An interference estimation method for an unmanned autonomous system based on bionic perception information fusion

Abstract

The present invention belongs to the field of unmanned autonomous system control, and particularly relates to a method for estimating interference of an unmanned autonomous system based on bionic perception information fusion, comprising the following steps: S1, building a neural network combining a gated recurrent unit and a multi-layer perceptron module to realize the prediction of interference information; S2, training the neural network model by means of knowledge distillation to reduce the complexity of the model and improve the generalization ability of the model; S3, combining the interference derivative information output by the neural network with a classical interference observer architecture to realize the accurate estimation of interference. The method for estimating interference of an unmanned autonomous system based on bionic perception information fusion proposed in the present invention is not only applicable to external force interference, but also applicable to external torque interference, and can realize the real-time and accurate estimation of external interference of an unmanned autonomous system in an unstructured environment, ensuring its high-precision operation control ability.

Description

Technical Field

[0001] The present invention belongs to the field of unmanned autonomous system control, and particularly relates to a method for estimating interference of an unmanned autonomous system based on bionic perception information fusion, which is applicable to an unmanned autonomous control system that needs to perform anti-interference safety operations in an unstructured environment. Background Art

[0002] In recent years, unmanned autonomous systems typified by rotor unmanned aerial vehicles, robotic dogs, and underwater unmanned vehicles have been deeply and widely applied in fields such as disaster warning, geological exploration, and emergency rescue. The application safety of unmanned autonomous systems has also received unprecedented attention. However, with the continuous enrichment of the application scenarios of unmanned autonomous systems, unknown interferences in various unstructured environments have posed severe challenges to the safe operation of unmanned autonomous systems. There are various interferences in unstructured environments, such as wind interference, rain interference, physical contact force, and sea wave fluctuation interference. These interferences are difficult to finely physically model and are often coupled with the state information of the unmanned autonomous system itself, making accurate interference estimation extremely challenging.

[0003] The research on methods related to interference estimation of unmanned autonomous systems has been relatively mature. Chinese Patent Application CN202411620235.0 proposes an interference observation method for small rotor unmanned aerial vehicles, but the interference observer used therein fails to effectively process the interference derivative information, resulting in delays and errors in interference estimation; Chinese Patent Application CN202010621554.9 proposes an interference learning network structure based on long short-term memory for the selection of anti-interference frequency points, but the direct use of neural networks therein places high requirements on the generalization ability of the network and it is difficult to ensure system stability. There are also some scientific researchers who have studied the estimation algorithms of external interferences from aspects such as interference observer design and interference learning. However, these works have not considered the problems that it is difficult for classical nonlinear interference observers to obtain external interference derivative information, resulting in interference estimation errors and delays, and the limited generalization ability of classical neural networks when tested in non-training datasets, resulting in estimation errors. Summary of the Invention

[0004] To overcome the deficiencies of existing research content and methods, for unmanned autonomous systems, the present invention provides a method for estimating interference of an unmanned autonomous system based on bionic perception information fusion, which solves the problem of reduced control accuracy caused by the difficulty of accurately estimating external interference in an unstructured environment for unmanned autonomous systems.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for estimating interference of an unmanned autonomous system based on bionic perception information fusion, comprising the following steps:

[0007] S1. Build a neural network combining a gated recurrent unit and a multi-layer perceptron module to predict interference information;

[0008] S2. Train the neural network model with knowledge distillation to reduce the complexity of the model and improve its generalization ability;

[0009] S3. Combine the interference derivative information output by the neural network with a classical interference observer architecture to achieve accurate estimation of interference.

[0010] Furthermore, step S1 is specifically as follows:

[0011] Build a gated recurrent unit as a neural network module for processing time-series data to encode the time-series features in the input data, and then introduce a multi-layer perceptron module to decode the encoded time-series features, so as to obtain the derivative estimation information of the external force interference and improve the overall performance and generalization of the model.

[0012] The gated recurrent unit contains two internal processing units, which are the reset gate and the update gate ; among them, the reset gate controls the influence of the previous hidden state on the current hidden state, and the update gate determines the amount of information inherited by the current hidden state from the previous hidden state; the specific expressions of the reset gate and the update gate are:

[0013] ,

[0014] ,

[0015] where, represents the input at the current time step, represents the hidden state at the previous moment, represents the trainable weight parameter, represents the Sigmoid activation function, represents the input data of the function, The specific expression of

[0016] ,

[0017] The update step of the current hidden state is as follows:

[0018] ,

[0019] ,

[0020] where, Represents an intermediate variable, represents a trainable weight parameter, represents element-wise multiplication, and tanh represents the hyperbolic tangent activation function tanh , and the specific expression is:

[0021] ,

[0022] The internal data processing method of the constructed multi-layer perceptron module is as follows:

[0023] ,

[0024] Among them, and represent the weight matrix and the bias term respectively, represents the output of the th layer, represents the output of the th layer. Finally, the neural network module applied in combination with the disturbance observer corresponds to , represents the activation function, and the specific expression is:

[0025] .

[0026] Furthermore, the specific steps of step S2 are as follows:

[0027] S2.1, based on the neural network model constructed in S1, design the input information of the neural network model;

[0028] The input information of the neural network model at the current moment includes the selected state information of the unmanned autonomous system from time to time, that is , where represents the state information of the drone at the current moment, respectively represent the state information of the drone at time, the th time, and the th time; the selected state data information of the unmanned autonomous system , where respectively represent the roll angle, pitch angle, and yaw angle of the unmanned autonomous system, represents the movement speed of the unmanned autonomous system on the x, y, and z axes, represents the expected total lift value of the unmanned autonomous system;

[0029] S2.2. Use the training set data to train a complex and powerful teacher model, which can effectively capture the non-linear mapping relationship between input and output information; with the help of knowledge distillation technology, transfer the rich and refined output information provided by the teacher model to a smaller and structurally simplified student model, and the student model learns more valuable features by imitating the output of the teacher model.

[0030] Furthermore, design the loss function used when training the student model , the output of the student model and the true label The difference penalty term between and the output of the student model and the output of the teacher model The difference penalty term between are both considered in the loss function, and combined with the application scenario of the neural network, a regularization term to improve the generalization ability of the model and a sign penalty term to ensure the accuracy of the prediction trend are added. The loss function is as follows:

[0031] ,

[0032] where is a dynamic adjustment parameter, which can be adjusted dynamically according to the training effect of the model, , is a constant value, representing the weight of each penalty term in the loss function;

[0033] ,

[0034] ,

[0035] ,

[0036] ,

[0037] where represents the number of time steps included in the time series data to be processed, represents the current time corresponding to the data being processed, represents the th weight parameter of the neural network model, represents the total number of weight parameters; represents the Euclidean norm of, The function is a common neural network hidden state activation function, and the specific expression is:

[0038] ,

[0039] Among them, represents the input data, represents comparison with , and selects the larger value among them.

[0040] Furthermore, the specific steps of step S3 are as follows: Incorporate the derivative information of the external force disturbance estimated by the neural network into the disturbance observer framework to improve the error and delay of the disturbance estimation in the traditional method;

[0041] The expression of the disturbance observer based on bionic perception information fusion is as follows:

[0042] ,

[0043] Among them, represents the derivative information of the disturbance estimated value, represents the derivative estimation value of the disturbance output by the neural network, represents the constant gain of the disturbance observer, represents the disturbance estimated value, represents the true value of the external force disturbance received by the unmanned autonomous system; Since in an unstructured environment it is difficult to obtain, it can be obtained through the dynamic derivation of the unmanned autonomous system;

[0044] The position control dynamic equation of the rotary-wing UAV is:

[0045] ,

[0046] Among them, represents the mass of the UAV, represents the acceleration of the unmanned autonomous system, represents the desired acceleration output by the controller of the unmanned autonomous system, represents the gravity received by the unmanned autonomous system, represents the true value of the external force disturbance received by the unmanned autonomous system;

[0047] Finally, based on the position control dynamic equation of the rotary-wing UAV, the designed structure of the disturbance observer is:

[0048] .

[0049] However, since the acceleration information obtained by the airborne accelerometer has large noise, the designed disturbance observer directly uses the acceleration information as output, which will cause the disturbance observer to estimate the disturbance with large noise and error. As a result, the disturbance estimate value is difficult to be directly used by the controller to suppress and compensate the disturbance. Based on this problem, the noise influence of acceleration information can be avoided by changing the disturbance observer structure, as follows.

[0050] Furthermore, the step S3 further includes the following steps:

[0051] Based on the designed disturbance observer structure, it is proposed to use additional variables , its derivative satisfy:

[0052] ,

[0053] Then we get:

[0054] ,

[0055] in, is the speed of translational motion of the unmanned autonomous system;

[0056] The original disturbance observer structure can be rewritten as:

[0057] ,

[0058] in, is the total lift generated by the unmanned autonomous system;

[0059] Therefore, the interference can be estimated with the help of speed information with less noise, thus avoiding the introduction of noise.

[0060] It can be shown that even if the estimated value of the disturbance derivative output by the neural network deviates from the estimated value of the true disturbance derivative ,Right now ,

[0061] when When the two-norm is bounded, the disturbance estimation error of the disturbance observer can still converge to a bounded range.

[0062] The beneficial effects of the present invention are:

[0063] (1) First, compared with the classic disturbance observer estimation method and the traditional disturbance learning method, in order to address the problem of limited available information in unstructured environments, it is proposed to use the state information and control input of the unmanned autonomous system over a period of time and at the current moment to effectively estimate the external disturbance derivative information.

[0064] (2) Secondly, a neural network combining a gated recurrent unit and a multi-layer perceptron module is built to predict interference information.

[0065] (3) Subsequently, since the complexity of the trained neural network model is relatively high and the computing resources required for operation are also large, the present invention proposes to adopt a model compression technology of knowledge distillation to significantly reduce the dimension of the neural network and reduce its computing consumption.

[0066] (4) Finally, the external interference derivative information obtained from the neural network is combined with a classical non-linear interference observer to achieve effective complementary enhancement and improve the accuracy of external interference estimation.

[0067] In summary, the present invention can achieve real-time and accurate estimation of external interference of an unmanned autonomous system in an unstructured environment, and ensure its high-precision operation control ability. Description of the Drawings

[0068] Figure 1 It is a flowchart of an interference estimation method for an unmanned autonomous system based on bionic perception information fusion according to the present invention. Specific Embodiments

[0069] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0070] When an unmanned autonomous system operates in an unstructured environment, it may encounter external force interference and external torque interference simultaneously. The interference estimation method for an unmanned autonomous system based on bionic perception information fusion proposed in the present invention is applicable not only to external force interference but also to external torque interference. To avoid repetition, the present invention takes an unmanned autonomous system as the application target and only considers realizing real-time and accurate estimation of external force interference.

[0071] As Figure 1 shown, an interference estimation method for an unmanned autonomous system based on bionic perception information fusion according to the present invention specifically includes the following steps:

[0072] S1, build a neural network combining a gated recurrent unit and a multi-layer perceptron module to predict interference information;

[0073] S2, train the neural network model by means of knowledge distillation to reduce the complexity of the model and improve the generalization ability of the model;

[0074] S3. Combine the interference derivative information output by the neural network with the classical interference observer architecture to achieve accurate estimation of interference.

[0075] Further, step S1 is specifically as follows:

[0076] Build a gated recurrent unit as a neural network module for processing time-series data to encode the time-series features in the input data, and then introduce a multi-layer perceptron module to decode the encoded time-series features, so as to obtain the derivative estimation information of the external force interference and improve the overall performance and generalization of the model.

[0077] The gated recurrent unit includes two internal processing units, and the two internal processing units are respectively a reset gate and an update gate ; among them, the reset gate controls the influence of the previous hidden state on the current hidden state, and the update gate determines the amount of information inherited by the current hidden state from the previous hidden state; the specific expressions of the reset gate and the update gate are:

[0078] ,

[0079] ,

[0080] where, represents the input at the current time step, represents the hidden state at the previous time step, represents the trainable weight parameter, represents the Sigmoid activation function, represents the input data of the function, The specific expression of

[0081] is:

[0082] The current hidden state is updated as follows:

[0083] ,

[0084] ,

[0085] where, represents the intermediate variable, represents the trainable weight parameter, represents element-wise multiplication, and tanh represents the hyperbolic tangent activation function tanh The specific expression is:

[0086] ,

[0087] The internal data processing method of the constructed multi-layer perceptron module is as follows:

[0088] ,

[0089] Among them, and represent the weight matrix and the bias term respectively, represents the output of the th layer, represents the output of the th layer. Finally, the neural network module applied in combination with the disturbance observer corresponds to , represents the activation function, The specific expression is:

[0090] .

[0091] Furthermore, the specific steps of step S2 are as follows:

[0092] S2.1, based on the neural network model constructed in S1, design the input information of the neural network model;

[0093] The input information of the neural network model at the current moment includes the selected state information of the unmanned autonomous system from the moment to the moment, that is , where represents the state information of the UAV at the current moment, respectively represent the UAV moment, the th moment and the th moment state information; the selected state data information of the unmanned autonomous system , where respectively represent the roll angle, pitch angle and yaw angle of the unmanned autonomous system, represents the movement speed of the unmanned autonomous system on the x, y, and z axes, represents the expected total lift value of the unmanned autonomous system;

[0094] S2.2, use the training set data to train a complex and powerful teacher model, which can effectively capture the non-linear mapping relationship between the input and output information; with the help of knowledge distillation technology, transfer the rich and refined output information provided by the teacher model to a smaller and structurally simplified student model, and the student model learns more valuable features by imitating the output of the teacher model.

[0095] Furthermore, design the loss function used in training the student model and incorporate the penalty term for the difference between the output of the student model and the ground truth label as well as the penalty term for the difference between the output of the student model and the output of the teacher model into the loss function. Considering the application scenario of the neural network, a regularization term for improving the generalization ability of the model and a sign penalty term for ensuring the accuracy of the prediction trend are added. The loss function is as follows:

[0096] ,

[0097] where is a dynamic adjustment parameter that can be adjusted dynamically according to the training effect of the model, , is a constant value representing the weight of each penalty term in the loss function;

[0098] ,

[0099] ,

[0100] ,

[0101] ,

[0102] where represents the number of time steps included in the time series data to be processed, represents the current time corresponding to the data being processed, represents the th weight parameter of the neural network model, represents the total number of weight parameters; represents 's Euclidean norm, The function is a common activation function for the hidden state of the neural network, and the specific expression is:

[0103] ,

[0104] where represents the input data, represents comparing with and selecting the larger value.

[0105] Further, the specific steps of step S3 are as follows: Incorporate the derivative information of the external force disturbance estimated by the neural network into the disturbance observer framework to improve the error and delay of the disturbance estimation in the traditional method;

[0106] The expression of the disturbance observer based on bionic perception information fusion is as follows:

[0107] ,

[0108] where, represents the derivative information of the disturbance estimation value, represents the estimated value of the disturbance derivative output by the neural network, represents the constant gain of the disturbance observer, represents the disturbance estimation value, represents the true value of the external force disturbance received by the unmanned autonomous system; Since in an unstructured environment is difficult to obtain, it can be derived through the dynamics of the unmanned autonomous system;

[0109] The position control dynamics equation of the rotary-wing UAV is:

[0110] ,

[0111] where, represents the mass of the UAV, represents the acceleration of the unmanned autonomous system, represents the desired acceleration output by the controller of the unmanned autonomous system, represents the gravity received by the unmanned autonomous system, represents the true value of the external force disturbance received by the unmanned autonomous system;

[0112] Finally, based on the position control dynamics equation of the rotary-wing UAV, the designed disturbance observer structure is:

[0113] .

[0114] However, due to the large noise in the acceleration information obtained by the on-board accelerometer, using this acceleration information directly as the output of the designed disturbance observer will result in large noise and errors in the disturbance estimation of the disturbance observer. As a result, it is difficult for the disturbance estimation value to be directly used by the controller for disturbance suppression and compensation. Based on this problem, the noise influence of the acceleration information can be avoided through the transformation of the disturbance observer structure, as follows.

[0115] Further, step S3 also includes the following steps:

[0116] Based on the designed disturbance observer structure, an additional variable is proposed, and its derivative Satisfy:

[0117] ,

[0118] Furthermore, obtain:

[0119] ,

[0120] wherein, is the velocity of the translational motion of the unmanned autonomous system;

[0121] Then the structure of the original disturbance observer can be rewritten as:

[0122] ,

[0123] wherein, is the total lift force generated by the unmanned autonomous system;

[0124] Thus, the estimation of the disturbance is realized by means of the velocity information with less noise, avoiding the introduction of noise.

[0125] It can be proved that even if there is a deviation between the estimated value of the disturbance derivative output by the neural network and the true estimated value of the disturbance derivative , that is ,

[0126] When satisfies the bounded two-norm, the disturbance estimation error of this disturbance observer can still converge to a bounded range.

[0127] Although the above description of the illustrative specific embodiments of the present invention is for the convenience of those skilled in the art of the present technology to understand the present invention, and it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

Claims

1. A method for estimating interference in an unmanned autonomous system based on bionic perception information fusion, characterized in that, It includes the following steps: S1. Build a neural network combining a gated recurrent unit and a multi-layer perceptron module to predict interference information; S2. Train the neural network model with knowledge distillation to reduce the complexity of the model and improve its generalization ability; S3. Combine the interference derivative information output by the neural network with a classical interference observer architecture to achieve accurate estimation of interference; The specific content of S3 is as follows: Incorporate the derivative information of the external force interference estimated by the neural network into the interference observer framework to improve the error and delay of interference estimation in traditional methods; The expression of the interference observer based on bionic perception information fusion is as follows: ， Among them, represents the derivative information of the interference estimation value, represents the interference derivative estimation value output by the neural network, represents the constant gain of the interference observer, represents the interference estimation value, represents the true value of the external force interference received by the unmanned autonomous system; since in an unstructured environment it is difficult to obtain, so it can be derived from the dynamics of the unmanned autonomous system; The position control dynamic equation of the rotor UAV is: ， Among them, represents the mass of the drone, represents the acceleration of the unmanned autonomous system, represents the desired acceleration output by the controller of the unmanned autonomous system, represents the gravity received by the unmanned autonomous system, represents the true value of the external force interference received by the unmanned autonomous system; Finally, based on the position control dynamic equation of the rotor UAV, the designed interference observer structure is: ； S3 also includes the following steps: Based on the designed disturbance observer structure, an additional variable is proposed , whose derivative satisfies: ， Furthermore, we get: ， Among them, is the velocity of the translational motion of the unmanned autonomous system; Then the original interference observer structure can be rewritten as: ， Among them, is the total lift generated by the unmanned autonomous system; Thus, estimate the interference by using the velocity information with less noise to avoid noise introduction.

2. The interference estimation method for the unmanned autonomous system based on bionic perception information fusion according to claim 1, wherein The specific content of S1 is as follows: Build a gated recurrent unit as a neural network module for processing time-series data to encode the time-series features in the input data, and then introduce a multi-layer perceptron module to decode the encoded time-series features, so as to obtain the derivative estimation information of the external force interference and improve the overall performance and generalization of the model.

3. The interference estimation method for an unmanned autonomous system based on bionic perception information fusion according to claim 2, wherein The gated recurrent unit includes two internal processing units, namely the reset gate and the update gate ; among them, the reset gate controls the influence of the previous hidden state on the current hidden state, and the update gate determines the amount of information inherited by the current hidden state from the previous hidden state.

4. The interference estimation method for an unmanned autonomous system based on bionic perception information fusion according to claim 1, wherein The specific content of S2 is as follows: S2.

1. Based on the neural network model built in S1, design the input information of the neural network model; The input information of the neural network model at the current moment contains the selected state information of the unmanned autonomous system from moment to moment, that is , where represents the state information of the UAV at the current moment, respectively represent the UAV moment, the th moment and the th moment of the state information; the selected state data information of the unmanned autonomous system , where respectively represent the roll angle, pitch angle and yaw angle of the unmanned autonomous system, represents the movement speed of the unmanned autonomous system on the x, y, and z axes, represents the expected total lift of the unmanned autonomous system; S2.

2. Use the training set data to train a teacher model, which can effectively capture the non-linear mapping relationship between input and output information; With the knowledge distillation technology, transfer the output information provided by the teacher model to a student model, and the student model learns more valuable features by imitating the output of the teacher model.

5. The interference estimation method for an unmanned autonomous system based on bionic perception information fusion according to claim 4, wherein The loss function used in the training of the student model , the output of the student model and the true label The difference penalty term between them And the difference penalty term between the output of the student model and the output of the teacher model between At the same time, it is considered into the loss function, and combined with the application scenario of the neural network, a regularization term to improve the generalization ability of the model is added And a sign penalty term to ensure the accuracy of the prediction trend , the loss function is as follows: ， Among them, is a dynamic adjustment parameter and can be dynamically adjusted according to the model training effect, , is a constant value, representing the weight of each penalty term in the loss function; ， ， ， ， Among them, represents the number of time steps included in the time series data to be processed, represents the moment corresponding to the currently processed data, represents the th weight parameter of the neural network model, represents the total number of weight parameters; represents the Euclidean norm of, The function is a common activation function for the hidden state of the neural network, and the specific expression is: ， Among them, represents the input data, represents comparison with , and selects the larger value among them.

Images