Flex-PLI knee mechanical property dynamic adjustment method based on physical guidance neural network

Abstract

The invention discloses a Flex-PLI knee mechanical property dynamic adjustment method based on a physical guidance neural network, and the method comprises the steps: building a state relation of four rows of springs of a knee, and analyzing the generalization capability of asymmetric spring configuration; establishing a physically-guided Flex-PLI knee static model, and obtaining a relational expression between the variation of the spring elongation and the variation of the ligament elongation; dynamically adjusting the weight coefficient of the weight distribution module on the influence of the Flex-PLI knee static model, enhancing the contribution of key variables and distinguishing the importance difference of input data; static and dynamic calibration curves are distinguished, auxiliary input of spring elongation is combined, and multi-task learning is carried out in combination with a feature fusion module; the driving neural network establishes a physical guidance neural network driving model, and training and hyper-parameter optimization are carried out; and establishing a prediction model and predicting a dynamic calibration peak value. According to the method, the passing rate and the efficiency of the Flex-PLI knee in the static and dynamic calibration process can be remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automobile crash test, in particular to a Flex-PLI knee mechanical performance dynamic adjustment method based on a physically guided neural network. BACKGROUND

[0002] With the rapid development of global automobile industry, traffic safety is increasingly concerned. The national pedestrian protection regulations and vehicle new evaluation procedures all use Flex-PLI leg impactors to evaluate the potential harm to the lower limbs when the vehicle hits the pedestrian. The calibration pass rate of the Flex-PLI knee is a key indicator to measure whether it can be used for real vehicle testing.

[0003] The traditional adjustment method is usually a trial-and-error method, which adjusts the elongation combination of the four rows of springs in the knee, recalibrates, and continues until the calibration is passed. This method has strong randomness and low efficiency. Sometimes, although the static calibration result meets the standard, the dynamic calibration result may not meet the requirements. SUMMARY

[0004] The purpose of the present application is to provide a Flex-PLI knee mechanical performance dynamic adjustment method based on a physically guided neural network to solve the problems raised in the background.

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] A Flex-PLI knee mechanical performance dynamic adjustment method based on a physically guided neural network, comprising the following steps:

[0007] Step 1: Based on Hooke's law, establish the state relationship of the four rows of springs in the knee, and analyze the generalization ability of the asymmetric spring configuration;

[0008] Step 2: Establish a physically guided Flex-PLI knee static model to obtain the relationship between the spring elongation change and the ligament extension change;

[0009] Step 3: Through a learnable weight distribution module, dynamically adjust the weight coefficient of the weight distribution module affecting the Flex-PLI knee static model according to the characteristics of the physical property data, strengthen the contribution of key variables and distinguish the importance difference of input data;

[0010] Step 4: Use multi-scale feature extraction to distinguish static and dynamic calibration curves, combine the auxiliary input of spring elongation, and use a joint feature fusion module for multi-task learning;

[0011] Step 5: Combine the static model and data enhancement strategy to drive the neural network to establish a physically guided neural network driving model, and perform training and hyperparameter optimization;

[0012] Step 6, based on the trained physical guide neural network driving model, a prediction model is established, the static calibration curve is input, the initial spring elongation is verified in reverse, and the dynamic calibration peak value is predicted.

[0013] Further, the step 1 comprises:

[0014] The spring elastic coefficient is measured, the large spring elastic coefficient is k1, and the small spring elastic coefficient is k2. Three states of the spring in the knee are established: no compression state, compression state and initial state;

[0015] The spring force formula is as follows:

[0016] F MCL =k1(6-X MCL +x MCL ) (1)

[0017] F LCL =k1(6-X LCL +x LCL ) (2)

[0018] F ACL =k2(1-X ACL +x ACL ) (3)

[0019] F PCL =k2(1-X PCL +x PCL ) (4)

[0020] Wherein, F MCL represents the spring force corresponding to the medial collateral ligament, F LCL represents the spring force corresponding to the lateral collateral ligament, F ACL represents the spring force corresponding to the anterior cruciate ligament, F PCL represents the spring force corresponding to the posterior cruciate ligament, X MCL represents the initial elongation of the spring corresponding to the medial collateral ligament, X LCL represents the initial elongation of the spring corresponding to the lateral collateral ligament, X ACL represents the initial elongation of the spring corresponding to the anterior cruciate ligament, X PCL represents the initial elongation of the spring corresponding to the posterior cruciate ligament, X MCL represents the spring elongation change of the medial collateral ligament during calibration, x LCL represents the spring elongation change of the medial collateral ligament during calibration, x ACL represents the spring elongation change of the anterior cruciate ligament during calibration, x PCL represents the spring elongation change of the posterior cruciate ligament during calibration.

[0021] Further, the step 2 comprises:

[0022] Based on the Flex-PLI knee static calibration, a physical model is established, and stress analysis is performed to obtain state 1 and state 2;

[0023] The torque conservation and force conservation formulas of state 1 are as follows:

[0024]

[0025] The torque conservation and force conservation formulas of state 2 are as follows:

[0026]

[0027] Where, M i represents the torque of the i-th force pair on the support point of the plane, F i represents the i-th external force acting on the object, F 压 represents the pressure of the thigh side of the knee on the calf side during the calibration process, F 载荷 represents the pressure applied by the pressure head to the calibration assembly, d represents the force arm length of the pressure head center to the support point of the plane, d ACL represents the force arm length of F ACL to the support point of the plane, d MCL represents the force arm length of F MCL to the support point of the plane, d LCL represents the force arm length of F LCL to the support point of the plane, f represents the frictional force between the thigh side and the calf side of the knee during the calibration process, d f represents the force arm length of f to the support point of the plane, d 压 represents the force arm length of F 压 to the support point of the plane, d PCL represents the force arm length of F PCL to the support point of the plane, represents the included angle of F ACL with the vertical direction, γ represents the included angle of F LCL with the horizontal direction, α represents the included angle of F MCL with the horizontal direction, β is the included angle of F 压 with the horizontal direction, F 支撑 represents the force on the support point during the calibration process, θ is the included angle of F PCL with the vertical direction;

[0028] Combined with the static calibration results, a motion coordinator is established to obtain the relationship between the ligament extension change and the spring elongation change:

[0029]

[0030] wherein ΔX represents the change in ligament extension, and Δx represents the change in spring elongation.

[0031] Further, the step 3 comprises:

[0032] Based on the feature correlation of the physical attribute data, a weight vector is dynamically generated to quantify the influence degree of the weight vector on the output of the Flex-PLI knee static model, focus on key physical variables and suppress the interference of secondary variables; then in the forward propagation, the calculated weight value is applied to the physical loss phys_loss to generate a weighted loss.

[0033] Further, the step 4 comprises:

[0034] For the one-dimensional curve obtained by static calibration, a CNN convolutional neural network is used to extract local features in the one-dimensional curve; for the time series curve obtained by pendulum calibration and impact calibration in dynamic calibration, an LSTM long short-term memory network is used to remember the long-term dependence of displacement and time and capture the dynamic pattern in the time series; for the four rows of spring elongations of the auxiliary input, they are stored as scalars in the embedding layer.

[0035] Further, the step 5 comprises:

[0036] The static physical model is combined by physical modeling and data-driven modeling, the physical modeling is responsible for generating constraint conditions, and the data-driven modeling is responsible for learning residual terms;

[0037] The data enhancement strategy includes static data enhancement and dynamic data enhancement, the static data is enhanced based on small distortion of Fourier disturbance, and the dynamic data is enhanced by extracting the maximum value in the dynamic signal through micro translation and inversion.

[0038] Compared with the prior art, the beneficial effects of the present application are: the present application proposes a Flex-PLI knee mechanical performance dynamic adjustment method based on physical guided neural network, the goal is to significantly improve the pass rate and efficiency of Flex-PLI knee in the process of static and dynamic calibration, through multi-scale feature extraction and joint feature fusion module, the nonlinear relationship between spring elongation, static calibration curve and dynamic calibration curve is established. At the same time, in the case of data scarcity, relying on the physical guided static modeling framework and data enhancement strategy, the reliability and adaptability of the model to the prediction results are improved; by comparing and analyzing the training results of pure data-driven, traditional neural network-driven and physical guided neural network-driven, the optimality of the model effect is ensured, the reliability and practicability of the method are verified, and more accurate technical support is provided for automobile crash test. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The method flowchart of the present application.

[0040] Figure 2 Flex-PLI knee spring different state corresponding elongation diagram.

[0041] Figure 3 Flex-PLI knee static model diagram for physical guidance.

[0042] Figure 4 Generalization ability diagram for asymmetric spring configuration.

[0043] Figure 5 Training result comparison diagram for pure data driven, traditional neural network driven and physically guided neural network driven.

[0044] Figure 6 Prediction result comparison diagram of hybrid model and actual result. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0046] Please refer to Figures 1-6 , a Flex-PLI knee mechanical property dynamic adjustment method based on a physically guided neural network, comprising the following steps:

[0047] Step 1, based on Hooke's law, the state relationship of the four-row springs in the knee is established, and the generalization ability of asymmetric spring configuration is analyzed, including:

[0048] The spring elastic coefficient is measured, and in the present embodiment, the elastic coefficient of the large spring is k1=76.6×10 3 N / m, the elastic coefficient of the small spring is k2=71.1×10 3 N / m, three states of the spring in the knee are established: no compression state, compression state and initial state. The no compression state and the initial state of the spring are as shown in Figure 2 .

[0049] The spring force formula is as follows:

[0050] F MCL =76.6(6-X MCL +x MCL ) (12)

[0051] F LCL =76.6(6-X LCL +xLCL (13)

[0052] F ACL =71.1(1-X ACL +x ACL (14)

[0053] F PCL =71.1(1-X PCL +x PCL (15)

[0054] Among them, F MCL F represents the spring force corresponding to the medial collateral ligament. LCL F represents the spring force corresponding to the lateral collateral ligament. ACL F represents the spring force corresponding to the anterior cruciate ligament. PCL X represents the spring force corresponding to the posterior cruciate ligament. MCL X represents the initial elongation of the spring corresponding to the medial collateral ligament. LCL X represents the initial elongation of the spring corresponding to the lateral collateral ligament. ACL X represents the initial elongation of the spring corresponding to the anterior cruciate ligament. PCL x represents the initial elongation of the spring corresponding to the posterior cruciate ligament. MCL x represents the change in spring elongation corresponding to the medial collateral ligament during the calibration process. LCL x represents the change in spring elongation corresponding to the medial collateral ligament during the calibration process. ACL x represents the change in spring elongation corresponding to the anterior cruciate ligament during the calibration process. PCL This represents the change in the spring elongation of the posterior cruciate ligament during the calibration process.

[0055] Step 2: Establish a physically guided Flex-PLI knee static model to obtain the relationship between the change in spring elongation and the change in ligament extension, including:

[0056] like Figure 3 As shown, a physical model of the Flex-PLI knee was established based on a static calibration experiment. The physical model divides the system into two different working states based on whether the LCL is tensioned: State 1 represents the complete knee force when the LCL is tensioned and bears tensile force; State 2 represents the complete knee force when the LCL relaxes and no longer bears tensile force as the load increases. The corresponding moment conservation and force conservation equations are listed for each state. Since the pressure exerted by the loading head on the Flex-PLI knee is usually distributed across both joint surfaces, only one side needs to be selected for force analysis when establishing the static physical model.

[0057] The formulas for torque conservation and force conservation in state 1 are as follows:

[0058]

[0059] The moment of inertia and force conservation formula in state 2 is as follows:

[0060]

[0061] Where, M i represents the moment of the i-th force pair about the support point of the plane, F i represents the i-th external force acting on the object, F 压 represents the pressure generated by the knee thigh side to the calf side in the calibration process, F 载荷 represents the pressure exerted by the pressure head on the calibration assembly, d represents the force arm length of the pressure head center to the support point of the plane, d ACL represents the force arm length of F ACL to the support point of the plane, d MCL represents the force arm length of F MCL to the support point of the plane, d LCL represents the force arm length of F LCL to the support point of the plane, f represents the friction generated by the knee thigh side to the calf side in the calibration process, d f represents the force arm length of f to the support point of the plane, d 压 represents the force arm length of F 压 to the support point of the plane, d PCL represents the force arm length of F PCL to the support point of the plane, represents the angle between F ACL and the vertical direction, γ represents the angle between F LCL and the horizontal direction, α represents the angle between F MCL and the horizontal direction, β is the angle between F 压 and the horizontal direction, F 支撑 represents the force on the support point in the calibration process, θ is the angle between F PCL and the vertical direction.

[0062] As Figure 4 shown, combined with the static calibration result, a motion coordinator is established to obtain the relationship between the change amount of the knee ligament extension and the change amount of the spring elongation:

[0063]

[0064] Where ΔX represents the change amount of the ligament extension, and Δx represents the change amount of the spring elongation.

[0065] Step 3, through the learnable weight distribution module, according to the characteristics of the physical property data, dynamically adjust the weight coefficient of the weight distribution module to the influence of the Flex-PLI knee static model, strengthen the contribution of key variables and distinguish the importance difference of input data, including:

[0066] Based on the feature correlation of the physical attribute data, a weight vector is dynamically generated to quantify the influence of the weight vector on the output of the Flex-PLI knee static model, focus on key physical variables and suppress the interference of secondary variables; then in the forward propagation, the calculated weight value is applied to the physical loss phys_loss to generate a weighted loss.

[0067] Step 4, multi-scale feature extraction is used to distinguish static and dynamic calibration curves, combined with auxiliary input of spring elongation, multi-task learning is performed by joint feature fusion module, including:

[0068] For the one-dimensional curve obtained by static calibration, a CNN convolutional neural network is used to extract local features in the one-dimensional curve; for the time series curve obtained by pendulum calibration and impact calibration in dynamic calibration, an LSTM long short-term memory network is used to remember the long-term dependence of displacement and time and capture the dynamic pattern in the time series; for the auxiliary input of four rows of spring elongation, it is stored as a scalar in the embedding layer.

[0069] Step 5, combined with the static model and data augmentation strategy, the neural network is driven to establish a physically guided neural network driven model, including training and hyperparameter optimization, including:

[0070] The static physical model is composed of physical modeling and data-driven modeling, the physical modeling is responsible for generating constraint conditions, and the data-driven modeling is responsible for learning residual terms;

[0071] The data augmentation strategy includes static data augmentation and dynamic data augmentation, the static data is augmented based on small distortion of Fourier perturbation, and the dynamic data is augmented by extracting the maximum value in the dynamic signal through micro translation and inversion.

[0072] Subsequently, as Figure 5 Comparing the training results of pure data-driven, traditional neural network-driven and physically guided neural network-driven, it can be seen that the best validation loss of pure data-driven and traditional neural network-driven is greater than the training loss, which may exist overfitting phenomenon; the validation loss of the physically guided neural network-driven model is the lowest, and the training effect is the best.

[0073] Step 6, based on the trained physically guided neural network-driven model, a prediction model is established. The static calibration curve is input into the physically guided neural network-driven model, the initial spring elongation is verified in reverse, and the dynamic calibration peak value is predicted.

[0074] As Figure 6The results obtained by the prediction model are compared with the actual results. In the figure, (a) is a comparison diagram of the prediction results and the actual results of the elongation of the first and fourth rows of springs; (b) is a comparison diagram of the prediction results and the actual results of the elongation of the second and third rows of springs; (c) is a comparison diagram of the prediction results and the actual results of the dynamic pendulum peak value; and (d) is a comparison diagram of the prediction results and the actual results of the dynamic impact peak value. The analysis results show that the prediction model based on the physically guided neural network driving model has high precision and high robustness in the spring and dynamic calibration tasks. The overall average absolute error of the reverse prediction of the spring elongation is 0.06 mm; and the overall average absolute percentage error of the dynamic peak value prediction is 1.87%. The elongation prediction of most samples falls within the confidence interval of the corresponding spring, and most of the predicted dynamic peak values are within the error bar range obtained by the repeated pendulum and impact tests.

[0075] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for dynamically adjusting the mechanical properties of the Flex-PLI knee based on a physically guided neural network, characterized in that, Includes the following steps: Step 1: Based on Hooke's Law, establish the state relationship of the four rows of springs in the knee and analyze the generalization ability of the asymmetric spring configuration; Step 2: Establish a physically guided Flex-PLI knee static model to obtain the relationship between the change in spring elongation and the change in ligament extension. Step 3: Through the learnable weight allocation module, the weight coefficients of the weight allocation module on the influence of the Flex-PLI knee static model are dynamically adjusted according to the characteristics of the physical attribute data, so as to strengthen the contribution of key variables and distinguish the differences in the importance of input data. Step 4: Multi-scale feature extraction is used to distinguish between static and dynamic calibration curves. Combined with the auxiliary input of spring elongation, multi-task learning is carried out in conjunction with the feature fusion module. Step 5: Combining the static model and data augmentation strategies, drive the neural network to establish a physically guided neural network driven model, and perform training and hyperparameter optimization; Step 6: Based on the trained physical guidance neural network driving model, establish a prediction model, input the static calibration curve, reverse verify the initial spring elongation, and predict the dynamic calibration peak value.

2. The method for dynamically adjusting the Flex-PLI knee mechanical properties based on a physically guided neural network according to claim 1, characterized in that, Step 1 includes: The spring constants of the springs are measured. The spring constant of the large spring is k1, and the spring constant of the small spring is k2. Three states of the spring in the knee are established: no compression, compression, and initial state. The formula for spring force is as follows: F MCL =k1(6-X MCL +x MCL ) (1) F LCL =k1(6-X LCL +x LCL ) (2) F ACL =k2(1-X ACL +x ACL ) (3) F PCL =k2(1-X PCL +x PCL ) (4) Among them, F MCL F represents the spring force corresponding to the medial collateral ligament. LCL F represents the spring force corresponding to the lateral collateral ligament. ACL F represents the spring force corresponding to the anterior cruciate ligament. PCL X represents the spring force corresponding to the posterior cruciate ligament. MCL X represents the initial elongation of the spring corresponding to the medial collateral ligament. LCL X represents the initial elongation of the spring corresponding to the lateral collateral ligament. ACL x represents the initial elongation of the spring corresponding to the anterior cruciate ligament. PCL x represents the initial elongation of the spring corresponding to the posterior cruciate ligament. MCL x represents the change in spring elongation corresponding to the medial collateral ligament during the calibration process. LCL x represents the change in spring elongation corresponding to the medial collateral ligament during the calibration process. ACL x represents the change in spring elongation corresponding to the anterior cruciate ligament during the calibration process. PCL This represents the change in the spring elongation of the posterior cruciate ligament during the calibration process.

3. The method for dynamically adjusting the Flex-PLI knee mechanical properties based on a physical guided neural network according to claim 1, characterized in that, Step 2 includes: A physical model of the knee is established based on the static calibration of Flex-PLI, and force analysis is performed to obtain state 1 and state 2. The formulas for torque conservation and force conservation in state 1 are as follows: The formulas for torque conservation and force conservation in state 2 are as follows: Among them, M i F represents the moment of the i-th force about its supporting point in the plane. i F represents the i-th external force acting on the object. 压 F represents the pressure generated on the thigh side of the knee against the calf side during calibration. 载荷 The pressure applied by the indenter to the calibration component is represented by d, where d represents the lever arm length from the center of the indenter to the fulcrum on the plane. ACL Representing F ACL The length of the lever arm from the fulcrum in its plane, d MCL Representing F MCL The length of the lever arm from the fulcrum in its plane, d LCL Representing F LCL The lever arm length to the fulcrum in its plane, f represents the frictional force generated between the thigh and calf sides of the knee during calibration, and d f d represents the length of the lever arm from f to the fulcrum in its plane. 压 Representing F 压 The length of the lever arm from the fulcrum in its plane, d PCL Representing F PCL The length of the lever arm to the fulcrum in its plane. Representing F ACL The angle between F and the vertical direction, γ represents F LCL The angle between F and the horizontal direction, α represents F PCL The angle between F and the horizontal direction, β is F 压 The angle between F and the horizontal direction 支撑 θ represents the force acting on the fulcrum during the calibration process. PCL The angle with the vertical direction; Based on the static calibration results, a motion coordination device was established, and the relationship between the change in ligament elongation and the change in spring elongation was obtained: Where ΔX represents the change in ligament elongation and Δx represents the change in spring elongation.

4. The method for dynamically adjusting the Flex-PLI knee mechanical properties based on a physically guided neural network according to claim 1, characterized in that, Step 3 includes: Based on the feature correlation of physical attribute data, a weight vector is dynamically generated to quantify the influence of the weight vector on the output of the Flex-PLI knee static model, focusing on key physical variables and suppressing the interference of secondary variables; then, in the forward propagation, the calculated weight values ​​are applied to the physical loss phys_loss to generate a weighted loss.

5. The method for dynamically adjusting the Flex-PLI knee mechanical properties based on a physically guided neural network according to claim 1, characterized in that, Step 4 includes: For the one-dimensional curve obtained from static calibration, a CNN convolutional neural network is used to extract local features from the one-dimensional curve; for the time series curves obtained from pendulum calibration and impact calibration in dynamic calibration, an LSTM long short-term memory network is used to remember the long-term dependence of displacement and time and capture the dynamic patterns in the time series; for the extension of the four rows of springs in the auxiliary input, it is stored as a scalar in the embedding layer.

6. The method for dynamically adjusting the Flex-PLI knee mechanical properties based on a physically guided neural network according to claim 1, characterized in that, Step 5 includes: The static physical model is composed of physical modeling and data-driven modeling. Physical modeling is responsible for generating constraints, while data-driven modeling is responsible for learning residual terms. Data augmentation strategies include static data augmentation and dynamic data augmentation. Static data augmentation is performed based on small distortions caused by Fourier perturbations, while dynamic data augmentation is performed by extracting the maximum value from the dynamic signal through micro-shifting and inversion.

Images