Machine learning-based assistive device service life prediction method and system

Abstract

The application discloses a kind of based on machine learning's auxiliary tool service life prediction method and system, including the use data of collection preset walking aid and material performance detection data, the use data and the material performance detection data are preprocessed;According to the use data and the material performance detection data, walking aid finite element model is established, the node bearing threshold data of auxiliary tool is acquired, mapping model based on pressure material characteristics is established, the node time sequence stress of auxiliary tool is acquired, when the node time sequence stress is greater than the node bearing threshold data, then according to the amount of frequency and excess value, calculate first cumulative damage;Conversely, then the node time sequence stress is input machine learning model and obtains use fatigue data, according to the use fatigue data, calculate second cumulative damage;The use fatigue data includes fatigue ratio and fatigue degree;According to the first cumulative damage and the second cumulative damage, construct auxiliary tool service life prediction model, output result.

Description

TECHNICAL FIELD

[0001] The application relates to the field of service life prediction, in particular to a walking aid service life prediction method and system based on machine learning. BACKGROUND

[0002] The walking aid is a key equipment for the daily travel of the action obstacle population, and the use safety and service life thereof are directly related to the health of the user. With the aggravation of population aging and the growth of rehabilitation demand, the amount of the walking aid is increasing rapidly, and accurate prediction of the service life has become a core demand for guaranteeing use safety and optimizing maintenance plan. Traditional prediction is mostly dependent on material theory or simple experience model, and it is difficult to cover the multi-dimensional dynamic influencing factors such as equipment sensor data, user behavior and environmental changes in actual use.

[0003] The existing method has single data dimension, and mostly only focuses on material performance (such as elastic modulus and yield strength), ignores key influencing factors such as maintenance records, gait cycle and ground environment; does not distinguish overload and regular fatigue damage, and cannot accurately calculate the cumulative damage under different stress scenarios; the model lacks physical constraints and time sequence feature capturing ability, has large prediction error, and is difficult to adapt to complex human-computer interaction scenarios.

[0004] The present application realizes double breakthroughs through multi-source data fusion, finite element model and machine learning combination: through node threshold division and overloading / regular damage separate calculation, the damage evaluation accuracy is improved; with the help of LSTM and convolutional neural network, the time sequence and frequency domain features are captured, and the model is optimized combined with physical constraints, and the data-driven and theoretical support are considered, the prediction accuracy is greatly improved, and scientific basis is provided for individualized maintenance. SUMMARY

[0005] The application aims to provide a walking aid service life prediction method based on machine learning.

[0006] In order to achieve the above-mentioned purpose, the application is implemented according to the following technical scheme:

[0007] The application comprises the following steps:

[0008] Collecting use data and material performance detection data of a preset walking aid, and preprocessing the use data and the material performance detection data; the use data includes equipment sensor data, maintenance data, user behavior data, environment data and working state; the material performance detection data includes elastic modulus, Poisson's ratio, density, yield strength, fatigue strength coefficient, ultimate tensile strength, fatigue strength index, porosity and crack position;

[0009] According to the use data and the material performance detection data, a finite element model of the walking aid is established, node bearing threshold data of the aid is obtained, a mapping model based on pressure material characteristics is established according to the maintenance data, user behavior data and environment data, node time sequence stress of the aid is obtained, when the node time sequence stress is greater than the node bearing threshold data, a first cumulative damage is calculated according to the excessive frequency and the excessive value; comprising:

[0010] Given the finite element control equation, the expression is:

[0011]

[0012]

[0013]

[0014] Wherein is the Cauchy stress tensor, is the elastic stiffness tensor, is the body force vector, E is the elastic modulus, is the stress tensor, T is the transpose, is the nabla operator, is the geometric region of the entire walking aid, is the double dot product, is the Poisson's ratio, is the first parameter of Lamé, is the second parameter of Lamé, is the Kronecker function of the i-th node index j, is the displacement gradient, is the displacement vector field, is the Kronecker function of the i-th node index k, is the Kronecker function of the j-th node index k, is the Kronecker function of the j-th node index , is the Kronecker function of the i-th node index ;

[0015] The node stress of the walking aid finite element is calculated:

[0016]

[0017] Wherein is the elastic tensor of the i-th node, is the stress tensor of the i-th node;

[0018] The bearing capacity, ultimate bearing capacity and allowable load of the node are calculated:

[0019]

[0020]

[0021]

[0022] wherein is the bearing capacity of the ith node, , , is the principal stress at the ith node, is the effective bearing area of the ith node, is the stress concentration factor of the ith node, is the ultimate bearing capacity of the ith node, is the ultimate tensile strength, is the static strength safety factor of the ith node, S is the safety factor, is the allowable load of the ith node, is the ultimate bearing capacity of the ith node, is the yield strength;

[0023] output the allowable load as the node bearing threshold data of the assistive device;

[0024] otherwise, the node time sequence force is input into the machine learning model to obtain the use fatigue data, and a second cumulative damage is calculated according to the use fatigue data;

[0025] According to the first cumulative damage and the second cumulative damage, an assistive device service life prediction model is constructed, and prediction data is input into the assistive device service life prediction model to output a prediction result.

[0026] Further, a method for establishing a mapping model based on the pressure exertion material characteristics according to the maintenance data, user behavior data and environmental data, comprising:

[0027] A nonlinear mapping relationship from multi-source data to the node force of the assistive device is established, and the expression is:

[0028]

[0029] wherein is the mapping model at time t, is the maintenance data at time t, is the user behavior data at time t, is the environmental data at time t, is the node force vector at time t;

[0030] Collecting environmental data, maintenance data and user behavior data; maintenance data includes usage time accumulation, connection component looseness rate, lubrication performance decay function and structural stiffness degradation index; user behavior data includes real-time gait cycle phase, foot-ground contact force estimation, motion acceleration amplitude, body weight distribution ratio and motion speed change rate; environmental data includes ground slope angle, ground unevenness coefficient, ground friction coefficient, environmental temperature change rate, wind speed and wind direction influence factor;

[0031]

[0032]

[0033] wherein is a system state matrix, is an input coupling matrix, is an external input vector of time t, is a nonlinear mapping function, is a node axial force of time t, is a node shear force of time t, is a node bending moment of time t, is a node torsion of time t, is a node vibration load amplitude of time t, and T is a transpose;

[0034] The frequency characteristics of the force signal are captured by using Fourier transformation, and the expression is:

[0035]

[0036] wherein is a resonance frequency, is a frequency domain characteristic, is a force vector of time t in the time domain, z is an imaginary unit, and t is time;

[0037] The predicted force and the actual force of the assistive device node are obtained, and the mapping model parameters are learned by using the minimum sequence prediction error, and the expression is:

[0038]

[0039] wherein is an observation time window, is a regularization term, is a regularization coefficient, is a loss function, is a force predicted by the mapping model, is an actual force of the time variable ;

[0040] ​The maintenance data, user behavior data and environment data to be predicted are input into a mapping model, and the predicted stress is output to the node time sequence of the assistive device according to the time sequence.

[0041] Further, the method for calculating the first cumulative damage according to the excessive frequency and the excessive value comprises:

[0042] The stress size of the target node overload is obtained, and for single overload, the damage is calculated:

[0043]

[0044]

[0045] wherein is the stress size of the kth overload, is the material yield strength, is the excessive value, is the damage increment of a single overload event, is the damage coefficient, is the damage index;

[0046] For multiple overload events, the overload events are grouped according to the stress level, and an interaction factor is introduced to calculate the first cumulative damage:

[0047]

[0048] wherein is the interaction factor, the enhancement effect of the yth stress level event on the damage of the bth stress level, is the frequency of the bth stress level, is the first cumulative damage, is the stress level frequency of the yth stress level event, and n is the number of stress level groups; the interaction factor is calibrated by analyzing the material accelerated fatigue test results under historical overload data;

[0049] The critical damage value is determined through experiments, and when the cumulative damage is greater than or equal to the critical damage value, the material fails, and the damage value after that is zero.

[0050] Further, the method for obtaining the fatigue data comprises:

[0051] The node time sequence stress meeting the condition is converted into a feature vector by using a feature extraction operator, the machine learning model comprises a feature engineering layer, a time sequence feature extraction layer and a frequency domain feature extraction layer, and the expression is:

[0052]

[0053]

[0054]

[0055]

[0056] wherein is an activation function, is a weight matrix, is a long short-term memory network, is a time series force vector, is a parameter set of the long short-term memory network, is an input feature vector, is a time series feature vector extracted from the long short-term memory network, is a frequency domain representation of the force signal, is a fused high-level feature vector, is a bias vector, is a frequency domain feature vector, is a modulus of the a-th Fourier coefficient, is a start time of the time window, is a time length of the time window;

[0057] According to the extracted features, multi-modal feature fusion and output fatigue data output are performed, and the expression is:

[0058]

[0059]

[0060] wherein , is a prediction output of the machine learning model, is a weight vector of the output layer, is a bias term of the output layer, is a fused feature vector of the previous layer, is a weight matrix, is a weighted frequency domain feature, is a nonlinear activation function; is a cumulative fatigue damage at time t, is a damage growth rate at time t, is a material performance degradation rate at time t;

[0061] In the machine learning model, a physical constraint is embedded, and the expression is:

[0062]

[0063]

[0064]

[0065] wherein The rate of change of cumulative damage, This is the derivative of the material property degradation rate with respect to time. For fatigue damage, This is the derivative of fatigue damage with respect to time. This refers to the critical energy absorption capacity of the material.

[0066] Given the loss function of a machine learning model, the expression is:

[0067]

[0068]

[0069] in For the loss function of the machine learning model, For physical loss function, This represents the prediction output of the machine learning model for the c-th sample. This represents the actual output of the c-th sample, where N is the number of training samples. Let be the square of the Euclidean norm. The regularization coefficient is . Here, G represents the penalty coefficient for physical constraints, and G is the set of parameters for the machine learning model. The cumulative fatigue damage is for the c-th sample.

[0070] The cumulative fatigue damage, damage growth rate, and material property degradation rate are output as fatigue data based on the machine learning model.

[0071] Further, the method for calculating the second cumulative damage using fatigue data includes:

[0072]

[0073]

[0074]

[0075] in For time The rate of damage growth, For time Cumulative fatigue damage, This is the upper limit of the monitoring time. This is the component of mechanical fatigue damage. This refers to the material aging damage component. For the second cumulative damage, For reference degradation rate, The instantaneous material property degradation rate, An aging process index. For coupling coefficient.

[0076] Further, the method for constructing the assistive device service life prediction model according to the first cumulative damage and the second cumulative damage comprises:

[0077] Objective weights of the first cumulative damage, the second cumulative damage and the loss function of the assistive device service life prediction model are obtained, a target function is constructed by using a weighted sum of the first cumulative damage, the second cumulative damage and the loss function based on the objective weights, and when a value of the target function reaches a critical value, a corresponding critical time is the service life of the assistive device;

[0078] The assistive device service life prediction model comprises a one-dimensional convolutional neural network algorithm, a multi-layer perception and a long short-term memory network.

[0079] The one-dimensional convolutional neural network algorithm automatically scans input data through a plurality of convolution kernels, and extracts abstract features from local edges to global patterns layer by layer.

[0080] The multi-layer perception learns a complex nonlinear mapping in the abstract features, and the long short-term memory network captures a dependency relationship and an evolution trend of the abstract features in a time sequence.

[0081] The multi-layer perception and the long short-term memory network work cooperatively, and realize an end-to-end precise prediction of the service life of the assistive device by optimizing the target function.

[0082] In a second aspect, an assistive device service life prediction system based on machine learning comprises:

[0083] A data acquisition module is configured to acquire usage data and material performance detection data of a preset walking assistive device, and to pre-process the usage data and the material performance detection data; the usage data comprises device sensor data, maintenance data, user behavior data, environmental data and working state; and the material performance detection data comprises assistive device type, material, initial quality score and manufacturer.

[0084] An excessive fatigue damage calculation module is configured to establish a finite element model of the walking assistive device according to the usage data and the material performance detection data, to acquire node bearing threshold data of the assistive device, to establish a mapping model based on pressure material characteristics according to the maintenance data, the user behavior data and the environmental data, to acquire node time sequence stress of the assistive device, and to calculate a first cumulative damage according to an excessive frequency and an excessive value when the node time sequence stress is greater than the node bearing threshold data; the calculation of the first cumulative damage comprises:

[0085] A finite element control equation is given, and the expression is as follows:

[0086]

[0087]

[0088]

[0089] wherein is the Cauchy stress tensor, is the elastic stiffness tensor, is the body force vector, E is the elastic modulus, is the stress tensor, T is the transpose, is the nabla operator, is the geometric area of the whole walking aid, is the double dot product, is the Poisson ratio, is the first Lamé parameter, is the second Lamé parameter, is the Kronecker function of the i-th node index j, is the displacement gradient, is the displacement vector field, is the Kronecker function of the i-th node index k, is the Kronecker function of the j-th node index k, is the Kronecker function of the j-th node index is the Kronecker function of the j-th node index is the Kronecker function of the i-th node index is the Kronecker function of the i-th node index

[0090] The nodes of the finite element of the walking aid are divided, and the node stress is calculated:

[0091]

[0092] wherein is the elastic tensor of the i-th node, is the stress tensor of the i-th node;

[0093] The bearing capacity, the ultimate bearing capacity and the allowable load of the node are calculated:

[0094]

[0095]

[0096]

[0097] wherein is the bearing capacity of the i-th node, , , is the principal stress at the i-th node, is the effective bearing area of the i-th node, is the stress concentration factor of the i-th node, is the ultimate bearing capacity of the i-th node, is the ultimate tensile strength, is the static strength safety factor of the i-th node, S is a safety factor, is the allowable load of the i-th node, is the ultimate bearing capacity of the i-th node, is the yield strength;

[0098] output the allowable load as the node bearing threshold data of the assistive device;

[0099] The specification fatigue calculation module is used to input the node time sequence stress into the machine learning model to obtain usage fatigue data, calculate a second cumulative damage according to the usage fatigue data; the usage fatigue data includes a fatigue ratio and a fatigue degree;

[0100] The modeling prediction output module is used to construct an assistive device service life prediction model according to the first cumulative damage and the second cumulative damage, input to-be-predicted data into the assistive device service life prediction model, and output a prediction result.

[0101] The beneficial effects of the present application are:

[0102] The present application breaks through the limitations of traditional single data by the preprocessing, establishing a finite element model of a walking assistive device, mapping a model, calculating a first cumulative damage, calculating a second cumulative damage, and model construction steps, fuses device sensor, maintenance, user behavior, environment, and material performance data, predicts a more realistic use scenario, distinguishes whether the node time sequence stress is over the threshold, respectively calculates the first and second cumulative damages, takes into account overload and regular fatigue, accurately calculates the damage, combines the finite element model and machine learning, embeds physical constraints, fuses time sequence and frequency domain features, improves prediction accuracy, optimizes maintenance plans, ensures user safety, and can also adapt to different assistive device needs, and has strong universality. BRIEF DESCRIPTION OF DRAWINGS

[0103] Figure 1 is a step flowchart of the assistive device service life prediction method based on machine learning. DETAILED DESCRIPTION

[0104] The present application will be further described below through specific embodiments, and the illustrative embodiments of the present application and the description are used to explain the present application, but not as a limitation of the present application.

[0105] The assistive device service life prediction method and system based on machine learning include the following steps:

[0106] As shown in the figure, Figure 1 in the present embodiment, the following steps are included:

[0107] Collect usage data of a preset walking aid and material performance detection data, preprocess the usage data and the material performance detection data; the usage data includes device sensor data, maintenance data, user behavior data, environment data, and working state; the material performance detection data includes elastic modulus, Poisson's ratio, density, yield strength, fatigue strength coefficient, ultimate tensile strength, fatigue strength index, porosity, and crack position;

[0108] In actual evaluation, a brand of aluminum alloy material walking stick is taken as a prediction object, walking stick usage data is collected, device sensor data includes handle pressure and rod body vibration signal; maintenance data includes 180 days of cumulative use and 5% of connection loosening rate; user behavior data includes 0.8s of user gait cycle and 60% of body weight distribution ratio (supporting leg); environment data includes 3° of daily use ground slope and 2℃ / day of temperature change rate; material performance detection data is 70GPa of aluminum alloy elastic modulus, 0.33 of Poisson's ratio, and 270MPa of yield strength; after denoising and complementing, the data is standardized;

[0109] According to the usage data and the material performance detection data, a walking aid finite element model is established, node bearing threshold data of the aid is obtained, a mapping model based on pressure applying material characteristics is established according to the maintenance data, user behavior data and environment data, node time sequence stress of the aid is obtained, when the node time sequence stress is greater than the node bearing threshold data, a first cumulative damage is calculated according to the excessive frequency and excessive value; comprising:

[0110] Given the finite element control equation, the expression is:

[0111]

[0112]

[0113]

[0114] Wherein is the Cauchy stress tensor, is the elastic stiffness tensor, is the body force vector, E is the elastic modulus, is the stress tensor, T is the transpose, is the Nabla operator, is the geometric region of the entire walking aid, and is the double dot product, is the Poisson's ratio, is the first parameter of Ram, is the second parameter of Ram, is the Kronecker function of the i-th node index j, is the displacement gradient, is the displacement vector field, the Kronecker function of the kth node index of the ith node, the Kronecker function of the kth node index of the jth node, the Kronecker function of the kth node index of the jth node, the Kronecker function of the kth node index of the jth node, the Kronecker function of the kth node index of the ith node, the Kronecker function of the kth node index of the ith node;

[0115] Carry out node division on the finite element of the walking aid, and calculate the node stress:

[0116]

[0117] wherein the elastic tensor of the ith node, the stress tensor of the ith node;

[0118] Calculate the bearing capacity, ultimate bearing capacity and allowable load of the node:

[0119]

[0120]

[0121]

[0122] wherein the bearing capacity of the ith node, , , the principal stress at the ith node, the effective bearing area of the ith node, the stress concentration coefficient of the ith node, the ultimate bearing capacity of the ith node, the ultimate tensile strength, the static strength safety factor of the ith node, S is the safety factor, the allowable load of the ith node, the ultimate bearing capacity of the ith node, the yield strength;

[0123] Output the allowable load as the node bearing threshold data of the aid;

[0124] In actual evaluation, when the node time sequence stress is greater than the allowable load, overload occurs; substitute the data to establish a finite element model of the walking stick, calculate the node stress, and obtain the ultimate bearing capacity 1200N and the static strength safety factor 1.8 (node bearing threshold data); establish a mapping model, input maintenance, user behavior and environmental data, and output the node time sequence stress (such as the average axial pressure 800N);

[0125] If not, the node time sequence stress input machine learning model is used to obtain the use fatigue data, and the second cumulative damage is calculated according to the use fatigue data;

[0126] In actual evaluation, the node time sequence stress is 800N < 1200N, the machine learning model (including the LSTM time sequence feature extraction layer) is input, the use fatigue data (cumulative fatigue damage 0.3, damage growth rate 0.002 / day, material performance degradation rate 0.261) is obtained, and the second cumulative damage 0.32 is calculated according to the formula;

[0127] The first cumulative damage and the second cumulative damage are used to build an assistive device service life prediction model, and the prediction result is output by inputting the to-be-predicted data into the assistive device service life prediction model;

[0128] In actual evaluation, the prediction model is built in combination with no first cumulative damage (not exceeding the threshold value), the to-be-predicted data is input, and the remaining service life of the walking stick is about 280 days.

[0129] In this embodiment, the method for establishing a mapping model based on the pressure application material characteristics according to the maintenance data, the user behavior data and the environment data comprises:

[0130] A nonlinear mapping relationship from multi-source data to assistive device node stress is established, and the expression is:

[0131]

[0132] Wherein is the mapping model of time t, is the maintenance data of time t, is the user behavior data of time t, is the environment data of time t, is the node stress vector of time t;

[0133] The environment data, the maintenance data and the user behavior data are collected; the maintenance data includes use time accumulation, connection component loosening rate, lubrication performance attenuation function and structure stiffness degradation index; the user behavior data includes real-time gait cycle phase, foot-ground contact force estimation, motion acceleration amplitude, body weight distribution ratio and motion speed change rate; the environment data includes ground slope angle, ground unevenness coefficient, ground friction coefficient, environment temperature change rate, wind speed and wind direction influence factor;

[0134]

[0135]

[0136] Wherein is a system state matrix, is an input coupling matrix, Let be the external input vector at time t. It is a nonlinear mapping function. Let be the axial pressure at the node at time t. Let be the nodal shear force at time t. Let be the nodal bending moment at time t. Let be the node torque at time t. Let T be the amplitude of the nodal vibration load at time t, where T is the transpose.

[0137] The frequency characteristics of the force signal are captured using Fourier transform, and the expression is as follows:

[0138]

[0139] in The resonant frequency, For frequency domain characteristics, Let z be the force vector in the time domain at time t, where z is the imaginary unit and t is time.

[0140] Obtain the predicted and actual forces on the assistive device nodes, and learn the mapping model parameters by minimizing the time-series prediction error. The expression is:

[0141]

[0142] in For the observation time window, For regularization terms, The regularization coefficient is . For loss function, Predicting time variables for mapping models The force, For time variables The actual force;

[0143] The model uses the maintenance data, user behavior data and environmental data to be predicted as inputs, and then outputs the time sequence of the stress on the nodes of the auxiliary device according to the time order.

[0144] In this embodiment, the method for calculating the first cumulative damage based on the excess frequency and excess value includes:

[0145] Obtain the stress magnitude of the target node under overload. For a single overload, calculate the damage:

[0146]

[0147]

[0148] in Let the stress be the magnitude of the k-th overload. For the material's yield strength, is an overvalue, is an incremental damage of a single overload event, is a damage coefficient, is a damage index;

[0149] For multiple overload events, the overload events are grouped according to the stress level, and an interaction factor is introduced to calculate the first cumulative damage:

[0150]

[0151] wherein is an interaction factor, the enhancement effect of the yth stress level event on the bth stress level time damage; is the bth stress level frequency, is the first cumulative damage, is the stress level frequency of the yth stress level event, and n is the number of stress level groups; the interaction factor is calibrated by analyzing the material accelerated fatigue test results under historical overload data;

[0152] The critical damage value is determined through experiments, and when the cumulative damage is greater than or equal to the critical damage value, the material fails, and the damage value after that is zero.

[0153] In the embodiment, the method for obtaining fatigue and damage data includes:

[0154] The time series stress of the node meeting the condition is converted into a feature vector by using a feature extraction operator, the machine learning model includes a feature engineering layer, a time series feature extraction layer and a frequency domain feature extraction layer, and the expression is:

[0155]

[0156]

[0157]

[0158]

[0159] wherein is an activation function, is a weight matrix, is a long short-term memory network, is a time series stress vector, is a parameter set of the long short-term memory network, is an input feature vector, is a time series feature vector extracted from the long short-term memory network, is a frequency domain representation of the force signal, is a fused high-level feature vector, is a bias vector, is the frequency domain feature vector, is the modulus of the a-th Fourier coefficient, is the start time of the time window, is the duration of the time window;

[0160] The multi-modal feature fusion and output fatigue data output are performed according to the extracted features, and the expression is:

[0161]

[0162]

[0163] wherein , is the prediction output of the machine learning model, is the weight vector of the output layer, is the bias term of the output layer, is the fusion feature vector of the previous layer, is the weight matrix, is the weighted frequency domain feature, is a nonlinear activation function; is the cumulative fatigue damage at time t, is the damage growth rate at time t, is the material performance degradation rate at time t;

[0164] The physical constraint is embedded in the machine learning model, and the expression is:

[0165]

[0166]

[0167]

[0168] wherein is the change rate of the cumulative damage, is the derivative of the material performance degradation rate with respect to time, is the fatigue damage, is the derivative of the fatigue damage with respect to time, is the critical energy absorption capacity of the material;

[0169] The physical constraint is realized by adding a penalty term in the loss function. Given the loss function of the machine learning model, the expression is:

[0170]

[0171]

[0172] wherein a loss function of a machine learning model, a physical loss function, a predicted output of the machine learning model for the cth sample, an actual output of the cth sample, N is the number of training samples, a square of the Euclidean norm, a regularization coefficient, a penalty coefficient of a physical constraint, G is a parameter set of the machine learning model, cumulative fatigue damage of the cth sample;

[0173] According to the machine learning model, the cumulative fatigue damage, the damage growth rate, and the material performance degradation rate are output as the service fatigue data.

[0174] In the embodiment, the method for calculating the second cumulative damage according to the service fatigue data comprises:

[0175]

[0176]

[0177]

[0178] wherein is the time damage growth rate, is the cumulative fatigue damage of the time , T is the upper limit of the monitoring time, is the mechanical fatigue damage component, is the material aging damage component, is the second cumulative damage, is the reference degradation rate, is the instantaneous material performance degradation rate, is the aging process index, is the coupling coefficient. In the embodiment, the method for constructing the assistive device service life prediction model according to the first cumulative damage and the second cumulative damage comprises:

[0179] Objective weights of the first cumulative damage, the second cumulative damage, and the loss function of the assistive device service life prediction model are obtained, and a target function is constructed by using a weighted sum of the first cumulative damage, the second cumulative damage, and the loss function based on the objective weights, and when the value of the target function reaches a critical value, the corresponding critical time is the service life of the assistive device;

[0180] The assistive device service life prediction model comprises a one-dimensional convolutional neural network algorithm, a multi-layer perception machine, and a long short-term memory network.

[0181] ​

[0182] The one-dimensional convolutional neural network algorithm automatically scans the input data (such as stress cloud maps and vibration signals) through multiple layers of convolution kernels, and extracts abstract features from local edges to global patterns layer by layer;

[0183] The multi-layer perception learns the complex nonlinear mapping in the abstract features; the long short-term memory network captures the dependence relationship and evolution trend of the abstract features in the time sequence;

[0184] The multi-layer perception and the long short-term memory network work cooperatively to realize the end-to-end accurate prediction of the service life of the assistive device through optimization of the objective function.

[0185] In a second aspect, a machine learning-based assistive device service life prediction system includes:

[0186] A data acquisition module is configured to acquire usage data and material performance detection data of a preset walking assistive device, and to preprocess the usage data and the material performance detection data; the usage data includes device sensor data, maintenance data, user behavior data, environmental data, and working state; and the material performance detection data includes assistive device type, material, initial quality score, and manufacturer;

[0187] An excessive fatigue calculation module is configured to establish a finite element model of the walking assistive device based on the usage data and the material performance detection data, to obtain node bearing threshold data of the assistive device, to establish a mapping model based on pressure material characteristics based on the maintenance data, user behavior data, and environmental data, to obtain node time-series stress of the assistive device, and to calculate a first cumulative damage based on excessive frequency and excessive value when the node time-series stress is greater than the node bearing threshold data;

[0188] A normal fatigue calculation module is configured to input the node time-series stress into a machine learning model to obtain usage fatigue data when the node time-series stress is less than the node bearing threshold data, and to calculate a second cumulative damage based on the usage fatigue data; the usage fatigue data includes fatigue proportion and fatigue degree;

[0189] A modeling prediction output module is configured to construct an assistive device service life prediction model based on the first cumulative damage and the second cumulative damage, to input to-be-predicted data into the assistive device service life prediction model, and to output a prediction result.

[0190] The above description is only a preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting the service life of assistive devices based on machine learning, characterized in that, Includes the following steps: The system collects usage data and material performance test data of a pre-set walking aid, and preprocesses the usage data and material performance test data. The usage data includes equipment sensor data, maintenance data, user behavior data, environmental data, and working status. The material performance test data includes elastic modulus, Poisson's ratio, density, yield strength, fatigue strength coefficient, ultimate tensile strength, fatigue strength index, porosity, and crack location. A finite element model of the walking aid is established based on the usage data and the material performance test data to obtain the node bearing threshold data of the aid. A mapping model based on the properties of the pressure material is established based on the maintenance data, user behavior data, and environmental data to obtain the temporal stress of the aid's nodes. When the temporal stress of the nodes exceeds the node bearing threshold data, the first cumulative damage is calculated based on the excess frequency and excess value; including: Given the finite element governing equations, the expression is: in For Cauchy stress tensor, Let be the elastic stiffness tensor. Let E be the volume force vector, and E be the elastic modulus. Let T be the stress tensor and T be the transpose. For the Naples operator, Let be the geometric region of the entire walking aid, and be the double dot product. Poisson's ratio, For Ram's first parameter, For Ram's second parameter, The Kronecker function for the index j of the i-th node. For displacement gradient, For displacement vector field, The Kronecker function for the index k of the i-th node. The Kronecker function for the index k of the j-th node, Index the j-th node The Kronecker function, Index of the i-th node The Kronecker function; The walking aid is divided into nodes using finite element method and the node stresses are calculated: in Let be the elastic tensor of the i-th node. Let be the stress tensor of the i-th node; Calculate the load-bearing capacity, ultimate bearing capacity, and allowable load of the nodes: in Let be the bearing capacity of the i-th node. , , Let be the principal stress at the i-th node. Let i be the effective carrying area of ​​the i-th node. Let be the stress concentration factor of the i-th node. Let be the ultimate carrying capacity of the i-th node. Ultimate tensile strength, Let S be the static strength safety factor of the i-th node, and S be the safety factor. Let be the allowable load for the i-th node. Let be the ultimate carrying capacity of the i-th node. Yield strength; The allowable load is output as the node bearing threshold data of the auxiliary device; Conversely, the node temporal stress is input into the machine learning model to obtain fatigue data, and the second cumulative damage is calculated based on the fatigue data. A prediction model for the service life of assistive devices is constructed based on the first cumulative damage and the second cumulative damage. The data to be predicted is input into the prediction model for the service life of assistive devices, and the prediction results are output.

2. The method for predicting the service life of assistive devices based on machine learning according to claim 1, characterized in that, A method for establishing a mapping model based on the properties of the pressure material using the maintenance data, user behavior data, and environmental data includes: Establish a nonlinear mapping relationship from multi-source data to the force on assistive device nodes, expressed as: in For the mapping model of time t, For maintenance data at time t, User behavior data at time t, For environmental data at time t, Let be the nodal force vector at time t; Collect environmental data, maintenance data, and user behavior data; maintenance data includes cumulative usage time, loosening rate of connecting parts, lubrication performance decay function, and structural stiffness degradation index; user behavior data includes real-time gait cycle phase, foot contact force estimation, motion acceleration amplitude, body weight distribution ratio, and motion speed change rate; environmental data includes ground slope angle, ground unevenness coefficient, ground friction coefficient, ambient temperature change rate, and wind speed and direction influencing factors; in The system state matrix, The input coupling matrix is... Let be the external input vector at time t. It is a nonlinear mapping function. Let be the axial pressure at the node at time t. Let be the nodal shear force at time t. Let be the nodal bending moment at time t. Let be the node torque at time t. Let T be the amplitude of the nodal vibration load at time t, where T is the transpose. The frequency characteristics of the force signal are captured using Fourier transform, and the expression is as follows: in The resonant frequency, For frequency domain characteristics, Let z be the force vector in the time domain at time t, where z is the imaginary unit and t is time. Obtain the predicted and actual forces on the assistive device nodes, and learn the mapping model parameters by minimizing the time-series prediction error. The expression is: in For the observation time window, For regularization terms, The regularization coefficient is . For loss function, Predicting time variables for mapping models The force, For time variables The actual force; The model uses the maintenance data, user behavior data and environmental data to be predicted as inputs, and then outputs the time sequence of the stress on the nodes of the auxiliary device according to the time order.

3. The method for predicting the service life of assistive devices based on machine learning according to claim 1, characterized in that, The method for calculating the first cumulative damage based on the excess frequency and excess value includes: Obtain the stress magnitude of the target node under overload. For a single overload, calculate the damage: in Let the stress be the magnitude of the k-th overload. For the material's yield strength, This is an excessive value. The damage increment for a single overload event. The damage coefficient is... Damage index; For multiple overload events, the overload events are grouped according to stress level, and an interaction factor is introduced to calculate the first cumulative damage: in , where is the interaction factor, representing the enhancing effect of the y-th stress level event on the time damage at the b-th stress level; For the b-th stress level frequency, For the first cumulative damage, Let y be the frequency of the stress level event, and n be the number of stress level groups; the interaction factor is calibrated by analyzing the accelerated fatigue test results of materials under historical overload data. The critical damage value is determined through experiments. When the cumulative damage is greater than or equal to the critical damage value, the material fails, and the damage value thereafter is zero.

4. The method for predicting the service life of assistive devices based on machine learning according to claim 1, characterized in that, Methods for obtaining fatigue data include: The temporal forces of nodes that meet the conditions are converted into feature vectors using feature extraction operators. The machine learning model includes a feature engineering layer, a temporal feature extraction layer, and a frequency domain feature extraction layer, expressed as follows: in For activation function, This is the weight matrix. For Long Short-Term Memory (LSTM) networks, For time-series force vectors, This is the parameter set for Long Short-Term Memory (LSTM) networks. The input feature vector, This is a temporal feature vector extracted from a Long Short-Term Memory network. For the frequency domain representation of the force signal, This is the fused high-level feature vector. For bias vectors, For frequency domain eigenvectors, Let be the modulus of the a-th Fourier coefficient. This is the start time of the time window. The duration of the time window; Based on the extracted features, multimodal feature fusion is performed, and fatigue data is output. The expression is as follows: in , This is the predicted output of the machine learning model. The weight vector of the output layer. For the bias term of the output layer, This is the fused feature vector from the previous layer. This is the weight matrix. For weighted frequency domain features, It is a non-linear activation function; The cumulative fatigue damage over time t. Let t be the rate of damage growth. The material property degradation rate over time t; Embedding physical constraints in a machine learning model, the expression is: in The rate of change of cumulative damage, This is the derivative of the material property degradation rate with respect to time. For fatigue damage, This represents the derivative of fatigue damage with respect to time. The critical energy absorption capacity of the material; Given the loss function of a machine learning model, the expression is: in For the loss function of the machine learning model, For physical loss function, This represents the prediction output of the machine learning model for the c-th sample. This represents the actual output of the c-th sample, where N is the number of training samples. Let be the square of the Euclidean norm. The regularization coefficient is . Here, G represents the penalty coefficient for physical constraints, and G is the set of parameters for the machine learning model. The cumulative fatigue damage is for the c-th sample. The cumulative fatigue damage, damage growth rate, and material property degradation rate are output as fatigue data based on the machine learning model.

5. The method for predicting the service life of assistive devices based on machine learning according to claim 1, characterized in that, The method for calculating the second cumulative damage using fatigue data includes: in For time The rate of damage growth, For time Cumulative fatigue damage, This is the upper limit of the monitoring time. This is the component of mechanical fatigue damage. This refers to the material aging damage component. For the second cumulative damage, For reference degradation rate, The instantaneous material property degradation rate, An aging process index. is the coupling coefficient.

6. The method for predicting the service life of assistive devices based on machine learning according to claim 1, characterized in that, A method for constructing an assistive device lifespan prediction model based on the first cumulative damage and the second cumulative damage includes: Obtain the objective weights of the first cumulative damage, the second cumulative damage, and the loss function of the assistive device service life prediction model. Based on the objective weights, construct the objective function by weighting the first cumulative damage, the second cumulative damage, and the loss function. When the objective function value reaches the critical value, the corresponding critical time is the service life of the assistive device. The assistive device lifespan prediction model includes a one-dimensional convolutional neural network algorithm, a multilayer perceptron, and a long short-term memory network. One-dimensional convolutional neural network algorithms automatically scan input data (such as stress cloud maps and vibration signals) through multiple layers of convolutional kernels, extracting abstract features from local edges to global patterns layer by layer; Multilayer perceptrons learn complex nonlinear mappings in abstract features; long short-term memory networks capture the dependencies and evolutionary trends of abstract features over time. Multilayer perceptrons and long short-term memory networks work together to achieve end-to-end accurate prediction of the lifespan of assistive devices by optimizing the objective function.

7. A machine learning-based assistive device lifespan prediction system, for performing the method according to any one of claims 1-6, characterized in that, include: Data acquisition module: used to collect usage data and material performance test data of preset walking aids, and to preprocess the usage data and material performance test data; the usage data includes equipment sensor data, maintenance data, user behavior data, environmental data, and working status; the material performance test data includes aid type, material, initial quality score, and manufacturer; Excessive fatigue calculation module: used to establish a finite element model of the walking aid based on the usage data and the material performance test data, obtain the node bearing threshold data of the aid, establish a mapping model based on the pressure material characteristics based on the maintenance data, user behavior data, and environmental data, obtain the node temporal stress of the aid, and when the node temporal stress is greater than the node bearing threshold data, calculate the first cumulative damage based on the excess frequency and excess value; including: Given the finite element governing equations, the expression is: in For Cauchy stress tensor, Let be the elastic stiffness tensor. Let E be the volume force vector, and E be the elastic modulus. Let T be the stress tensor and T be the transpose. For the Naples operator, Let be the geometric region of the entire walking aid, and be the double dot product. Poisson's ratio, For Ram's first parameter, For Ram's second parameter, The Kronecker function for the index j of the i-th node. For displacement gradient, For displacement vector field, The Kronecker function for the index k of the i-th node. The Kronecker function for the index k of the j-th node, Index the j-th node The Kronecker function, Index of the i-th node The Kronecker function; The walking aid is divided into nodes using finite element method and the node stresses are calculated: in Let be the elastic tensor of the i-th node. Let be the stress tensor of the i-th node; Calculate the load-bearing capacity, ultimate bearing capacity, and allowable load of the nodes: in Let be the bearing capacity of the i-th node. , , Let be the principal stress at the i-th node. Let i be the effective carrying area of ​​the i-th node. Let be the stress concentration factor of the i-th node. Let be the ultimate carrying capacity of the i-th node. Ultimate tensile strength, Let S be the static strength safety factor of the i-th node, and S be the safety factor. Let be the allowable load for the i-th node. Let be the ultimate carrying capacity of the i-th node. Yield strength; The allowable load is output as the node bearing threshold data of the auxiliary device; Standardized fatigue calculation module: Conversely, if the node's temporal stress is input into a machine learning model to obtain fatigue data, the second cumulative damage is calculated based on the fatigue data; the fatigue data includes fatigue ratio and fatigue degree; Modeling and prediction output module: used to construct an assistive device service life prediction model based on the first cumulative damage and the second cumulative damage, input the data to be predicted into the assistive device service life prediction model, and output the prediction results.

Images