A method and system for controlling the tension of a grain conveyor belt

Abstract

The application belongs to the technical field of tension control, and particularly relates to a kind of grain conveying belt tension control method and system, to solve the technical problem that the control method of prior art cannot consider quick response capability and long-term stability.The control method comprises: S1, based on the obtained motor current and actual tension data, the input vector of recursive RBF network is constructed;S2, the input vector is input into the recursive RBF network, and the target control quantity is generated according to the hidden layer output vector and the output layer weight vector;S3, when the minimum distance between the input vector and all centers is greater than the first threshold, a new center is generated;S4, the penalty term in the cost function of recursive least squares method is determined according to the change amount of the output layer weight vector at adjacent time.The detection method of the application can control the tension under various complex working conditions, and improve the running stability of conveying and the service life of equipment.

Description

Technical Field

[0001] This invention belongs to the technical field of tension control, specifically relating to a method and system for controlling the tension of a grain conveyor belt. Background Technology

[0002] As a core piece of equipment in grain storage, logistics, and processing, the tension of the grain conveyor belt is one of the key parameters ensuring its normal operation. The grain conveying system is a complex, nonlinear, and time-varying system. As a bulk material, grain's flow rate, density, and moisture content, among other factors, will exert nonlinear loads on the conveyor belt. Furthermore, the physical properties of the conveyor belt itself, such as its elastic modulus and coefficient of friction, will slowly change with temperature, humidity, usage time, and degree of wear.

[0003] For the control of tension in grain conveyor belts, existing technologies mainly employ traditional PID control methods and some preliminary intelligent control methods. PID control is widely used due to its simple structure and ease of implementation; however, a PID controller is essentially a linear controller. For a system like grain conveying, which has strong nonlinear and time-varying characteristics, a PID controller with fixed parameters struggles to maintain good control performance under various operating conditions.

[0004] To address this, some studies have attempted to introduce intelligent algorithms from radial basis function neural networks (RBF networks) to leverage their powerful nonlinear approximation capabilities in constructing control models. However, traditional RBF networks typically have fixed network structures, requiring pre-setting parameters such as the number, center, and width of neurons in their hidden layers, which makes it difficult to determine optimal values ​​in practical applications. Although some RBF networks can adjust their parameters online, they still have shortcomings in terms of self-organization of network structure. For example, when facing complex scenarios with both sudden changes and slow parameter drift, it is difficult to balance rapid response capabilities with long-term stability. Summary of the Invention

[0005] This invention provides a method and system for controlling the tension of a grain conveyor belt, in order to solve the technical problem that existing control methods cannot simultaneously achieve rapid response capability and long-term stability.

[0006] In a first aspect, the present invention provides a method for controlling the tension of a grain conveyor belt, comprising: S1. Based on the acquired motor current and actual tension force data, construct the input vector of the recursive RBF network: The input vector includes the mean of the first difference of the motor current within the time window, and the deviation coefficient of the belt material properties obtained by nonlinear mapping of the tension force prediction error and the static load. S2, input the input vector into the recursive RBF network, generate the target control quantity based on the hidden layer output vector and the output layer weight vector, and update the center vector, basis function width and the output layer weight vector of the recursive RBF network based on the control error formed by the tension setpoint and the actual tension after adjustment by the target control quantity. S3, when the minimum distance between the input vector and all centers is greater than the first threshold, an interpolation point is determined on the line connecting the input vector and the nearest center according to the magnitude of the control error, thereby generating a new center; when the distance between any two centers is less than the second threshold, the cosine similarity of their corresponding output layer weight vectors is greater than the third threshold, and the product of their local complexity indices is less than the fourth threshold, the two centers are merged. S4. Determine the penalty term in the cost function of the recursive least squares method based on the change of the output layer weight vector at adjacent time points, and determine the forgetting factor of the penalty term based on the tension difference at adjacent time points. Then, update the output layer weight vector using the recursive least squares method.

[0007] Further, in S1, the input vector for constructing the recursive RBF network includes the mean of the first difference of the motor current within the time window, and the deviation coefficient of the belt material properties obtained by nonlinear mapping of the tension prediction error and the static load, including: exist At time , the input vector Construct as a two-dimensional vector ,in: The mean value of the primary difference of the motor current within the time window is calculated using the following formula: ;in, The length of the time window. for The motor current at any given moment; The deviation coefficient of the material properties of the belt is calculated using the following formula: ;in, For mapping gain coefficients, The preset static load value, for The tension prediction error at time t is calculated using the following formula: ,in, for The actual tension at any given moment for Predicting network tension at any given moment.

[0008] Further, in S2, updating the basis function width includes: exist At time t, the recursive RBF network is at time t. individual centers Local complexity index The calculation formula is: ;in, Center The average distance to its five nearest neighbor centers in Euclidean distance; for Time and the The output layer weights corresponding to each center; The width of the basis functions The update formula is: ;in, This is the proportionality coefficient. To prevent constants with a denominator of zero.

[0009] Furthermore, S3 includes: Generate new centers: if the input vector With the nearest center If the distance is greater than the first threshold, then the new center Generated by interpolation: ;in, To control errors; Merge Centers: If Centers and The spacing between them is less than the second threshold, and the corresponding output layer weight vectors are... and The cosine similarity is greater than the third threshold and the product of their local complexity indices. If the value is less than the fourth threshold, a new merged center is generated. The width of the basis functions of the new center The output layer weight vector of the new center Generate a new center Delete the original center and .

[0010] Furthermore, S4 includes: The cost function for:

[0011] in The coefficient for the penalty term. for The output layer weight vector at time 1. for The actual tension at any given moment for The hidden layer output vector at time step 1. for The forgetting factor of time; Forgetting factor The absolute value of the tension difference between adjacent time points The adjusted calculation formula is as follows: .

[0012] Secondly, the present invention provides a grain conveyor belt tension control system, comprising the following modules: The module is used to acquire motor current and actual tension force data, and construct the input vector of the recursive RBF network. The input vector includes the mean of the first difference of the motor current within the time window, and the deviation coefficient of the belt material properties obtained by nonlinear mapping of tension force prediction error and static load. The generation module is used to input the input vector into the recursive RBF network, generate a target control quantity based on the hidden layer output and the output layer weight vector, and update the center vector, basis function width and output layer weight vector of the network based on the control error formed by the target control quantity and the actual tension force. The optimization module is used to generate a new center by determining an interpolation point on the line connecting the input vector and the nearest center based on the magnitude of the control error when the minimum distance between the input vector and all existing centers is greater than a first threshold; and to merge the two centers when the distance between any two centers is less than a second threshold, the cosine similarity of their corresponding output weights is greater than a third threshold, and the product of their local complexity indices is less than a fourth threshold. The update module is used to determine the penalty term in the cost function of the recursive least squares method based on the change of the output layer weight vector at adjacent time points, and to determine the forgetting factor of the penalty term based on the tension difference at adjacent time points, and to update the output layer weight vector using the recursive least squares method.

[0013] Furthermore, in the construction module, the input vector for constructing the recursive RBF network includes the mean of the first difference of the motor current within the time window, and the deviation coefficient of the belt material properties obtained by nonlinear mapping of the tension prediction error and the static load, including: exist At time , the input vector Construct as a two-dimensional vector ,in: The mean value of the primary difference of the motor current within the time window is calculated using the following formula: ;in, The length of the time window. for The motor current at any given moment; The deviation coefficient of the material properties of the belt is calculated using the following formula: ;in, For mapping gain coefficients, The preset static load value, for The tension prediction error at time t is calculated using the following formula: ,in, for The actual tension at any given moment for Predicting network tension at any given moment.

[0014] Furthermore, in the generation module, updating the basis function width includes: exist At time t, the recursive RBF network is at time t. individual centers Local complexity index The calculation formula is: ;in, Center The average distance to its five nearest neighbor centers in Euclidean distance; for Time and the The output layer weights corresponding to each center; The width of the basis functions The update formula is: ;in, This is the proportionality coefficient. To prevent constants with a denominator of zero.

[0015] Furthermore, the optimization modules include: Generate new centers: if the input vector With the nearest center If the distance is greater than the first threshold, then the new center Generated by interpolation: ;in, To control errors; Merge Centers: If Centers and The spacing between them is less than the second threshold, and the corresponding output layer weight vectors are... and The cosine similarity is greater than the third threshold and the product of their local complexity indices. If the value is less than the fourth threshold, a new merged center is generated. The width of the basis functions of the new center The output layer weight vector of the new center Generate a new center Delete the original center and .

[0016] Furthermore, the update module includes: The cost function for:

[0017] in The coefficient for the penalty term. for The output layer weight vector at time 1. for The actual tension at any given moment for The hidden layer output vector at time step 1. for The forgetting factor of time; Forgetting factor The absolute value of the tension difference between adjacent time points The adjusted calculation formula is as follows: .

[0018] The beneficial effects are as follows: This invention constructs an input vector that can simultaneously reflect both the instantaneous changes in motor load and the long-term deviations in belt material properties. The internal structure of the neural network is optimized based on the distribution and complexity of the operating data. New operating modes are learned by adding neurons, and the structure remains compact by merging redundant neurons. Simultaneously, the adjustment of the neuron's sensing range is linked to the complexity of local data, enabling the characterization of nonlinear relationships. During the generation of the control law, the update of the output weights not only ensures a smooth transition of the control quantity through a penalty term but also utilizes a forgetting mechanism related to the amplitude of tension force changes. This allows for tension force control under various complex operating conditions, improving the operational stability of the conveyor and extending equipment lifespan. Attached Figure Description

[0019] Figure 1 A flowchart for the method of controlling the tension of a grain conveyor belt. Detailed Implementation

[0020] An embodiment of the grain conveyor belt tension control method provided by the present invention: like Figure 1 As shown, a method for controlling the tension of a grain conveyor belt includes: S1. Based on the acquired motor current and actual tension force data, construct the input vector of the recursive RBF network: The input vector includes the mean of the first difference of the motor current within the time window, and the deviation coefficient of the belt material properties obtained by nonlinear mapping of the tension force prediction error and the static load. Specifically, in each control cycle The stator current of the motor is collected by a Hall current sensor, and the actual tension of the conveyor belt is collected by a tension sensor. A length of... Time windows, for example Equals 10. Calculate the difference between the motor current at the current moment and the previous moment; this is the first difference of the motor current. The first component of the input vector. The primary differential current of the motor within the time window The arithmetic mean within.

[0021] In an optional embodiment, in S1, the input vector for constructing the recursive RBF network includes the mean of the first difference of the motor current within a time window, and the deviation coefficient of the belt material properties obtained by nonlinear mapping of the tension prediction error and the static load, including: exist At time , the input vector Construct as a two-dimensional vector ,in: The mean value of the primary difference of the motor current within the time window is calculated using the following formula: ;in, The length of the time window. for The motor current at any given moment; The deviation coefficient of the material properties of the belt is calculated using the following formula: ;in, For mapping gain coefficients, The preset static load value, for The tension prediction error at time t is calculated using the following formula: ,in, for The actual tension at any given moment for Predicting network tension at any given moment.

[0022] Specifically, the first component of the input vector is the local average of the rate of change of the motor current, reflecting the changing trend of the conveyor belt load. For example, in At any given time, 11 motor current values ​​were collected, including the current time. =5.2A, =5.1A, until =4.6A. Calculate the current difference values ​​over these 10 sampling periods, sum these 10 difference values, and then divide by the time window length of 10 to obtain an average value, for example, 0.06. This value is the input component. This value can smooth out noise and capture rapid load changes caused by material buildup or unloading.

[0023] The second component of the input vector, normalized by a sigmoid function, is the tension prediction error, used to calculate the real-time deviation of physical properties of the belt material, such as the elastic modulus. For example, in... At that moment, the actual measured tension force It is 150N, while at the previous moment Network Predicted Tension If the value is 142N, then the current prediction error is... The value is 8N. Assume a preset static load. The mapping gain coefficient is 100N. =5, then the calculation yields The value is approximately 0.6, which is greater than 0.5, indicating that the actual tension response of the belt exceeds the model's expectations, possibly due to elastic changes caused by temperature variations or material fatigue.

[0024] S2, input the input vector into the recursive RBF network, generate the target control quantity based on the hidden layer output vector and the output layer weight vector, and update the center vector, basis function width and output layer weight vector of the recursive RBF network based on the control error formed by the tension setpoint and the actual tension after adjustment by the target control quantity. Specifically, the input vector is fed into the hidden layer of an RBF network with M neurons. For the j-th neuron, its output... The Gaussian function is used to calculate its value as an exp function. The outputs of all M neurons form the hidden layer output vector Φ(k). The target control quantity output by the network is... Let Φ(k) be the inner product of the hidden layer output vector and the output layer weight vector W(k). Calculate the control error between the tension setpoint and the actual tension after adjustment by the target control variable.

[0025] In an optional instance, in S2, updating the basis function width includes: exist At time t, the recursive RBF network is at time t. individual centers Local complexity index The calculation formula is: ;in, Center The average distance to its five nearest neighbor centers in Euclidean distance; for Time and the The output layer weights corresponding to each center; The width of the basis functions The update formula is: ;in, This is the proportionality coefficient. To prevent constants with a denominator of zero.

[0026] For example, for the j-th neuron center in the network, calculate its Euclidean distance to its 5 nearest neighbors, assuming the average distance is obtained. The output layer weights corresponding to this neuron. =-2.0. This is the local complexity index of the neuron. The local complexity index of this neuron is 8.0. =8.0, if the proportionality coefficient If the value is 0.8, then its new basis function width is... It is approximately 0.156. This value is a relatively small width value, which makes the radial basis function of this neuron concentrated. Conversely, if another neuron has a low complexity index, such as 0.5, its basis function width will increase accordingly to approximately 2.49, forming a smoother response with a wider coverage.

[0027] S3, when the minimum distance between the input vector and all existing centers is greater than the first threshold, an interpolation point is determined on the line connecting the input vector and the nearest center according to the magnitude of the control error, thereby generating a new center; when the distance between any two centers is less than the second threshold, the cosine similarity of their corresponding output layer weight vectors is greater than the third threshold, and the product of their local complexity indices is less than the fourth threshold, the two centers are merged. Specifically, at each time k, the Euclidean distance between the current input vector X(k) and all M centers in the network is calculated, and the minimum distance is found. and the corresponding nearest center Set the first threshold. ,like > It is assumed that new pattern data has been input, in X(k) and A new center is generated on the connection between them. Traverse any two centers in the network. and If their distance is less than the second threshold Its corresponding output layer weight vector and The cosine similarity (i.e., the product of their dot products divided by their respective moduli) is greater than the third threshold. Furthermore, both have local complexity indices. × <Fourth Threshold If the two neurons are considered redundant and located in non-critical regions, they are merged. The new center after merging is the arithmetic mean of the two neurons, and the new width and weight are also the arithmetic mean of the two neurons. The original two centers are then deleted.

[0028] In an optional embodiment, S3 includes: Generate new centers: if the input vector With the nearest center If the distance is greater than the first threshold, then the new center Generated by interpolation: ;in, To control errors; Merge Centers: If Centers and The spacing between them is less than the second threshold, and the corresponding output layer weight vectors are... and The cosine similarity is greater than the third threshold and the product of their local complexity indices. If the value is less than the fourth threshold, a new merged center is generated. The width of the basis functions of the new center The output layer weight vector of the new center Generate a new center Delete the original center and .

[0029] Specifically, suppose that at time k, the new input vector X(k) is [0.9, 0.8], and the nearest center in the network is... The values ​​are [0.2, 0.3]. The Euclidean distance between them is 0.86, which is greater than the set first threshold (e.g., 0.8). If the control error... If the value is relatively large, for example, 12, then the interpolation coefficient is close to 1, and the newly generated center... It will be very close to the position of the input vector X(k).

[0030] Network pruning mechanisms are used to remove redundant neurons, keeping the network concise and efficient. For example, a network may have two centers. and Their coordinates are [0.3, 0.4] and [0.32, 0.45], respectively. The distance between them is 0.054, which is less than the second threshold of 0.1. Their corresponding output weights are 1.5 and 1.6, respectively, and the cosine similarity is 0.99, which is greater than the third threshold of 0.9. Their local complexity index product is 0.04, which is less than the fourth threshold of 0.05, indicating that these two neurons are not only geographically close and functionally similar, but also that the complexity of their respective regions is low, indicating redundancy. Therefore, these two centers are merged to generate a new center located at the midpoint of the two, with its width and weights also averaged.

[0031] S4. Determine the penalty term in the cost function of the recursive least squares method based on the change of the output layer weight vector at adjacent time points, and determine the forgetting factor of the penalty term based on the tension difference at adjacent time points. Then, update the output layer weight vector using the recursive least squares method.

[0032] Specifically, the output layer weight vector W(k) is updated using a recursive least squares method with a regularization term and a variable forgetting factor. The forgetting factor is calculated. In one embodiment, the value of the forgetting factor is based on the absolute value of the difference in actual tension between adjacent time points. In one embodiment, the specific formula is as follows: ,in It is 0.999. It is 0.96, and g is a positive coefficient. When the tension changes drastically, Increase Reduce old data and forget old data; conversely, focus on historical data. The update formula is derived from the cost function, which includes a penalty term. This cost function adds a penalty term to the traditional least squares sum of squared errors. Multiply by the weight vector Time and The squared norm of the time difference. The update process is as follows: calculate the gain matrix, and use the gain matrix and control error... Update the weight vector W(k) according to the variable forgetting factor. and penalty coefficient Update the covariance matrix.

[0033] In an optional instance, S4 includes: The cost function for: ;in, The coefficient for the penalty term. for The output layer weight vector at time 1. for The actual tension at any given moment for The hidden layer output vector at time step 1. for The forgetting factor of time; Forgetting factor The absolute value of the tension difference between adjacent time points The adjusted calculation formula is as follows: .

[0034] Specifically, the update of the output layer weights employs an improved recursive least squares method, which adds a weight change penalty term to the cost function. This penalty term is used to limit the output layer weight vector between two adjacent time steps. and The magnitude of the change between them. For example, at a certain moment, a large instantaneous prediction error may occur due to measurement noise. Due to the existence of the penalty term, while minimizing the prediction error, it also suppresses excessive jumps in the weights, making the weight update process smoother and more stable.

[0035] Forgetting factor in this invention It can adjust according to the stability of the working conditions, realizing intelligent control of the learning rate. For example, when the conveyor belt is in a stable operating state, the tension readings at adjacent moments are monitored. and Very close, for example, 120N and 120.5N respectively, the absolute value of their difference The calculated forgetting factor is 0.5. It will be very close to its upper limit of 0.999. The large forgetting factor makes the weight updates slow and stable. Conversely, when the working conditions change abruptly, such as a large amount of material being loaded instantly, the tension force suddenly increases from 120N to 140N. When the value is 20, the calculated forgetting factor will rapidly decrease to near its lower limit of 0.95.

[0036] An embodiment of the grain conveyor belt tension control system provided by the present invention: A tension control system for a grain conveyor belt includes the following modules: The module is used to acquire motor current and actual tension force data, and construct the input vector of the recursive RBF network. The input vector includes the mean of the first difference of the motor current within the time window, and the deviation coefficient of the belt material properties obtained by nonlinear mapping of tension force prediction error and static load. The generation module is used to input the input vector into the recursive RBF network, generate a target control quantity based on the hidden layer output and the output layer weight vector, and update the center vector, basis function width and output layer weight vector of the network based on the control error formed by the target control quantity and the actual tension force. The optimization module is used to generate a new center by determining an interpolation point on the line connecting the input vector and the nearest center based on the magnitude of the control error when the minimum distance between the input vector and all existing centers is greater than a first threshold; and to merge the two centers when the distance between any two centers is less than a second threshold, the cosine similarity of their corresponding output weights is greater than a third threshold, and the product of their local complexity indices is less than a fourth threshold. The update module is used to determine the penalty term in the cost function of the recursive least squares method based on the change of the output layer weight vector at adjacent time points, and to determine the forgetting factor of the penalty term based on the tension difference at adjacent time points, and to update the output layer weight vector using the recursive least squares method.

[0037] Furthermore, in the construction module, the input vector for constructing the recursive RBF network includes the mean of the first difference of the motor current within the time window, and the deviation coefficient of the belt material properties obtained by nonlinear mapping of the tension prediction error and the static load, including: exist At time , the input vector Construct as a two-dimensional vector ,in: The mean value of the primary difference of the motor current within the time window is calculated using the following formula: ;in, The length of the time window. for The motor current at any given moment; The deviation coefficient of the material properties of the belt is calculated using the following formula: ;in, For mapping gain coefficients, The preset static load value, for The tension prediction error at time t is calculated using the following formula: ,in, for The actual tension at any given moment for Predicting network tension at any given moment.

[0038] Furthermore, in the generation module, updating the basis function width includes: exist At time t, the recursive RBF network is at time t. individual centers Local complexity index The calculation formula is: ;in, Center The average distance to its five nearest neighbor centers in Euclidean distance; for Time and the The output layer weights corresponding to each center; The width of the basis functions The update formula is: ;in, This is the proportionality coefficient. To prevent constants with a denominator of zero.

[0039] Furthermore, the optimization modules include: Generate new centers: if the input vector With the nearest center If the distance is greater than the first threshold, then the new center Generated by interpolation: ;in, To control errors; Merge Centers: If Centers and The spacing between them is less than the second threshold, and the corresponding output layer weight vectors are... and The cosine similarity is greater than the third threshold and the product of their local complexity indices. If the value is less than the fourth threshold, a new merged center is generated. The width of the basis functions of the new center The output layer weight vector of the new center Generate a new center Delete the original center and .

[0040] Furthermore, the update module includes: The cost function for:

[0041] in The coefficient for the penalty term. for The output layer weight vector at time 1. for The actual tension at any given moment for The hidden layer output vector at time step 1. for The forgetting factor of time; Forgetting factor The absolute value of the tension difference between adjacent time points The adjusted calculation formula is as follows: .

[0042] In addition, in the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

Claims

1. A method for controlling the tension of a grain conveyor belt, characterized in that, include: S1. Based on the acquired motor current and actual tension force data, construct the input vector of the recursive RBF network: The input vector includes the mean of the first difference of the motor current within the time window, and the deviation coefficient of the belt material properties obtained by nonlinear mapping of the tension force prediction error and the static load. S2, input the input vector into the recursive RBF network, generate the target control quantity based on the hidden layer output vector and the output layer weight vector, and update the center vector, basis function width and the output layer weight vector of the recursive RBF network based on the control error formed by the tension setpoint and the actual tension after adjustment by the target control quantity. S3, when the minimum distance between the input vector and all centers is greater than the first threshold, an interpolation point is determined on the line connecting the input vector and the nearest center according to the magnitude of the control error, thereby generating a new center; when the distance between any two centers is less than the second threshold, the cosine similarity of their corresponding output layer weight vectors is greater than the third threshold, and the product of their local complexity indices is less than the fourth threshold, the two centers are merged. S4. Determine the penalty term in the cost function of the recursive least squares method based on the change of the output layer weight vector at adjacent time points, and determine the forgetting factor of the penalty term based on the tension difference at adjacent time points. Then, update the output layer weight vector using the recursive least squares method.

2. The method for controlling the tension of a grain conveyor belt according to claim 1, characterized in that, In S1, the input vector for constructing the recursive RBF network includes the mean of the first difference of the motor current within the time window, and the deviation coefficient of the belt material properties obtained by nonlinear mapping of the tension prediction error and the static load, including: exist At time , the input vector Construct as a two-dimensional vector ,in: The mean value of the primary difference of the motor current within the time window is calculated using the following formula: ;in, The length of the time window. for The motor current at any given moment; The deviation coefficient of the material properties of the belt is calculated using the following formula: ;in, For mapping gain coefficients, The preset static load value, for The tension prediction error at time t is calculated using the following formula: ,in, for The actual tension at any given moment for Predicting network tension at any given moment.

3. The method for controlling the tension of a grain conveyor belt according to claim 1, characterized in that, In S2, updating the width of the basis function includes: exist At time t, the recursive RBF network is at time t. individual centers Local complexity index The calculation formula is: ;in, Center The average distance to its five nearest neighbor centers in Euclidean distance; for Time and the The output layer weights corresponding to each center; The width of the basis function The update formula is: ;in, This is the proportionality coefficient. To prevent constants with a denominator of zero.

4. The method for controlling the tension of a grain conveyor belt according to claim 1, characterized in that, S3 includes: Generate new centers: if the input vector With the nearest center If the distance is greater than the first threshold, then the new center Generated by interpolation: ;in, To control errors; Merge Centers: If Centers and The spacing between them is less than the second threshold, and the corresponding output layer weight vectors are... and The cosine similarity is greater than the third threshold and the product of their local complexity indices. If the value is less than the fourth threshold, a new merged center is generated. The width of the basis functions of the new center The output layer weight vector of the new center Generate a new center Delete the original center and .

5. The method for controlling the tension of a grain conveyor belt according to any one of claims 1-4, characterized in that, S4 include: The cost function for: ;in, The coefficient for the penalty term. for The output layer weight vector at time 1. for The actual tension at any given moment, for The hidden layer output vector at time step 1. for The forgetting factor of time; Forgetting factor The absolute value of the tension difference between adjacent time points The adjusted calculation formula is as follows: .

6. A grain conveyor belt tension control system, characterized in that, Includes the following modules: The module constructs the input vector of the recursive RBF network based on the acquired motor current and actual tension force data. The input vector includes the mean of the first difference of the motor current within the time window, and the deviation coefficient of the belt material properties obtained by nonlinear mapping of the tension force prediction error and the static load. The generation module inputs the input vector into the recursive RBF network, generates a target control quantity based on the hidden layer output vector and the output layer weight vector, and updates the center vector, basis function width and output layer weight vector of the recursive RBF network based on the control error formed by the tension setpoint and the actual tension after adjustment by the target control quantity. The optimization module determines an interpolation point on the line connecting the input vector and the nearest center based on the magnitude of the control error when the minimum distance between the input vector and all centers is greater than a first threshold, thereby generating a new center; when the distance between any two centers is less than a second threshold, the cosine similarity of their corresponding output layer weight vectors is greater than a third threshold, and the product of their local complexity indices is less than a fourth threshold, the two centers are merged. The update module determines the penalty term in the cost function of the recursive least squares method based on the change of the output layer weight vector at adjacent time points, and determines the forgetting factor of the penalty term based on the tension difference at adjacent time points, and updates the output layer weight vector using the recursive least squares method.

7. The grain conveyor belt tension control system according to claim 6, characterized in that, In the construction module, the input vector for building the recursive RBF network is defined as follows: the input vector includes the mean of the first difference of the motor current within the time window, and the deviation coefficient of the belt material properties obtained by nonlinear mapping of the tension prediction error and the static load, including: exist At time , the input vector Construct as a two-dimensional vector ,in: The mean value of the primary difference of the motor current within the time window is calculated using the following formula: ;in, The length of the time window. for The motor current at any given moment; The deviation coefficient of the material properties of the belt is calculated using the following formula: ;in, For mapping gain coefficients, The preset static load value, for The tension prediction error at time t is calculated using the following formula: ,in, for The actual tension at any given moment, for Predicting network tension at any given moment.

8. The grain conveyor belt tension control system according to claim 6, characterized in that, In the generation module, updating the basis function width includes: exist At time t, the recursive RBF network is at time t. individual centers Local complexity index The calculation formula is: ;in, Center The average distance to its five nearest neighbor centers in Euclidean distance; for Time and the The output layer weights corresponding to each center; The width of the basis functions The update formula is: ;in, This is the proportionality coefficient. To prevent constants with a denominator of zero.

9. The grain conveyor belt tension control system according to claim 6, characterized in that, The optimization modules include: Generate new centers: if the input vector With the nearest center If the distance is greater than the first threshold, then the new center Generated by interpolation: ;in, To control errors; Merge Centers: If Centers and The spacing between them is less than the second threshold, and the corresponding output layer weight vectors are... and The cosine similarity is greater than the third threshold and the product of their local complexity indices. If the value is less than the fourth threshold, a new merged center is generated. The width of the basis functions of the new center The output layer weight vector of the new center Generate a new center Delete the original center and .

10. The grain conveyor belt tension control system according to any one of claims 6-9, characterized in that, The updated modules include: The cost function for: ;in, The coefficient for the penalty term. for The output layer weight vector at time 1. for The actual tension at any given moment, for The hidden layer output vector at time step 1. for The forgetting factor of time; Forgetting factor The absolute value of the tension difference between adjacent time points The adjusted calculation formula is as follows: .

Images