CN-121994633-A - Reversing loading force impact inhibition method based on Fourier neural operator

CN121994633ACN 121994633 ACN121994633 ACN 121994633ACN-121994633-A

Abstract

The invention provides a reversing loading force impact suppression method based on a Fourier nerve operator, which aims at impact of an electrohydraulic servo loading platform, a servo motor lead screw loading device and a material fatigue testing machine when loading force instructions are switched positively and negatively, and is used for carrying out time alignment sampling on the loading force instructions, feedback, actuator speed and previous period control input in a fixed control period, judging a reversing starting point and an adjacent period, intercepting historical fragments, taking the fragments as input, generating a reversing impact disturbance prediction sequence aligned with the adjacent period through frequency domain characteristics, correlating operator with time reduction by the Fourier nerve operator, estimating total disturbance by an expansion state observer, linearly fusing the total disturbance with the prediction disturbance according to a fusion coefficient in the adjacent period, driving active disturbance rejection control to calculate control input, and adopting total disturbance closed-loop control in a non-adjacent period.

Inventors

YANG WEN
JIANG RENHUA
LI YUTE
JIANG TONGMING
YANG CHANGQUN
HUAN JUNJUN
NIU DAODONG
LIN YUANWEN
SHI BAOHU
QIU SHUI
LI LINGBO

Assignees

国家石油天然气管网集团有限公司
国家石油天然气管网集团有限公司华南分公司
扬州恒春电子有限公司

Dates

Publication Date: 20260508
Application Date: 20260122

Claims (9)

1. The reversing loading force impact suppression method based on the Fourier neural operator is characterized by comprising the following steps of: S1, acquiring a loading force instruction, loading force feedback, actuator speed and a control input instruction of a previous control period in an electrohydraulic servo loading table, a servo motor lead screw loading device or a material fatigue testing machine to form a current loading sampling sequence; S2, judging that the loading force instruction is switched in the positive and negative directions and the absolute value of the speed of the actuator is smaller than a zero-speed judgment threshold value according to the current loading sampling sequence, and determining the starting point moment of the reversing and the duration of the reversing adjacent time period to obtain reversing adjacent time period parameters; S3, intercepting a preset history time before a reversing starting point time from a current loading sampling sequence according to reversing adjacent time period parameters to form a reversing adjacent state segment; S4, inputting the reversing adjacent state segments into a Fourier operator for calculation, wherein the Fourier operator sequentially comprises a frequency domain feature generation layer, a Fourier operator association layer and a time waveform reduction layer, the frequency domain feature generation layer converts the reversing adjacent state segments into frequency domain representation for describing reversing instantaneous change, the Fourier operator association layer carries out association calculation on different frequency components in the frequency domain representation to describe disturbance forms corresponding to static friction debonding and friction direction overturning, and the time waveform reduction layer reduces association calculation results into reversing impact disturbance prediction sequences consistent with duration of reversing adjacent time periods to obtain reversing impact disturbance prediction sequences; S5, controlling an input instruction based on loading force feedback and a previous control period, estimating total disturbance by adopting an extended state observer, taking out a corresponding disturbance value from a reversing impact disturbance prediction sequence according to the position of the current moment in a reversing adjacent period when the current moment meets the reversing adjacent period condition defined by reversing adjacent period parameters, fusing the corresponding disturbance value with the total disturbance estimation to obtain impact compensation disturbance quantity, calculating a control input instruction based on the loading force instruction, the loading force feedback and the impact compensation disturbance quantity, and calculating the control input instruction based on the loading force instruction, the loading force feedback and the total disturbance estimation when the current moment does not meet the reversing adjacent period condition; s6, applying a control input instruction to the actuator to drive the loading mechanism to generate loading force, acquiring updated loading force feedback and updated actuator speed, and forming a loading sampling sequence of the next control period together with the loading force instruction and the control input instruction.
2. The method for suppressing reverse loading force impact based on fourier nerve operator according to claim 1, wherein S1 specifically comprises: Setting control periods, triggering sampling by a control period timer, acquiring a loading force instruction in each control period, acquiring loading force feedback by a force sensor, acquiring the speed of an actuator by an encoder, and enabling the time deviation between the sampling moment of the loading force feedback and the sampling moment of the speed of the actuator to be smaller than a time alignment threshold; reading a control input instruction of a previous control period when the control period starts, and associating the control input instruction of the previous control period with a loading force instruction, loading force feedback and actuator speed corresponding to the control period according to the same control period mark to form a sampling point; writing sampling points into a sliding window according to a control period sequence, wherein the sliding window covers continuous sampling points with the length not less than a preset history time length so as to meet the intercepting length requirement of a subsequent reversing adjacent state segment; and forming the sampling points in the sliding window into a current loading sampling sequence according to a time sequence, and outputting the current loading sampling sequence to a reversing judgment and reversing impact disturbance prediction sequence to generate a link.
3. The method for suppressing reverse loading force impact based on fourier nerve operator according to claim 1, wherein S2 specifically comprises: The direction judgment is carried out on the loading force instruction in the current loading sampling sequence according to the control cycle sequence, and the loading force instruction in the current control cycle and the loading force instruction in the last control cycle are obtained by comparing the loading force instruction with the instruction direction judgment threshold; When the loading force instruction direction is inconsistent with the loading force instruction direction of the previous control period, the speed of the current control period executor is obtained and compared with a zero speed judgment threshold value, and a reversing trigger mark is generated when the speed amplitude is smaller than the zero speed judgment threshold value; In a control period generated by the reversing trigger mark, determining the sampling time of the control period as a reversing starting time, and executing reversing adjacent period tracking on the subsequent control period from the reversing starting time; and determining the duration of the commutation adjacent time period according to the first moment when the speed amplitude exceeds the commutation exit speed threshold value in the commutation adjacent time period tracking, and determining the duration of the commutation adjacent time period by using the duration threshold value of the commutation adjacent time period under the condition that the speed amplitude does not exceed the commutation exit speed threshold value, so as to obtain the parameters of the commutation adjacent time period including the moment of the start of the commutation and the duration of the commutation adjacent time period.
4. The method for suppressing reverse loading force impact based on fourier nerve operator according to claim 1, wherein S3 specifically comprises: According to the reversing starting point time in the reversing adjacent period parameter, and combining the control period to determine the historical sampling point number corresponding to the preset historical time length; positioning a reversing starting point sampling point corresponding to the reversing starting point time in the current loading sampling sequence, and taking the sampling point of the continuous historical sampling point before the reversing starting point sampling point as a intercepting range to enable the intercepting range to cover the continuous sampling point before the loading force instruction is switched in the positive and negative directions; when the number of sampling points before the reversing starting point sampling points in the current loading sampling sequence is smaller than the number of historical sampling points, all sampling points before the reversing starting point sampling points are determined to be starting sampling points, and the interception range is kept continuous; and forming a reversing adjacent state segment from a starting sampling point to a reversing starting sampling point according to a time sequence, and keeping the channel sequence of three numerical characteristics of a loading force instruction, loading force feedback and an actuator speed by each sampling point so that the reversing adjacent state segment meets the point-by-point linear mapping input requirement of a frequency domain characteristic generation layer.
5. The method for suppressing reverse loading force impact based on fourier nerve operator according to claim 1, wherein S4 specifically comprises: Taking the reversing adjacent state segment as the input of a Fourier neural operator, generating a speed direction change characteristic based on the speed of an actuator on the input side, generating an instruction direction change characteristic based on a loading force instruction, and jointly participating the speed direction change characteristic, the instruction direction change characteristic and the three-way characteristic of the reversing adjacent state segment into subsequent mapping calculation; The frequency domain feature generation layer is adopted to conduct point-by-point linear mapping on the reversing adjacent state segments, and sixty-four neuron groups are used for conducting weighted summation on loading force instructions, loading force feedback and actuator speed and conducting nonlinear transformation to obtain sixty-four-channel internal representation arranged in time sequence; Generating a frequency domain representation along the time direction based on the sixty-four channel internal representation, wherein the frequency domain representation is expressed by adopting a real part channel and an imaginary part channel and is kept as a frequency domain structure, so that the frequency domain representation carries information of changing the speed and the amplitude of change before and after; Inputting the frequency domain representation into a Fourier operator association layer and sequentially passing through multi-stage Fourier association units, wherein each stage Fourier association unit sets a leachable weight for each frequency position, and performs weighted combination on sixty four channels of the frequency position to obtain a frequency domain response; Performing channel mixing calculation on the frequency domain response, performing inter-channel linear combination on each frequency position by adopting sixty-four times sixty-four scale weight connection, performing nonlinear transformation, and outputting the updated frequency domain representation as the input of a next-stage Fourier association unit until an association calculation result is obtained; inputting the correlation calculation result into a time waveform reduction layer for performing time waveform reduction, determining the predicted sampling point number according to the duration time and the control period of the commutation adjacent time period in the commutation adjacent time period parameter, and generating a time sequence representation consistent with the predicted sampling point number; And performing point-by-point linear mapping compression on the time sequence representation to obtain a single-channel disturbance representation, mapping sixty-four channels into sixteen channels by adopting two-stage fully-connected neuron groups, mapping the sixteen channels into a single channel, outputting a reversing impact disturbance prediction sequence according to time sequence, and outputting no control input instruction.
6. The method for suppressing reverse loading force impact based on a fourier nerve operator according to claim 5, wherein a reverse adjacent state segment is used as an input of the fourier nerve operator, a speed direction change feature is generated on the input side based on an actuator speed, an instruction direction change feature is generated based on a loading force instruction, the speed direction change feature and the instruction direction change feature are used together with three channel features of the reverse adjacent state segment for subsequent mapping calculation, the speed direction change feature and the instruction direction change feature are obtained by calculation of a change function, wherein the change function is specifically: ; Wherein, the Is the first The speed direction change characteristics of the individual sample points, Is the first The instruction direction change characteristics of the individual sample points, Is the first The actuator speed of the individual sample points, Is the first The actuator speed of the individual sample points, Is the first A loading force instruction for a single sample point, Is the first A loading force instruction for a single sample point, A threshold value is determined for the zero speed, A threshold value is determined for the direction of the command, To indicate a function and to take a value of one when the condition in brackets is satisfied and zero when the condition is not satisfied, for The first sampling point of (2), the controller will And (3) with Set to zero and recursively calculate in the manner described starting from the second sample point, the controller will And (3) with With the same sampling point 、、 The method is used for participating in the subsequent mapping calculation together, so that the point-by-point linear mapping obtains a computable input difference source at sampling points of the switching of the loading force instruction direction and the overturning of the speed direction of the actuator.
7. The method for suppressing reverse loading force impact based on fourier nerve operator according to claim 1, wherein S5 specifically comprises: Taking loading force feedback and a control input instruction of the previous control period as the input of the extended state observer, executing observer state update on the extended state observer and outputting total disturbance estimation; According to the commutation starting point time and the commutation adjacent time period duration in the commutation adjacent time period parameter, and combining the current time, judging whether the current time meets the commutation adjacent time period condition; when the current moment meets the condition of the commutation adjacent time period, determining the position of the current moment in the commutation adjacent time period according to the time difference between the current moment and the commutation starting point moment and combining the control period, and determining the predicted sampling point number according to the duration of the commutation adjacent time period and the control period; Selecting a corresponding disturbance value aligned with the current position from the reversing impact disturbance prediction sequence based on the position of the current moment in a reversing adjacent period, wherein the corresponding disturbance value is a single-channel disturbance prediction value of the reversing impact disturbance prediction sequence; Performing linear fusion on the total disturbance estimation and the corresponding disturbance value by adopting a preset impact fusion coefficient to obtain impact compensation disturbance quantity, wherein the impact compensation disturbance quantity is used as disturbance compensation input of active disturbance rejection control; when the current moment meets the condition of the reversing adjacent period, generating a loading force error by feeding back a loading force instruction and a loading force, and calculating a control input instruction based on the loading force error, the impact compensation disturbance quantity and the control input instruction of the previous control period; And when the current moment does not meet the condition of the reversing adjacent period, generating a loading force error by feeding back the loading force instruction and the loading force, and calculating a control input instruction based on the loading force error, the total disturbance estimation and the control input instruction of the previous control period.
8. The method for suppressing reverse loading force impact based on fourier nerve operator according to claim 7, wherein the loading force feedback and the previous control period control input command are used as the extended state observer to input and output the total disturbance estimation, the method determines whether the current moment satisfies the condition of the reverse adjacent period according to the duration of the reverse start point moment and the duration of the reverse adjacent period in the reverse adjacent period parameter and combines the current moment, selects the corresponding disturbance value from the reverse impact disturbance prediction sequence according to the position of the current moment in the reverse adjacent period when the current moment satisfies the condition of the reverse adjacent period, and adopts the preset impact fusion coefficient to linearly fuse the total disturbance estimation and the corresponding disturbance value to obtain the impact compensation disturbance quantity, and calculates the control input command based on the loading force error, the impact compensation disturbance quantity or the total disturbance estimation and the previous control period control input command when the condition of the reverse adjacent period is satisfied or not satisfied, wherein the operation is represented by a comprehensive function, and the comprehensive function is specifically: ; Wherein, the In order to load the force error, In order to load the force instruction, In order to load the force feedback, Is the judgment result of the commutation adjacent time period condition and takes the value as Or (b) , For the current moment of time, For the moment of the start of the commutation, For commutation adjacent period duration, In order to indicate the function, In order to predict the number of sampling points, In order to control the period of time, In order to perform the rounding operation upwards, In order to perform the rounding down operation, In order to take the minimum value to operate on, In order to take the maximum value to operate, For indexing the position of the current instant within the commutation neighborhood, Predicting sequences for commutation impact disturbances Middle (f) Single-channel disturbance prediction values of the sampling points, In order to correspond to the value of the disturbance, For the total disturbance estimation output by the extended state observer, In order to preset the impact fusion coefficient, In order to compensate the disturbance quantity for the impact, The control input command is for the last control period, The control input command is for the current control period, 、、 In order to control the gain of the gain control, For the purpose of equivalent control gain, And (3) with The loading force errors of the previous control period and the previous two control periods are respectively, 、、 The observer gain of the extended state observer is respectively applied to error correction of the loading force state estimation channel, the loading force change rate estimation channel and the total disturbance estimation channel.
9. The method for suppressing reverse loading force impact based on fourier nerve operator according to claim 1, wherein step S6 specifically comprises: Converting a control input instruction into an actuator driving signal through a driving interface and applying the actuator driving signal to an actuator, and driving a loading mechanism to generate loading force in the current control period; the force sensor collects update loading force feedback and the encoder collects update actuator speed in a sampling window of the next control period, and the time deviation between the update loading force feedback sampling time and the update actuator speed sampling time is smaller than a time alignment threshold; Forming sampling points by updating loading force feedback, updating the speed of an actuator and a loading force instruction of the control period and a current control period control input instruction according to the control period sequence; And writing the sampling points into a sliding window of the current loading sampling sequence and forming a next control period loading sampling sequence to be used as a subsequent commutation adjacent period parameter to determine the input intercepted by the commutation adjacent state segment.

Description

Reversing loading force impact inhibition method based on Fourier neural operator Technical Field The invention relates to the technical field of loading force control and intelligent control, in particular to a reversing loading force impact inhibition method based on a Fourier neural operator. Background The electrohydraulic servo loading table, the servo motor lead screw loading device and the material fatigue testing machine are always loaded alternately in the positive and negative directions in the force control test. At the moment of reversing, the speed of the actuator is close to zero, and static friction debonding and friction direction overturning are easy to cause loading force impact, so that test waveform and structural safety are affected. Meanwhile, hydraulic elasticity, transmission clearance, sampling and driving delay are overlapped, so that the traditional steady state or track smoothing strategy is difficult to cover the non-steady disturbance. The engineering needs to give consideration to synchronous acquisition of instructions, feedback and speed and local period identification facing reversing. The prior art mainly adopts closed-loop control and feedforward shaping, namely, PID/PI or self-disturbance-rejection control restraining force error based on an extended state observer is adopted, the change of instructions is slowed down by matching with jerk limit, S curve or filtering, friction/dead zone compensation is introduced, zero speed detection and switching logic daemon reversing process is carried out, robustness is enhanced by using a disturbance observer or a speed loop in part of schemes, and a method is also adopted for judging direction change and control switching by utilizing sliding window record history data. The whole is used for improving tracking and disturbance rejection by time domain adjustment and parameterization compensation. The existing scheme has the defects that special prediction of the impact disturbance time sequence form in the 'reversing adjacent period' is lacked, the compensation quantity is always constant or an empirical function, the offset is easy to be misplaced with the actual impact waveform, unified sampling time alignment and strict direction and zero speed judgment are lacked, a triggering and compensation window is inaccurate, a compensation channel and an observer are not fused, disturbance estimation and impact prediction are not cooperated in the same closed loop, and targeted inhibition of reversing transient is limited. Therefore, a method of suppressing the impact of the reverse loading force, which can solve the above-mentioned deficiencies of the prior art, is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The application aims to provide a reversing loading force impact suppression method based on a Fourier neural operator, which aims to solve the core technical problems of how to generate impact disturbance prediction aligned with a time-aligned multi-channel historical segment based on a reversing adjacent time period in which a loading force instruction is subjected to positive-negative switching and the speed of an actuator is close to zero, and to fuse the impact disturbance prediction with the total disturbance of an extended state observer for control input calculation so as to suppress reversing impact on the premise of not changing a non-adjacent time period control strategy. The reversing loading force impact suppression method based on the Fourier neural operator comprises the following steps: S1, acquiring a loading force instruction, loading force feedback, actuator speed and a control input instruction of a previous control period in an electrohydraulic servo loading table, a servo motor lead screw loading device or a material fatigue testing machine to form a current loading sampling sequence; S2, judging that the loading force instruction is switched in the positive and negative directions and the absolute value of the speed of the actuator is smaller than a zero-speed judgment threshold value according to the current loading sampling sequence, and determining the starting point moment of the reversing and the duration of the reversing adjacent time period to obtain reversing adjacent time period parameters; S3, intercepting a preset history time before a reversing starting point time from a current loading sampling sequence according to reversing adjacent time period parameters to form a reversing adjacent state segment; S4, inputting the reversing adjacent state segments into a Fourier operator for calculation, wherein the Fourier operator sequentially comprises a frequency domain feature generation layer, a Fourier operator association layer and a time waveform reduction layer, the frequency domain feature generation layer converts the reversing adjacent state segments into frequency domain representation for describing reversing instantaneous change, the Fourier operator a