Multi-sensor fusion method and system for intelligent analysis of spine gait characteristics

Abstract

The invention relates to the technical field of multi-sensor fusion, in particular to a multi-sensor fusion method and system for spine gait feature intelligent analysis, and the method comprises the steps: obtaining human body motion information through a plurality of inertial sensors distributed at different parts of a human body; extracting multi-scale gait features based on the human body motion information; according to the multi-scale gait features, analyzing gait cycle, symmetry and stability parameters; based on the human body movement information, establishing a relation model between spine movement and lower limb movement; learning a normal gait mode according to the relation model; a spine gait feature analysis result is generated, a comprehensive analysis framework is constructed, multi-source data can be processed at the same time, multi-scale features can be captured, and multi-task collaborative optimization is achieved through an advanced machine learning algorithm. Due to the introduction of a multi-task learning framework, the system can make full use of internal association among different gait parameters, and the accuracy and robustness of overall analysis are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of multi - sensor fusion, and more specifically, to a multi - sensor fusion method and system for intelligent analysis of spinal gait characteristics. Background Art

[0002] With the increasing prominence of the problems of population aging and sub - health, spinal health has become the focus of global attention. Spinal diseases not only affect the quality of life but may also lead to serious complications. Therefore, accurately and timely evaluating the spinal health status, especially early abnormal detection through gait analysis, has important clinical significance and social value.

[0003] In recent years, with the rapid development of sensor technology and artificial intelligence, gait analysis methods based on wearable devices have made remarkable progress. In the prior art, common methods include using a single inertial sensor to collect data and performing gait feature extraction and classification through simple signal processing and machine learning algorithms. These methods have achieved certain results in some specific scenarios, but there are still many limitations when facing complex spinal gait analysis tasks.

[0004] Firstly, the data acquisition method of a single sensor is difficult to comprehensively capture the complexity of human movement. Spinal gait involves the coordinated movement of multiple joints throughout the body, and relying solely on single - point data is difficult to accurately reflect the overall movement state. Secondly, traditional feature extraction methods are often limited to single - scale analysis in the time domain or frequency domain and cannot effectively capture multi - scale gait features. This leads to difficulties in ensuring the accuracy and stability of the analysis results when facing large individual differences or complex gait patterns.

[0005] In addition, most existing gait analysis systems use independent models to process different gait parameters, such as period, symmetry, and stability. This method ignores the internal relationship between parameters and is difficult to make full use of the advantages of multi - task learning. At the same time, for the relationship between spinal and lower - limb movements, the prior art often uses a simplified linear model and cannot accurately describe its complex non - linear dynamic characteristics.

[0006] Finally, in terms of the prediction of abnormal gait, the prior art is mostly limited to the identification of existing abnormalities and lacks the ability to prospectively predict potential risks. This greatly limits the implementation of preventive intervention measures and reduces the effectiveness of spinal health management. Summary of the Invention

[0007] In view of the above problems, the present invention proposes an innovative multi - sensor fusion method and system for intelligent analysis of spinal gait characteristics. The method aims to comprehensively improve the accuracy, stability, and prediction ability of spinal gait analysis through multi - sensor data fusion, multi - scale feature extraction, multi - task learning, and an innovative prediction model.

[0008] The present invention provides a multi-sensor fusion method for intelligent analysis of spinal gait characteristics, including:

[0009] An acquisition step, including:

[0010] Obtaining human motion information through multiple inertial sensors distributed at different parts of the human body;

[0011] A processing step, including:

[0012] Extracting multi-scale gait characteristics based on the human motion information;

[0013] Analyzing gait cycle, symmetry, and stability parameters according to the multi-scale gait characteristics;

[0014] Establishing a relationship model between spinal motion and lower limb motion based on the human motion information;

[0015] Learning normal gait patterns according to the relationship model;

[0016] An output step, including:

[0017] Generating spinal gait characteristic analysis results.

[0018] Preferably, the acquisition step specifically includes:

[0019] Installing IMU sensors on the soles of the feet, calves, thighs, waists, and chests respectively;

[0020] Collecting acceleration, angular velocity, and magnetic field data through the IMU sensors.

[0021] Preferably, the extraction of multi-scale gait characteristics specifically includes:

[0022] Fusing acceleration and angular velocity to obtain an angular acceleration signal;

[0023] Preprocessing the original data, including removing linear acceleration and gravity;

[0024] Extracting time-domain characteristics and frequency-domain characteristics.

[0025] Preferably, the analysis of gait cycle, symmetry, and stability parameters specifically includes:

[0026] Inputting three-dimensional angular acceleration into a neural network to output the gait cycle;

[0027] Inputting the acceleration signal into a neural network to output gait symmetry;

[0028] Inputting three-dimensional angular acceleration into a neural network to output gait stability.

[0029] Preferably, the establishment of the relationship model between spinal movement and lower limb movement specifically includes: preprocessing the original data, including removing linear acceleration and gravity;

[0030] Extracting lower limb movement features and establishing a collaborative relationship model between lower limb and spinal movement.

[0031] Preferably, the learning of the normal gait pattern specifically includes:

[0032] Preprocessing the historical data;

[0033] Obtaining gait features and performing multi-task learning;

[0034] Introducing a multi-scale learning framework to learn the normal gait pattern.

[0035] Preferably, it further includes an abnormal prediction step:

[0036] Constructing a collaborative gait prediction model;

[0037] Using the collaborative gait prediction model to perform early prediction on spinal abnormal gait.

[0038] Preferably, the construction of the collaborative gait prediction model specifically includes:

[0039] Using a spatio-temporal graph convolutional network to analyze the motion data and learn the normal activity pattern;

[0040] Introducing a self-supervised learning method to further learn and optimize the extracted features;

[0041] Constructing a motion chain multi-task regression neural network to comprehensively model different kinematic parameters. Preferably, the motion chain multi-task regression neural network includes:

[0042] A multi-class neural network for judging the joint activity state;

[0043] An abnormality degree neural network for calculating the gait abnormality degree;

[0044] A classification confidence neural network for evaluating the reliability of the classification result.

[0045] A multi-sensor fusion system for realizing the intelligent analysis of spinal gait features of the method, including:

[0046] A data acquisition module for obtaining human motion information through multiple inertial sensors;

[0047] A feature extraction module for extracting multi-scale gait features;

[0048] A parameter analysis module for analyzing gait cycle, symmetry, and stability parameters;

[0049] A relationship modeling module for establishing a relationship model between spinal movement and lower limb movement;

[0050] A pattern learning module for learning normal gait patterns;

[0051] An abnormal prediction module for predicting spinal abnormal gaits;

[0052] An output module for generating spinal gait feature analysis results.

[0053] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0054] The core of the present invention lies in constructing a comprehensive analysis framework that can simultaneously process multi-source data, capture multi-scale features, and achieve multi-task collaborative optimization through advanced machine learning algorithms. This method not only overcomes the limitations of single-sensor data insufficiency but also improves the richness of feature expression through multi-scale analysis. The introduction of the multi-task learning framework enables the system to fully utilize the internal correlations between different gait parameters, improving the overall analysis accuracy and robustness.

[0055] Especially worth mentioning is that the present invention innovatively establishes a collaborative relationship model between the spine and lower limb movements, which provides a new perspective for comprehensively understanding the human movement mechanism. By introducing spatio-temporal graph convolutional networks and self-supervised learning strategies, this method can better capture complex non-linear dynamic characteristics and improve the generalization ability of the model.

[0056] In terms of abnormal prediction, the method of the present invention realizes prospective prediction of potential abnormalities by constructing a coordinated gait prediction model. This innovation greatly enhances the preventive ability of the system, provides the possibility for early intervention, and is expected to significantly improve the effect of spinal health management.

[0057] Generally speaking, the method of the present invention has achieved innovative breakthroughs in multiple aspects such as data acquisition, feature extraction, model construction, and prediction analysis. These innovation points complement each other and work synergistically to form a comprehensive and efficient spinal gait analysis solution. Compared with the prior art, the present invention not only improves the accuracy and reliability of the analysis but also expands the application scope, especially showing great potential in early abnormal detection and preventive health management.

[0058] In addition, the method of the present invention has good scalability and adaptability. By adjusting the sensor configuration and model parameters, this method can easily adapt to different application scenarios, such as clinical diagnosis, rehabilitation training, daily health monitoring, etc. This flexibility gives the method of the present invention broad application prospects and is expected to play an important role in improving the public health level and reducing medical costs.

[0059] In summary, the multi-sensor fusion method and system for intelligent analysis of spinal gait characteristics provided by the present invention not only achieve multiple innovations technically, but also demonstrate significant advantages in practical applications. It provides a new, more comprehensive and effective technical means for spinal health management, and is expected to bring a revolutionary change to the prevention and treatment of spinal diseases, thereby greatly improving people's quality of life and health level. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a flowchart of the method of the present invention.

[0061] Figure 2 is a logical block diagram of the feature extraction module of the present invention.

[0062] Figure 3 is a logical block diagram of the parameter analysis module of the present invention.

[0063] Figure 4 is a logical block diagram of the relationship modeling module of the present invention.

[0064] Figure 5 is a logical block diagram of the pattern learning module of the present invention.

[0065] Figure 6 is a logical block diagram of the anomaly prediction module of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0066] Please refer to Figure 1-6 , the present invention provides a multi-sensor fusion method and system for intelligent analysis of spinal gait characteristics. The method collects human motion information through multiple distributed inertial sensors, and combines multi-scale feature extraction and multi-task learning technologies to achieve comprehensive analysis and anomaly prediction of spinal gait. The technical solutions of the present invention will be described in detail below.

[0067] First, the multi-sensor fusion method for intelligent analysis of spinal gait characteristics of the present invention includes an acquisition step, a processing step, and an output step. In the acquisition step, human motion information is acquired through multiple inertial sensors distributed at different parts of the human body. Preferably, these inertial sensors may be inertial measurement units (IMUs), including accelerometers, gyroscopes, and magnetometers. In one embodiment of the present invention, these sensors are respectively installed on the soles of the feet, calves, thighs, waist, and chest to comprehensively capture the human motion state.

[0068] In the processing step, first, multi-scale gait features are extracted based on the acquired human motion information. This step is one of the cores of the present invention. Through multi-scale feature extraction, gait information within different time scales and frequency ranges can be captured, thereby improving the comprehensiveness and robustness of feature expression. Specifically, this method first fuses the acceleration and angular velocity signals to obtain the angular acceleration signal. This fusion process can be achieved through a Kalman filter, where the estimation of the angular acceleration α can be expressed as:

[0069]

[0070] where ω is the angular velocity, is the linear acceleration, and t is the time.

[0071] Next, the raw data is preprocessed, including linear acceleration removal and gravity removal. Linear acceleration removal can be achieved through a high-pass filter, while gravity removal can use a low-pass filter. In a preferred embodiment of the present invention, a high-pass filter with a cut-off frequency of 0.3 Hz is used to remove the linear acceleration, and a low-pass filter with a cut-off frequency of 0.8 Hz is used to remove the gravitational acceleration. These parameters are empirically selected based on the normal walking frequency range of the human body (about 0.6 - 2 Hz) and can effectively separate the required motion information.

[0072] In the feature extraction stage, this method simultaneously extracts time-domain features and frequency-domain features. Time-domain features include statistics such as mean, standard deviation, peak value, and valley value, while frequency-domain features are obtained through the fast Fourier transform (FFT) and include main frequency, frequency energy distribution, etc. This multi-dimensional feature extraction method can comprehensively characterize gait features and lay a foundation for subsequent analysis.

[0073] Based on the extracted multi-scale gait features, this method further analyzes gait cycle, symmetry, and stability parameters. This step adopts an innovative multi-task learning framework to simultaneously predict multiple related parameters and fully utilizes the correlation information between different tasks. Specifically, this method uses three parallel neural networks to process these three tasks respectively:

[0074] 1. Gait cycle prediction: The three-dimensional angular acceleration signal is input into a long short-term memory (LSTM) network, and the gait cycle is output. The selection of the LSTM network is based on its ability to effectively capture temporal dependence relationships.

[0075] 2. Gait symmetry analysis: The acceleration signal is input into a convolutional neural network (CNN), and the gait symmetry index is output. The multi-layer convolutional structure of the CNN can automatically learn hierarchical feature representations and is very suitable for processing this complex pattern recognition task.

[0076] 3. Gait stability assessment: The three-dimensional angular acceleration signal is input into another LSTM network, and the gait stability index is output.

[0077] These three networks share some underlying features and are optimized simultaneously through backpropagation, thus achieving the goal of multi-task learning. Preferably, the loss function adopted in the present invention is:

[0078] L = αL cycle + βL symmetry + γL stability + λR

[0079] where L cycle , L symmetry and L stability are the task-specific loss functions for period prediction, symmetry analysis, and stability assessment respectively. R is the regularization term used to prevent overfitting. α, β, γ, and λ are weight coefficients used to balance the importance of different tasks. In one embodiment, the initial values of these weight coefficients can be set as α = 0.4, β = 0.3, γ = 0.3, λ = 0.01, and are dynamically adjusted through the gradient descent method during the training process to adapt to the characteristics of different datasets.

[0080] Next, based on the human motion information, this method establishes a relationship model between spinal motion and lower limb motion. This innovative modeling method can better describe the coordination of the overall human motion and contribute to more accurate analysis and prediction of spinal gait characteristics. Specifically, this method first preprocesses the original data, including linear acceleration removal and gravity removal, then extracts the lower limb motion characteristics, and establishes a collaborative relationship model between the lower limb and spinal motion.

[0081] This collaborative relationship model can be constructed through the dynamic time warping (DTW) algorithm. The DTW algorithm can calculate the similarity between two time series and find the best alignment. Assuming the spinal motion sequence is S = (s 1 , s 2 ,..., s m ), and the lower limb motion sequence is L = (l 1 , l 2 ,..., l n ), then the DTW distance can be solved through the dynamic programming method:

[0082] DTW(i, j) = d(s i , l j ) + min{DTW(i - 1, j), DTW(i, j - 1), DTW(i - 1, j - 1)1)}

[0083] where d(s i , l j) is a distance metric between two data points. By analyzing the DTW distance matrix, the synergistic relationship between spinal movement and lower limb movement can be obtained.

[0084] On the basis of establishing the relationship model, this method further learns the normal gait pattern. This step includes preprocessing historical data, obtaining gait features and performing multi-task learning, and introducing a multi-scale learning framework. The introduction of the multi-scale learning framework is to capture gait patterns at different time scales and improve the generalization ability of the model. Specifically, this method adopts the multi-scale convolutional neural network (MS-CNN) structure, which contains multiple parallel convolutional layers, and the convolutional kernel sizes of each convolutional layer are different, capturing short-term, medium-term, and long-term gait patterns respectively.

[0085] Preferably, the output of MS-CNN can be expressed as:

[0086]

[0087] where Conv i represents the convolutional operation of the i-th scale, w i is the corresponding weight, and f is the activation function (such as ReLU). This multi-scale learning method can effectively extract gait features at different time scales and provide more comprehensive information for anomaly detection.

[0088] Through the above steps, the method of the present invention can comprehensively analyze spinal gait features and output the analysis results. These results can include quantitative indicators of gait cycle, symmetry, and stability, as well as information such as the degree of deviation from the normal gait pattern, providing important references for clinical diagnosis and rehabilitation evaluation.

[0089] The method of the present invention can not only accurately analyze normal gait but also has the ability to predict abnormal gait. This predictive ability is of great significance for early intervention and prevention of spinal diseases. In practical applications, this method can be integrated into wearable devices to achieve real-time monitoring and early warning, providing timely health advice for users.

[0090] In summary, the multi-sensor fusion method for intelligent analysis of spinal gait features provided by the present invention realizes a comprehensive and accurate analysis of spinal gait by innovatively combining technologies such as multi-sensor data acquisition, multi-scale feature extraction, multi-task learning, and anomaly prediction. This method is not only advanced in technology but also has broad prospects in practical applications, providing strong technical support for spinal health management. To further elaborate on the technical solution of the present invention, the technical features involved in claims 4-7 will be described in detail below.

[0091] In a preferred embodiment of the present invention, the analysis of gait cycle, symmetry, and stability parameters employs an innovative neural network architecture. Specifically, this method uses three parallel neural networks to separately handle the prediction tasks of these three key parameters.

[0092] First, for the prediction of the gait cycle, this method inputs the three-dimensional angular acceleration signal into a carefully designed long short-term memory (LSTM) network. The choice of the LSTM network is based on its excellent ability to process time-series data, which can effectively capture the long-term dependencies in the gait cycle. In the embodiment of the present invention, the structure of the LSTM network includes two layers of bidirectional LSTM layers, each layer containing 128 hidden units, followed by a fully connected layer. This structural design can process time series in both forward and backward directions, thereby more comprehensively capturing the characteristics of the gait cycle.

[0093] For the analysis of gait symmetry, this method innovatively adopts a one-dimensional convolutional neural network (1D-CNN). The acceleration signal is input into this 1D-CNN, and the network can automatically learn to extract the key features of gait symmetry. In the preferred embodiment of the present invention, the structure of the 1D-CNN includes 3 convolutional layers, with the kernel sizes of each layer being 5, 3, and 3 respectively, and the number of kernels being 32, 64, and 64 respectively. Each convolutional layer is followed by a max-pooling layer with a pooling size of 2. This design of gradually increasing the number of convolutional kernels can gradually extract higher-level features, which is beneficial for capturing the subtle changes in gait symmetry.

[0094] In terms of gait stability assessment, this method also uses an LSTM network, but its structure is slightly different from the cycle prediction network. Specifically, this network includes one layer of bidirectional LSTM layer (128 hidden units) and two layers of unidirectional LSTM layers (64 hidden units each). This structural design aims to first capture the bidirectional time-series information and then further extract the features of gait stability in the unidirectional LSTM layers.

[0095] It is worth noting that these three parallel networks are not completely independent but share some underlying features. By designing a shared feature extraction layer, this method achieves the goal of multi-task learning and fully utilizes the correlation information between different tasks. Preferably, the shared layer can be a 1D-CNN containing two convolutional layers, and its output is respectively connected to the three task-specific networks.

[0096] During the training process, this method adopts a comprehensive loss function to simultaneously optimize these three tasks:

[0097] L total = αL cycle + βL symmetry + γL stability + λR

[0098] Among them, L cycle、 L symnetry and L stability are respectively the task-specific loss functions for periodic prediction, symmetry analysis, and stability evaluation. R is the regularization term used to prevent overfitting. α, β, γ, and λ are weight coefficients used to balance the importance of different tasks. In one embodiment of the present invention, the initial values of these weight coefficients can be set as α = 0.4, β = 0.3, γ = 0.3, λ = 0.01. During the training process, these weights can be dynamically adjusted by the gradient descent method to adapt to the characteristics of different datasets.

[0099] Another innovation point in the method of the present invention is to establish a relationship model between spinal motion and lower limb motion. The establishment of this model is crucial for comprehensively understanding human motion coordination. In specific implementation, first, the original data needs to be preprocessed, including linear acceleration removal and gravity removal. The removal of linear acceleration can be achieved by designing a high-pass filter with a cut-off frequency of 0.3 Hz, while the gravity removal can use a low-pass filter with a cut-off frequency of 0.8 Hz. The selection of these parameters is based on the frequency range of normal human walking (about 0.6 - 2 Hz).

[0100] After the preprocessing is completed, this method extracts lower limb motion features and establishes a collaborative relationship model between the lower limbs and spinal motion. The core of this collaborative relationship model is to capture the temporal dependence relationship between lower limb motion and spinal motion. For this purpose, the preferred embodiment of the present invention adopts the dynamic time warping (DTW) algorithm. The DTW algorithm can calculate the similarity between two time series and find the best alignment method, which is particularly suitable for processing signals such as gait that are periodic but may have time distortion.

[0101] Based on the established relationship model, this method further learns normal gait patterns. This step includes preprocessing historical data, obtaining gait features and performing multi-task learning, and introducing a multi-scale learning framework. The introduction of the multi-scale learning framework is another innovation point of the present invention, aiming to capture gait patterns at different time scales and improve the generalization ability of the model.

[0102] Specifically, this method adopts the multi-scale convolutional neural network (MS-CNN) structure. The MS-CNN contains multiple parallel convolutional layers, and the convolutional kernel sizes of each convolutional layer are different, respectively capturing short-term, medium-term, and long-term gait patterns. For example, three parallel convolutional layers can be set, and the convolutional kernel sizes are 3, 5, and 7 respectively, corresponding to capturing gait features of about 0.1 second, 0.2 second, and 0.3 second (assuming a sampling frequency of 50 Hz).

[0103] The method of the present invention also includes an innovative abnormal prediction step. This step first constructs a coordinated gait prediction model and then uses this model to predict spinal abnormal gait at an early stage. The core idea of the coordinated gait prediction model is to compare the normal gait pattern with the gait data collected in real time, so as to identify potential abnormalities.

[0104] In a specific implementation, the coordinated gait prediction model adopts a recurrent neural network based on the attention mechanism (Attention-based RNN). This structure can automatically learn the importance of different time steps and is particularly suitable for processing time series data such as gait with long-term dependencies. The inputs of the model include the multi-scale features extracted in the previous steps and the output of the spine-lower limb coordination relationship model.

[0105] Through this comprehensive method, the present invention can not only accurately analyze normal gait but also has the ability to predict abnormal gait. This predictive ability is of great significance for early intervention and prevention of spinal diseases. In practical applications, this method can be integrated into wearable devices to achieve real-time monitoring and early warning, providing timely health advice for users.

[0106] In summary, the multi-sensor fusion method for intelligent analysis of spinal gait features provided by the present invention realizes a comprehensive and accurate analysis of spinal gait by innovatively combining technologies such as multi-sensor data acquisition, multi-scale feature extraction, multi-task learning, and abnormal prediction. This method is not only advanced technically but also has broad prospects in practical applications, and can provide strong technical support for spinal health management. In a further implementation scheme of the present invention, the construction process of the coordinated gait prediction model involves multiple innovative technical means. First, this method uses a spatio-temporal graph convolutional network (ST-GCN) to analyze motion data and learn normal activity patterns. The advantage of ST-GCN is that it can capture information in both spatial and temporal dimensions simultaneously and is particularly suitable for processing data with complex spatio-temporal dependencies such as human motion.

[0107] Specifically, ST-GCN models the human body as a graph structure, where nodes represent body joints and edges represent the connections between joints. The features of each node include the three-dimensional coordinates and velocity information of the joint. In a preferred embodiment of the present invention, the structure of ST-GCN includes 9 ST-GCN layers, with every 3 layers forming a block, and residual connections are used between blocks to enhance feature propagation. The output of each ST-GCN layer can be expressed as:

[0108]

[0109] where H (l) is the input of the l-th layer, is the normalized adjacency matrix, is a learnable weight matrix, and σ is an activation function (such as ReLU). This structural design can effectively learn and represent the spatio-temporal patterns of normal gait.

[0110] To further improve the model's representation ability and generalization performance, this method introduces a self-supervised learning strategy. The core idea of self-supervised learning is to utilize the structural information of the data itself to construct learning tasks without the need for a large amount of labeled data. In the embodiments of the present invention, the Temporal Contrastive Learning (TCL) method is adopted. The goal of TCL is to learn an encoder such that the representations of different time segments from the same gait sequence are similar, while the representations of segments from different sequences are different.

[0111] Specifically, given a gait sequence x, two time segments x i and x j are randomly selected, and and are obtained through data augmentation. The loss function of TCL is defined as:

[0112]

[0113] where is the output of the encoder f, sin is the cosine similarity function, τ is the temperature parameter, and n is the batch size. By minimizing this loss function, the model can learn more discriminative and robust feature representations.

[0114] On the basis of constructing the synchronous gait prediction model, the present invention further introduces a motion chain multi-task regression neural network for comprehensively modeling different kinematic parameters. The design inspiration of this network comes from the concept of motion chain in human kinematics, aiming to capture the mutual influence and constraint relationships between the motions of different body parts.

[0115] The core structure of this network is a multi-branch Recurrent Neural Network (RNN). Each branch corresponds to a main body part (such as the spine, hip, knee, etc.), and information interaction is carried out between branches through lateral connections. At each time step, the network not only predicts the state of each part at the next moment but also predicts the relative motion relationship between them. The loss function of the network consists of multiple components:

[0116] L total = λ 1 L state + λ 2 L relation + λ 3 L consistency + λ 4 R,

[0117] where, Lstate is the state prediction loss, \(L\) relation is the relationship prediction loss, \(L\) consistency is the consistency constraint loss (ensuring that the predicted states and relationships are mutually consistent), \(R\) is the regularization term. \(\lambda\) 1 , \(\lambda\) 2 , \(\lambda\) 3 , \(\lambda\) 4 is the weight coefficient for balancing these loss terms. In one embodiment of the present invention, the initial values of these weights can be set to \(\lambda\) 1 = 0.4, \(\lambda\) 2 = 0.3, \(\lambda\) 3 = 0.2, \(\lambda\) 4 = 0.1, and are dynamically adjusted during the training process.

[0118] The motion chain multi-task regression neural network of the present invention further includes three key components: a multi-class neural network, an anomaly degree neural network, and a classification confidence neural network. The multi-class neural network is used to judge the joint activity state and can classify joint movements into four categories: normal, mildly abnormal, moderately abnormal, and severely abnormal. This network uses a softmax classifier, and its output can be expressed as:

[0119]

[0120] where \(x\) is the input feature, \(w\) i and \(b\) i are the weights and biases of the \(i\)-th class.

[0121] The anomaly degree neural network is used to calculate the degree of gait abnormality and adopts a regression model structure. The output of this network is a continuous value between 0 and 1, representing the degree of gait abnormality. The last layer of the network uses a sigmoid activation function:

[0122]

[0123] The classification confidence neural network is used to evaluate the reliability of the classification result, which is very important for decision-making in practical applications. This network also outputs a value between 0 and 1, representing the confidence in the classification result. The network structure is similar to that of the anomaly degree neural network, but the training objectives are different.

[0124] These three networks work together, not only being able to identify abnormal gaits, but also providing a quantitative index of the degree of abnormality and an evaluation of the reliability of the results, providing comprehensive information support for clinical decision-making.

[0125] Finally, the present invention also provides a multi-sensor fusion system for intelligent analysis of spinal gait characteristics for implementing the above method. The system includes a data acquisition module 1, a feature extraction module 2, a parameter analysis module 3, a relationship modeling module 4, a pattern learning module 5, an anomaly prediction module 6, and an output module 7.

[0126] The data acquisition module 1 adopts a distributed design and includes multiple inertial sensors, which are respectively installed at key parts of the human body, such as the soles of the feet, calves, thighs, waist, and chest. Each sensor includes a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer, and the sampling frequency can reach 100 Hz to ensure capturing subtle motion changes.

[0127] The feature extraction module 2 implements the multi-scale feature extraction algorithm described above. This module is implemented using FPGA (Field Programmable Gate Array) to meet the requirements of real-time processing. The parallel processing ability of FPGA enables the system to process data streams from multiple sensors simultaneously, greatly improving the computing efficiency.

[0128] The parameter analysis module 3, the relationship modeling module 4, and the pattern learning module 5 are integrated in a high-performance embedded processor, such as Nvidia Jetson Xavier NX. This kind of processor has powerful AI computing capabilities and can efficiently run complex neural network models.

[0129] The anomaly prediction module 6 is implemented based on an edge computing framework, and can perform real-time prediction on a local device, reducing data transmission latency and improving the system response speed. This module also includes a lightweight model update mechanism, which can continuously optimize the prediction model according to user feedback.

[0130] The output module 7 includes a user-friendly graphical interface, which can intuitively display gait analysis results, including key parameters such as gait cycle, symmetry, stability, etc., as well as anomaly warning information. In addition, this module also provides an API interface for easy integration with other health management systems.

[0131] In summary, the multi-sensor fusion method and system for intelligent analysis of spinal gait characteristics provided by the present invention, through innovatively combining a variety of advanced technologies, realizes comprehensive and accurate analysis and anomaly prediction of spinal gait. This method and system not only have significant technical advantages, but also show great potential in practical applications, can provide strong technical support for spinal health management, and open up new ways for preventing and early intervening in spinal diseases.

[0132] It should be noted that: the above are only the preferred embodiments of the present invention, and are not used to limit the present invention. Any modifications, equivalent replacements, and improvements made within the principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A multi-sensor fusion method for intelligent analysis of spinal gait characteristics, characterized in that: include: The acquisition steps include: The human body motion information is obtained through multiple inertial sensors distributed in different parts of the human body; Processing steps include: Extracting multi-scale gait features based on the human motion information; Analyzing gait period, symmetry, and stability parameters according to the multi-scale gait characteristics; Based on the human body movement information, a relationship model between spinal movement and lower limb movement is established; learning a normal gait pattern according to the relationship model; Output steps include: Generate spinal gait feature analysis results.

2. The method according to claim 1, characterized in that The acquisition step specifically includes: Install IMU sensors on the soles of the feet, calves, thighs, waist, and chest; The IMU sensor collects acceleration, angular velocity and magnetic field data.

3. The method according to claim 1, characterized in that The extracting of multi-scale gait features specifically includes: Fuse the acceleration and angular velocity to obtain the angular acceleration signal; Preprocess the raw data, including removing linear acceleration and gravity; Extract time domain features and frequency domain features.

4. The method according to claim 1, characterized in that: The analysis of gait cycle, symmetry and stability parameters specifically includes: The three-dimensional angular acceleration is input into the neural network, and the gait cycle is output; The acceleration signal is input into the neural network, and the gait symmetry is output; The three-dimensional angular acceleration is input into the neural network and the gait stability is output.

5. The method according to claim 1, characterized in that The establishment of the relationship model between spinal motion and lower limb motion specifically includes: Preprocess the raw data, including removing linear acceleration and gravity; Extract the movement characteristics of lower limbs and establish a coordinated relationship model between lower limbs and spine movements.

6. The method according to claim 1, characterized in that The learning of the normal gait pattern specifically includes: Preprocess historical data; Obtain gait features and perform multi-task learning; A multi-scale learning framework is introduced to learn normal gait patterns.

7. The method according to claim 1, characterized in that It also includes anomaly prediction steps: Constructing collaborative gait prediction model; The collaborative gait prediction model is used to perform early prediction of abnormal spinal gait.

8. The method according to claim 7, characterized in that The constructing of the collaborative gait prediction model specifically includes: Use spatiotemporal graph convolutional networks to analyze motion data and learn normal activity patterns; Introduce self-supervised learning methods to further learn and optimize the extracted features; A kinematic chain multi-task regression neural network is constructed to comprehensively model different kinematic parameters.

9. The method according to claim 8, characterized in that The kinematic chain multi-task regression neural network comprises: Multi-classification neural network, used to determine the state of joint activity; Abnormality neural network, used to calculate the degree of gait abnormality; Classification confidence neural network is used to evaluate the reliability of classification results.

10. A multi-sensor fusion system for intelligent analysis of spinal gait characteristics according to any one of claims 1 to 9, characterized in that: include: A data acquisition module, used to obtain human motion information through multiple inertial sensors; Feature extraction module, used to extract multi-scale gait features; Parameter analysis module, used to analyze gait cycle, symmetry, and stability parameters; A relationship modeling module is used to establish a relationship model between spinal motion and lower limb motion; Pattern learning module, used to learn normal gait patterns; An abnormality prediction module, used to predict abnormal gait of the spine; The output module is used to generate spinal gait feature analysis results.

Images