Control method and system for intelligent cradle

Abstract

The invention provides a control method and system for an intelligent cradle, and the method comprises the steps: collecting multi-modal physiological data of an infant in real time through a sensor group, integrating the multi-modal physiological data into a composite comfort index Q, and calculating a time sequence change gradient # imgabs0 # of the composite comfort index Q to construct a dynamic evaluation model; acquiring internal and external parameters of the cradle by an environment sensing sensor, and performing signal noise reduction and space-time alignment by an edge computing node to form a digital twinborn model; inputting the space-time correlation feature matrix into a hierarchical state recognition model, and outputting an infant state grade, emotion judgment and a comfort level score through common feature extraction and individual adaptability fine adjustment; the abnormal state prediction module monitors Q value change and environment sudden change based on the time convolutional network; a three-dimensional control vector is dynamically generated according to an infant state, a differentiation strategy is adopted for different sleep-awakening states, and a control parameter group is continuously optimized. According to the scheme, the adaptability and individuation level of the intelligent cradle are improved, and the intelligence, safety and reliability of the control system are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control technology, and in particular to a control method and system for an intelligent cradle. Background Art

[0002] With the advancement of technology and the improvement of living standards, users have higher requirements for comfort in the process of taking care of infants and young children. Most existing cradles are manually rocked, which is laborious and its functions can no longer meet people's diverse functional needs in the process of taking care of infants and young children. Summary of the Invention

[0003] Based on the above problems, the present invention proposes a control method and system for a smart cradle. Through the solution of the present invention, not only the adaptability and personalization level of the smart cradle are improved, but also the intelligence, safety and reliability of the system are improved through advanced algorithms and architectural design.

[0004] In view of this, one aspect of the present invention provides a control method for an intelligent cradle, comprising:

[0005] The distributed biometric sensor group collects the infant's multimodal physiological data in real time;

[0006] Adaptive fusion algorithm is used to integrate the collected multimodal physiological data into a composite comfort index Q, and the temporal change gradient of the Q value is calculated at the same time. Establish a dynamic assessment model for infant status;

[0007] Acquire the environmental parameter group inside and outside the cradle through the environmental perception sensor group;

[0008] All multimodal physiological data and environmental parameter groups are processed through edge computing nodes for signal noise reduction and spatiotemporal alignment, establishing a spatiotemporal correlation feature matrix of multi-source heterogeneous data and forming a full-dimensional digital twin model of the cradle and its surrounding environment;

[0009] Input the spatiotemporal correlation feature matrix into the pre-trained hierarchical state recognition model;

[0010] The hierarchical state recognition model outputs the current infant's wakefulness-sleep state level, emotion type determination, and comprehensive comfort score through the collaborative calculation of the common feature extraction layer across infant groups and the individual adaptive fine-tuning layer;

[0011] An abnormal state prediction module is built based on a temporal convolutional network to continuously monitor Q-value change patterns and environmental parameter mutation events to predict possible abnormal states.

[0012] The composite comfort index Q and the environmental parameter group are input into the pre-trained deep reinforcement learning model, and a three-dimensional control vector is dynamically generated according to the recognized infant state level;

[0013] Combined with the three-dimensional control vector, differentiated control strategies are adopted for different sleep-wake states;

[0014] The control parameter group is continuously optimized through the online Q-learning algorithm to maximize the comfort score function while balancing the immediate response and long-term optimization goals.

[0015] Optionally, all multimodal physiological data and environmental parameter groups are processed through edge computing nodes for signal noise reduction and spatiotemporal alignment, a spatiotemporal correlation feature matrix of multi-source heterogeneous data is established, and a full-dimensional digital twin model of the cradle and its surrounding environment is formed, including the following steps:

[0016] Configure a time synchronization protocol to synchronize the clocks of all sensor nodes in the biometric sensor group and the environmental perception sensor group at the microsecond level using the Precision Time Protocol or the Network Time Protocol to establish a unified time base.

[0017] Execute sensor-specific denoising algorithms on the multimodal physiological data collected by the biometric sensor group; perform calibration and filtering on the environmental parameter group collected by the environmental perception sensor group;

[0018] Establish a unified timestamp reference system to map sensor data with different sampling frequencies to a unified time axis;

[0019] Implement a delay compensation mechanism to solve the problem of signal transmission and processing delays of different sensors;

[0020] Construct a three-dimensional coordinate system with the geometric center of the cradle as the origin, and uniformly map all spatially related data to this coordinate system;

[0021] Construct a mapping model for infant physiological parameters: Based on a standard infant human body model, map infrared thermal imaging data to a human body surface model; combine data from force-sensitive sensors within the cradle to estimate the infant's posture and position within the cradle; and establish a dynamic transformation relationship between the cradle coordinate system and the infant's body coordinate system.

[0022] Extract features from multi-source heterogeneous data after spatiotemporal alignment;

[0023] Construct the spatiotemporal correlation feature matrix M;

[0024] Construct a digital mapping of the cradle's physical system based on a spatiotemporal correlation feature matrix;

[0025] Construct a parametric model of the infant's physiological state;

[0026] Establish an environment-cradle-infant three-way interaction model.

[0027] Optionally, the adaptive fusion algorithm is used to integrate the collected multimodal physiological data into a composite comfort index Q, and the time series change gradient of the Q value is calculated. The steps to establish a dynamic assessment model of infant status include:

[0028] Adaptive wavelet transform is applied to the original heart rate signal in the multimodal physiological data for noise reduction. The characteristic sequence of heart rate variability is extracted, including the standard deviation of consecutive heartbeat intervals, the root mean square of the difference between adjacent heartbeat intervals, and the percentage of adjacent heartbeat intervals with a difference greater than 50ms, to form the heart rate feature vector F_HRV(t);

[0029] The infrared thermal imaging data in the multimodal physiological data is used to construct the infant's body surface temperature distribution matrix T(x, y, t). The environmental interference is eliminated by two-dimensional Gaussian filtering, and the neck-limb temperature difference ΔT_NL and the core area temperature gradient vector are calculated. Forming the temperature characteristic vector F_TEMP(t);

[0030] Perform spectrum analysis on the respiratory waveform in the multimodal physiological data, extract the respiratory frequency RF, respiratory depth RD, inspiration / expiration ratio RI / RE and respiratory irregularity index RII, and form the respiratory feature vector F_RESP(t);

[0031] Mel-frequency cepstral coefficients and acoustic event detection algorithms are used to extract sound features from the sound data in multimodal physiological data, identify crying, comfort sounds and environmental noise, calculate the acoustic comfort index ASI, and form the acoustic feature vector F_ACOU(t);

[0032] The aforementioned eigenvectors are adaptively normalized so that physiological indicators of different scales are mapped to a unified [0,1] interval:

[0033] Among them, min_i and max_i are the historical minimum and maximum values updated online, and are dynamically adjusted through the exponential moving average algorithm to adapt to individual differences and developmental changes of infants;

[0034] Construct a multi-layer attention network to calculate the feature weight matrix W(t);

[0035] Establish four sub-comfort index mapping functions;

[0036] The composite comfort index Q(t) is calculated by combining the feature weight matrix W(t) with the mapping functions of each sub-comfort index;

[0037] The central difference method is used to calculate the time series gradient of the Q value: Wherein, Δt is the sampling time interval;

[0038] Combining Q(t) and Construct a two-dimensional state space and divide it into multiple state regions;

[0039] A hybrid model combining long short-term memory network and Gaussian process regression is used to predict the future trend of Q(t).

[0040] Optionally, the method for generating the hierarchical state recognition model includes:

[0041] Establish a three-level nested neural network architecture, including a low-level physiological state recognition network, a mid-level emotional state analysis network, and a high-level demand intention reasoning network;

[0042] Design a common feature extraction layer, using a universal feature mapping network across infant groups to capture common physiological response patterns among different infants;

[0043] Constructing an individual adaptive fine-tuning layer, using a parameter-efficient meta-learning structure that can quickly adapt to the individual differences of specific infants;

[0044] Deploy local training modules on the local edge computing nodes of each cradle to ensure that the original physiological data does not leave the local device;

[0045] Establish a secure aggregation protocol so that local nodes only share model gradients instead of raw data;

[0046] Deploy a differential privacy mechanism to add calibration noise to the gradient information to prevent user data from being reconstructed through the gradient;

[0047] Use pre-trained convolutional neural networks to extract time-frequency features from multimodal physiological signals;

[0048] Establish a universal representation space for physiological signals through self-supervised learning methods;

[0049] Apply attention mechanism to dynamically adjust and fuse the weights of different sensor signals;

[0050] Gradient descent is used to optimize a specific parameter subset on the local edge device, keeping the common characteristic network parameters unchanged;

[0051] Using a meta-learning framework to achieve few-shot learning, model adaptation can be completed with only a small amount of individual data; designing a memory-enhanced neural network structure to store and utilize the historical response patterns of specific infants;

[0052] Build a multi-task learning architecture that shares underlying representations and simultaneously outputs wakefulness-sleep state levels, emotion types, and comfort scores;

[0053] Using a soft parameter sharing strategy, each task network maintains some independent parameters to capture task-specific features;

[0054] Formulate a cross-attention mechanism between tasks so that the outputs of different tasks can promote and constrain each other.

[0055] Optionally, the calculation formula of the composite comfort index Q is:

[0056]

[0057] Among them, the weight w_i(t) reflects the contribution of each physiological feature to the overall comfort at the current moment. The specific steps of constructing a multi-layer attention network to calculate the feature weight matrix W(t) are:

[0058] Capturing temporal dependencies within a single modality through self-attention mechanisms;

[0059] Establish correlations between different physiological features through cross-modal attention mechanism;

[0060] Adopt gated recursive units to fuse historical weight information to ensure the temporal continuity of weight changes;

[0061] Output dynamic weight vector W(t)=[w_HRV(t), w_TEMP(t), w_RESP(t), w_ACOU(t)], where each weight satisfies ∑w_i(t)=1.

[0062] Optionally, the dynamic optimization process of the three-dimensional control vector includes:

[0063] Establish the state-action value function Q(s, a) between the cradle control parameters and the infant's physiological feedback;

[0064] Reduce the deviation of Q-value estimation through double temporal difference learning algorithm;

[0065] Adopting the experience replay mechanism to store historical state transition samples and improve sample utilization efficiency;

[0066] A curiosity-driven exploration strategy is introduced to actively explore new control parameter combinations to avoid falling into local optimal solutions.

[0067] Optionally, the working mechanism of the abnormal state prediction module includes:

[0068] Construct a multi-scale temporal convolutional network to simultaneously capture short-term, medium-term, and long-term physiological parameter change patterns; design a multi-parameter fusion algorithm based on the attention mechanism to dynamically adjust the weights of different physiological indicators;

[0069] The Bayesian probability graphical model is used to establish the conditional probability distribution of abnormal states, and the early warning mechanism is triggered when the probability of abnormal states exceeds the threshold.

[0070] Optionally, the IoT device collaborative control further includes:

[0071] Adjust home environment parameters based on the baby's sleep cycle and state prediction;

[0072] Build a personalized environmental preference model based on the baby's emotional response, including optimal temperature range, light intensity, and background music type;

[0073] When sudden interference from the external environment is detected, a multi-layer protection mechanism is activated, including active noise reduction, vibration isolation, and environmental parameter stabilization control.

[0074] Optionally, the process also includes quantum optimization execution and trajectory control steps, specifically:

[0075] The actuator control parameters are optimized by quantum annealing algorithm, and the three-dimensional control vector is converted into a pulse sequence of the multi-axis servo motor;

[0076] The Bezier curve interpolation algorithm is used to smooth the swing trajectory to simulate the three-dimensional composite trajectory of the mother's embrace;

[0077] Dynamically adjust the electromagnetic suspension position of the counterweight according to the baby's weight distribution to achieve precise motion control and adaptive balance adjustment with optimal energy consumption;

[0078] When the abnormal state prediction module triggers an early warning, the progressive braking mechanism and emergency posture adjustment are immediately activated to ensure the safety of the baby.

[0079] Another aspect of the present invention provides a control system for a smart cradle, for executing a control method for a smart cradle, comprising: a distributedly deployed biometric sensor group, an environmental perception sensor group, an edge computing node, and a server;

[0080] The biometric sensor group is configured to: collect multimodal physiological data of the infant in real time;

[0081] The environmental perception sensor group is configured to: obtain an environmental parameter group inside and outside the cradle;

[0082] The edge computing node is configured to: perform signal noise reduction and spatiotemporal alignment processing on all multimodal physiological data and environmental parameter groups, establish a spatiotemporal correlation feature matrix of multi-source heterogeneous data, and form a full-dimensional digital twin model of the cradle and its surrounding environment;

[0083] The server is configured to:

[0084] Adaptive fusion algorithm is used to integrate the collected multimodal physiological data into a composite comfort index Q, and the temporal change gradient of the Q value is calculated at the same time. Establish a dynamic assessment model for infant status;

[0085] Input the spatiotemporal correlation feature matrix into the pre-trained hierarchical state recognition model;

[0086] The hierarchical state recognition model outputs the current infant's wakefulness-sleep state level, emotion type determination, and comprehensive comfort score through the collaborative calculation of the common feature extraction layer across infant groups and the individual adaptive fine-tuning layer;

[0087] An abnormal state prediction module is built based on a temporal convolutional network to continuously monitor Q-value change patterns and environmental parameter mutation events to predict possible abnormal states.

[0088] The composite comfort index Q and the environmental parameter group are input into the pre-trained deep reinforcement learning model, and a three-dimensional control vector is dynamically generated according to the recognized infant state level;

[0089] Combined with the three-dimensional control vector, differentiated control strategies are adopted for different sleep-wake states;

[0090] The control parameter group is continuously optimized through the online Q-learning algorithm to maximize the comfort score function while balancing the immediate response and long-term optimization goals.

[0091] The control method for the smart cradle using the technical solution of the present invention includes: collecting the multimodal physiological data of the baby in real time through a distributed biometric sensor group; integrating the collected multimodal physiological data into a composite comfort index Q using an adaptive fusion algorithm, and calculating the time series change gradient of the Q value. A dynamic assessment model for the infant's state is established. Environmental parameters inside and outside the cradle are acquired through an environmental perception sensor group. All multimodal physiological data and environmental parameter groups are processed through edge computing nodes for signal noise reduction and spatiotemporal alignment. A spatiotemporal correlation feature matrix of multi-source heterogeneous data is established, forming a full-dimensional digital twin model of the cradle and its surrounding environment. This spatiotemporal correlation feature matrix is input into a pre-trained hierarchical state recognition model. The hierarchical state recognition model, through the collaborative computation of a common feature extraction layer across infant groups and an individual adaptive fine-tuning layer, outputs the infant's current wakefulness-sleep state level, emotion type determination, and comprehensive comfort score. An abnormal state prediction module is constructed based on a temporal convolutional network to continuously monitor Q-value change patterns and environmental parameter mutation events to predict possible abnormal states. The composite comfort index Q and environmental parameter group are input into a pre-trained deep reinforcement learning model, which dynamically generates a three-dimensional control vector based on the identified infant state level. Combined with the three-dimensional control vector, differentiated control strategies are adopted for different sleep-wake states. An online Q-learning algorithm is used to continuously optimize the control parameter group to maximize the comfort score function while balancing immediate response and long-term optimization goals. By integrating multi-source biosensors with environmental perception units, an environmental-dynamic digital twin model is constructed, providing a more comprehensive data foundation for control decisions. A three-layer state recognition architecture (physiological-emotional-demand) is introduced, combined with a temporal convolutional network to predict abnormal states and provide early warning of potential risks. Differentiated control strategies are designed for infants' different wakefulness and sleep states, directly incorporating physiological parameters such as heart rate variability into the control function for personalized adaptation. This multi-dimensional integrated control approach not only improves the adaptability and personalization of the smart cradle, but also enhances the system's intelligence, safety, and reliability through advanced algorithms and architectural design. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 is a flow chart of a control method for a smart cradle provided by one embodiment of the present invention;

[0093] Figure 2 The figure is a schematic block diagram of a control system for an intelligent cradle provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0094] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0095] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0096] The terms "first," "second," and so on, in the specification and claims of this application and the accompanying drawings are used to distinguish between different objects, not to describe a specific order. Furthermore, the terms "including," "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements but may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to the process, method, product, or apparatus.

[0097] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0098] Refer to the following Figures 1 to 2 A control method and system for a smart cradle provided according to some embodiments of the present invention will be described.

[0099] like Figure 1 As shown, one embodiment of the present invention provides a control method for an intelligent cradle, comprising:

[0100] The distributed biometric sensor group collects the infant's multimodal physiological data in real time;

[0101] In this step, the biometric sensor suite includes a millimeter-wave radar vital sign monitoring module, an infrared thermal imaging body temperature module, a flexible piezoelectric respiratory sensor strip, and a non-contact acoustic signature analysis unit, respectively used to collect data such as heart rate variability, body surface temperature gradient, respiratory rhythm, and crying sound spectrum characteristics. The biometric sensor suite also includes a photoplethysmography sensor for accurately detecting changes in infant microblood flow; a high-sensitivity acoustic sensor array that uses a deep neural network to classify and identify infant sounds, distinguishing at least eight types of crying and four emotional states; and a near-infrared spectral imaging system that analyzes changes in blood oxygen levels in the infant's brain to assist in determining sleep quality.

[0102] Adaptive fusion algorithm is used to integrate the collected multimodal physiological data into a composite comfort index Q, and the temporal change gradient of the Q value is calculated at the same time. Establish a dynamic assessment model of the infant's status (to provide a comprehensive, continuous and accurate description of the physiological status);

[0103] The environmental parameters inside and outside the cradle are acquired through the environmental perception sensor group (including the cradle's motion trajectory collected by the built-in six-axis inertial measurement unit, the three-dimensional data of the surrounding three-dimensional space collected by the lidar, and the environmental parameters collected by sensors such as temperature and humidity, light intensity, and volatile organic compound concentration sensors);

[0104] All multimodal physiological data and environmental parameter groups are processed through edge computing nodes for signal noise reduction and spatiotemporal alignment, establishing a spatiotemporal correlation feature matrix for multi-source heterogeneous data and forming a full-dimensional digital twin model of the cradle and surrounding environment (providing a high-dimensional spatiotemporal correlation data foundation for state recognition and decision-making);

[0105] Input the spatiotemporal correlation feature matrix into a pre-trained hierarchical state recognition model (this model combines transfer learning and federated learning frameworks and includes at least three levels: low-level physiological state recognition, mid-level emotional state analysis, and high-level demand intention reasoning);

[0106] The hierarchical state recognition model uses a collaborative calculation between a common feature extraction layer across infant groups and an individual adaptive fine-tuning layer to output the current infant's wakefulness-sleep state level, emotion type determination, and a comprehensive comfort score (achieving continuous model optimization while protecting privacy).

[0107] An abnormal state prediction module is built based on a temporal convolutional network to continuously monitor Q-value change patterns and environmental parameter mutation events, and predict possible abnormal states (200-500ms in advance);

[0108] The composite comfort index Q and the environmental parameter set are input into a pre-trained deep reinforcement learning model. Based on the identified infant status level, a three-dimensional control vector (including rocking frequency F, amplitude A, and trajectory curvature K) is dynamically generated, and the adaptive weight coefficient of each parameter is calculated.

[0109] Combined with the three-dimensional control vector, differentiated control strategies are adopted for different sleep-wake states;

[0110] In an embodiment of the present invention, a differentiated control strategy may be: in the wakeful state, a high-frequency, low-amplitude oscillation mode is adopted, wherein the amplitude A(t) = k·exp(-α·t)+C and the frequency f(t) = β·HRV(t) (HRV is real-time heart rate variability); in the light sleep state, a medium-frequency variable-amplitude swing mode is adopted, wherein the frequency forms a nonlinear resonant relationship with the infant's breathing frequency; in the deep sleep state, a low-frequency variable-amplitude swing mode is adopted, wherein the amplitude is regulated by a quasi-periodic chaotic function, and a tiny random perturbation factor is generated by a quantum random number generator to prevent the formation of mechanical motion dependence.

[0111] The control parameter group is continuously optimized through the online Q-learning algorithm to maximize the comfort score function while balancing the immediate response and long-term optimization goals.

[0112] It can be understood that the control parameter set includes: the three-dimensional control vector (F, A, K) mentioned above, the adaptive weight coefficients of each parameter, and other more fine-grained execution parameters. The role of the online Q-learning algorithm is to continuously optimize this entire parameter set so that it can: maximize the infant's comfort score, balance the system's immediate responsiveness and long-term optimization goals. In short, the three-dimensional control vector (F, A, K) is the core subset of the control parameter set. The deep reinforcement learning model first generates this three-dimensional control vector, and then the online Q-learning algorithm will further optimize the more complete control parameter set including this vector to achieve continuous improvement in system performance.

[0113] The technical solution of this embodiment integrates multi-source biosensors and environmental perception units to construct an environmental-dynamic digital twin model, providing a more comprehensive data foundation for control decisions. It introduces a three-layer state recognition architecture (physiological-emotional-demand), combined with a temporal convolutional network to predict abnormal states and provide early warning of potential risks. It also designs differentiated control strategies for the infant's different wakefulness and sleep states, directly incorporating physiological parameters such as heart rate variability into the control function to achieve personalized adaptation. This multi-dimensional integrated control method not only improves the adaptability and personalization of the smart cradle, but also enhances the system's intelligence, safety, and reliability through advanced algorithms and architectural design.

[0114] In some possible implementations of the present invention, quantum optimization execution and trajectory control steps are further included, specifically:

[0115] The actuator control parameters are optimized by quantum annealing algorithm, and the three-dimensional control vector is converted into a pulse sequence of the multi-axis servo motor;

[0116] The Bezier curve interpolation algorithm is used to smooth the swing trajectory to simulate the three-dimensional composite trajectory of the mother's embrace;

[0117] Dynamically adjust the electromagnetic suspension position of the counterweight according to the baby's weight distribution to achieve precise motion control and adaptive balance adjustment with optimal energy consumption;

[0118] When the abnormal state prediction module triggers an early warning, the progressive braking mechanism and emergency posture adjustment are immediately activated to ensure the safety of the baby.

[0119] In this embodiment, a quantum annealing algorithm is used to optimize execution parameters, and precise and smooth cradle movement is achieved through Bezier curve interpolation and electromagnetic suspension counterweight technology.

[0120] In some possible implementations of the present invention, the steps of collaborative control of IoT devices and collaborative computing of edge and cloud are further included, specifically:

[0121] When the baby is detected entering a specific state or a state transition is predicted, the blockchain smart contract triggers the associated devices to execute chain operations, including: adjusting the intelligent temperature control system to establish a radiant temperature field centered on the cradle; activating the directional sound wave transmitter to form an active noise reduction zone above the cradle; and pre-adjusting the indoor light intensity and color temperature based on the baby's sleep cycle prediction.

[0122] Execute real-time control algorithms at edge computing nodes, and regularly upload desensitized status data and control effect evaluations to the cloud analysis platform.

[0123] The cloud platform trains deep learning models based on multi-instance data and regularly sends updated model parameters to edge nodes to achieve continuous iterative optimization of the model.

[0124] In this embodiment, the collaborative operation of the cradle and the environmental control equipment is achieved through blockchain smart contracts, and a layered computing architecture combining edge computing and cloud analysis is established.

[0125] In some possible implementations of the present invention, the following further aspects are included:

[0126] Establish a verifiable log of the control process on the blockchain platform, encrypting the timestamps of key control decisions, sensor data fingerprints, and execution parameters and storing them in a distributed ledger;

[0127] Data traceability and operation auditing are achieved through smart contracts to ensure the system's tamper-proof operation and medical-grade safety compliance;

[0128] Differential privacy technology is used to desensitize data uploaded to the cloud, balancing data availability and privacy protection.

[0129] In this embodiment, blockchain verifiable logs and differential privacy technology are integrated to balance system security and data privacy protection.

[0130] In some possible implementations of the present invention, all multimodal physiological data and environmental parameter groups are processed through edge computing nodes for signal noise reduction and spatiotemporal alignment, a spatiotemporal correlation feature matrix of multi-source heterogeneous data is established, and a full-dimensional digital twin model of the cradle and its surrounding environment is formed, including the following steps:

[0131] Configure a time synchronization protocol to synchronize the clocks of all sensor nodes in the biometric sensor group and the environmental perception sensor group at the microsecond level using the Precision Time Protocol (PTP) or the Network Time Protocol (NTP) to establish a unified time base (ensuring the temporal consistency of sensor data with different sampling frequencies, laying the foundation for subsequent spatiotemporal alignment).

[0132] Sensor-specific denoising algorithms are applied to the multimodal physiological data collected by the biometric sensor group (including millimeter-wave radar, infrared thermal imaging, flexible piezoelectric sensor strips, and acoustic sensor arrays): adaptive bandpass filtering and phase noise suppression algorithms are applied to the millimeter-wave radar signals to improve the signal-to-noise ratio of weak vital sign signals; non-local mean filtering and background temperature compensation algorithms are applied to the infrared thermal imaging data to eliminate interference from ambient temperature fluctuations; wavelet threshold denoising and baseline drift correction are performed on the flexible piezoelectric sensor strip signals to extract a stable respiratory waveform; beamforming and blind source separation algorithms are applied to the acoustic sensor array signals to isolate infant sounds from ambient noise;

[0133] This step is a customized preprocessing for the characteristics of various sensors, maximizing the retention of effective signals while suppressing noise interference.

[0134] The environmental parameter set collected by the environmental perception sensor group (including a six-axis inertial measurement unit, lidar, temperature and humidity sensor, light intensity sensor, and air quality sensor) is calibrated and filtered. The accelerometer and gyroscope data are fused through Kalman filtering to calculate the cradle's precise three-dimensional posture and motion trajectory. Voxel filtering and the RANSAC algorithm are applied to the lidar point cloud data to remove outliers and extract environmental structural features. Median filtering and trend analysis are used to eliminate instantaneous fluctuations in environmental parameters and extract stable environmental features.

[0135] This step can reduce environmental parameter noise, extract stable features, and ensure the accuracy of environmental status assessment.

[0136] Establish a unified timestamp reference system to map sensor data with different sampling frequencies to a unified time axis: define the system master clock frequency fs (usually 100 Hz) as the reference frequency for data synchronization; apply the cubic spline interpolation algorithm to upsample sensor data below fs (such as infrared thermal imaging and environmental parameters); and apply wavelet decomposition and reconstruction to downsample sensor data above fs (such as acoustic signals and millimeter-wave radar) to preserve signal characteristics.

[0137] Implement a delay compensation mechanism to address signal transmission and processing delays for different sensors: System calibration measures the delay characteristics of each sensor and establishes a delay compensation table. An adaptive delay estimation algorithm is used to adjust compensation parameters in real time to ensure time consistency in dynamic environments.

[0138] This step can eliminate the time deviation of different sensors, establish an accurate time sequence relationship, and provide a time basis for multimodal data fusion.

[0139] A three-dimensional coordinate system with the cradle's geometric center as its origin is constructed, and all spatially relevant data is uniformly mapped to this coordinate system. The cradle's own geometric parameters and sensor installation location information are used to establish a transformation matrix between the sensor's local coordinate system and the cradle's global coordinate system. An iterative closest point (ICP) registration algorithm is applied to the 3D point cloud of the environment generated by the lidar to align it with the cradle's coordinate system.

[0140] Construct a mapping model for infant physiological parameters: Based on a standard infant human body model, map infrared thermal imaging data to a human body surface model; combine data from force-sensitive sensors within the cradle to estimate the infant's posture and position within the cradle; and establish a dynamic transformation relationship between the cradle coordinate system and the infant's body coordinate system.

[0141] This step can achieve a unified spatial representation of data from different sensors and establish an accurate mapping relationship between physical space and data space.

[0142] Feature extraction is performed on multi-source heterogeneous data after spatiotemporal alignment: adaptive window segmentation is applied to physiological parameter data to calculate time domain features (mean, variance, peak), frequency domain features (power spectral density, main frequency component), and nonlinear features (sample entropy, Poincare plot parameters); statistical features, gradient features, and time-varying features are calculated for environmental parameter data to characterize environmental status and changing trends; trajectory features, vibration features, and stability features are extracted from cradle motion data;

[0143] Constructing a spatiotemporal correlation feature matrix M: Define the structure of the feature matrix M, where the row dimension corresponds to the time series and the column dimension corresponds to the feature vectors of different sensors. For each time point t, select the sub-feature set with the highest information gain from each sensor feature to form the feature vector of time point t. Calculate the correlation and mutual information between feature vectors at adjacent time points to capture temporal dependencies. Continuously update the feature matrix using a sliding window method, with the window size dynamically adjusted based on the analysis task (short windows are used for real-time state assessment, long windows are used for pattern recognition).

[0144] This step converts discrete sensor data into structured spatiotemporal feature representations, reduces data dimensionality, extracts key information, and provides a compact representation for subsequent analysis.

[0145] A digital mapping of the cradle's physical system is constructed based on a spatiotemporal correlation matrix. A parametric physical model of the cradle's mechanical system is established, including dynamic models of the rocking mechanism, shock absorption system, and drive unit. Combined with real-time motion data, the recursive least squares method is used to estimate the physical model parameters and achieve dynamic identification of physical characteristics.

[0146] Construct a parametric model of the infant's physiological state: Based on physiological parameter characteristics, establish a physiological feedback model that describes the infant's heart rate regulation, body temperature regulation, and respiratory regulation; use state-space representation to describe the coupling relationship and dynamic characteristics between various physiological systems;

[0147] Establish an environment-cradle-infant ternary interaction model: define the transfer function from environmental parameters to the microenvironment within the cradle; establish a response model from cradle movement to infant physiological feedback; and represent the causal relationship and information flow in the ternary system through a graph structure.

[0148] In this embodiment, the time deviation and spatial reference system inconsistency problems of different sensor data are solved through clock synchronization and coordinate system conversion, which provides a basic guarantee for multimodal data fusion and significantly improves the accuracy of state recognition; customized preprocessing algorithms are used for different sensor characteristics, combined with a multi-layer anomaly detection mechanism, which effectively suppresses the interference caused by environmental noise and sensor failure, improves the robustness of the system in complex environments, and makes comfort assessment more reliable; by constructing a spatiotemporal correlation feature matrix, the system not only captures the static characteristics of each parameter, but also retains the spatiotemporal correlation relationship and dynamic evolution characteristics between parameters, providing a more comprehensive information basis for subsequent infant state recognition and improving recognition accuracy; the digital twin model realizes the virtual expression of the cradle-infant-environment ternary system. The system can accurately reflect the behavior and state of the physical system in a virtual environment, providing a reliable platform for predictive control and simulation analysis; based on the constructed digital twin model, the system can simulate predictive Measure the response of the system under different control strategies, evaluate the control effect in advance, transform the control decision from passive response to active prediction, and reduce control delay; through feature extraction and matrix structure optimization, convert massive amounts of raw sensor data into structured, moderate-dimensional feature representations, significantly reducing the computing and storage burden while retaining key information, enabling edge computing nodes to process complex multimodal data streams in real time; the system continuously optimizes model parameters by comparing the differences between model predictions and actual observations in real time, so that the digital twin model can adaptively track changes in the physical system and maintain long-term accuracy and effectiveness; the multi-strategy data compensation mechanism enables the system to maintain its function when some sensors fail or data is missing, and fills information gaps through data complementarity and model prediction, greatly improving the system's reliability and continuous operation capability; the parameterized representation of the digital twin model enables the system to be personalized according to the physiological characteristics and behavioral patterns of different infants, achieving precise adaptation of "one infant, one model". This embodiment realizes the conversion from discrete and heterogeneous raw sensor data to a unified and structured spatiotemporal information representation by constructing a high-quality spatiotemporal correlation feature matrix and digital twin model, laying a solid foundation for the precise state perception and predictive control of the smart cradle, and significantly improving the adaptability, accuracy and reliability of the smart cradle.

[0149] In some possible implementations of the present invention, the following further aspects are included:

[0150] Implement online adaptive updates of the model: Calculate the update amount for model parameters by comparing the error between measured data and the model's predicted output; Use a gradient descent algorithm with an adaptive learning rate to dynamically adjust the parameter update step size based on the rate of state change; Set parameter change rate limits to ensure the smoothness and stability of model updates; This step constructs a digital twin model that reflects the dynamic characteristics of the physical world, achieving real-time mapping and synchronization between virtual and reality.

[0151] Design a multi-layered anomaly detection mechanism: at the sensor level, detect sensor failures or abnormal readings based on historical statistical characteristics; at the feature level, use the isolation forest algorithm to identify anomalies in the feature space; at the model level, monitor the deviation between the digital twin model prediction and actual observations to detect system anomalies;

[0152] Implement multi-strategy data compensation: For short-term data loss, use forward filling or linear interpolation methods to supplement; for long-term data loss, use the predicted values generated by the digital twin model to replace it; for the coordinated loss of multi-source data, start the degradation mode and reconfigure the feature matrix structure; this step improves the system's fault tolerance to sensor failures and data anomalies, ensuring the continuity and reliability of the digital twin model.

[0153] Implement a tiered data storage strategy: maintain a high sampling rate for key features required for real-time control and store them in a fast-access cache; compress and store historical data required for long-term analysis using principal component analysis dimensionality reduction and time series compression algorithms;

[0154] Design feature indexing and fast retrieval mechanisms: Establish a multi-level index structure based on timestamps and feature types; implement efficient query methods for specific time windows and feature subsets; this step balances storage efficiency and access speed to support the dual needs of real-time control and historical analysis.

[0155] In some possible embodiments of the present invention, the adaptive fusion algorithm is used to integrate the collected multimodal physiological data into a composite comfort index Q, and the time series change gradient of the Q value is calculated. The steps to establish a dynamic assessment model of infant status include:

[0156] Adaptive wavelet transform is applied to the original heart rate signal in the multimodal physiological data for noise reduction. The heart rate variability (HRV) feature sequence is extracted, including SDNN (standard deviation of consecutive heartbeat intervals), RMSSD (root mean square of the difference between adjacent heartbeat intervals) and pNN50 (the percentage of adjacent heartbeat intervals with a difference greater than 50ms), to form the heart rate feature vector F_HRV(t);

[0157] The infrared thermal imaging data in the multimodal physiological data is used to construct the infant's body surface temperature distribution matrix T(x, y, t). The environmental interference is eliminated by two-dimensional Gaussian filtering, and the neck-limb temperature difference ΔT_NL and the core area temperature gradient vector are calculated. Forming the temperature characteristic vector F_TEMP(t);

[0158] Perform spectrum analysis on the respiratory waveform in the multimodal physiological data (collected by a flexible piezoelectric respiratory sensor belt) to extract the respiratory frequency RF, respiratory depth RD, inspiration / expiration ratio RI / RE and respiratory irregularity index RII to form the respiratory feature vector F_RESP(t);

[0159] Mel-frequency cepstral coefficients (MFCC) and acoustic event detection algorithms are used to extract sound features from the sound data in multimodal physiological data, identify crying, comfort sounds and environmental noise, calculate the acoustic comfort index ASI, and form the acoustic feature vector F_ACOU(t);

[0160] The aforementioned eigenvectors (including heart rate eigenvector, temperature eigenvector, respiration eigenvector and acoustic eigenvector) are adaptively normalized so that physiological indicators of different scales are mapped to a unified [0, 1] interval:

[0161]

[0162] Among them, min_i and max_i are the historical minimum and maximum values updated online, and are dynamically adjusted through the exponential moving average algorithm to adapt to individual differences and developmental changes of infants;

[0163] Construct a multi-layer attention network to calculate the feature weight matrix W(t);

[0164] Establish four sub-comfort index mapping functions;

[0165] In this embodiment, the specific method of establishing the four sub-comfort index mapping functions is:

[0166] The normalized heart rate variability characteristics are mapped into a cardiac comfort sub-index, and a quadratic function model is used to reflect the nonlinear relationship between HRV and comfort.

[0167] The normalized temperature characteristics are mapped into thermal comfort sub-indexes, and the Gaussian function model is used to express the optimal temperature distribution;

[0168] The normalized breathing characteristics are mapped into a breathing comfort sub-index, and a piecewise function model is used to reflect the comfort state corresponding to different breathing patterns.

[0169] Normalized acoustic features are mapped into acoustic comfort sub-indexes, and support vector regression model is used to infer comfort level from acoustic features.

[0170] The composite comfort index Q(t) is calculated by combining the feature weight matrix W(t) with the mapping functions of each sub-comfort index;

[0171] The central difference method is used to calculate the time series gradient of the Q value: Where Δt is the sampling time interval (usually set to 100-500 milliseconds);

[0172] Combining Q(t) and Construct a two-dimensional state space and divide it into multiple state regions:

[0173] a) Stable comfort zone: high Q value and Close to zero, indicating that the infant is in a state of continuous comfort;

[0174] b) Comfort improvement zone: Q value is medium to high and If it is positive, it means that the baby's comfort level is improving;

[0175] c) Comfortable descent zone: Q value is medium to high but If it is negative, it means that the baby's comfort level is beginning to decrease and needs to be closely monitored;

[0176] d) Uncomfortable area: low Q value and A negative value indicates that the baby is in an uncomfortable state and the trend is worsening, requiring intervention;

[0177] e) Recovery zone: Q value is low but If it is positive, it means that the discomfort is improving;

[0178] A hybrid model combining long short-term memory (LSTM) and Gaussian process regression (GPR) is used to predict the future trend of Q(t). The LSTM network captures the long-term and short-term dependencies of the Q value sequence and predicts short-term changes in comfort. The GPR provides a probabilistic prediction framework and outputs the predicted mean and uncertainty. The prediction model is updated using a sliding window method, and the window length is dynamically adjusted according to the infant's sleep-wake cycle.

[0179] In this embodiment, by integrating heart rate variability, body surface temperature distribution, respiratory characteristics and acoustic characteristics, a comprehensive assessment of the infant's physiological state is achieved, overcoming the limitations of single parameter monitoring and providing a more comprehensive description of comfort. A multi-layer attention network is used to dynamically adjust the weights of each physiological characteristic, enabling the system to automatically adjust the assessment strategy according to different scenarios and individual characteristics, thereby improving the adaptability and personalization of the index. By calculating the temporal gradient of the comfort index Combined with the hybrid model of LSTM and GPR, the system can predict the trend of comfort changes in advance, provide an early warning time window for active intervention, and realize the transformation from passive response to active prevention; based on Q and The constructed two-dimensional state space finely divides the infant's state into multiple regions, enabling the control system to adopt differentiated strategies for different state regions, significantly improving control accuracy and efficiency. Through baseline calibration and online learning mechanisms, the system can adapt to individual differences and developmental changes in infants, continuously self-optimizing as usage increases, thereby enhancing its long-term value. Multimodal data fusion improves the system's robustness to single sensor failure or interference, while layered feature extraction and normalization enhance the system's resistance to noise and outliers. Through online feature selection and attention mechanisms, the system can adjust computing resource allocation in real time, reducing the computational burden while ensuring evaluation accuracy, enabling edge nodes to process complex physiological data streams in real time. This composite comfort index construction method transforms discrete multimodal sensor data into continuous, interpretable state indicators through a systematic and intelligent data processing and analysis process, laying a solid foundation for the precise control of the smart cradle and significantly improving the smart cradle's ability to perceive and respond to infant needs.

[0180] In some possible implementations of the present invention, the following further aspects are included:

[0181] Establish an individualized baseline calibration mechanism: record the statistical characteristics of the infant's Q value distribution in different states and construct an individualized Q value reference spectrum; automatically adjust the mapping function parameters according to age and developmental milestones to adapt to changes in the infant's developmental stage;

[0182] Implement an online learning mechanism to continuously optimize the evaluation model: evaluate model accuracy by comparing the time difference between Q-value prediction and actual state transition through cross-validation; use guardian feedback and intervention effect data as supervision signals to optimize model parameters; and use the Bayesian optimization algorithm to adjust the hyperparameters of the attention network and state mapping function to maximize prediction accuracy.

[0183] In some possible implementations of the present invention, the method for generating the hierarchical state recognition model includes:

[0184] Establish a three-level nested neural network architecture, including a low-level physiological state recognition network, a mid-level emotional state analysis network, and a high-level demand intention reasoning network;

[0185] Design a common feature extraction layer, using a universal feature mapping network across infant groups to capture common physiological response patterns among different infants;

[0186] Constructing an individual adaptive fine-tuning layer, using a parameter-efficient meta-learning structure that can quickly adapt to the individual differences of specific infants;

[0187] Deploy local training modules on the local edge computing nodes of each cradle to ensure that the original physiological data does not leave the local device;

[0188] Establish a secure aggregation protocol so that local nodes only share model gradients instead of raw data;

[0189] Deploy a differential privacy mechanism to add calibration noise to the gradient information to prevent user data from being reconstructed through the gradient;

[0190] Use pre-trained convolutional neural networks to extract time-frequency features from multimodal physiological signals;

[0191] Establish a universal representation space for physiological signals through self-supervised learning methods;

[0192] Apply attention mechanism to dynamically adjust and fuse the weights of different sensor signals;

[0193] Gradient descent is used to optimize a specific parameter subset on the local edge device, keeping the common characteristic network parameters unchanged;

[0194] Using a meta-learning framework to achieve few-shot learning, model adaptation can be completed with only a small amount of individual data; designing a memory-enhanced neural network structure to store and utilize the historical response patterns of specific infants;

[0195] Build a multi-task learning architecture that shares underlying representations and simultaneously outputs wakefulness-sleep state levels, emotion types, and comfort scores;

[0196] Using a soft parameter sharing strategy, each task network maintains some independent parameters to capture task-specific features;

[0197] Formulate a cross-attention mechanism between tasks so that the outputs of different tasks can promote and constrain each other.

[0198] In some possible implementations of the present invention, the method for generating the hierarchical state recognition model further includes:

[0199] Privacy-preserving model updates: Design a secure multi-party computing protocol to aggregate model updates from different edge nodes; apply homomorphic encryption technology to protect model parameter transmission and aggregation processes; implement a verifiable computing mechanism to ensure the correctness and integrity of model updates.

[0200] Continuous optimization and knowledge transfer: Deploy a model version control system to track the evolution path of the global model and individual models; use knowledge distillation technology to compress and transfer the knowledge of the global model to the individual model; establish a model quality assessment framework to monitor model performance changes through preset indicators to ensure that the optimization process does not deviate from the target.

[0201] The embodiments of the present invention can realize local data processing, and the original sensitive physiological data does not leave the device; the differential privacy mechanism ensures that individual data cannot be reconstructed even when model parameters are shared; homomorphic encryption ensures data security during the model update process; the common feature extraction layer takes advantage of group data to solve the problem of insufficient data of a single infant; the individual adaptation layer realizes fine-tuning accurate to the individual, improving recognition accuracy; the multi-task learning framework enables different recognition tasks to promote each other and reduce the risk of overfitting; the hierarchical design reduces computational complexity and enables complex models to run on edge devices; the meta-learning framework achieves rapid adaptation and reduces the amount of data and time required for personalized training; knowledge distillation technology compresses complex models and reduces storage and computing requirements; the continuous learning architecture enables the model to be dynamically adjusted as the infant grows and develops, and knowledge transfer across infant groups accelerates model convergence for new users. Model version control ensures system stability and prevents catastrophic forgetting; the three-level output (wakefulness-sleep state, emotion type, comfort score) provides a comprehensive state assessment and provides an explainable decision basis for subsequent control strategy generation.

[0202] In some possible implementations of the present invention, the calculation formula of the composite comfort index Q is:

[0203]

[0204] Among them, the weight w_i(t) reflects the contribution of each physiological feature to the overall comfort at the current moment. The specific steps of constructing a multi-layer attention network to calculate the feature weight matrix W(t) are:

[0205] Capturing temporal dependencies within a single modality (e.g., periodic changes in respiratory rhythm) through self-attention mechanisms;

[0206] Establish correlations between different physiological features (e.g., the synergistic relationship between heart rate changes and respiratory rate) through cross-modal attention mechanisms;

[0207] The gated recurrent unit (GRU) is used to fuse historical weight information to ensure the temporal continuity of weight changes;

[0208] Output dynamic weight vector W(t)=[w_HRV(t), w_TEMP(t), w_RESP(t), w_ACOU(t)], where each weight satisfies ∑w_i(t)=1.

[0209] In some possible implementations of the present invention, the dynamic optimization process of the three-dimensional control vector includes:

[0210] Establish the state-action value function Q(s, a) between the cradle control parameters and the infant's physiological feedback;

[0211] Reduce the deviation of Q-value estimation through the Double Temporal Difference Learning algorithm;

[0212] Adopting the experience replay mechanism to store historical state transition samples and improve sample utilization efficiency;

[0213] A curiosity-driven exploration strategy is introduced to actively explore new control parameter combinations to avoid falling into local optimal solutions.

[0214] In some possible implementations of the present invention, the working mechanism of the abnormal state prediction module includes:

[0215] Construct a multi-scale temporal convolutional network to simultaneously capture short-term, medium-term, and long-term physiological parameter change patterns; design a multi-parameter fusion algorithm based on the attention mechanism to dynamically adjust the weights of different physiological indicators;

[0216] The Bayesian probability graphical model is used to establish the conditional probability distribution of abnormal states, and the early warning mechanism is triggered when the probability of abnormal states exceeds the threshold.

[0217] In some possible implementations of the present invention, the collaborative control of IoT devices further includes:

[0218] Adjust home environment parameters based on the baby's sleep cycle and status prediction (15-30 minutes in advance);

[0219] Build a personalized environmental preference model based on the baby's emotional response, including optimal temperature range, light intensity, and background music type;

[0220] When sudden interference from the external environment is detected, a multi-layer protection mechanism is activated, including active noise reduction, vibration isolation, and environmental parameter stabilization control.

[0221] In some possible implementations of the present invention, the edge-cloud collaborative computing architecture further includes:

[0222] The local edge computing node continuously optimizes the personalized model through online incremental learning methods, and can adapt to changes in the baby's condition without relying on cloud connection;

[0223] The cloud platform optimizes the control strategy library through a multi-objective genetic algorithm, taking into account comfort improvement, energy consumption reduction, and adaptability enhancement;

[0224] Compressed knowledge representation technology is used to reduce the communication overhead of model updates and achieve efficient model iteration under low bandwidth conditions.

[0225] In some possible implementations of the present invention, quantum optimization execution and trajectory control steps are further included, specifically:

[0226] The actuator control parameters are optimized by quantum annealing algorithm, and the three-dimensional control vector is converted into a pulse sequence of the multi-axis servo motor;

[0227] The Bezier curve interpolation algorithm is used to smooth the swing trajectory to simulate the three-dimensional composite trajectory of the mother's embrace;

[0228] Dynamically adjust the electromagnetic suspension position of the counterweight according to the baby's weight distribution to achieve precise motion control and adaptive balance adjustment with optimal energy consumption;

[0229] When the abnormal state prediction module triggers an early warning, the progressive braking mechanism and emergency posture adjustment are immediately activated to ensure the safety of the baby.

[0230] See Figure 2 ,Another embodiment of the present invention provides a control system for a smart cradle, for executing a control method for a smart cradle, comprising: a distributedly deployed biometric sensor group, an environmental perception sensor group, an edge computing node, and a server;

[0231] The biometric sensor group is configured to: collect multimodal physiological data of the infant in real time;

[0232] The environmental perception sensor group is configured to: obtain an environmental parameter group inside and outside the cradle;

[0233] The edge computing node is configured to: perform signal noise reduction and spatiotemporal alignment processing on all multimodal physiological data and environmental parameter groups, establish a spatiotemporal correlation feature matrix of multi-source heterogeneous data, and form a full-dimensional digital twin model of the cradle and its surrounding environment;

[0234] The server is configured to:

[0235] Adaptive fusion algorithm is used to integrate the collected multimodal physiological data into a composite comfort index Q, and the temporal change gradient of the Q value is calculated at the same time. Establish a dynamic assessment model for infant status;

[0236] Input the spatiotemporal correlation feature matrix into the pre-trained hierarchical state recognition model;

[0237] The hierarchical state recognition model outputs the current infant's wakefulness-sleep state level, emotion type determination, and comprehensive comfort score through the collaborative calculation of the common feature extraction layer across infant groups and the individual adaptive fine-tuning layer;

[0238] An abnormal state prediction module is built based on a temporal convolutional network to continuously monitor Q-value change patterns and environmental parameter mutation events to predict possible abnormal states.

[0239] The composite comfort index Q and the environmental parameter group are input into the pre-trained deep reinforcement learning model, and a three-dimensional control vector is dynamically generated according to the recognized infant state level;

[0240] Combined with the three-dimensional control vector, differentiated control strategies are adopted for different sleep-wake states;

[0241] The control parameter group is continuously optimized through the online Q-learning algorithm to maximize the comfort score function while balancing the immediate response and long-term optimization goals.

[0242] It should be known that Figure 2 The block diagram of the control system for the smart cradle is for illustration only, and the number of modules shown does not limit the scope of protection of the present invention. The control system for the smart cradle provided in this embodiment can be used to execute the corresponding embodiments of the control method for the smart cradle. For the specific implementation process, please refer to the description of the embodiments of each method, which will not be repeated here.

[0243] It should be noted that for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application.

[0244] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0245] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the above-mentioned units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0246] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0247] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0248] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a memory, including a number of instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the above-mentioned methods of each embodiment of the present application. The aforementioned memory includes: various media that can store program codes, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0249] Those skilled in the art will appreciate that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program. The program can be stored in a computer-readable memory, and the memory can include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0250] The above is a detailed introduction to the embodiments of the present application. Specific examples are used herein to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core idea of the present application. At the same time, for those skilled in the art, according to the idea of the present application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

[0251] Although the present invention is disclosed above, it is not limited thereto. Any person skilled in the art may readily conceive of variations or substitutions, and may make various modifications and alterations without departing from the spirit and scope of the present invention. Combinations of the above-described functions and implementation steps, including software and hardware implementations, are all within the scope of protection of the present invention.

Claims

1. A control method for an intelligent cradle, characterized in that: include: The distributed biometric sensor group collects the infant's multimodal physiological data in real time; Adaptive fusion algorithm is used to integrate the collected multimodal physiological data into a composite comfort index Q, and the temporal change gradient of the Q value is calculated at the same time. Establish a dynamic assessment model for infant status; Acquire the environmental parameter group inside and outside the cradle through the environmental perception sensor group; All multimodal physiological data and environmental parameter groups are processed through edge computing nodes for signal noise reduction and spatiotemporal alignment, establishing a spatiotemporal correlation feature matrix of multi-source heterogeneous data and forming a full-dimensional digital twin model of the cradle and its surrounding environment; Input the spatiotemporal correlation feature matrix into the pre-trained hierarchical state recognition model; The hierarchical state recognition model outputs the current infant's wakefulness-sleep state level, emotion type determination, and comprehensive comfort score through the collaborative calculation of the common feature extraction layer across infant groups and the individual adaptive fine-tuning layer; An abnormal state prediction module is built based on a temporal convolutional network to continuously monitor Q-value change patterns and environmental parameter mutation events to predict possible abnormal states. The composite comfort index Q and the environmental parameter group are input into the pre-trained deep reinforcement learning model, and a three-dimensional control vector is dynamically generated according to the recognized infant state level; Combined with the three-dimensional control vector, differentiated control strategies are adopted for different sleep-wake states; The control parameter group is continuously optimized through the online Q-learning algorithm to maximize the comfort score function while balancing the immediate response and long-term optimization goals.

2. The control method for the smart cradle according to claim 1, characterized in that: The steps for performing signal noise reduction and spatiotemporal alignment processing on all multimodal physiological data and environmental parameter groups through edge computing nodes, establishing a spatiotemporal correlation feature matrix of multi-source heterogeneous data, and forming a full-dimensional digital twin model of the cradle and its surrounding environment include: Configure a time synchronization protocol to synchronize the clocks of all sensor nodes in the biometric sensor group and the environmental perception sensor group at the microsecond level using the Precision Time Protocol or the Network Time Protocol to establish a unified time base. performing sensor-specific denoising algorithms on multimodal physiological data collected by a biometric sensor array; Performing calibration and filtering on the environmental parameter group collected by the environmental perception sensor group; Establish a unified timestamp reference system to map sensor data with different sampling frequencies to a unified time axis; Implement a delay compensation mechanism to solve the problem of signal transmission and processing delays of different sensors; Construct a three-dimensional coordinate system with the geometric center of the cradle as the origin, and uniformly map all spatially related data to this coordinate system; Construct a mapping model for infant physiological parameters: Based on a standard infant human body model, map infrared thermal imaging data to a human body surface model; combine data from force-sensitive sensors within the cradle to estimate the infant's posture and position within the cradle; and establish a dynamic transformation relationship between the cradle coordinate system and the infant's body coordinate system. Extract features from multi-source heterogeneous data after spatiotemporal alignment; Construct the spatiotemporal correlation feature matrix M; Construct a digital mapping of the cradle's physical system based on a spatiotemporal correlation feature matrix; Construct a parametric model of the infant's physiological state; Establish an environment-cradle-infant three-way interaction model.

3. The control method for the smart cradle according to claim 2, characterized in that: The adaptive fusion algorithm is used to integrate the collected multimodal physiological data into a composite comfort index Q, and the time series change gradient of the Q value is calculated. The steps to establish a dynamic assessment model of infant status include: Adaptive wavelet transform was applied to the original heart rate signal in the multimodal physiological data for noise reduction. The characteristic sequence of heart rate variability was extracted, including the standard deviation of consecutive heartbeat intervals, the root mean square of the difference between adjacent heartbeat intervals, and the percentage of adjacent heartbeat intervals with a difference greater than 50ms, to form the heart rate feature vector F_HRV(t). The infrared thermal imaging data in the multimodal physiological data was used to construct the infant's body surface temperature distribution matrix T(x,y,t). The environmental interference was eliminated through two-dimensional Gaussian filtering, and the neck-limb temperature difference ΔT_NL and the core area temperature gradient vector were calculated. Forming the temperature characteristic vector F_TEMP(t); The respiratory waveforms in the multimodal physiological data were spectrally analyzed to extract the respiratory frequency (RF), respiratory depth (RD), inspiratory / expiratory ratio (RI / RE), and respiratory irregularity index (RII), forming a respiratory feature vector (F_RESP(t)). Mel-frequency cepstral coefficients and an acoustic event detection algorithm were used to extract sound features from the sound data in the multimodal physiological data, identify crying, comfort sounds, and ambient noise, and calculate the acoustic comfort index (ASI) to form an acoustic feature vector (F_ACOU(t). Adaptively normalize the aforementioned eigenvectors to map physiological indicators of different scales to a unified [0, 1] interval: Among them, min_i and max_i are the historical minimum and maximum values updated online, and are dynamically adjusted through the exponential moving average algorithm to adapt to individual differences and developmental changes of infants; Construct a multi-layer attention network to calculate the feature weight matrix W(t); Establish four sub-comfort index mapping functions; The composite comfort index Q(t) is calculated by combining the feature weight matrix W(t) with the mapping functions of each sub-comfort index; The central difference method is used to calculate the time series gradient of the Q value: Wherein, Δt is the sampling time interval; Combining Q(t) and Construct a two-dimensional state space and divide it into multiple state regions; A hybrid model combining long short-term memory network and Gaussian process regression is used to predict the future trend of Q(t).

4. The control method for the intelligent cradle according to claim 3, characterized in that: The method for generating the hierarchical state recognition model includes: Establish a three-level nested neural network architecture, including a low-level physiological state recognition network, a mid-level emotional state analysis network, and a high-level demand intention reasoning network; Design a common feature extraction layer, using a universal feature mapping network across infant groups to capture common physiological response patterns among different infants; Constructing an individual adaptive fine-tuning layer, using a parameter-efficient meta-learning structure that can quickly adapt to the individual differences of specific infants; Deploy local training modules on the local edge computing nodes of each cradle to ensure that the original physiological data does not leave the local device; Establish a secure aggregation protocol so that local nodes only share model gradients instead of raw data; Deploy a differential privacy mechanism to add calibration noise to the gradient information to prevent user data from being reconstructed through the gradient; Use pre-trained convolutional neural networks to extract time-frequency features from multimodal physiological signals; Establish a universal representation space for physiological signals through self-supervised learning methods; Apply attention mechanism to dynamically adjust and fuse the weights of different sensor signals; Gradient descent is used to optimize a specific parameter subset on the local edge device, keeping the common characteristic network parameters unchanged; Using a meta-learning framework to achieve few-shot learning, only a small amount of individual data is needed to complete model adaptation; Designing memory-enhancing neural network structures that store and utilize historical response patterns of specific infants; Build a multi-task learning architecture that shares underlying representations and simultaneously outputs wakefulness-sleep state levels, emotion types, and comfort scores; Using a soft parameter sharing strategy, each task network maintains some independent parameters to capture task-specific features; Formulate a cross-attention mechanism between tasks so that the outputs of different tasks can promote and constrain each other.

5. The control method for the smart cradle according to claim 4, characterized in that: The calculation formula of the composite comfort index Q is: Among them, the weight w_i(t) reflects the contribution of each physiological feature to the overall comfort at the current moment. The specific steps of constructing a multi-layer attention network to calculate the feature weight matrix W(t) are: Capturing temporal dependencies within a single modality through self-attention mechanisms; Establish correlations between different physiological features through cross-modal attention mechanism; Adopt gated recursive units to fuse historical weight information to ensure the temporal continuity of weight changes; Output dynamic weight vector W(t)=[w_HRV(t), w_TEMP(t), w_RESP(t), w_ACOU(t)], where each weight satisfies ∑w_i(t)=1.

6. The control method for the intelligent cradle according to claim 5, characterized in that: The dynamic optimization process of the three-dimensional control vector includes: Establish the state-action value function Q(s, a) between the cradle control parameters and the infant's physiological feedback; Reduce the deviation of Q-value estimation through double temporal difference learning algorithm; Adopting the experience replay mechanism to store historical state transition samples and improve sample utilization efficiency; A curiosity-driven exploration strategy is introduced to actively explore new control parameter combinations to avoid falling into local optimal solutions.

7. The control method for the intelligent cradle according to claim 6, characterized in that: The working mechanism of the abnormal state prediction module includes: Construct a multi-scale temporal convolutional network to simultaneously capture short-term, medium-term, and long-term physiological parameter change patterns; design a multi-parameter fusion algorithm based on the attention mechanism to dynamically adjust the weights of different physiological indicators; The Bayesian probability graphical model is used to establish the conditional probability distribution of abnormal states, and the early warning mechanism is triggered when the probability of abnormal states exceeds the threshold.

8. The control method for the intelligent cradle according to claim 7, characterized in that: The collaborative control of IoT devices also includes: Adjust home environment parameters based on the baby's sleep cycle and state prediction; Build a personalized environmental preference model based on the baby's emotional response, including optimal temperature range, light intensity, and background music type; When sudden interference from the external environment is detected, a multi-layer protection mechanism is activated, including active noise reduction, vibration isolation, and environmental parameter stabilization control.

9. The control method for the intelligent cradle according to claim 8, characterized in that: It also includes quantum optimization execution and trajectory control steps, specifically: The actuator control parameters are optimized by quantum annealing algorithm, and the three-dimensional control vector is converted into a pulse sequence of the multi-axis servo motor; The Bezier curve interpolation algorithm is used to smooth the swing trajectory to simulate the three-dimensional composite trajectory of the mother's embrace; Dynamically adjust the electromagnetic suspension position of the counterweight according to the baby's weight distribution to achieve precise motion control and adaptive balance adjustment with optimal energy consumption; When the abnormal state prediction module triggers an early warning, the progressive braking mechanism and emergency posture adjustment are immediately activated to ensure the safety of the baby.

10. A control system for an intelligent cradle, configured to execute the control method for an intelligent cradle according to any one of claims 1 to 9, characterized in that: include: Distributed deployment of biometric sensor groups, environmental perception sensor groups, edge computing nodes and servers; The biometric sensor group is configured to: collect multimodal physiological data of the infant in real time; The environmental perception sensor group is configured to: obtain an environmental parameter group inside and outside the cradle; The edge computing node is configured to: perform signal noise reduction and spatiotemporal alignment processing on all multimodal physiological data and environmental parameter groups, establish a spatiotemporal correlation feature matrix of multi-source heterogeneous data, and form a full-dimensional digital twin model of the cradle and its surrounding environment; The server is configured to: Adaptive fusion algorithm is used to integrate the collected multimodal physiological data into a composite comfort index Q, and the temporal change gradient of the Q value is calculated at the same time. Establish a dynamic assessment model for infant status; Input the spatiotemporal correlation feature matrix into the pre-trained hierarchical state recognition model; The hierarchical state recognition model outputs the current infant's wakefulness-sleep state level, emotion type determination, and comprehensive comfort score through the collaborative calculation of the common feature extraction layer across infant groups and the individual adaptive fine-tuning layer; An abnormal state prediction module is built based on a temporal convolutional network to continuously monitor Q-value change patterns and environmental parameter mutation events to predict possible abnormal states. The composite comfort index Q and the environmental parameter group are input into the pre-trained deep reinforcement learning model, and a three-dimensional control vector is dynamically generated according to the recognized infant state level; Combined with the three-dimensional control vector, differentiated control strategies are adopted for different sleep-wake states; The control parameter group is continuously optimized through the online Q-learning algorithm to maximize the comfort score function while balancing the immediate response and long-term optimization goals.

Images