Intelligent shoe system based on motion detection data analysis

Abstract

The invention relates to the technical field of detection data analysis, in particular to an intelligent shoe system based on motion detection data analysis, which comprises a data collection module, a feature extraction and model training module and a model prediction module, and is characterized in that the data collection module collects historical and real-time user gait data, and the historical data comprises normal and abnormal gait data; the feature extraction and model training module automatically extracts low-dimensional potential features of normal gaits by using a variational auto-encoder, and records reconstruction errors; the normal gait low-dimensional data and the abnormal gait data with the abnormal labels are used for training a recurrent neural network model; the model prediction module calculates a dynamic error threshold according to the reconstruction error record, extracts low-dimensional features of the gait data in real time and performs anomaly judgment; and if the real-time reconstruction error exceeds a threshold value, determining that the gait is an abnormal gait, inputting the low-dimensional features into the recurrent neural network model, and outputting an abnormal type.

Description

Technical Field

[0001] The present invention relates to the technical field of detection data analysis, and in particular to a smart shoe system based on motion detection data analysis. Background Art

[0002] Traditional gait analysis systems rely on fixed feature extraction methods when processing high-dimensional gait data, which often lack adaptability and robustness. Specifically, although manually designed features perform well in certain specific environments or users, they cannot effectively capture the key information in all gait data, causing the system to perform poorly when faced with new data or different environments.

[0003] In addition, traditional systems usually do not have an effective method to evaluate and record the reconstruction error of the data, which makes the setting of anomaly detection thresholds lack of scientific basis and prone to false detection and missed detection. For example, a fixed threshold may perform well in the gait data of some users, but may fail in other users because the normal gait data of different users have significant differences in reconstruction errors. At the same time, traditional systems usually use simple rules or classifiers when classifying abnormal types. These methods have limited effectiveness in dealing with complex and multi-type abnormal gait data and are prone to confusing different types of abnormalities. For example, fatigue gait and injured gait may have similarities in features, resulting in inaccurate classification.

[0004] Therefore, the defects of traditional gait analysis systems in feature extraction, error assessment and anomaly classification make them lack robustness and accuracy in practical applications, making it difficult to meet the complex and changeable gait detection needs. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent shoe system based on motion detection data analysis to solve the problems raised in the above background technology.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solution: a smart shoe system based on motion detection data analysis, comprising a data collection module, a feature extraction and model training module and a model prediction module, wherein:

[0007] The data collection module collects historical user gait data and real-time user gait data, wherein the historical user gait data includes normal gait data and abnormal gait data;

[0008] The feature extraction and model training module uses a variational autoencoder to extract low-dimensional potential features of normal gait data as normal gait low-dimensional data, and records the reconstruction error between the normal gait data and the normal gait low-dimensional data as a reconstruction error record; and uses the normal gait low-dimensional data as the input of the recursive neural network model, and the abnormal gait data as an additional label in the recursive neural network model, and trains the recursive neural network model by defining and optimizing the loss function;

[0009] The model prediction module calculates an error threshold based on the reconstruction error record; uses a variational autoencoder to extract low-dimensional potential features from real-time user gait data as real-time low-dimensional features; and makes an abnormal judgment on the real-time low-dimensional features based on the error threshold; inputs the abnormal real-time low-dimensional features into a recursive neural network model, and outputs the abnormal type.

[0010] As a further improvement of the present technical solution, the abnormal gait data in the data collection module includes different abnormal types.

[0011] As a further improvement of the technical solution, the feature extraction and model training module includes a feature extraction unit, and the process of extracting normal gait low-dimensional data by the feature extraction unit specifically includes:

[0012] The variational autoencoder consists of multiple fully connected layers, each of which is responsible for compressing the normal gait data z into a lower-dimensional feature representation; through multiple fully connected layers, the normal gait data is gradually compressed into a lower-dimensional space, and the output of each layer of the network is nonlinearly transformed through an activation function;

[0013] In the last fully connected layer, the encoder generates two vectors, including the mean vector μ and the variance vector σ, which are used to represent the distribution of normal gait data z in the low-dimensional latent space;

[0014] A vector ε is randomly generated from a standard normal distribution, and the final low-dimensional latent features are calculated using the method z = u + ε·σ as the normal gait low-dimensional data z′.

[0015] As a further improvement of the technical solution, the process of calculating the reconstruction error record by the feature extraction unit specifically includes:

[0016] Record the reconstruction error between the normal gait data z and the normal gait low-dimensional data z′. For each feature dimension at each time point in the data, calculate the square of the difference between the normal gait data z and the normal gait low-dimensional data z′, sum all these square differences, and obtain the total reconstruction error as the reconstruction error record.

[0017] As a further improvement of the present technical solution, the feature extraction and model training module includes a model training unit, and the training process of the recursive neural network model performed by the model training unit specifically includes:

[0018] Filter out the user's gait data in abnormal state from the historical gait data, and add a corresponding abnormal type label to each abnormal gait data;

[0019] Calculate the cross entropy loss between the gait type probability distribution output by the recurrent neural network model and the true gait type label. The cross entropy loss is used to measure the difference between the probability distribution predicted by the model and the true label. The total gait recognition loss is obtained by summing the cross entropy losses at all time points.

[0020] Calculate the cross entropy loss between the anomaly type probability distribution output by the recurrent neural network model and the true anomaly type label. The cross entropy loss is used to measure the difference between the anomaly type probability distribution predicted by the model and the true label. The total anomaly detection loss is obtained by summing the cross entropy losses at all time points.

[0021] The total gait recognition loss and the total gait recognition loss are combined into a total loss function by weighted summation; the gradient of the total loss function with respect to the recursive neural network model parameters is calculated by the back propagation algorithm.

[0022] As a further improvement of the technical solution, the process of calculating the error threshold by the model prediction module specifically includes:

[0023] Divide all reconstruction error records in the feature extraction unit by the total number of records to obtain the average reconstruction error; calculate the standard deviation of all reconstruction error records, which is used to measure the fluctuation of the reconstruction error; add K times the standard deviation to the average reconstruction error to obtain the error threshold, where K is a constant.

[0024] As a further improvement of the technical solution, the process of the model prediction module outputting the abnormal type specifically includes:

[0025] Use the encoder part of the trained variational autoencoder to extract low-dimensional latent features from real-time user gait data as real-time low-dimensional features;

[0026] The real-time low-dimensional features are compared with the error threshold. For each feature dimension at each time point in the data, the square of the difference between the real-time user gait data and the real-time low-dimensional features is calculated, and all these square differences are summed to obtain the real-time reconstruction error. The calculated real-time reconstruction error is compared with the error threshold. If the real-time reconstruction error is greater than the error threshold, the real-time gait data is judged to be an abnormal gait; if the real-time reconstruction error is less than or equal to the error threshold, the real-time gait data is judged to be a normal gait.

[0027] For real-time gait data judged to be abnormal, the real-time low-dimensional features are input into the trained recursive neural network model. The recursive neural network model outputs the probability distribution of the abnormal type based on the input real-time low-dimensional features. The abnormal type of the real-time gait data is determined by selecting the abnormal type corresponding to the maximum value in the probability distribution.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] This intelligent shoe system based on motion detection data analysis automatically extracts low-dimensional potential features of gait data through variational autoencoders, overcoming the limitation of traditional systems that rely on manually designed features. The variational autoencoder can capture key information in gait data through multiple fully connected layers and nonlinear transformations, improving the adaptability and robustness of feature extraction. The system also records the reconstruction error between normal gait data and low-dimensional features, providing a scientific basis for subsequent threshold calculation, effectively avoiding the false detection and missed detection problems of fixed thresholds.

[0030] In addition, in terms of anomaly classification, the recursive neural network model can accurately identify multiple abnormal gait types by optimizing the total gait recognition loss and anomaly detection loss; the recursive neural network model is trained using historical gait data, can process time series data, and better capture the dynamic characteristics of gait changes; the system determines whether the gait is abnormal by calculating the real-time reconstruction error and comparing it with the dynamic error threshold, and further inputs the abnormal low-dimensional features into the recursive neural network model to output the anomaly type, which significantly improves the accuracy of anomaly detection and classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the overall module of the present invention;

[0032] Figure 2 Schematic diagram of the feature extraction and model training module unit of the present invention.

[0033] In the figure: 100, data collection module; 200, feature extraction and model training module; 201, feature extraction unit; 202, model training unit; 300, model prediction module. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0035] Next, see Figure 1-Figure 2 The present invention provides a technical solution: an intelligent shoe system based on motion detection data analysis, including a data collection module 100, a feature extraction and model training module 200 and a model prediction module 300.

[0036] The data collection module 100 collects real-time user gait data through sensors integrated in the smart shoes; collects historical user gait data using a public database, wherein the historical user gait data includes normal gait data and abnormal gait data; and the abnormal gait data includes different abnormal types.

[0037] The feature extraction unit 201 in the feature extraction and model training module 200 uses a variational autoencoder to extract low-dimensional potential features of the normal gait data as normal gait low-dimensional data, and records the reconstruction error between the normal gait data and the normal gait low-dimensional data as a reconstruction error record, specifically including:

[0038] The variational autoencoder consists of multiple fully connected layers, each of which is responsible for compressing the normal gait data z into a lower-dimensional feature representation; through multiple fully connected layers, the normal gait data is gradually compressed into a lower-dimensional space, for example, from 100 dimensions to 8 dimensions, and the output of each layer of the network is nonlinearly transformed by an activation function (such as ReLU);

[0039] In the last fully connected layer, the encoder generates two vectors, including the mean vector μ and the variance vector σ, which are used to represent the distribution of normal gait data z in the low-dimensional latent space;

[0040] In order to make the subsequent model easier to train, the reparameterization technique is used to sample the feature vector z′ from the latent space. Specifically, a vector ε is randomly generated from a standard normal distribution, and then the final low-dimensional latent feature is calculated using the method z = u + ε·σ as the normal gait low-dimensional data z′;

[0041] Record the reconstruction error between the normal gait data z and the normal gait low-dimensional data z′. For each feature dimension at each time point in the data, calculate the square of the difference between the normal gait data z and the normal gait low-dimensional data z′, sum all these square differences, and obtain the total reconstruction error as the reconstruction error record.

[0042] The model training unit 202 in the feature extraction and model training module 200 uses the normal gait low-dimensional data as the input of the recursive neural network model, and the abnormal gait data as an additional label in the recursive neural network model, and trains the recursive neural network model by defining and optimizing the loss function, specifically including:

[0043] Filter out the user's gait data in abnormal states from historical gait data, such as gait data in fatigue, injury, etc., and add a corresponding abnormal type label, such as "fatigue", "injury", etc., to each abnormal gait data;

[0044] Calculate the cross entropy loss between the gait type probability distribution output by the recurrent neural network model and the true gait type label. The cross entropy loss is used to measure the difference between the probability distribution predicted by the model and the true label. The total gait recognition loss is obtained by summing the cross entropy losses at all time points.

[0045] Calculate the cross entropy loss between the anomaly type probability distribution output by the recurrent neural network model and the true anomaly type label. The cross entropy loss is used to measure the difference between the anomaly type probability distribution predicted by the model and the true label. The total anomaly detection loss is obtained by summing the cross entropy losses at all time points.

[0046] The total gait recognition loss and the total gait recognition loss are combined into a total loss function by weighted summation; the gradient of the total loss function with respect to the recursive neural network model parameters is calculated by the back propagation algorithm, and the model parameters are updated using the gradient to achieve high accuracy and robustness in both gait recognition and anomaly detection tasks; for example, the Adam optimization algorithm can be used to update the parameters to gradually reduce the value of the total loss function.

[0047] The model prediction module 300 calculates the error threshold according to the reconstruction error record, specifically including:

[0048] Divide all reconstruction error records in the feature extraction unit 201 by the total number of records to obtain an average reconstruction error; and calculate the standard deviation of all reconstruction error records, the standard deviation is used to measure the fluctuation of the reconstruction error;

[0049] The error threshold is obtained by adding K times the standard deviation to the average reconstruction error. This ensures that the reconstruction error of most normal gait data is within the threshold range, while the reconstruction error of abnormal gait data usually exceeds this threshold. K is a constant, usually 2 or 3.

[0050] The model prediction module 300 uses a variational autoencoder to extract low-dimensional potential features from real-time user gait data as real-time low-dimensional features; and performs abnormal judgment on the real-time low-dimensional features according to the error threshold; inputs the abnormal real-time low-dimensional features into the recursive neural network model, and outputs the abnormal type, which specifically includes:

[0051] Use the encoder part of the trained variational autoencoder to extract low-dimensional latent features from real-time user gait data as real-time low-dimensional features;

[0052] The real-time low-dimensional features are compared with the error threshold. For each feature dimension at each time point in the data, the square of the difference between the real-time user gait data and the real-time low-dimensional features is calculated, and all these square differences are summed to obtain the real-time reconstruction error. The calculated real-time reconstruction error is compared with the error threshold. If the real-time reconstruction error is greater than the error threshold, the real-time gait data is judged to be an abnormal gait; if the real-time reconstruction error is less than or equal to the error threshold, the real-time gait data is judged to be a normal gait.

[0053] For real-time gait data that is judged to be abnormal, the real-time low-dimensional features are input into the trained recursive neural network model. The recursive neural network model outputs the probability distribution of the abnormal type based on the input real-time low-dimensional features. For example, if the abnormal types include "fatigue", "injury", "drunk", etc., a probability vector containing these abnormal types is output. The abnormal type of the real-time gait data is determined by selecting the abnormal type corresponding to the maximum value in the probability distribution. For example, if the abnormal type probability vector is [0.1, 0.6, 0.3], the abnormal type is determined to be "injury" because the probability of "injury" is the highest.

[0054] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. The smart shoe system based on motion detection data analysis is characterized by: The method comprises a data collection module (100), a feature extraction and model training module (200) and a model prediction module (300), wherein: The data collection module (100) collects historical user gait data and real-time user gait data, wherein the historical user gait data includes normal gait data and abnormal gait data; The feature extraction and model training module (200) uses a variational autoencoder to extract low-dimensional potential features of normal gait data as normal gait low-dimensional data, and records a reconstruction error between the normal gait data and the normal gait low-dimensional data as a reconstruction error record; and uses the normal gait low-dimensional data as an input of a recursive neural network model, and the abnormal gait data as an additional label in the recursive neural network model, and trains the recursive neural network model by defining and optimizing a loss function; The model prediction module (300) calculates an error threshold based on the reconstruction error record; uses a variational autoencoder to extract low-dimensional potential features from real-time user gait data as real-time low-dimensional features; and performs abnormality judgment on the real-time low-dimensional features based on the error threshold; inputs the abnormal real-time low-dimensional features into a recursive neural network model, and outputs the abnormality type.

2. The smart shoe system based on motion detection data analysis according to claim 1, characterized in that: The abnormal gait data in the data collection module (100) includes different abnormal types.

3. The intelligent shoe system based on motion detection data analysis according to claim 2, characterized in that: The feature extraction and model training module (200) comprises a feature extraction unit (201), and the process of extracting normal gait low-dimensional data by the feature extraction unit (201) specifically comprises: The variational autoencoder consists of multiple fully connected layers, each of which is responsible for compressing the normal gait data z into a lower-dimensional feature representation; through multiple fully connected layers, the normal gait data is gradually compressed into a lower-dimensional space, and the output of each layer of the network is nonlinearly transformed through an activation function; In the last fully connected layer, the encoder generates two vectors, including the mean vector μ and the variance vector σ, which are used to represent the distribution of normal gait data z in the low-dimensional latent space; A vector ε is randomly generated from a standard normal distribution, and the final low-dimensional latent features are calculated using the method z = u + ε·σ as the normal gait low-dimensional data z′.

4. The intelligent shoe system based on motion detection data analysis according to claim 3, characterized in that: The process of calculating the reconstruction error record by the feature extraction unit (201) specifically includes: Record the reconstruction error between the normal gait data z and the normal gait low-dimensional data z′. For each feature dimension at each time point in the data, calculate the square of the difference between the normal gait data z and the normal gait low-dimensional data z′, sum all these square differences, and obtain the total reconstruction error as the reconstruction error record.

5. The intelligent shoe system based on motion detection data analysis according to claim 4, characterized in that: The feature extraction and model training module (200) comprises a model training unit (202), and the training process of the recursive neural network model performed by the model training unit (202) specifically comprises: Filter out the user's gait data in abnormal state from the historical gait data, and add a corresponding abnormal type label to each abnormal gait data; Calculate the cross entropy loss between the gait type probability distribution output by the recurrent neural network model and the true gait type label. The cross entropy loss is used to measure the difference between the probability distribution predicted by the model and the true label. The total gait recognition loss is obtained by summing the cross entropy losses at all time points. Calculate the cross entropy loss between the anomaly type probability distribution output by the recurrent neural network model and the true anomaly type label. The cross entropy loss is used to measure the difference between the anomaly type probability distribution predicted by the model and the true label. The total anomaly detection loss is obtained by summing the cross entropy losses at all time points. The total gait recognition loss and the total gait recognition loss are combined into a total loss function by weighted summation; the gradient of the total loss function with respect to the recursive neural network model parameters is calculated by the back propagation algorithm.

6. The intelligent shoe system based on motion detection data analysis according to claim 5, characterized in that: The process of calculating the error threshold by the model prediction module (300) specifically includes: All reconstruction error records in the feature extraction unit (201) are divided by the total number of records to obtain an average reconstruction error; and the standard deviation of all reconstruction error records is calculated, and the standard deviation is used to measure the fluctuation of the reconstruction error; the average reconstruction error is added with K times the standard deviation to obtain an error threshold, where K is a constant.

7. The intelligent shoe system based on motion detection data analysis according to claim 6, characterized in that: The process of the model prediction module (300) outputting the abnormal type specifically includes: Use the encoder part of the trained variational autoencoder to extract low-dimensional latent features from real-time user gait data as real-time low-dimensional features; The real-time low-dimensional features are compared with the error threshold. For each feature dimension at each time point in the data, the square of the difference between the real-time user gait data and the real-time low-dimensional features is calculated, and all these square differences are summed to obtain the real-time reconstruction error. The calculated real-time reconstruction error is compared with the error threshold. If the real-time reconstruction error is greater than the error threshold, the real-time gait data is judged to be an abnormal gait; if the real-time reconstruction error is less than or equal to the error threshold, the real-time gait data is judged to be a normal gait. For real-time gait data judged to be abnormal, the real-time low-dimensional features are input into the trained recursive neural network model. The recursive neural network model outputs the probability distribution of the abnormal type based on the input real-time low-dimensional features. The abnormal type of the real-time gait data is determined by selecting the abnormal type corresponding to the maximum value in the probability distribution.

Images