Gait phase recognition method based on neural network

Abstract

The present invention discloses a gait phase recognition method based on a neural network. The operation steps of the method are: a. constructing a gait data acquisition device, b. gait data acquisition experiment, c. lower limb walking mechanism analysis and gait phase division, d. building a gait phase recognition model based on a multi-layer perceptron neural network algorithm to perform gait phase recognition, e. building a gait data set, and f. multi-motion mode gait phase recognition. The purpose of the present invention is to build a plantar pressure measurement system for collecting motion gait data, use a gait phase recognition algorithm based on a neural network to carry out gait phase recognition research in walking mode and running mode, and optimize and adjust the neural network recognition model by comparing the influence of different network structures on the recognition accuracy.

Description

Technical Field

[0001] The present invention relates to a gait phase recognition method based on a neural network model, which is applied to the research field of plantar pressure information perception and human body motion state prediction. Background Art

[0002] As a powerful human-machine coupling system, the lower limb exoskeleton robot needs to accurately perceive the operator's action state and behavioral intention. The gait consistency of the exoskeleton is mainly determined by the following three aspects: gait trajectory generation, gait execution and gait evaluation, all of which are related to the gait phase. The gait phase must be accurately identified by the human-machine system to accurately control the exoskeleton to produce motions to assist the wearer. Generally, if one leg is used as a reference, a gait cycle mainly consists of two phases, namely the stance phase and the swing phase; if two legs are used as a reference, a complete gait cycle includes a unipedal stance phase, a bipedal stance phase (walking) or a bipedal swing phase (running). So far, several gait phase division models with different levels of accuracy have been proposed according to different clinical purposes. However, it is intuitive and reasonable to divide the gait cycle into the stance phase and the swing phase by judging whether the currently observed foot is in contact with the ground. However, there is no accurate division principle for the division of gait sub-phases. Taking the lower limbs as a reference and considering the actual application scenarios of powered exoskeletons, it is very necessary to recognize gait in different motion modes. It is challenging to develop an algorithm that can recognize corresponding gait phases in different motion modes. Summary of the invention

[0003] In order to solve the problems of the prior art, the purpose of the present invention is to overcome the shortcomings of the existing technology and provide a gait phase recognition method based on a neural network. A gait phase recognition algorithm based on a neural network is used to carry out gait phase recognition research in walking mode and running mode. By comparing the influence of different network structures on the recognition accuracy, the neural network recognition model is optimized and adjusted.

[0004] In order to achieve the above invention purpose, the present invention adopts the following technical solution:

[0005] A gait phase recognition method based on a neural network, the operation steps are as follows:

[0006] a. Build gait data collection equipment;

[0007] b. Gait data collection experiment;

[0008] c. Analysis of lower limb walking mechanism and gait phase division;

[0009] d. Establish a gait phase recognition model based on a multi-layer perceptron neural network algorithm to perform gait phase recognition;

[0010] e. Establishment of gait dataset;

[0011] f. Multi-motion mode gait phase recognition.

[0012] Preferably, the gait data acquisition device system of step a is composed of a sensor module, a signal conditioning module, a single-chip control module, a data storage module and a power supply module; it is installed at a fixed position on a human limb for easy carrying to collect signals; the integrated circuit board module is designed according to the package size of the single-chip microcomputer Arduino Uno, and all the modules are integrated into a whole; after the signal is processed by the signal conditioning module, it enters the single-chip microcomputer module through the multiplexer unit under the control of the single-chip microcomputer, and is processed by the analog-to-digital conversion module built in the Arduino microcontroller to become a digital signal, and finally stored in the data storage module SD card for subsequent data analysis and processing; a flexible sensing insole designed based on the FSR sensing unit is used to monitor the plantar pressure signal, with 8 sensing units for one foot and 16 sensing units for both feet, wherein the position of the sensing unit is arranged according to the physiological characteristics and load-bearing characteristics of the foot, and is divided according to the sole area; 3 sensing units are arranged in the metatarsal and heel areas, and 1 sensing unit is placed in the toe and arch areas.

[0013] Preferably, the specific steps of the gait data collection experiment in step b are:

[0014] (1) Recruiting multiple testers who have never suffered from gait-related diseases to participate in the wearable test;

[0015] (2) Place the designed plantar pressure measurement system in the shoes of each test subject, and then use a nylon buckle belt to fix the signal acquisition module around the waist;

[0016] (3) To prevent the wires from swinging and affecting normal walking and causing unnatural movements, and to facilitate wearing and taking off the experimental equipment, a cable tie with a nylon buckle was used above the knee joint to restrain the wires to the back of the lower limbs;

[0017] (4) Place a power bank with sufficient power in the collector's pocket to power the collection system, ensuring that the tester can achieve the natural swinging state of both hands and feet without wearing sensor equipment, while the gait data can be recorded by the signal collection module at the same time to avoid interference with the natural gait caused by wearing sensors;

[0018] (5) The tester starts collecting gait data on the treadmill, and conducts experiments at different set speeds. At each speed setting, the participant enters the motion state from the standing position according to their own walking / running habits. While starting the treadmill, the power switch of the signal acquisition system is turned on. Each set of gait motion collection experiments lasts for 1 minute, and the sampling frequency is 100 Hz.

[0019] Preferably, the lower limb walking mechanism analysis and gait phase division in step c: before the gait phase recognition algorithm is designed and applied, the gait phase in the gait data needs to be labeled offline first, and the data set used to train the classification recognition algorithm needs to be truth-processed so that the supervised machine learning method can map an input to an output, which usually requires a set of labeled data so that the model can continuously adjust the learning algorithm from learning; the threshold-based method is usually used to distinguish the "switching state" of the pressure sensor signal, by setting the corresponding signal threshold for each sensor unit separately to determine whether the sensor unit is in an "idle state" or a "working state".

[0020] Preferably, the step d is based on a multi-layer perceptron neural network algorithm to establish a gait phase recognition model and perform gait phase recognition. The input vector is an input layer composed of pressure sensor signals, the vector dimension is 16, the output vector is an output layer composed of gait phase categories, and the output vector size - the number of gait phase categories is 6; the training set recognition accuracy and gait phase recognition accuracy are defined to quantify the recognition results of the evaluation model.

[0021] Preferably, the establishment of the gait data set in step e: establishing a plantar pressure experimental data set from four experimental participants under eight groups of different gait movement speed settings; the gait data is recorded from the standing state, and the gait signal in the process of switching from standing to the specified gait movement speed is not a valid signal. Before carrying out gait phase classification, it is necessary to preliminarily organize the data set, consider the valid motion gait signals, and remove the signals during the preparatory action period - the period from the standing state to the specified walking speed, from the data set, leaving the gait data information in the stable motion state.

[0022] Preferably, the multi-motion mode gait phase recognition in step f: a cross-validation method is used to comprehensively train and verify the gait phase recognition model, and to verify the performance of the model; the gait data set of the tester is divided into two groups according to the ratio of "3+1". Each time the gait phase recognition model is trained, the gait data of three participants are used as training sets to train the gait phase recognition model, and the gait data of the fourth participant are used as a test set for the model to recognize the gait phase, and all data can be traversed in four cycles; during the training process, the gait data of the three participants are further divided into Training, Validation, and Testing data sets according to the ratio of "70:15:15", which are used for network training and control of the machine learning training process respectively; the gait phase recognition neural network is trained using the MATLAB toolbox nprtool.

[0023] Compared with the prior art, the present invention has the following obvious outstanding substantive features and significant advantages:

[0024] 1. The present invention is based on a plantar pressure information acquisition sensing system based on a pressure sensing unit FSR (Feedback Shift Register), and recruits multiple experimental participants to conduct multiple groups of gait motion experiments with different movement rates. The relevant gait experiment data are recorded and used to establish a data set;

[0025] 2. The present invention combines the analysis of the characteristics of plantar pressure data, and the gait experiment data is first divided into multiple gait cycles. According to the characteristics of the plantar pressure data during the process of the sole of the foot contacting the ground, the gait cycle is further subdivided into two gait sub-phases and marked with corresponding labels, and multiple different stages of the lower limbs and legs are defined in combination with different movement patterns;

[0026] 3. The present invention does not require manual intervention to select relevant signal features. By using the original input signals of 16 sensor channels to train the neural network model to obtain a six-dimensional output vector, the accuracy of the gait phase recognition model is improved as the number of nodes in the hidden layer of the neural network is optimized. The recognition of gait phases in various motion modes based on plantar pressure signals has certain effectiveness.

[0027] 4. The method of the present invention is simple, easy to implement, low in cost, and suitable for popularization and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flowchart of the gait phase recognition method of the present invention.

[0029] Figure 2 Figure 2. A flexible sensing insole designed based on the FSR pressure sensing unit.

[0030] Figure 3 Figure 1 shows the gait data collection process: (a) the front of the sensor when worn; (b) the back of the sensor when worn; (c) collecting gait data on a treadmill.

[0031] Figure 4 Schematic diagram of plantar pressure signal and phase in walking mode.

[0032] Figure 5 Schematic diagram of plantar pressure signal and phase in running mode.

[0033] Figure 6 Schematic diagram of the neural network model structure.

[0034] Figure 7 The statistical results of gait frequency at different walking speeds.

[0035] Figure 8 This is a statistical result diagram of the time proportion of the main gait phase in the gait cycle.

[0036] Fig. 9 This is the gait phase recognition result diagram in walking mode (5km / h).

[0037] Fig.10 This is the gait phase recognition result diagram in running mode (9km / h). DETAILED DESCRIPTION

[0038] The above scheme is further described below in conjunction with specific embodiments. The preferred embodiments of the present invention are described in detail as follows:

[0039] Embodiment 1:

[0040] In this embodiment, see Figure 1 , a gait phase recognition method based on neural network, the operation steps are as follows:

[0041] a. Build gait data collection equipment;

[0042] b. Gait data collection experiment;

[0043] c. Analysis of lower limb walking mechanism and gait phase division;

[0044] d. Based on the multi-layer perceptron neural network algorithm, a gait phase recognition model is established to perform gait phase recognition;

[0045] e. Establishment of gait dataset;

[0046] f. Multi-motion mode gait phase recognition.

[0047] This embodiment uses a neural network-based gait phase recognition algorithm to carry out gait phase recognition research in walking mode and running mode, and optimizes and adjusts the neural network recognition model by comparing the impact of different network structures on the recognition accuracy.

[0048] Embodiment 2:

[0049] This embodiment is basically the same as the first embodiment, except that:

[0050] In this embodiment, the gait data acquisition device system in step a is composed of a sensor module, a signal conditioning module, a single-chip control module, a data storage module and a power supply module; in order to facilitate carrying and installation at a fixed position of a human limb for signal acquisition, an integrated circuit board module is designed according to the package size of the single-chip microcomputer Arduino Uno, and all modules are integrated into a whole; after being processed by the signal conditioning module, the signal enters the single-chip microcomputer module through a multiplexer unit under the control of the single-chip microcomputer, and is processed by the analog-to-digital conversion module built into the Arduino microcontroller to become a digital signal, and finally stored in the data storage module SD card for subsequent data analysis and processing; a flexible sensing insole designed based on the FSR pressure sensing unit is used to monitor the plantar pressure signal, with 8 sensing units for one foot and a total of 16 sensing units for both feet, wherein the position of the sensing unit is arranged according to the physiological characteristics and load-bearing characteristics of the foot, and is divided according to the sole area; 3 sensing units are arranged in the metatarsal and heel areas, and 1 sensing unit is placed in the toe and arch areas.

[0051] The specific steps of the gait data collection experiment in step b are:

[0052] (1) Recruiting multiple testers who have never suffered from gait-related diseases to participate in the wearable test;

[0053] (2) Place the designed plantar pressure measurement system in the shoes of each test subject, and then use a nylon buckle belt to fix the signal acquisition module around the waist;

[0054] (3) To prevent the wire from swinging and affecting normal walking and causing unnatural movements, and to facilitate wearing and taking off the experimental equipment, a cable tie with a nylon buckle was used in the experiment to restrain the wire on the back of the lower limb above the knee joint;

[0055] (4) Place a power bank with sufficient power in the collector's pocket to power the collection system, ensuring that the tester can achieve the natural swinging state of both hands and feet without wearing sensor equipment, while the gait data can be recorded by the signal collection module at the same time to avoid interference with the natural gait caused by wearing sensors;

[0056] (5) The tester starts collecting gait data on the treadmill and conducts experiments at different set speeds. At each speed setting, the participant enters the motion state from a standing position according to their own walking / running habits. While starting the treadmill, the power switch of the signal acquisition system is turned on. Each set of gait motion acquisition experiments lasts for 1 minute, and the sampling frequency is 100 Hz.

[0057] The step c is about the analysis of the walking mechanism of the lower limbs and the division of the gait phases: before the design and application of the gait phase recognition algorithm, the gait phases in the gait data need to be labeled offline first, and the data set used to train the classification recognition algorithm needs to be processed with true values, so that the supervised machine learning method can map an input to an output. This usually requires a set of labeled data so that the model can continuously adjust the learning algorithm during learning; the threshold-based method is usually used to distinguish the "switching state" of the pressure sensor signal, and the corresponding signal threshold is set for each sensor unit separately to determine whether the sensor unit is in the "idle state" or "working state".

[0058] In step d, a gait phase recognition model is established based on a multi-layer perceptron neural network algorithm to perform gait phase recognition. The input vector is an input layer composed of pressure sensing signals, and the vector dimension is 16. The output vector is an output layer composed of gait phase categories, and the output vector size - the number of gait phase categories is 6. The training set recognition accuracy and gait phase recognition accuracy are defined to quantify the recognition results of the evaluation model.

[0059] The step e is to establish a gait data set: establish a plantar pressure experimental data set from four experimental participants under eight different gait movement speed settings; the gait data is recorded from the standing state, and the gait signal in the process of switching from standing to the specified gait movement speed is not a valid signal. Before carrying out gait phase classification, it is necessary to preliminarily organize the data set, and consider the valid movement gait signal to remove the signal during the preparatory action period - from the standing state to the specified walking speed from the data set, leaving the gait data information in the stable movement state.

[0060] The step f multi-motion mode gait phase recognition: a cross-validation method is used to comprehensively train and verify the gait phase recognition model, and to verify the performance of the model; the gait data set of the tester is divided into two groups according to the ratio of "3+1"; each time the gait phase recognition model is trained, the gait data of three participants are used as training sets to train the gait phase recognition model, and the gait data of the fourth participant is used as a test set for the model to recognize the gait phase, and all data can be traversed by four cycles; during the training process, the gait data of the three participants are further divided into Training, Validation, and Testing data sets according to the ratio of "70:15:15", which are used for network training and control of the machine learning training process respectively; the gait phase recognition neural network is trained using the MATLAB toolbox nprtool.

[0061] This embodiment is based on a sensing system for collecting plantar pressure information of an FSR pressure sensing unit, and recruits multiple experimental participants to conduct multiple groups of gait movement experiments at different movement rates. The relevant gait experiment data are recorded and used to establish a data set. In this embodiment, the gait experiment data is first divided into multiple gait cycles in combination with the analysis of the characteristics of the plantar pressure data. According to the characteristics of the plantar pressure data during the process of the sole of the foot contacting the ground, the gait cycle is further subdivided into two gait sub-phases, and marked with corresponding labels. Multiple different stages of the lower limbs and legs are defined in combination with different movement modes; this embodiment does not require manual intervention to select relevant signal features. By using the original input signals of 16 sensing channels, the neural network model is trained to obtain a six-dimensional output vector. The accuracy of the gait phase recognition model is improved as the number of nodes in the hidden layer of the neural network is optimized. The recognition of gait phases under various movement modes based on plantar pressure signals has certain effectiveness.

[0062] Embodiment three:

[0063] This embodiment is basically the same as the above embodiment, with the following features:

[0064] (1) In this embodiment, step a is a gait data collection system.

[0065] The gait data acquisition system consists of a sensor module, a signal conditioning module, a single-chip control module, a data storage module and a power supply module. The main components of the system are shown in the attached figure. Figure 2As shown, in order to facilitate carrying and installation at a fixed position on the human body for signal collection, the integrated circuit board module is designed according to the package size of the single-chip microcomputer Arduino Uno, and all modules are integrated into a whole. After being processed by the signal conditioning module, the signal enters the single-chip microcomputer module through the multiplexer unit under the control of the single-chip microcomputer, and is processed by the analog-to-digital conversion module built into the Arduino microcontroller to become a digital signal, which is finally stored in the data storage module SD card for subsequent data analysis and processing.

[0066] The quality of the sensor module design determines from the source whether the collected data can provide effective research information. Therefore, the sensor needs to be selected in combination with the characteristics of human body movement and stability. FSR sensors are widely used in plantar measurement applications due to their light weight, small size, thinness, high sensitivity and good flexibility. The present invention adopts a flexible sensing system designed based on FSR sensing units to collect plantar pressure signals, with 8 sensing units for one foot and 16 sensing units for both feet. The positions of the sensing units are arranged according to the physiological characteristics and load-bearing characteristics of the foot and divided according to the sole area. Since the metatarsal and heel areas play a major load-bearing role during walking, the pressure distribution information is relatively wide. Therefore, the number of layouts is relatively large, namely 3 sensing units, and 1 sensing unit is placed in each of the remaining toes and arch areas. The manufacturing technology of FSR sensors is already quite mature, the linearity consistency of the sensing units is good, and the sensing units do not require additional calibration processes.

[0067] (2) Step b is a gait data collection experiment.

[0068] During the movement of the lower limbs of the human body, different areas of the sole of the foot come into contact with the ground and show regular pressure changes. Measuring and analyzing the gait data of the lower limbs of the human body in different motion states is the most direct way to study the laws of human movement. First, recruit a number of testers who have never suffered from gait-related diseases to participate in the wearable test; secondly, place the designed plantar pressure measurement system in the shoes worn by each tester, and then use a nylon buckle belt to fix the signal acquisition module to the waist. In order to avoid the swing of the wires affecting normal walking and causing unnatural movements, and to facilitate the wearing and taking off of experimental equipment, a cable tie with a nylon buckle is used in the experiment to restrain the wires to the back of the lower limbs above the knee joint, and a power bank with sufficient power is placed in the collector's pocket to power the collection system. The wearable method is attached Figure 3As shown in the figure, the tester can achieve the natural swinging state of both hands and feet without wearing sensor equipment, and the gait data can be recorded by the signal acquisition module at the same time to avoid the interference of wearing sensors on the natural gait; next, the tester starts to collect gait data on the treadmill, referring to the attached Figure 3 , the experiment was carried out at different set speeds. At each speed setting, the participants entered the motion state from the standing position according to their respective walking / running habits. While starting the treadmill, the power switch of the signal acquisition system was turned on. Each set of gait motion acquisition experiments lasted for 1 minute, and the sampling frequency was 100 Hz.

[0069] (3) Step c is the analysis of the walking mechanism of the lower limbs and the division of gait phases.

[0070] Generally, a gait cycle is defined as follows: the time from the heel touchdown of the observed foot (e.g., the right foot) to the next heel touchdown of the observed foot, that is, the heel touchdown event means the end of the previous gait cycle and the beginning of the next gait cycle. There are two main gait phases for a single leg during walking: the standing phase (STP) and the swing phase (SWP). Since the sole of the foot does not contact the ground when the leg is in the swing phase and there is no plantar pressure, the plantar pressure signal changes mainly in the standing phase. The gait phase division based on the plantar pressure information can be more finely divided into multiple gait sub-phases for the standing phase.

[0071] The development of gait phase recognition needs to be based on the offline classification of gait data. Before the design and application of the gait phase recognition algorithm, the gait phases in the gait data need to be labeled offline first, and the data set used to train the classification and recognition algorithm needs to be processed with true values, so that the supervised machine learning method can map an input to an output. This usually requires a set of labeled data so that the model can continuously adjust the learning algorithm from learning.

[0072] The threshold-based method is usually used to distinguish the "on / off state" of the pressure sensor signal. By setting the corresponding signal threshold for each sensor unit separately, it is determined whether the sensor unit is in the "idle state" or "working state". The gait cycle of a single leg can be distinguished by the "on / off state" of the sensor unit, using the following formula:

[0073]

[0074] Where D threshold is the threshold of the FSR sensor unit in the gait data, S maxis the maximum value in the gait data, S min is the minimum value in the gait data, δ is an adjustable threshold coefficient, which is used to indicate the percentage of the threshold value between the peak values. Here, δ is 20%. The threshold data determines the state of the sensor unit and is used to establish a truth table for dividing the gait phases, as shown in the attached figure.

[0075] (4) Step d is a multi-layer perceptron neural network algorithm.

[0076] The present invention establishes a gait phase recognition model based on a multilayer perceptron neural network algorithm (Multilayer Perceptron, MLP) to perform gait phase recognition. The multilayer perceptron is a forward-structured artificial neural network that maps a set of input vectors to a set of output vectors. Its structure is shown in the accompanying figure. Its model is as follows:

[0077]

[0078] Among them, P Input It is the input layer composed of input vector (pressure sensor signal), the vector dimension is 16, G Output It is the output layer composed of output vectors (gait phase categories), and the output vector size (number of gait phase categories) is 6. According to the definition of MLP neural network, f(x) is expressed as:

[0079] h(x)ss(b (1) +W (1) p i ) (3)

[0080] f(x)=G(b (2) +W (2) h(x)) (4)

[0081] Here, h(x) is the hidden layer output vector, the function G(·) is softmax(), which is used for multi-category classification, and the activation function s(·) is tanh(·), which usually makes training faster. (1) , b (1) is the offset vector, the matrix is the weight matrix used to connect the input layer nodes and the hidden layer nodes. represents the weight matrix used to connect the hidden layer nodes and the output layer nodes. Before the gait recognition model training begins, the parameter β={W (1) , W (2) , b (1) , b (2)The present invention defines the correct rate of set (CRS) and the correct rate of phase (CRP) to quantify the recognition results of the evaluation model.

[0082]

[0083] Among them, N s is the total number of gait data sampling points in the training set or test set, is the number of correctly identified gait data sampling points in the training set or test set.

[0084]

[0085] Among them, N p is the number of sampling points in each gait phase, It is the number of sampling points that are accurately identified in the corresponding gait phase stage. CRS is used to describe the overall performance of the gait phase recognition model. The improvement of the gait phase model parameters to improve CRS can reflect the optimization process of the gait phase recognition model from the side.

[0086] (5) Step e is to establish a gait dataset.

[0087] A plantar pressure experimental data set was established from four experimental participants under eight different gait movement speed settings. The gait data was recorded from the standing state. The gait signal during the process of switching from standing to the specified gait movement speed is not a valid signal. Before conducting gait phase classification, the data set needs to be preliminarily sorted. Considering the valid motion gait signal, the signal during the preparatory action period (from the standing state to the specified walking speed) is removed from the data set, leaving the gait data information in the stable motion state.

[0088] (6) Step f is multi-motion mode gait phase recognition.

[0089] The present invention adopts a cross-validation method to comprehensively train and verify the gait phase recognition model, and verify the performance ability of the model. The gait data set of the tester is divided into two groups according to the ratio of "3+1". Each time the gait phase recognition model is trained, the gait data of three participants are used as training sets to train the gait phase recognition model, and the gait data of the fourth participant is used as a test set for the model to recognize the gait phase. All data can be traversed in four cycles. During the training process, the gait data of the three participants are further divided into Training, Validation, and Testing data sets according to the ratio of "70:15:15", which are used for network training and control of the machine learning training process respectively. The gait phase recognition neural network is trained using the toolbox nprtool (Neural Network Pattern Recognition app) of MATLAB (version number: R2020A).

[0090] Embodiment 4:

[0091] This embodiment is basically the same as the above embodiment, with the following features:

[0092] In this embodiment, a gait phase recognition method based on a neural network is provided, and the operation steps are as follows:

[0093] a. Gait data collection equipment

[0094] According to the invention content a, a gait data collection device is designed and constructed.

[0095] b. Gait data collection experiment

[0096] The age and gender of different people will more or less affect the duration of the gait phase. In order to ensure the feasibility of the proposed model and avoid other influencing factors, narrow the scope of recruitment, and 4 members from the laboratory participated in the gait data collection experiment. All members participating in the experiment did not suffer from gait-related diseases. These participants were all male, aged 24±2 years old, 175±3cm tall, with similar shoe sizes. The specific physical fitness data are shown in Tables 1 and 2.

[0097] Table 1. Average recognition accuracy of walking mode (5 km / h) and recognition accuracy of each gait phase

[0098]

[0099] Table 2. Average recognition accuracy of running mode (9 km / h) and recognition accuracy of each gait phase

[0100]

[0101] The gait motion collection experiment was conducted by setting the treadmill speed at 8 different speeds: 3km / h, 4km / h, 5km / h, 6km / h, 7km / h, 8km / h, 9km / h, and 10km / h. In the experiment, most participants were in walking mode at speeds below 7km / h or 8km / h, and in running mode at speeds above 8km / h.

[0102] c. Analysis of the walking mechanism of lower limbs

[0103] According to the lower limb walking mechanism described in content c of the present invention, the method of defining the gait phase with both feet as reference is different from the method of defining the gait phase of a single foot. During walking, due to the slow walking speed, in order to stably switch from the standing state of one leg to the standing state of the other leg, both feet touch the ground at the same time; during running, it is also switching from the standing state of one leg to the standing state of the other leg. Due to the fast movement speed, the limbs rely on inertia to present a "flying" posture, and both feet leave the ground at the same time. The division of gait sub-phases is shown in the attached figure. Figure 4 and attached Figure 5 As shown. Figure 4 Plantar pressure signal indicating walking pattern, Figure 4 Plantar pressure signal indicating running mode.

[0104] d. Gait dataset establishment

[0105] According to the process of inventive method e, the collected data were preliminarily sorted, and the resulting data set contained nearly 200,000 sets of sampling points (sampling frequency 100 Hz) from 3 km / h to 10 km / h for four participants, and each set of sampling points contained plantar pressure signals from 16 sensing channels (8 for the left foot and 8 for the right foot).

[0106] According to the gait cycle division method, the plantar pressure signal is intercepted at the starting point according to the marking action of both feet (the starting moment of the left heel touching the ground, which is also the end moment of the left foot swing phase). In the subsequent signal sequence, every time the left foot repeats the marking action, it is marked as a gait cycle division point. With the help of the UI interface of the Signal Labeler toolbox of MATLAB (version number: R2020A), the corresponding position in the plantar pressure signal curve is selected for marking. The gait data set of the four experimental participants contains 1052 gait cycles. The gait frequency statistics under different walking speed settings are shown in the attached figure. Figure 7 .

[0107] The results of visualizing the gait frequency statistics according to the main phases (stance phase and swing phase) are shown in the attached Figure 7For all the experimental subjects, when the walking speed did not exceed 7 km / h, the proportion of time in the standing phase was longer than that in the swing phase; when the walking speed was 7 km / h and 8 km / h, the average proportions of the swing phase and the standing phase were almost the same; when the walking speed exceeded 8 km / h, the proportion of time in the swing phase was longer than that in the standing phase.

[0108] e. Multi-motion mode gait phase recognition

[0109] Generally speaking, different numbers of nodes in the hidden layer of a neural network will have different effects on the recognition accuracy. In order to select the appropriate number of neural network nodes, neural network models with different numbers of nodes are used to evaluate the gait phase recognition of the above data sets. The recognition results of each neural network model can be referred to in the attached Fig. 9 and attached Fig.10 Attached Fig. 9 The data recorded are the average recognition accuracy of walking motion patterns and the recognition accuracy of each gait phase during the training of the gait phase recognition model; Fig.10 The records are the average recognition accuracy of the running movement pattern and the recognition accuracy of each gait phase. Fig. 9 and attached Fig.10 The gait phase recognition results obtained by analyzing a gait cycle selected from four experiments in walking mode and running mode are shown respectively. The results show that the gait phase recognition model can accurately recognize the gait phases of different participants in the walking motion mode (slow speed). Although there are more or less recognition lags in the switching stage of the gait phase, accurate recognition results are shown in the main stages of the gait phase. In the running motion mode (fast speed), for different experimental subjects, there are more or less misrecognitions and recognition fluctuations in the gait phase recognition when both feet are off the ground, but correct recognition can be completed in most gait data.

[0110] The above describes the embodiments of the present invention in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Various changes can be made according to the purpose of the invention of the present invention. Any changes, modifications, substitutions, combinations or simplifications made according to the spirit and principle of the technical solution of the present invention should be equivalent replacement methods. As long as they meet the purpose of the invention of the present invention and do not deviate from the technical principles and inventive concepts of the present invention, they belong to the protection scope of the present invention.

Claims

1. A gait phase recognition method based on neural network, characterized in that: The steps are as follows: a. Build gait data collection equipment; b. Gait data collection experiment; c. Analysis of lower limb walking mechanism and gait phase division; d. Based on the multi-layer perceptron neural network algorithm, a gait phase recognition model is established to perform gait phase recognition; e. Establishment of gait dataset; f. Multi-motion mode gait phase recognition; The gait data acquisition device system in step a is composed of a sensor module, a signal conditioning module, a single-chip control module, a data storage module and a power supply module; in order to facilitate carrying and installation at a fixed position of a human limb for signal acquisition, an integrated circuit board module is designed according to the package size of the single-chip microcomputer Arduino Uno, and all modules are integrated into a whole; after being processed by the signal conditioning module, the signal enters the single-chip microcomputer module through a multiplexer unit under the control of the single-chip microcomputer, and is processed by the analog-to-digital conversion module built into the Arduino microcontroller to become a digital signal, and finally stored in the data storage module SD card for subsequent data analysis and processing; A flexible sensing insole designed based on FSR sensing units is used to monitor the plantar pressure signal. There are 8 sensing units on one foot and 16 sensing units on both feet. The positions of the sensing units are arranged according to the physiological characteristics and load-bearing characteristics of the foot, and are divided according to the sole area; 3 sensing units are arranged in the metatarsal and heel areas, and 1 sensing unit is placed in the toe and arch areas respectively. The specific steps of the gait data collection experiment in step b are: (1) Recruit multiple testers who have never suffered from gait-related diseases to participate in the wearable test; (2) Place the designed plantar pressure measurement system in the shoes of each test subject, and then use a nylon buckle belt to fix the signal acquisition module around the waist; (3) To prevent the wires from swinging and affecting normal walking and causing unnatural movements, and to facilitate wearing and taking off the experimental equipment, a cable tie with a nylon buckle was used to restrain the wires to the back of the lower limbs above the knee joint. (4) Place a power bank with sufficient power in the collector's pocket to power the collection system, ensuring that the tester can achieve the natural swinging state of both hands and feet without wearing sensor equipment, while the gait data can be recorded by the signal acquisition module at the same time to avoid interference with the natural gait caused by wearing sensors; (5) The tester starts to collect gait data on the treadmill, and the experiment is carried out at different set speeds. At each speed setting, the participants enter the motion state from the standing position according to their own walking / running habits. When the treadmill is started, the power switch of the signal acquisition system is turned on. Each set of gait motion collection experiments lasts for 1 minute, and the sampling frequency is 100 Hz. The step c includes analysis of the walking mechanism of the lower limbs and division of gait phases; before the gait phase recognition algorithm is designed and applied, the gait phases in the gait data need to be labeled offline first, and the data set used to train the classification recognition algorithm needs to be processed with true values, so that the supervised machine learning method can map an input to an output, which usually requires a set of labeled data so that the model can continuously adjust the learning algorithm from learning; the threshold-based method is usually used to distinguish the "switching state" of the pressure sensor signal, and the corresponding signal threshold is set for each sensor unit separately to determine whether the sensor unit is in the "idle state" or "working state"; In the step d, a gait phase recognition model is established based on a multilayer perceptron neural network algorithm to perform gait phase recognition; based on a multilayer perceptron neural network algorithm, a gait phase recognition model is established to perform gait phase recognition; wherein, the input vector is an input layer composed of pressure sensing signals, the vector dimension is 16, the output vector is an output layer composed of gait phase categories, and the output vector size - the number of gait phase categories is 6; the training set recognition accuracy and the gait phase recognition accuracy are defined to quantify the recognition results of the evaluation model; The establishment of the gait data set in step e: establishing a plantar pressure experimental data set from four experimental participants under eight different gait movement speed settings; the gait data is recorded from the standing state, and the gait signal in the process of switching from standing to the specified gait movement speed is not a valid signal. Before carrying out gait phase classification, it is necessary to preliminarily organize the data set, and consider the valid movement gait signal to remove the signal during the preparatory action period - the period from the standing state to the specified walking speed from the data set, leaving the gait data information in the stable movement state; The step f includes multi-motion mode gait phase recognition; a cross-validation method is used to comprehensively train and verify the gait phase recognition model, and to verify the performance of the model; the gait data set of the tester is divided into two groups according to the ratio of "3+1"; each time the gait phase recognition model is trained, the gait data of three participants are used as the training set to train the gait phase recognition model, and the gait data of the fourth participant is used as the test set for the model to recognize the gait phase, and all the data can be traversed by four cycles; During the training process, the gait data of the three participants were further divided into Training, Validation, and Testing data sets in a ratio of "70:15:15", which were used for network training and control of the machine learning training process, respectively; the gait phase recognition neural network was trained using MATLAB's toolbox nprtool.

Images