Earphone heart rate error calibration method based on motion knowledge graph and related device

Abstract

The present application relates to the technical field of wearable devices, in particular to a method for calibrating earphone heart rate error based on a sports knowledge graph and related devices. The method comprises the following steps: obtaining multi-modal physiological data of a user wearing an earphone during sports, the multi-modal physiological data including photoplethysmographic wave signals collected by an optical heart rate sensor and the like. The present application structures the complex mapping relationship between different sports state types, sports intensity levels and heart rate error characteristics by constructing a sports knowledge graph, thereby achieving explicit modeling of the correlation between sports artifacts and heart rate signals. Compared with traditional heart rate calibration methods based on statistical filtering or a single deep learning model, the present application uses the topological structure information of the knowledge graph to better capture the deep semantic correlation between sports states and heart rate errors, thereby achieving higher heart rate monitoring accuracy in various sports scenarios.

Description

Technical Field

[0001] This invention relates to the field of wearable device technology, specifically to a headphone heart rate error calibration method and related device based on a motion knowledge graph. Background Technology

[0002] Mainstream headphone heart rate monitoring technology is primarily based on photoplethysmography (PPG), which involves illuminating the skin with light of a specific wavelength, using a photoelectric sensor to receive the reflected light signal, and then extracting the PPG signal reflecting the heartbeat to calculate the heart rate. However, when users use headphones for heart rate monitoring during exercise, the acquired PPG signal often contains numerous motion artifacts. These artifacts mainly originate from factors such as skin vibration, muscle contraction, changes in blood flow velocity, and the relative displacement between the optical sensor and the skin during exercise. The introduction of motion artifacts distorts the waveform of the PPG signal, leading to significant errors in heart rate calculations. Experimental data shows that during running, the error rate of traditional headphone heart rate monitoring devices can be as high as 20% to 30%, and in complex exercise scenarios such as high-intensity interval training, the error rate may even exceed 40%. This seriously affects the user's accurate assessment of their health status and exercise effectiveness, reducing the product's practical value.

[0003] To address the problem of motion artifact interference, various solutions have been proposed in the industry. Traditional methods mainly include adaptive filtering, independent component analysis (ICA), and noise removal based on acceleration signals. Adaptive filtering methods estimate and remove motion artifacts by constructing a reference noise channel, but their filtering parameters need to be preset, making it difficult to adapt to scenarios with dynamically changing motion states. ICA methods assume that the heart rate signal and motion artifacts are statistically independent, but this assumption often fails to hold under motion conditions because the heart rate signal itself experiences frequency drift with changes in motion intensity. Noise removal methods based on acceleration signals use motion signals acquired by accelerometers as a reference and eliminate artifacts generated by synchronous motion through mathematical operations. However, the relationship between acceleration signals and motion artifacts in photoplethysmography (PPG) waves is not a simple linear one, limiting the removal effectiveness.

[0004] Therefore, the technical problem proposed by this invention is: how to construct a headphone heart rate error calibration method that can adapt to various sports states, has strong generalization ability, and supports real-time processing. Summary of the Invention

[0005] This disclosure proposes a headphone heart rate error calibration method and related device based on motion knowledge graph, aiming to overcome at least one of the defects in the prior art.

[0006] To achieve the above objectives, the technical solution disclosed in this invention is as follows: According to one aspect of this disclosure, a headphone heart rate error calibration method based on a motion knowledge graph is provided, comprising the following steps: The system acquires multimodal physiological data of users wearing headphones during exercise. The multimodal physiological data includes photoplethysmography (PPG) signals collected by an optical heart rate sensor, acceleration data collected by a triaxial accelerometer, and blood oxygen saturation data collected by a blood oxygen sensor. The acceleration data is subjected to time-domain feature extraction and frequency-domain feature extraction. Based on the extracted features, a pre-trained motion state classification model is used to identify the user's current motion state type and motion intensity level. Retrieve a knowledge subgraph from a pre-constructed motion knowledge graph that matches the current motion state type and motion intensity level. The motion knowledge graph includes motion state nodes, physiological parameter nodes, error feature nodes, and edge relationships connecting the nodes. The relationship weights of the edge relationships are calculated based on the sample frequency and statistical characteristics in the historical motion physiological dataset. Based on the associations in the retrieved knowledge subgraph, the embedding vectors of artifact-type nodes are fused with the time-frequency features of the currently acquired photoplethysmography pulse wave signal through a graph attention mechanism. The fused features are then input into a preset error prediction neural network model to calculate the heart rate error prediction value. The original heart rate data is calibrated based on the predicted heart rate error value. A Kalman filter is used to perform time-series filtering on the initial calibrated heart rate value, and a second correction is performed based on the blood oxygen saturation data to output the calibrated heart rate data.

[0007] Furthermore, the acquisition of multimodal physiological data specifically includes: The photoplethysmography (PPG) signal, acceleration data, and blood oxygen saturation data are synchronously acquired at a preset sampling frequency. When the user is detected to be stationary, the sampling frequency is adjusted to a first preset frequency. When the user is detected to be in motion, the sampling frequency is adjusted to a second preset frequency that is higher than the first preset frequency.

[0008] Furthermore, the method for constructing the motion knowledge graph includes: Collect historical users' exercise physiological datasets, which include the original waveforms of photoplethysmography (PPG) signals under different exercise types and intensities, the acceleration data, the blood oxygen saturation data, and the corresponding true heart rate values; Motion artifact features of photoplethysmography (PPG) signals are extracted from the exercise physiological dataset, and a mapping relationship between the motion state and the motion artifact features is established. Establish a set of node types, including exercise type nodes, exercise intensity nodes, artifact type nodes, heart rate zone nodes, and environmental parameter nodes; Based on the mapping relationship, a set of edges connecting the nodes is constructed. Each edge is assigned a relation weight calculated based on the sample frequency, and the relation weight is iteratively optimized using a backpropagation algorithm. A graph attention network is used to embed the initially constructed motion knowledge graph into a vector representation of each node in the motion knowledge graph.

[0009] Furthermore, the retrieval of knowledge subgraphs from the pre-constructed motion knowledge graph includes: Convert the currently identified motion state type and motion intensity level into a graph query vector; Based on the graph query vector, an approximate nearest neighbor search is performed in the motion knowledge graph to determine a preset number of knowledge nodes that are most similar to the current motion state. The subgraph is expanded based on the adjacency relationships of the retrieved knowledge nodes in the graph. The associated nodes directly connected to the knowledge nodes and their edge relationships are included in the knowledge subgraph. The number of nodes in the knowledge subgraph is adaptively adjusted according to the similarity threshold between the current motion state and the historical motion state.

[0010] Furthermore, the loss function of the error prediction neural network model includes: The mean squared error loss term is used to measure the Euclidean distance between the predicted heart rate error value and the actual heart rate error value. The graph structure loss term is used to maintain the distance relationship between the topological structure information in the motion knowledge graph and the embedding space. The diversity regularization term is used to encourage the model to learn different artifact feature representations under the same motion state.

[0011] Furthermore, the method for calculating the predicted heart rate error includes: The embedding vectors corresponding to artifact type nodes and heart rate interval nodes are extracted from the retrieved knowledge subgraph. The embedding vector of the artifact type node is fused with the time-frequency features of the currently acquired photoplethysmography pulse wave signal through an attention mechanism to obtain a fused feature vector. The fused feature vector is input into an error prediction neural network model using a long short-term memory network architecture to obtain a preliminary heart rate error value; The initial heart rate error value is corrected based on the current exercise intensity level to obtain the final heart rate error prediction value.

[0012] Furthermore, the calibration process for the raw heart rate data includes: The difference between the original heart rate data and the predicted heart rate error is calculated to obtain the preliminary calibrated heart rate value. The initial calibrated heart rate value is subjected to time-series filtering using a Kalman filter, wherein the state transition matrix of the Kalman filter is designed according to the physiological characteristics of heart rate. The filtered heart rate value is then corrected based on the blood oxygen saturation data to address the impact of blood oxygen changes in high-altitude environments or during strenuous exercise on heart rate monitoring.

[0013] Furthermore, the motion state classification model is structured as a lightweight convolutional neural network, comprising two one-dimensional convolutional layers and one fully connected layer.

[0014] Furthermore, the environmental parameter nodes include temperature nodes, humidity nodes, air pressure nodes, and altitude nodes.

[0015] According to another aspect of this disclosure, a headphone heart rate error calibration system based on motion knowledge graph is provided to implement the headphone heart rate error calibration method based on motion knowledge graph as described above, comprising: The multimodal data acquisition module is used to acquire multimodal physiological data of users wearing headphones during exercise. The motion state recognition module is connected to the multimodal data acquisition module and is used to extract time-domain features and frequency-domain features from the acceleration data. Based on the extracted features, a pre-trained motion state classification model is used to identify the user's current motion state type and motion intensity level. The knowledge graph retrieval module, connected to the motion state recognition module, is used to retrieve knowledge subgraphs from the pre-built motion knowledge graph that match the current motion state type and motion intensity level. The error prediction module, connected to the knowledge graph retrieval module, is used to fuse the embedding vector of the artifact type node with the time-frequency features of the currently acquired photoplethysmography pulse wave signal through the graph attention mechanism, and input the fused features into the preset error prediction neural network model to calculate the heart rate error prediction value. The heart rate calibration module, connected to the error prediction module and the multimodal data acquisition module, is used to calibrate the original heart rate data according to the heart rate error prediction value, perform time-series filtering using a Kalman filter and perform secondary correction based on blood oxygen saturation data, and output calibrated heart rate data. The knowledge graph construction module, connected to the knowledge graph retrieval module, is used to construct the sports knowledge graph based on the historical users' sports physiology dataset, and to use a graph attention network to embed the nodes in the sports knowledge graph.

[0016] According to another aspect of this disclosure, a computer-readable storage medium is provided that stores a computer program, which, when executed by a processor, implements the headphone heart rate error calibration method based on motion knowledge graph as described above.

[0017] According to another aspect of this disclosure, a headphone device is provided, including the headphone heart rate error calibration system based on motion knowledge graph as described above, and further comprising: The shell structure is provided with an ear hook for accommodating the ear during movement and an in-ear part for embedding into the ear canal; The main control processor, located inside the housing structure, is used to perform scheduling and data processing of various modules in the heart rate error calibration system; An optical heart rate sensor is located inside the ear canal to collect the photoplethysmography (PPG) signal of the wearer. A triaxial accelerometer is installed inside the housing structure to collect acceleration data of the user during movement. A blood oxygen sensor is located inside the ear canal to collect the wearer's blood oxygen saturation data. A Bluetooth communication module is located inside the housing structure and connected to the main control processor, used to wirelessly transmit calibrated heart rate data to an external terminal device; The power management module is connected to the main control processor and each sensor module, and is used to manage the power supply and low power consumption control of each module. The environmental parameter monitoring module is used to collect environmental temperature, environmental humidity and environmental air pressure values, and send the environmental parameters to the main control processor for the heart rate calibration module to perform environmental adaptability calibration.

[0018] The beneficial effects of this invention are: This invention constructs a motion knowledge graph to structurally represent the complex mapping relationship between different motion state types, motion intensity levels, and heart rate error characteristics, thereby achieving explicit modeling of the correlation between motion artifacts and heart rate signals.

[0019] Compared to traditional heart rate calibration methods based on statistical filtering or single deep learning models, this invention utilizes the topological structure information of knowledge graphs to better capture the deep semantic relationship between motion state and heart rate error, thereby achieving higher heart rate monitoring accuracy in various motion scenarios.

[0020] Meanwhile, the introduction of knowledge graphs gives the system good interpretability, allowing users and developers to clearly understand the root cause of heart rate errors under specific exercise conditions.

[0021] In addition, the knowledge graph supports incremental updates. When new exercise physiological data is collected, the knowledge graph can be easily expanded and optimized without retraining the entire model, effectively reducing the system's maintenance costs and update cycle.

[0022] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0023] Figure 1 This is a flowchart of a headphone heart rate error calibration method based on motion knowledge graph according to an embodiment of the present invention; Figure 2 This is a comparison diagram of photoplethysmography (PPG) signal processing and error in one embodiment of the present invention. Figure 3 This is a diagram illustrating the weight distribution effect of the attention mechanism in one embodiment of the present invention. Figure 4 This is a diagram illustrating the heart rate error prediction effect in one embodiment of the present invention; Figure 5 This is a visualization of the spatial distribution and similarity matrix of knowledge graph node embeddings in one embodiment of the present invention. Figure 6 This is a system performance diagram according to one embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0025] In embodiments of the present invention, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0026] The present invention provides the following preferred embodiments: In one embodiment, reference Figures 1 to 2 As shown, the headphone heart rate error calibration method based on motion knowledge graph of the present invention includes the following steps: Referring to step S101, multimodal physiological data of the user wearing headphones during exercise is acquired. The optical heart rate sensor on the headphone device emits green light and receives reflected light signals, extracting the photoplethysmography (PPG) signal. This optical heart rate sensor employs a dual-wavelength design; the ratio of the two wavelengths' signals effectively eliminates some motion artifact interference. The core idea of ​​the dual-wavelength design is to utilize the differences in penetration depth of different wavelengths of light in tissues and the differences in absorption characteristics of hemoglobin for different wavelengths of light to construct a reference channel to estimate the intensity of motion artifacts. When the user is in motion, skin vibration and muscle contraction cause dynamic changes in the contact state between the optical sensor and the skin, and these changes have varying degrees of impact on different wavelengths of light. By comparing the signal changes of the two wavelengths, the relative intensity of motion artifacts can be estimated, thus allowing for compensation in subsequent processing. A triaxial accelerometer collects the user's triaxial acceleration data during exercise for subsequent motion state analysis. The triaxial accelerometer can simultaneously capture the user's acceleration in three orthogonal directions. By analyzing the amplitude and direction changes of the acceleration signals, the user's current motion posture and intensity can be determined. The pulse oximeter sensor collects the user's blood oxygen saturation data, which is primarily used to address the impact of blood oxygen fluctuations on heart rate monitoring, especially in high-altitude environments or during strenuous exercise. Blood oxygen saturation reflects the proportion of oxyhemoglobin in the blood and is an important indicator for assessing a user's respiratory function and exercise intensity. In high-altitude environments, the oxygen content in the air is low, leading to a corresponding decrease in the user's blood oxygen saturation. This can affect the morphology of the photoplethysmography (PPG) signal and the accuracy of heart rate calculation.

[0027] Referring to step S102, during exercise, the sampling frequency is adaptively adjusted according to the exercise intensity. When the exercise intensity is high, the sampling frequency of each sensor will increase accordingly to ensure that sufficient physiological data can be captured for subsequent error analysis. Adaptive adjustment of the sampling frequency is one of the key factors in ensuring the accuracy of heart rate monitoring. When the user is at rest or in a low-intensity exercise state, the heart rate signal is relatively stable, and a lower sampling frequency is sufficient for analysis while also reducing system power consumption. When the user enters a high-intensity exercise state, the heart rate signal changes drastically, and motion artifact interference increases, requiring a higher sampling frequency to capture the detailed changes in the signal so as to distinguish the heart rate signal from motion artifacts in subsequent processing.

[0028] Furthermore, time-domain and frequency-domain features are extracted from the acceleration data. Based on the extracted features, a pre-trained motion state classification model is used to identify the user's current motion state type and motion intensity level. Specifically, the acceleration data is first preprocessed, including zero-mean normalization and low-pass filtering to remove high-frequency noise. Zero-mean normalization adjusts the mean of the acceleration signal to zero, eliminating the influence of the sensor's static bias and making signal analysis more accurate. Low-pass filtering removes high-frequency noise components from the acceleration signal, which mainly originate from the sensor's own electronic noise and interference from high-frequency vibrations in the environment. The preprocessed acceleration signal is smoother, facilitating subsequent feature extraction operations.

[0029] Furthermore, time-domain features are extracted from the preprocessed acceleration data, and statistical characteristics such as the mean, variance, number of peaks, and zero-crossing rate of the acceleration signal are calculated. The mean reflects the average level of the acceleration signal and is related to the user's motion intensity. The variance reflects the dispersion of the acceleration signal; high variance indicates vigorous motion, while low variance indicates gentle motion. The number of peaks reflects the number of peaks in the acceleration signal exceeding a preset threshold and is related to the user's stride frequency and motion rhythm. The zero-crossing rate reflects the frequency at which the acceleration signal crosses the zero line and is related to the type of motion; different types of motion have different zero-crossing rate characteristics. By comprehensively analyzing these time-domain features, the user's current motion state can be preliminarily determined.

[0030] Furthermore, frequency domain feature extraction is performed on the acceleration data to obtain the spectral distribution of the acceleration signal. The dominant frequency component and spectral energy distribution are extracted as frequency domain features. Frequency domain feature extraction employs a Fast Fourier Transform (FFT) algorithm to convert the time-domain signal into a frequency-domain signal, obtaining the energy distribution of the signal at different frequency components. The dominant frequency component is the frequency component with the highest energy in the acceleration signal, corresponding to the user's motion frequency, such as the stride frequency during running or the pedal frequency during cycling. The spectral energy distribution reflects the allocation of motion energy at different frequencies; the spectral energy of high-intensity motion is mainly distributed in the low-frequency region, while the spectral energy distribution of low-intensity motion is more uniform. The extracted time-domain and frequency-domain features are concatenated to form a comprehensive feature vector, which contains comprehensive information about the motion state.

[0031] Furthermore, this comprehensive feature vector is input into a pre-trained motion state classification model, which employs a lightweight convolutional neural network architecture comprising two one-dimensional convolutional layers and one fully connected layer. The lightweight architecture is designed to enable real-time inference on embedded devices while maintaining high classification accuracy. The one-dimensional convolutional layers effectively extract local features from the acceleration signal, and the stacking of two one-dimensional convolutional layers abstracts features layer by layer, progressively extracting motion semantic features from lower-level raw signal features to higher-level motion semantic features. The fully connected layer non-linearly combines the features extracted by the convolutional layers, outputting the probability distribution of various motion states. The motion state classification model outputs five basic types: running, walking, strength training, cycling, and sitting.

[0032] Furthermore, the exercise intensity level is determined by comparing the spectral energy of the acceleration signal with a preset intensity threshold. The intensity levels are divided into three categories: low, medium, and high. Low-intensity exercise corresponds to everyday walking and slow walking, where the acceleration signal energy is low and the frequency distribution is relatively uniform. Medium-intensity exercise corresponds to brisk walking and jogging, where the acceleration signal energy is moderate and the frequency distribution exhibits a certain peak. High-intensity exercise corresponds to sprinting and running, where the acceleration signal energy is high and the frequency distribution exhibits a distinct peak.

[0033] Referring to step S103, a knowledge subgraph matching the current motion state type and motion intensity level is retrieved from the pre-built motion knowledge graph. In this embodiment, the motion knowledge graph is pre-built during the system initialization phase, and its construction process includes the following sub-steps: First, a large dataset of historical users' exercise physiology was collected. This dataset includes raw waveform data of photoplethysmography (PPG) signals, acceleration data, blood oxygen saturation data, and true heart rate values ​​synchronously acquired via medical-grade heart rate monitors under different exercise types and intensities. The size and diversity of the dataset are crucial to ensuring the generalization ability of the knowledge graph. The dataset should cover users of different ages, genders, and physical conditions to ensure that the knowledge graph can adapt to various user groups. The dataset should include samples of various exercise types and intensities to ensure that the knowledge graph can handle various exercise scenarios. Simultaneously, the data collection process should cover different environmental conditions as much as possible, such as indoor and outdoor environments, different temperatures, and different altitudes, to enhance the environmental adaptability of the knowledge graph.

[0034] Motion artifact features of the photoplethysmography (PPG) signal were extracted from this dataset, including artifact amplitude, artifact frequency components, and the degree of frequency overlap between the artifact and the heart rate signal. Motion artifact feature extraction is a core step in constructing the knowledge graph. Artifact amplitude reflects the intensity of interference from motion artifacts on the PPG signal; a larger amplitude indicates more severe interference. Artifact frequency components reflect the frequency distribution characteristics of motion artifacts; artifacts generated by different motion types have different frequency characteristics. The degree of frequency overlap between the artifact and the heart rate signal reflects the degree of interference with heart rate calculation; interference is particularly severe when the artifact frequency is close to the heart rate frequency.

[0035] A set of node types is established, including exercise type nodes, exercise intensity nodes, artifact type nodes, heart rate zone nodes, and environmental parameter nodes. Exercise type nodes include subtypes such as running, walking, strength training, and cycling, with each subtype node encoding typical characteristics of its corresponding exercise type. Exercise intensity nodes include three levels: low intensity, moderate intensity, and high intensity, each corresponding to a different acceleration energy threshold range. Artifact type nodes are classified according to the frequency characteristics of artifacts, including low-frequency artifact nodes, medium-frequency artifact nodes, and high-frequency artifact nodes; different frequencies of artifacts require different elimination strategies. Heart rate zone nodes are divided according to the range of resting heart rate and exercise heart rate, including four levels: resting zone, low-intensity zone, moderate-intensity zone, and high-intensity zone. Environmental parameter nodes include environments such as normal temperature, low temperature, high temperature, low altitude, and high altitude; heart rate monitoring under different environmental conditions needs to consider the influence of environmental factors on physiological signals.

[0036] Furthermore, an edge set connecting the nodes is constructed based on the extracted mapping relationships. Each edge is assigned a corresponding relationship weight, which reflects the frequency and confidence of the mapping relationship in historical data. The construction of edge relationships is the core of the knowledge graph, encoding the semantic associations between different types of nodes. For example, there is a strong association between running and moderate intensity exercise because running typically corresponds to a moderate intensity exercise load. There is a strong association between cycling and muscle movement artifacts because the periodic contractions of leg muscles during cycling produce significant artifact interference.

[0037] Furthermore, a graph attention network is employed to perform embedding learning on the initially constructed knowledge graph, mapping each node to a low-dimensional vector space. This ensures that semantically similar nodes are closer together in the vector space, thus completing the construction of the knowledge graph. The core idea of ​​the graph attention network is to utilize the attention mechanism to learn the importance weights of different neighboring nodes for the current node, thereby generating more accurate node embedding vectors. Through embedding learning, semantically similar nodes are closer together in the vector space; for example, running nodes and walking nodes are closer together in the embedding space because they both involve periodic movements.

[0038] In the knowledge graph retrieval stage, the currently identified motion state type and motion intensity level are first converted into a graph query vector. This query vector is constructed by concatenating the corresponding motion type node embedding, motion intensity node embedding, and the time-frequency feature embedding of the current photoplethysmography (PPG) signal. Then, based on this query vector, an approximate nearest neighbor search is performed in the motion knowledge graph. A fast retrieval algorithm based on vector similarity is used to determine the top twenty knowledge nodes most similar to the current motion state. Next, a subgraph expansion is performed based on the adjacency relationships of the retrieved knowledge nodes in the graph, incorporating directly connected nodes and their edge relationships into the knowledge subgraph. The depth of the subgraph expansion is adaptively adjusted based on the similarity threshold between the current motion state and historical motion states. A shallower expansion depth is used when the similarity is high to improve retrieval speed, while a deeper expansion depth is used when the similarity is low to obtain richer contextual information. The final knowledge subgraph contains information such as motion features, artifact patterns, and historical calibration experience associated with the current motion state.

[0039] Referring to step S104, the predicted heart rate error is calculated based on the retrieved knowledge subgraph. First, embedding vectors corresponding to artifact type nodes and heart rate interval nodes are extracted from the knowledge subgraph. These embedding vectors encode the mapping relationship between artifact patterns and heart rate errors under historical motion states. Then, the embedding vectors of artifact type nodes are fused with the time-frequency features of the currently acquired photoplethysmography (PPG) signal. The fusion method employs an attention mechanism, learning the importance weights of different embedding dimensions and performing a weighted summation to obtain a fused feature vector. Next, the fused feature vector is input into a pre-defined error prediction neural network model. This model uses a long short-term memory (LSTM) network architecture, containing two LSM layers and a fully connected output layer, effectively capturing the time-series dependence of heart rate errors. The model outputs a preliminary heart rate error value, reflecting the degree of influence of motion artifacts in the current PPG signal on heart rate calculation. The preliminary heart rate error value is corrected based on the current exercise intensity level. The correction coefficient is determined based on the statistical information of the relationship between exercise intensity and error in the knowledge subgraph, ultimately yielding an accurate predicted heart rate error value.

[0040] Referring to step S105, the raw heart rate data is calibrated. First, the difference between the raw heart rate data and the predicted heart rate error is calculated to obtain a preliminary calibrated heart rate value. Then, a Kalman filter is used to perform time-series filtering on the preliminary calibrated heart rate value. The state transition matrix of the Kalman filter is designed according to the physiological characteristics of heart rate, which can preserve the dynamic change characteristics of the heart rate signal while ensuring the filtering effect. The filtered heart rate value is then corrected a second time based on blood oxygen saturation data. When blood oxygen saturation is lower than a preset threshold, it indicates that the user may be in a high-altitude environment or in a state of strenuous exercise. At this time, the heart rate value is further corrected according to the correlation between blood oxygen and heart rate to cope with the impact of blood oxygen changes on heart rate monitoring. Finally, the second-corrected heart rate value is output as the final calibrated heart rate data and transmitted to the user's mobile terminal application for display and recording via Bluetooth communication module.

[0041] In another embodiment, reference Figure 3 As shown, this invention provides a detailed explanation of the weight distribution of the attention mechanism. In the heart rate error prediction process, the multi-head attention mechanism plays a crucial role, adaptively learning the importance weights of different knowledge nodes for the current prediction task. The specific attention weight allocation strategy is as follows: For a given knowledge subgraph, attention scores are first calculated in different subspaces using three parallel attention heads. Each attention head independently learns a set of query, key, and value transformation matrices to capture semantic relationships at different levels. The query vector is obtained by linearly transforming the current input features, the key vector by linearly transforming the embedding vectors of knowledge nodes, and the value vector by linearly transforming the embedding vectors of knowledge nodes. The attention score is calculated by the dot product of the query vector and the key vector, reflecting the correlation between the current input and each knowledge node. The attention scores are normalized by applying the softmax function to obtain the final attention weight, which represents the contribution of each knowledge node to the current prediction task.

[0042] The first attention head focuses on capturing the correlation between motion type and artifact features. In this head, the attention scores between motion type nodes and artifact type nodes are generally high. In particular, the attention weight between running type and high-frequency artifacts reaches a high level, reflecting the significant impact of high-frequency artifacts generated by skin vibrations during running on heart rate calculation. This finding is consistent with theoretical expectations in exercise physiology; the impact vibrations generated by the foot striking the ground during running are transmitted to the head through body tissues, causing relative displacement between the optical sensor worn in the ear and the skin, thus introducing high-frequency motion artifacts.

[0043] The second attention head focuses on capturing the correlation between exercise intensity and heart rate zones. In this attention head, the attention weight between high-intensity exercise nodes and high-intensity heart rate zone nodes is relatively high, indicating that the combined effect of increased heart rate and motion artifacts on heart rate calculation is more significant during high-intensity exercise. During high-intensity exercise, not only does the amplitude of motion artifacts increase, but the user's heart rate itself is also at a higher level, and the frequency overlap between the heart rate signal and the artifact signal also increases accordingly. These factors combined make heart rate error prediction more complex. The second attention head also pays attention to environmental parameter nodes, and it maintains a high attention weight between environmental parameter nodes related to blood oxygen saturation, indicating that environmental factors have a significant impact on the accuracy of heart rate monitoring.

[0044] The third attention head focuses on capturing the cross-correlation between artifact types and heart rate zones. In this attention head, high-frequency artifacts have a greater impact on low heart rate zones than high heart rate zones, a finding with significant practical value. When users are at rest or in a low-intensity state, their heart rate frequency is lower, and the frequency difference with high-frequency artifacts is larger, theoretically making them easier to separate. However, actual test results show that high-frequency artifacts cause greater interference in low heart rate zones. This may be because the amplitude of the heart rate signal is smaller at low heart rates, making it relatively easier to be overwhelmed by artifact signals.

[0045] In the feature fusion stage, the node embedding vectors output by the three attention heads are weighted and concatenated to form the final fused feature representation. The fusion weights are dynamically adjusted based on the prediction performance of each attention head on the validation set, enabling the model to adaptively select the most relevant attention head output. Visual analysis reveals a significant positive correlation between the distribution of attention weights and the predefined edge relation weights in the motion knowledge graph, indicating that the attention mechanism designed in this invention can effectively reuse the structured knowledge encoded in the knowledge graph. Furthermore, through an end-to-end training process, the attention mechanism can also learn implicit association rules not explicitly encoded in the knowledge graph, further improving prediction performance.

[0046] In another embodiment, reference Figure 4 As shown, this invention provides a detailed explanation of the heart rate error prediction effect. In the time-series comparative analysis, the heart rate value predicted by the method of this invention is continuously compared with the true heart rate value collected by a medical-grade heart rate belt. Experimental results show that during continuous exercise monitoring, the heart rate curve predicted by the method of this invention exhibits a highly consistent trend with the true heart rate curve, and the deviation between the two remains within a small range.

[0047] Specifically, in the initial stage of exercise, the user's heart rate is rapidly increasing and fluctuates significantly. This invention's method, by incorporating historical calibration experience from a knowledge graph, can quickly track the trend of heart rate changes, keeping the error between the predicted and true values ​​within an acceptable range. The initial stage of exercise is one of the challenges in heart rate error prediction, as the user's physiological state is transitioning from rest to exercise, and the characteristics of heart rate signals and motion artifacts are changing rapidly. Traditional statistical filtering-based methods struggle to adapt to this rapid change, while this invention's method, leveraging historical calibration experience stored in a knowledge graph, can better address this challenge.

[0048] During sustained exercise, the user's heart rate enters a steady-state range, where the predictive advantage of the method in this invention becomes more pronounced, with significantly lower errors compared to traditional deep learning methods. While heart rate changes are relatively stable during sustained exercise, motion artifacts still exist. The method in this invention maintains stable predictive performance by continuously utilizing information from the knowledge graph. Although traditional deep learning methods can also track heart rate changes, their prediction errors are more than twice that of the method in this invention, especially when motion artifact interference is severe, leading to significant fluctuations in the predicted values.

[0049] Statistical analysis of the error distribution of a large number of test samples reveals that the error distribution of the method of this invention exhibits typical normal distribution characteristics, with a mean error close to zero and a small standard deviation, indicating that the prediction results have high accuracy and stability. The normal distribution characteristic suggests that the prediction error of the method of this invention is mainly caused by random factors rather than systematic bias. This means that averaging multiple measurements can further reduce errors and improve prediction accuracy.

[0050] Comparing the error distribution of traditional convolutional neural network methods reveals a significant right skewness, meaning that a certain proportion of samples exhibit significantly larger errors. These samples are primarily concentrated in scenarios involving high-intensity motion and transitions between motion states. This right skewness indicates that traditional methods are prone to generating large prediction errors in certain specific scenarios, which are precisely the problems that the method of this invention aims to address.

[0051] This invention effectively solves the problem of heart rate error prediction in challenging scenarios by introducing a motion knowledge graph, making the error distribution more concentrated near zero and significantly improving prediction performance. The introduction of the knowledge graph provides the model with rich prior knowledge, enabling it to make reasonable predictions based on historical experience of similar scenarios when facing unseen motion scenarios, thereby effectively improving the model's generalization ability and robustness.

[0052] In another embodiment, reference Figure 5As shown, this invention provides a detailed description of the spatial distribution and similarity matrix of knowledge graph node embeddings. In the embedding space visualization, different types of nodes exhibit obvious clustering characteristics, which reflect the successful modeling of the semantic relationship between motion state and heart rate error by the knowledge graph.

[0053] The movement type nodes are mainly distributed in the upper left region of the embedding space, with a moderate distance maintained between them. This distribution reflects the similarities and differences between different movement types. Running and walking nodes are relatively close in the embedding space because they are both periodic movements with similar motion characteristics. Strength training nodes are farther from running and walking nodes because strength training is a non-periodic movement, and its acceleration signal characteristics differ significantly from periodic movements. Cycling nodes are located between running and strength training nodes, reflecting that cycling combines both periodic and strength-based characteristics.

[0054] The motion intensity nodes are mainly distributed in the lower right region of the embedded space. The three nodes—low intensity, medium intensity, and high intensity—are arranged sequentially in a triangular spatial distribution, which conforms to the hierarchical relationship between motion intensity levels. The low intensity node is located at the edge of the lower right region, the medium intensity node is located in the middle, and the high intensity node is located in the upper left direction. This spatial arrangement makes the transition between adjacent intensity levels smoother.

[0055] Artifact type nodes are mainly distributed in the central region of the embedding space, close to motion type nodes and motion intensity nodes, reflecting the close relationship between artifact features and motion states. The distribution of artifact type nodes is also related to the frequency characteristics of the artifacts; low-frequency artifact nodes are close to low-intensity motion nodes, and high-frequency artifact nodes are close to high-intensity motion nodes. This distribution pattern verifies the rationality of the edge relationship between artifact type nodes and motion intensity nodes in the knowledge graph.

[0056] Heart rate zone nodes are mainly distributed in the upper right region of the embedding space, and their distribution position corresponds to that of the exercise intensity nodes. The resting zone node is located in the upper right position, with a lower and more stable resting heart rate, which is significantly different from low-intensity exercise. The low-intensity exercise zone node is located at the edge of the heart rate zone node cluster, while the moderate-intensity and high-intensity exercise zone nodes are arranged sequentially to the lower left, corresponding to the distribution trend of the exercise intensity nodes.

[0057] The node similarity matrix heatmap visually illustrates the semantic similarity between any two nodes. Analysis of the similarity matrix reveals that nodes of the same type generally exhibit high similarity; for example, nodes of different motion types show moderate similarity, reflecting their commonalities as motion states. The similarity between adjacent node types is slightly lower; for instance, the similarity between motion-type nodes and artifact-type nodes is positively correlated with the edge relationship weights defined in the knowledge graph. Environmental parameter nodes generally show low similarity to other node types, indicating that the influence mechanism of environmental factors on heart rate monitoring is relatively independent and requires separate processing. This visualization analysis of the embedding space allows for intuitive verification of whether the motion knowledge graph constructed in this invention successfully encodes the semantic association between motion states and heart rate errors.

[0058] In another embodiment, reference Figure 6 As shown, this invention provides a comprehensive evaluation of system performance. In performance tests under different signal-to-noise ratio conditions, the method of this invention demonstrates significant advantages in accuracy and robustness across various noise environments.

[0059] When the signal-to-noise ratio (SNR) is low, i.e., when motion artifact interference is severe, the error of the method in this invention increases slightly, but remains at a low level, significantly outperforming the traditional method in comparison. Low SNR environments are a challenging scenario in heart rate monitoring, where the intensity of motion artifacts can even exceed the heart rate signal itself, making it difficult for traditional methods to effectively distinguish them. The method in this invention, by incorporating historical calibration experience stored in a knowledge graph, can identify the type and intensity of current motion artifacts and adjust the prediction strategy accordingly, thereby maintaining a low error level.

[0060] As the signal-to-noise ratio (SNR) gradually improves, the error of the method in this invention continues to decrease, achieving prediction accuracy close to the theoretical limit under high SNR conditions. In high SNR environments, motion artifact interference is weak, and the heart rate signal quality is high, allowing both the method in this invention and traditional methods to achieve good prediction performance. However, even under these conditions, the error of the method in this invention is still slightly lower than that of the traditional method, thanks to the refined calibration capabilities provided by the knowledge graph.

[0061] Traditional methods exhibit significant performance degradation and a sharp increase in error under low signal-to-noise ratio conditions, indicating limited ability to model motion artifacts and difficulty in handling heart rate calibration requirements in complex motion scenarios. Traditional methods primarily rely on statistical assumptions and empirical parameters, leading to a significant performance decline when the motion scenario differs considerably from the distribution of training data. The method of this invention, however, leverages the generalization capabilities of knowledge graphs to mitigate this deficiency to some extent.

[0062] In long-term monitoring tests under continuous exercise scenarios, the method of this invention demonstrated excellent stability and adaptability. Throughout the monitoring process, the heart rate error remained at a low level, with no significant error accumulation or drift. Long-term monitoring is an important scenario for testing the practical value of heart rate monitoring devices. Users typically want to obtain accurate heart rate data continuously throughout an exercise session, without experiencing significant error drift.

[0063] In contrast, the error of traditional deep learning methods shows a fluctuating upward trend during monitoring, which may be due to the model gradually deviating from the optimal parameter space during long-term inference. Traditional deep learning models typically lack effective long-term stability guarantees and are susceptible to noise accumulation during long-term operation.

[0064] The stability of the method in this invention primarily benefits from the support of the knowledge graph. The structured knowledge embedded in the knowledge graph provides stable prior information for the model, enabling it to maintain consistent predictive behavior over long-term operation. Furthermore, the incremental update capability of the knowledge graph also ensures long-term stable operation. When model performance shows signs of degradation, the knowledge graph and prediction model can be updated through incremental learning without requiring complete retraining. This incremental update mechanism allows the method to adapt to changes in user habits and long-term evolution of environmental conditions.

[0065] In another embodiment, the headphone heart rate error calibration system and headphone device based on motion knowledge graph provided by the present invention will be described in detail. The system comprises six core modules, which are connected through data streams to form a complete processing link.

[0066] The multimodal data acquisition module is responsible for collecting raw physiological data from various sensors in the earphone, including photoplethysmography (PPG) signals, acceleration data, and blood oxygen saturation data. This module uses a streaming approach, acquiring sensor data in real time and caching it in a local data buffer. The buffer employs a circular queue structure to ensure data continuity and real-time performance. When the buffer is full, new data overwrites the oldest data, ensuring the system always processes the latest sensor data.

[0067] The motion state recognition module reads acceleration data from the data buffer, performs preprocessing and feature extraction, and then inputs it into the motion state classification model, outputting the current motion state type and motion intensity level. Preprocessing includes zero-mean normalization and low-pass filtering, while feature extraction includes both time-domain and frequency-domain feature extraction. The motion state classification model employs a lightweight convolutional neural network architecture, enabling real-time inference on embedded devices with inference latency controlled within milliseconds.

[0068] The knowledge graph retrieval module receives the motion state type and motion intensity level output by the motion state recognition module. After converting this information into a graph query vector, it performs a retrieval operation in a pre-constructed motion knowledge graph to obtain the knowledge subgraph that best matches the current motion state. The retrieval algorithm uses approximate nearest neighbor search, enabling large-scale knowledge graph retrieval operations to be completed within milliseconds. The retrieved knowledge subgraph contains node and edge information related to the current motion state, providing knowledge support for subsequent error prediction.

[0069] The error prediction module receives the knowledge subgraph and the time-frequency features of the current photoplethysmography (PPG) pulse wave signal, performs feature fusion and time-series modeling, and outputs a predicted heart rate error value. The built-in error prediction neural network model in this module has been trained offline before the system leaves the factory, and the model parameters are stored in a local model file. The model uses a long short-term memory (LSTM) network architecture, which can capture the time-series dependence of heart rate error and output stable error prediction values.

[0070] The heart rate calibration module calibrates the raw heart rate data based on the error prediction value, outputting the final calibrated heart rate data. The calibration process includes two steps: error compensation and filtering. Error compensation calculates the difference between the raw heart rate data and the error prediction value to eliminate systematic measurement bias. The filtering process uses a Kalman filter, which can eliminate noise while preserving the dynamic characteristics of the heart rate signal.

[0071] The knowledge graph construction module is used to build a motion knowledge graph during the system initialization phase. This module receives historical users' motion physiological datasets, extracts nodes, constructs edge relationships, and performs embedding learning to generate a complete motion knowledge graph structure, which is then stored in a local knowledge base. Knowledge graph construction is an offline process, typically completed before the product leaves the factory. Once completed, the knowledge graph can be continuously queried and used during device operation.

[0072] The headphones feature an ear-hook design and include a shell structure, main control processor, optical heart rate sensor, three-axis accelerometer, blood oxygen sensor, Bluetooth communication module, power management module, and environmental parameter monitoring module. The shell structure is ergonomically designed, featuring an ear hook for accommodating the ear during exercise and an in-ear section for fitting into the ear canal, ensuring a stable fit in various sports scenarios. The ear hook is made of flexible material to adapt to different ear shapes. The in-ear section has a tapered design for a close fit to the ear canal while ensuring comfortable wear.

[0073] The main control processor uses a low-power Bluetooth audio chip, integrating a processor core, digital signal processor, and Bluetooth baseband module, enabling efficient execution of heart rate error calibration algorithms. This chip employs advanced manufacturing processes, achieving low power consumption while maintaining computational performance. The digital signal processor is specifically designed to accelerate neural network inference operations, quickly completing motion state recognition and error prediction calculations.

[0074] The optical heart rate sensor is located inside the ear canal and uses dual-wavelength green and yellow-green light sources to acquire higher-quality photoplethysmography (PPG) signals during exercise. The dual-wavelength design effectively counteracts some motion artifacts, improving signal quality. The sensor also integrates low-noise analog front-end circuitry, which amplifies weak physiological signals and filters out high-frequency noise.

[0075] The triaxial accelerometer is housed within the casing and has a measurement range of ±16g, covering acceleration levels for various activities from walking to high-intensity running. The sensor employs a low-power design, consuming only microamps of current in normal motion monitoring mode. It also supports multiple sampling frequency configurations, dynamically adjusting according to the motion state to reduce power consumption while maintaining signal quality.

[0076] The pulse oximeter sensor is positioned inside the ear canal and uses dual-wavelength infrared and red light sources. It calculates blood oxygen saturation by measuring the absorption ratio of different wavelengths of light by oxyhemoglobin and deoxyhemoglobin. The ear canal is rich in blood vessels and has relatively thin tissue, making it an ideal location for measuring blood oxygen saturation. The pulse oximeter sensor achieves medical-grade accuracy, accurately reflecting the user's blood oxygen status.

[0077] The Bluetooth communication module supports Bluetooth Low Energy technology, enabling real-time transmission of calibrated heart rate data to mobile applications with latency controlled within milliseconds. Employing adaptive frequency hopping technology, the module maintains a stable connection even in congested wireless environments. Furthermore, it supports multiple broadcast and connection modes, facilitating data synchronization with various types of terminal devices.

[0078] The power management module is powered by a 120mAh lithium battery, supporting over 12 hours of continuous operation in normal motion monitoring mode. The module employs a multi-level power management strategy, dynamically adjusting the power supply status of each module under different operating conditions. In a stationary state, the system enters deep sleep mode, with most modules shut down, retaining only the timed wake-up function. In low-power motion monitoring mode, each sensor operates at the lowest sampling frequency, and the main control processor runs at a low clock frequency. In normal motion monitoring mode, all modules operate at full speed, providing complete monitoring functionality.

[0079] The environmental parameter monitoring module includes a temperature acquisition unit, a humidity acquisition unit, and an air pressure acquisition unit, which are used to collect ambient temperature, ambient humidity, and ambient air pressure values, respectively. The air pressure acquisition unit can obtain the current ambient air pressure value and convert it into the corresponding altitude, providing a basis for environmental adaptation adjustments for heart rate calibration. Environmental parameters are particularly important for heart rate monitoring in high-altitude sports scenarios, as low-oxygen environments can affect the user's physiological state and heart rate signal characteristics.

[0080] In a typical use case, when a user begins exercise, the main control processor activates the multimodal data acquisition module to collect photoplethysmography (PPG) signals, acceleration data, and blood oxygen saturation data at appropriate sampling frequencies. Acceleration data is input in real-time to the motion state recognition module to identify the user's current exercise state and intensity level. The recognition result is sent to the knowledge graph retrieval module, which retrieves a matching knowledge subgraph from a pre-built motion knowledge graph. The knowledge subgraph, along with the time-frequency characteristics of the PPG signal, is input to the error prediction module to calculate the current heart rate error prediction value. The heart rate calibration module calibrates the raw heart rate data based on the error prediction value and sends the calibrated heart rate data to a mobile application for display via Bluetooth communication. The latency of the entire processing flow is controlled within milliseconds, meeting the user's real-time heart rate monitoring needs during exercise.

[0081] Although the present invention has been specifically described above with reference to preferred embodiments, it should be understood that the present invention is not limited to the embodiments described above. Various modifications and variations can be made by those skilled in the art without departing from the spirit of the present invention, and such modifications and variations should fall within the scope defined by the appended claims and their equivalents.

Claims

1. A method for calibrating headphone heart rate error based on motion knowledge graph, characterized in that, Includes the following steps: The system acquires multimodal physiological data of users wearing headphones during exercise. The multimodal physiological data includes photoplethysmography (PPG) signals collected by an optical heart rate sensor, acceleration data collected by a triaxial accelerometer, and blood oxygen saturation data collected by a blood oxygen sensor. The acceleration data is subjected to time-domain feature extraction and frequency-domain feature extraction. Based on the extracted features, a pre-trained motion state classification model is used to identify the user's current motion state type and motion intensity level. Retrieve a knowledge subgraph from a pre-constructed motion knowledge graph that matches the current motion state type and motion intensity level. The motion knowledge graph includes motion state nodes, physiological parameter nodes, error feature nodes, and edge relationships connecting the nodes. The relationship weights of the edge relationships are calculated based on the sample frequency and statistical characteristics in the historical motion physiological dataset. Based on the associations in the retrieved knowledge subgraph, the embedding vectors of artifact-type nodes are fused with the time-frequency features of the currently acquired photoplethysmography pulse wave signal through a graph attention mechanism. The fused features are then input into a preset error prediction neural network model to calculate the heart rate error prediction value. The original heart rate data is calibrated based on the predicted heart rate error value. A Kalman filter is used to perform time-series filtering on the initial calibrated heart rate value, and a second correction is performed based on the blood oxygen saturation data to output the calibrated heart rate data.

2. The headphone heart rate error calibration method based on motion knowledge graph according to claim 1, characterized in that, The acquisition of multimodal physiological data specifically includes: The photoplethysmography (PPG) signal, acceleration data, and blood oxygen saturation data are synchronously acquired at a preset sampling frequency. When the user is detected to be stationary, the sampling frequency is adjusted to a first preset frequency. When the user is detected to be in motion, the sampling frequency is adjusted to a second preset frequency that is higher than the first preset frequency.

3. The headphone heart rate error calibration method based on motion knowledge graph according to claim 1, characterized in that, The method for constructing the motion knowledge graph includes: Collect historical users' exercise physiological datasets, which include the original waveforms of photoplethysmography (PPG) signals under different exercise types and intensities, the acceleration data, the blood oxygen saturation data, and the corresponding true heart rate values; Motion artifact features of photoplethysmography (PPG) signals are extracted from the exercise physiological dataset, and a mapping relationship between the motion state and the motion artifact features is established. Establish a set of node types, including exercise type nodes, exercise intensity nodes, artifact type nodes, heart rate zone nodes, and environmental parameter nodes; Based on the mapping relationship, a set of edges connecting the nodes is constructed. Each edge is assigned a relation weight calculated based on the sample frequency, and the relation weight is iteratively optimized using a backpropagation algorithm. A graph attention network is used to embed the initially constructed motion knowledge graph into a vector representation of each node in the motion knowledge graph.

4. The headphone heart rate error calibration method based on motion knowledge graph according to claim 3, characterized in that, The process of retrieving knowledge subgraphs from a pre-constructed motion knowledge graph includes: Convert the currently identified motion state type and motion intensity level into a graph query vector; Based on the graph query vector, an approximate nearest neighbor search is performed in the motion knowledge graph to determine a preset number of knowledge nodes that are most similar to the current motion state. The subgraph is expanded based on the adjacency relationships of the retrieved knowledge nodes in the graph. The associated nodes directly connected to the knowledge nodes and their edge relationships are included in the knowledge subgraph. The number of nodes in the knowledge subgraph is adaptively adjusted according to the similarity threshold between the current motion state and the historical motion state.

5. The headphone heart rate error calibration method based on motion knowledge graph according to claim 1, characterized in that, The loss function of the error prediction neural network model includes: The mean squared error loss term is used to measure the Euclidean distance between the predicted heart rate error value and the actual heart rate error value. The graph structure loss term is used to maintain the distance relationship between the topological structure information in the motion knowledge graph and the embedding space. The diversity regularization term is used to encourage the model to learn different artifact feature representations under the same motion state.

6. The headphone heart rate error calibration method based on motion knowledge graph according to claim 1, characterized in that, The method for calculating the heart rate error prediction value includes: The embedding vectors corresponding to artifact type nodes and heart rate interval nodes are extracted from the retrieved knowledge subgraph. The embedding vector of the artifact type node is fused with the time-frequency features of the currently acquired photoplethysmography pulse wave signal through an attention mechanism to obtain a fused feature vector. The fused feature vector is input into an error prediction neural network model using a long short-term memory network architecture to obtain a preliminary heart rate error value; The initial heart rate error value is corrected based on the current exercise intensity level to obtain the final heart rate error prediction value.

7. The headphone heart rate error calibration method based on motion knowledge graph according to claim 1, characterized in that, The calibration process for the raw heart rate data includes: The difference between the original heart rate data and the predicted heart rate error is calculated to obtain the preliminary calibrated heart rate value. The initial calibrated heart rate value is subjected to time-series filtering using a Kalman filter, wherein the state transition matrix of the Kalman filter is designed according to the physiological characteristics of heart rate. The filtered heart rate value is then corrected based on the blood oxygen saturation data to address the impact of blood oxygen changes in high-altitude environments or during strenuous exercise on heart rate monitoring.

8. A headphone heart rate error calibration system based on motion knowledge graph, used to implement the headphone heart rate error calibration method based on motion knowledge graph as described in any one of claims 1 to 7, characterized in that, include: The multimodal data acquisition module is used to acquire multimodal physiological data of users wearing headphones during exercise. The motion state recognition module is connected to the multimodal data acquisition module and is used to extract time-domain features and frequency-domain features from the acceleration data. Based on the extracted features, a pre-trained motion state classification model is used to identify the user's current motion state type and motion intensity level. The knowledge graph retrieval module, connected to the motion state recognition module, is used to retrieve knowledge subgraphs from the pre-built motion knowledge graph that match the current motion state type and motion intensity level. The error prediction module, connected to the knowledge graph retrieval module, is used to fuse the embedding vector of the artifact type node with the time-frequency features of the currently acquired photoplethysmography pulse wave signal through the graph attention mechanism, and input the fused features into the preset error prediction neural network model to calculate the heart rate error prediction value. The heart rate calibration module, connected to the error prediction module and the multimodal data acquisition module, is used to calibrate the original heart rate data according to the heart rate error prediction value, perform time-series filtering using a Kalman filter and perform secondary correction based on blood oxygen saturation data, and output calibrated heart rate data. The knowledge graph construction module, connected to the knowledge graph retrieval module, is used to construct the sports knowledge graph based on the historical users' sports physiology dataset, and to use a graph attention network to embed the nodes in the sports knowledge graph.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the headphone heart rate error calibration method based on motion knowledge graph as described in any one of claims 1 to 7.

10. A headphone device, characterized in that, The headphone heart rate error calibration system based on motion knowledge graph as described in claim 8 further includes: The shell structure is provided with an ear hook for accommodating the ear during movement and an in-ear part for embedding into the ear canal; The main control processor, located inside the housing structure, is used to perform scheduling and data processing of various modules in the heart rate error calibration system; An optical heart rate sensor is located inside the ear canal to collect the photoplethysmography (PPG) signal of the wearer. A triaxial accelerometer is installed inside the housing structure to collect acceleration data of the user during movement. A blood oxygen sensor is located inside the ear canal to collect the wearer's blood oxygen saturation data. A Bluetooth communication module is located inside the housing structure and connected to the main control processor, used to wirelessly transmit calibrated heart rate data to an external terminal device; The power management module is connected to the main control processor and each sensor module, and is used to manage the power supply and low power consumption control of each module. The environmental parameter monitoring module is used to collect environmental temperature, environmental humidity and environmental air pressure values, and send the environmental parameters to the main control processor for the heart rate calibration module to perform environmental adaptability calibration.

Images