Multi-dimensional force sensor decoupling method based on robust loss function and neural network

Abstract

The invention relates to a multi-dimensional force sensor decoupling method based on a robust loss function and a neural network, and belongs to the technical field of sensor measurement, and the method comprises the steps: constructing a coupling data set of a noise-containing multi-dimensional force sensor; designing a neural network model based on robustness; training a neural network model based on robustness by using the noise-containing coupling data set; and performing decoupling processing on a real-time output signal of the multi-dimensional force sensor by using the trained robustness-based neural network model, and the like. According to the method, a neural network model based on robustness is trained by using a coupled data set of a noise-containing multi-dimensional force sensor, and model training fits actual working conditions by using a robust loss function, so that the anti-noise generalization ability is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of sensor measurement, and particularly relates to a multi-dimensional force sensor decoupling method based on a robust loss function and a neural network. BACKGROUND

[0002] Multi-dimensional force sensors are widely used in static measurement scenarios (such as material strength testing, precision weighing, static assembly force monitoring, nuclear power oxide film detection robots, and bolt hole jamming processing robots), and the core requirement is to accurately obtain the static force / torque values of each dimension. However, the structural characteristics (such as geometric asymmetry and material unevenness) of the sensor elastic body can cause serious static coupling, and the static force of a certain dimension can cause the output signal deviation of other dimensions, and the noise (such as circuit thermal noise and mechanical micro-motion interference) in the static measurement environment can further amplify the coupling error and reduce the measurement accuracy.

[0003] In the traditional static decoupling method, the linear calibration method cannot handle nonlinear coupling, and the decoupling effect is limited. Although the method based on the neural network can fit the nonlinear relationship, the mean square error (MSE) loss function is usually used, such as the three-dimensional force sensor decoupling model determination method disclosed in the patent application with the application publication number CN116380321A. This method is based on training data, iteratively updates the parameters of the three-dimensional force sensor decoupling model until the value of the loss function MSE continuously decreases, and obtains the three-dimensional force sensor decoupling model in actual measurement. However, this model is sensitive to static noise and abnormal values. In actual static scenarios, the sensor output signal inevitably contains noise (such as signal drift in long-term static measurement) and occasional abnormal values (such as slight collision from the outside world), and the MSE loss will pay excessive attention to these noise points, causing the model to learn the wrong coupling rule, and the static decoupling accuracy is difficult to guarantee.

[0004] Therefore, it is of great significance to design a decoupling method with strong noise resistance for static working conditions to improve the static measurement performance of multi-dimensional force sensors. SUMMARY

[0005] The application aims to provide a multi-dimensional force sensor decoupling method based on a robust loss function and a neural network to solve the problem of difficult guarantee of static decoupling accuracy of existing multi-dimensional force sensors.

[0006] In order to achieve the above-mentioned purpose, the technical scheme of the application is as follows: The application relates to a multi-dimensional force sensor decoupling method based on a robust loss function and a neural network, comprising the following steps: S1. Constructing a coupling data set of a multi-dimensional force sensor containing noise; S2. Designing a neural network model based on robustness; S3. Train a robust neural network model using a noisy coupled dataset; S4. Decouple the real-time output signal of the multidimensional force sensor using a trained robust neural network model.

[0007] Preferably, the specific steps of constructing the coupled dataset of the noisy multidimensional force sensor in step S1 include: for the multidimensional force sensor, setting m equally spaced measurement points in each dimension; applying a multidimensional force / torque from 0 to full scale to the force sensor under static working conditions without dynamic interference, repeatedly collecting the output signal n times at each point to form a basic dataset; adding 1%-5% static random noise and 5%-15% outliers to the basic data to form the coupled dataset of the noisy multidimensional force sensor.

[0008] Preferably, the robust neural network model designed in S2 adopts a fully connected neural network, including an input layer, a hidden layer, and an output layer. The dimension of the input layer is consistent with the dimension of the force sensor output signal. The hidden layer is a two-layer fully connected layer, with 64 neurons in the first layer and 32 neurons in the second layer, activated by the ReLU activation function. The dimension of the output layer is consistent with the force / torque dimension and adopts linear activation. The Huber loss function is added as a robust loss term to suppress the influence of noise and outliers. The output layer also adds a physical constraint term based on the principle of force balance.

[0009] Preferably, the expression for the Huber loss function is: , in, Represents the loss function. This represents the static actual force or torque value. The force or torque value output by the model. The threshold value is used.

[0010] Preferably, in step S3, when training a robust neural network model using a noisy coupled dataset, the coupled dataset of the noisy multidimensional force sensor is divided into a training set and a test set. Training stops when the rate of change of the loss values ​​of the training set and the test set is less than a set threshold for several consecutive rounds.

[0011] Preferably, the specific method for real-time decoupling using the trained robust neural network model in S4 is as follows: the robust neural network model is deployed to the measurement system, the force sensor output signal is preprocessed and then input into the robust neural network model, and noise-resistant force / torque measurement values ​​are output.

[0012] Preferably, in step S1, after forming the coupled dataset of the noisy multidimensional force sensor, the data in the dataset is further normalized.

[0013] Compared with the prior art, the technical solution provided by this invention has the following advantages: 1. The multidimensional force sensor decoupling method based on robust loss function and neural network of the present invention designs a robust neural network model, trains the model, and uses the trained robust neural network model to decouple the real-time output signal of the multidimensional force sensor. The robust loss function is used to make the model training fit the actual working conditions and improve the noise resistance and generalization ability.

[0014] 2. The multidimensional force sensor decoupling method based on robust loss function and neural network of the present invention constructs a coupled dataset of noisy multidimensional force sensors, and uses the coupled dataset of noisy multidimensional force sensors to train a robust neural network model, so that the model training fits the actual working conditions and improves the noise resistance and generalization ability. Attached Figure Description

[0015] Figure 1 This is a flowchart of a multidimensional force sensor decoupling method based on robust loss function and neural network; Figure 2 This is a schematic diagram of the data acquisition device for constructing coupled datasets; Figure 3 This is a schematic diagram of a robust neural network structure; Figure 4 F under different decoupling methods x Box plot of component decoupling error to verify decoupling effect; Figure 5 F under different decoupling methods y Box plot of component decoupling error to verify decoupling effect; Figure 6 F under different decoupling methods z Box plot of component decoupling error, verification diagram of decoupling effect; Figure 7 M under different decoupling methods x Box plot of component decoupling error to verify decoupling effect; Figure 8 M under different decoupling methods y Box plot of component decoupling error to verify decoupling effect; Figure 9 M under different decoupling methods z Box plot of component decoupling error, verification diagram of decoupling effect; Figure 10 It is F xA comparison chart of normal dimension decoupling, robust optimized decoupling, and actual load volume; Figure 11 It is F y A comparison chart of normal dimension decoupling, robust optimized decoupling, and actual load volume; Figure 12 It is F z A comparison chart of normal dimension decoupling, robust optimized decoupling, and actual load volume; Figure 13 It is M x A comparison chart of normal dimension decoupling, robust optimized decoupling, and actual load volume; Figure 14 It is M y A comparison chart of normal dimension decoupling, robust optimized decoupling, and actual load volume; Figure 15 It is M z A comparison chart of normal dimensional decoupling, robust optimized decoupling, and actual load volume. Detailed Implementation

[0016] To further understand the content of this invention, the invention will be described in detail with reference to the embodiments. The following embodiments are used to illustrate the invention, but are not intended to limit the scope of the invention.

[0017] See attached document Figure 1 As shown, this invention relates to a method for decoupling a multidimensional force sensor based on a robust loss function and a neural network, comprising the following steps: S1. Construct a coupled dataset of a noisy multidimensional force sensor. Specific steps include: for example... Figure 2 The multi-dimensional force sensor shown has 32 equally spaced measurement points in each dimension. Under static conditions without dynamic interference, corresponding weights are used to apply multi-dimensional forces / torques from 0 to full scale to the force sensor. The output signal is repeatedly collected 5 times at each point, and the average value is taken as the basic data for that point. The six-dimensional data obtained by the circuit and data converter, together with the input force / torque information, form the basic dataset. To simulate interference in actual measurements (such as circuit thermal noise and mechanical fretting), 1%-5% static random noise is added to the basic data (in this embodiment, Gaussian noise with an amplitude of 3% of the signal is added). 10% of the data ± 5% of the pulse disturbance is randomly selected as outliers, forming a coupled dataset of the noisy multidimensional force sensor. .

[0018] Because the calibration force of a multidimensional force sensor includes both force and torque, there are certain numerical differences. These differences may affect the training effect and efficiency of the neural network. To eliminate the influence of different units on data analysis and improve the training efficiency of the neural network, the data needs to be normalized. The normalization formula is as follows: , in, For the values ​​of the sample data, The minimum value in the sample data. This represents the maximum value in the sample data.

[0019] S2. Design a robust neural network model: such as Figure 3 As shown, the robust neural network model employs a fully connected neural network, including an input layer, a hidden layer, and an output layer. The dimension of the input layer is consistent with the dimension of the force sensor output signal. The hidden layer consists of two fully connected layers, with 64 neurons in the first layer and 32 neurons in the second layer, activated using the ReLU activation function. The output layer has the same dimension as the force / torque dimension and uses linear activation, with a Huber loss function added as a robust loss term to suppress the influence of noise and outliers. The expression for the Huber loss function is as follows: , in, Represents the loss function. This represents the static actual force or torque value. The force or torque value output by the model. The threshold is used when the prediction error is small. The squared error is used, which is sensitive to small errors and ensures the model's fitting accuracy to normal data. When the prediction error is large ( Linear error is used to avoid excessive amplification of the squared error and reduce the model's focus on outliers.

[0020] Thus, the total loss function of the robust neural network model is... for: .

[0021] S3. Train a robust neural network model using a noisy coupled dataset: Divide the coupled dataset of a noisy multidimensional force sensor into a training set and a test set in a 7:3 ratio. When the rate of change of the loss values ​​of the training set and the test set is less than a set threshold (10... -5 When training is interrupted, training should be stopped to avoid overfitting.

[0022] S4. Decouple the real-time output signal of the multi-dimensional force sensor using a trained robust neural network model: Deploy the robust neural network model to the measurement system, and input the pre-processed force sensor output signal into the robust neural network model to output noise-resistant force / torque measurements.

[0023] likeFigures 4-9 As shown, the error box plots of the six-axis force sensor under three decoupling algorithms are presented, where the error calculation formula is: in, Indicates error. For the decoupled force and torque, This represents the actual load applied by the force sensor. The error box plots of the six-axis force sensor under three decoupling algorithms show that: Under this decoupling algorithm, the six-axis force sensor F... x F y and F z ( Figures 4-9 The error of the robust method is significantly lower than that of the least squares method and the traditional MSE neural network. The error advantage of the robust method is even more significant in the torque components Mx, My, and Mz (Figure 7-9), with the box almost completely below the other two methods. This indicates that the robust loss neural network has a smaller mean decoupling error in each dimension.

[0024] Figures 10-15 This is a comparison chart of normal decoupling, robust optimized decoupling, and actual loading. Blue represents the actual loading of the sensor, green represents the force / torque under normal decoupling, and red represents the force / torque under the robust model. The overlap between the blue curve (actual loading) and the red curve (robust decoupling data) is significantly higher than that of the green curve (normal decoupling data). Considering the characteristics of the dataset (1%-5% static noise and 5%-15% outliers), the green curve (normal decoupling) shows significant deviation in some intervals (such as the Fy dimension). Figure 11 200-250 sample index, Fz dimension Figure 12 The red curve (robust decoupling) is less affected by disturbances and its fluctuation amplitude is significantly reduced. This shows that the robust loss function effectively suppresses the interference of noise and outliers on the decoupling results, enabling the model to maintain stable output in complex measurement environments.

[0025] The present invention has been described in detail above with reference to the embodiments, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. All equivalent changes and improvements made in accordance with the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A decoupling method for a multidimensional force sensor based on a robust loss function and a neural network, characterized in that, Includes the following steps: S1. Construct a coupled dataset of a noisy multidimensional force sensor; S2. Design a robust neural network model; S3. Train a robust neural network model using a noisy coupled dataset; S4. Decouple the real-time output signal of the multidimensional force sensor using a trained robust neural network model.

2. The decoupling method for a multidimensional force sensor based on a robust loss function and a neural network according to claim 1, characterized in that: The specific steps of constructing the coupled dataset of the noisy multidimensional force sensor in S1 include: for a multidimensional force sensor, setting m equally spaced measurement points in each dimension; applying a multidimensional force / torque from 0 to full scale to the force sensor under static working conditions without dynamic interference, repeatedly collecting the output signal n times at each point to form a basic dataset; adding 1%-5% static random noise and 5%-15% outliers to the basic data to form the coupled dataset of the noisy multidimensional force sensor.

3. The decoupling method for a multidimensional force sensor based on a robust loss function and a neural network according to claim 1, characterized in that: The robust neural network model designed in S2 adopts a fully connected neural network, including an input layer, a hidden layer, and an output layer. The dimension of the input layer is consistent with the dimension of the force sensor output signal. The hidden layer consists of two fully connected layers, with 64 neurons in the first layer and 32 neurons in the second layer, activated by the ReLU activation function. The dimension of the output layer is consistent with the force / torque dimension, using linear activation, and adding the Huber loss function as a robust loss term to suppress the influence of noise and outliers. The output layer also adds a physical constraint term based on the principle of force balance.

4. The decoupling method for a multidimensional force sensor based on a robust loss function and a neural network according to claim 3, characterized in that: The expression for the Huber loss function is as follows: ， in, Represents the loss function. This represents the static actual force or torque value. The force or torque value output by the model. The threshold value is used.

5. The decoupling method for a multidimensional force sensor based on a robust loss function and a neural network according to claim 1, characterized in that: In step S3, when training a robust neural network model using a noisy coupled dataset, the coupled dataset of the noisy multidimensional force sensor is divided into a training set and a test set. Training stops when the rate of change of the loss values ​​of the training set and the test set is less than a set threshold for several consecutive rounds.

6. The decoupling method for a multidimensional force sensor based on a robust loss function and a neural network according to claim 1, characterized in that: The specific method for real-time decoupling using the trained robust neural network model in S4 is as follows: the robust neural network model is deployed to the measurement system, the force sensor output signal is preprocessed and then input into the robust neural network model, and the output is noise-resistant force / torque measurement value.

7. The decoupling method for a multidimensional force sensor based on a robust loss function and a neural network according to claim 2, characterized in that: In step S1, after forming the coupled dataset of the noisy multidimensional force sensor, the data in the dataset is normalized.

Images