Intelligent pet collar based on unconstrained inertial sensing and physiological signal processing method thereof

Abstract

The invention provides an intelligent pet necklace based on unconstrained inertial sensing and a physiological signal processing method thereof. The intelligent pet necklace is worn on the neck of a pet in an unconstrained mode, a gap change of 0-30 mm is allowed to exist between the necklace and the skin, and the necklace is allowed to rotate freely relative to the body. An inertial measurement unit is arranged in the necklace to collect multi-axis acceleration data, and the multi-axis acceleration data is converted into scalar signals irrelevant to the wearing direction through spatial irrelevant fusion; self-adaptive frequency domain adjustment is carried out according to the physiological features of the target pet, and motion noise and breathing interference are inhibited; further separating cardiovascular micro-vibration components from the mixed signals based on signal statistical independence characteristics through a blind source separation algorithm; and finally, cardiovascular activity parameters including heart rate, heart rate variability, heart beat interval sequence and the like are extracted. According to the method, high-robustness cardiovascular micro-vibration signal extraction and health monitoring are realized only by depending on a general inertial sensor and an advanced signal processing algorithm under the conditions of random wearing posture and uncertain direction.

Description

Technical Field

[0001] This invention relates to the field of wearable smart pet devices, and in particular to a smart pet collar based on unconstrained inertial sensing and its physiological signal processing method. Background Technology

[0002] With the development of the pet economy, the demand for health monitoring of companion animals is increasing. Smart pet collars, as the most common form of wearable pet device, offer multiple functions such as activity monitoring, health tracking, and location tracking. Cardiac impaction imaging is a physiological signal reflecting the mechanical pulsation of the heart and the micro-vibrations caused by blood jets. It has the advantages of being non-invasive and electrode-free, making it suitable for long-term health monitoring of animals.

[0003] There are various smart pet collar solutions in the existing technology, but all of them have certain technical limitations:

[0004] Existing technology 1: This method uses an elastic protrusion on the sensor housing, requiring the protruding part of the sensor to face the animal's neck skin. This approach has the following problems: the elastic protrusion needs to directly contact the skin to transmit vibration signals, which places strict requirements on the tightness of the fit; the sensor orientation is fixed and cannot adapt to the rotation of the collar during animal activity; the protruding structure may cause discomfort to the animal's skin.

[0005] Existing technology 2: Integrating a speech recognition translation unit, attempting to translate animal sounds into human language. This approach suffers from the following problems: high system complexity and high power consumption; the speech translation function lacks scientific basis, and its practicality is questionable; stacking multiple functions into a single device makes it difficult to optimize the core functionality.

[0006] Existing technology 3: A dual-speed comparison scheme using a collar and leg-wearing components, which determines the animal's movement status by comparing the collar's movement speed with the leg's movement speed. This scheme has the following problems: it requires additional leg-wearing components, increasing system complexity and the burden on the animal; the collaborative operation of the two components increases potential points of failure; and it is not suitable for animals that find it inconvenient to wear leg-wearing components.

[0007] Existing technology 4: Using shape memory alloys to achieve adaptive adjustment of collar tightness. This approach has the following problems: shape memory alloys are expensive and require precise temperature control; mechanical adjustment mechanisms increase system complexity and weight; frequent adjustments may affect the animal's wearing comfort.

[0008] Therefore, there is an urgent need in this field for a smart pet collar with a simple structure, comfortable wear, comprehensive functions, and no need for special sensor orientation or additional wearable components, as well as a method for processing its physiological signals. Summary of the Invention

[0009] To address at least one of the technical problems mentioned above, this invention provides an intelligent pet collar based on unconstrained inertial sensing and its physiological signal processing method. Under unconstrained wearing conditions where the device posture is random and the direction is uncertain, it can achieve multiple functions such as cardiovascular micro-vibration signal extraction, activity monitoring, sleep detection, posture recognition, and health warning by relying solely on general-purpose inertial sensors.

[0010] The present invention solves the technical problem by adopting the following technical solution:

[0011] A smart pet collar based on unconstrained inertial sensing, comprising:

[0012] The smart pet collar body is wrapped around the pet's neck in a non-restrained manner, and an inertial measurement unit is installed inside the smart pet collar body;

[0013] The signal processing unit, which is communicatively connected to the smart pet collar body, is configured to execute:

[0014] (a) Data acquisition: Receive multi-axis acceleration time series data from the inertial measurement unit;

[0015] (b) Spatial Independence Fusion: Performing a direction-invariant transformation on the multi-axis acceleration data to generate a composite energy signal independent of the wearing direction of the smart pet collar; the direction-invariant transformation includes calculating at least one of the Euclidean norm, weighted sum of squares, or time rate of change of vector magnitude of the multi-axis acceleration.

[0016] (c) Adaptive frequency domain adjustment: Configure bandpass filter parameters according to the physiological characteristics of the target pet to suppress non-cardiovascular motion noise and respiratory interference; the physiological characteristics include at least one of species, body size, a priori range of cardiac rate, and a priori range of respiratory rate;

[0017] (d) Blind source separation: The filtered signal is input into the computational inference model, and the cardiovascular micro-vibration component is separated from the mixed signal based on the statistical independence or non-Gaussianity characteristics of the signal.

[0018] (e) Parameter extraction: Based on the separated cardiovascular micro-vibration components, the cardiovascular activity parameters of the pet are determined and output.

[0019] Furthermore, the unconstrained method allows the distance between the smart pet collar body and the pet's skin to vary within the range of 0~30mm; the unconstrained method also allows the smart pet collar body to rotate relative to the pet's body in any direction.

[0020] Furthermore, the signal processing unit adopts any of the following structural forms:

[0021] An embedded processor integrated within the smart pet collar body;

[0022] A mobile terminal device that wirelessly communicates with the smart pet collar body;

[0023] A cloud computing server connected via the Internet.

[0024] Furthermore, the computational inference model includes at least one of the following algorithms: independent component analysis; principal component analysis; singular value decomposition; convolutional neural network; recurrent neural network or long short-term memory network; and attention-based Transformer architecture.

[0025] Furthermore, it also includes a motion state determination module, which is configured as follows:

[0026] Determining whether a pet is at rest based on statistical characteristics of multi-axis acceleration data;

[0027] The blind source separation step is performed only when the system is at rest.

[0028] Furthermore, methods for extracting heart rate from cardiovascular micro-vibration components include:

[0029] Perform autocorrelation analysis on the separated signal and take the time delay corresponding to the first significant peak as the heartbeat cycle; or perform FFT transformation on the separated signal and take the peak frequency within the heartbeat frequency band as the heart rate; or use a peak detection algorithm to detect characteristic peaks and calculate the interval between adjacent peaks.

[0030] A physiological signal processing method for a smart pet collar based on unconstrained inertial sensing includes the following steps:

[0031] S1, acquire inertial sensor data from the body of the smart pet collar worn around the pet's neck in an unrestrained manner, the inertial sensor data including a multi-axis acceleration time series; wherein the unrestrained manner allows the distance between the collar and the skin to vary in the range of 0-30mm, and allows the collar to rotate arbitrarily relative to the body;

[0032] S2, Perform a spatially independent fusion operation on the multi-axis acceleration data to generate a scalar energy signal that is independent of the wearing direction by merging the multi-axis components;

[0033] S3, apply an adaptive bandpass filter based on the physiological characteristics of the target pet to the scalar energy signal to suppress motion noise and respiratory interference;

[0034] S4 inputs the filtered signal into the blind source separation model, and separates the cardiovascular micro-vibration component from the mixed signal based on the statistical independence or non-Gaussianity characteristics of the signal.

[0035] S5, quantify cardiovascular activity parameters based on the cardiovascular micro-vibration components;

[0036] S6, output the cardiovascular activity parameters.

[0037] Furthermore, the spatially independent fusion operation includes at least one of the following operations:

[0038] Calculate the Euclidean norm of triaxial acceleration;

[0039] Calculate the magnitude of the weighted vector;

[0040] After performing coordinate transformation to eliminate the gravitational component, the dynamic acceleration amplitude is calculated.

[0041] Furthermore, the blind source separation model is trained based on any of the following methods:

[0042] Supervised learning was performed using a labeled cardiovascular signal dataset.

[0043] Unsupervised learning based on signal statistical properties using unlabeled data;

[0044] Adaptive parameter configuration is performed for different species.

[0045] Furthermore, it also includes:

[0046] The quality of the extracted cardiovascular activity parameters was assessed.

[0047] When the quality is below a preset threshold, the data for that period is marked as low confidence.

[0048] The beneficial effects of this invention are:

[0049] This invention eliminates the need for sensors to be in close contact with the skin or in a specific orientation, improving pet comfort; it eliminates the need for elastic protruding sensing structures, leg-wearing components, or SMA adjustment mechanisms, reducing cost and complexity; it ensures consistent signal extraction across any rotational direction through spatially independent fusion; all functions can be performed with a single gonioscope, eliminating the need for multiple components; it integrates cardiovascular monitoring, activity monitoring, sleep monitoring, body temperature monitoring, and location tracking into a single device; the signal processing algorithm can be optimized through software updates without requiring hardware replacement; and it has broad applicability, suitable for various pets such as cats, dogs, and rabbits. Attached Figure Description

[0050] Figure 1 This is a structural block diagram of the intelligent pet collar based on unconstrained inertial sensing according to the present invention.

[0051] Figure 2 This is a flowchart of the physiological signal processing method for the intelligent pet collar based on unconstrained inertial sensing according to the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] refer to Figure 1 This invention provides a smart pet collar based on unconstrained inertial sensing, comprising:

[0054] The smart pet collar body is wrapped around the pet's neck in a non-restrained manner, and an inertial measurement unit is installed inside the smart pet collar body;

[0055] The signal processing unit, which is communicatively connected to the smart pet collar body, is configured to execute:

[0056] (a) Data acquisition: Receive multi-axis acceleration time series data from the inertial measurement unit;

[0057] (b) Spatial Independence Fusion: Performing a direction-invariant transformation on the multi-axis acceleration data to generate a composite energy signal independent of the wearing direction of the smart pet collar; the direction-invariant transformation includes calculating at least one of the Euclidean norm, weighted sum of squares, or time rate of change of vector magnitude of the multi-axis acceleration.

[0058] (c) Adaptive frequency domain adjustment: Configure bandpass filter parameters according to the physiological characteristics of the target pet to suppress non-cardiovascular motion noise and respiratory interference; the physiological characteristics include at least one of species, body size, a priori range of cardiac rate, and a priori range of respiratory rate;

[0059] (d) Blind source separation: The filtered signal is input into the computational inference model, and the cardiovascular micro-vibration component is separated from the mixed signal based on the statistical independence or non-Gaussianity characteristics of the signal.

[0060] (e) Parameter extraction: Based on the separated cardiovascular micro-vibration components, determine and output the cardiovascular activity parameters of the pet. The output of cardiovascular activity parameters includes at least: BCG waveform / feature points, instantaneous / average heart rate, HRV, and heart interval sequence.

[0061] The technical solution is further optimized so that the unconstrained method allows the distance between the smart pet collar body and the pet's skin to vary within the range of 0~30mm; the unconstrained method also allows the smart pet collar body to rotate relative to the pet's body in any direction.

[0062] To further optimize the technical solution, the signal processing unit adopts any of the following structural forms:

[0063] An embedded processor integrated within the smart pet collar body;

[0064] A mobile terminal device that wirelessly communicates with the smart pet collar body;

[0065] A cloud computing server connected via the Internet.

[0066] Further optimize the technical solution. The bandpass filter parameters in the adaptive frequency domain adjustment step are configured according to the pet species as follows: 1.5-4.5Hz for felines; 1.2-3.5Hz for small canines (weight <15kg); and 0.7-2.0Hz for large canines (weight ≥15kg).

[0067] A complete and reproducible bandpass filter configuration typically includes at least:

[0068] A. Frequency boundary parameter: Lower passband cutoff frequency f p1 , (Hz); passband cutoff frequency f p2 (Hz); Stopband cutoff frequency f s1 (Hz); Stopband cutoff frequency f s2 , (Hz).

[0069] B. Filter structure parameters: Filter type, IIR (Butterworth / Chebyshev / Elliptic) or FIR (window function method / equiripple); Order N (IIR order or FIR tap number); Passband ripple A P (dB)

[0070] Stopband attenuation A S (dB); digitization method, bilinear transform (IIR) or Parks–McClellan (FIR).

[0071] C. Implementation details: whether to remove phase distortion, forward-backward zero-phase filtering (such as the FLTFILT method), window and overlap, window length T, step size / overlap rate (for streaming processing); boundary processing, mirror extension / zero padding (to reduce endpoint artifacts).

[0072] To further optimize the technical solution, the method for configuring bandpass filter parameters based on the physiological characteristics of the target pet is as follows:

[0073] (1) Convert the heart rate range to the cardiac frequency range

[0074] Convert heart rate (bpm) to frequency (Hz):

[0075]

[0076] Where HR stands for heart rate, measured in bpm; f HR This represents the frequency corresponding to heart rate, measured in Hz.

[0077] If the target pet's psychological pre-hoc range (HR) is known. min ~HR max The range of heart rate is:

[0078] ,

[0079] (2) Passband settings (add margin to enhance robustness)

[0080] Considering individual differences, transient heart rate changes, and algorithm errors, a margin is set. , :

[0081] ,

[0082] in: The cutoff frequency in the passband (Hz); The cutoff frequency in the passband (Hz); , All values ​​are margins (Hz), which can be taken as 0.2–0.8 Hz (by species / body size). To prevent excessively low frequencies from causing motion / drift (e.g., 0.5–1.0 Hz); To prevent excessive noise introduction (e.g., 6–8 Hz, depending on f) s (and sensor noise), f s The noise is caused by the specifications of the hardware sensor.

[0083] (3) Stopband settings (used to suppress motion noise and breathing interference)

[0084] The respiratory rate is usually lower. Convert respiratory rate (breaths / minute) to Hz:

[0085]

[0086] Where RR is the respiratory rate, measured in bpm; f RR This represents the frequency corresponding to breathing, measured in Hz.

[0087] To suppress the respiratory fundamental frequency and its slow body motion components, it is recommended that the lower boundary of the blockade band cover the vicinity of the upper respiratory boundary and leave a margin:

[0088]

[0089] in, The passband and stopband spacing (Hz) is, for example, 0.2 to 0.5 Hz; Respiratory inhibition margin (Hz), for example, 0.1–0.3 Hz;

[0090]

[0091] in, This is the upper limit of respiratory rate (breaths / minute); This is the frequency corresponding to the upper bound of the respiratory rate.

[0092] High-frequency stopbands are used to suppress mechanical vibrations, collar impacts, and high-frequency noise from sensors.

[0093]

[0094] in, The frequency range can be 0.5 to 2.0 Hz (depending on the sampling rate and noise conditions) (to prevent the filter design from being too aggressive, this uses the buffer band reserved during the filter design, similar to mechanical vibration, collar collision, and high-frequency noise from the sensor).

[0095] Further optimization of the technical solution to determine whether a pet is in a resting state includes: calculating the standard deviation of triaxial acceleration within a time window T; when the standard deviation is less than a threshold σ th When T is 3-30 seconds, it is determined to be a resting state; where T is 3-30 seconds, σ th It is 0.005g-0.05g.

[0096] The technical solution was further optimized, and the sampling parameters of the inertial measurement unit were configured as follows: sampling 50Hz-500Hz; acceleration range ±2g to ±8g; data bit width 12-bit-16-bit ADC.

[0097] Further optimize the technical solution, and the computational inference model includes at least one of the following algorithms: independent component analysis; principal component analysis; singular value decomposition; convolutional neural network; recurrent neural network or long short-term memory network; and Transformer architecture based on attention mechanism.

[0098] 1) ICA (Independent Component Analysis) analysis process

[0099] Input construction: The three-axis acceleration and its derivatives (such as amplitude fusion, detrended / bandpass signal) are combined to form a multi-channel observation matrix.

[0100] Preprocessing: remove mean and standardize; if necessary, perform whitening to reduce inter-channel correlation and improve separation stability.

[0101] Independent component estimation: Find statistically independent components through iterative optimization and output several independent source signal candidates.

[0102] Component selection: Calculate the energy percentage, spectral peak significance, period consistency and other quality indicators of each component in the target heart rate frequency band, and select the component that best matches the characteristics of cardiovascular micro-vibration.

[0103] Output: Outputs cardiovascular micro-vibration components (for subsequent peak detection / autocorrelation / spectral peak extraction of heart rate and HRV).

[0104] 2) PCA (Principal Component Analysis) Analysis Process

[0105] Input construction: The multi-channel signals are aligned in time to form a feature matrix.

[0106] Centralization / standardization: Eliminate the influence of dimensions and DC bias.

[0107] Dimensionality reduction decomposition: Identify the principal directions (principal components) and extract the common change patterns with the highest energy.

[0108] Component selection: Select principal components that match the heart rate frequency and have weak motion artifacts (usually one or a combination of the first few principal components).

[0109] Output: Output the denoised principal component signal as the input for heart rate / HRV extraction.

[0110] 3) SVD (Singular Value Decomposition) Analysis Process

[0111] Constructing data matrices: Matrices can be constructed using channel × time or delay embedding / sliding window stacking to enhance periodic structures.

[0112] Decomposition and Energy Sorting: The matrix is ​​decomposed into several modes, and the modes are sorted according to their energy.

[0113] Modal screening and reconstruction: The modes that are most stable in the target frequency band and are clearly distinguishable from the motion-related modes are selected for reconstruction.

[0114] Output: A smoother, more clearly defined cardiovascular component is obtained for subsequent parameter extraction.

[0115] 4) CNN (Convolutional Neural Network) Analysis Process

[0116] Input format: Use the filtered single-channel or multi-channel time series, or its time-frequency plot (such as short-time spectrum) as the model input.

[0117] Feature extraction: Convolutional layers automatically learn local waveform morphology (heartbeat micro-vibration features) and noise reduction modes.

[0118] Task output (choose one or a combination):

[0119] Regression output: Directly outputs heart rate / respiratory rate and confidence level;

[0120] Sequence labeling output: Output the probability sequence of cardiac events, and then obtain the IBI sequence and HRV by peak localization.

[0121] Quality control: Use output confidence / consistency rules to eliminate high-motion windows or low-quality results.

[0122] Output: Heart rate, IBI sequence (if labeled), and confidence level.

[0123] 5) RNN / LSTM Analysis Process

[0124] Input sequence: Input the time-series features (original / filtered signal, energy, spectral peaks, motion intensity, etc.) after sliding windowing according to time steps.

[0125] Temporal modeling: RNN / LSTM uses memory to capture the smooth changes in heart rate over time, reducing jumps caused by transient noise.

[0126] Output: Continuous heart rate estimation sequence, abnormal state markers (optional), and confidence level; prediction of interpeak sequence can also be output.

[0127] Post-processing: Smoothing and threshold constraints are applied to the estimated sequence (within the physiological range of the species).

[0128] 6) Transformer (Attention Mechanism) Analysis Process

[0129] Input representation: The time series is sliced ​​into several segments / Tokens (which can be original sample point segments or time-frequency blocks), and position encoding is added.

[0130] Attention modeling: By learning long-range dependencies through self-attention mechanisms, it is possible to capture stable heart rhythms even in the presence of intermittent motion artifacts.

[0131] Output format:

[0132] Directly output heart rate / HRV metrics and confidence levels; or

[0133] Output the heartbeat event sequence (probability / location), and then calculate IBI and HRV.

[0134] Robustness strategy: Motion intensity / posture category can be used as an additional token or conditional input to improve adaptability to body movement disturbances.

[0135] The above algorithms can be used individually or in combination (e.g., PCA / SVD pre-denoising + ICA separation or ICA separation + CNN peak detection). The system outputs heart rate, IBI sequence, and confidence level within each sliding window; when the confidence level is below a threshold, it is marked as invalid or only a trend value is output to ensure the reliability of the results.

[0136] The present invention also includes a motion state determination module, configured as follows:

[0137] Determining whether a pet is at rest based on statistical characteristics of multi-axis acceleration data;

[0138] The blind source separation step is performed only when the system is at rest.

[0139] The present invention also includes an activity monitoring module, configured as follows:

[0140] The number of steps or the distance traveled by the pet is counted based on the multi-axis acceleration data;

[0141] Calculate the pet's exercise intensity index, which is based on the time integral or frequency domain energy of the acceleration amplitude;

[0142] Statistical analysis of pet activity levels within a preset time period;

[0143] Generate pet activity reports or activity trend analyses.

[0144] The present invention also includes a sleep detection module, configured as follows:

[0145] The sleep state of a pet is identified based on the motion characteristics of the multi-axis acceleration data;

[0146] Distinguish between a pet's deep sleep, light sleep, and awake state;

[0147] Statistical analysis of pets' sleep duration and sleep quality indicators;

[0148] Detect your pet's sleep cycles and sleep patterns.

[0149] The present invention also includes a posture recognition module, configured as follows:

[0150] The pet's body posture is identified based on the gravity and dynamic components in the multi-axis acceleration data.

[0151] Body posture includes at least one of standing, walking, running, lying down, curling up, and lying on one's side;

[0152] Record the percentage of time the pet spends in various postures;

[0153] Detect abnormal changes in a pet's posture.

[0154] The present invention also includes an anomaly warning module, configured as follows:

[0155] When a pet's heart rate is detected to exceed the preset normal range, an abnormal heart rate warning is generated.

[0156] When a pet's activity level is detected to be significantly lower than the historical average, an abnormal activity alert is generated.

[0157] When the time a pet remains still exceeds a preset threshold, an alert for prolonged stillness is generated.

[0158] When a pet's body temperature is detected to be outside the preset normal range, an abnormal body temperature warning is generated.

[0159] The warning information is sent to the user terminal device through the wireless communication module.

[0160] The present invention also includes a positioning assistance module, which includes at least one of the following positioning methods:

[0161] The GPS satellite positioning module, Beidou satellite positioning module, base station positioning module, WiFi fingerprint positioning module, Bluetooth beacon positioning module, and positioning assistance module are configured to obtain the pet's geographical location information and report it to the user terminal or cloud server via the wireless communication module.

[0162] The wireless communication module supports at least one of the following communication protocols: Bluetooth Low Energy (BLE); WiFi; LoRa; NB-IoT; 4G / 5G cellular networks.

[0163] The present invention also includes a body temperature monitoring module, configured as follows:

[0164] The temperature of the pet's skin surface is collected by a temperature sensor located inside the collar body;

[0165] The temperature sensor is a non-contact infrared temperature sensor or a contact thermistor temperature sensor.

[0166] The collected skin surface temperatures were corrected to estimate the pet's core body temperature.

[0167] Record the trend of pet's body temperature changes;

[0168] An alert is triggered when the body temperature exceeds the preset normal range.

[0169] Further optimization of the technical solution, methods for extracting heart rate from cardiovascular micro-vibration components include:

[0170] Perform autocorrelation analysis on the separated signal and take the time delay corresponding to the first significant peak as the heartbeat cycle; or perform FFT transformation on the separated signal and take the peak frequency within the heartbeat frequency band as the heart rate; or use a peak detection algorithm to detect characteristic peaks and calculate the interval between adjacent peaks.

[0171] Method 1, the autocorrelation method, is as follows:

[0172] For window signals Calculate autocorrelation:

[0173]

[0174] Where: n is the sampling point index; k is the delay (points);

[0175]

[0176] Where T is the analysis window length (e.g., 10~30s); is the sampling rate; N is the total number of sampling points in the window signal that participate in the autocorrelation calculation.

[0177] Within the corresponding heart rate range, the delay interval Find the first significant peak:

[0178]

[0179] in:

[0180] ,

[0181] Obtain cardiac cycle B and heart rate HR:

[0182] , .

[0183] Method 2: Frequency domain peak method (simple to implement, suitable for stable resting states) as follows:

[0184] The power spectrum is obtained by performing an FFT on the window. Find the peak within the passband:

[0185]

[0186] Suitable for quiet environments with good signal quality.

[0187] Method 3: Time-domain peak detection method (used to output IBI sequence and HRV) is as follows:

[0188] Peak detection: In Or detect characteristic peaks (e.g., the J peak) on its envelope.

[0189] Set the minimum peak interval (corresponding to maximum heart rate):

[0190]

[0191] Setting an adaptive threshold (example):

[0192]

[0193] in, 2 to 4 are acceptable.

[0194] Peak time sequence: ,but

[0195] , .

[0196] It is the time when the i-th characteristic peak appears. It is the time interval between the i-th heartbeat and the (i+1)-th heartbeat. It is the instantaneous heart rate.

[0197] HRV index calculation (based on IBI sequence)

[0198] Let M be the number of heartbeat intervals (in seconds). For i=1...M, the following can be calculated:

[0199] (1) SDNN (Standard Deviation)

[0200]

[0201] in, for The mean.

[0202] (2) RMSSD (Root Mean Square Difference Between Adjacent Differences)

[0203]

[0204] (3) pNNx

[0205]

[0206] x can be 0.02s, 0.05s, etc. (adjustable according to the pet's heart rate scale).

[0207] pNNx is the percentage (indicating that the absolute value of the difference between two adjacent IBIs (or NN intervals) is greater than a threshold x. It represents a standard parameter of HRV (Heart Rate Variability) time-domain indices, used to measure the severity of changes in adjacent heartbeat intervals.

[0208] Further optimize the technical solution, output BCG waveform feature points (to improve the completeness of parameter extraction):

[0209] Define a local window within each cardiac cycle. Use derivatives / second derivatives or wavelets to locate feature points (e.g., I, J, K points). Example (primarily J peak):

[0210]

[0211] in Here, is a time parameter, representing the time position of the J-peak within the j-th cardiac cycle;

[0212] Amplitude:

[0213] Morphological characteristics: slope of the rising edge, peak width, interpeak spacing, etc.

[0214] These features can be used as input for subsequent health analysis or anomaly detection.

[0215] The blind source separation step employs the FastICA algorithm for Independent Component Analysis (ICA), which includes:

[0216] (a) Preprocessing of FastICA

[0217] FastICA typically involves two steps: centralization and whitening. These steps are implementation details but are important for patent enforceability.

[0218] Centering (mean removal): Subtract the mean from each channel within the current window to remove the DC component and slow drift bias.

[0219] Whitening (decorrelation and normalization): This process whitens the centered multi-channel data to minimize linear correlation between channels and unify variance across directions. PCA / SVD can be used for whitening in engineering applications.

[0220] First, calculate the covariance and then perform eigenvalue decomposition or SVD;

[0221] Then, the data is linearly transformed using the inverse square root of the eigenvalues.

[0222] Whitening output: Obtain the whitened data matrix, which serves as the input for the main iteration of FastICA.

[0223] (II) FastICA Main Iteration Process

[0224] The core of FastICA is to iteratively find a weight vector that maximizes the non-Gaussianity of the projected signal, thereby approximating statistical independence.

[0225] 1) Number of independent components: How to implement 2-4?

[0226] In each window, set the number of independent components to be estimated. .

[0227] Choose a strategy:

[0228] Fixed method: default m=3 (corresponding to three axes), or take m=min(4, number of channels) according to the number of channels of the device;

[0229] Adaptive method: Select based on the energy ratio of the characteristic values ​​of the whitening stage (excluding directions with lower energy), limiting m to 2~4.

[0230] Output: m independent candidate signals are obtained.

[0231] 2) Choice of nonlinear function: tanh(u) or u·exp(-u² / 2)

[0232] In FastICA's weight update, a non-linear function is required. To characterize non-Gaussianity. Implementation depends on the scenario:

[0233] Option A: tanh(u)

[0234] Suitable for signals that have some outliers but are generally stable;

[0235] It has fast convergence speed and good numerical stability, and is commonly used in embedded systems.

[0236] Option B: u·exp(-u² / 2)

[0237] More sensitive to small morphological changes against a background of near-Gaussian noise;

[0238] It is helpful to extract weak periodic micro-vibrations from strong noise, but it is more sensitive to the numerical range, so it is recommended to use it in conjunction with standardization.

[0239] In practice, a switch parameter is usually provided: nonlinearity="tanh" or "gauss" (corresponding to the second function), which is selected by the system based on the pet's size / signal quality or offline calibration results.

[0240] 3) Weight vector update and orthogonalization (must be clearly written when there are multiple components)

[0241] For each independent component, maintain a weight vector w. i .

[0242] After each iteration update, w needs to be updated. i Perform normalization and apply it to multiple w. i Perform decorrelation / orthogonalization (e.g., Gram-Schmidt or symmetric orthogonalization of the weight matrix) to prevent different components from converging in the same direction.

[0243] This is crucial for achieving stable output of 2 to 4 components in engineering; otherwise, duplicate components will appear.

[0244] 4) Iterative convergence criterion: Cosine value of the included angle > 0.99999.

[0245] The given criterion can be directly implemented as follows:

[0246] Let the weight vectors before and after the iteration be respectively and Calculate the cosine of the angle between the two (which is essentially the absolute value of the normalized inner product).

[0247] When the value is greater than 0.99999, the component is considered to have converged; if all components have converged, the iteration stops.

[0248] Project Supplement:

[0249] At the same time, a maximum number of iterations (e.g., 200~1000 times) is set as a fallback to avoid blocking caused by extreme noise windows failing to converge.

[0250] If the threshold is not met even after reaching the maximum iteration, the current optimal weight is output and the window is marked as having low confidence.

[0251] (III) Output of Separation Results and Selection of Cardiovascular Components

[0252] FastICA outputs 2-4 independent components, but the system needs to automatically identify which one is the cardiovascular micro-vibration component. This can be achieved as follows:

[0253] Frequency band energy ratio screening: Calculate the energy ratio (or spectral peak significance) of each component within the target heart rate frequency band, and prioritize the component with the highest ratio.

[0254] Periodic consistency screening: Perform short-time autocorrelation or interpeak period statistics on each component, and prioritize the component with the most stable period.

[0255] Motion artifact elimination: If the component has excessively high energy at low frequencies (body movement / posture changes), or is highly correlated with the acceleration amplitude channel, then reduce its priority.

[0256] Output: The selected cardiovascular components are fed into subsequent parameter extraction (heart rate, IBI, HRV, etc.), and the confidence level and component number are output for debugging / traceability.

[0257] Further optimization of the technical solution may yield 2-4 components after blind source separation (range given in the document). The component most similar to a heartbeat should be selected, and its confidence level should be provided.

[0258] (1) Percentage of energy in the heartbeat frequency band (SQI example)

[0259] For components power spectrum ,choose The largest component is the cardiovascular component.

[0260]

[0261] (2) Peak sequence consistency

[0262] For the IBI sequence obtained from peak detection, outliers (such as those exceeding the median ± 3MAD) are removed, and the effective peak rate is calculated; if the effective peak rate is low, the output is marked as low confidence (consistent with the subordinate scheme of document quality assessment / low confidence marking).

[0263] refer to Figure 2 The present invention also provides a physiological signal processing method for a smart pet collar based on unconstrained inertial sensing, comprising the following steps:

[0264] S1, acquire inertial sensor data from the body of the smart pet collar worn around the pet's neck in an unrestrained manner, the inertial sensor data including a multi-axis acceleration time series; wherein the unrestrained manner allows the distance between the collar and the skin to vary in the range of 0-30mm, and allows the collar to rotate arbitrarily relative to the body;

[0265] S2, Perform a spatially independent fusion operation on the multi-axis acceleration data to generate a scalar energy signal that is independent of the wearing direction by merging the multi-axis components;

[0266] S3, apply an adaptive bandpass filter based on the physiological characteristics of the target pet to the scalar energy signal to suppress motion noise and respiratory interference;

[0267] S4 inputs the filtered signal into the blind source separation model, and separates the cardiovascular micro-vibration component from the mixed signal based on the statistical independence or non-Gaussianity characteristics of the signal.

[0268] S5, quantify cardiovascular activity parameters based on the cardiovascular micro-vibration components;

[0269] S6, output the cardiovascular activity parameters.

[0270] To further optimize the technical solution, the spatially independent fusion operation includes at least one of the following operations:

[0271] Calculate the Euclidean norm of triaxial acceleration;

[0272] Calculate the magnitude of the weighted vector;

[0273] After performing coordinate transformation to eliminate the gravitational component, the dynamic acceleration amplitude is calculated.

[0274] To further optimize the technical solution, the blind source separation model is trained based on any of the following methods:

[0275] Supervised learning was performed using a labeled cardiovascular signal dataset.

[0276] Unsupervised learning based on signal statistical properties using unlabeled data;

[0277] Adaptive parameter configuration is performed for different species.

[0278] The present invention also includes:

[0279] The quality of the extracted cardiovascular activity parameters was assessed.

[0280] When the quality is below a preset threshold, the data for that period is marked as low confidence.

[0281] Quality assessment is based on at least one or more of the following categories of indicators:

[0282] Signal layer quality indicators: energy concentration of the cardiovascular component within the target frequency band; distinguishability from low-frequency body motion components; periodic stability of the separated components.

[0283] Parameter consistency indicators: whether the change in heart rate between adjacent time windows is smooth and whether it exceeds the physiologically reasonable range; the proportion of abnormal intervals (too short or too long) in the cardiac interval sequence.

[0284] Motion disturbance indicators: whether the overall amplitude or variance of acceleration exceeds a preset threshold within the same time period; the percentage of duration of rapid changes in posture or violent movements.

[0285] Overall score: The above indicators are combined according to preset weights or rules to obtain the quality score or grade for this time window.

[0286] Threshold comparison: The quality score is compared with a preset quality threshold to determine whether the reliable output conditions are met.

[0287] Time continuity check: can further check the score changes of multiple consecutive windows, avoiding frequent state switching caused by a single abnormal window.

[0288] Low confidence labeling: When the quality is below the threshold, the cardiovascular parameters for the corresponding time period are labeled as low confidence.

[0289] Output strategy adjustment: You can output only the trend value instead of the precise value; or pause the update of heart rate / HRV and only keep the last high confidence result.

[0290] Downstream modules are instructed to synchronously transmit confidence level markers to display, storage, or health assessment modules to avoid misjudgments.

[0291] This invention eliminates the need for sensors to be in close contact with the skin or in a specific orientation, improving pet comfort; it eliminates the need for elastic protruding sensing structures, leg-wearing components, or SMA adjustment mechanisms, reducing cost and complexity; it ensures consistent signal extraction across any rotational direction through spatially independent fusion; all functions can be performed with a single gonioscope, eliminating the need for multiple components; it integrates cardiovascular monitoring, activity monitoring, sleep detection, body temperature monitoring, and location tracking into a single device; the signal processing algorithm can be optimized through software updates without requiring hardware replacement; and it has broad applicability, suitable for various pets such as cats, dogs, and rabbits.

[0292] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart pet collar based on unconstrained inertial sensing, characterized in that, include: The smart pet collar body is wrapped around the pet's neck in a non-restrained manner, and an inertial measurement unit is installed inside the smart pet collar body; The signal processing unit, which is communicatively connected to the smart pet collar body, is configured to execute: (a) Data acquisition: Receive multi-axis acceleration time series data from the inertial measurement unit; (b) Spatial Independence Fusion: Performing a direction-invariant transformation on the multi-axis acceleration data to generate a composite energy signal independent of the wearing direction of the smart pet collar; the direction-invariant transformation includes calculating at least one of the Euclidean norm, weighted sum of squares, or time rate of change of vector magnitude of the multi-axis acceleration. (c) Adaptive frequency domain adjustment: Configure bandpass filter parameters according to the physiological characteristics of the target pet to suppress non-cardiovascular motion noise and respiratory interference; the physiological characteristics include at least one of species, body size, a priori range of cardiac rate, and a priori range of respiratory rate; (d) Blind source separation: The filtered signal is input into the computational inference model, and the cardiovascular micro-vibration component is separated from the mixed signal based on the statistical independence or non-Gaussianity characteristics of the signal. (e) Parameter extraction: Based on the separated cardiovascular micro-vibration components, the cardiovascular activity parameters of the pet are determined and output.

2. The smart pet collar based on unconstrained inertial sensing according to claim 1, characterized in that, The unconstrained method allows the distance between the smart pet collar body and the pet's skin to vary within the range of 0~30mm; the unconstrained method also allows the smart pet collar body to rotate relative to the pet's body in any direction.

3. The smart pet collar based on unconstrained inertial sensing according to claim 1, characterized in that, The signal processing unit adopts any of the following structural forms: An embedded processor integrated within the smart pet collar body; A mobile terminal device that wirelessly communicates with the smart pet collar body; A cloud computing server connected via the Internet.

4. The smart pet collar based on unconstrained inertial sensing according to claim 1, characterized in that, The computational inference model includes at least one of the following algorithms: independent component analysis; principal component analysis; singular value decomposition; convolutional neural network; recurrent neural network or long short-term memory network; and attention-based Transformer architecture.

5. A smart pet collar based on unconstrained inertial sensing according to claim 1, characterized in that, It also includes a motion state determination module, which is configured as follows: Determining whether a pet is at rest based on statistical characteristics of multi-axis acceleration data; The blind source separation step is performed only when the system is at rest.

6. The smart pet collar based on unconstrained inertial sensing according to claim 1, characterized in that, Methods for extracting heart rate from cardiovascular micro-vibration components include: Perform autocorrelation analysis on the separated signal and take the time delay corresponding to the first significant peak as the heartbeat cycle; or perform FFT transformation on the separated signal and take the peak frequency within the heartbeat frequency band as the heart rate; or use a peak detection algorithm to detect characteristic peaks and calculate the interval between adjacent peaks.

7. A physiological signal processing method for an intelligent pet collar based on unconstrained inertial sensing, characterized in that, Includes the following steps: S1, acquire inertial sensor data from the body of the smart pet collar worn around the pet's neck in an unrestrained manner, the inertial sensor data including a multi-axis acceleration time series; wherein the unrestrained manner allows the distance between the collar and the skin to vary in the range of 0-30mm, and allows the collar to rotate arbitrarily relative to the body; S2, Perform a spatially independent fusion operation on the multi-axis acceleration data to generate a scalar energy signal that is independent of the wearing direction by merging the multi-axis components; S3, apply an adaptive bandpass filter based on the physiological characteristics of the target pet to the scalar energy signal to suppress motion noise and respiratory interference; S4 inputs the filtered signal into the blind source separation model, and separates the cardiovascular micro-vibration component from the mixed signal based on the statistical independence or non-Gaussianity characteristics of the signal. S5, quantify cardiovascular activity parameters based on the cardiovascular micro-vibration components; S6, output the cardiovascular activity parameters.

8. The physiological signal processing method for a smart pet collar based on unconstrained inertial sensing according to claim 7, characterized in that, The space-independent fusion operation includes at least one of the following operations: Calculate the Euclidean norm of triaxial acceleration; Calculate the magnitude of the weighted vector; After performing coordinate transformation to eliminate the gravitational component, the dynamic acceleration amplitude is calculated.

9. The physiological signal processing method for a smart pet collar based on unconstrained inertial sensing according to claim 8, characterized in that, The blind source separation model is trained based on any of the following methods: Supervised learning was performed using a labeled cardiovascular signal dataset. Unsupervised learning based on signal statistical properties using unlabeled data; Adaptive parameter configuration is performed for different species.

10. The physiological signal processing method for a smart pet collar based on unconstrained inertial sensing according to claim 9, characterized in that, Also includes: The quality of the extracted cardiovascular activity parameters was assessed. When the quality is below a preset threshold, the data for that period is marked as low confidence.

Images