Decoupling method of six-dimensional force sensor based on improved grey wolf optimization algorithm

Abstract

The application is a method for decoupling six-dimensional force sensor based on improved grey wolf algorithm optimization, relates to the technical field of six-dimensional force sensor, and comprises calibrating six-dimensional force sensor to obtain data under different direction force input; then, a BP neural network model is established, the output voltage signals of 6 channels are taken as the input vectors of the BP neural network, and the force acting on the sensor is taken as the output vector; finally, the grey wolf optimization algorithm is improved by using three strategies of adjusting dynamic distribution parameters, dynamic weight position updating strategy, setting diversity threshold and reinitializing population based on diversity, so that the accurate decoupling of the calibration data is realized; in the improved grey wolf algorithm, the diversity of the population is maintained or improved, so that the effective exploration of the algorithm in the global search space is promoted, the algorithm is prevented from falling into local optimum, the probability of finding global optimal solution is increased, the measurement precision and adaptability of the six-dimensional force sensor can be effectively improved, and the inter-dimension coupling is reduced.

Description

Technical Field

[0001] This invention relates to the field of six-dimensional force sensor technology, and in particular to a decoupling method for six-dimensional force sensors based on an improved gray wolf algorithm. Background Technology

[0002] With the advancement of technology, six-dimensional force sensors have an urgent need and broad application prospects in high-tech fields such as industrial engineering. In the aerospace field, six-dimensional force sensors helped Tiangong-1 and Shenzhou-9 achieve their smooth space docking. They can also be used to simulate the aerodynamic measurement of spacecraft propellers, enabling precise measurement of the changing six-dimensional forces experienced by spacecraft under aerodynamic loads. In the robotics field, six-dimensional force sensors are often installed in the joints of robots' fingers and wrists to achieve intelligent robot control. Therefore, the development and application of multi-dimensional force sensors have significant theoretical and practical implications. Due to the influence of factors such as the sensor's own structure and manufacturing errors, applying force in one dimension of the sensor will also result in outputs in other directions, leading to severe inter-dimensional coupling. This inter-dimensional coupling significantly affects the sensor's accuracy. To reduce the impact of coupling errors, a decoupling algorithm for six-dimensional force sensors is studied using a combination of error analysis, theoretical derivation, and experimental verification.

[0003] When using intelligent algorithms to decouple six-dimensional force sensors, commonly used algorithms include Backpropagation Neural Network (BPNN), Extreme Learning Machine (ELM), and Support Vector Regression (SVR). However, applying these algorithms to decouple six-dimensional force sensors during training can easily lead to problems such as oscillations, slow convergence, premature convergence, and getting trapped in local maxima, resulting in decoupling failures. This poses a challenge to research in this field, necessitating more intelligent and universal methods to address the interdimensional coupling problem of the sensor. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, such as oscillations, slow convergence, premature convergence, and getting trapped in local optima during training, which lead to decoupling failures, the main objective of this invention is to provide a decoupling method for a six-dimensional force sensor based on an improved Grey Wolf Optimization algorithm. This invention improves the Grey Wolf Optimization algorithm by employing three strategies: adjusting dynamic distribution parameters, nonlinear convergence factors, setting diversity thresholds, and re-initializing the population based on diversity. These strategies address the problems of traditional algorithms easily getting trapped in local optima and slow convergence in existing technologies.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A decoupling method for a six-dimensional force sensor based on an improved gray wolf algorithm is characterized by: using the output voltage signals U = [U...] of the six channels of the six-dimensional force sensor. Fx U Fy U Fz M Fx M Fy M Fz ], as the input vector of the neural network, acts on the sensor's F = [F x F y F z M x M y M z As the output vector, a six-dimensional force sensor decoupling model is established. The GreyWolf Optimization algorithm is improved by three strategies: adjusting the dynamic distribution parameters, nonlinear convergence factor, setting diversity threshold, and re-initializing the population based on diversity. The improved GreyWolf Optimization algorithm is then used to optimize the BP neural network to decouple the six-dimensional force sensor model.

[0007] In this invention example, the six-dimensional force sensor data is first calibrated by collecting and saving the data using a data acquisition card. Next, a BP neural network decoupling model is established, as follows:

[0008] A static loading method was used for the six-dimensional force sensor. The hardware system included a USB3200 data acquisition card, weights, and a six-dimensional force testing platform. First, the sensor was mounted on the calibration platform and secured. After completing all necessary hardware and software for calibration, loading the calibration table, and confirming the system connection was normal and usable, the experimental preparation was complete, and calibration began. During calibration loading, a principle of gradual increase was followed, starting from the minimum loading point and then progressively increasing the calibration load until the maximum range. After loading was completed, the values ​​were calibrated and saved. For multi-directional calibration of the sensor, the calibration loading space became three-dimensional, using simultaneous loading of force and torque in three directions. This facilitated recording the nonlinear mechanical characteristics under simultaneous multi-dimensional loads and prevented outliers. Three calibrations were performed in each direction, and the average value was taken to ensure data accuracy. Force and torque were applied in all directions. All data were recorded and saved, completing the calibration experiment.

[0009] In one embodiment of this invention, a neural network decoupling model is first constructed. Then, parameters are adjusted through orthogonal experiments, and the model's decoupling capability is used to evaluate whether the optimized weights and thresholds can achieve more accurate decoupling, thus establishing the final decoupling model. The improvement of the Grey Wolf Optimization Algorithm (GWO) using three strategies—adjusting dynamic distribution parameters, nonlinear convergence factors, setting diversity thresholds, and re-initializing the population based on diversity—is detailed below:

[0010] (1) Dynamically adjust the distribution parameters

[0011] To improve the performance of the gray wolf optimization algorithm in high-dimensional and complex search spaces, a method of dynamically adjusting the distributed parameters is adopted, as shown in formula (3), to generate the initial position of the search agent (wolf pack) and ensure that the initial wolf pack position does not exceed the predetermined boundary, as shown in formula (6).

[0012]

[0013] Where L represents the number of wolf pack layers, S refers to the number of search agents, and L e Let C represent the number of agents in each layer, C represent the standard deviation, and l represent the total number of wolf pack layers. Initialize agent positions. For even-numbered dimensions, position P is initialized using a normal distribution model, as shown in formula (4):

[0014] P(idx,d)=Ib(d)-(ub(d)-Ib(d))·n(0.5,C) (4)

[0015] For odd-dimensional positions, the initialization of position P adopts a uniform distribution model, as shown in formula (5):

[0016] P(idx,d)=Ib(d)+(ub(d)-Ib(d))·h (5)

[0017] To ensure that the location of each distribution is within the search space, as shown in formula (6):

[0018] P(idx,d)=max(Ib(d),min·(P(idx,d),ub(d))) (6)

[0019] Where ub(d) and Ib(d) represent the upper and lower boundaries, respectively, and h represents a random number.

[0020] (2) Nonlinear convergence factor

[0021] The update mechanism of the control parameter 'a' has a decisive impact on the overall performance of the algorithm. Traditional linear convergence cannot give full play to its optimization advantage, as shown in formula (7). Therefore, the convergence factor of the Grey Wolf algorithm is adjusted nonlinearly, such as... Figure 1 Specifically, as shown in formula (8):

[0022]

[0023] α=a max -(α min -cos(π·t / T max ))·k0-0.5 (8)

[0024] Among them, bmax =2, b min =0, t is the current iteration number, T is the maximum iteration number, and k0 is the adjustment coefficient, which takes values ​​in the range of (0, 1).

[0025] (3) Set diversity threshold and reinitialize the population based on diversity.

[0026] By setting diverse thresholds and re-initializing the population based on diversity, the diversity of the gray wolf population in the gray wolf optimization algorithm can be improved, thus promoting the algorithm's efficient search in the global search space. The specific steps are as follows:

[0027] First, the standard deviation of the wolf pack's position in each dimension is calculated. Then, the standard deviation of the position in each dimension is used to evaluate the diversity of the population. If the diversity value is lower than a pre-set threshold (indicating that the wolf pack may be too concentrated and the search range is limited), a portion of the population is selected for initialization to increase the diversity of the population, thereby improving search efficiency and preventing the algorithm from converging to a local optimum too early.

[0028] The improved gray wolf algorithm is a novel metaheuristic search algorithm. Its basic idea is to find the optimal solution through strict social hierarchy within the population, surrounding, tracking, and attacking prey. The specific steps are as follows:

[0029] 1. First, determine the neural network model, and then output the voltage signals U = [U] from the six channels of the sensor. Fx U Fy U Fz M Fx M Fy M Fz ], as the input vector of the BP neural network, acts on the sensor's F = [F x F y F z M x M y M z As the output vector, the input layer parameters m=6 and the output layer parameters n=6 of the neural network are represented by the empirical formula. (a is an integer between 1 and 10) and the trial-and-error method are used to determine the number of nodes in the neural network input layer and the neural network output layer. The number of nodes is set to 6. The data is loaded and normalized. The number of training samples, the number of test samples, the target error, etc. are set.

[0030] 2. Dynamically adjust the distribution parameters to initialize the wolf pack size, individual positions, number of iterations, and the values ​​of a, A, and C.

[0031] 3. Calculate the fitness values ​​between individual gray wolves and their prey, and designate the three wolves that are closest to their prey as α wolf, β wolf, and δ wolf.

[0032] The mathematical model of the algorithm is:

[0033] D = |C·X P (t)-X(t)| (9)

[0034] X(t+1)=X P (t)-A·D (10)

[0035] Where: D is the distance vector between the individual gray wolf and its prey; t is the current iteration number; X p X(t) is the position vector of the prey; X(t) is the position vector of the individual gray wolf; convergence factors A = 2a·r1-a, C = 2·r2, r1 and r2 are random vectors in [0,1], and a is a control parameter.

[0036] When surrounding prey, α, β, and δ wolves are the closest and most capable of sensing the prey's presence. In the abstract search space, the position vector of ω is updated by calculating the position vectors of α, β, and δ. Its mathematical expression is:

[0037]

[0038] Where: X represents the current position of the individual gray wolf; C1, C2, and C3 are random perturbation vectors; X α X β X δ The positions of α, β, and δ are respectively; D α D β D δ Let α, β, and δ be the distances to the current gray wolf individual, respectively. The formula below defines the step length and direction of the ω wolf individual in the pack as it moves towards the α, β, and δ wolves, as well as the final position of the next generation of gray wolves.

[0039]

[0040] In the Grey Wolf algorithm, the update mechanism of the control parameter 'a' has a decisive impact on the overall performance of the algorithm. However, the linear adjustment method fails to fully realize the potential for balancing the search for the global optimum with a precise search for local optima, as shown in the formula:

[0041]

[0042] Therefore, this invention proposes a nonlinear update strategy. Through this method, the adjustment of 'a' no longer follows a single linear path, but adopts a more complex nonlinear pattern, aiming to more precisely regulate the wolf pack's exploration and exploitation behavior, thereby achieving a better balance between global and local searches. The specific mathematical expression of this nonlinear strategy is as follows:

[0043] α=a max -(α min -cos(π·t / T max ))·k0-0.5 (14)

[0044] Among them, b max =2, b min =0, t is the current iteration number, T is the maximum iteration number, and k0 is the adjustment coefficient, which takes values ​​in the range of (0, 1).

[0045] Using this formula, the deceleration rate of the control parameter 'a' is reduced in the early stage, enhancing the global search capability; the deceleration rate is accelerated in the later stage, improving the convergence speed of the algorithm in local optimization problems. This adjustment mechanism is conducive to balancing exploration and development at different stages of the algorithm, improving optimization efficiency and effectiveness.

[0046] 4. Use the update formula α = a max -(α min -cos(π·t / T max ))·k0-0.5, convergence factor formula: A=2a·r1-a, and random vector formula: C=2·r2, update the values ​​of a, A, and C.

[0047] 5. Improve global search capabilities and avoid getting trapped in local optima by refining the position update mechanism and dynamic weights, and setting diversity thresholds.

[0048] Output the optimal solution X α and the optimal fitness value J α ;

[0049] 6. Calculate the distance between the new generation of wolves and their prey. Update the positions and new fitness values ​​of the α, β, and δ wolves based on their fitness values ​​and the formula X(t+1)=(X1+X2+X3) / 3. Determine if the maximum number of iterations has been reached. If so, output the optimization result: the optimal solution X. α The program will terminate if the condition is not met, and will return to recalculate if the condition is not met.

[0050] 7. The global optimal solution obtained in the entire improved Grey Wolf algorithm is back-mapped to the neural network. That is, the initial parameter values ​​of the number of hidden layer neurons, neuron connection weights and thresholds are solved by the Grey Wolf algorithm. A BP neural network is established, and normalized training sample data is used for learning and training. The trained network is tested with normalized test sample dataset, and the inverse normalized test result value is output. The output result is rounded, and the algorithm ends.

[0051] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention establishes a decoupling model based on the output and input of a six-dimensional force sensor. The Grey Wolf Optimization Algorithm (GWO) is improved using three strategies: adjusting dynamic distribution parameters, nonlinear convergence factors, setting diversity thresholds, and re-initializing the population based on diversity. Results show that the improved algorithm controls the Type I and Type II errors to within 0.21% and 0.068%, respectively (specifically as follows...). Figure 5 and Figure 6 This demonstrates higher decoupling accuracy and faster convergence speed than traditional intelligent algorithms such as BP neural networks and least squares methods. The improved GWO-BP algorithm exhibits excellent versatility and stability across different datasets, providing an effective approach for optimizing complex sensor systems. The correctness and superiority of the proposed model are verified. Attached Figure Description

[0052] The accompanying drawings, which are provided to further illustrate this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application.

[0053] Figure 1 This invention improves the gray wolf optimization algorithm and optimizes the BP neural network solution process.

[0054] Figure 2 To improve the convergence factor comparison before and after

[0055] Figure 3 Iteration curves of five optimization algorithms

[0056] Figure 4 Convergence curves of five optimization algorithms

[0057] Figure 5 Comparison curves of Type I error for different optimization algorithms

[0058] Figure 6 Comparison curves of Type II errors for different optimization algorithms Detailed Implementation

[0059] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0060] Example:

[0061] See Figures 1-6This invention provides an optimized decoupling method for a six-dimensional force sensor based on an improved gray wolf algorithm. It solves the problems encountered when using existing algorithms for decoupling six-dimensional force sensors, such as oscillations during training, slow convergence, premature convergence, and getting trapped in local maxima, leading to decoupling failures. Furthermore, this method adds difficulty to research in this field, necessitating a more intelligent and universal method to address the interdimensional coupling problem of the sensor. The method can be summarized as follows:

[0062] S1. Data calibration of the six-dimensional force sensor; S2. Construction of the BP neural network model; S3. Improved Grey Wolf algorithm; S4. Case study analysis.

[0063] S1 includes two steps:

[0064] S1.1 First, statically load the six-dimensional force sensor. The hardware system includes a USB3200 data acquisition card, weights, and a six-dimensional force testing platform. Next, install the sensor on the calibration platform and fix it in place. After completing all the necessary hardware and software equipment for calibration and loading the calibration table, and finally confirming that the system connection is normal and usable, the experimental preparation is complete, and calibration begins.

[0065] S1.2 During calibration loading, the principle of gradual increase is followed, starting from the minimum loading point and then increasing the calibration load step by step until the maximum range is reached. After loading is completed, the values ​​are calibrated and saved. When calibrating the sensor in multiple directions, the calibration loading space becomes a three-dimensional space. Forces and torques are applied simultaneously in three directions to facilitate recording the nonlinear mechanical characteristics under simultaneous multi-dimensional loads and to prevent singular values. Three calibrations are required in each direction, and the average value is taken to ensure data accuracy. Forces and torques are applied in all directions. All data are recorded and saved to complete the calibration experiment.

[0066] S2 includes a step:

[0067] The input force F = [F] of the six-dimensional force sensor x F y F z M x M y M z ] and output voltage U = [U Fx U Fy U Fz M Fx M Fy M Fz A decoupling model for a six-dimensional force sensor is established, where the input layer parameters m = 6, the output layer parameters n = 6, and the hidden layer parameters are determined by empirical formulas. (a is an integer between 1 and 10) and the trial-and-error method are used to determine it.

[0068] S3 consists of seven steps:

[0069] S3.1. Individuals are divided into four tiers based on their fitness values: α, β, δ, and ω. α represents the optimal solution, β and δ represent the second and third optimal solutions, respectively, and ω is the candidate solution.

[0070] S3.2 Initialize the relevant parameters of the improved gray wolf algorithm, and set the population size, adjustment coefficient, maximum number of iterations, etc.;

[0071] S3.3 Calculate the fitness value between individual gray wolves and their prey, and set the three wolves with the shortest distance from their prey as α wolf, β wolf, and δ wolf;

[0072] S3.4 Update the positions of α wolf, β wolf, and δ wolf individuals, and calculate the fitness value of the current gray wolf individual according to the formula: X(t+1)=(X1+X2+X3) / 3;

[0073] S3.5. Generate 'a' according to the improved nonlinear factor convergence control strategy, and calculate the cooperative parameters A and C according to the cooperative parameter adaptive optimization strategy.

[0074] S3.6. Improve the global search capability and avoid getting trapped in local optima by improving the position update mechanism and dynamic weights and setting diversity thresholds;

[0075] Output the optimal solution X α and the optimal fitness value J α ;

[0076] S3.7, The optimal solution X α The optimal weights and biases of the BP neural network are saved, and the optimized neural network is used to decouple the calibration data to determine the optimal solution X. α If it is feasible, the algorithm ends; otherwise, the calculation is repeated.

[0077] In step S2, the established decoupled data model is divided into a training set and a test set, with 80% used as the training set and 20% as the test set. The dataset division is performed randomly using the randperm function in Matlab to verify the stability of the algorithm and enhance its reliability through multiple calculations.

[0078] The specific steps for initializing the dynamically adjusted distribution parameters of the improved Grey Wolf algorithm in step S3.2 are as follows:

[0079] To improve the performance of the gray wolf optimization algorithm in high-dimensional and complex search spaces, a method of dynamically adjusting the distributed parameters is adopted, as shown in formula (3), to generate the initial position of the search agent (wolf pack) and ensure that the initial wolf pack position does not exceed the predetermined boundary, as shown in formula (6).

[0080]

[0081]

[0082] Where L represents the number of wolf pack layers, S refers to the number of search agents, and L e Let C represent the number of agents in each layer, C represent the standard deviation, and l represent the total number of wolf pack layers. Initialize agent positions. For even-numbered dimensions, position P is initialized using a normal distribution model, as shown in formula (4):

[0083] P(idx,d)=Ib(d)-(ub(d)-Ib(d))·n(0.5,C) (4)

[0084] For odd-dimensional positions, the initialization of position P adopts a uniform distribution model, as shown in formula (5):

[0085] P(idx,d)=Ib(d)+(ub(d)-Ib(d))·h (5)

[0086] To ensure that the location of each distribution is within the search space, as shown in formula (6):

[0087] P(idx,d)=max(Ib(d),min·(P(idx,d),ub(d))) (6)

[0088] Where ub(d) and Ib(d) represent the upper and lower boundaries, respectively, and h represents a random number.

[0089] In step S3.5, initializing the improved Grey Wolf algorithm parameters requires optimizing the control parameter a. Traditional linear convergence cannot fully leverage its optimization advantage, as shown in formula (7). Therefore, the convergence factor of the Grey Wolf algorithm is nonlinearly adjusted, as shown in formula (7). Figure 2 Specifically, as shown in formula (8):

[0090]

[0091] α=a max -(α min -cos(π·t / T max ))·k0-0.5 (8)

[0092] Among them, b max =2, b min=0, t is the current iteration number, T is the maximum iteration number, and k0 is the adjustment coefficient, which takes values ​​in the range of (0, 1).

[0093] In step S3.6, setting diverse thresholds and re-initializing the population based on diversity can improve the diversity of the gray wolf population in the gray wolf optimization algorithm, promoting efficient search in the global search space. The specific steps are as follows:

[0094] First, the standard deviation of the wolf pack's position in each dimension is calculated. Then, the standard deviation of the position in each dimension is used to evaluate the diversity of the population. If the diversity value is lower than a pre-set threshold (indicating that the wolf pack may be too concentrated and the search range is limited), then a portion of the population is selected for initialization to increase the diversity of the population, thereby improving search efficiency and preventing the algorithm from converging to a local optimum too early.

[0095] S4. Case Analysis

[0096] Taking the calibration data of a six-dimensional force sensor as an example, a total of 160 sets of calibration data were analyzed, with 80% used as the training set and 20% as the test set. Considering the case of randomly applied forces in various directions, the dataset was randomly divided using the `randperm` function in Matlab to verify the algorithm's stability. The algorithm's reliability was enhanced by averaging multiple calculations. The parameters for the IGWO algorithm are as follows: gray wolf population size N = 50, maximum number of iterations T... max =200, the maximum value of the convergence factor a max =2, minimum value a min =0, set the diversity threshold to 0.2.

[0097] The IGWO algorithm was used in Matlab software for programming calculations. The BP neural network was optimized using the IGWO algorithm to decouple the six-dimensional force sensor model. The fitness curve of the IGWO algorithm is shown below. Figure 3 As shown, for comparative analysis, Figure 3 The paper also presents the convergence curves of the BP neural network algorithm, the Genetic Algorithm (GA), the Particle Swarm Optimization (PSO) algorithm, the GWO algorithm, and the IGWO algorithm for solving this model. Figure 3 It can be seen that the IGWO-BP algorithm converges to the optimal fitness value after about 29 generations and requires the fewest iterations.

[0098] Figure 4A comparison of the training error results of five algorithms is presented. The IGWO-BP algorithm, while achieving the same mean squared error, demonstrates fewer iterations and faster runtime compared to the other algorithms, effectively overcoming their shortcomings in convergence speed and local optima. The correctness of adjusting the dynamic distribution parameters, the nonlinear convergence factor, setting the diversity threshold, and the diversity-based population re-initialization strategy are verified.

[0099] Table 1 shows the errors of the six algorithms. Based on Table 1, Figure 5 and Figure 6 It can be concluded that the least squares method has relatively poor linear decoupling capability, with a maximum Type I error of 11.11% in all directions after decoupling. In contrast, the maximum Type I errors of BP, GA-BP, and the improved GWO-BP in each direction are 3.268%, 1.3%, and 0.341%, respectively, showing a significant reduction in Type I error. The improved GWO-BP algorithm controls the Type I error to within 0.267%, and the Type I errors in different directions are smaller than those of other algorithms. Regarding Type II error, the maximum error of the improved GWO-BP algorithm is 0.049%, significantly improving decoupling accuracy compared to other algorithms, especially in M... x M y M z In terms of direction, Type II error was controlled within 0.05%, and coupling error was well controlled.

[0100] Table 1 - Errors of Various Algorithms

[0101]

[0102] In summary, the improved gray wolf algorithm of this invention, which optimizes the BP neural network to decouple a six-dimensional force sensor, effectively reduces inter-dimensional coupling and thus improves the sensor's measurement accuracy. By employing three strategies—adjusting dynamic distribution parameters, nonlinear convergence factors, setting diversity thresholds, and re-initializing the population based on diversity—the convergence of the algorithm is effectively improved, the search space is expanded, and the probability of escaping local optima is increased. Experimental results not only demonstrate the advantages of this algorithm in the decoupling control of six-dimensional force sensors but also show its applicability to the decoupling control of other models.

[0103] It should be noted that, in this invention, relational terms such as error I and error II are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0104] The above embodiments are merely illustrative examples of the present invention and do not constitute a limitation on the scope of protection of the present invention. Any designs that are the same as or similar to the present invention are within the scope of protection of the present invention.

Claims

1. A method for decoupling of a six-dimensional force sensor based on an improved grey wolf optimization algorithm, characterized in that, The method comprises the following steps: Calibration of six-dimensional force sensor data, data acquisition card is used to collect data and save, and a BP neural network decoupling model is constructed; The grey wolf algorithm is improved by dynamically adjusting the distribution parameters, nonlinear convergence factor, setting the diversity threshold and reinitializing the population based on diversity; An optimal weight matrix and threshold vector are obtained by using the improved grey wolf algorithm, and the BP neural network decoupling model is optimized to obtain an optimized BP neural network decoupling model; The six-dimensional force sensor data is input into the optimized BP neural network decoupling model to obtain the decoupled force components; The population initialization of the grey wolf algorithm is improved by using the dynamic adjustment of the distribution parameters, comprising the following steps: S1, the total number of preset search wolf groups is S、 the upper boundary of the search space is the lower boundary of the search space is wherein d denotes a dimension index; S2, according to the total number of search wolf packs S Calculate the number of layers of wolf packs L And the number of agents per layer The expression is as follows: ；（1） ；（2） According to L calculating the dynamic distribution parameters corresponding to the layer C ; ；（3） wherein, L denotes the number of layers of the wolf pack, S refers to the total number of search wolves, denotes the number of agents per layer, denotes the total number of layers of the wolf pack; S3, initializing the agent position of each layer, specifically comprising: For even dimensions d , the initialization of the positions P uses a normal distribution model, formula as (4): ；（4） For odd dimensions d , the initialization of positions P uses a uniform distribution model, as in (5): ；（5） wherein h represents a random number; S4, ensure that the position generated by each distribution is within the search space, and the formula is as formula (6): ；（6） Convergence factor of grey wolf algorithm using the nonlinear convergence factor Improvements are made to the convergence factor The expression is as follows: ；（7） wherein, is a maximum value of the convergence factor, is a minimum value of the convergence factor, t is a current iteration number, is a maximum iteration number, k 0 is an adjustment coefficient, which is in the range of (0, 1); The grey wolf algorithm is improved by setting the diversity threshold and reinitializing the population based on diversity, comprising the following steps: The standard deviation of the position of the wolf group in each dimension is obtained; The diversity of the population is evaluated according to the standard deviation of the position of each dimension, wherein if the diversity value is lower than the pre-set threshold, a part of the population is initialized.

2. The method for decoupling of a six-dimensional force sensor based on an improved grey wolf optimization algorithm according to claim 1, wherein, The BP neural network decoupling model is constructed, comprising: Based on the input force F = [F x , F y , F z , M x , M y , M z ] of the six-dimensional force sensor and the output voltage U = [U Fx , U Fy , U Fz , M Fx , M Fy , M Fz ], the input layer m = 6, the output layer n = 6 and the number of hidden layer neurons of the neural network are determined by and trial method, wherein, is an integer between 1 and 10, and the structure of the BP neural network model is constructed.

3. The method for decoupling of a six-dimensional force sensor based on an improved grey wolf optimization algorithm according to claim 1, wherein, The optimal weight matrix and threshold vector include the initial parameter values of the number of hidden layer neurons, neuron connection weights and threshold values solved by the grey wolf algorithm.

4. The method for decoupling of a six-dimensional force sensor based on an improved grey wolf optimization algorithm according to claim 1, wherein, The improved grey wolf algorithm comprises the following steps: Individuals are divided into four hierarchical levels according to the fitness value: alpha, beta, delta and omega, wherein alpha is the optimal solution, beta and delta are the second and third optimal solutions respectively, and the candidate solution is omega; The related parameters of the improved grey wolf algorithm are initialized, and the population size, adjustment coefficient and maximum iteration number are set; The fitness value between the grey wolf individual and the prey is calculated, and the first three wolves with the shortest distance to the prey are set as alpha wolf, beta wolf and delta wolf; Update the positions of the alpha wolf, beta wolf, and delta wolf individuals according to the equations: Calculate the fitness value of the current grey wolf individual; According to the improved nonlinear factor convergence control strategy generation , the cooperative parameters A and C are calculated according to the cooperative parameter adaptive optimization strategy. According to the position updating mechanism and the improvement of dynamic weight, the threshold of population diversity is set. outputting the optimal solution and the optimal fitness value .

Images