Human body lower limb posture monitoring system based on flexible sensor

Abstract

The invention discloses a human body lower limb posture monitoring system based on a flexible sensor. The human body lower limb posture monitoring system comprises an array type flexible piezoresistive sensor module, a data acquisition and processing module, a feature selection module, a pressure data color mapping module, a wavelet neural network module and a joint bending angle result output module. An array type flexible piezoresistive sensor used for obtaining initial plantar pressure distribution data is designed, a method based on gait intervals is adopted to select data points, key feature points most related to joint angle changes are screened out based on time sequence information of gait events, the dimensionality of input data is reduced, and the accuracy of the input data is improved. And the structure and the training process of the network are optimized, and high-precision prediction of the lower limb joint angle is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of human motion feature recognition, and in particular to a human lower limb posture monitoring system based on a flexible sensor. Background Art

[0002] The human lower limb posture monitoring system is of great significance in the fields of rehabilitation medicine, sports science, virtual reality, augmented reality, and ergonomics. It can not only help doctors evaluate and formulate rehabilitation plans, but also provide athletes and coaches with feedback to optimize training. At the same time, in virtual reality and augmented reality applications, the system can enhance the user's immersion and interactive experience. In ergonomics, it helps to design products and environments that better meet human needs.

[0003] At present, traditional human posture monitoring systems face many technical challenges:

[0004] First, although the existing optical motion capture system can provide high-precision human lower limb posture motion data, its high cost and strict environmental requirements limit its application scope. Such system equipment is expensive and can usually only be operated in professional laboratories or specific controlled environments, which limits the popularity of the equipment in daily life. At the same time, because the system is easily affected by occlusion, resulting in data loss or error, its portability is insufficient and it is difficult to meet the application requirements in dynamic and mobile environments.

[0005] Second, as a traditional motion posture monitoring tool, the inertial measurement unit (IMU) faces problems such as drift, error accumulation, and environmental sensitivity. IMU sensors are prone to drift and error accumulation during long-term use, resulting in inaccurate data. In addition, IMU is greatly affected by the environmental magnetic field and requires a complex calibration process to ensure the accuracy of the measurement. This high sensitivity to the environment increases the complexity of data processing and fusion algorithms, and increases the difficulty of system implementation and maintenance.

[0006] Third, as a commonly used sensor technology, resistance strain gauges have problems such as insufficient sensitivity and dynamic range in the monitoring of fast dynamic movements. For large-scale and rapidly changing movements, the sensor may not provide sufficient sensitivity and response speed. In addition, resistance strain gauges need to fit tightly to specific parts, which affects the user's comfort and limits their application in wearable devices. Summary of the invention

[0007] In view of the above-mentioned deficiencies in the prior art, the present invention provides a human lower limb posture monitoring system based on flexible sensors, which can accurately estimate the bending angles of the hip joint, knee joint and ankle joint in real time, thereby realizing more real-time, accurate and space-time-free human lower limb posture monitoring and prediction.

[0008] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0009] Provided is a human lower limb posture monitoring system based on flexible sensors, which includes an array-type flexible piezoresistive sensor module, a data acquisition and processing module, a feature selection module, a pressure data color mapping module, a wavelet neural network module and a joint bending angle result output module;

[0010] An array-type flexible piezoresistive sensor module is used to obtain initial plantar pressure distribution data; wherein the initial plantar pressure distribution data is vertical ground reaction force data;

[0011] A data acquisition and processing module, used to receive and pre-process the initial plantar pressure distribution data from the array-type flexible piezoresistive sensor module, and send the pre-processed data to the feature selection module;

[0012] A feature selection module, used for selecting feature data from the preprocessed data and sending it to the pressure data color mapping module;

[0013] A pressure data color mapping module is used to map the selected feature data to a color space to form a pressure color map;

[0014] A wavelet neural network module is used to use the pressure color map as the input of the trained wavelet neural network, and obtain the angle points of the ankle, knee and hip corresponding to the pressure color map through the trained wavelet neural network;

[0015] The joint bending angle result output module is used to reconstruct the joint angle through the spline interpolation method based on the output of the wavelet neural network module to complete the lower limb posture monitoring of the human body.

[0016] Furthermore, the array-type flexible piezoresistive sensor module includes 18 flexible piezoresistive sensors flatly embedded in the sole, and each flexible piezoresistive sensor is connected to the input end of the data acquisition and processing module.

[0017] Furthermore, the 18 flexible piezoresistive sensors are divided into 6 rows, of which the first 2 rows each contain 4 flexible piezoresistive sensors, the middle 2 rows each contain 3 flexible piezoresistive sensors, and the last 2 rows each contain 2 flexible piezoresistive sensors. The measuring range of a single flexible piezoresistive sensor is 70 kg.

[0018] Furthermore, the data acquisition and processing module preprocesses the initial plantar pressure distribution data including data cleaning, filtering and normalization.

[0019] Furthermore, the specific method of filtering through the data acquisition and processing module is:

[0020] The cleaned data is filtered through a 1st-order Butterworth low-pass filter; the cutoff frequency of the 1st-order Butterworth low-pass filter is -3dB; the expression corresponding to the 1st-order Butterworth low-pass filter is:

[0021]

[0022] Where y[n] is the output signal after filtering at time n; f c is the cut-off frequency; f s is the sampling frequency; x[n] is the input signal at time n; x[n-1] is the input signal at time n-1; x[n-2] is the input signal at time n-2; y[n-1] is the output signal after filtering at time n-1; y[n-2] is the output signal after filtering at time n-2.

[0023] Furthermore, the specific method of selecting feature data from the preprocessed data is:

[0024] A single gait cycle is identified from the preprocessed data, and 22 time points are selected from each gait cycle. The pressure data corresponding to each time point is used as its characteristic value, and the 22 characteristic values ​​corresponding to each gait cycle are used as the characteristic data corresponding to the gait cycle; the 22 time points are respectively:

[0025] Initial contact IC: 0% The beginning of the gait cycle, the time point when the foot first touches the ground;

[0026] Contralateral toe-off (OTO): the time point when 10% of the contralateral foot's toes leave the ground;

[0027] Heel lift HR: 30% of the time when the heel of the supporting leg leaves the ground;

[0028] Contralateral initial contact OIC: the time point when 50% of the contralateral foot is in contact with the ground;

[0029] Initial toe-off (ITO): the time when the toe of the 60% swing leg leaves the ground;

[0030] Next toe-off NTO: the time when 87% of the contralateral foot enters the swing phase;

[0031] Foot adjustment FA: 73% foot adjustment phase, which is the point in time when the foot swing is about to end;

[0032] Tibial vertical TV: the time point when the tibia of the 50% stance leg reaches the vertical position;

[0033] Next initial contact NIC: the end point of the 100% gait cycle, i.e. the point when the foot contacts the ground again;

[0034] Load response time point LR: 5% duration of a single gait cycle;

[0035] The first mid-stance time point MST1: 15% of the duration of a single gait cycle;

[0036] The second mid-stance time point MST2: 25% of the duration of a single gait cycle;

[0037] The first end support time point TST1: 35% of the duration of a single gait cycle;

[0038] The second end support time point TST2: 45% of the duration of a single gait cycle;

[0039] Post-swing time point PSW: 55% of the duration of a single gait cycle;

[0040] The first initial swing time point ISW1: 62% of the duration of a single gait cycle;

[0041] The second initial swing time point ISW2: 69% of the duration of a single gait cycle;

[0042] The first mid-swing time point MSW1: 74% of the duration of a single gait cycle;

[0043] The second mid-swing time point MSW2: 79% of the duration of a single gait cycle;

[0044] The third mid-swing time point MSW3: 84% of the duration of a single gait cycle;

[0045] The first terminal swing time point TSW1: 89% of the duration of a single gait cycle;

[0046] The second terminal swing time point TSW2 is 95% of the duration of a single gait cycle.

[0047] Further, the wavelet neural network includes an input layer, a hidden layer, and an output layer connected in sequence, and a direct connection layer connecting the input layer and the output layer;

[0048] The calculation expression of wavelet neural network is:

[0049]

[0050] in is the jth output of the output layer of the wavelet neural network; N o and N l Respectively represent the number of nodes in the hidden layer and the number of input variables; w ij is the weight between hidden layer node i and output layer node j; w kjis the weight between the input layer node k and the output layer node j, that is, the weight of the direct connection layer; x k represents the value of the kth input variable; b j is the bias of the jth output layer node; ψ i (x) represents the multidimensional wavelet function, through N l The product of scalar wavelets, x is the input vector; ψ(·) represents the mother wavelet function; x i is the i-th input variable of the input layer; is the scaling parameter.

[0051] Furthermore, it also includes a wavelet neural network training module for training the wavelet neural network.

[0052] Furthermore, the specific method for training the wavelet neural network includes the following sub-steps:

[0053] A1. Let the trainee wear shoes equipped with array-type flexible piezoresistive sensor modules, and place reflective marks on the trainee's body parts and joints. Use a motion camera to capture the trainee's walking motion to obtain several image frames.

[0054] A2. Use motion capture data analysis software to digitize the walking motion captured by the motion camera, and obtain the joint angles in the image frame corresponding to the feature data selection points as training labels;

[0055] A3, receiving and preprocessing the initial plantar pressure distribution data of the trainee during walking through the data acquisition and processing module, and obtaining the pressure data corresponding to each image frame after preprocessing;

[0056] A4, selecting feature data from the pressure data corresponding to each image frame after preprocessing through a feature selection module;

[0057] A5, mapping the feature data selected in step A4 to a color space through a pressure data color mapping module to form a pressure color map, and using the pressure color map as a training sample;

[0058] A6. Initialize the parameters of the wavelet neural network, use the training samples as the input of the wavelet neural network, and obtain the angle points of the ankle, knee, and hip corresponding to the output of the wavelet neural network;

[0059] A7, based on the angle points of the ankle, knee and hip output by the wavelet neural network, the spline interpolation method is used to reconstruct the joint angle, that is, the joint angle predicted by the wavelet neural network is obtained;

[0060] A8. According to the formula:

[0061]

[0062] Calculate the loss function L of the wavelet neural network; where N is the total number of training samples;

[0063] A9. With the goal of minimizing the loss function L, the back propagation algorithm is used to update the parameters of the wavelet neural network until the loss function L converges, completing the training of the wavelet neural network.

[0064] Furthermore, the specific method for initializing the parameters of the wavelet neural network is:

[0065] Randomly initialize the weights and biases in the wavelet neural network according to the formula:

[0066] w (ξ)ki =0.5(N k +M k )

[0067] w (ζ)ki =100(N k -M k )

[0068] Get parameter w (ξ)ki and parameter w (ζ)ki The initial value of w (ζ)ki ≤1;M k =max P＝1,...,N (x KP ), N k =min P＝1,...,N (x KP ), x KP represents the Kth input variable in the Pth training sample; N is the total number of training samples; max represents the maximum value, min represents the minimum value, M k and N k are the maximum and minimum values ​​in the input variables, respectively.

[0069] The beneficial effects of the present invention are:

[0070] 1. The present invention designs an array-type flexible piezoresistive sensor for obtaining initial plantar pressure distribution data. It adopts a gait interval-based method to select data points, and screens out key feature points most relevant to joint angle changes based on the timing information of gait events. This not only reduces the dimension of input data, but also optimizes the network structure and training process, thereby achieving high-precision prediction of lower limb joint angles.

[0071] 2. The flexible piezoresistive sensor can adhere closely to the human body surface and obtain accurate plantar pressure distribution data, while the wavelet neural network effectively processes and analyzes this data to accurately estimate the bending angles of the hip, knee and ankle joints.

[0072] 3. The system has the ability to capture and process lower limb motion data in real time, and can provide users with instant posture feedback. This feature makes the system particularly useful in application scenarios such as sports training and rehabilitation therapy, and can help users adjust their posture in time, optimize training effects, and promote rehabilitation progress.

[0073] 4. Compared with traditional optical motion capture systems and inertial measurement units, this system has lower hardware costs, which enables the system to be promoted and applied at a lower cost while ensuring accuracy, and has a high market penetration. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 This is the structural diagram of the system;

[0075] Figure 2 It is a schematic diagram of the flexible piezoresistive sensor points;

[0076] Figure 3 It is a schematic diagram of the structure of wavelet neural network;

[0077] Figure 4 Schematic diagram of joint bending angle. DETAILED DESCRIPTION

[0078] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0079] like Figure 1 As shown, the human lower limb posture monitoring system based on flexible sensors includes an array flexible piezoresistive sensor module, a data acquisition and processing module, a feature selection module, a pressure data color mapping module, a wavelet neural network module and a joint bending angle result output module;

[0080] An array-type flexible piezoresistive sensor module is used to obtain initial plantar pressure distribution data; wherein the initial plantar pressure distribution data is vertical ground reaction force data;

[0081] A data acquisition and processing module, used to receive and pre-process the initial plantar pressure distribution data from the array-type flexible piezoresistive sensor module, and send the pre-processed data to the feature selection module;

[0082] A feature selection module, used for selecting feature data from the preprocessed data and sending it to the pressure data color mapping module;

[0083] A pressure data color mapping module is used to map the selected feature data to a color space to form a pressure color map;

[0084] A wavelet neural network module is used to use the pressure color map as the input of the trained wavelet neural network, and obtain the angle points of the ankle, knee and hip corresponding to the pressure color map through the trained wavelet neural network;

[0085] The joint bending angle result output module is used to reconstruct the joint angle through the spline interpolation method based on the output of the wavelet neural network module to complete the lower limb posture monitoring of the human body.

[0086] In this embodiment, the vertical ground reaction forces (vGRF) selected from the initial plantar pressure distribution data are represented by the stance feature as f and the swing feature as f′. During the swing phase, the foot leaves the ground and the vGRF of the swing foot is zero. When the swing foot leaves the ground, the contralateral foot is in the stance phase. During stance, the foot contacts the ground and produces a non-zero vGRF. Therefore, the vGRF feature of the swing foot is selected relative to the vGRF signal f′ of the contralateral foot. This process provides a unique set of features that avoids the zero vGRF of the swing foot. The measured angle is represented as the target and is recorded as θ.

[0087] This embodiment is visualized by a pressure color map, where the pressure value of each region is represented by a color. The relationship between color and pressure value is usually illustrated by a colorbar, which shows the range of pressure (e.g., low to high pressure) and the colors corresponding to these ranges. In a specific implementation, assuming that the pressure value is between the interval [Pmin, Pmax], the mapping function can be defined as follows:

[0088]

[0089] Among them, colormap refers to the color mapping function, P0 is the pressure value of the current point, P min and P max are the minimum and maximum values ​​of pressure.

[0090] like Figure 2 As shown, the array-type flexible piezoresistive sensor module includes 18 flexible piezoresistive sensors flatly embedded in the sole, and each flexible piezoresistive sensor is connected to the input end of the data acquisition and processing module. When embedding, ensure that the sensor area is flat and fits the test surface to avoid sharp objects from acting on the sensing area.

[0091] 18 flexible piezoresistive sensors are divided into 6 rows, of which the first 2 rows are 4 flexible piezoresistive sensors, the middle 2 rows are 3 flexible piezoresistive sensors, and the last 2 rows are 2 flexible piezoresistive sensors. The range of a single flexible piezoresistive sensor is 70 kg, the thickness of the flexible piezoresistive sensor is less than 0.3 mm, the terminal extension length is 5 mm, the center distance of the outlet line is 2.54 mm, and the terminal type is a male terminal Pin. The wiring parameters are 1 group of 10 pins. The static resistance is greater than 1 megohm, the hysteresis is less than 6%, the drift is less than 8%, the sensor repeatability is ±10%, the operating voltage range is 3.3 to 5 volts, the operating temperature range is -50 to +50 degrees Celsius, the operating humidity range is 0 to 95%, the response time is less than 20 milliseconds, the numerical stabilization time is less than 5 seconds, there is no electromagnetic interference and electrostatic discharge, and the service life exceeds 500,000 positive pressure cycle tests.

[0092] In this embodiment, the data acquisition and processing module includes a Bluetooth module HC-05, and the data of the array-type flexible piezoresistive sensor module is collected and stored in real time to the host computer through the Bluetooth module HC-05, so as to facilitate subsequent preprocessing.

[0093] The data acquisition and processing module preprocesses the initial plantar pressure distribution data including data cleaning, filtering and normalization.

[0094] The specific method of filtering through the data acquisition and processing module is:

[0095] The cleaned data is filtered through a 1st-order Butterworth low-pass filter; the cutoff frequency of the 1st-order Butterworth low-pass filter is -3dB, which is determined by the fast Fourier transform spectrum; the expression corresponding to the 1st-order Butterworth low-pass filter is:

[0096]

[0097] Where y[n] is the output signal after filtering at time n; f c is the cut-off frequency; f s is the sampling frequency; x[n] is the input signal at time n; x[n-1] is the input signal at time n-1; x[n-2] is the input signal at time n-2; y[n-1] is the output signal after filtering at time n-1; y[n-2] is the output signal after filtering at time n-2.

[0098] The specific method for selecting feature data from preprocessed data is:

[0099] A single gait cycle is identified from the preprocessed data, and 22 time points are selected from each gait cycle. The pressure data corresponding to each time point is used as its characteristic value, and the 22 characteristic values ​​corresponding to each gait cycle are used as the characteristic data corresponding to the gait cycle; the 22 time points are respectively:

[0100] Initial Contact (IC): 0% The beginning of the gait cycle, the time when the foot first touches the ground;

[0101] Opposite Toe Off (OTO): The time point when 10% of the toes of the contralateral foot leave the ground;

[0102] Heel Rise HR: The time when the heel of the 30% supporting leg leaves the ground;

[0103] Opposite initial contact (OIC): the time point when 50% of the contralateral foot contacts the ground;

[0104] Initial Toe Off (ITO): The time when the toe of the 60% swing leg leaves the ground;

[0105] Next Toe Off (NTO): The time point when 87% of the contralateral foot enters the swing phase;

[0106] Foot Adjustment FA (Foot Adjustment): 73% of the foot adjustment phase, which is the time point when the foot swing is about to end;

[0107] Tibia Vertical TV: The time point when the tibia of the 50% stance leg reaches the vertical position;

[0108] Next Initial Contact (NIC): The end point of the 100% gait cycle, i.e. the time when the foot touches the ground again;

[0109] Loading Response time point LR (Loading Response): 5% duration of a single gait cycle;

[0110] The first mid-stance time point MST1 (MidStance 1): 15% of the duration of a single gait cycle;

[0111] The second mid-stance time point MST2 (MidStance 2): 25% of the duration of a single gait cycle;

[0112] Terminal Stance 1 (TST1): 35% of the duration of a single gait cycle.

[0113] The second terminal support time point TST2 (TerminalStance 2): 45% of the duration of a single gait cycle;

[0114] Pre-swing time point PSW (Pre Swing): 55% of the duration of a single gait cycle;

[0115] The first initial swing time point ISW1 (Initial Swing1): 62% of the duration of a single gait cycle;

[0116] The second initial swing time point ISW2 (Initial Swing2): 69% of the duration of a single gait cycle;

[0117] The first mid-swing time point MSW1 (Mid Swing 1): 74% of the duration of a single gait cycle;

[0118] The second mid-swing time point MSW2 (Mid Swing 2): 79% of the duration of a single gait cycle;

[0119] The third mid-swing time point MSW3 (Mid Swing 3): 84% of the duration of a single gait cycle;

[0120] The first terminal swing time point TSW1 (Terminal Swing 1): 89% of the duration of a single gait cycle;

[0121] The second terminal swing time point TSW2 (Terminal Swing 2): 95% of the duration of a single gait cycle.

[0122] like Figure 3 As shown, the wavelet neural network includes an input layer, a hidden layer, and an output layer connected in sequence, as well as a direct connection layer connecting the input layer and the output layer;

[0123] In this embodiment, the wavelet neural network models the relationship between input and output variables by considering the linear and nonlinear aspects of the input and output variables respectively. The linear relationship between the input and output is adjusted by adding a direct connection layer, while the nonlinear relationship is adjusted by the activation of the hidden layer nodes. The hidden layer nodes are activated by the wavelet activation function, and the input of the hidden layer nodes is nonlinearly transformed by the wavelet mother function to generate the required output, that is, the calculation expression of the wavelet neural network is:

[0124]

[0125] in is the jth output of the output layer of the wavelet neural network; N o and N l Respectively represent the number of nodes in the hidden layer and the number of input variables; w ij is the weight between hidden layer node i and output layer node j; w kj is the weight between the input layer node k and the output layer node j, that is, the weight of the direct connection layer; x k represents the value of the kth input variable; b j is the bias of the jth output layer node; ψ i (x) represents the multidimensional wavelet function, through N l The product of scalar wavelets, x is the input vector; ψ(·) represents the mother wavelet function; x i is the i-th input variable of the input layer; is the scaling parameter.

[0126] In this embodiment, each hidden layer node is activated by a wavelet activation function. The choice of the wavelet mother function depends on the specific application. The Mexican hat wavelet is a real-valued symmetric function that performs well in identifying the peaks and troughs of nonlinear data structures. Therefore, the Mexican hat wavelet function will be an ideal choice for capturing the nonlinear patterns of plantar pressure distribution data and joint angles:

[0127]

[0128] In the output layer, the linear activation ψ i (x)w ij and x k w ij The weighted sum is calculated and the bias term b is added j , generating the final network output.

[0129] The number of hidden layer nodes is chosen based on the minimum prediction risk principle, where prediction risk is the expected performance of the network on the test data. The number of hidden layer nodes can be set to 5.

[0130]

[0131] in, is the target output, is the network output for the test sample p. N tst is the total number of test samples. In this embodiment, N tst The number of samples is set to 90. This embodiment selects the number of hidden layer nodes that generates the minimum predicted risk. According to the reported minimum predicted risk, this wavelet neural network uses five hidden layer nodes.

[0132] In this embodiment, a wavelet neural network training module is also included for training the wavelet neural network. The specific method for training the wavelet neural network includes the following sub-steps:

[0133] A1. Have the trainees wear shoes equipped with array-type flexible piezoresistive sensor modules, and place reflective markers on the trainees' body parts and joints, such as Figure 4 As shown, the walking motion of the trainee is captured by a motion camera to obtain several image frames;

[0134] Specifically, 30 young adults (aged between 20 and 30 years old, with a height of 160–190 cm and a weight of 50–100 kg) with no history of lower limb injury, surgery, and gait disorder were selected as subjects for human dynamic gait data collection;

[0135] A2. Use motion capture data analysis software to digitize the walking motion captured by the motion camera, and obtain the joint angles in the image frame corresponding to the feature data selection points as training labels;

[0136] Specifically, the corresponding digitization can be achieved through Visual3D motion capture data analysis software; gait parameters are calculated by converting dynamic Qualisys motion files into Visual 3D static bone and joint models;

[0137] A3, receiving and preprocessing the initial plantar pressure distribution data of the trainee during walking through the data acquisition and processing module, and obtaining the pressure data corresponding to each image frame after preprocessing;

[0138] A4. Select feature data from the pressure data corresponding to each image frame after preprocessing through a feature selection module; the structure of the feature data is:

[0139]

[0140] Where x is the feature data set; n is the total number of samples; m is the total number of feature data, which is 22; f m,n Represents the mth feature data of the nth sample;

[0141] A5, mapping the feature data selected in step A4 to a color space through a pressure data color mapping module to form a pressure color map, and using the pressure color map as a training sample;

[0142] A6. Initialize the parameters of the wavelet neural network, use the training samples as the input of the wavelet neural network, and obtain the angle points of the ankle, knee and hip corresponding to the output of the wavelet neural network; the expression is:

[0143]

[0144] Where y is the output set of the wavelet neural network; Indicates the ankle angle point corresponding to the mth feature data of the nth sample; Indicates the knee angle point corresponding to the mth feature data of the nth sample; Indicates the hip angle point corresponding to the mth feature data of the nth sample;

[0145] A7, based on the angle points of the ankle, knee and hip output by the wavelet neural network, the spline interpolation method is used to reconstruct the joint angle, that is, the joint angle predicted by the wavelet neural network is obtained;

[0146] A8. According to the formula:

[0147]

[0148] Calculate the loss function L of the wavelet neural network; where N is the total number of training samples;

[0149] A9. With the goal of minimizing the loss function L, the back propagation algorithm is used to update the parameters of the wavelet neural network until the loss function L converges, completing the training of the wavelet neural network.

[0150] In this embodiment, the loss function lower limit L min is set to 0.00001, epoch(E) is set to 1, and the step size is 1. The maximum training cycle Emax is set to 50. The training performance is evaluated at the 50th cycle to verify whether the WNN has reached its target state, otherwise the network will continue to train until it reaches its optimal state. The samples are shuffled and randomly grouped for training and testing. The random selection process is repeated three times to generate three groups (sequences) of input data, namely: (1) sequence 1, (2) sequence 2, and (3) sequence 3. The network is trained and tested using these three sets of data sequences to study the repeatability of the network's estimated performance. 70% of the shuffled data samples (equivalent to 210 samples) are used to train N trn The remaining 30% of the samples (equivalent to 90 samples) are used to test N tst .

[0151] The specific method to initialize the parameters of the wavelet neural network is:

[0152] Randomly initialize the weights and biases in the wavelet neural network according to the formula:

[0153] w (ξ)ki =0.5(N k +M k )

[0154] w (ζ)ki =100(N k -M k )

[0155] Get parameter w (ξ)ki and parameter w (ζ)ki The initial value of w (ζ)ki ≤1;M k =max P＝1,...,N (x KP ), N k =min P＝1,...,N (x KP ), x KP represents the Kth input variable in the Pth training sample; N is the total number of training samples; max represents the maximum value, min represents the minimum value, M k and N k are the maximum and minimum values ​​in the input variables, respectively.

[0156] In summary, the present invention designs an array-type flexible piezoresistive sensor for obtaining initial plantar pressure distribution data, adopts a gait interval-based method to select data points, and screens out key feature points most relevant to joint angle changes based on the timing information of gait events. This not only reduces the dimension of input data, but also optimizes the network structure and training process, thereby achieving high-precision prediction of lower limb joint angles.

Claims

1. A human lower limb posture monitoring system based on flexible sensors, characterized in that: It includes an array-type flexible piezoresistive sensor module, a data acquisition and processing module, a feature selection module, a pressure data color mapping module, a wavelet neural network module and a joint bending angle result output module; An array-type flexible piezoresistive sensor module is used to obtain initial plantar pressure distribution data; wherein the initial plantar pressure distribution data is vertical ground reaction force data; A data acquisition and processing module, used to receive and pre-process the initial plantar pressure distribution data from the array-type flexible piezoresistive sensor module, and send the pre-processed data to the feature selection module; A feature selection module, used for selecting feature data from the preprocessed data and sending it to the pressure data color mapping module; A pressure data color mapping module is used to map the selected feature data to a color space to form a pressure color map; A wavelet neural network module is used to use the pressure color map as the input of the trained wavelet neural network, and obtain the angle points of the ankle, knee and hip corresponding to the pressure color map through the trained wavelet neural network; The joint bending angle result output module is used to reconstruct the joint angle through the spline interpolation method based on the output of the wavelet neural network module to complete the lower limb posture monitoring of the human body.

2. The human lower limb posture monitoring system based on flexible sensors according to claim 1 is characterized in that: The array-type flexible piezoresistive sensor module includes 18 flexible piezoresistive sensors flatly embedded in the sole, and each flexible piezoresistive sensor is connected to the input end of the data acquisition and processing module.

3. The human lower limb posture monitoring system based on flexible sensors according to claim 2 is characterized in that: The 18 flexible piezoresistive sensors are divided into 6 rows, of which the first 2 rows each contain 4 flexible piezoresistive sensors, the middle 2 rows each contain 3 flexible piezoresistive sensors, and the last 2 rows each contain 2 flexible piezoresistive sensors. The measuring range of a single flexible piezoresistive sensor is 70 kg.

4. The human lower limb posture monitoring system based on flexible sensors according to claim 1 is characterized in that: The data acquisition and processing module preprocesses the initial plantar pressure distribution data including data cleaning, filtering and normalization.

5. The human lower limb posture monitoring system based on flexible sensors according to claim 4 is characterized in that: The specific method of filtering through the data acquisition and processing module is: The cleaned data is filtered through a 1st-order Butterworth low-pass filter; the cutoff frequency of the 1st-order Butterworth low-pass filter is -3dB; the expression corresponding to the 1st-order Butterworth low-pass filter is: Where y[n] is the output signal after filtering at time n; f c is the cut-off frequency; f s is the sampling frequency; x[n] is the input signal at time n; x[n-1] is the input signal at time n-1; x[n-2] is the input signal at time n-2; y[n-1] is the output signal after filtering at time n-1; y[n-2] is the output signal after filtering at time n-2.

6. The human lower limb posture monitoring system based on flexible sensors according to claim 1 is characterized in that: The specific method for selecting feature data from preprocessed data is: A single gait cycle is identified from the preprocessed data, and 22 time points are selected from each gait cycle. The pressure data corresponding to each time point is used as its characteristic value, and the 22 characteristic values ​​corresponding to each gait cycle are used as the characteristic data corresponding to the gait cycle; the 22 time points are respectively: Initial contact IC: 0% The beginning of the gait cycle, the time point when the foot first touches the ground; Contralateral toe-off (OTO): the time point when 10% of the contralateral foot's toes leave the ground; Heel lift HR: 30% of the time when the heel of the supporting leg leaves the ground; Contralateral initial contact OIC: the time point when 50% of the contralateral foot is in contact with the ground; Initial toe-off (ITO): the time when the toe of the 60% swing leg leaves the ground; Next toe-off NTO: the time when 87% of the contralateral foot enters the swing phase; Foot adjustment FA: 73% foot adjustment phase, which is the point in time when the foot swing is about to end; Tibial vertical TV: the time point when the tibia of the 50% stance leg reaches the vertical position; Next initial contact NIC: the end point of the 100% gait cycle, i.e. the point when the foot contacts the ground again; Load response time point LR: 5% duration of a single gait cycle; The first mid-support time point MST1: 15% of the duration of a single gait cycle; The second mid-stance time point MST2: 25% of the duration of a single gait cycle; The first end support time point TST1: 35% of the duration of a single gait cycle; The second end support time point TST2: 45% of the duration of a single gait cycle; Post-swing time point PSW: 55% of the duration of a single gait cycle; The first initial swing time point ISW1: 62% of the duration of a single gait cycle; The second initial swing time point ISW2: 69% of the duration of a single gait cycle; The first mid-swing time point MSW1: 74% of the duration of a single gait cycle; The second mid-swing time point MSW2: 79% of the duration of a single gait cycle; The third mid-swing time point MSW3: 84% of the duration of a single gait cycle; The first terminal swing time point TSW1: 89% of the duration of a single gait cycle; The second terminal swing time point TSW2 is 95% of the duration of a single gait cycle.

7. The human lower limb posture monitoring system based on flexible sensors according to claim 1 is characterized in that: The wavelet neural network includes an input layer, a hidden layer, and an output layer connected in sequence, as well as a direct connection layer connecting the input layer and the output layer; The calculation expression of wavelet neural network is: in is the jth output of the output layer of the wavelet neural network; N o and N l Respectively represent the number of nodes in the hidden layer and the number of input variables; w ij is the weight between hidden layer node i and output layer node j; w kj is the weight between the input layer node k and the output layer node j, that is, the weight of the direct connection layer; x k represents the value of the kth input variable; b j is the bias of the jth output layer node; ψ i (x) represents the multidimensional wavelet function, through N l The product of scalar wavelets, x is the input vector; ψ(·) represents the mother wavelet function; x i is the i-th input variable of the input layer; is the scaling parameter.

8. The human lower limb posture monitoring system based on flexible sensors according to claim 7 is characterized in that: It also includes a wavelet neural network training module for training the wavelet neural network.

9. The human lower limb posture monitoring system based on flexible sensors according to claim 8 is characterized in that: The specific method for training the wavelet neural network includes the following sub-steps: A1. Let the trainee wear shoes equipped with array-type flexible piezoresistive sensor modules, and place reflective marks on the trainee's body parts and joints. Use a motion camera to capture the trainee's walking motion to obtain several image frames. A2. Use motion capture data analysis software to digitize the walking motion captured by the motion camera, and obtain the joint angles in the image frame corresponding to the feature data selection points as training labels; A3, receiving and preprocessing the initial plantar pressure distribution data of the trainee during walking through the data acquisition and processing module, and obtaining the pressure data corresponding to each image frame after preprocessing; A4, selecting feature data from the pressure data corresponding to each image frame after preprocessing through a feature selection module; A5, mapping the feature data selected in step A4 to a color space through a pressure data color mapping module to form a pressure color map, and using the pressure color map as a training sample; A6. Initialize the parameters of the wavelet neural network, use the training samples as the input of the wavelet neural network, and obtain the angle points of the ankle, knee, and hip corresponding to the output of the wavelet neural network; A7, based on the angle points of the ankle, knee and hip output by the wavelet neural network, the spline interpolation method is used to reconstruct the joint angle, that is, the joint angle predicted by the wavelet neural network is obtained; A8. According to the formula: Calculate the loss function L of the wavelet neural network; where N is the total number of training samples; A9. With the goal of minimizing the loss function L, the back propagation algorithm is used to update the parameters of the wavelet neural network until the loss function L converges, completing the training of the wavelet neural network.

10. The human lower limb posture monitoring system based on flexible sensors according to claim 9, characterized in that: The specific method to initialize the parameters of the wavelet neural network is: Randomly initialize the weights and biases in the wavelet neural network according to the formula: w (ξ)ki =0.5(N k +M k ) w (ζ)ki =100(N k -M k ) Get parameter w (ξ)ki and parameter w (ζ)ki The initial value of w (ζ)ki ≤1;M k =max P＝1,...,N (x KP ), N k =min P＝1,...,N (x KP ), x KP represents the Kth input variable in the Pth training sample; N is the total number of training samples; max represents the maximum value, min represents the minimum value, M k and N k are the maximum and minimum values ​​in the input variables, respectively.

Images