Reversing loading force impact suppression method based on Fourier neural operator

Abstract

The invention provides a reversing loading force impact suppression method based on a Fourier neural operator. Aiming at the impact of an electro-hydraulic servo loading table, a servo motor lead screw loading device and a material fatigue testing machine during positive and negative switching of a loading force instruction, the method comprises the following steps: performing time alignment sampling on the loading force instruction, feedback, actuator speed and control input of a previous period in a fixed control period, judging a reversing starting point and an adjacent time period, and intercepting a historical fragment; taking the segment as input, and performing frequency domain feature generation, operator association and time reduction on a Fourier neural operator to obtain a commutation impact disturbance prediction sequence aligned with an adjacent time period; the extended state observer estimates the total disturbance and linearly fuses the total disturbance with the predicted disturbance according to a fusion coefficient in an adjacent time period, and the active disturbance rejection control is driven to calculate control input; total disturbance closed-loop control is adopted in a non-adjacent period; according to the method, impact can be inhibited in the reversing process, the loading stability is improved, and the method is suitable for a force control scene of the loading equipment.

Description

Technical Field

[0001] This invention relates to the field of loading force control and intelligent control technology, and in particular to a method for suppressing commutated loading force impact based on Fourier neural operators. Background Technology

[0002] Electro-hydraulic servo loading stages, servo motor lead screw loading devices, and material fatigue testing machines often require alternating positive and negative loading in force-controlled tests. At the moment of reversal, the actuator speed approaches zero, and static friction disbondment and friction direction reversal can easily cause loading force impacts, affecting the test waveform and structural safety. Simultaneously, the superposition of hydraulic elasticity, transmission backlash, sampling, and drive delay makes it difficult for traditional steady-state or trajectory smoothing strategies to cover this non-stationary disturbance. In engineering, it is necessary to consider the synchronous acquisition of commands, feedback, and speed, as well as the identification of local time periods oriented towards reversal.

[0003] Existing technologies mostly employ closed-loop control and feedforward shaping: using PID / PI or active disturbance rejection control based on extended state observers to suppress force errors; combining jerk limiting, S-curves, or filtering to mitigate command changes; introducing friction / dead-zone compensation, zero-speed detection, and switching logic to protect the commutation process; some schemes use disturbance observers or velocity loops to enhance robustness; and some methods utilize sliding windows to record historical data to determine direction changes and control switching. Overall, time-domain adjustment and parameterized compensation are used to improve tracking and disturbance rejection.

[0004] The shortcomings of the existing scheme are: lack of specific prediction of the timing pattern of the impact disturbance in the "adjacent period of commutation", the compensation amount is often a constant or empirical function, which is easy to be misaligned with the actual impact waveform; lack of unified sampling time alignment and strict direction and zero velocity determination, and inaccurate triggering and compensation windows; insufficient integration of compensation channel and observer, disturbance estimation and impact prediction are not coordinated in the same closed loop, and the targeted suppression of commutation transients is limited.

[0005] Therefore, a method for suppressing the impact of commutation loading force that can overcome the shortcomings of the prior art is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a commutation loading force impact suppression method based on Fourier neural operators. The core technical problem to be solved by this application is: how to generate an impact disturbance prediction aligned with the time period based on time-aligned multi-channel historical segments during the commutation adjacent time period when the loading force command switches between positive and negative and the actuator speed is close to zero, and to fuse it with the total disturbance of the extended state observer for control input calculation, so as to suppress the commutation impact without changing the control strategy of the non-adjacent time period.

[0007] The commutation loading force impact suppression method based on Fourier neural operators according to embodiments of the present invention includes: S1. In an electro-hydraulic servo loading stage, servo motor screw loading device, or material fatigue testing machine, acquire loading force command, loading force feedback, actuator speed, and control input command of the previous control cycle to form the current loading sampling sequence. S2. Based on the current loading sampling sequence, determine that the loading force command has switched between positive and negative directions and the absolute value of the actuator speed is less than the zero speed determination threshold. Determine the starting time of the reversal and the duration of the reversal adjacent time period to obtain the parameters of the reversal adjacent time period. S3. Based on the commutation proximity time period parameter, extract the preset historical duration before the commutation start time from the currently loaded sampling sequence to form a commutation proximity state segment. S4. Input the commutation proximity state segment into the Fourier neural operator for calculation. The Fourier neural operator includes a frequency domain feature generation layer, a Fourier operator correlation layer, and a time waveform restoration layer. The frequency domain feature generation layer converts the commutation proximity state segment into a frequency domain representation to describe the instantaneous changes in commutation. The Fourier operator correlation layer performs correlation calculations on different frequency components in the frequency domain representation to characterize the disturbance patterns corresponding to static friction debonding and friction direction reversal. The time waveform restoration layer restores the correlation calculation results into a commutation impact disturbance prediction sequence consistent with the duration of the commutation proximity period, thus obtaining the commutation impact disturbance prediction sequence. S5. Based on the loading force feedback and the control input command of the previous control cycle, the total disturbance is estimated using an extended state observer. When the current time meets the conditions of the commutation adjacent time period defined by the commutation adjacent time period parameter, the corresponding disturbance value is taken from the commutation impact disturbance prediction sequence according to the current time position within the commutation adjacent time period, and fused with the total disturbance estimate to obtain the impact compensation disturbance amount. Then, the control input command is calculated based on the loading force command, loading force feedback and impact compensation disturbance amount. When the current time does not meet the conditions of the commutation adjacent time period, the control input command is calculated based on the loading force command, loading force feedback and total disturbance estimate. S6. Apply the control input command to the actuator to drive the loading mechanism to generate loading force, obtain updated loading force feedback and update the actuator speed, and form the loading sampling sequence for the next control cycle together with the loading force command and the control input command.

[0008] Optionally, S1 is as follows: The control cycle is set and sampling is triggered by the control cycle timer. The loading force command is obtained in each control cycle, and the loading force feedback is collected by the force sensor and the actuator speed is collected by the encoder. The time deviation between the loading force feedback sampling time and the actuator speed sampling time is less than the time alignment threshold. At the beginning of the control cycle, the control input command of the previous control cycle is read, and the control input command of the previous control cycle is associated with the loading force command, loading force feedback and actuator speed corresponding to the current control cycle to form a sampling point according to the same control cycle identifier; The sampling points are written into the sliding window in the order of the control cycle. The sliding window covers a number of consecutive sampling points that are not less than the preset historical duration, so as to meet the truncation length requirements of the subsequent adjacent state segments. The sampling points within the sliding window are arranged in chronological order to form the currently loaded sampling sequence and output to the commutation determination and commutation shock disturbance prediction sequence generation link.

[0009] Optionally, S2 is as follows: The loading force commands in the current loading sampling sequence are determined according to the control cycle order. The loading force commands are compared with the command direction determination threshold to obtain the loading force command direction of the current control cycle and the loading force command direction of the previous control cycle. When the direction of the loading force command is inconsistent with the direction of the loading force command in the previous control cycle, the speed amplitude of the actuator in the current control cycle is taken and compared with the zero speed determination threshold. When the speed amplitude is less than the zero speed determination threshold, a reversal trigger flag is generated. During the control cycle in which the commutation trigger flag is generated, the sampling time of this control cycle is determined as the commutation start time, and commutation proximity time tracking is performed on subsequent control cycles starting from the commutation start time. The duration of the near-turning period is determined based on the first moment when the speed amplitude exceeds the turning exit speed threshold during the near-turning period tracking. If the speed amplitude does not exceed the turning exit speed threshold, the duration of the near-turning period is determined based on the near-turning period duration threshold, thus obtaining the near-turning period parameters that include the turning start time and the duration of the near-turning period.

[0010] Optionally, S3 specifically refers to: The number of historical sampling points corresponding to the preset historical duration is determined based on the reversal start time in the reversal proximity time parameters and in combination with the control cycle. Locate the commutation start sampling point corresponding to the commutation start time in the current loading sampling sequence, and use the number of consecutive historical sampling points before the commutation start sampling point as the interception range, so that the interception range covers the consecutive sampling points before the positive and negative directions of the loading force command are switched. When the number of sampling points before the reversal starting point in the current loading sampling sequence is less than the number of historical sampling points, all sampling points before the reversal starting point are determined as the starting sampling points and the truncation range is kept continuous. The continuous sampling points from the starting sampling point to the commutation starting point are arranged in chronological order to form a commutation adjacent state segment. Each sampling point maintains the channel order of the three numerical features of loading force command, loading force feedback, and actuator speed, so that the commutation adjacent state segment meets the point-by-point linear mapping input requirements of the frequency domain feature generation layer.

[0011] Optionally, S4 specifically refers to: The commutation neighbor state segment is used as the input of the Fourier neural operator. On the input side, velocity direction change features are generated based on the actuator velocity and command direction change features are generated based on the loading force command. The velocity direction change features and command direction change features are combined with the three-channel features of the commutation neighbor state segment to participate in the subsequent mapping calculation. A frequency domain feature generation layer is used to perform point-by-point linear mapping on the commutation neighbor state segments. Sixty-four groups of neurons perform weighted summation and nonlinear transformation on the loading force command, loading force feedback and actuator speed to obtain the internal representation of sixty-four channels arranged in time order. Based on the internal representation of the 64 channels, a frequency domain representation is generated along the time direction. The frequency domain representation is expressed using real and imaginary channels and maintained as a frequency domain structure, so that the frequency domain representation carries information about the rate and magnitude of change before and after the change. The frequency domain representation is input into the Fourier operator correlation layer and passed through multiple levels of Fourier correlation units in sequence. Each level of Fourier correlation unit sets a learnable weight for each frequency position, and the sixty-four channels at that frequency position are weighted and combined to obtain the frequency domain response. Channel hybrid computation is performed on the frequency domain response. A weighted connection of 64 x 64 is used to perform linear combination of channels at each frequency position and nonlinear transformation. The updated frequency domain representation is output and used as the input of the next level Fourier correlation unit until the correlation calculation result is obtained. The correlation calculation results are input into the time waveform restoration layer and the time waveform is restored. The number of predicted sampling points is determined based on the duration of the adjacent time period and the control cycle in the adjacent time period parameter of the commutation, and a time series representation consistent with the number of predicted sampling points is generated. The time series representation is compressed into a single-channel perturbation representation by point-by-point linear mapping. A two-level fully connected neuron group is used to map the sixty-four channels into sixteen channels and then map the sixteen channels into single channels. The commutation shock perturbation prediction sequence is output in chronological order without outputting control input commands.

[0012] Optionally, the commutation neighbor state segment is used as the input to the Fourier neural operator, and velocity direction change features are generated based on the actuator velocity and command direction change features are generated based on the loading force command on the input side. The velocity direction change features and command direction change features are combined with the three-channel features of the commutation neighbor state segment to participate in the subsequent mapping calculation. The velocity direction change features and command direction change features are obtained by calculating a change function, wherein the change function is specifically: ; in, For the first The velocity direction change characteristics of each sampling point For the first The command direction change characteristics of each sampling point For the first Executor speed at each sampling point For the first Executor speed at each sampling point For the first The force command for each sampling point For the first The force command for each sampling point The threshold for zero speed determination. A threshold for determining the direction of the instruction. This is an indicator function that takes the value of 1 when the condition within the parentheses is true and 0 when the condition is false. The controller will at the first sampling point and Set it to zero and recursively calculate from the second sampling point in the manner described. The controller will and With the same sampling point , , They jointly participate in subsequent mapping calculations, enabling the point-by-point linear mapping to obtain a computable source of input difference at the sampling points where the direction of the applied force command is switched and the direction of the actuator velocity is reversed.

[0013] Optional, S5 specifically includes: Using the loading force feedback and the control input command of the previous control cycle as inputs to the extended state observer, the observer state is updated and the total disturbance estimate is output. Based on the start time of the reversal and the duration of the reversal adjacent time period in the reversal adjacent time period parameters, and combined with the current time, determine whether the current time meets the reversal adjacent time period conditions; When the current time meets the condition of the adjacent time period of the reversal, the position of the current time in the adjacent time period of the reversal is determined according to the time difference between the current time and the starting time of the reversal and the control cycle, and the number of predicted sampling points is determined by the duration of the adjacent time period of the reversal and the control cycle. Based on the current position within the adjacent time period of the reversal, a corresponding disturbance value aligned with the current position is selected from the reversal impact disturbance prediction sequence. The corresponding disturbance value is the single-channel disturbance prediction value of the reversal impact disturbance prediction sequence. A preset impact fusion coefficient is used to linearly fuse the total disturbance estimate and the corresponding disturbance value to obtain the impact compensation disturbance amount, which is used as the disturbance compensation input for active disturbance rejection control. When the conditions for the adjacent period of reversal are met at the current moment, the loading force command and loading force feedback are used to generate the loading force error, and the control input command is calculated based on the loading force error, the impact compensation disturbance amount and the control input command of the previous control cycle. If the conditions for the adjacent period of reversal are not met at the current moment, the loading force command and loading force feedback are used to generate the loading force error, and the control input command is calculated based on the loading force error, the total disturbance estimate and the control input command of the previous control cycle.

[0014] Optionally, the loading force feedback and the control input command of the previous control cycle are used as inputs to the extended state observer, and the total disturbance estimate is output. Based on the commutation start time and the duration of the commutation adjacent time period in the commutation adjacent time period parameters, and combined with the current time, it is determined whether the current time meets the commutation adjacent time period condition. If the current time meets the commutation adjacent time period condition, the corresponding disturbance value is selected from the commutation impact disturbance prediction sequence according to the current time's position within the commutation adjacent time period, and the total disturbance estimate and the corresponding disturbance value are linearly fused using a preset impact fusion coefficient to obtain the impact compensation disturbance amount. If the commutation adjacent time period condition is met or not, the control input command is calculated based on the loading force error, the impact compensation disturbance amount, or the total disturbance estimate and the control input command of the previous control cycle, respectively. The above calculations are represented by a synthesis function, which is specifically: ; in, For the applied force error, For loading force command, For force feedback, The result of determining the condition of the reversal period is given and the value is [value]. or , For the current moment, The starting point of the reversal, The duration of the adjacent time period for the reversal, For indicator functions, To predict the number of sampling points, To control the cycle, For floor operations, For floor operations, To perform the minimum value operation, To perform the maximum value operation, This is the position index of the current time within the adjacent time period of the reversal. For the prediction sequence of commutation shock disturbance The Middle Single-channel perturbation prediction value for each sampling point For the corresponding disturbance value, This is the total perturbation estimate from the output of the extended state observer. To preset the impact fusion coefficient, This is the amount of disturbance to compensate for the impact. This is the control input command for the previous control cycle. For the current control cycle, control input commands are provided. , , To control the gain, For equivalent control gain, and These represent the loading force errors of the previous control cycle and the two control cycles prior, respectively. , , The observer gain of the extended state observer is applied to the error correction of the loading force state estimation channel, the loading force rate of change estimation channel, and the total disturbance estimation channel, respectively.

[0015] Optionally, step S6 specifically includes: The control input command is converted into an actuator drive signal via the drive interface and applied to the actuator, driving the loading mechanism to generate a loading force in the current control cycle; Within the sampling window of the next control cycle, the force sensor collects and updates the loading force feedback, and the encoder collects and updates the actuator speed. The time deviation between the sampling time of updating the loading force feedback and the sampling time of updating the actuator speed is less than the time alignment threshold. The updated loading force feedback, updated actuator speed, loading force command of the current control cycle, and control input command of the current control cycle are combined into sampling points in the order of the control cycle; The sampling points are written into the sliding window of the currently loaded sampling sequence and form the next control cycle loading sampling sequence, which serves as the input for subsequent commutation adjacent time period parameter determination and commutation adjacent state segment extraction.

[0016] The beneficial effects of this invention are: (1) This proposal puts forward an improved method for fusing active disturbance rejection control and extended state observer. Focusing on the impact suppression of the commutation adjacent time period, the single-channel disturbance prediction value output by the Fourier neural operator is selected according to the current position index and linearly fused with the total disturbance of the observer with a preset fusion coefficient. Under the unified discrete timing, the current control input is calculated by the combined function using the loading force error, the fused disturbance and the control input of the previous cycle. The non-adjacent time period is backed up to the total disturbance closed loop. This improvement achieves time-period and position-period compensation for commutation transients without changing the conventional control structure, reducing over-compensation or under-compensation caused by the misalignment between the compensation and the actual impact waveform.

[0017] (2) This proposal puts forward a novel method for predicting commutation shock disturbances. It adopts a cascaded structure of a frequency domain feature generation layer, a Fourier operator correlation layer, and a time waveform reconstruction layer of Fourier neural operators, and introduces velocity direction change features and command direction change features on the input side. This technique uses multi-channel historical segments extracted by a sliding window to establish learnable correlations across frequencies and channels in the frequency domain. Then, it restores the disturbance prediction sequence aligned with the time period according to the duration constraint of the commutation adjacent time period. It can characterize the typical disturbance patterns of static friction debonding and friction direction reversal, and provide time series quantities that can be directly used for control, thereby enhancing the targeted compensation in commutation scenarios.

[0018] (3) This proposal puts forward a sampling and judgment method for commutation scenarios. It uses timestamp alignment to collect the loading force command, feedback and actuator speed. Based on the command direction judgment threshold and zero speed judgment threshold, the commutation start point is accurately triggered, and the commutation exit speed threshold or the longest duration is used to limit the adjacent time period. On this basis, the preset historical time before the start point is extracted to form the commutation adjacent state segment. This overall approach aligns the triggering, prediction and compensation within the same discrete closed loop, reduces the error of the commutation compensation time window, improves the consistency between the compensation amount and the actual disturbance in time sequence, and avoids unnecessary intervention in non-adjacent time periods. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a commutation loading force impact suppression method based on Fourier neural operators proposed in this invention; Figure 2 The flowchart shows the loading sampling and sequence generation process of a commutation loading force impact suppression method based on Fourier neural operators proposed in this invention. Figure 3 This is a flowchart of the commutation determination and commutation proximity time parameter generation process for a commutation loading force impact suppression method based on Fourier neural operators proposed in this invention. Detailed Implementation

[0020] In Example 1, reference Figures 1 to 3 A method for suppressing commutated loading force impact based on Fourier neural operators, comprising: S1. In an electro-hydraulic servo loading stage, servo motor screw loading device, or material fatigue testing machine, acquire loading force command, loading force feedback, actuator speed, and control input command of the previous control cycle to form the current loading sampling sequence. S2. Based on the current loading sampling sequence, determine that the loading force command has switched between positive and negative directions and the absolute value of the actuator speed is less than the zero speed determination threshold. Determine the starting time of the reversal and the duration of the reversal adjacent time period to obtain the parameters of the reversal adjacent time period. S3. Based on the commutation proximity time period parameter, extract the preset historical duration before the commutation start time from the currently loaded sampling sequence to form a commutation proximity state segment. S4. Input the commutation proximity state segment into the Fourier neural operator for calculation. The Fourier neural operator includes a frequency domain feature generation layer, a Fourier operator correlation layer, and a time waveform restoration layer. The frequency domain feature generation layer converts the commutation proximity state segment into a frequency domain representation to describe the instantaneous changes in commutation. The Fourier operator correlation layer performs correlation calculations on different frequency components in the frequency domain representation to characterize the disturbance patterns corresponding to static friction debonding and friction direction reversal. The time waveform restoration layer restores the correlation calculation results into a commutation impact disturbance prediction sequence consistent with the duration of the commutation proximity period, thus obtaining the commutation impact disturbance prediction sequence. S5. Based on the loading force feedback and the control input command of the previous control cycle, the total disturbance is estimated using an extended state observer. When the current time meets the conditions of the commutation adjacent time period defined by the commutation adjacent time period parameter, the corresponding disturbance value is taken from the commutation impact disturbance prediction sequence according to the current time position within the commutation adjacent time period, and fused with the total disturbance estimate to obtain the impact compensation disturbance amount. Then, the control input command is calculated based on the loading force command, loading force feedback and impact compensation disturbance amount. When the current time does not meet the conditions of the commutation adjacent time period, the control input command is calculated based on the loading force command, loading force feedback and total disturbance estimate. S6. Apply the control input command to the actuator to drive the loading mechanism to generate loading force, obtain updated loading force feedback and update the actuator speed, and form the loading sampling sequence for the next control cycle together with the loading force command and the control input command.

[0021] In this embodiment, step S1 specifically includes: In electro-hydraulic servo loading stages, servo motor screw loading devices, or material fatigue testing machines, the controller is first set to a fixed control cycle, denoted as [missing information]. The controller uses a control cycle timer to generate periodic trigger signals. At the trigger moment of each control cycle, it enters the sampling process, ensuring that the sampling process is in the same discrete time sequence as the subsequent extended state observer and active disturbance rejection control calculations. The applied force command is denoted as... The applied force feedback is denoted as The actuator speed is denoted as Control input commands are denoted as During the control cycle Internally, the controller reads the loading force command corresponding to the current control cycle from the host computer or waveform generation module. Within the same control cycle, the force sensor is triggered to sample and obtain the loading force feedback. The encoder is triggered to sample and obtain the actuator speed. ,in For control cycle identification; To ensure that subsequent commutation-adjacent state segments accurately reflect the instantaneous commutation process, the controller imposes a time consistency constraint on the sampling time of the applied force feedback and the sampling time of the actuator speed. The controller assigns a timestamp to the sampling time of the applied force feedback. Assign a timestamp to the moment when the actuator speed is sampled. And set a time alignment threshold. The controller uses the absolute value of the difference between two timestamps as the time deviation, and then compares the time deviation with... Comparison, when the time deviation is less than At that time, and The results are considered valid synchronous sampling results within the same control cycle if the time deviation is not less than [value missing]. When the controller re-triggers the next sampling within the current control cycle until the threshold condition is met, it then enters the sampling point construction process, so that the actuator speed amplitude judgment required for the reversal determination and the instantaneous offset judgment of the loading force feedback have the same time reference. At the beginning of the control cycle, the controller reads the control input command from the previous control cycle, which is denoted as [previous control input command]. The controller will Acquired with the current control cycle , , Identified according to the same control cycle Perform association and construct sampling points Sampling points Includes loading force instructions Force feedback Actuator speed Control input command from the previous control cycle By introducing and maintain with Same period alignment, sampling points It can be directly used as the input basis for estimating the total perturbation by the extended state observer, while maintaining the consistency of the three-channel data source required for the commutation neighbor state segment, so that the input of the subsequent Fourier neural operator can still be obtained from the sampling point. Extraction , , Three numerical features form a commutative neighbor state segment; The controller uses a sliding window to hold consecutive sampling points. The length of the sliding window is represented by the number of consecutive sampling points, denoted as . The preset historical duration is recorded as The controller according to With control cycle Determine the number of historical sampling points The method of determination is to... Divide by The result is rounded up to the nearest integer number of sampling points, and then... Not less than Before the end of each control cycle, the controller samples the points in the order of the control cycle. Write to the sliding window, and when the sliding window capacity reaches... The earliest written sampling point is used as the object to be removed, ensuring that the sampling points within the window remain continuous, thus ensuring that the sliding window always covers a period of time no less than the preset history duration. The number of consecutive sampling points meets the length requirement for subsequently extracting adjacent state segments based on the starting time of the reversal; The controller arranges the sampling points within the sliding window into a currently loaded sampling sequence in chronological order. This currently loaded sampling sequence is denoted as... Currently loading sampling sequence It is updated once in each control cycle and output to the commutation determination and commutation shock disturbance prediction sequence generation link, enabling the commutation determination to be based on... Switching between positive and negative directions and The zero-velocity threshold is compared to determine the commutation start time, while ensuring that subsequent Fourier neural operators can start from the currently loaded sampling sequence before the determined commutation start time. Extracting and covering preset historical durations The continuous three-channel sequence is used as a commutation neighbor state segment.

[0022] In this embodiment, step S2 specifically includes: In step S2, the controller uses the currently loaded sampling sequence as input to perform commutation determination and commutation proximity time parameter generation. The currently loaded sampling sequence is denoted as... ,in This serves as the identifier for the current control cycle, denoted as . , It consists of sampling points arranged in chronological order, and each sampling point contains the applied force command. Force feedback Actuator speed With sampling time timestamp ,in As the control cycle identifier corresponding to the sampling point, the controller pre-sets a command direction determination threshold. Zero speed determination threshold , reversing exit speed threshold Duration threshold of the period adjacent to the reversal ,in Used to suppress directional misjudgments caused by minute jitters in the applied force command near zero. Used for zero-velocity determination Used to describe the speed amplitude threshold for exiting a commutation phase. Used to limit the maximum duration of tracking in the vicinity of a reversal; The controller determines the direction of the loading force commands in the current loading sampling sequence according to the control cycle order. Read the current control cycle loading force command Force command from the previous control cycle and will and The direction of the applied force command in the current control cycle is obtained through comparison. ,Will and The direction of the applied force command from the previous control cycle is obtained by comparison. ,in, As a directional marker, Taking the positive direction indicates that the loading force command is greater than , Taking the negative value indicates that the applied force command is less than When the loading force command is located and In between, the controller will To maintain continuity of direction determination, the direction flag is kept at the most recent non-zero value. During system startup, the controller initializes the direction flag to the value where the first applied force command satisfies a value greater than 0.5%. or less The corresponding direction flag is set, and the continuous judgment process is entered after initialization; In obtaining and Then, the controller determines whether the direction of the applied force command has switched between positive and negative. and When the signs are opposite, the zero-speed condition is determined, and the actuator speed for the current control cycle is adjusted. Take velocity amplitude Among them, the velocity amplitude for The absolute value of the speed amplitude will be used by the controller. Zero speed determination threshold Comparison, in terms of velocity amplitude Less than Generate commutation trigger flag at time ,in The trigger state indicates that the combined conditions of switching the direction of the applied force command and the zero-speed determination threshold are met, when the speed amplitude... Not less than When the controller does not generate a reversal trigger flag, it continues to repeat the direction determination process in the next control cycle, so that the reversal start time falls within the period when the actuator speed meets the zero speed determination threshold constraint. When the reversing trigger flag During the control cycle During generation, the controller determines the sampling time of this control cycle as the commutation start time, which is denoted as . and order The controller starts from the commutation start time. The system performs commutation proximity time tracking for subsequent control cycles. This commutation proximity time tracking reads the actuator speed of subsequent sampling points sequentially within each control cycle and calculates the corresponding speed amplitude in each subsequent control cycle. ,in This is the control cycle offset after the start of the reversal. for The absolute value, the reversal of the adjacent time period tracking and maintenance with the sampling point timestamp Synchronous updates are implemented to ensure that the duration of adjacent reversal periods can be directly obtained from the timestamp difference and remains consistent with the conversion of subsequent prediction sampling point counts. The controller determines the duration of the commutation proximity period based on the tracking of the commutation proximity period. The controller checks the speed amplitude cycle by cycle after the commutation start time. Does it exceed the reversing exit speed threshold? And will meet the speed amplitude Not less than minimum offset Corresponding sampling time Determined as the end time of the adjacent reversal period When no velocity amplitude appears during the tracking process after the starting moment of the reversal, Not less than During the control cycle, the controller uses the duration threshold of the adjacent time period for switching. Determine the end time of the adjacent reversal period ,in Take as Continued At that moment, the controller and the starting point of the reversal The time difference determines the duration of adjacent time periods during the reversal. And will switch to the starting point. Duration of the period adjacent to the reversal Together they form the parameters of the adjacent time period of the reversal. When it is necessary to provide discrete length constraints to the subsequent time waveform reconstruction layer, the controller is based on With control cycle Will Converted to the number of prediction sampling points The conversion method is to convert... Divide by Round up to obtain the integer number of sampling points, making the predicted number of sampling points... Parameters of the period adjacent to the reversal Maintaining the same commutation start time reference is used to constrain the output length of the commutation shock disturbance prediction sequence to align with the time series.

[0023] In this embodiment, step S3 specifically includes: In step S3, the controller uses the commutation proximity time interval parameter and the currently loaded sampling sequence as input to generate a commutation proximity state segment for use by the Fourier neural operator. The commutation proximity time interval parameter is denoted as... This includes the start time of the reversal. Duration of the period adjacent to the reversal The currently loaded sampling sequence is denoted as ,in This serves as the identifier for the current control cycle. It consists of sampling points arranged in chronological order, with each sampling point containing a force command. Force feedback Actuator speed With sampling time timestamp ,in This is the control period identifier corresponding to the sampling point; the control period is denoted as... ; The controller's preset historical duration is recorded as and based on With control cycle Determine the number of historical sampling points Historical sampling point count The method of determination is to Divide by The result is rounded up to the nearest integer number of sampling points, thus making The corresponding time coverage is no less than the preset historical duration. The controller will record the number of historical sampling points. As the target length benchmark for extracting the commutation adjacent state segment, the commutation adjacent state segment can cover the continuous historical change process before the positive and negative directions of the loading force command are switched, providing a time window for the subsequent frequency domain feature generation layer to extract the frequency domain information related to the direction change. The controller is currently loading the sampling sequence. Mid-positioning and reversal start time To ensure deterministic positioning, the controller timestamps the sampling points at the corresponding reversal starting points. and the starting point of the reversal Perform matching and set a matching time threshold. ,in To allow timestamp matching deviation thresholds, the controller in Search for sampling points in chronological order and calculate and The absolute value of the time difference, and when that absolute value is less than The corresponding sampling point is included in the candidate set. The controller determines the sampling point with the smallest absolute value of time difference in the candidate set as the commutation start point sampling point. The control cycle identifier of the commutation start point sampling point is denoted as... This is to ensure that the sampling point of the reversing start point is consistent with the generation time of the reversing trigger flag; After determining the commutation start point sampling point, the controller uses the number of consecutive historical sampling points preceding the commutation start point sampling point. The sampling points are used as the interception range, and the controller records the start control cycle identifier of the interception as... And determine as follows :when The number of previously available sampling points was not less than At that time, take ,when Previously, the number of available sampling points was less than At that time, the earliest control cycle identifier corresponding to all sampling points before the reversal starting point sampling point is determined as... and keep from to The sampling points are continuous and uninterrupted, thus meeting the requirement of continuous interception range and avoiding the introduction of interpolated data; The controller will identify the start of the control cycle. Control cycle identifier to the sampling point of the reversal start point Continuous sampling points are arranged in chronological order to form a commutation neighbor state segment, which is denoted as . In formation At that time, the controller only extracts the loading force command for each sampling point. Force feedback Actuator speed Three numerical characteristics, and a fixed channel sequence of loading force command, loading force feedback, and actuator speed, make It is represented by a three-channel sequence arranged in chronological order. The channel order is consistent with the point-by-point linear mapping input order of the frequency domain feature generation layer, which enables the frequency domain feature generation layer to perform a three-channel to multi-channel mapping for each time sampling point. In the subsequent generation of frequency domain representation, it retains the joint temporal information of the switching of the loading force command direction, the instantaneous offset of the loading force feedback, and the change of the actuator speed near zero speed, thereby supporting the input requirements of the Fourier neural operator for the prediction sequence of the commutation impact disturbance.

[0024] In this embodiment, step S4 specifically includes: In step S4, the controller uses the commutation neighbor state segment as input to the Fourier neural operator to generate a commutation shock perturbation prediction sequence, where the commutation neighbor state segment is denoted as... It consists of consecutive sampling points in chronological order, with each sampling point containing a force command. Force feedback Actuator speed ,in for The sampling point number in the sample is denoted as the parameter for the adjacent time period after the reversal. This includes the duration of adjacent reversal periods. The control cycle is denoted as The controller according to and Determine the number of prediction sampling points Predict the number of sampling points As a constraint on the length of the output sequence of the Fourier neural operator; To enable the frequency domain feature generation layer to explicitly carry direction reversal information, the controller generates velocity direction change features and command direction change features on the input side, and uses these features as additional inputs for point-by-point linear mapping. The controller sets a zero-velocity determination threshold. Threshold for determining instruction direction and in The velocity direction change characteristics and command direction change characteristics are obtained for each sampling point in chronological order based on the difference between adjacent sampling points. Specifically, a change function is used, and the calculation method is as follows: ; in, For the first The velocity direction change characteristics of each sampling point For the first The command direction change characteristics of each sampling point For the first Executor speed at each sampling point For the first Executor speed at each sampling point For the first The force command for each sampling point For the first The force command for each sampling point The threshold for zero speed determination. A threshold for determining the direction of the instruction. This is an indicator function that takes the value of 1 when the condition within the parentheses is true and 0 when the condition is false. The controller will at the first sampling point and Set it to zero and recursively calculate from the second sampling point in the manner described. The controller will and With the same sampling point , , They jointly participate in subsequent mapping calculations, enabling the point-by-point linear mapping to obtain a computable source of input difference at the sampling points where the direction of the applied force command is switched and the direction of the actuator velocity is reversed; Frequency domain feature generation layer The controller performs a point-by-point linear mapping to generate a 64-channel internal representation. The controller configures 64 groups of neurons for the frequency domain feature generation layer, with each group of neurons corresponding to the same sampling point. , , Perform a weighted summation and then consider the characteristics of the velocity direction change. Characteristics of changes in command direction By incorporating a weighted term into the weighted summation result and then performing a nonlinear transformation on the weighted summation result, an internal channel output for that sampling point is obtained. The sixty-four neuron groups sequentially generate the sixty-four channel internal outputs for that sampling point. The controller concatenates the sixty-four channel internal outputs from all sampling points in chronological order to form a sixty-four channel internal representation sequence, denoted as [the sequence is missing from the original text]. Its duration and Consistent; Frequency domain feature generation layer based on Generate frequency domain representation along the time direction, and the controller... Perform a Discrete Fourier Transform along the time direction for each channel to obtain the frequency domain structure containing complex coefficients at each frequency position. This frequency domain structure is denoted as... The controller will Organizing the frequency domain representation by using real and imaginary channels ensures that the frequency domain representation retains its frequency domain structure and phase information. This allows the frequency domain representation to simultaneously carry frequency distribution information corresponding to the rate of change before and after the transformation, as well as spectral amplitude information corresponding to the magnitude of change. The Fourier operator correlation layer performs multi-level correlation calculations on the frequency domain representation to obtain the correlation calculation result. The Fourier operator correlation layer is composed of multiple Fourier correlation units connected in series. Each Fourier correlation unit receives the frequency domain representation output from the previous layer and outputs the updated frequency domain representation. In each Fourier correlation unit, the controller first performs frequency domain weight calculation, configures learnable weights for each frequency position, and performs weighted combination of the 64 channels at that frequency position to obtain the frequency domain response. Then, channel mixing calculation is performed, and a 64x64 scale linear combination of channels is performed on the frequency domain response at each frequency position, and a nonlinear transformation is performed to make different frequency components form a correlation in the same frequency domain structure. The controller sends the updated frequency domain representation to the next level Fourier correlation unit and repeats the above process until the correlation calculation result is obtained. The correlation calculation result is maintained as a frequency domain structure that can be directly used for time waveform reconstruction. The time waveform reconstruction layer restores the correlation calculation results into a commutation shock disturbance prediction sequence. The controller performs time waveform reconstruction on the correlation calculation results in the time waveform reconstruction layer to obtain a multi-channel time series representation aligned with the time series, and then determines the prediction sampling point based on the number of sampling points. Constrain the output length so that the length of the time series representation is equal to the length of the time series representation. Correspondingly, the controller performs a pointwise linear mapping on the time series representation to obtain a single-channel perturbation representation. The pointwise linear mapping employs a two-level fully connected neuron group. The first fully connected neuron group maps sixty-four channels to sixteen channels and performs a nonlinear transformation. The second fully connected neuron group maps the sixteen channels to a single channel and performs a nonlinear transformation. The controller outputs a length of [length missing] in chronological order. The commutation shock disturbance prediction sequence is denoted as... Each sampling point represents a predicted perturbation value, and the Fourier neural operator only outputs the predicted sequence of commutation shock perturbations. And it does not output control input commands, so that It can enter the perturbation compensation link of the extended state observer in subsequent steps and be constrained by the loading force feedback closed loop.

[0025] In this embodiment, step S5 specifically includes: In step S5, the controller calculates the control input command based on the loading force feedback closed loop in each control cycle, and introduces the commutation impact disturbance prediction sequence into the disturbance compensation channel when the commutation proximity time condition is met. The loading force command is denoted as... The applied force feedback is denoted as The control input command of the previous control cycle is denoted as The control input command for the current control cycle is denoted as The control cycle is denoted as The current time is recorded as The parameters for the adjacent time period after the reversal are denoted as The reversing proximity time parameter includes the reversing start time. Duration of the period adjacent to the reversal The predicted sequence of commutation shock disturbances is denoted as , This is a sequence of single-channel perturbation predictions arranged in chronological order. The controller uses force feedback. Control input command from the previous control cycle The extended state observer is used as input to perform observer state updates. The extended state observer takes a discrete state form, and the observer state includes the applied force state estimate. Estimation of the rate of change of loading force With total disturbance estimate The controller uses the difference between the applied force feedback and the applied force state estimate as the observation error, and multiplies the observation error by the observer gain. , , Three correction values ​​are generated, and the controller adds these correction values ​​to the values ​​generated by the observer state from the previous control cycle. The obtained state prediction increments are calculated according to the control cycle. Perform discrete integral updates to output the total disturbance estimate for the current control cycle. ; The controller completes the determination of commutation proximity conditions, selection of corresponding disturbance values, fusion of impact compensation disturbances, and calculation of control input commands within the same computational link, specifically calculated using a comprehensive function: ; in, For the applied force error, For loading force command, For force feedback, The result of determining the condition of the reversal period is given and the value is [value]. or , For the current moment, The starting point of the reversal, The duration of the adjacent time period for the reversal, For indicator functions, To predict the number of sampling points, To control the cycle, For floor operations, For floor operations, To perform the minimum value operation, To perform the maximum value operation, This is the position index of the current time within the adjacent time period of the reversal. For the prediction sequence of commutation shock disturbance The Middle Single-channel perturbation prediction value for each sampling point For the corresponding disturbance value, This is the total perturbation estimate from the output of the extended state observer. To preset the impact fusion coefficient, This is the amount of disturbance to compensate for the impact. This is the control input command for the previous control cycle. For the current control cycle, control input commands are provided. , , To control the gain, For equivalent control gain, and These represent the loading force errors of the previous control cycle and the two control cycles prior, respectively. , , The observer gain of the extended state observer is applied to the error correction of the loading force state estimation channel, the loading force rate of change estimation channel, and the total disturbance estimation channel, respectively. when At that time, the controller uses impact compensation for disturbance. Enter the disturbance compensation input item and complete the control input command. Calculate, when At that time, the controller uses total disturbance estimation. Enter the disturbance compensation input item and complete the control input command. To ensure computability during the system startup phase, the controller performs calculations before a complete computation is formed. or During the control cycle, the missing loading force error term is set to zero and the calculation link is updated as described.

[0026] In this embodiment, step S6 specifically includes: In step S6, the controller uses the control input command obtained in step S5 to drive the loading mechanism, and writes the loading force feedback and actuator speed collected in the next control cycle into the sliding window of the current loading sampling sequence to form the loading sampling sequence for the next control cycle. The current control cycle is identified as... The control cycle is The current control cycle control input command is The current control cycle loading force command is ; Control input commands The signal is converted into an actuator drive signal via the drive interface and applied to the actuator. The actuator drive signal is denoted as... The driver interface is for Perform amplitude limiting, quantization, and output format conversion to enable... The controller meets the input range and interface type requirements of the actuator driver during the control cycle. General Output to actuator driver, actuator driver based on Generate a current or voltage driving force and apply it to the actuator, causing the loading mechanism to operate within the control cycle. The corresponding loading force output process is generated internally; In the next control cycle Within the sampling window, the force sensor collects and updates the loading force feedback, and the encoder collects and updates the actuator speed. The updated loading force feedback is denoted as... The speed of updating the executor is recorded as The sampling time for the applied force feedback is recorded as The actuator speed sampling time is recorded as The controller presets a time alignment threshold. and to and In comparison, At that time, and Determined to be the same control cycle The valid update value, in order to keep the timestamp of the written sampling point unique, the controller will control the cycle. The sampling time timestamp is recorded as and will Take as and The earlier sampling time ensures that the written sampling points are in the same order as the control cycle; Will update loading force feedback Update executor speed With control cycle Loading force command and control cycle Control input commands Sampling points are arranged in sequence according to the control cycle, and these sampling points are denoted as... , Includes loading force instruction field Loading force feedback field Actuator speed field Sampling time timestamp field And record the control input instruction fields. Among them, the loading force command field, loading force feedback field, and actuator speed field serve as the three-channel numerical feature source for the subsequent commutation adjacent state segment, and the control input command field serves as the control quantity recording field required for the input association of the extended state observer, so that the sampling point can simultaneously support the determination of the commutation adjacent time period parameters in step S2, the commutation adjacent state segment truncation in step S3, and the observer input tracing in step S5. sampling points A sliding window is written into the currently loaded sampling sequence to form the loading sampling sequence for the next control cycle. The currently loaded sampling sequence is denoted as... The capacity of the sliding window is denoted as , The controller will use a preset integer to limit the number of sampling points retained in the loaded sampling sequence. Added in chronological order The tail, and the number of sampling points after the addition exceeds The earliest sampling point is deleted to maintain a constant window length, thus obtaining the loading sampling sequence for the next control cycle. The controller will As inputs for determining the parameters of the adjacent switching time period and extracting the adjacent switching state segments, the calculation of the parameters of the adjacent switching time period, the adjacent switching state segments, and the prediction sequence of the switching shock disturbance are continuously updated within the closed loop of the control cycle.

[0027] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for suppressing commutated loading force impact based on Fourier neural operators, characterized in that, include: S1. In an electro-hydraulic servo loading stage, servo motor screw loading device, or material fatigue testing machine, acquire loading force command, loading force feedback, actuator speed, and control input command of the previous control cycle to form the current loading sampling sequence. S2. Based on the current loading sampling sequence, determine that the loading force command has switched between positive and negative directions and the absolute value of the actuator speed is less than the zero speed determination threshold. Determine the starting time of the reversal and the duration of the reversal adjacent time period to obtain the parameters of the reversal adjacent time period. S3. Based on the commutation proximity time period parameter, extract the preset historical duration before the commutation start time from the currently loaded sampling sequence to form a commutation proximity state segment. S4. Input the commutation proximity state segment into the Fourier neural operator for calculation. The Fourier neural operator includes a frequency domain feature generation layer, a Fourier operator correlation layer, and a time waveform restoration layer. The frequency domain feature generation layer converts the commutation proximity state segment into a frequency domain representation to describe the instantaneous changes in commutation. The Fourier operator correlation layer performs correlation calculations on different frequency components in the frequency domain representation to characterize the disturbance patterns corresponding to static friction debonding and friction direction reversal. The time waveform restoration layer restores the correlation calculation results into a commutation impact disturbance prediction sequence consistent with the duration of the commutation proximity period, thus obtaining the commutation impact disturbance prediction sequence. S5. Based on the loading force feedback and the control input command of the previous control cycle, the total disturbance is estimated using an extended state observer. When the current time meets the conditions of the commutation adjacent time period defined by the commutation adjacent time period parameter, the corresponding disturbance value is taken from the commutation impact disturbance prediction sequence according to the current time position within the commutation adjacent time period, and fused with the total disturbance estimate to obtain the impact compensation disturbance amount. Then, the control input command is calculated based on the loading force command, loading force feedback and impact compensation disturbance amount. When the current time does not meet the conditions of the commutation adjacent time period, the control input command is calculated based on the loading force command, loading force feedback and total disturbance estimate. S6. Apply the control input command to the actuator to drive the loading mechanism to generate loading force, obtain updated loading force feedback and update the actuator speed, and form the loading sampling sequence for the next control cycle together with the loading force command and the control input command.

2. The method for suppressing commutated loading force impact based on Fourier neural operators according to claim 1, characterized in that, S1 specifically refers to: The control cycle is set and sampling is triggered by the control cycle timer. The loading force command is obtained in each control cycle, and the loading force feedback is collected by the force sensor and the actuator speed is collected by the encoder. The time deviation between the loading force feedback sampling time and the actuator speed sampling time is less than the time alignment threshold. At the beginning of the control cycle, the control input command of the previous control cycle is read, and the control input command of the previous control cycle is associated with the loading force command, loading force feedback and actuator speed corresponding to the current control cycle to form a sampling point according to the same control cycle identifier; The sampling points are written into the sliding window in the order of the control cycle. The sliding window covers a number of consecutive sampling points that are not less than the preset historical duration, so as to meet the truncation length requirements of the subsequent adjacent state segments. The sampling points within the sliding window are arranged in chronological order to form the currently loaded sampling sequence and output to the commutation determination and commutation shock disturbance prediction sequence generation link.

3. The method for suppressing commutated loading force impact based on Fourier neural operators according to claim 1, characterized in that, S2 specifically refers to: The loading force commands in the current loading sampling sequence are determined according to the control cycle order. The loading force commands are compared with the command direction determination threshold to obtain the loading force command direction of the current control cycle and the loading force command direction of the previous control cycle. When the direction of the loading force command is inconsistent with the direction of the loading force command in the previous control cycle, the speed amplitude of the actuator in the current control cycle is taken and compared with the zero speed determination threshold. When the speed amplitude is less than the zero speed determination threshold, a reversal trigger flag is generated. During the control cycle in which the commutation trigger flag is generated, the sampling time of this control cycle is determined as the commutation start time, and commutation proximity time tracking is performed on subsequent control cycles starting from the commutation start time. The duration of the near-turning period is determined based on the first moment when the speed amplitude exceeds the turning exit speed threshold during the near-turning period tracking. If the speed amplitude does not exceed the turning exit speed threshold, the duration of the near-turning period is determined based on the near-turning period duration threshold, thus obtaining the near-turning period parameters that include the turning start time and the duration of the near-turning period.

4. The method for suppressing commutated loading force impact based on Fourier neural operators according to claim 1, characterized in that, S3 specifically refers to: The number of historical sampling points corresponding to the preset historical duration is determined based on the reversal start time in the reversal proximity time parameters and in combination with the control cycle. Locate the commutation start sampling point corresponding to the commutation start time in the current loading sampling sequence, and use the number of consecutive historical sampling points before the commutation start sampling point as the interception range, so that the interception range covers the consecutive sampling points before the positive and negative directions of the loading force command are switched. When the number of sampling points before the reversal starting point in the current loading sampling sequence is less than the number of historical sampling points, all sampling points before the reversal starting point are determined as the starting sampling points and the truncation range is kept continuous. The continuous sampling points from the starting sampling point to the commutation starting point are arranged in chronological order to form a commutation adjacent state segment. Each sampling point maintains the channel order of the three numerical features of loading force command, loading force feedback, and actuator speed, so that the commutation adjacent state segment meets the point-by-point linear mapping input requirements of the frequency domain feature generation layer.

5. The method for suppressing commutation loading force impact based on Fourier neural operators according to claim 1, characterized in that, S4 specifically refers to: The commutation neighbor state segment is used as the input of the Fourier neural operator. On the input side, velocity direction change features are generated based on the actuator velocity and command direction change features are generated based on the loading force command. The velocity direction change features and command direction change features are combined with the three-channel features of the commutation neighbor state segment to participate in the subsequent mapping calculation. A frequency domain feature generation layer is used to perform point-by-point linear mapping on the commutation neighbor state segments. Sixty-four groups of neurons perform weighted summation and nonlinear transformation on the loading force command, loading force feedback and actuator speed to obtain the internal representation of sixty-four channels arranged in time order. Based on the internal representation of the 64 channels, a frequency domain representation is generated along the time direction. The frequency domain representation is expressed using real and imaginary channels and maintained as a frequency domain structure, so that the frequency domain representation carries information about the rate and magnitude of change before and after the change. The frequency domain representation is input into the Fourier operator correlation layer and passed through multiple levels of Fourier correlation units in sequence. Each level of Fourier correlation unit sets a learnable weight for each frequency position, and the sixty-four channels at that frequency position are weighted and combined to obtain the frequency domain response. Channel hybrid computation is performed on the frequency domain response. A weighted connection of 64 x 64 is used to perform linear combination of channels at each frequency position and nonlinear transformation. The updated frequency domain representation is output and used as the input of the next level Fourier correlation unit until the correlation calculation result is obtained. The correlation calculation results are input into the time waveform restoration layer and the time waveform is restored. The number of predicted sampling points is determined based on the duration of the adjacent time period and the control cycle in the adjacent time period parameter of the commutation, and a time series representation consistent with the number of predicted sampling points is generated. The time series representation is compressed into a single-channel perturbation representation by point-by-point linear mapping. A two-level fully connected neuron group is used to map the sixty-four channels into sixteen channels and then map the sixteen channels into single channels. The commutation shock perturbation prediction sequence is output in chronological order without outputting control input commands.

6. The method for suppressing commutation loading force impact based on Fourier neural operators according to claim 5, characterized in that, The commutation neighbor state segment is used as the input to the Fourier neural operator. On the input side, velocity direction change features are generated based on the actuator velocity, and command direction change features are generated based on the applied force command. These velocity direction change features and command direction change features, along with the three-channel features of the commutation neighbor state segment, participate in subsequent mapping calculations. The velocity direction change features and command direction change features are obtained by calculating a change function, specifically: ； in, For the first The velocity direction change characteristics of each sampling point For the first The command direction change characteristics of each sampling point For the first Executor speed at each sampling point For the first Executor speed at each sampling point For the first The force command for each sampling point For the first The force command for each sampling point The threshold for zero speed determination. A threshold for determining the direction of the instruction. This is an indicator function that takes the value of 1 when the condition within the parentheses is true and 0 when the condition is false. The controller will at the first sampling point and Set it to zero and recursively calculate from the second sampling point in the manner described. The controller will and With the same sampling point , , They jointly participate in subsequent mapping calculations, enabling the point-by-point linear mapping to obtain a computable source of input difference at the sampling points where the direction of the applied force command is switched and the direction of the actuator velocity is reversed.

7. The method for suppressing commutated loading force impact based on Fourier neural operators according to claim 1, characterized in that, S5 specifically refers to: Using the loading force feedback and the control input command of the previous control cycle as inputs to the extended state observer, the observer state is updated and the total disturbance estimate is output. Based on the start time of the reversal and the duration of the reversal adjacent time period in the reversal adjacent time period parameters, and combined with the current time, determine whether the current time meets the reversal adjacent time period conditions; When the current time meets the condition of the adjacent time period of the reversal, the position of the current time in the adjacent time period of the reversal is determined according to the time difference between the current time and the starting time of the reversal and the control cycle, and the number of predicted sampling points is determined by the duration of the adjacent time period of the reversal and the control cycle. Based on the current position within the adjacent time period of the reversal, a corresponding disturbance value aligned with the current position is selected from the reversal impact disturbance prediction sequence. The corresponding disturbance value is the single-channel disturbance prediction value of the reversal impact disturbance prediction sequence. A preset impact fusion coefficient is used to linearly fuse the total disturbance estimate and the corresponding disturbance value to obtain the impact compensation disturbance amount, which is used as the disturbance compensation input for active disturbance rejection control. When the conditions for the adjacent period of reversal are met at the current moment, the loading force command and loading force feedback are used to generate the loading force error, and the control input command is calculated based on the loading force error, the impact compensation disturbance amount and the control input command of the previous control cycle. If the conditions for the adjacent period of reversal are not met at the current moment, the loading force command and loading force feedback are used to generate the loading force error, and the control input command is calculated based on the loading force error, the total disturbance estimate and the control input command of the previous control cycle.

8. The method for suppressing commutated loading force impact based on Fourier neural operators according to claim 7, characterized in that, The load force feedback and the control input command of the previous control cycle are used as inputs to the extended state observer, and the total disturbance estimate is output. Based on the commutation start time and the duration of the commutation adjacent time period in the commutation adjacent time period parameters, and combined with the current time, it is determined whether the current time meets the commutation adjacent time period condition. If the current time meets the commutation adjacent time period condition, the corresponding disturbance value is selected from the commutation impact disturbance prediction sequence according to the current time's position within the commutation adjacent time period. The total disturbance estimate and the corresponding disturbance value are linearly fused using a preset impact fusion coefficient to obtain the impact compensation disturbance amount. If the commutation adjacent time period condition is met or not, the control input command is calculated based on the load force error, the impact compensation disturbance amount, or the total disturbance estimate and the control input command of the previous control cycle, respectively. The above calculations are represented by a synthesis function, which is specifically: ; in, For the applied force error, For loading force command, For force feedback, The result of determining the condition of the reversal period is given and the value is [value]. or , For the current moment, The starting point of the reversal, The duration of the adjacent time period for the reversal, For indicator functions, To predict the number of sampling points, To control the cycle, For floor operations, For floor operations, To perform the minimum value operation, To perform the maximum value operation, This is the position index of the current time within the adjacent time period of the reversal. For the prediction sequence of commutation shock disturbance The Middle Single-channel perturbation prediction value for each sampling point For the corresponding disturbance value, This is the total perturbation estimate from the output of the extended state observer. To preset the impact fusion coefficient, This is the amount of disturbance to compensate for the impact. This is the control input command for the previous control cycle. For the current control cycle, control input commands are provided. , , To control the gain, For equivalent control gain, and These represent the loading force errors of the previous control cycle and the two control cycles prior, respectively. , , The observer gain of the extended state observer is applied to the error correction of the loading force state estimation channel, the loading force rate of change estimation channel, and the total disturbance estimation channel, respectively.

9. The method for suppressing commutated loading force impact based on Fourier neural operators according to claim 1, characterized in that, Step S6 is as follows: The control input command is converted into an actuator drive signal via the drive interface and applied to the actuator, driving the loading mechanism to generate a loading force in the current control cycle; Within the sampling window of the next control cycle, the force sensor collects and updates the loading force feedback, and the encoder collects and updates the actuator speed. The time deviation between the sampling time of updating the loading force feedback and the sampling time of updating the actuator speed is less than the time alignment threshold. The updated loading force feedback, updated actuator speed, loading force command of the current control cycle, and control input command of the current control cycle are combined into sampling points in the order of the control cycle; The sampling points are written into the sliding window of the currently loaded sampling sequence and form the next control cycle loading sampling sequence, which serves as the input for subsequent commutation adjacent time period parameter determination and commutation adjacent state segment extraction.

Images