An AI-powered smart sleep-aid fan light based on multimodal perception and its control method
By identifying the user's sleep state through a multimodal perception module and an edge AI processing unit, a light-wind collaborative control strategy is generated, which solves the problem of the single perception dimension of existing fan lights and achieves personalized sleep aid effects.
Patent Information
- Application Number
- CN202611589330.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-14
- Publication Date
- 2026-11-17
AI Technical Summary
Existing fan lights have a single sensing dimension, cannot identify the user's sleep state, and the light and wind control are not coordinated, making it difficult to provide a personalized sleep-aiding environment.
By employing a multimodal perception module combined with an edge AI processing unit, user data is collected through radar sensors, temperature sensors, sound sensors, and photosensitive sensors. A multimodal fusion sleep stage model is used to identify sleep stages and generate a light-wind collaborative control strategy to achieve adaptive adjustment of light and wind speed.
It achieves personalized, adaptive sleep aid control, providing gradually dimming light and natural wind to avoid light and wind interference and improve sleep quality.
Smart Images

Figure CN123110113A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of fan lights, and in particular to an AI-powered intelligent sleep-aid fan light based on multimodal perception and its control method. Background Technology
[0002] With the fast pace of life, sleep disorders are becoming increasingly common. Existing fan-light products typically only offer lighting and airflow functions. While some products integrate human body sensors or temperature detection, their sensing capabilities are limited, and their control strategies are simplistic, failing to adapt to the user's actual sleep state. For example, some existing technologies use pyroelectric sensors to detect human presence, achieving only a crude "on when someone is there, off when no one is there" control; others adjust fan speed using temperature sensors, but cannot identify whether the user is asleep or in what sleep stage. Furthermore, the lighting modules of existing fan-lights are usually fixed, lacking dynamic lighting effects linked to sleep stages; there is also a lack of coordinated control mechanisms between the fan and the light, making it difficult to maintain a low-wind, low-light-stimulation sleep environment simultaneously after the user falls asleep. Summary of the Invention
[0003] In view of this, the present invention provides an AI-powered smart sleep-aid fan light based on multimodal perception and its control method, aiming to solve the problems of existing fan lights having a single perception dimension, being unable to identify sleep stages, and having uncoordinated light and wind control.
[0004] The technical solution adopted by this invention to solve its technical problem is as follows: An AI-powered smart sleep-aid fan light based on multimodal perception includes a base for fixing to a wall, a housing disposed at the bottom of the base, a drive mechanism disposed at the bottom of the base and located within the housing, and a plurality of fan blades disposed at the bottom of the drive mechanism. It also includes a light module, a multimodal perception module, and a control module. The multimodal sensing module includes a radar sensor, a temperature sensor, a sound sensor, and a photosensitive sensor disposed on the outer periphery of the housing. The radar sensor is used to collect the user's breathing waveform signal and body movement signal in a non-contact manner. The sound sensor is used to collect environmental noise and user snoring signal. The temperature sensor is used to collect the ambient temperature. The photosensitive sensor is used to collect the ambient light intensity. The control module is located inside the base and includes an edge AI processing unit, a remote communication unit, a timing unit, and a storage unit, and is electrically connected to the multimodal perception module, the lighting module, and the drive mechanism, respectively. The edge AI processing unit is configured to perform the following steps: (a) Perform phase calculation on the respiratory waveform signal collected by the radar sensor, extract respiratory frequency and respiratory rhythm variability features, extract time domain features from the body movement signal to obtain body movement intensity and body movement frequency, and perform spectrum analysis on the signal collected by the sound sensor to extract snoring event features; (b) After aligning the breathing features, body movement features, snoring features, and ambient temperature features and ambient light intensity features extracted in step (a) with time, input them into a pre-trained multimodal fusion sleep staging model. The multimodal fusion sleep staging model includes a feature encoding layer and a cross-modal attention fusion layer. The feature encoding layer encodes each modality feature independently, and the cross-modal attention fusion layer calculates the correlation weights between each modality feature through an attention mechanism and performs weighted fusion. (c) The multimodal fusion sleep staging model outputs the user's current sleep stage, which includes at least three of the following: wakefulness, sleep onset, light sleep, deep sleep, and REM sleep. (d) Generate a light-wind coordinated control strategy based on the sleep stage. The light-wind coordinated control strategy includes: when the user is in the sleep stage, controlling the color temperature of the light module to gradually decrease from a first color temperature value to a second color temperature value, and controlling the brightness to gradually decrease from a first brightness value according to a preset dimming curve, while controlling the drive mechanism to rotate forward to perform fanning; when the user enters the deep sleep stage, controlling the light module to turn off, and controlling the drive mechanism to rotate in reverse at a second wind speed to perform ventilation.
[0005] In a preferred embodiment of the present invention, the radar sensor is a 24GHz or 60GHz millimeter-wave radar, disposed on the outer peripheral surface of the housing and tilted downward toward the user, and the detection beam of the radar sensor covers the user's head and chest area; the sound sensor is a MEMS microphone, disposed adjacent to the radar sensor; the temperature sensor and the photosensitive sensor are disposed on the outer peripheral surface of the housing and spaced apart from the radar sensor to avoid electromagnetic interference from the radar sensor.
[0006] In a preferred embodiment of the present invention, the multimodal fusion sleep staging model is deployed on the edge AI processing unit, the feature encoding layer includes convolutional neural network branches corresponding to features of each modality, the cross-modal attention fusion layer includes a multi-head self-attention module and a feedforward network; the storage unit also stores the user's historical sleep records, and the edge AI processing unit is configured to fine-tune the multimodal fusion sleep staging model online based on the user's historical sleep records.
[0007] In a preferred embodiment of the present invention, step (a) of performing phase calculation on the respiratory waveform signal includes: performing range-dimensional FFT transformation on the intermediate frequency signal output by the radar sensor to determine the distance cell where the user's chest cavity is located; performing bandpass filtering on the phase signal within the distance cell; performing peak detection on the filtered phase signal; extracting the respiratory cycle sequence; and calculating the respiratory frequency and respiratory rhythm variability index based on the respiratory cycle sequence.
[0008] In a preferred embodiment of the present invention, the lighting module includes an upward-emitting module disposed on the top of the housing and surrounding the outside of the base, and a downward-emitting module disposed at the bottom of the fan blade; The housing is provided with a rotating seat, and the upward-emitting module is disposed on the rotating seat; the control module is also configured to: control the rotating seat to drive the upward-emitting module to rotate according to the sleep stage; when the user is in the sleep stage, the upward-emitting module is rotated to generate dynamic light effects to guide the user to fall asleep; when the user enters the light sleep stage, the rotation is stopped to reduce light stimulation.
[0009] In a preferred embodiment of the present invention, the driving mechanism includes a fixed shaft disposed at the bottom of the base and a rotor portion rotatably connected to the fixed shaft, and the fan blades are disposed on the bottom surface of the rotor portion; the housing is provided with a rotating seat for assembling the upward-emitting module and a transmission mechanism disposed between the driving mechanism and the rotating seat and capable of driving the rotating seat to rotate; the transmission mechanism includes a transmission seat rotatably connected to the fixed shaft and drivenly connected to the rotor portion, and a transmission assembly disposed between the transmission seat and the rotating seat.
[0010] In a preferred embodiment of the present invention, the transmission base includes a support shaft rotatably connected to the fixed shaft, and a base plate disposed at the bottom of the support shaft and drivenly connected to the rotor; the transmission assembly includes an upper support plate rotatably connected to the upper end of the support shaft, a lower support plate rotatably connected to the lower end of the support shaft, and a gear assembly disposed between the upper support plate and the lower support plate; the gear assembly includes a first gear fixed to the support shaft and a plurality of second gears meshing with the first gear, the upper support plate and the lower support plate being respectively disposed on the upper and lower sides of the second gear, so that the second gear drives the upper support plate and the lower support plate to rotate when revolving around the first gear; the rotating seat is disposed on the upper support plate; An electromagnetic clutch for controlling the power transmission between the base plate and the support shaft is provided between the base plate and the support shaft. The control module is configured to: when the multimodal fusion sleep staging model outputs that the user is in the sleep stage, control the electromagnetic clutch to engage, causing the rotor to sequentially drive the base plate, the support shaft, the transmission assembly, the rotating seat, and the upward-emitting module to rotate; when the multimodal fusion sleep staging model outputs that the user has entered the deep sleep stage, control the electromagnetic clutch to disengage, the upward-emitting module to stop rotating, and the fan blades to continue running at the second wind speed under the drive of the rotor.
[0011] In a preferred embodiment of the present invention, the top of the housing and the outer side surrounding the base body are provided with a downwardly recessed area. The upward-emitting module includes an annular lamp plate disposed in the recessed area, a plurality of low-power lamp beads disposed on the annular lamp plate, and a light-transmitting shell covering the low-power lamp beads. Each fan blade has a groove at the bottom of its end connected to the drive mechanism. When each fan blade is installed on the rotor, the grooves are connected end to end to form an annular lamp groove. The downward-emitting module includes a lamp strip disposed in the annular lamp groove and a lamp cover covering the lamp strip.
[0012] In a preferred embodiment of the present invention, the control module is further configured to: execute a wake-up strategy when the timing unit reaches a preset wake-up time, or when the multimodal fusion sleep staging model outputs that the user has transitioned from deep sleep to light sleep and the current time is within a preset wake-up window; the wake-up strategy includes: controlling the color temperature of the light module to gradually increase and the brightness to gradually increase from the lowest value to 50% within 15-30 minutes according to a preset gradual brightening curve; the multimodal fusion sleep staging model dynamically adjusts the slope of the gradual brightening curve based on the trend of changes in the user's body movement signals during the wake-up process.
[0013] This invention also provides a control method for an AI-powered smart sleep-aid fan light based on multimodal perception, employing the fan light described above, including: S1. Non-contact acquisition of user's breathing waveform signals and body movement signals through radar sensors, acquisition of environmental noise and user snoring signals through sound sensors, acquisition of ambient temperature through temperature sensors, and acquisition of ambient illuminance through photosensitive sensors. S2. Perform phase calculation on the respiratory waveform signal to extract respiratory frequency and respiratory rhythm variability features; perform time-domain feature extraction on the body movement signal to obtain body movement intensity and body movement frequency; perform spectrum analysis on the signal collected by the sound sensor to extract snoring event features; S3. After aligning the breathing features, body movement features, snoring features, and ambient temperature features and ambient light intensity features extracted in step S2 with time, input them into a pre-trained multimodal fusion sleep staging model. The multimodal fusion sleep staging model includes a feature encoding layer and a cross-modal attention fusion layer, and outputs the user's current sleep stage. S4. Generate a light-wind coordinated control strategy based on the sleep stage; S5. Control the lighting module and drive mechanism according to the light-wind coordinated control strategy. When the user is in the sleep stage, control the color temperature of the lighting module to gradually decrease and the brightness to decrease according to the preset dimming curve. At the same time, control the drive mechanism to rotate forward to provide ventilation and control the rotating seat to drive the upward light-emitting module to rotate. When the user enters the deep sleep stage, control the lighting module to turn off, control the drive mechanism to reverse to provide ventilation at the second wind speed, and control the rotating seat to stop rotating. S6. Collect respiratory characteristics, body movement characteristics, and sleep stage data after execution, update the user's historical sleep profile, and fine-tune the multimodal fusion sleep staging model online based on the updated user historical sleep profile.
[0014] The beneficial effects of this invention are: the AI smart sleep-aid fan light achieves personalized and adaptive sleep-aid control through non-contact multimodal perception and edge AI sleep staging; and through the corresponding light-wind collaborative control strategy, it can provide gradually dimming light and natural wind during the sleep-on period, and turn off the light and maintain low wind speed ventilation during the deep sleep period to avoid light and wind interference. Attached Figure Description
[0015] Figure 1 This is an exploded view of the present invention; Figure 2 This is an exploded view of the fan blades and the lower light-emitting module in this invention; Figure 3 This is an exploded view of the rotating seat, transmission mechanism, and upper light-emitting module in the second embodiment of the present invention; Figure 4 This is a framework diagram of the control method of the present invention. Detailed Implementation
[0016] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.
[0017] Reference Figures 1 to 3 An AI-powered smart sleep-aid fan light based on multimodal perception includes a base 1, a housing 2, a drive mechanism 3, fan blades 4, a light module 5, a multimodal perception module 6, and a control module.
[0018] The base 1 is used to fix it to a wall or ceiling, and can be a ceiling mount or a suspended mount. The housing 2 is located at the bottom of the base 1, is cylindrical and has an internal receiving space. The drive mechanism 3 is located at the bottom of the base 1 and is located within the receiving space of the housing. The drive mechanism 3 includes a fixed shaft 31 and a rotor 32. The fixed shaft 31 is fixed to the bottom of the base 1, and the rotor 32 is rotatably connected to the fixed shaft 31. The fan blade 4 is provided with several blades, for example, 3 to 5 blades, which are evenly arranged on the bottom surface of the rotor 32.
[0019] The lighting module 5 includes an upward-emitting module 51 and a downward-emitting module 52. The upward-emitting module 51 is located on the top of the housing 2 and surrounds the outer side of the base 1. The top of the housing 2 and the outer side of the base 1 have a downward-recessed area 21. The upward-emitting module 51 includes an annular lamp plate 511 located in the recessed area 21, several low-power LED beads 512 located on the annular lamp plate 511, and a light-transmitting shell 513 covering the low-power LED beads 512. The low-power LED beads 512 are used to emit warm-colored light that helps promote sleep. The downward-emitting module 52 is located at the bottom of the fan blades 4. Each fan blade 4 has a groove 41 at the bottom of the end connected to the drive mechanism 3. When each fan blade 4 is installed on the rotor part 32, the grooves 41 are connected end to end to form an annular lamp groove 42. The downward-emitting module 52 includes a light strip 521 located in the annular lamp groove 42 and a lamp cover 522 covering the light strip 521.
[0020] The multimodal sensing module 6 includes a radar sensor 61, a temperature sensor 62, a sound sensor 63, and a photosensor 64. The radar sensor 61 is a 24GHz or 60GHz millimeter-wave radar, disposed on the outer peripheral surface of the housing 2 and facing the user. Its detection beam covers the user's head and chest area. The radar sensor 61 is used for non-contact acquisition of the user's breathing waveform and body movement signals. The sound sensor 63 is a MEMS microphone, disposed adjacent to the radar sensor 61, used to acquire ambient noise and the user's snoring signals. The temperature sensor 62 is used to acquire ambient temperature, and the photosensor 64 is used to acquire ambient light intensity. The temperature sensor 62 and the photosensor 64 are disposed on the outer peripheral surface of the housing 2 and spaced apart from the radar sensor 61 to avoid electromagnetic interference from the radar sensor 61.
[0021] The control module is located inside the seat 1 and includes an edge AI processing unit, a remote communication unit, a timing unit, and a storage unit. The edge AI processing unit is electrically connected to the multimodal perception module 6, the lighting module 5, and the drive mechanism 3. The remote communication unit can use at least one of WiFi, Bluetooth, ZigBee, or NB-IoT. The timing unit can use an RTC real-time clock chip. The storage unit can use a Flash memory to store the user's historical sleep records, multimodal fusion sleep staging model parameters, and control strategies.
[0022] As a second embodiment of this solution, the housing 2 is provided with a rotating seat 7 and a transmission mechanism 8. The rotating seat 7 is used to assemble the upward light-emitting module 51. The upward light-emitting module 51 is disposed on the rotating seat 7. The transmission mechanism 8 is disposed between the driving mechanism 3 and the rotating seat 7, and can drive the rotating seat 7 to rotate.
[0023] The transmission mechanism 8 includes a transmission base 81 and a transmission component 82. The transmission base 81 is rotatably connected to the fixed shaft 31 and drivenly connected to the rotor part 32. The transmission component 82 is disposed between the transmission base 81 and the rotating base 7.
[0024] The transmission base 81 includes a support shaft 811 and a base plate 812. The support shaft 811 is rotatably connected to a fixed shaft 31. The base plate 812 is disposed at the bottom of the support shaft 811 and is drivenly connected to the rotor part 32. An electromagnetic clutch 9 is provided between the base plate 812 and the support shaft 811. The electromagnetic clutch 9 is used to control the power transmission between the base plate 812 and the support shaft 811. When the electromagnetic clutch 9 is engaged, the base plate 812 and the support shaft 811 rotate synchronously; when the electromagnetic clutch 9 is disengaged, the base plate 812 and the support shaft 811 can rotate relative to each other.
[0025] The transmission assembly 82 includes an upper support plate 821, a lower support plate 822, and a gear assembly 823. The upper support plate 821 is rotatably connected to the upper end of the support shaft 811, and the lower support plate 822 is rotatably connected to the lower end of the support shaft 811. The gear assembly 823 is disposed between the upper support plate 821 and the lower support plate 822. The gear assembly 823 includes a first gear 8231 fixed to the support shaft 811 and a plurality of second gears 8232 meshing with the first gear 8231. The upper support plate 821 and the lower support plate 822 are respectively disposed on the upper and lower sides of the second gear 8232, and the second gear 8232 is connected to the upper support plate 821 and the lower support plate 822 respectively through gear shafts. When the support shaft 811 rotates, the first gear 8231 rotates accordingly, driving the second gear 8232 to revolve around the first gear 8231, which in turn drives the upper support plate 821 and the lower support plate 822 to rotate. The rotating seat 7 is set on the upper support plate 821, so the rotating seat 7 rotates with the upper support plate 821, thereby driving the upward light-emitting module 51 to rotate.
[0026] The control module is configured to: when the user is in the sleep stage according to the multimodal fusion sleep staging model, control the electromagnetic clutch 9 to engage, causing the rotor 32 to sequentially drive the base plate 812, support shaft 811, first gear 8231, second gear 8232, upper support plate 821, rotating seat 7, and upward-emitting light module 51 to rotate; when the user enters the deep sleep stage according to the multimodal fusion sleep staging model, control the electromagnetic clutch 9 to disengage, the upward-emitting light module 51 to stop rotating, and the fan blades 4 continue to run at the second wind speed under the drive of the rotor 32. This achieves independent control of the light rotation and fan blade rotation, stopping the light rotation during deep sleep to reduce light stimulation and mechanical noise.
[0027] As a third embodiment of this solution, the edge AI processing unit is configured to perform the following steps: (a) Phase calculation is performed on the respiratory waveform signal acquired by radar sensor 61. Specifically, a range-dimensional FFT transform is performed on the intermediate frequency signal output by radar sensor 61 to determine the range cell where the user's chest cavity is located. The phase signal within the range cell is bandpass filtered with a filtering frequency band of 0.1-0.5Hz. Peak detection is performed on the filtered phase signal to extract the respiratory cycle sequence. The respiratory rate and respiratory rhythm variability index are calculated based on the respiratory cycle sequence. The respiratory rate is the reciprocal of the respiratory cycle per unit time. The respiratory rhythm variability index can be the standard deviation or root mean square of the difference between adjacent respiratory cycles.
[0028] Time-domain feature extraction is performed on the body motion signal to obtain the body motion intensity and frequency. The body motion intensity can be expressed as the integral of the signal amplitude per unit time, and the body motion frequency can be expressed as the number of body motion events per unit time.
[0029] The signal collected by the sound sensor 63 is subjected to spectral analysis to extract snoring event features. Specifically, a short-time Fourier transform is performed on the sound signal to extract spectral features. A pre-trained snoring classifier is used to determine whether snoring events exist, and the frequency and duration of snoring events are statistically analyzed.
[0030] (b) After aligning the respiratory features, body movement features, snoring features, and ambient temperature and ambient light intensity features extracted in step (a) with time, input them into the pre-trained multimodal fusion sleep staging model. Time alignment can be achieved using a sliding time window, for example, with a window length of 30 seconds, aligning each modality feature to the same timestamp.
[0031] The multimodal sleep staging model comprises a feature encoding layer and a cross-modal attention fusion layer. The feature encoding layer independently encodes features for each modality. For example, respiratory and body movement features can utilize a one-dimensional convolutional neural network branch, sound features a two-dimensional convolutional neural network branch, and temperature and light features a fully connected network branch. The cross-modal attention fusion layer includes a multi-head self-attention module and a feedforward network. The multi-head self-attention module calculates the correlation weights between features of each modality through an attention mechanism and performs weighted fusion. For instance, when respiratory rhythm variability increases and body movement frequency increases, the attention mechanism increases the weights of the respiratory and body movement modalities, thereby more accurately determining sleep stages.
[0032] The multimodal fusion sleep staging model is deployed on an edge AI processing unit with no more than 500KB of parameters and a single inference latency of no more than 50ms. The model can be deployed using TensorFlow Lite or PyTorch Mobile.
[0033] (c) The multimodal fusion sleep staging model outputs the user's current sleep stage. Sleep stages include at least three of the following: wakefulness, sleep onset, light sleep, deep sleep, and REM sleep. For example, the model output is a five-dimensional vector, each corresponding to a probability value for one of the five sleep stages, with the category with the highest probability being taken as the current sleep stage.
[0034] (d) Generate a light-wind coordinated control strategy based on the sleep stage. The light-wind coordinated control strategy includes: when the user is in the sleep stage, controlling the color temperature of the light module 5 to gradually decrease from a first color temperature value to a second color temperature value, and the brightness to gradually decrease from a first brightness value according to a preset dimming curve, while controlling the drive mechanism 3 to rotate forward to provide ventilation. When the drive mechanism 3 rotates forward, the fan blades 4 blow air downwards to achieve the ventilation effect.
[0035] When the user enters a deep sleep state, the control light module 5 turns off, and the control drive mechanism 3 reverses to provide ventilation at a second wind speed. This second wind speed is lower than the first wind speed during sleep. When the drive mechanism 3 reverses, the fan blades 4 draw air upwards or create a low-speed air circulation, achieving ventilation while avoiding direct airflow onto the user.
[0036] Furthermore, the light-wind coordinated control strategy also includes: when the multimodal fusion sleep staging model outputs the transition phase from light sleep to deep sleep, the wind speed of the control drive mechanism 3 decreases at a first preset rate, while the brightness of the light module 5 decreases at a second preset rate. The first preset rate is greater than the second preset rate, so as to reduce wind interference before the light is completely turned off. When the multimodal fusion sleep staging model outputs the transition phase from deep sleep to REM sleep, the control drive mechanism 3 switches to an intermittent light breeze mode, with an on / off cycle of 30-120 seconds.
[0037] As a fourth embodiment of this solution, the storage unit also stores the user's historical sleep records, and the edge AI processing unit is configured to perform online fine-tuning of the multimodal fusion sleep staging model based on the user's historical sleep records. The online fine-tuning can employ an incremental learning approach. For example, the user's sleep data for the current night can be used as new samples, the output of the multimodal fusion sleep staging model can be used as pseudo-labels, and a low learning rate can be used for small-scale iterative updates, allowing the model to gradually adapt to the user's individual characteristics.
[0038] The control module is also configured to execute a wake-up strategy. When the timing unit reaches a preset wake-up time, or when the multimodal fusion sleep staging model outputs that the user has transitioned from deep sleep to light sleep and the current time falls within a preset wake-up window, the wake-up strategy is executed. The wake-up strategy includes: gradually increasing the color temperature of the light module 5 and gradually increasing the brightness from its lowest value to 50% within 15-30 minutes according to a preset gradual brightening curve. The multimodal fusion sleep staging model dynamically adjusts the slope of the gradual brightening curve based on the user's body movement signal changes during the wake-up process. For example, if the body movement signal gradually increases, the brightness increase rate is accelerated; if the body movement signal shows no significant change, the brightness increase rate is slowed down.
[0039] Reference Figure 4 The present invention also discloses a control method for an AI-powered smart sleep-aid fan light based on multimodal perception, used in the aforementioned fan light, comprising: S1. The user's breathing waveform signal and body movement signal are collected non-contactly through radar sensor 61, the ambient noise and user snoring signal are collected through sound sensor 63, the ambient temperature is collected through temperature sensor 62, and the ambient light intensity is collected through photosensitive sensor 64.
[0040] S2. Perform phase calculation on the respiratory waveform signal to extract respiratory frequency and respiratory rhythm variability features; extract time-domain features from the body movement signal to obtain body movement intensity and body movement frequency; perform spectrum analysis on the signal collected by the sound sensor 63 to extract snoring event features.
[0041] S3. After aligning the breathing features, body movement features, snoring features, and ambient temperature and ambient light intensity features extracted in step S2 with time, input them into a pre-trained multimodal fusion sleep staging model. The multimodal fusion sleep staging model includes a feature encoding layer and a cross-modal attention fusion layer, and outputs the user's current sleep stage.
[0042] S4. Generate a light-wind coordinated control strategy based on sleep stages.
[0043] S5. Control the lighting module 5 and the drive mechanism 3 according to the light-wind coordinated control strategy. Specifically, when the user is in the sleep stage, the color temperature of the lighting module 5 is gradually reduced and the brightness is reduced according to the preset dimming curve. At the same time, the drive mechanism 3 is controlled to rotate forward to provide ventilation, and the rotating seat 7 is controlled to drive the upward-emitting module 51 to rotate. When the user enters the deep sleep stage, the lighting module 5 is controlled to turn off, the drive mechanism 3 is controlled to reverse to provide ventilation at the second wind speed, and the rotating seat 7 is controlled to stop rotating.
[0044] S6. Collect respiratory characteristics, body movement characteristics, and sleep stage data after execution, update the user's historical sleep profile, and fine-tune the multimodal fusion sleep staging model online based on the updated user historical sleep profile.
[0045] Furthermore, the training method for the multimodal fusion sleep staging model includes: collecting radar respiratory signals, body movement signals, sound signals, temperature data, and illumination data from multiple users as training samples, and using sleep stages labeled with a polysomnography map as training labels. The training samples undergo time alignment and normalization preprocessing. The multimodal fusion sleep staging model is trained using a cross-entropy loss function, and a domain adversarial training mechanism is used to transfer respiratory signal feature knowledge collected by contact sensors to respiratory signal features collected by radar sensors. Domain adversarial training can employ a gradient inversion layer to enable the feature encoding layer to extract domain-invariant features.
[0046] Furthermore, the control method also includes communication with a mobile app or cloud server via a remote communication unit to achieve remote control, sleep report viewing, control strategy updates, and OTA upgrades. The cloud server receives anonymized sleep data from multiple fan lights, performs federated learning updates on the multimodal fusion sleep staging model, and distributes the updated model parameters to the edge AI processing units of each fan light. Federated learning can employ the FedAvg algorithm, where each fan light encrypts and uploads its model gradients after local training, and the cloud aggregates and distributes the global model.
[0047] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.
Claims
1. An AI-powered smart sleep-aid fan light based on multimodal perception, comprising a base (1) for fixing to a wall, a housing (2) disposed at the bottom of the base (1), a drive mechanism (3) disposed at the bottom of the base (1) and located within the housing (2), and a plurality of fan blades (4) disposed at the bottom of the drive mechanism (3), characterized in that, It also includes a lighting module (5), a multimodal sensing module (6), and a control module; The multimodal sensing module (6) includes a radar sensor (61), a temperature sensor (62), a sound sensor (63), and a photosensitive sensor (64) disposed on the outer periphery of the housing (2). The radar sensor (61) is used to collect the user's breathing waveform signal and body movement signal in a non-contact manner. The sound sensor (63) is used to collect the ambient noise and the user's snoring signal. The temperature sensor (62) is used to collect the ambient temperature. The photosensitive sensor (64) is used to collect the ambient light intensity. The control module is located inside the seat (1) and includes an edge AI processing unit, a remote communication unit, a timing unit and a storage unit, and is electrically connected to the multimodal perception module (6), the lighting module (5) and the driving mechanism (3) respectively. The edge AI processing unit is configured to perform the following steps: (a) Perform phase calculation on the respiratory waveform signal collected by the radar sensor (61), extract respiratory frequency and respiratory rhythm variability features, extract time domain features of the body movement signal to obtain body movement intensity and body movement frequency, and perform spectrum analysis on the signal collected by the sound sensor (63) to extract snoring event features; (b) After aligning the breathing features, body movement features, snoring features, and ambient temperature features and ambient light intensity features extracted in step (a) with time, input them into a pre-trained multimodal fusion sleep staging model. The multimodal fusion sleep staging model includes a feature encoding layer and a cross-modal attention fusion layer. The feature encoding layer encodes each modality feature independently, and the cross-modal attention fusion layer calculates the correlation weights between each modality feature through an attention mechanism and performs weighted fusion. (c) The multimodal fusion sleep staging model outputs the user's current sleep stage, which includes at least three of the following: wakefulness, sleep onset, light sleep, deep sleep, and REM sleep. (d) Generate a light-wind coordinated control strategy according to the sleep stage. The light-wind coordinated control strategy includes: when the user is in the sleep stage, control the color temperature of the light module (5) to gradually decrease from the first color temperature value to the second color temperature value, and control the brightness to gradually decrease from the first brightness value according to the preset dimming curve. At the same time, control the drive mechanism (3) to rotate forward to fan the air. When the user enters the deep sleep stage, control the light module (5) to turn off, and control the drive mechanism (3) to reverse to ventilate at the second wind speed.
2. The AI-powered smart sleep-aid fan light based on multimodal perception according to claim 1, characterized in that, The radar sensor (61) is a 24GHz or 60GHz millimeter-wave radar, which is set on the outer peripheral surface of the housing (2) and tilted downwards towards the user. The detection beam of the radar sensor (61) covers the user's head and chest area. The sound sensor (63) is a MEMS microphone, which is set adjacent to the radar sensor (61). The temperature sensor (62) and the photosensitive sensor (64) are set on the outer peripheral surface of the housing (2) and arranged at intervals from the radar sensor (61) to avoid electromagnetic interference from the radar sensor (61).
3. The AI-powered smart sleep-aid fan light based on multimodal perception according to claim 1, characterized in that, The multimodal fusion sleep staging model is deployed on the edge AI processing unit. The feature encoding layer includes convolutional neural network branches corresponding to the features of each modality. The cross-modal attention fusion layer includes a multi-head self-attention module and a feedforward network. The storage unit also stores the user's historical sleep records. The edge AI processing unit is configured to fine-tune the multimodal fusion sleep staging model online based on the user's historical sleep records.
4. The AI-powered smart sleep-aid fan light based on multimodal perception according to claim 1, characterized in that, Step (a) involves performing phase calculation on the respiratory waveform signal, which includes: performing a range-dimensional FFT transformation on the intermediate frequency signal output by the radar sensor (61) to determine the distance cell where the user's chest cavity is located; performing bandpass filtering on the phase signal within the distance cell; performing peak detection on the filtered phase signal; extracting the respiratory cycle sequence; and calculating the respiratory frequency and respiratory rhythm variability index based on the respiratory cycle sequence.
5. An AI-powered smart sleep-aid fan light based on multimodal perception according to any one of claims 1 to 4, characterized in that, The lighting module (5) includes an upward-emitting module (51) disposed on the top of the housing (2) and surrounding the outside of the base (1) and a downward-emitting module (52) disposed at the bottom of the fan blade (4). The housing (2) is provided with a rotating seat (7), and the upward light-emitting module (51) is disposed on the rotating seat (7); the control module is also configured to: control the rotating seat (7) to drive the upward light-emitting module (51) to rotate according to the sleep stage; when the user is in the sleep stage, the upward light-emitting module (51) is rotated to generate dynamic light effects to guide the user to fall asleep; when the user enters the light sleep stage, the rotation is stopped to reduce light stimulation.
6. The AI-powered smart sleep-aid fan light based on multimodal perception according to claim 5, characterized in that, The drive mechanism (3) includes a fixed shaft (31) disposed at the bottom of the base (1) and a rotor (32) rotatably connected to the fixed shaft (31). The fan blade (4) is disposed on the bottom surface of the rotor (32). The housing (2) is provided with a rotating seat (7) for assembling the upward light-emitting module (51) and a transmission mechanism (8) disposed between the drive mechanism (3) and the rotating seat (7) and capable of driving the rotating seat (7) to rotate. The transmission mechanism (8) includes a transmission seat (81) rotatably connected to the fixed shaft (31) and drivenly connected to the rotor (32), and a transmission assembly (82) disposed between the transmission seat (81) and the rotating seat (7).
7. The AI-powered smart sleep-aid fan light based on multimodal perception according to claim 6, characterized in that, The transmission base (81) includes a support shaft (811) rotatably connected to the fixed shaft (31), and a base plate (812) disposed at the bottom of the support shaft (811) and drivenly connected to the rotor (32); the transmission assembly (82) includes an upper support plate (821) rotatably connected to the upper end of the support shaft (811), a lower support plate (822) rotatably connected to the lower end of the support shaft (811), and a gear assembly (823) disposed between the upper support plate (821) and the lower support plate (822); the gear assembly ( 823) includes a first gear (8231) fixed on the support shaft (811) and a plurality of second gears (8232) meshing with the first gear (8231). The upper support plate (821) and the lower support plate (822) are respectively disposed on the upper and lower sides of the second gear (8232) so that the second gear (8232) drives the upper support plate (821) and the lower support plate (822) to rotate when the second gear (8232) revolves around the first gear (8231); the rotating seat (7) is disposed on the upper support plate (821); An electromagnetic clutch (9) for controlling the power transmission between the base plate (812) and the support shaft (811) is provided between the base plate (812) and the support shaft (811). The control module is configured to: when the multimodal fusion sleep staging model outputs that the user is in the sleep stage, control the electromagnetic clutch (9) to engage, so that the rotor (32) sequentially drives the base plate (812), the support shaft (811), the transmission assembly (82), the rotating seat (7) and the upward light-emitting module (51) to rotate; when the multimodal fusion sleep staging model outputs that the user enters the deep sleep stage, control the electromagnetic clutch (9) to disengage, the upward light-emitting module (51) to stop rotating, and the fan blade (4) continues to run at the second wind speed under the drive of the rotor (32).
8. The AI-powered smart sleep-aid fan light based on multimodal perception according to claim 6, characterized in that, The top of the housing (2) and the outer side of the base (1) are provided with a recessed area (21) that is recessed downwards. The upward light-emitting module (51) includes an annular lamp plate (511) disposed in the recessed area (21), a plurality of low-power lamp beads (512) disposed on the annular lamp plate (511), and a light-transmitting shell (513) covering the low-power lamp beads (512). Each fan blade (4) has a groove (41) at the bottom of the end connected to the drive mechanism (3). When each fan blade (4) is installed on the rotor (32), the grooves (41) are connected end to end to form an annular lamp groove (42). The downward light-emitting module (52) includes a lamp strip (521) disposed in the annular lamp groove (42) and a lamp cover (522) covering the lamp strip (521).
9. The AI-powered smart sleep-aid fan light based on multimodal perception according to claim 1, characterized in that, The control module is further configured to execute a wake-up strategy when the timing unit reaches the preset wake-up time, or when the multimodal fusion sleep staging model outputs that the user has entered the light sleep stage from the deep sleep stage and the current time is within the preset wake-up window; the wake-up strategy includes: controlling the color temperature of the light module (5) to gradually increase and the brightness to gradually increase from the lowest value to 50% within 15-30 minutes according to the preset gradual brightening curve; the multimodal fusion sleep staging model dynamically adjusts the slope of the gradual brightening curve based on the change trend of the user's body movement signal during the wake-up process.
10. A control method for an AI-powered smart sleep-aid fan light based on multimodal perception, employing the fan light as described in any one of claims 5 to 9, characterized in that, include: S1. The user's breathing waveform signal and body movement signal are collected non-contactly by radar sensor (61), the ambient noise and user snoring signal are collected by sound sensor (63), the ambient temperature is collected by temperature sensor (62), and the ambient light intensity is collected by photosensitive sensor (64). S2. Perform phase calculation on the respiratory waveform signal to extract respiratory frequency and respiratory rhythm variability features; The body movement signal is subjected to time-domain feature extraction to obtain body movement intensity and frequency; the signal collected by the sound sensor (63) is subjected to spectrum analysis to extract snoring event features; S3. After aligning the breathing features, body movement features, snoring features, and ambient temperature features and ambient light intensity features extracted in step S2 with time, input them into a pre-trained multimodal fusion sleep staging model. The multimodal fusion sleep staging model includes a feature encoding layer and a cross-modal attention fusion layer, and outputs the user's current sleep stage. S4. Generate a light-wind coordinated control strategy based on the sleep stage; S5. Control the light module (5) and the drive mechanism (3) according to the light-wind coordinated control strategy. When the user is in the sleep stage, control the color temperature of the light module (5) to gradually decrease and the brightness to decrease according to the preset dimming curve. At the same time, control the drive mechanism (3) to rotate forward to fan the air and control the rotating seat (7) to drive the upward light-emitting module (51) to rotate. When the user enters the deep sleep stage, control the light module (5) to turn off, control the drive mechanism (3) to reverse at the second wind speed to ventilate, and control the rotating seat (7) to stop rotating. S6. Collect respiratory characteristics, body movement characteristics, and sleep stage data after execution, update the user's historical sleep profile, and fine-tune the multimodal fusion sleep staging model online based on the updated user historical sleep profile.
