A bionic intelligent powered walker based on multi-layer improved ELM classifier

Abstract

The present invention is applicable to the field of intelligent walking assistance technology, and provides a bionic intelligent power walker based on a multi-layer improved ELM classifier. The present invention collects data by installing IMU sensors on the soles of two feet, and performs phase recognition in combination with plantar pressure sensors. When the wearer's heels leave the ground, the phase recognition module is started, and the human motion intention prediction model performs pattern recognition on the motion feature vector. The model greatly improves the prediction accuracy of the wearer's motion intention by adding an input weight intervention layer and an output weight similarity prediction layer. The present invention not only ensures a high-accuracy prediction, but also adapts to the needs of actual application scenarios. By streamlining the number of sensors and optimizing the algorithm design, the equipment cost is effectively reduced, and the real-time performance and efficiency of the system are improved, so that the bionic intelligent power walker can better meet the needs of human motion intention recognition in daily life, and provide the wearer with a more convenient and efficient assisted walking experience.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent walking aids, and in particular relates to a bionic intelligent power walking aid based on a multi-layer improved ELM classifier. Background Art

[0002] The human motion intention recognition system is a high-level controller of the hierarchical control strategy of the intelligent portable lower limb rehabilitation robot, responsible for sensing the wearer's motion intention based on user, environment and prosthesis signals. The ideal high-level controller output should allow the device to smoothly switch between different motion modes without any conscious input from the wearer. Motion mode recognition combined with direct volitional control provides users with the ability to adjust the device in a specific motion state.

[0003] The steps for implementing the human motion intention recognition method applied to bionic intelligent power walkers include: first, accurately collect and measure human motion information such as joint angles, accelerations and angular velocities, and establish a human motion database containing different motion modes; preprocess, segment, extract features and reduce dimensions of sensor data to construct feature vectors; use the feature vector set to train the human motion intention recognition model and output the prediction function; embed the prediction function program into the walker development board, when the wearer's heel leaves the ground, the plantar pressure sensor value reaches the threshold, activate the classifier to recognize the human motion intention, and input the prediction result to the walker control system to switch the motion mode; drive the motor to provide the wearer with the power required for movement, reduce the energy consumption required for movement, and improve the wearer's sports experience. Although scientific research teams have conducted research on such walkers, there is still a lack of mature commercial products on the market. Research walkers have problems such as sensor redundancy (such as Inertial Measurement Unit, IMU), superposition of multi-source sensor networks such as pressure sensors15, encoders, and electromyography sensors, which increase the burden of wearing and equipment cost), complex prediction model structure (many parameters to be saved, which puts a burden on the control board memory), and low accuracy in human motion intention recognition. It is difficult to achieve real-time and effective prediction and powered walking assistance functions in various complex motion scenes in daily life. Existing deep learning algorithms for human motion intention recognition, such as recurrent neural networks, back propagation neural networks, convolutional neural networks, etc., also have problems such as many parameters to be saved and low recognition accuracy. For this reason, the present invention proposes a bionic intelligent powered walker based on a multi-layer improved ELM classifier. Summary of the invention

[0004] The purpose of the present invention is to provide a bionic intelligent power walker based on a multi-layer improved ELM classifier, aiming to solve the problems raised in the above background technology.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] A bionic intelligent powered walker based on a multi-layer improved ELM classifier, comprising a main foot system and an auxiliary foot system;

[0007] The main foot system includes:

[0008] Main Foot Development Board;

[0009] Signal acquisition module 1, composed of an IMU sensor 1 integrated on the main force foot development board, used to collect three-axis acceleration data at the main force foot surface;

[0010] The signal receiving module is composed of a Bluetooth chip 1 integrated on the main foot development board, and is used to receive the three-axis acceleration data at the auxiliary foot surface through the Bluetooth chip 1;

[0011] The signal storage module is composed of a Micro SD card module integrated on the main foot development board, and is used to store the motion data collected by the main foot and the auxiliary foot through the MicroSD card module;

[0012] The phase recognition module consists of a pressure sensor installed on the sole of the foot, which is used to determine which movement phase the wearer is currently in through the pressure sensor;

[0013] The feature integration module is used to integrate the data collected by the main foot system and the auxiliary foot system within the 250ms time window and extract the feature values, including the maximum value, minimum value, mean value and standard deviation, to form a motion feature vector, and transmit it to the intention recognition module;

[0014] The intention recognition module uses the human motion intention prediction model based on the multi-layer improved ELM classifier to recognize the human motion intention of the input motion feature vector and output the current motion mode to the power drive module;

[0015] A power drive module, the power drive module is used to receive the output of the intention recognition module and realize motion control according to the output; the power drive module is composed of an Elmo motor driver, a DC motor, a water drop wheel and a high-horsepower fishing line; the Elmo motor driver is used to supply power to the DC motor and control it, the output end of the DC motor is connected to the input end of the water drop wheel, and the water drop wheel is connected to the walker through a plurality of high-horsepower fishing lines; the DC motor drives the water drop wheel by forward or reverse rotation, and the water drop wheel realizes motion control by driving the high-horsepower fishing lines connected to the walker at multiple locations;

[0016] The main power module, which consists of a 24v lithium battery, is used to power the main foot development board, Elmo motor driver and pressure sensor;

[0017] Pressure sensor monitoring module, used to monitor gait status and provide auxiliary prediction conditions;

[0018] The auxiliary foot system comprises:

[0019] Assistive Foot Development Board;

[0020] Signal acquisition module 2, composed of IMU sensor 2 integrated on the auxiliary foot development board, is used to collect three-axis acceleration data at the instep of the auxiliary foot;

[0021] The signal transmission module is composed of a Bluetooth chip 2 integrated on the auxiliary foot development board, and is used to send the three-axis acceleration data of the auxiliary foot surface to the main foot system through the Bluetooth chip 2;

[0022] Auxiliary power supply module, consisting of a 5v lithium battery, is used to power the auxiliary foot development board;

[0023] The human motion intention prediction model includes a multi-layer improved extreme learning machine classifier (multi-layer improved ELM classifier), a KNN classifier, an output weight similarity prediction layer and a prediction output layer; the multi-layer improved ELM classifier includes an input weight intervention layer, an input layer, a hidden layer and an output layer, wherein the input weight intervention layer uses an LR classifier, and uses the output coefficient and intercept of the LR classifier, combined with the GWO-SMA algorithm, to obtain the optimal output coefficient and the optimal intercept as the input weight and the hidden layer node bias; the output weight similarity prediction layer uses an improved KNN classifier; the human motion intention prediction model identifies the human motion intention of the input motion feature vector through three prediction methods, prediction method one is KNN classifier prediction, prediction method two is multi-layer improved ELM classifier direct prediction, and prediction method three is to use the improved KNN classifier to compare the output weights of the multi-layer improved ELM classifier with the output weight matrix set obtained by multiple training using all training data and labels Perform similarity comparison and output the prediction result; use prediction method 1, prediction method 2 and prediction method 3 to predict the motion feature vector, and get prediction value 1, prediction value 2 and prediction value 3 respectively. The result of the prediction output layer is obtained by hard voting of prediction value 1, prediction value 2 and prediction value 3. If the voting results are unified (the minority obeys the majority to determine the result), the prediction result is directly output. If the voting results cannot be unified (the voting results are 1:1:1), the prediction value of the previous gait is combined for judgment.

[0024] Furthermore, the training process of the LR classifier is as follows:

[0025] The Xsens MVN Awinda human motion capture system was used to collect data from healthy subjects in five daily motion patterns, including horizontal walking, uphill walking, downhill walking, walking upstairs, and walking downstairs.

[0026] The collected data was preprocessed, including denoising, removing abnormal data with missing items, and adding classification labels to normal data: 1 for horizontal walking, 2 for uphill walking, 3 for downhill walking, 4 for upstairs walking, and 5 for downstairs walking; a fixed 250ms time window was used to intercept the sensor data after the heel left the ground in each gait cycle, and four time domain feature values ​​of each dimension of data were extracted, including maximum, minimum, mean and standard deviation:

[0027] Maximum value: ;

[0028] Minimum: ;

[0029] Mean: ;

[0030] Standard Deviation: ;

[0031] in: For sampling samples; is the number of samples sampled in the 250ms time window;

[0032] Use the preprocessed subject data to train the LR classifier and save the output coefficients of the LR classifier and the intercept .

[0033] Furthermore, the optimal output coefficient and the optimal intercept The calculation formula is as follows:

[0034] ;

[0035] ;

[0036] in: and is the weight coefficient, and ; The optimal output coefficient obtained by optimizing the GWO-SMA algorithm; The optimal intercept obtained by optimizing the GWO-SMA algorithm; and are the output coefficient and intercept of the LR classifier respectively.

[0037] Furthermore, when the pressure sensor monitoring module detects that the foot is off the ground for more than 2.5 seconds, it determines that the current gait recognition has failed and clears all prediction results of the gait; if the gaits in the previous movement stages have been correctly recognized, the prediction value of the last successfully recognized gait will be used as a preset condition and participate in the auxiliary prediction of the current gait.

[0038] Furthermore, in the GWO-SMA algorithm, the slime mold weight The update formula is:

[0039] ;

[0040] ;

[0041] in: and represents random parameters; Indicates the individual number of slime mold; represents the number of slime molds; and Represent the slime mold weight and The fitness value of each slime mold individual; and Respectively represent the best fitness value and the worst fitness value in the current iteration; and Respectively represent the second best fitness value and the second best fitness value obtained in the current iteration; is the average of the three best fitness values ​​before the current iteration; is the sorting function, defined as follows:

[0042] ;

[0043] in: Express Sort the sequence;

[0044] Update individual location , divided into three situations:

[0045] ;

[0046] ;

[0047] in, is the ratio of randomly distributed slime mold individuals to the total; random parameter ; is a control parameter used to control the update of the slime mold position; is the upper bound of the search range; is the lower bound of the search range; for Random numbers within for Random numbers within is the average position of the three slime mold individuals with the highest fitness; is the slime mold weight; For the The top three slime mold individuals with the highest fitness at the iteration; The value is A slime mold individual is randomly selected in the iteration; the optimal fitness value in all iterations is defined as .

[0048] Furthermore, the hard voting mathematical formula used in the prediction output layer is expressed as:

[0049] ;

[0050] in: is the final prediction result; is the index of the category, from 1 to the total number of categories ; is the number of models participating in the vote; It is i Model pair samples x The prediction results; is the indicator function; It is The label value of the class; when The value is 1 when , otherwise it is 0, and the category with the most votes is selected as the final prediction result;

[0051] If the previous step is successfully identified, the prediction results of the three prediction methods are the same as the results of the previous step as the prediction results of the prediction model; if the prediction results of the three prediction methods are different from the results successfully identified in the previous step, the weight distribution of the three prediction methods is used for judgment, and the weight distribution relationship is shown in the following formula:

[0052] ;

[0053] in: Improve the weights of ELM classifier for multiple layers; is the weight of the KNN classifier; The weights of the improved KNN classifier for the output weight similarity prediction layer.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] The present invention proposes an innovative lightweight solution, which collects data by installing IMU sensors only on the soles of the feet, and combines them with plantar pressure sensors for phase recognition. When the wearer's heel leaves the ground, the human body movement intention prediction model performs pattern recognition on the motion feature vector to achieve accurate prediction of human body movement intention. The model improves the prediction accuracy of the algorithm by adding an input weight intervention layer and an output weight similarity prediction layer, thereby greatly improving the prediction accuracy of the wearer's movement intention. This lightweight solution not only ensures high-accuracy prediction, but also adapts to the needs of actual application scenarios. By streamlining the number of sensors and optimizing the algorithm design, the equipment cost is effectively reduced, and the real-time performance and efficiency of the system are improved, so that the bionic intelligent power walker can better meet the needs of human body movement intention recognition in daily life, and provide the wearer with a more convenient and efficient assisted walking experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a schematic diagram of the structure of the bionic intelligent powered walker.

[0057] Figure 2 This is the framework diagram of the main foot system.

[0058] Figure 3 This is a framework diagram of the auxiliary foot system.

[0059] Figure 4 Prediction flow chart of the human motion intention prediction model.

[0060] Figure 5 Confusion matrix for the prediction performance test of the human motion intention prediction model.

[0061] Figure 6 ROC curve for the prediction performance test of the human movement intention prediction model.

[0062] Figure 7 The prediction time and prediction accuracy results of the human motion intention prediction model.

[0063] In the figure: main foot development board 11, IMU sensor 112, Bluetooth chip 113, Micro SD card module 14, pressure sensor 15, Elmo motor driver 16, DC motor 17, water drop wheel 18, high-horsepower fishing line 19, 24v lithium battery 110, auxiliary foot development board 21, IMU sensor 2 22, Bluetooth chip 2 23, 5v lithium battery 24. DETAILED DESCRIPTION

[0064] In order to have a clearer understanding of the technical features, purposes and beneficial effects of the present invention, the technical solution of the present invention is now described in detail below, but it should not be construed as limiting the applicable scope of the present invention.

[0065] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0066] like Figure 1 As shown, one embodiment of the present invention provides a bionic intelligent power walker based on a multi-layer improved ELM classifier, including a main foot system and an auxiliary foot system;

[0067] The main foot system includes:

[0068] Main foot development board 11;

[0069] The signal acquisition module 1 is composed of an IMU sensor 12 integrated on the main foot development board 11, and uses the IMU sensor 12 to collect the three-axis acceleration data of the main foot surface, with a range of -16g~16g;

[0070] The signal receiving module is composed of a Bluetooth chip 13 integrated on the main foot development board 11, and receives the three-axis acceleration data of the auxiliary foot surface through the Bluetooth chip 13, with a range of -16g~16g;

[0071] The signal storage module is composed of a Micro SD card module 14 integrated on the main foot development board 11, and the motion data collected by the main foot and the auxiliary foot are stored through the MicroSD card module 14;

[0072] The phase recognition module is composed of a pressure sensor 15 installed on the sole of the foot, and determines which movement phase (stance phase or swing phase) the wearer is currently in through the pressure sensor 15;

[0073] The feature integration module integrates the data collected by the main foot system and the auxiliary foot system within the 250ms time window and extracts the feature values ​​(including maximum value, minimum value, mean value and standard deviation) to form a motion feature vector, which is then transmitted to the intention recognition module;

[0074] The intention recognition module uses the human motion intention prediction model based on the multi-layer improved ELM classifier to identify the human motion intention of the input motion feature vector and output the current motion mode to the power drive module;

[0075] The power drive module is used to receive the output of the intention recognition module and realize motion control according to the output; the power drive module is composed of an Elmo motor driver 16, a DC motor 17, a water drop wheel 18 and a high-powered fishing line 19. The Elmo motor driver 16 supplies power to the DC motor 17 and controls it. The output end of the DC motor 17 is connected to the input end of the water drop wheel 18, and the water drop wheel 18 is connected to the walker through a plurality of high-powered fishing lines 19; the water drop wheel 18 is driven by the forward or reverse rotation of the DC motor 17, thereby driving the high-powered fishing lines 19 connected to the walker at multiple locations, so as to realize precise motion control;

[0076] The main power supply module, which consists of a 24V lithium battery 110, is used to power the main foot development board 11, the Elmo motor driver 16 and the pressure sensor 15;

[0077] Pressure sensor monitoring module, used to monitor gait status and provide auxiliary prediction conditions.

[0078] The auxiliary foot system comprises:

[0079] Assisted Foot Development Board 21;

[0080] The signal acquisition module 2 is composed of an IMU sensor 22 integrated on the auxiliary foot development board 21, and is used to collect three-axis acceleration data at the auxiliary foot surface, with a range of -16g~16g;

[0081] The signal sending module is composed of a Bluetooth chip 23 integrated on the auxiliary foot development board 21, and sends the three-axis acceleration data (range: -16g~16g) of the auxiliary foot surface to the main foot system through the Bluetooth chip 23;

[0082] The auxiliary power supply module, composed of a 5V lithium battery 24, is used to power the auxiliary foot development board 21.

[0083] In the embodiment of the present invention, the electronic framework diagram of the main force foot system is as follows: Figure 2 As shown, the dotted line represents the power supply line, and the solid line represents the signal transmission line. A 24v lithium battery 110 supplies power to the main foot development board 11, the Elmo motor driver 16 and the plantar pressure sensor 15. The Elmo motor driver 16 supplies power to the DC motor 17 and controls it. The control chip of the main foot development board 11 is STM32F429, which communicates with the Elmo motor driver 16 through the CAN bus and communicates with the pressure sensor 15 through RS485. The main foot development board 11 integrates an HC42 Bluetooth chip 13 to realize Bluetooth data transmission between boards; the main foot development board 11 integrates a Micro SD card module 14 to realize data storage. The electronic framework diagram of the auxiliary foot system is shown in Figure 3As shown, it is powered by a 5V lithium battery 24. The auxiliary foot development board 21 adopts an Arduino development board, and is externally connected to a six-axis IMU sensor 22 and a HC42 Bluetooth chip 23 to realize signal collection and signal transmission functions.

[0084] During exercise, when the wearer's heel leaves the ground (the value of the pressure sensor 15 is less than 20N), the phase recognition module is started and the swing period bending stage is entered; then the 250ms sensor data collected by the main foot system and the auxiliary foot system are integrated and the characteristic values ​​(including maximum value, minimum value, mean value and standard deviation) are extracted and transmitted to the intention recognition module, and the human motion intention prediction model is used for prediction; according to the human motion intention prediction results, the wearer's current motion mode is accurately identified, and the corresponding torque control parameters are selected according to different motion modes, and efficient communication is achieved through the CAN bus, and the forward and reverse rotation of the DC motor 17 is accurately controlled, thereby driving the water drop wheel 18, driving the high-horsepower fishing line 19 connected to multiple points of the walker, and realizing precise motion control.

[0085] The human motion intention prediction model based on the multi-layer improved ELM classifier is described in detail below;

[0086] The schematic diagram of the human motion intention prediction model is as follows: Figure 4 As shown in FIG. 1 , it includes a multilayer improved ELM classifier (Multilayer Extreme Learning Machine, MELM), a KNN classifier, an output weight similarity prediction layer and a prediction output layer. The multilayer improved ELM classifier includes an input weight intervention layer, an input layer, a hidden layer and an output layer. The input weight intervention layer uses a logistic regression analysis (LR) classifier, uses the output coefficient and intercept of the LR classifier, and combines the improved slime mold algorithm (SMA) with the Grey Wolf Optimization (GreyWolf Optimization, GWO) algorithm (i.e., GWO-SMA algorithm) is used to obtain the optimal output coefficient and the optimal intercept as the weight from the input layer to the hidden layer (i.e., input weight) and the hidden layer node bias. This strategy aims to effectively reduce the number of hidden layer nodes in the ELM algorithm and improve model efficiency and performance. The output weight similarity prediction layer includes an improved KNN classifier, which realizes the prediction function by performing similarity comparison with multiple sets of output weight matrices obtained by pre-experimental training. The human motion intention prediction model recognizes the human motion intention of the input motion feature vector through three prediction methods. Prediction method one is KNN classifier prediction, prediction method two is direct prediction by multi-layer improved ELM classifier, and prediction method three is to use the improved KNN classifier to compare the output weights of the multi-layer improved ELM classifier with the output weight matrix set obtained by multiple training using all training data and labels. Perform similarity comparison and output prediction results; use prediction method 1, prediction method 2 and prediction method 3 to predict the motion feature vector, and obtain prediction value 1, prediction value 2 and prediction value 3 respectively. The result of the prediction output layer is obtained by hard voting of these three prediction values. The present invention improves the prediction accuracy of the algorithm by adding an input weight intervention layer and an output weight similarity prediction layer, thereby greatly improving the prediction accuracy of the wearer's motion intention.

[0087] 1. In the data collection phase, the Xsens MVN Awinda human motion capture system was used to collect IMU sensor triaxial acceleration data from 17 locations (including head, shoulder, upper arm, lower arm, thigh, calf and foot) of 10 healthy subjects of different ages in five daily motion modes, and a total of 51 channels of sensor data motion data were obtained with a sampling frequency of 60Hz. In each motion mode, each subject collected 600 steps of data.

[0088] 2. Preprocess the collected data, including denoising, removing abnormal data with missing items, and adding classification labels to normal data (1 for horizontal walking, 2 for uphill walking, 3 for downhill walking, 4 for upstairs walking, and 5 for downstairs walking). Use a fixed 250ms time window to intercept the sensor data after the heel leaves the ground in each gait cycle, and extract four time domain feature values ​​of each dimension of data, including maximum, minimum, mean and standard deviation:

[0089] Maximum value: ;

[0090] Minimum: ;

[0091] Mean: ;

[0092] Standard Deviation: ;

[0093] in: For sampling samples; is the number of samples sampled in the 250ms time window;

[0094] 3. Use the preprocessed subject data to train the LR classifier and save the output coefficients of the LR classifier and the intercept Then the GWO-SMA algorithm is used to optimize the input weights of the multi-layer improved ELM classifier. Another component of and the hidden layer node bias Another component of , output coefficient With intercept as well as and Combined, the optimal output coefficient and the optimal intercept are obtained as input weights and hidden layer node biases. The formula is as follows:

[0095] Formula 1: ;

[0096] Formula 2: ;

[0097] in: and is the weight coefficient, and ; The optimal output coefficient obtained by optimizing the GWO-SMA algorithm; It is the optimal intercept obtained by optimizing the GWO-SMA algorithm.

[0098] 4. By comparing the classification performance under different feature combinations and following the principle of minimizing the number of sensors, it is determined that the bionic intelligent power walker is constructed using IMU sensors placed on both sides of the foot. The bionic intelligent power walker consists of two core components: one is the IMU sensor 12 integrated on the main foot development board 11, and the other is the IMU sensor 22 located on the auxiliary foot development board 21. These sensors are responsible for collecting the three-axis acceleration data of the human feet during exercise.

[0099] When the pressure sensor 15 detects that the heel is off the ground (the value of the pressure sensor 15 is less than 20N), the system will start a 250ms time window to segment the experimental data. In this time window, the maximum value, minimum value, average value and standard deviation are extracted from the data obtained from each channel, and these parameters are combined into a motion feature vector. Finally, the motion feature vector will be input into the human motion intention prediction model for comprehensive prediction to obtain prediction value 1, prediction value 2 and prediction value 3. The result of the prediction output layer is obtained by hard voting of these three prediction values. If the voting results are unified, the prediction results are directly output. If the voting results cannot be unified, the prediction value of the previous gait is combined for judgment.

[0100] When the pressure sensor monitoring module detects that the foot is off the ground for more than 2.5 seconds, it determines that the current gait recognition has failed and clears all the prediction results of the gait to ensure that these data will not be used as a prerequisite for subsequent predictions. If the gaits in the previous movement stages have been correctly recognized, the prediction value of the last successfully recognized gait will be used as a preset condition and participate in the auxiliary prediction of the current gait to help solve the problem of inconsistent hard voting results. This improvement is intended to improve the accuracy and reliability of gait recognition. According to the prediction results, the power drive module performs corresponding motion control.

[0101] 1) Improved SMA algorithm To address the problem of lack of randomness in the later stage of the original algorithm, the GWO-SMA algorithm was proposed by combining the advantages of the Grey Wolf Optimization Algorithm, such as strong convergence performance, few parameters, and easy implementation. The update formula is:

[0102] Formula 3: ;

[0103] Formula 4: ;

[0104] in: and represents random parameters; Indicates the individual number of slime mold; represents the number of slime molds; and Represent the slime mold weight and The fitness value of each slime mold individual; and Respectively represent the best fitness value and the worst fitness value in the current iteration; and Respectively represent the second best fitness value and the second best fitness value obtained in the current iteration; is the average of the three best fitness values ​​before the current iteration; is the sorting function, defined as follows:

[0105] Formula 5: ;

[0106] in: Express Sort the sequence;

[0107] Update individual location , divided into three situations:

[0108] Formula 6: ;

[0109] Formula 7: ;

[0110] in, is the ratio of randomly distributed slime mold individuals to the total; the present invention sets ; Random parameters ; is a control parameter used to control the update of the slime mold position; is the upper bound of the search range; is the lower bound of the search range; for Random numbers within for Random numbers within is the average position of the three slime mold individuals with the highest fitness; is the slime mold weight; For the The top three slime mold individuals with the highest fitness at the iteration; The value is A slime mold individual is randomly selected in the iteration; the optimal fitness value in all iterations is defined as .

[0111] 2) The input of the improved KNN classifier is the output weight matrix of the multi-layer improved ELM classifier , and the output weight matrix set obtained by multiple trainings using all training data and labels Calculate the distance and assign the class label with the smallest mean distance to the sample to be tested. In addition, the improved KNN classifier combines the K-Means clustering algorithm as a preprocessing step. First, use the K-Means algorithm to determine an initial cluster center in each category, thereby forming a set of initial centers of each category. This method can optimize the performance of the KNN algorithm and improve the accuracy and efficiency of classification. Then cluster the samples in each class. The triangular inequality is used to improve the classification decision rule of the KNN algorithm: first, the distance between the unlabeled sample and the initial center of each class is calculated and sorted from the smallest to the largest, and the first class with the smallest distance is selected. the initial center of the class; then In the clusters, the secondary clustering points closer to the sample than the initial center are selected through the triangle inequality; finally, in the selected In each class, calculate the distance from the selected secondary cluster point to the unlabeled sample, and select the first The class label with the smallest distance mean is assigned to the unlabeled sample as the class label.

[0112] 3) The mathematical formula for hard voting used in the prediction output layer is expressed as:

[0113] Formula 8: ;

[0114] in: is the final prediction result; is the index of the category, from 1 to the total number of categories ; is the number of models participating in the voting; in the present invention, it is 3; It is i Model pair samples x The prediction results; is the indicator function; It is The label value of the class; when The value is 1 when , otherwise it is 0, and the category with the most votes is selected as the final prediction result.

[0115] Due to the current gait hard voting mechanism, the prediction results of the three classifiers are often inconsistent, resulting in the inability to effectively perform gait recognition. If the previous step is successfully recognized, the prediction results of the three prediction methods are the same as the results of the previous step as the prediction results of the prediction model; if the prediction results of the three prediction methods are different from the results of the previous step, the weight distribution of the three prediction methods is used for judgment. The weight distribution relationship is shown in Formula 9:

[0116] ;

[0117] in: Improve the weights of ELM classifier for multiple layers; is the weight of the KNN classifier; The weights of the improved KNN classifier for the output weight similarity prediction layer.

[0118] Embodiment 1, prediction performance test of human body movement intention prediction model;

[0119] The confusion matrix can clearly show the relationship between the actual value and the predicted value, and can be used to evaluate the performance of the human motion intention prediction model. The values ​​on the diagonal indicate the accuracy of the human motion intention prediction model in successfully predicting each category. The full name of the ROC curve is the Receiver Operating Characteristic (ROC) curve. The curve is drawn with the true positive rate as the vertical coordinate and the false positive rate as the horizontal coordinate. The closer the curve is to the upper left, the better the performance of the human motion intention prediction model. The area under the curve and the coordinate axis (Area Under the ROC Curve, AUC) is also a standard for evaluating the quality of the human motion intention prediction model. The closer the value is to 1, the more perfect the performance of the human motion intention prediction model.

[0120] The human motion intention prediction model based on the multi-layer improved ELM classifier proposed in the present invention is used to predict human motion intention, and the average classification accuracy can reach 98.28%. Figure 5 As shown in the figure, the confusion matrix shows that movement mode 0 (uphill) is easily confused with movement mode 1 (downhill) and movement mode 2 (horizontal walking). Although these misclassifications may seem potentially dangerous, they are actually harmless and the subjects will not notice them in these cases. Movement mode 3 (upstairs) and movement mode 4 (downstairs) are also slightly confused with movement mode 2, which still needs to be improved in the future. Figure 6As shown in the figure, the area enclosed by the average ROC curve of the five motion modes and the coordinate axis is 0.9891, which is very close to 1, indicating that the human motion intention prediction model based on the multi-layer improved ELM classifier proposed in the present invention has good classification performance for the five motion modes. The lowest AUC value of the proposed model for classifying the five motion modes appears in motion mode 0, which is only 0.9785, indicating that the effective classification between walking on flat ground and walking on a slope with a smaller slope is still a difficult problem to solve.

[0121] The proposed human motion intention prediction model has a prediction time of 1ms and a prediction accuracy of 98.28% for 6000 test data sets when running Python code on a 12th Gen Intel(R) Core(TM) i5-12500 3.00 GHz processor. Figure 7 ), while achieving accurate prediction, it also ensures real-time performance.

[0122] The above are only preferred embodiments of the present invention. It should be pointed out that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention. These should also be regarded as the protection scope of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A bionic intelligent powered walker based on a multi-layer improved ELM classifier, characterized in that: Including main foot system and auxiliary foot system; The main foot system includes: Main Foot Development Board; Signal acquisition module 1, composed of an IMU sensor 1 integrated on the main force foot development board, used to collect three-axis acceleration data at the main force foot surface; The signal receiving module is composed of a Bluetooth chip 1 integrated on the main foot development board, and is used to receive the three-axis acceleration data at the auxiliary foot surface through the Bluetooth chip 1; The signal storage module is composed of a Micro SD card module integrated on the main foot development board, and is used to store the motion data collected by the main foot and the auxiliary foot through the Micro SD card module; The phase recognition module consists of a pressure sensor installed on the sole of the foot, which is used to determine which movement phase the wearer is currently in through the pressure sensor; The feature integration module is used to integrate the data collected by the main foot system and the auxiliary foot system within the 250ms time window and extract the feature values, including the maximum value, minimum value, mean value and standard deviation, to form a motion feature vector, and transmit it to the intention recognition module; The intention recognition module uses the human motion intention prediction model based on the multi-layer improved ELM classifier to recognize the human motion intention of the input motion feature vector and output the current motion mode to the power drive module; A power drive module, the power drive module is used to receive the output of the intention recognition module and realize motion control according to the output; the power drive module is composed of an Elmo motor driver, a DC motor, a water drop wheel and a high-horsepower fishing line; the Elmo motor driver is used to supply power to the DC motor and control it, the output end of the DC motor is connected to the input end of the water drop wheel, and the water drop wheel is connected to the walker through a plurality of high-horsepower fishing lines; the DC motor drives the water drop wheel by forward or reverse rotation, and the water drop wheel realizes motion control by driving the high-horsepower fishing lines connected to the walker at multiple locations; The main power module, which consists of a 24v lithium battery, is used to power the main foot development board, Elmo motor driver and pressure sensor; Pressure sensor monitoring module, used to monitor gait status and provide auxiliary prediction conditions; The auxiliary foot system comprises: Assistive Foot Development Board; Signal acquisition module 2, composed of IMU sensor 2 integrated on the auxiliary foot development board, is used to collect three-axis acceleration data at the auxiliary foot surface; The signal sending module is composed of a Bluetooth chip 2 integrated on the auxiliary foot development board, and is used to send the three-axis acceleration data of the auxiliary foot surface to the main foot system through the Bluetooth chip 2; Auxiliary power supply module, consisting of a 5v lithium battery, is used to power the auxiliary foot development board; The human motion intention prediction model includes a multi-layer improved ELM classifier, a KNN classifier, an output weight similarity prediction layer and a prediction output layer; the multi-layer improved ELM classifier includes an input weight intervention layer, an input layer, a hidden layer and an output layer, wherein the input weight intervention layer uses an LR classifier, and uses the output coefficient and intercept of the LR classifier, combined with the GWO-SMA algorithm, to obtain the optimal output coefficient and the optimal intercept as the input weight and the hidden layer node bias; the output weight similarity prediction layer uses an improved KNN classifier; the human motion intention prediction model identifies the human motion intention of the input motion feature vector through three prediction methods, prediction method one is KNN classifier prediction, prediction method two is multi-layer improved ELM classifier direct prediction, and prediction method three is to use the improved KNN classifier to compare the output weights of the multi-layer improved ELM classifier with the output weight matrix set obtained by multiple training using all training data and labels Perform similarity comparison and output the prediction result; use prediction method 1, prediction method 2 and prediction method 3 to predict the motion feature vector, and get prediction value 1, prediction value 2 and prediction value 3 respectively. The result of the prediction output layer is obtained by hard voting of prediction value 1, prediction value 2 and prediction value 3. If the minority obeys the majority and can judge the result, the prediction result is directly output. If the voting result is 1:1:1, the prediction value of the previous gait is combined for judgment.

2. The bionic intelligent powered walker based on the multi-layer improved ELM classifier according to claim 1, characterized in that: The training process of the LR classifier is as follows: The Xsens MVN Awinda human motion capture system was used to collect data from healthy subjects in five daily motion patterns, including horizontal walking, uphill walking, downhill walking, walking upstairs, and walking downstairs. The collected data was preprocessed, including denoising, removing abnormal data with missing items, and adding classification labels to normal data: 1 for horizontal walking, 2 for uphill walking, 3 for downhill walking, 4 for upstairs walking, and 5 for downstairs walking; a fixed 250ms time window was used to intercept the sensor data after the heel left the ground in each gait cycle, and four time domain feature values ​​of each dimension of data were extracted, including maximum, minimum, mean and standard deviation: Maximum value: ; Minimum: ; Mean: ; Standard Deviation: ; in: For sampling samples; is the number of samples sampled in the 250ms time window; Use the preprocessed subject data to train the LR classifier and save the output coefficients of the LR classifier and the intercept .

3. The bionic intelligent powered walker based on the multi-layer improved ELM classifier according to claim 2, characterized in that: The optimal output coefficient and the optimal intercept The calculation formula is as follows: ； ； in: and is the weight coefficient, and ; The optimal output coefficient obtained by optimizing the GWO-SMA algorithm; The optimal intercept obtained by optimizing the GWO-SMA algorithm; and are the output coefficient and intercept of the LR classifier respectively.

4. The bionic intelligent powered walker based on the multi-layer improved ELM classifier according to claim 1, characterized in that: When the pressure sensor monitoring module detects that the foot is off the ground for more than 2.5 seconds, it determines that the current gait recognition has failed and clears all prediction results of the gait; if the gaits in the previous movement stages have been correctly recognized, the prediction value of the last successfully recognized gait will be used as a preset condition and participate in the auxiliary prediction of the current gait.

5. The bionic intelligent powered walker based on the multi-layer improved ELM classifier according to claim 1, characterized in that: In the GWO-SMA algorithm, the slime mold weight The update formula is: ； ； in: and represents random parameters; Indicates the individual number of slime mold; represents the number of slime molds; and Represent the slime mold weight and The fitness value of each slime mold individual; and Respectively represent the best fitness value and the worst fitness value in the current iteration; and Respectively represent the second best fitness value and the second best fitness value obtained in the current iteration; is the average of the three best fitness values ​​before the current iteration; is the sorting function, defined as follows: ； in: Express Sort the sequence; Update individual location , divided into three situations: ； ； in, is the ratio of randomly distributed slime mold individuals to the total; random parameter ; is a control parameter used to control the update of the slime mold position; is the upper bound of the search range; is the lower bound of the search range; for Random numbers within for Random numbers within is the average position of the three slime mold individuals with the highest fitness; is the slime mold weight; For the The top three slime mold individuals with the highest fitness at the iteration; The value is A slime mold individual is randomly selected in the iteration; the optimal fitness value in all iterations is defined as .

6. The bionic intelligent powered walker based on the multi-layer improved ELM classifier according to claim 1, characterized in that: The hard voting mathematical formula used in the prediction output layer is expressed as: ； in: is the final prediction result; is the index of the category, from 1 to the total number of categories ; is the number of models participating in the vote; It is i Model pair samples x The prediction results; is the indicator function; It is The label value of the class; when The value is 1 when , otherwise it is 0, and the category with the most votes is selected as the final prediction result; If the previous step is successfully identified, the prediction results of the three prediction methods are the same as the results of the previous step as the prediction results of the prediction model; if the prediction results of the three prediction methods are different from the results successfully identified in the previous step, the weight distribution of the three prediction methods is used for judgment, and the weight distribution relationship is shown in the following formula: ； in: Improve the weights of ELM classifier for multiple layers; is the weight of the KNN classifier; The weights of the improved KNN classifier for the output weight similarity prediction layer.

Images