data to the data processing device.

[0075] To achieve sleep position and posture recognition, the sleep posture recognition pillow employs a deep learning method. First, the collected data needs to be preprocessed, which involves three steps: normalization, sliding windowing, and time-domain / frequency-domain feature extraction. This process obtains the features of the collected data, providing input information for the deep learning algorithm. The steps include, for example... Figure 3 As shown:

[0076] Step S1: Turn on the speaker 3. First, the microphone 2 collects the sound signals of the subject in different sleeping positions and different sleeping postures. The chip (4) converts the sound signals collected by the four microphones 2 into four electrical signals and transmits them to the data processing device.

[0077] Step S2: Then, the electrical signal is preprocessed using a data processing device and a deep learning algorithm to extract the recognition feature values ​​under different sleep positions and sleep postures, and then a pose training dataset is constructed.

[0078] Step S3: Establish a neural network model for sleep posture and position recognition. Use the pose training data set obtained in step S2 for training. After training is completed, obtain the trained neural network model.

[0079] Step S4: Identify human sleeping posture and position

[0080] When the user is asleep, the system acquires newly collected actual sound signals, uses a trained neural network model to identify the user's sleeping posture and position, and determines and records the actual sleeping position and posture of the human body.

[0081] Step S2 is as follows:

[0082] Step S2.1: The four electrical signals output by chip 4 are used as the initial electrical signals of the four channels. The data processing device is used to normalize the initial electrical signals of the four channels to obtain four continuous electrical signals; wherein, one microphone 2 corresponds to one channel.

[0083] Step S2.2: Use s windows to perform sliding window processing on the four continuous electrical signals, so that the j-th continuous electrical signal is truncated into s time series P by the s windows. jn Each window contains four time series P obtained by extracting four continuous electrical signals respectively. jn P 1n P 2n P 3n and P 4n In the window, for each time series P jn After performing a Fast Fourier Transform, the corresponding frequency domain sequence Q is obtained. jm A time series P jn and a corresponding frequency domain sequence Q jm Form a set of discrete sequence groups P jn -Q jm This results in each window containing four discrete sequence groups, namely P 1n -Q 1m P 2n -Q 2m P 3n- Q 3m and P 4n- Q 4m :

[0084] P jn =p j1 ,p j2 ,...,p ji ,...,p jn

[0085] Q jm =q j1 ,q j2 ,...,q ji ,...,q jm

[0086] Wherein, time series P jn The subscript j in the text indicates that the time series is formed by extracting the j-th continuous electrical signal, where j = 1, 2, 3, 4; p ji and q ji They are time series P jn and frequency domain sequence Q jm The i-th element in the sequence, where the index i represents the ordinal number of the element and the index n represents the time series P. jn The total number of elements in the sequence, where the subscript m represents the frequency domain sequence Q. jm The total number of elements in the middle;

[0087] In the sliding window operation, the window size is first set. The number of windows, 's', is obtained from the length of the electrical signal. Four continuous electrical signals are truncated using 's' windows, and each continuous signal is truncated into 's' time series. Therefore, the sliding window operation yields a total of 4's time series. Each continuous signal, after being truncated by a window, is converted into a time series within that window, so each window includes four time series. One time series corresponds to one frequency domain sequence, and the two are combined into a discrete sequence group. Each discrete sequence group contains 10 identification feature values. Each window contains four discrete sequence groups, therefore each window has 40 identification feature values. In addition, each window includes a category ID, and the identification feature values ​​and category ID are used as input data for the neural network model.

[0088] Each frequency domain sequence Q jm From a time series P jn It is obtained through Fast Fourier Transform.

[0089] Step S2.3: Extract time-domain and frequency-domain features from the four discrete sequence groups in each window to obtain recognition feature values, and then construct a pose training dataset.

[0090] The specific method for obtaining a continuous electrical signal using normalization in step S2.1 is as follows:

[0091] The amplitude x of the continuous electrical signal is obtained by processing it according to the following formula. norm :

[0092]

[0093] Where x is the signal amplitude of the initial electrical signal, x min and x max These are the minimum and maximum amplitudes of the initial electrical signal, respectively.

[0094] The normalization process is used to normalize the result of each channel of the acquired four-channel electrical signal, and the continuous electrical signal is mapped to the interval [-1,1].

[0095] Step S2.3 specifically includes:

[0096] Step S2.3.1: Extract the recognition feature values ​​from the four discrete sequence groups in the window. The recognition feature values ​​for each discrete sequence group include the maximum value (max), minimum value (min), and average value. Variance D, Median M, Energy E, Entropy S, Frequency Domain Mean The center frequency f, frequency domain variance YD, and the identification feature values ​​of the j-th discrete sequence group corresponding to the j-th continuous electrical signal are obtained by processing in the following way:

[0097] Max is the time series Pjn p ji The maximum value; Min is the time series P jn p ji The minimum value of P; M is the time series P jn p ji The minimum value;

[0098]

[0099]

[0100]

[0101]

[0102]

[0103]

[0104]

[0105] Step S2.3.2: First, set the 4×4 matrix corresponding to the window. The element corr(j,k) in the j-th row and k-th column of the 4×4 matrix is ​​obtained as follows:

[0106]

[0107] Where j, k = 1, 2, 3, 4;

[0108] Then, K-means clustering is performed on the 4×4 matrix to obtain the class IDs of each window;

[0109] Step S2.3.3: Construct a pose training dataset, which includes all the recognition feature values ​​and category IDs in s windows. Use the pose training dataset as input data for the deep learning algorithm and input it into the neural network model.

[0110] Each window contains a time series of 4 channels, which is four discrete sequence groups. Each discrete sequence group contains 10 identification feature values. Each window contains 1 category ID. Therefore, the sample collected by each sliding window contains 41 features, including identification feature values ​​and category IDs.

[0111] The topology of the neural network model in step S3 is as follows:

[0112] like Figure 6 As shown, the identification feature value and category ID are connected to the input end of the input layer, the output end of the input layer is connected to the input end of the hidden module, and the output end of the hidden module is connected to the output layer; the hidden module is mainly composed of five hidden layers connected in sequence.

[0113] The sliding window method is used to acquire discrete samples from a time series of continuous electrical signals. In this invention, the sliding window size is set to 0.5s and the overlap rate is 0.75, which can convert a two-minute time series of each sleep position and sleep posture combination into 960 windows.

[0114] Time-domain and frequency-domain feature extraction is used to extract features from samples acquired through a sliding window. Each audio signal time series can be processed by Fast Fourier Transform (FFT) to obtain a spectrum, thereby calculating the features of the sequence in the time and frequency domains.

[0115] This invention uses a seven-layer fully connected neural network model to classify the calculated features. The first layer is the input layer, which receives 41 features from each sample. The middle five layers are hidden layers, and the last layer is the output layer. The final output is in the form of sleep location ID and sleep posture ID.

[0116] To collect the dataset needed for training the neural network model, this invention plans to collect three sets of data. Two types of sounds will be played using the built-in speaker: a fixed-frequency sound at 200Hz and soft, hypnotic music. The volume will be adjusted to a level audible when the user is lying supine in the center of a latex pillow. The three sets of data will include no sound, a fixed-frequency sound at 200Hz, and hypnotic music, respectively, to enhance the application effect of the sleep posture recognition model in different scenarios. For each type of sound, data will be collected as a control without anyone lying down. The data to be collected includes sound signals from ten scenarios, representing five sleep positions and two combinations of sleep postures. Each sleep position and posture combination will be collected for 120 seconds with an overlap rate of 0.75. The dataset will be randomly divided into a training set (70%) and a test set (30%). After training, a neural network model will be obtained, which can be used to verify the model's recognition accuracy and for sleep posture recognition in practical application scenarios.

Claims

1. A sleep posture recognition pillow based on acoustic characteristics, characterized in that: The device includes a latex pillow (1), a built-in microphone array for receiving sound signals, a built-in speaker array for playing sound, a chip (4), and a data processing device. The bottom of the latex pillow (1) has a cutout. The built-in microphone array, the built-in speaker array, and the chip (4) are all located at the cutout at the bottom of the latex pillow (1). The built-in microphone array and the built-in speaker array are located at the front and middle of the latex pillow (1), respectively. The input end and the output end of the chip (4) are connected to the built-in microphone array and the data processing device, respectively. The built-in speaker array is electrically connected to the data processing device. The built-in microphone array is mainly formed by two sets of built-in microphones arranged at intervals along the length of the latex pillow (1). Each set of built-in microphones is mainly formed by two microphones (2) arranged at intervals along the length of the latex pillow (1). The built-in speaker array is mainly formed by two speakers (3) arranged at intervals along the length of the latex pillow (1). The speaker (3) is located in the middle of the two microphones (2) in the built-in microphone set. The speaker (3) in the built-in speaker array outputs a sound signal, the microphone (2) receives the sound signal output by the speaker (3), and then transmits the received sound signal to the chip (4). The chip (4) converts the input sound signal into an electrical signal and transmits it to the data processing device. The data processing device processes and identifies the input electrical signal and finally outputs the human body's sleeping posture and sleeping position. The outer surface of the latex pillow (1) is made of knitted fabric, and the inner core of the latex pillow (1) is made of latex and slow rebound memory foam.

2. The sleep posture recognition pillow based on acoustic characteristics according to claim 1, characterized in that: The sleep posture recognition pillow is used to identify the sleep posture and position of a person during sleep, including supine and lateral sleeping postures.

3. A sleep posture recognition method based on acoustic characteristics applied to the sleep posture recognition pillow according to any one of claims 1-2, characterized in that, Includes the following steps: Step S1: First, turn on the speaker (3) and collect four sound signals from the subject in different sleep positions and different sleep postures through four microphones (2). The chip (4) converts the four sound signals into four electrical signals and transmits them to the data processing device. Step S2: Then, the electrical signal is preprocessed using a data processing device to extract the recognition feature values ​​under different sleep positions and sleep postures, thereby constructing a pose training dataset. Step S3: Establish a neural network model for sleep posture and position recognition. Use the pose training dataset from Step S2 to train the model. After training, obtain the trained neural network model. Step S4: Identify the subject's sleeping posture and location. When the subject is asleep, the newly acquired actual sound signal is obtained, and the sleep posture and position are identified according to the pre-trained neural network model. The subject's actual sleep position and sleep posture are then determined and recorded.

4. The sleep posture recognition method based on acoustic characteristics according to claim 3, characterized in that: Step S2 specifically involves: Step S2.1: Take the four electrical signals output by the chip (4) as the initial electrical signals of the four channels respectively, and use the data processing device to normalize the initial electrical signals of the four channels respectively to obtain four continuous electrical signals. Step S2.2: Use s windows to perform sliding window processing on the four continuous electrical signals. The j-th continuous electrical signal is truncated into s time series P by the window. jn Each window contains four time series P jn For each time series P jn After performing a Fast Fourier Transform, the corresponding frequency domain sequence Q is obtained. jm A time series P jn and the corresponding frequency domain sequence Q jm Form a set of discrete sequence groups P jn- Q jm This results in each window containing four discrete sequence groups, each discrete sequence group P jn- Q jm Time series P in jn and frequency domain sequence Q jm They are represented as follows: P jn =pj1,pj2,...,p ji ,...,p jn Q jm =q j1 ,q j2 ,...,q ji ,...,q jm Where, p ji and q ji They are time series P jn and frequency domain sequence Q jm The i-th element in the sequence, where the index i represents the ordinal number of the element and the index n represents the time series P. jn The total number of elements in the sequence, where the subscript m represents the frequency domain sequence Q. jm The total number of elements in the middle; Step S2.3: Extract time-domain and frequency-domain features from the four discrete sequence groups in each window to obtain recognition feature values, and then construct a pose training dataset.

5. The sleep posture recognition method based on acoustic characteristics according to claim 4, characterized in that: The specific method for obtaining a continuous electrical signal using normalization processing in step S2.1 is as follows: The amplitude x of the continuous electrical signal is obtained by processing it according to the following formula. norm : Where x is the signal amplitude of the initial electrical signal, x min and x max These are the minimum and maximum amplitudes of the initial electrical signal, respectively.

6. The sleep posture recognition method based on acoustic characteristics according to claim 4, characterized in that: Step S2.3 specifically includes: Step S2.3.1: Extract the recognition feature values ​​from the four discrete sequence groups of each window. The recognition feature values ​​of each discrete sequence group include the maximum value (max), minimum value (min), and average value. Variance D, Median M, Energy E, Entropy S, Frequency Domain Mean The center frequency f, frequency domain variance YD, and the identification feature values ​​of each discrete sequence group are obtained by processing them in the following way: Max is the time series P jn p ji The maximum value; Min is the time series P jn p ji The minimum value of P; M is the time series P jn p ji The minimum value; Step S2.3.2: First, set the window to a 4×4 matrix. The element corr(j,k) in the j-th row and k-th column of the 4×4 matrix is ​​obtained as follows: Where j, k = 1, 2, 3, 4; Then, K-means clustering is performed on the 4×4 matrix to obtain the class ID of each window; Step S2.3.3: Construct a pose training dataset, which includes all the recognition feature values ​​and category IDs in s windows, and use the pose training dataset as the input data for the neural network model.

Images