Integrated infant sleep and colic active intelligent intervention system and control method

Abstract

The application discloses an integrated infant sleep and intestinal distension active intelligent intervention system in the technical field of infant monitoring, which is used for monitoring and regulating an infant cabin, and comprises a multi-modal sensing module, an intelligent recognition processing module, an active intervention execution module and a safe interaction module.The multi-modal sensing module is used for collecting physiological state data of an infant in the infant cabin.The intelligent recognition processing module is used for receiving the physiological state data, recognizing an infant sleep state based on the physiological state data, and generating a control instruction.The active intervention execution module is used for executing the control instruction and regulating the infant cabin.The safe interaction module is used for feeding back the infant sleep state.Through the cooperative work of the multi-modal sensing module, the intelligent recognition processing module and the active intervention execution module, the active intervention execution module can automatically execute intestinal distension double-effect intervention and bionic lull intervention according to the recognition result, so that the intelligent level and intervention effect of infant monitoring are improved.

Description

Technical Field

[0001] This invention relates to the field of infant and toddler monitoring, specifically to an integrated intelligent intervention system and control method for infant sleep and colic. Background Technology

[0002] In infant and toddler home care, sleep safety monitoring and colic management are two high-frequency essential needs and core pain points. Infants aged 0-3 months have a high incidence of physiological colic, which can easily lead to persistent crying and sleep disturbances. At the same time, there are safety risks during infant sleep, such as mouth and nose obstruction, abnormal breathing, and suffocation from prone sleeping, placing higher demands on real-time monitoring and rapid response.

[0003] However, existing sleep monitoring systems suffer from functional fragmentation and limited monitoring methods: First, they are functionally limited, with most products only providing monitoring alarms or passive soothing functions, failing to simultaneously achieve sleep safety monitoring and active intervention for intestinal gas. Parents need to use multiple devices, resulting in cumbersome operation and a poor user experience. Second, monitoring methods are mainly wearable or contact-based, which are prone to falling off or shifting due to infant movement, leading to data interruption or distortion, and long-term use poses a risk of skin irritation. Third, they lack the ability to accurately identify and adaptively intervene in intestinal gas, failing to integrate multi-dimensional information such as respiratory rhythm, abdominal pressure, body movement frequency, and crying characteristics to automatically determine the state of intestinal gas. Intervention methods rely on manual activation, and the intensity and duration cannot be adaptively adjusted, resulting in unstable intervention effects. Summary of the Invention

[0004] The purpose of this invention is to provide an integrated active intelligent intervention system and control method for infant sleep and gas, which solves the problem that existing infant monitoring systems are unable to effectively intervene in sleep and gas.

[0005] The present invention achieves the above objectives through the following technical solutions: An integrated intelligent intervention system for infant sleep and gas control is used to monitor and regulate the infant pod. The system includes: A multimodal sensing module is used to collect physiological state data of infants and young children inside the infant cabin; The intelligent recognition and processing module is used to receive physiological state data and identify the infant's sleep state based on the physiological state data in order to generate control commands; An active intervention execution module is used to execute the control commands and regulate the baby cabin; A safe interactive module is used to provide feedback on the sleep status of infants and young children; The control commands include dual-effect intervention commands for intestinal gas and bionic sleep-inducing intervention commands; When the active intervention execution module executes the dual-effect intervention command for intestinal gas, it heats and vibrates the baby cabin through a heating film and a vibration motor to relieve intestinal gas; when it executes the bionic soothing intervention command, it performs a bionic soothing action on the baby cabin, the bionic soothing action including reciprocating movement of the baby cabin.

[0006] As a further optimization of the invention, the baby cabin includes a cabin body and an airbag sleeping pad assembly disposed inside the cabin body. The airbag sleeping pad assembly includes a head airbag, a waist and abdomen airbag, and a leg airbag. The head airbag is higher than the leg airbag, and the air pressure of the waist and abdomen airbag is lower than that of the head airbag and the leg airbag. The cabin body includes a main structure and an outer cover. The main structure is a hollow shell with an opening at the top, and the outer cover is engaged with the opening of the shell.

[0007] As a further optimization of the invention, the multimodal sensing module includes a physiological state monitoring component, which includes a radar mounted on the outer casing, a linear drive for moving the radar back and forth, a pitch adjustment component for adjusting the radar pitch angle, and a flexible pressure sensor mounted on the top of the waist and abdomen airbag.

[0008] As a further optimization of the invention, the airbag sleeping pad assembly also includes a contact layer, wherein the head airbag, waist and abdomen airbag and leg airbag are all disposed inside the contact layer, and the active intervention execution module includes a dual-effect intestinal bloating intervention mechanism, wherein the dual-effect intestinal bloating intervention mechanism includes a heating film disposed on the top of the contact layer and a vibration motor disposed inside the waist and abdomen airbag.

[0009] As a further optimization of the invention, the active intervention execution module also includes a bionic sleep-inducing intervention mechanism, which includes a base and a lifting component disposed between the base and the main structure.

[0010] A control method includes the following steps: Step S1: Collect physiological state data of infants and young children in the baby cabin based on the multimodal sensing module; Step S2: Receive physiological state data through the intelligent recognition and processing module, and identify the infant's sleep state based on the physiological state data to generate control commands; Step S3: Execute the control command and regulate the baby cabin through the active intervention execution module; Step S4: Feedback on the infant's sleep status based on the security interaction module; The control commands include a dual-effect intervention command for intestinal colic, a bionic soothing intervention command, and a safety warning intervention command; when the dual-effect intervention command for intestinal colic is executed, the baby cabin is heated and vibrated; when the bionic soothing intervention command is executed, the baby cabin is moved back and forth; when the safety warning intervention command is executed, an alarm is sounded and the baby cabin is vibrated when the infant's sleep state is abnormal.

[0011] As a further optimization of the invention, the physiological state data includes respiration, heart rate, body movement, body position, abdominal pressure fluctuations, and crying audio data; in step S1, the infant's respiration, heart rate, body movement, and body position data are collected based on radar; the infant's abdominal pressure fluctuations and body movement data are collected based on a flexible pressure sensor; and the crying audio data are collected based on a microphone.

[0012] As a further optimization of the invention, step S2 includes: Step S21: Preprocess the respiratory, heart rate, body movement and body position data, determine the infant's position by distance dimension FFT, extract phase data and unwrap to obtain chest wall displacement waveform, separate respiratory data and heart rate data, and calculate respiratory rate and heart rate in real time; Step S22: Time-align the data of abdominal pressure fluctuations, crying audio, and body movement, and splice them together to form a multi-dimensional temporal feature vector; Step S23: Input the multidimensional temporal feature vector into the pre-trained CNN+LSTM fusion network model and output the classification result of the infant's current state.

[0013] As a further optimization of the invention, the classification results include respiratory state categories and intestinal distension state categories; the fusion network model includes CNN layers, LSTM layers, and fully connected layers; the CNN layer adopts a two-layer one-dimensional convolutional structure to extract local discriminative features of the temporal feature vector; the LSTM layer adopts a single-layer or two-layer structure to construct a temporal evolution law model of the state; the fully connected layer is used to combine the Softmax activation function to output the confidence scores of the respiratory state category and the intestinal distension state category.

[0014] As a further optimization of the invention, step S2 further includes: step S24, generating corresponding control instructions based on the classification results, and performing transfer learning fine-tuning; the transfer learning fine-tuning includes the following steps: The parameters of the CNN and LSTM layers of the fusion network model are retained, and the fully connected layers of the fusion network model are replaced. The intervention data collected within a set time period during system deployment is used as new samples to gradually unfreeze the high-level parameters of the fusion network model, perform small-sample fine-tuning, and update the fusion network model. After a set time period, the automatically accumulated intervention data will be used as new samples for incremental learning, continuously updating the fusion network model.

[0015] The beneficial effects of this invention are as follows: 1) This invention breaks through the limitations of traditional monitoring and intervention separation by working together a multimodal perception module, an intelligent recognition and processing module, and an active intervention execution module. The system can not only identify sleep abnormalities based on physiological state data, but also automatically generate and execute dual-effect intervention instructions for intestinal gas and bionic soothing intervention instructions based on the recognition results. It realizes closed-loop control of the entire process of "perception-recognition-decision-execution", which improves the intelligence level and real-time intervention of infant care. 2) This invention adopts a sensing scheme that combines non-contact millimeter-wave radar with flexible pressure sensors. It can accurately collect multi-dimensional physiological data such as respiration, heart rate, body movement, body position and abdominal pressure fluctuations without the need for the baby to wear any devices. This unrestricted monitoring method not only avoids the stimulation and discomfort of traditional wearable devices to the baby's skin, but also overcomes the influence of clothing and body position on the monitoring signal, significantly improving the accuracy of physiological state data. 3) This invention utilizes a CNN+LSTM fusion network model to perform in-depth analysis of multidimensional temporal features, which can accurately identify states such as intestinal gas and abnormal breathing. Based on this, the system automatically matches different intervention strategies according to the identification results, and dynamically adjusts parameters such as hot compress temperature, vibration frequency, and bionic soothing mode based on closed-loop feedback mechanism and reinforcement learning. This intervention method can form the optimal personalized care plan for the individual differences of different infants, improve the intervention effect and avoid over-intervention. 4) This invention employs comprehensive safety protection measures at the structural, electrical, material, and software levels. For example, the cabin adopts a fully curved, edgeless design to prevent scratches; the lifting mechanism incorporates mechanical safety mechanisms such as anti-pinch and automatic locking upon power failure; the heating module is equipped with precise temperature control and timer protection; and the millimeter-wave radar radiation power density meets the national first-class health standard. These designs fundamentally solve the safety hazards existing in current equipment, providing infants with a safe and comfortable sleeping and care environment. Attached Figure Description

[0016] Figure 1 This is the overall flowchart of the present invention; Figure 2 This is a system block diagram of the present invention; Figure 3 This is a schematic diagram of the external structure of the baby compartment of the present invention; Figure 4 This is a schematic diagram of the internal structure of the baby compartment of the present invention; Figure 5 This is a schematic diagram of the structure of the airbag sleeping pad assembly of the present invention; Figure 6 This is a schematic diagram of the main structure of the baby compartment of the present invention; Figure 7 for Figure 6 Enlarged view of point A in the middle; Figure 8 This is a flowchart illustrating the state classification process of the present invention. In the diagram: 1. Cabin; 2. Airbag sleeping pad assembly; 3. Physiological status monitoring assembly; 4. Dual-effect intestinal gas intervention mechanism; 5. Bionic sleep-inducing intervention mechanism; 6. Alarm; 7. Display; 8. Emergency stop button; 11. Main structure; 12. Outer cover; 13. Baffle; 14. T-connector; 15. Air valve; 16. Air pressure transmitter; 21. Head airbag; 22. Waist and abdomen airbag; 23. Leg airbag; 24. Contact layer; 25. Buffer layer; 26. Load-bearing frame; 27. Foot pad; 31. Radar; 32. Linear drive component; 33. Pitch adjustment component; 34. Pressure sensor; 41. Heating film; 42. Waterproof and breathable membrane; 43. Vibration motor; 44. Flexible sound insulation sleeve; 45. Electrical feeder; 51. Base; 52. Lifting component. Detailed Implementation

[0017] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0018] First Embodiment like Figures 1 to 6 As shown, this embodiment relates to an integrated active intelligent intervention system for infant sleep and colic, employing a five-layer architecture of perception-transmission-processing-execution-interaction. It combines multimodal sensing, intelligent algorithms, and flexible mechanical structures to achieve automated monitoring of infants throughout the entire process. The active intelligent intervention system monitors and regulates the infant cabin, which consists of a cabin body 1 and an airbag sleeping pad assembly 2, for infant sleep. The system includes a multimodal perception module, an intelligent recognition and processing module, an active intervention execution module, and a safety interaction module.

[0019] The multimodal perception module includes a physiological state monitoring component 3 and a microphone, used to collect physiological state data of infants in the baby cabin; the intelligent recognition and processing module is an industrial control computer, used to receive physiological state data and identify the infant's sleep state based on the physiological state data to generate control commands; the active intervention execution module includes a dual-effect gas intervention mechanism 4 and a bionic soothing intervention mechanism 5 and an alarm 6, used to execute control commands to regulate the baby cabin; the safety interaction module includes a display 7 and an emergency stop button 8, used to provide feedback on the infant's sleep state to the caregiver.

[0020] like Figure 3 As shown, the overall structure of the aforementioned baby cabin is as follows: The infant cabin includes a cabin body 1 and an airbag sleeping pad assembly 2 disposed inside the cabin body 1. The airbag sleeping pad assembly 2 includes a head airbag 21, a waist and abdomen airbag 22, and a leg airbag 23. The head airbag 21 is higher than the leg airbag 23, and the tangent between the top of the head airbag 21 and the top of the leg airbag 23 forms a natural 10° angle. The overall design conforms to ergonomics and helps promote intestinal peristalsis in infants. The air pressure of the waist and abdomen airbag 22 is lower than that of the head airbag 21 and the leg airbag 23, so that the contact surface between the airbag sleeping pad assembly 2 and the infant is C-shaped. The cabin body 1 includes a main structure 11 and an outer cover 12. The main structure 11 is a hollow shell with an opening at the top, and the outer cover 12 is engaged with the opening of the shell.

[0021] The infant cabin is ellipsoidal in shape and consists of three main areas: a monitoring area, an intervention area, and a support area. The monitoring area is located above the head airbag 21 and is used to install the physiological status monitoring component 3. The intervention area is located at the waist and abdomen airbag 22 of the airbag sleeping pad component 2 and is used to install the dual-effect intestinal gas intervention mechanism 4. The support area is located below the head airbag 21, waist and abdomen airbag 22, and leg airbag 23 and is used to install the airbag support structure.

[0022] The airbag sleeping pad assembly 2 also includes a contact layer 24 and a cushioning layer 25. The head airbag 21, waist and abdomen airbag 22, and leg airbag 23 are all located inside the contact layer 24, and the cushioning layer 25 is located between the head airbag 21, waist and abdomen airbag 22, and leg airbag 23 and the contact layer 24. The contact layer 2 is made of medical-grade antibacterial cotton, and the cushioning layer 25 is made of high-density sponge. A load-bearing frame 26, i.e., the airbag support structure, is located below the contact layer 2. The load-bearing frame 26 is made of ABS engineering plastic, with edges finished using a fully arc-shaped, non-sharp grinding process with a grinding radius of not less than 15mm to eliminate the risk of sharp protrusions. Hidden handles are provided around the load-bearing frame 26 for easy movement within the home. Rubber anti-slip pads 27 are provided between the bottom of the load-bearing frame 26 and the main structure 11. The cushioning layer 25, the load-bearing frame 26, and the pads 27 provide triple-layer combined shock absorption and support for the airbags.

[0023] In addition, such as Figures 4 to 6 As shown, the cabin 1 also includes three T-joints 14, each T-joint 14 being connected to an air valve 15 and a pressure transmitter 16. The head airbag 21, chest and abdomen airbag 22, and leg airbag 23 are respectively connected to the three T-joints 14. The system also has an air source connected to the three air valves 15. The pressure transmitter 16 monitors the air pressure of the head airbag 21, chest and abdomen airbag 22, and leg airbag 23 in real time, and controls the airflow status of the airbags through the air valves 15, thereby ensuring that the air pressure of the chest and abdomen airbag 22 is lower than that of the head airbag 21 and leg airbag 23. Alternatively, in some other embodiments, air sources corresponding one-to-one with the three air valves 15 can be provided, and the head airbag 21, chest and abdomen airbag 22, and leg airbag 23 can all be configured as closed airbags.

[0024] like Figure 1 and Figure 6 As shown, the multimodal sensing module serves as the system's sensing layer, comprising a physiological state monitoring component 3 and a microphone. The physiological state monitoring component 3 includes a radar component and a flexible pressure sensor 34. The radar component includes a radar 31 mounted on the outer casing 12, a linear drive 32 for moving the radar 31 back and forth, and a pitch adjustment component 33 for adjusting the pitch angle of the radar 31. The flexible pressure sensor 34 is located on the top of the sleeping pad assembly 2.

[0025] A pressure signal acquisition circuit is built using an HX71 analog-to-digital converter chip to sense whether the infant is on the airbag sleeping pad assembly 2. The flexible pressure sensor 34 is preferably an FSR402 model, which can be seamlessly embedded between the abdominal airbag 22 and the contact layer 24 without affecting wearing comfort. It has a lifespan of over 1 million cycles, meeting the needs of long-term family use. This sensor is mainly used to collect dynamic pressure fluctuations in the abdominal area to assess the degree of intestinal gas and can be used in conjunction with an industrial control computer to detect whether a person is sitting down.

[0026] The physiological state monitoring component 3 also includes an infrared body temperature sensor, preferably the MLX90614 model. This embodiment uses this infrared body temperature sensor to measure the body temperature of infants and young children. Considering the sensitivity of infants' skin and to avoid causing allergic rashes, a non-contact temperature measurement method is used to improve the accuracy of the measurement data. The MLX90614 infrared body temperature sensor has a measurement accuracy of ±0.2℃ in the range of 36℃ to 38℃.

[0027] The linear drive component 32 is preferably a linear guide rail, with one end hinged to the outer cover 12 and the other end hinged to the pitch adjustment component 33. The pitch adjustment component 33 is preferably an electric push rod, with both ends hinged to the cover 12 and the linear guide rail, respectively. The radar 31 is fixed to the bottom of the slide table of the linear guide rail. The radar 31 is preferably a 24GHz millimeter-wave radar, capable of non-contact acquisition of respiratory, heart rate, body movement, and body position data, unaffected by clothing or other obstructions. A flexible pressure sensor 34 captures abdominal pressure fluctuations and bowel sounds-related body movement data in real time, and a microphone collects crying audio data so that the intelligent recognition and processing module can extract the crying audio features.

[0028] The radar 31 can be adjusted laterally by 10cm using an electric guide rail, and vertically by 15° using an electric push rod, thereby adjusting the pitch angle of the radar 31 to accommodate infants of different heights and various sleeping positions such as supine and lateral, ensuring unobstructed millimeter-wave radar monitoring. Additionally, in some other embodiments, if the outer casing 12 is made of a wave-transparent material such as glass, the radar assembly can be positioned above the outer casing 12.

[0029] In addition, the multimodal sensing module also includes an environmental monitoring component, which includes a temperature and humidity sensor and a CO2 concentration sensor built into the inside of the cabin 1, for real-time monitoring of changes in the internal environment of the cabin 1.

[0030] The intelligent recognition processing module in this embodiment is the system's processing layer. This module has a built-in core controller, which uses STM32 series chips such as the STM32F103C8T6. It is equipped with a CNN+LSTM fusion network model, supporting two core tasks: millimeter-wave radar infant respiratory / positional abnormality recognition and multimodal gas state recognition. This achieves a complete closed loop from data acquisition to decision output and supports transfer learning for fine-tuning to adapt to individual differences. The specific recognition process is as follows: Breathing / Postural Abnormality Detection: This detection process can assess the risk of suffocation and respiratory arrest, triggering a high-level warning. It employs a 24GHz FMCW millimeter-wave radar, coupled with an STM32 microcontroller, to achieve non-contact monitoring. The specific process is as follows: Data Acquisition and Preprocessing: Radar 31 emits FMCW electromagnetic waves and acquires echo I / Q data. The echo I / Q data undergoes DC removal, bandpass filtering, and validity detection to remove invalid data. During bandpass filtering, filtering is performed based on the difference between respiratory rate and heart rate to identify breathing and heartbeat separately. The respiratory rate is 0.1–0.8 Hz, and the heart rate is 1.0–3.0 Hz.

[0031] Target localization and data separation: The infant's position is determined by distance dimension FFT, the phase is extracted and unwrapped to obtain the chest wall displacement waveform, and the breathing and heartbeat data are separated and the real-time frequency is calculated.

[0032] Status determination: Based on the frame data of radar 31, the status of being in bed / out of bed is determined, and the supine, lateral, and prone positions are identified by the echo power distribution, while violent body movements are detected.

[0033] Anomaly Detection: A multi-dimensional temporal feature model incorporating respiration, heart rate, body position, and body movement is constructed and input into a pre-trained CNN+LST fusion network model. This CNN+LST fusion network model has CNN layers, LSTM layers, and fully connected layers. Local anomaly patterns are extracted through the CNN layers, and temporal evolution is captured through the LSTM layers. The output is a respiratory state classification, which comprehensively assesses the asphyxiation risk level.

[0034] Early warning feedback: Triggers audible and visual alarms, app push notifications, and gentle wake-up when there is a high risk; continuously records data and generates a sleep quality assessment report.

[0035] Multimodal flatulence recognition: By integrating abdominal pressure fluctuations, crying audio, and body movement data, it accurately determines the onset of flatulence. The specific process is as follows: Data acquisition and preprocessing: Abdominal pressure data was acquired using a flexible pressure sensor 34, and crying audio data was acquired using a microphone. Spectrograms and MFCC features were extracted, and body movement frequencies were statistically analyzed in conjunction with radar data. The MFCC feature is the Mel-frequency cepstral coefficient.

[0036] Feature fusion: Time alignment is performed on data with different sampling rates, and multi-dimensional features are spliced ​​together using a feature-level early fusion method.

[0037] State classification: Local features such as crying audio, abdominal pressure fluctuations, and body movements are extracted through CNN layers. An LSTM layer is used to construct a dynamic evolution model of gas episodes. The model is then output with confidence scores for "normal quiet / gas episode / crying for other reasons" through a fully connected layer. "Normal quiet," "gas episode," and "crying for other reasons" represent three categories of sleep states.

[0038] Intervention and control: Different intervention strategies are matched according to the three levels of "mild / moderate / severe", the intervention effect is evaluated in real time, parameters are automatically adjusted or modes are switched, and personalized intervention plans are optimized through reinforcement learning.

[0039] The structure and working principle of the CNN+LSTM fusion network model are as follows: CNN layers excel at extracting local features, while LSTM layers excel at capturing temporal dependencies; their fusion performs exceptionally well in physiological and acoustic data processing. The network structure of this system is as follows: The CNN layer is a two-layer one-dimensional CNN architecture, containing batch normalization layers and dropout layers. The CNN layer can automatically extract discriminative features from respiratory waveforms, crying spectra, and stress sequences based on the ReLU activation function.

[0040] The LSTM layer can be a single-layer or two-layer architecture with 64 or 128 hidden units. It is used to establish a model of the temporal evolution of the state and overcome the false alarm problem of single-point threshold judgment.

[0041] Fully connected layers can combine the Softmax activation function to output the confidence scores of each state.

[0042] The system implements transfer learning fine-tuning based on the following process: To address the issues of scarce infant scene samples and significant individual differences, a transfer learning approach is adopted: Source domain pre-training: The base model is trained using public datasets such as adult sleep apnea and infant cries, and learns general feature extraction and temporal modeling capabilities.

[0043] Model Reconstruction: The parameters of the CNN and LSTM layers are retained, and the top fully connected layer is replaced to adapt to the classification task of this system. Specifically, the original top fully connected layer of the pre-trained model contains two sequentially connected layers with 512 and 256 neurons respectively, ultimately outputting two neurons for binary classification. The replaced top layer, adapted for the multi-class classification task of this system, contains three sequentially connected layers with 512, 256, and 128 neurons respectively, ultimately outputting K neurons corresponding to the number of classification categories. Dropout (0.5) and batch normalization are used to prevent overfitting. K is the number of target categories, including respiratory abnormalities, flatulence, and normal. Dropout (0.5) means that during model training, a portion of neurons and their connections are randomly and temporarily ignored with a 50% probability in each iteration.

[0044] Small-sample fine-tuning: A layered freeze + step-by-step unfreezing strategy is adopted. First, the CNN and LSTM backbone networks are frozen, and the top-level classifier is trained using only a small amount of individualized data (e.g., using 50-200 labeled infant physiological signal sequences); then, the data is gradually unfrozen in the following order: Phase 1: Unfreeze the LSTM layer, with a learning rate of 1e-4, and train for 20-30 epochs; Phase 2: Unfreeze the mid-to-high-level CNN feature layers with a learning rate of 5e-5; Phase 3: Fine-tuning of the entire network with a learning rate of 1e-5, using an early stopping mechanism to prevent overfitting.

[0045] Incremental learning: The system automatically accumulates intervention data as new samples and regularly updates the model in the background to achieve long-term adaptation. During actual operation, the system automatically accumulates pre-labeled data, including radar data, abdominal pressure fluctuations, body temperature data, crying audio data, user intervention behavior records (such as adjusting posture, changing position, initiating massage, alarm confirmation, etc.), as well as system operation logs and risk level assessment results. Incremental training is performed daily or weekly in the background. The specific process is as follows: (1) New sample collection and pseudo-label generation; (2) Key samples of manual / semi-supervised audits; (3) Use the experience replay buffer in combination with old samples for replay training; (4) Only update the top-level and some higher-order parameters (learning rate 1e-5 to 1e-6) to achieve long-term model adaptation. Through this mechanism, the system achieves an accuracy rate of no less than 92% in identifying respiratory abnormalities and intestinal bloating with a small amount of individualized data, and a false alarm rate of less than 3%.

[0046] The active intervention module includes a dual-effect gas intervention mechanism 4 and a bionic soothing intervention mechanism 5. The dual-effect gas intervention mechanism 4 is used for hot compresses and vibration of the baby cabin to relieve gas; the bionic soothing intervention mechanism 5 is used for reciprocating movement of the baby cabin to achieve bionic soothing.

[0047] The system uses the active intervention execution module as the execution layer to perform active interventions. The active intervention execution module includes a dual-effect intestinal gas intervention mechanism 4. For example... Figure 7 As shown, the dual-effect intestinal gas intervention mechanism 4 includes a heating film 41 disposed on top of the contact layer 24, and a vibration motor 43 disposed inside the abdominal air bladder 22. A waterproof and breathable membrane 42 is provided on the upper surface of the heating film 41, and a flexible sound-insulating sleeve 44 covers the outside of the vibration motor 43. The dual-effect intestinal gas intervention mechanism 4 also includes an electrical feedthrough 45 disposed between the abdominal air bladder 22 and the contact layer 24, through which communication between the industrial control computer and the vibration motor 43 and power supply to the vibration motor 43 are realized.

[0048] The heating film 41 uses a graphene heating film as the heating element, achieving a heating uniformity of no less than 90% and reaching the set temperature within 5 minutes. Equipped with a high-precision NTC temperature sensor, the heating film 41 achieves precise temperature control within ±0.5℃ of the 36℃ to 38℃ range, avoiding the risk of burns. A waterproof and breathable membrane 42 is installed on the surface of the heating film 41 to prevent the seepage of baby sweat and urine, which could cause short circuits. An industrial control computer works with the heating film 41 to provide multi-level temperature adjustment and a timer function; the default operating time is 15 minutes, after which heating automatically stops.

[0049] The vibration motor 42 is preferably a flat vibration motor. To achieve miniaturization of the equipment, this embodiment uses an ultra-thin flat vibration motor with an adaptive vibration frequency range of 20Hz to 50Hz and an amplitude not exceeding 2mm. By setting a flexible sound insulation sleeve 44 on the outside of the vibration motor 42, the vibration motor 42 is flexibly connected to the airbag sleeping pad assembly 2, which can effectively reduce resonance noise. The material of the flexible sound insulation sleeve 44 is preferably silicone.

[0050] The active intervention execution module also includes a bionic sleep-inducing intervention mechanism 5, which includes a base 51 and a lifting component 52 located between the base 51 and the main structure 11. The lifting component 52 is preferably an electric push rod. An industrial control computer is mounted on the base 51. The electric push rod includes a stepper motor and a ball screw. The stepper motor is connected via a flexible damping coupling, enabling stepless adjustment of the lifting amplitude (5-8cm) and frequency (0.5-1Hz), with a smooth and seamless movement trajectory. The electric push rod incorporates a travel limit switch and a torque sensor, immediately stopping operation when abnormal resistance is detected to prevent mechanical jamming and potential safety hazards.

[0051] An alarm 6 is provided on one side of the main structure 11, which is preferably an audible and visual alarm. The alarm 6 is fixed to the outside of the main structure 11 by a mounting plate, which is also provided with a display screen 7 and a rotary reset emergency stop button 8.

[0052] During the dual-effect intervention for intestinal bloating, the heating film 41 applies a warm compress to the intervention area, keeping the temperature in the middle of the airbag sleeping pad component 2 within the range of 36°C to 38°C. The vibration motor 43 can perform silent low-frequency vibration, and promotes intestinal gas expulsion through the gentle physical intervention of automatically triggering the warm compress vibration.

[0053] During the bionic sleep intervention process, a bionic sleep intervention mechanism is set at the bottom of the cabin, which can imitate the "carrot squat" sleep-inducing action. During the sleep-inducing process, the lifting component 52 has a lifting range of 5-8cm and a lifting frequency of 0.5-1Hz. It is driven by flexible damping through the airbag sleeping pad component 2, with noise less than 30dB, to achieve non-stimulating sleep-inducing.

[0054] During the safety warning and intervention process, when an abnormal breathing is detected, alarm 6 immediately triggers an audible and visual alarm, pushes a notification to the parent's app, and simultaneously initiates a gentle wake-up.

[0055] Specifically, the industrial control computer automatically adjusts the vibration intensity and mode of the vibration motor 43 according to the severity of intestinal bloating. Mild bloating uses an intermittent vibration mode, moderate bloating uses a continuous vibration mode, and severe bloating uses a combination of heat therapy and vibration.

[0056] When the industrial control computer detects an episode of intestinal bloating, it assesses the severity of the bloating based on abdominal pressure fluctuations, crying audio, and body movement data. When the system assesses severe bloating, it first activates the heating film 41 for heat application. Once the temperature reaches the set value, it automatically activates the vibration motor 43 for vibration intervention.

[0057] During the intervention, data on abdominal pressure fluctuations, crying audio, and body movement are continuously collected. The intervention effect is assessed in real time based on the probability of gas relief. If the probability of gas relief is less than the set probability after 15 minutes of intervention, the system automatically adjusts the vibration frequency and heat application temperature, or switches to a radish-squatting sleep-inducing mode. The adjustment process uses data changes as a closed-loop feedback mechanism to dynamically judge the intervention effect and iteratively optimize parameters.

[0058] The system uses reinforcement learning based on historical intervention data to develop personalized intervention plans for different infants, including parameters such as the best intervention time, temperature, vibration frequency and duration.

[0059] The adjustment criteria, rules, and feedback mechanism for the vibration frequency and heat therapy temperature are as follows: (1) Basis for adjustment: ① The abdominal pressure baseline and fluctuation amplitude collected by the flexible pressure sensor 34; the duration, sound pressure level, and paroxysmal characteristics of crying collected by the microphone; and the body movement frequency and amplitude collected by the millimeter-wave radar are used as core inputs to directly determine the severity of intestinal gas and the degree of infant discomfort. The aforementioned abdominal pressure baseline is the average value of the baseline pressure collected by the flexible pressure sensor 34 at the waist and abdomen when the infant is in a normal state of quiet stability, without intestinal gas, and without violent crying.

[0060] The method for obtaining sound pressure level and paroxysmal features from crying audio data is as follows: First, the crying audio collected by the microphone is preprocessed by noise reduction, frame windowing, and removal of silent segments; the effective sound pressure is calculated, and the frame sound pressure level is solved by combining acoustic formulas. The average and peak sound pressure levels of the entire audio segment are statistically analyzed to complete the sound pressure level feature extraction; then, the energy envelope is drawn by short-time audio energy, the crying burst segments and intermittent segments are divided, and parameters such as crying duration, pause interval, burst frequency, and pulse period are extracted to obtain paroxysmal features.

[0061] The method for obtaining body movement amplitude is as follows: The method for obtaining body movement amplitude based on body movement data is as follows: the raw body movement signal collected by millimeter-wave radar is filtered and preprocessed to eliminate environmental interference; the displacement waveform of the infant's limbs is calculated based on the phase change of the radar echo; the displacement value in the normal resting state is used as a reference, and the difference between the real-time displacement and the reference value is calculated to obtain the body movement offset of a single frame; the maximum value of the displacement offset within a specified time window is counted, which is the body movement amplitude of that period.

[0062] This system relies on a CNN+LSTM fusion model to determine the severity of "intestinal gas" and the degree of "infant discomfort" respectively, and completes the iterative optimization of the model through transfer learning, small sample fine-tuning and incremental learning.

[0063] In the mapping between input and labels, the system performs multi-dimensional feature decoupling and fusion: Severity of intestinal bloating: mainly focuses on physiological signs. The model extracts the abdominal pressure baseline drift, pressure fluctuation amplitude and body movement frequency collected by flexible pressure sensors as core input features to define the pressure level inside the intestine.

[0064] Infant discomfort level: mainly focusing on behavior and emotional expression, the model focuses on extracting the sound pressure level of crying, paroxysmal characteristics, duration and body movement amplitude collected by the microphone as core input features to quantify the infant's pain and irritability index.

[0065] The model processing flow is as follows: First, the model was pre-trained in the source domain using a publicly available dataset. After deployment, the parameters of the CNN and LSTM layers were retained, while the fully connected layers were replaced. Fine-tuning was then performed using a small number of on-site samples with stratified thawing of parameters. During long-term use, monitoring and intervention data were automatically accumulated for continuous incremental learning. The model was uniformly input with the aforementioned time-aligned multi-dimensional temporal feature vector, using mild, moderate, and severe as independent classification labels for "severity of intestinal gas" and "infant discomfort level," respectively. After extracting local features and temporal patterns, the model outputs the confidence score for each category using the Softmax function to determine the final intervention level.

[0066] ② Algorithm model evaluation: Through the CNN+LSTM temporal fusion algorithm, the "probability of gas relief" and "infant discomfort index" are output in real time to predict the symptom improvement trend under the current parameters and guide the adjustment direction.

[0067] (2) Adjustment level setting: The system uses the initial intervention parameters as a benchmark and evaluates the effect every 5 minutes. If there is no improvement after 15 minutes, the graded adjustment is initiated: ① Adjustment of hot compress temperature: The initial default temperature is 37℃, which is a neutral and comfortable temperature. If the abdominal pressure decreases by ≤5% from the baseline, the crying does not shorten, and the body movement does not decrease, it indicates that the symptoms are not relieved. At this time, the temperature is increased in increments of 0.5℃, with the maximum temperature not exceeding 38℃. If there is still no improvement after the temperature is increased to 38℃, the upper limit of the temperature is maintained to avoid the risk of burns. ② Adjustment of vibration frequency: The parameters are initially set according to the severity of intestinal gas (intermittent vibration at 20-30Hz for mild gas and continuous vibration at 30-40Hz for moderate gas). If there is no significant improvement in symptoms, the frequency is increased in increments of +5Hz. After each adjustment, it is maintained for 5 minutes and then evaluated again, with the maximum not exceeding 50Hz. If there is still no improvement, the upper limit of the frequency is maintained, and the combined intervention of "hot compress + vibration + imitation radish squatting to lull to sleep" is initiated.

[0068] (3) Adjusted feedback loop and effect judgment: After the parameter adjustment takes effect, the system continuously collects data at a 5-minute interval, judges the effect by quantifying the change in indicators, and forms a closed-loop control: If abdominal pressure decreases by ≥10% from baseline, crying duration shortens by ≥50%, and body movement frequency decreases by ≥30%, it is judged as "significant symptom relief trend". Continue intervention while maintaining the current parameters until the intestinal gas state is recognized as "relieved", then gradually reduce the intensity of intervention and stop. If the change in the indicator is between 5% and 10%, it is judged as "there is a trend of easing", and the current parameters are maintained without additional adjustments. If the indicator decreases by ≤5%, it is judged as "no improvement", and the parameters continue to be adjusted according to the grading rules; If any abnormal reactions are detected in the infant during the adjustment process (violent body movement, persistent high-pitched crying, abnormal skin temperature), all active interventions will be immediately suspended, triggering an audio-visual alert and an app notification. The criteria for violent body movement, persistent high-pitched crying, and abnormal skin temperature are as follows: ① Criteria for judging physical exertion: A triple-threshold approach—relative amplitude, absolute amplitude, and movement frequency—is employed to mitigate baseline drift. The dual thresholds include the following movement frequency and amplitude thresholds: a movement frequency threshold of 10 movements / minute; and a movement amplitude threshold comprising a relative threshold and an absolute threshold. The relative threshold is 2.5 times the baseline amplitude for an infant at rest, and the absolute threshold is 5 cm. The determination criteria are as follows: Body movement frequency greater than 10 times / minute; The real-time displacement amplitude is greater than 2.5 times the baseline amplitude for an infant in a resting position; A single limb displacement amplitude greater than 5cm; If any of the above conditions are met and the duration is greater than or equal to 3 seconds, it is considered to be strenuous physical activity.

[0069] ② Criterion for persistent, shrill crying: This technology differentiates between normal crying and screaming in pain based on infant acoustic characteristics, enhancing the infant's ability to resist interference. It determines "continuously sharp crying" by using a sound pressure level threshold of 85 dB.

[0070] When the average sound pressure level is greater than or equal to 85 dB and the duration is greater than or equal to 5 seconds without any silent intervals, it is judged as a continuous and sharp crying sound.

[0071] ③ Criteria for judging abnormal skin temperature: Based on the safe skin temperature standards for infants and young children, the high-temperature threshold has been tightened and the low-temperature warning value has been optimized. High temperature abnormality (overheating / burn risk): If the local skin temperature exceeds 37.5℃, the power supply to the heating film will be cut off immediately upon triggering; a fault tolerance limit is reserved, and the short-term peak temperature must not exceed 37.8℃.

[0072] Low temperature abnormality (low body temperature): local skin temperature is less than 35.5℃.

[0073] Duration: If the temperature deviates from the safe range for more than or equal to 2 seconds, it is considered an abnormal skin temperature.

[0074] The bionic sleep intervention mechanism 5 can work independently or in conjunction with heat therapy and vibration modules to simulate the physical sensation of a parent holding and soothing a child to sleep.

[0075] Bionic sleep intervention center 5 supports the following multiple sleep-inducing modes: Gentle mode: Lifting range is 5cm, frequency is 0.5Hz; Standard mode: Lifting range is 6.5cm, frequency is 0.75Hz; Depth mode: Lifting range is 8cm, frequency is 1Hz.

[0076] The system automatically switches modes by monitoring the baby's sleep state: when the baby enters light sleep, the lifting amplitude and frequency of the lifting component 52 are gradually reduced; when the baby enters deep sleep, the lifting component 52 is automatically stopped.

[0077] The system uses a safety interaction module as its interaction layer, which includes the aforementioned display screen 7 and emergency stop button 8. The safety interaction module can perform both local and remote interaction.

[0078] During local interaction, the monitoring personnel can stop the active intervention execution module by pressing the emergency stop button 8 and manually adjust the intervention parameters by pressing the display screen 7.

[0079] During remote interaction, the industrial control computer connects wirelessly to a parent's mobile device carried by the caregiver. This parent's device is a mobile app that allows for real-time data viewing, remote control, and receiving alerts.

[0080] Both sides of the inner wall of the main structure 11 are provided with baffles 13. The two baffles 13 serve as a two-way anti-pinch structure to fill the gap between the cabin 1 and the airbag sleeping pad assembly 2.

[0081] The system has the following security mechanisms: In terms of mechanical safety protection: a two-way anti-pinch structure prevents infants' and caregivers' hands from being pinched by the cabin 1 and the airbag sleeping pad assembly 2. When the active intervention execution module is stopped by pressing the emergency stop button 8, the lifting component 52 is de-energized and automatically locked, keeping the cabin 1 stable and preventing sudden abnormal changes in the infant's position.

[0082] Regarding electrical safety protection: the electromagnetic radiation of Radar 31 meets the Class I limit requirements of GB 9175-88 "Environmental Electromagnetic Wave Hygiene Standard", with a radiation power density ≤10μW / cm². 2 It is safe and harmless for infants and young children to be exposed for extended periods. The heat therapy vibration has a timed protection function. After the heat therapy and vibration have reached the preset time, the industrial control computer automatically cuts off the power supply circuit of the heating film 41 and the vibration motor 43. When the temperature of the heating film 41 exceeds 40°C or the current is abnormal, its power supply is immediately cut off.

[0083] Regarding material safety: All components that come into direct contact with an infant's skin are made of materials that comply with Class A standards of GB / T 31701-2015 "Safety Technical Specifications for Infant and Child Textile Products". Contact layer 24 is made of medical-grade antibacterial cotton with an antibacterial rate of ≥99%. The flexible pressure sensor 34's shell is made of food-grade silicone, odorless, and free of harmful substances, and has passed professional skin irritation testing. All system materials comply with RoHS environmental standards and are free of formaldehyde, heavy metals, and other harmful substances. The surface of chamber 1 is coated with water-based paint, releasing no volatile organic compounds.

[0084] Regarding software security: The parent app supports multi-user permission management, allowing the primary user to assign different permissions to other family members. The system monitors the operational status of each module in real time. When any sensor malfunction, communication interruption, or other abnormal situation is detected, an audible and visual alarm is immediately triggered via alarm 6, and an abnormality information is pushed to the parent app.

[0085] Second Embodiment This embodiment relates to a control method applicable to the integrated infant sleep and colic intervention system of the first embodiment, comprising the following steps: Step S1, Multimodal Sensing: The physiological status data of the infants in the infant cabin are collected based on the multimodal perception module. The physiological status data includes respiratory data, heart rate data, body movement data, body position data, abdominal and lumbar pressure fluctuation data, and cry audio data. Step S1 includes the following process: Step S11, Non-contact monitoring: The head monitoring area is equipped with a 24GHz millimeter-wave radar, which collects respiratory, heart rate, body movement and body position data without contact and is not affected by clothing or obstruction.

[0086] Step S12, Intestinal distension feature acquisition: A flexible pressure sensor is embedded in the lumbar and abdominal intervention area to capture abdominal pressure fluctuations and bowel sounds-related body movements in real time; a microphone is used to collect the audio features of crying.

[0087] Step S13, Environmental Monitoring: Built-in temperature, humidity, and CO2 sensors monitor the cabin environment in real time.

[0088] Step S2, Intelligent Recognition Processing: Through the intelligent recognition and processing module, based on the CNN+LSTM fusion model, millimeter-wave radar is used to identify abnormal infant breathing and posture, as well as multimodal gas state recognition, realizing a complete closed loop from data acquisition to decision output, and supporting transfer learning to fine-tune and adapt to individual differences. Step S2 specifically includes the following steps: Step 21: Millimeter wave respiratory monitoring and abnormality identification: Abnormal breathing and body position recognition is used to assess the risk of suffocation and sleep apnea, triggering high-level warnings. A 24GHz FMCW millimeter-wave radar, coupled with an STM32 microcontroller, enables non-contact monitoring. The specific process is as follows: Step S211, Data Acquisition and Preprocessing: The radar emits FMCW electromagnetic waves, and the echo I / Q data is acquired. After DC removal, bandpass filtering, and validity detection, invalid data is removed. During the bandpass filtering process, filtering is performed based on the difference between respiratory rate and heart rate to identify breathing and heart rate separately. The respiratory rate is 0.1–0.8 Hz, and the heart rate is 1.0–3.0 Hz.

[0089] Step S212, Target localization and data separation: Determine the infant's position using distance dimension FFT, extract and unwrap the phase to obtain the chest wall displacement waveform, separate the breathing and heartbeat data and calculate the real-time frequency.

[0090] Step S213, Status Determination: Combine radar frame data to determine the status of being in bed or out of bed, identify supine, lateral, and prone positions through echo power distribution, and detect violent body movements at the same time.

[0091] Step S214, Anomaly Identification: Construct a multi-dimensional temporal feature including respiration, heart rate, body position, and body movement, input it into a CNN+LSTM fusion network model, extract local anomaly patterns in the CNN layer, capture temporal evolution in the LSTM layer, output respiratory state classification, and comprehensively assess the asphyxiation risk level.

[0092] Step S215, Early Warning Feedback: Trigger audible and visual alarms, App push notifications, and gentle wake-up when there is a high risk; continuously record data and generate a sleep quality assessment report.

[0093] Step S22, Multimodal flatulence recognition: By integrating abdominal pressure, crying sounds, and body movement frequency, an accurate diagnosis of intestinal gas attacks can be made. The specific identification process is as follows: Step S221, Data Acquisition and Preprocessing: Abdominal pressure is acquired using a flexible pressure sensor, crying sounds are collected using a microphone and spectrograms and MFCC features are extracted, and body movement frequencies are statistically analyzed using radar data.

[0094] Step S222, Feature Fusion: Time alignment is performed on data with different sampling rates, and multi-dimensional temporal feature vectors are spliced ​​together using a feature-level early fusion method.

[0095] Step S23, State Classification: The multidimensional temporal feature vector is input into a pre-trained CNN+LSTM fusion network model, which outputs a classification result of the infant's current state. The CNN+LSTM fusion network model includes CNN layers and LSTM layers. CNN layers excel at extracting local features, while LSTM layers excel at capturing temporal dependencies; the fusion of the two performs excellently in physiological and acoustic data processing. The specific state classification process is as follows: Step S231: Automatically extract discriminative features from respiratory waveforms, cry spectra, and stress sequences using a CNN layer. The CNN layer is a two-layer one-dimensional CNN, containing batch normalization layers, dropout layers, and the ReLU activation function.

[0096] Step S232: Construct a temporal evolution model of the state using an LSTM layer to overcome the false alarm problem of single-point threshold judgment. The LSTM layer can be a single-layer or two-layer architecture, with 64 or 128 hidden units.

[0097] Step S233: The fully connected layer, combined with the Softmax activation function, outputs three state confidence scores. The three states are: normal quiet state, gas attack state, and crying state due to other reasons.

[0098] Step S24, Intervention and Control: Different intervention strategies are matched according to three levels: mild, moderate, and severe. The intervention effect is evaluated in real time, parameters are automatically adjusted or modes are switched, and personalized intervention plans are optimized through reinforcement learning to achieve fine-tuning through transfer learning. The fine-tuning process of transfer learning is as follows: To address the issues of scarce infant scene samples and significant individual differences, a transfer learning approach is adopted: Step S241, Model Reconstruction: Retain the parameters of the CNN and LSTM layers, and replace the top fully connected layer to adapt to the classification task of this system.

[0099] Step S242, Small Sample Fine-tuning: A layered freeze + thaw strategy is adopted. First, the top-level classifier is trained with a small amount of individual data, and then the high-level parameters are gradually thawed to fine-tune the entire network with a low learning rate to prevent overfitting.

[0100] Step S243, Incremental Learning: Automatically accumulate intervention data as new samples, and regularly update the model in the background to achieve long-term adaptation.

[0101] Through this mechanism, the system achieves an accuracy rate of no less than 92% in identifying respiratory abnormalities and intestinal bloating with a false alarm rate of less than 3% even with limited individualized data.

[0102] It should be noted that the CNN+LSTM fusion model is built through source domain pre-training. The source domain pre-training process is as follows: a base model is trained using publicly available datasets such as adult sleep apnea and infant cries to learn general feature extraction and temporal modeling capabilities, thus establishing a pre-trained CNN+LSTM fusion network model. Step S3: Active intervention in the execution system: The active intervention execution module executes control commands and regulates the infant pod. Step S3 specifically includes the following process: Step S31, Dual-Effect Intervention for Intestinal Distension: A warm compress module and a silent low-frequency vibration module are integrated into the lumbar and abdominal intervention area to automatically trigger gentle physical intervention and promote intestinal gas expulsion. The warm compress module controls the temperature of the lumbar and abdominal intervention area at 36-38℃.

[0103] Step S32, Bionic Sleep-Inducing Intervention: A radish-like squatting lifting mechanism is installed at the bottom of the infant cabin. A flexible damping transmission mechanism is installed between the lifting mechanism and the bottom of the infant cabin, with noise <30dB, providing non-stimulating sleep-inducing effects. The lifting range of the mechanism is 5-8cm, and the lifting frequency is 0.5-1Hz.

[0104] Step S33, Safety Warning Intervention: When breathing is abnormal, i.e. when breathing cannot be identified in step S2, an audible and visual alarm is immediately triggered, and information about the abnormal breathing is pushed to the parent's APP. At the same time, the gentle wake-up mechanism is activated to wake the baby.

[0105] Step S4, Human-Computer Interaction and Security Protection: The process of using a safe interactive module to provide feedback on the infant's sleep status is as follows: Step S41, Local Interaction: An emergency stop button and a status display screen are installed on the outside of the infant cabin. The system receives and responds to manual adjustment data and adjusts the intervention parameters.

[0106] Step S42, Remote Interaction: The accompanying parent app displays data in real time, allows parents to remotely send control commands, and receives alerts.

[0107] Step S43, Safety Mechanism: The baby cabin has a two-way anti-pinch structure and automatic locking when the power is off. The electromagnetic radiation of the baby cabin meets the first-level limit of GB 9175-88. The baby cabin has heat therapy and vibration timer protection functions.

[0108] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. An integrated intelligent intervention system for infant sleep and colic, characterized in that: Used for monitoring and regulating the baby pod, including: A multimodal sensing module is used to collect physiological state data of infants and young children inside the infant cabin; The intelligent recognition and processing module is used to receive physiological state data and identify the infant's sleep state based on the physiological state data in order to generate control commands; An active intervention execution module is used to execute the control commands and regulate the baby cabin; A safe interactive module is used to provide feedback on the sleep status of infants and young children; The control commands include dual-effect intervention commands for intestinal gas and bionic sleep-inducing intervention commands; When the active intervention execution module executes the dual-effect intervention command for intestinal gas, it heats and vibrates the baby cabin through a heating film and a vibration motor; when it executes the bionic soothing intervention command, it performs a bionic soothing action on the baby cabin, the bionic soothing action including reciprocating movement of the baby cabin.

2. The active intelligent intervention system according to claim 1, characterized in that: The infant cabin includes a cabin body and an airbag sleeping pad assembly disposed inside the cabin body. The airbag sleeping pad assembly includes a head airbag, a waist and abdomen airbag, and a leg airbag. The head airbag is higher than the leg airbag, and the air pressure of the waist and abdomen airbag is lower than that of the head airbag and the leg airbag. The cabin body includes a main structure and an outer cover. The main structure is a hollow shell with an opening at the top, and the outer cover is engaged with the opening of the shell.

3. The active intelligent intervention system according to claim 2, characterized in that: The multimodal sensing module includes a physiological state monitoring component, which includes a radar mounted on the outer casing, a linear drive for moving the radar back and forth, a pitch adjustment component for adjusting the radar pitch angle, and a flexible pressure sensor mounted on the top of the lumbar airbag.

4. The active intelligent intervention system according to claim 2, characterized in that: The airbag sleeping pad assembly also includes a contact layer, and the head airbag, waist and abdomen airbag and leg airbag are all located inside the contact layer. The active intervention execution module includes a dual-effect intestinal bloating intervention mechanism, which includes a heating film located on the top of the contact layer and a vibration motor located inside the waist and abdomen airbag.

5. The active intelligent intervention system according to claim 4, characterized in that: The active intervention execution module also includes a bionic sleep-inducing intervention mechanism, which includes a base and a lifting component located between the base and the main structure.

6. A control method for an active intelligent intervention system based on any one of claims 1-5, comprising the following steps: Step S1: Collect physiological state data of infants and young children in the baby cabin based on the multimodal sensing module; Step S2: Receive physiological state data through the intelligent recognition and processing module, and identify the infant's sleep state based on the physiological state data to generate control commands; Step S3: Execute the control command and regulate the baby cabin through the active intervention execution module; Step S4: Feedback on the infant's sleep status based on the security interaction module; The control commands include a dual-effect intervention command for intestinal colic, a bionic soothing intervention command, and a safety warning intervention command; when the dual-effect intervention command for intestinal colic is executed, the baby cabin is heated and vibrated; when the bionic soothing intervention command is executed, the baby cabin is moved back and forth; when the safety warning intervention command is executed, an alarm is sounded and the baby cabin is vibrated when the infant's sleep state is abnormal.

7. The control method according to claim 6, characterized in that: The physiological data includes respiration, heart rate, body movement, body position, abdominal pressure fluctuations, and cry audio data; in step S1, the infant's respiration, heart rate, body movement, and body position data are collected based on radar; and the infant's abdominal pressure fluctuations and body movement data are collected based on a flexible pressure sensor. Audio data of crying is collected using a microphone.

8. The control method according to claim 6, characterized in that: Step S2 includes: Step S21: Preprocess the respiratory, heart rate, body movement and body position data, determine the infant's position by distance dimension FFT, extract phase data and unwrap to obtain chest wall displacement waveform, separate respiratory data and heart rate data, and calculate respiratory rate and heart rate in real time; Step S22: Time-align the data of abdominal pressure fluctuations, crying audio, and body movement, and splice them together to form a multi-dimensional temporal feature vector; Step S23: Input the multidimensional temporal feature vector into the pre-trained CNN+LSTM fusion network model and output the classification result of the infant's current state.

9. The control method according to claim 8, characterized in that: The classification results include respiratory state categories and intestinal distension state categories; the fusion network model includes CNN layers, LSTM layers, and fully connected layers; the CNN layer adopts a two-layer one-dimensional convolutional structure to extract local discriminative features of the temporal feature vector; the LSTM layer adopts a single-layer or two-layer structure to construct a temporal evolution law model of the state; the fully connected layer is used to combine the Softmax activation function to output the confidence scores of the respiratory state category and the intestinal distension state category.

10. The control method according to claim 9, characterized in that: Step S2 further includes: Step S24, generating corresponding control instructions based on the classification results and performing transfer learning fine-tuning; the transfer learning fine-tuning includes the following steps: The parameters of the CNN and LSTM layers of the fusion network model are retained, and the fully connected layers of the fusion network model are replaced. The intervention data collected within a set time period during system deployment is used as new samples to gradually unfreeze the high-level parameters of the fusion network model, perform small-sample fine-tuning, and update the fusion network model. After a set time period, the automatically accumulated intervention data will be used as new samples for incremental learning, continuously updating the fusion network model.

Images