Grading type rotary closed-loop regulation and control method and system based on fuzzy neural network

Abstract

The invention discloses a progressive rotary closed-loop regulation and control method and system based on a fuzzy neural network, and the method comprises the following steps: collecting the angle and angular velocity of a rotary system and the feedback of a driving side, carrying out the alignment verification, and forming state, disturbance and credibility information; stage boundary self-evolution is carried out, and regulation and control stages are divided; inputting the state into the adaptive fuzzy neural network according to the current stage to obtain rule activation, candidate output and model credibility; weighting the rule in combination with the credibility and the stability to generate a rule weight; according to rule weight fusion control output, limited residual error micro-correction is carried out, and a control instruction is formed; and executing the control instruction and obtaining feedback to realize closed-loop regulation and control of self-correction and rollback. The adaptive fuzzy neural network and the progressive rotation closed-loop regulation and control method are adopted, staged stable control and self-correction rollback of the rotation system are achieved, and the method has the advantages of being high in robustness, stability and control precision.

Description

Technical Field

[0001] This invention relates to the field of industrial automation equipment, and in particular to a progressive rotary closed-loop control method and system based on fuzzy neural networks. Background Technology

[0002] In the field of existing rotary system control, methods such as proportional-integral-derivative (PID) control, adaptive control, fuzzy control, and neural network control are widely used for closed-loop regulation of rotating objects such as motors, turntables, and servo actuators. Some schemes compensate for system nonlinearity and uncertainty by introducing fuzzy rules or neural networks, while others combine multi-sensor feedback to jointly control angle, angular velocity, or drive-side signals. However, the above-mentioned existing technologies usually assume that the system operating state is relatively stable on the time scale, and the control strategy maintains a uniform structure across the entire operating range. They lack a clear distinction and dynamic switching mechanism for different control stages, making it difficult to balance rapid convergence of large errors with high-precision stability of small errors.

[0003] On the other hand, existing technologies often employ independent designs for handling sensor reliability, sudden disturbances, and closed-loop stability. Common practices include simply filtering out abnormal data or applying fixed limits to control inputs, lacking a unified weighting framework that systematically couples sensor reliability, model reliability, and stability risk. Furthermore, while some solutions introduce adaptive or online learning mechanisms, they lack clear and executable triggering conditions and fallback paths for parameter updates and anomaly rollback. This makes them prone to control oscillations or performance degradation under complex disturbances or sensor degradation, hindering the formation of a safe, stable, and self-calibrating rotating closed-loop control system.

[0004] Therefore, how to provide a progressive rotational closed-loop control method and system based on fuzzy neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a progressive rotational closed-loop control method and system based on fuzzy neural networks. This invention employs an adaptive fuzzy neural network and a progressive rotational closed-loop control method to achieve staged stable control and self-correction backoff of the rotational system, possessing the advantages of strong robustness, high stability, and high control accuracy.

[0006] A stepwise rotational closed-loop control method based on a fuzzy neural network according to an embodiment of the present invention includes the following steps: The system acquires the angle feedback, angular velocity feedback and drive-side feedback of the rotating system and performs time alignment and consistency verification to generate a set of rotational states, a set of disturbance characterizations and a set of sensing reliability. The set of rotational states, the set of disturbance characterizations, and the set of sensing reliability are input into the boundary self-evolution unit, and the set of progressive control stages and the set of stage boundaries are output. The current control stage is determined based on the set of progressive control stages and the set of stage boundaries. The set of rotation states and the set of sensing confidence are input into the adaptive Takagi–Sugeno fuzzy neural network, which outputs the set of rule activation intensity, the set of local model candidate outputs, and the set of model confidence. The rule activation intensity set, the sensor credibility set, and the model credibility set are input into a credibility-stability coupled rule activation weighting mechanism to generate a rule weight set. Based on the set of rule weights, the candidate output set of the local model is continuously weighted and fused, and a set of stage control instructions is generated through a constrained residual micro-correction structure. The stage control instruction set is applied to the rotating system and closed-loop feedback is collected to trigger a rotating closed-loop intelligent control system with self-correction and back-off capabilities. Parameter update information is generated and fed back to the boundary self-evolution unit and the adaptive Takagi–Sugeno fuzzy neural network.

[0007] Optionally, the generation of the rotation state set, the disturbance characterization set, and the sensing reliability set specifically includes: The original sequences of angles, angular velocities, and drive side of the rotating system are collected, and the collected time stamp information is added to obtain the angle time stamp sequence, angular velocity time stamp sequence, and drive side time stamp sequence; Perform deburring, saturated segment removal and short-term missing filling to obtain the angle cleaning sequence, angular velocity cleaning sequence and drive-side cleaning sequence, and write the saturated segment position set and the missing filling position set into the quality mark set. Using the time scale corresponding to the angle cleaning sequence as the alignment reference, the angular velocity cleaning sequence and the driving side cleaning sequence are aligned based on time offset estimation to obtain the angular velocity aligned sequence and the driving side aligned sequence; Perform unified resampling and bandpass constraint filtering on the angle cleaning sequence, angular velocity alignment sequence and drive-side alignment sequence to generate the angle alignment sequence, angular velocity unified sequence and drive-side unified sequence. Write the resampling factor and the effective segment index of the filter into the alignment configuration set. A kinematic consistency check is constructed based on the angle alignment sequence and the angular velocity unified sequence. The angle consistency residual sequence and the angular velocity consistency residual sequence are calculated. The set of consistency abnormal segments is determined based on the continuous out-of-limit segments of the residuals. The driver-side execution representation sequence is extracted based on the unified sequence of the driver side, and a perturbation representation set is jointly constructed with the angular velocity consistent residual sequence. The sensing confidence level is calculated based on the quality label set, the consistency anomaly fragment set, and the perturbation characterization set. The confidence level corresponding to the consistency anomaly fragment set is then suppressed and updated to obtain the sensing confidence level set. The angle alignment sequence, the unified angular velocity sequence, and the unified drive-side sequence are fused into a rotation state set according to the same time scale. The disturbance characterization set and the sensing reliability set are indexed and bound to the rotation state set.

[0008] Optionally, the generation of the progressive control stage set and the stage boundary set specifically includes: The set of rotational states is windowed and sliced, and then time-aligned with the set of disturbance characterization and the set of sensing reliability to obtain the set of stage determination features. In the boundary self-evolution unit, the error amplitude statistics, error change rate statistics, and disturbance intensity statistics are calculated for the stage judgment feature set, and then multiplicatively weighted with the sensing credibility set to generate a credibility weighted feature set. A candidate set for progressive control stages is constructed based on a set of credible weighted features, and the set of stage membership degrees is calculated for each window. The set of progressive control stages is determined based on the set of stage membership degrees, and the initial values ​​of the stage boundaries are calculated and written into the set of stage boundaries. Perform boundary self-evolution update on the stage boundary set to generate a boundary update quantity set and obtain the updated stage boundary set; Hysteresis locking and dwell constraints are applied to the updated stage boundary set to generate a boundary stability constraint set, and the progressive control stage set is modified to obtain the final progressive control stage set. Bind the final set of progressive control stages to the updated set of stage boundaries using time-stamped indexes.

[0009] Optionally, the generation of the rule activation intensity set, the local model candidate output set, and the model credibility set specifically includes: Read the set of progressive control stages and the set of stage boundaries, determine the current control stage at the current window index, output the current control stage identifier and write it to the stage index binding set; Based on the current control stage identifier, the set of rotational states is extracted in stages and scaled uniformly to obtain the stage input set, which is then multiplicatively fused with the sensing confidence set according to the same window index to generate a reliable stage input set. In the adaptive Takagi–Sugeno fuzzy neural network, a set of antecedent conditions is constructed based on the set of credible stage inputs and the set of stage boundaries. Write the set of antecedent conditions into the rule base set to form the antecedent condition part of the rule base set, thus completing the initial construction of the rule base set; For each antecedent condition in the rule base set, a corresponding consequent local linear model set is established based on the trusted stage input set and bound to the antecedent condition; Based on the membership values ​​of the input set in the trusted stage under each antecedent condition, calculate the rule activation intensity of each rule in the current window, perform normalization processing, and generate a set of rule activation intensities; For each rule in the rule base set, perform candidate output calculations on the consequent local linear model set to generate a local model candidate output set; A set of prediction residuals is constructed based on the local model candidate output set and the rotation state set, and a set of model confidence is calculated. The set of rule activation strengths, the set of local model candidate outputs, and the set of model credibility are time-stamped and then output.

[0010] Optionally, the generation of the rule weight set specifically includes: Read the set of rule activation intensity, the set of local model candidate outputs, and the set of model confidence, and simultaneously read the set of sensor confidence. Align them according to the rule index to generate the set of rule evaluation input. A closed-loop stability evaluation feature set is constructed based on the local model candidate output set and the rotation state set. The feature set is then bound to the corresponding local model candidate output set according to the rule index to generate a rule-level stability feature set. Stability scores are calculated on the rule-level stability feature set to generate a stability score set. The first stage of weighting is performed on the set of rule activation strengths to generate an initial set of rule weights; The initial rule weight set and the sensor confidence set are modified by confidence constraint correction to generate a confidence-corrected rule weight set. Input the set of trustworthy correction rule weights and the set of stability scores into the trustworthiness-stability coupling operator to generate the set of stability coupling rule weights. Perform rule-level smoothing on the stability coupling rule weight set to generate a rule weight set, and write it into the rule weight set storage area; The rule weight set is time-stamped and bound to the local model candidate output set, and the stability score set and the reliable correction rule weight set are written into the feedback buffer.

[0011] Optionally, the generation of the stage control instruction set specifically includes: Read the set of rule weights and the set of candidate outputs of the local model, align them according to the rule index and the time index, and generate a fused and aligned input set; Perform continuous weighted fusion based on the fusion-aligned input set to generate a fusion control output set; The fusion control output set and the rotation state set are time-stamped and bound together to construct a fusion prediction residual set, which is then written into the residual buffer. Read the sensor confidence set and stability score set, align them with the fusion prediction residual set, and generate a micro-correction trigger judgment set; In the constrained residual micro-correction structure, a micro-correction configuration set is generated based on the micro-correction trigger determination set and the current control stage identifier; Based on the micro-correction configuration set, constrained residual micro-correction is performed on the fused prediction residual set to generate a micro-correction quantity set; The set of micro-correction values ​​and the set of fused control outputs are superimposed and stage consistency constraints are executed to generate a set of stage control instructions, which are then written into the stage control instruction set storage area. The set of stage control instructions is time-stamped and bound to the current control stage identifier, the set of fusion prediction residuals, and the set of micro-correction trigger judgments, and then output. The set of fusion control outputs, the set of micro-correction quantities, and the set of micro-correction configurations are written into the feedback buffer.

[0012] Optionally, the generation of the parameter update specifically includes: The set of stage control commands is sent to the drive execution interface of the rotating system and the timing of the command action is recorded. Angle feedback, angular velocity feedback and drive-side feedback are collected to obtain a closed-loop feedback set, which is indexed and bound to the set of stage control commands according to the same timing and written into the feedback buffer. Read the fusion prediction residual set in the residual buffer, the stability score set and the sensor reliability set in the feedback buffer, and calculate the closed-loop consistency evaluation set by combining the closed-loop feedback set; Generate a trigger judgment set based on the closed-loop consistency evaluation set and output the self-correction trigger identifier and the rollback trigger identifier; Under the condition that the self-calibration trigger flag is valid and the rollback trigger flag is invalid, parameter update information is generated based on the closed-loop consistency evaluation set, and indexed and bound with the current control stage flag and written into the parameter update information cache. If the rollback trigger flag is valid, generate a set of rollback actions and execute the rollback. The backoff configuration formed by the parameter update information or backoff action set is written into the feedback buffer and fed back to the boundary self-evolution unit and the adaptive Takagi-Sugeno fuzzy neural network. When the parameter update information is written, it is used to update the boundary update quantity set generation strategy, membership function set and consequent local linear model set update gating. When the backoff configuration is written, it is used to reset the input state of stage judgment and rule weighting and continue to output the stage control instruction set in the next window.

[0013] A cascaded rotational closed-loop control system based on a fuzzy neural network according to an embodiment of the present invention includes: The rotation state acquisition module is used to acquire the angle feedback, angular velocity feedback and drive-side feedback of the rotation system, and generate the corresponding rotation state data. The stage determination module is used to generate progressive control stages and stage boundaries based on rotation state data, disturbance characterization data, and confidence information, and output the current control stage. The fuzzy neural network inference module is used to perform fuzzy rule inference on the rotation state data under the current control stage, and output the rule activation strength, local model candidate output, and model credibility. The rule weighting module is used to generate rule weights based on rule activation strength, model credibility, and sensor credibility. The control instruction generation module is used to continuously weight and fuse the candidate outputs of the local model and perform constrained residual micro-correction to generate stage control instructions. The closed-loop feedback and update module is used to apply stage control commands to the rotating system, collect closed-loop feedback, and generate parameter update or rollback control information to achieve closed-loop regulation.

[0014] The beneficial effects of this invention are: This invention performs time alignment and kinematic consistency verification on angle feedback, angular velocity feedback, and drive-side feedback to form a set of rotational states, a set of disturbance characteristics, and a set of sensing reliability. Furthermore, it generates a set of progressive control stages and a set of stage boundaries within the boundary self-evolution unit. This enables the control process to dynamically determine candidates for coarse adjustment stages, transition stages, and fine adjustment stages based on the trends of error amplitude statistics, error rate of change statistics, and load disturbance intensity. This avoids stage misjudgments and frequent switching caused by fixed segments or single thresholds in existing technologies, thereby maintaining the stability and continuity of stage determination under different operating conditions, disturbances, and sensing quality fluctuations.

[0015] Meanwhile, based on the set of rule activation intensity, local model candidate output set, and model credibility set of the adaptive Takagi–Sugeno fuzzy neural network, this invention introduces a credibility-stability coupled rule activation weighting mechanism to generate a set of rule weights, and performs continuous weighted fusion on the local model candidate output set. Furthermore, through a constrained residual micro-correction structure, a set of staged control commands is generated under constraints such as limited micro-correction amplitude, limited micro-correction rate of change, limited micro-correction start / stop conditions, and limited micro-correction trigger conditions. This ensures that the control output maintains the continuity of error convergence while suppressing jitter and overshoot. In closed-loop feedback and updates, parameter update information or a set of backoff actions is generated based on the combination of the residual over-limit segment set, the stability low-score segment set, and the credibility suppression segment set. This achieves gated updates and abnormal backoff of the rule base set and the consequent local linear model set, solving the problem of lack of self-correction and backoff links for sensing distortion, disturbance mutations, and model mismatch in existing technologies, thereby improving the robustness, stability, and long-term operational consistency of rotational closed-loop control. Attached Figure Description

[0016] 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 progressive rotational closed-loop control method based on a fuzzy neural network proposed in this invention. Figure 2 This is a schematic diagram of the adaptive Takagi–Sugeno fuzzy neural network, rule weighting, and control command generation structure of the progressive rotational closed-loop control method based on fuzzy neural network proposed in this invention; Figure 3 This invention presents a flowchart illustrating the kinematic consistency verification and disturbance characterization generation process of a progressive rotational closed-loop control method based on a fuzzy neural network, comprising a set of rotational states, a set of disturbance characterizations, and a set of sensing reliability. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A stepwise rotational closed-loop control method based on a fuzzy neural network includes the following steps: The system acquires the angle feedback, angular velocity feedback and drive-side feedback of the rotating system and performs time alignment and consistency verification to generate a set of rotational states, a set of disturbance characterizations and a set of sensing reliability. The set of rotational states, the set of disturbance characterizations, and the set of sensing reliability are input into the boundary self-evolution unit, and the set of progressive control stages and the set of stage boundaries are output. The current control stage is determined based on the set of progressive control stages and the set of stage boundaries. The set of rotation states and the set of sensing confidence are input into the adaptive Takagi–Sugeno fuzzy neural network, which outputs the set of rule activation intensity, the set of local model candidate outputs, and the set of model confidence. The rule activation intensity set, the sensor credibility set, and the model credibility set are input into a credibility-stability coupled rule activation weighting mechanism to generate a rule weight set. Based on the set of rule weights, the candidate output set of the local model is continuously weighted and fused, and a set of stage control instructions is generated through a constrained residual micro-correction structure. The stage control instruction set is applied to the rotating system and closed-loop feedback is collected to trigger a rotating closed-loop intelligent control system with self-correction and back-off capabilities. Parameter update information is generated and fed back to the boundary self-evolution unit and the adaptive Takagi–Sugeno fuzzy neural network.

[0019] In this embodiment, the generation of the rotation state set, the disturbance characterization set, and the sensing reliability set specifically includes: The original sequences of angles, angular velocities, and drive side of the rotating system are collected, and the collected time stamp information is added to obtain the angle time stamp sequence, angular velocity time stamp sequence, and drive side time stamp sequence; Among them, the original angle sequence is collected from the angle feedback channel of the rotating system, the original angular velocity sequence is collected from the angular velocity feedback channel, and the original drive-side sequence is collected from the drive-side feedback channel. Perform deburring, saturated segment removal and short-term missing filling to obtain the angle cleaning sequence, angular velocity cleaning sequence and drive-side cleaning sequence, and write the saturated segment position set and the missing filling position set into the quality mark set. Using the time scale corresponding to the angle cleaning sequence as the alignment reference, the angular velocity cleaning sequence and the driving side cleaning sequence are aligned based on time offset estimation to obtain the angular velocity aligned sequence and the driving side aligned sequence; The alignment based on time offset estimation is obtained by searching within the candidate time offset range to maximize the cumulative sum of the first-order difference sequences of the angular velocity cleaning sequence and the angle cleaning sequence at the corresponding time scale; Perform unified resampling and bandpass constraint filtering on the angle cleaning sequence, angular velocity alignment sequence and drive-side alignment sequence to generate the angle alignment sequence, angular velocity unified sequence and drive-side unified sequence. Write the resampling factor and the effective segment index of the filter into the alignment configuration set. A kinematic consistency check is constructed based on the angle alignment sequence and the angular velocity unified sequence. The angle consistency residual sequence and the angular velocity consistency residual sequence are calculated. The set of consistency abnormal segments is determined based on the continuous out-of-limit segments of the residuals. The kinematic consistency check is obtained by incrementally integrating the unified angular velocity sequence between adjacent time scales over time intervals and comparing it with the angle-aligned sequence using a difference comparison. The driver-side execution representation sequence is extracted based on the unified sequence of the driver side, and a perturbation representation set is jointly constructed with the angular velocity consistent residual sequence. The disturbance characterization set includes a load disturbance intensity sequence, a frictional mutation indication sequence, and a drive-side fluctuation indication sequence. The load disturbance intensity sequence is obtained by combining the absolute value of the proportional mapping difference between the drive-side execution characterization sequence and the angular velocity unified sequence with the absolute value of the angle-consistent residual sequence according to the residual coupling coefficient. The sensing confidence level is calculated based on the quality label set, the consistency anomaly fragment set, and the perturbation characterization set. The confidence level corresponding to the consistency anomaly fragment set is then suppressed and updated to obtain the sensing confidence level set. The sensing reliability set includes an angle reliability sequence, an angular velocity reliability sequence, and a drive-side reliability sequence. The angle reliability sequence is obtained by applying an exponential decay penalty to the absolute value of the angle consistency residual sequence and the label value of the quality label set, and then performing multiplicative fusion. The angle alignment sequence, the unified angular velocity sequence, and the unified drive-side sequence are fused into a rotation state set according to the same time scale. The disturbance characterization set and the sensing reliability set are indexed and bound to the rotation state set.

[0020] In this embodiment, the generation of the progressive control stage set and the stage boundary set specifically includes: The set of rotational states is windowed and sliced, and then time-aligned with the set of disturbance characterization and the set of sensing reliability to obtain the set of stage determination features. The set of stage judgment features is obtained by concatenating the angle alignment sequence, angular velocity unification sequence, driving side unification sequence, load disturbance intensity sequence, friction mutation indication sequence, driving side fluctuation indication sequence, and angle confidence sequence, angular velocity confidence sequence, and driving side confidence sequence according to the same window index. In the boundary self-evolution unit, the error amplitude statistics, error change rate statistics, and disturbance intensity statistics are calculated for the stage judgment feature set, and then multiplicatively weighted with the sensing credibility set to generate a credibility weighted feature set. The error amplitude statistics are obtained by multiplying the absolute values ​​of the angle deviation and angular velocity deviation of each sampling point within the same window by the corresponding values ​​of the angle confidence sequence and angular velocity confidence sequence, and then summing them up. The angle deviation is the deviation value of the angle alignment sequence relative to the target angle, and the angular velocity deviation is the deviation value of the angular velocity unification sequence relative to the target angular velocity. A candidate set for progressive control stages is constructed based on a set of credible weighted features, and the set of stage membership degrees is calculated for each window. The candidate set for the progressive control stage is automatically generated based on the dominant relationship and distribution of error amplitude statistics, error rate of change statistics and disturbance intensity statistics in the credible weighted feature set, forming candidates for the coarse adjustment stage, the transition stage, and the fine adjustment stage. The set of stage membership degrees is obtained by applying stage sensitivity coefficients to the stage discriminant values ​​of each stage, taking the negative values, and then performing exponential mapping. The result is then normalized by the sum of the three-stage exponential mapping values. The stage discriminant values ​​are obtained by normalizing and fusing the error amplitude statistics, error change rate statistics, and disturbance intensity statistics. The set of progressive control stages is determined based on the set of stage membership degrees, and the initial values ​​of the stage boundaries are calculated and written into the set of stage boundaries. The initial value of the stage boundary is jointly determined by the quantile threshold of the error amplitude statistic on the window sequence and the quantile threshold of the disturbance intensity statistic on the window sequence; Perform boundary self-evolution update on the stage boundary set to generate a boundary update quantity set and obtain the updated stage boundary set; The set of boundary update quantities is jointly driven by the sliding trend of the error magnitude statistics, the sliding trend of the error change rate statistics, and the sliding trend of the load disturbance intensity sequence. The updated stage boundary set is obtained by superimposing the boundary update step size coefficient and the boundary update quantity benchmark on the boundary value of the previous window to obtain candidate boundary values. Then, the candidate boundary values ​​are subjected to the amplitude constraint of the lower boundary value and the upper boundary value. The boundary update quantity benchmark is obtained by fusing the error amplitude statistical trend, the error change rate statistical trend and the load disturbance intensity trend in the same direction. Hysteresis locking and dwell constraints are applied to the updated stage boundary set to generate a boundary stability constraint set, and the progressive control stage set is modified to obtain the final progressive control stage set. The boundary self-evolution update is limited to three factors: the trend of error amplitude statistics, the trend of error change rate statistics, and the trend of load disturbance intensity. The update process is constrained by boundary value limiting, hysteresis locking, and the number of shortest dwell windows. The hysteresis locking and dwell constraint are achieved by setting an entry threshold and an exit threshold and combining them with the shortest dwell window number. When the stage judgment quantity meets the entry threshold and the shortest dwell window number is met consecutively, the stage judgment result is locked as an entry state. When the stage judgment quantity meets the exit threshold and the shortest dwell window number is met consecutively, the stage judgment result is locked as an exit state. In other cases, the stage judgment result of the previous window remains unchanged. Bind the final set of progressive control stages to the updated set of stage boundaries using time-stamped indexes.

[0021] In this embodiment, the generation of the rule activation intensity set, the local model candidate output set, and the model credibility set specifically includes: Read the set of progressive control stages and the set of stage boundaries, determine the current control stage at the current window index, output the current control stage identifier and write it to the stage index binding set; The current control stage is determined by the dominant stage corresponding to the stage membership set and the hysteresis locking result; Based on the current control stage identifier, the set of rotational states is extracted in stages and scaled uniformly to obtain the stage input set, which is then multiplicatively fused with the sensing confidence set according to the same window index to generate a reliable stage input set. The stage input set is obtained by splicing together segments of the angle-aligned sequence, the angular velocity unified sequence, and the driving-side unified sequence within the current window after processing them with the same normalization rule. In the adaptive Takagi–Sugeno fuzzy neural network, a set of antecedent conditions is constructed based on the set of credible stage inputs and the set of stage boundaries. The fuzzy rule antecedent structure performs fuzzy partitioning on the angle alignment sequence, angular velocity unified sequence, and driving-side unified sequence in the trusted stage input set based on the stage boundary set, generating corresponding antecedent fuzzy interval sets. A membership function set is established for each fuzzy interval set. The center position and coverage of the membership function set are limited by the stage boundary set and are adjusted synchronously with the update of the stage boundary set. The fuzzy intervals corresponding to the angle alignment sequence, angular velocity unified sequence, and driving-side unified sequence within the same window are combined to generate a multidimensional antecedent condition term set. Write the set of antecedent conditions into the rule base set to form the antecedent condition part of the rule base set, thus completing the initial construction of the rule base set; For each antecedent condition in the rule base set, a corresponding consequent local linear model set is established based on the trusted stage input set and bound to the antecedent condition; The consequent local linear model set describes the local linear mapping relationship between the set of rotational states and the control output under the constraints of the corresponding antecedent condition terms. Based on the membership values ​​of the input set in the trusted stage under each antecedent condition, calculate the rule activation intensity of each rule in the current window, perform normalization processing, and generate a set of rule activation intensities; For each rule in the rule base set, perform candidate output calculations on the consequent local linear model set to generate a local model candidate output set; The local model candidate output set is obtained by mapping the credible stage input set through each consequent local linear model, and the current control stage identifier is written into the stage attribute field of the local model candidate output set to form a stage-consistent candidate output description; A set of prediction residuals is constructed based on the local model candidate output set and the rotation state set, and a set of model confidence is calculated. The predicted residual set is obtained by subtracting the local model candidate output set within the current window from the corresponding angular velocity unified sequence fragment in the rotation state set; The model confidence set is obtained by applying exponential decay penalties to the statistical results of the residual magnitude and the statistical results of the residual change rate of the prediction residual set and then multiplicatively fusing them. The set of rule activation strengths, the set of local model candidate outputs, and the set of model credibility are time-stamped and then output.

[0022] In this embodiment, the generation of the rule weight set specifically includes: Read the set of rule activation intensity, the set of local model candidate outputs, and the set of model confidence, and simultaneously read the set of sensor confidence. Align them according to the rule index to generate the set of rule evaluation input. A closed-loop stability evaluation feature set is constructed based on the local model candidate output set and the rotation state set. The feature set is then bound to the corresponding local model candidate output set according to the rule index to generate a rule-level stability feature set. The closed-loop stability evaluation feature set includes the control output change amplitude sequence, the control output change rate sequence, and the short-time fluctuation statistics of the angular velocity uniform sequence in the rotation state set; Stability scores are calculated on the rule-level stability feature set to generate a stability score set. The stability score set is obtained by applying a preset decay function to the short-time fluctuation statistics of the control output change amplitude sequence, the control output change rate sequence, and the angular velocity unified sequence and performing multiplicative fusion. The control output change amplitude sequence reflects the amplitude difference of candidate control outputs in adjacent windows, and the control output change rate sequence reflects the smoothness of the change of candidate control outputs on the time axis. The first stage of weighting is performed on the set of rule activation strengths to generate an initial set of rule weights; The initial rule weight set is obtained by multiplicatively fusing the rule activation intensity set and the model credibility set rule by rule, and then normalizing the fusion result. The initial rule weight set and the sensor confidence set are modified by confidence constraint correction to generate a confidence-corrected rule weight set. The set of reliable correction rule weights is obtained by applying a joint suppression factor related to the angle reliability sequence, angular velocity reliability sequence and drive side reliability sequence to the initial rule weights corresponding to each rule, thereby reducing the effective proportion of rule weights under the condition of insufficient sensing reliability. Input the set of trustworthy correction rule weights and the set of stability scores into the trustworthiness-stability coupling operator to generate the set of stability coupling rule weights. The credibility-stability coupling operator applies a monotonic constraint mapping of the stability score set to the credibility correction rule weight set at the rule level, so that the rule weights with lower stability scores are continuously compressed, while the rule weights with higher stability scores are maintained or enhanced. Perform rule-level smoothing on the stability coupling rule weight set to generate a rule weight set, and write it into the rule weight set storage area; The rule-level smoothing process suppresses weight jitter by imposing a rate-of-change limit on the weight changes of the same rule between adjacent windows; The rule weight set is time-stamped and bound to the local model candidate output set, and the stability score set and the reliable correction rule weight set are written into the feedback buffer.

[0023] In this embodiment, the generation of the stage control instruction set specifically includes: Read the set of rule weights and the set of candidate outputs of the local model, align them according to the rule index and the time index, and generate a fused and aligned input set; The fusion alignment input set includes the rule weight corresponding to each rule and the candidate control output sequence of that rule in the current window; Perform continuous weighted fusion based on the fusion-aligned input set to generate a fusion control output set; The continuous weighted fusion is obtained by weighting and summing the candidate control outputs of all rules at the same time scale according to the corresponding rule weights, and the fused control output set is subjected to amplitude consistency processing to keep the dimensions consistent with the control interface of the drive side feedback. The fusion control output set and the rotation state set are time-stamped and bound together to construct a fusion prediction residual set, which is then written into the residual buffer. The fusion prediction residual set is obtained by aligning and differencing the short-time change results of the unified sequence of angular velocities in the set of rotational states driven by the fusion control output set with the target angular velocity deviation within the current window. Read the sensor confidence set and stability score set, align them with the fusion prediction residual set, and generate a micro-correction trigger judgment set; The micro-correction trigger determination set includes a confidence deficiency determination result, a stability risk determination result, and a residual bias determination result. The confidence deficiency determination result is determined by the joint suppression state of the angle confidence sequence, the angular velocity confidence sequence, and the driving side confidence sequence. The stability risk determination result is determined by the continuous low-score segments of the stability score set, and the residual bias determination result is determined by the continuous same-direction bias segments of the fused prediction residual set. In the constrained residual micro-correction structure, a micro-correction configuration set is generated based on the micro-correction trigger determination set and the current control stage identifier; The micro-correction configuration set defines the restricted dimensions of micro-correction and serves as the sole constraint input for the restricted residual micro-correction structure. The restricted dimensions include micro-correction amplitude restriction, micro-correction rate of change restriction, micro-correction start / stop condition restriction, and micro-correction trigger condition restriction. Among them, the micro-correction amplitude restriction is determined by the stage boundary set corresponding to the current control stage identifier, which determines the upper and lower limits of the amplitude. The micro-correction rate of change restriction is determined by the suppression level corresponding to the stability risk judgment result, which determines the upper limit of the rate of change. The micro-correction start / stop condition restriction is determined by the confidence deficiency judgment result, which determines the start / stop gating state. The micro-correction trigger condition restriction is determined by the residual bias judgment result, which determines the trigger threshold and the number of duration windows. Based on the micro-correction configuration set, constrained residual micro-correction is performed on the fused prediction residual set to generate a micro-correction quantity set; The restricted residual micro-correction is obtained by generating a micro-correction amount opposite to the bias direction for the residual bias segment that meets the restricted micro-correction triggering conditions, under the condition that the micro-correction start and stop conditions are restricted, and then applying restrictions on the micro-correction amplitude and the micro-correction rate of change to the micro-correction amount in sequence. The set of micro-correction values ​​and the set of fused control outputs are superimposed and stage consistency constraints are executed to generate a set of stage control instructions, which are then written into the stage control instruction set storage area. The phase consistency constraint allows for larger amplitude values ​​but keeps the rate of change limited under the coarse adjustment phase candidate, allows smaller amplitude values ​​but strengthens the limitation on the rate of change under the fine adjustment phase candidate, and adopts a hybrid constraint method under the transition phase candidate. The set of stage control instructions is time-stamped and bound to the current control stage identifier, the set of fusion prediction residuals, and the set of micro-correction trigger judgments, and then output. The set of fusion control outputs, the set of micro-correction quantities, and the set of micro-correction configurations are written into the feedback buffer.

[0024] In this embodiment, the generation of the parameter update specifically includes: The set of stage control commands is sent to the drive execution interface of the rotating system and the timing of the command action is recorded. Angle feedback, angular velocity feedback and drive-side feedback are collected to obtain a closed-loop feedback set, which is indexed and bound to the set of stage control commands according to the same timing and written into the feedback buffer. Read the fusion prediction residual set in the residual buffer, the stability score set and the sensor reliability set in the feedback buffer, and calculate the closed-loop consistency evaluation set by combining the closed-loop feedback set; The closed-loop consistency evaluation set includes a residual excess segment set, a stability low-score segment set, and a credibility suppression segment set. The residual excess segment set is determined by the excess state of the amplitude statistics and change rate statistics of the fused prediction residual set within a continuous window. The stability low-score segment set is determined by the continuous low-score state of the stability score set. The credibility suppression segment set is determined by the joint suppression state of the angle credibility sequence, the angular velocity credibility sequence, and the driving side credibility sequence. Generate a trigger judgment set based on the closed-loop consistency evaluation set and output the self-correction trigger identifier and the rollback trigger identifier; The triggering conditions of the triggering judgment set are limited to at least two of the following three types: residual over-limit fragment set, stability low-score fragment set, and credibility suppression fragment set. At the same window index, the combined triggering condition is satisfied and continuously reaches the shortest dwell window number. When the combined trigger first appears during the dwell period corresponding to the current control stage, a self-correction triggering mark is generated. When the combined trigger repeatedly appears during the continuous dwell period or continues to exist under hysteresis lock state, a rollback triggering mark is generated. Under the condition that the self-calibration trigger flag is valid and the rollback trigger flag is invalid, parameter update information is generated based on the closed-loop consistency evaluation set, and indexed and bound with the current control stage flag and written into the parameter update information cache. The parameter update information includes the suppression level of the boundary update set of the stage boundary set, the fine-tuning configuration of the center position and coverage of the membership function set, the update gating configuration of the consequent local linear model set, and the configuration of the change rate constraint strength of the rule-level smoothing process. If the rollback trigger flag is valid, generate a set of rollback actions and execute the rollback. The rollback action set includes stage boundary set rollback action, rule weight set rollback action, rule base set freeze action, and micro-correction configuration set shutdown action. The stage boundary set rollback action is used to roll back the stage boundary set to the updated stage boundary set when the previous hysteresis lock took effect and keep the boundary stability constraint set unchanged. The rule weight set rollback action is used to roll back the rule weight set to the version written to the rule weight set storage area in the previous window and keep the change rate limit of rule-level smoothing. The rule base set freeze action is used to freeze the updates of the rule base set and the consequent local linear model set in the subsequent window. The micro-correction configuration set shutdown action is used to set the micro-correction start / stop condition to the off state to suspend the restricted residual micro-correction. The backoff configuration formed by the parameter update information or backoff action set is written into the feedback buffer and fed back to the boundary self-evolution unit and the adaptive Takagi-Sugeno fuzzy neural network. When the parameter update information is written, it is used to update the boundary update quantity set generation strategy, membership function set and consequent local linear model set update gating. When the backoff configuration is written, it is used to reset the input state of stage judgment and rule weighting and continue to output the stage control instruction set in the next window.

[0025] A cascaded rotational closed-loop control system based on a fuzzy neural network, comprising: The rotation state acquisition module is used to acquire the angle feedback, angular velocity feedback and drive-side feedback of the rotation system, and generate the corresponding rotation state data. The stage determination module is used to generate progressive control stages and stage boundaries based on rotation state data, disturbance characterization data, and confidence information, and output the current control stage. The fuzzy neural network inference module is used to perform fuzzy rule inference on the rotation state data under the current control stage, and output the rule activation strength, local model candidate output, and model credibility. The rule weighting module is used to generate rule weights based on rule activation strength, model credibility, and sensor credibility. The control instruction generation module is used to continuously weight and fuse the candidate outputs of the local model and perform constrained residual micro-correction to generate stage control instructions. The closed-loop feedback and update module is used to apply stage control commands to the rotating system, collect closed-loop feedback, and generate parameter update or rollback control information to achieve closed-loop regulation.

[0026] Example 1: To verify the feasibility of this invention in practice, it was applied to an online closed-loop control scenario of a high-precision industrial rotary actuator. This scenario is located in a long-term, continuously operating equipment manufacturing site. Under conditions of high-speed start-stop, frequent load changes, and long-term stable operation, the rotary system is prone to problems such as angle micro-oscillations, control overshoot, and sensor anomalies. Traditional fixed-parameter control methods struggle to simultaneously balance response speed and stability, and are prone to performance degradation when sensor noise or drive-side disturbances increase.

[0027] In practical applications, the system first continuously acquires angle feedback, angular velocity feedback, and drive-side feedback through a rotation state acquisition module. Within the control cycle, it completes time alignment and kinematic consistency verification, forming a rotation state set, a disturbance representation set, and a sensing reliability set. Subsequently, these sets are input into a boundary self-evolutionary unit, which automatically identifies different control stages during continuous operation and generates adaptively adjustable stage boundaries that change with operating conditions, enabling smooth switching between coarse, transitional, and fine-tuning states. Based on this, an adaptive Takagi–Sugeno fuzzy neural network, combined with the current control stage, performs rule-based reasoning on the rotation state, dynamically outputting rule activation strength, local model candidate outputs, and model reliability, preventing single-model failure under complex operating conditions.

[0028] In the control execution phase, the system uses a rule-weighted mechanism that couples reliability and stability to continuously weight and fuse candidate outputs of multiple fuzzy rules. It then uses a constrained residual micro-correction structure to controllably correct potential systematic deviations, thereby generating stage control commands and issuing them to the rotating system. After the control commands are executed, the system collects closed-loop feedback data in real time and performs a comprehensive consistency evaluation of the fused prediction residuals, stability scores, and sensor reliability. When abnormal trends are detected, a self-correction or backoff mechanism is automatically triggered to adjust stage boundaries, rule weights, and local model update strategies, ensuring that the control behavior always remains within a safe and stable closed-loop range.

[0029] Long-term comparisons of field operation data show that, within continuous operating cycles, the angular velocity fluctuations of the rotating system during multi-condition switching are significantly reduced, the angle convergence process is smoother, the switching during the control phase is more stable, and the impact of sensor anomalies on control performance is effectively suppressed. Simultaneously, the system maintains consistent control performance even under conditions of increased drive-side disturbances and frequent load changes, without exhibiting control oscillations or sudden performance drops due to parameter mismatch. These results demonstrate that the fuzzy neural network-based progressive closed-loop control method for rotating systems proposed in this invention can effectively solve the technical problem of balancing accuracy, stability, and robustness in complex operating conditions, possessing good engineering applicability and continuous operation capability.

[0030] Table 1. Comparison of overall performance of the cascaded rotating closed-loop control method based on fuzzy neural network.

[0031] As shown in Table 1, in terms of the core accuracy indicators of the rotating system, the maximum steady-state error of the method in this invention is significantly lower than that of the comparative methods. Traditional fixed-parameter control lacks the ability to adapt to disturbances and changes in operating conditions, resulting in a high level of steady-state error. Although conventional fuzzy control introduces rule-based reasoning, its rule weights are fixed and cannot be dynamically adjusted by combining model credibility and sensor credibility. This invention, through a credibility-stability coupled rule activation weighting mechanism, enables highly reliable and stable rules to dominate the fusion process, thereby significantly reducing the steady-state error.

[0032] In terms of angular velocity dynamic characteristics, this invention demonstrates significant advantages in both the peak angular velocity fluctuation amplitude and the average rate of change of control commands. This is mainly due to the synergistic effect of continuous weighted fusion and the constrained residual micro-correction structure, which ensures the control output maintains a fast response while strictly constraining the rate of change, thus preventing high-frequency jitter from being amplified in the closed loop.

[0033] From the perspective of operational stability, the continuous stable operation time of the method of this invention is significantly extended, while the number of switching between control stages is significantly reduced. This indicates that the progressive control stage set and the self-evolution mechanism of stage boundaries can reasonably divide the control stages according to the error, disturbance, and confidence status, avoiding frequent switching caused by fixed stage boundaries, and improving operational stability at the system level.

[0034] Regarding the ability to handle abnormal operating conditions, this invention achieves significant improvements in both the recovery time from sensor anomalies and the number of stable recovery windows after self-calibration triggering. This is directly related to the closed-loop consistency evaluation set, self-calibration triggering, and rollback mechanism introduced in the claims. By jointly judging residuals, stability, and reliability, the system can identify potential instability trends earlier and complete parameter correction or rollback within a shorter window, thereby shortening the duration of abnormal impacts.

[0035] Based on the above data, it can be confirmed that this invention achieves simultaneous improvements in multiple dimensions such as accuracy, stability, robustness, and continuous operation capability through collaborative modeling of the rotation state set, disturbance representation set, and sensing reliability set, combined with fuzzy neural network inference, rule weighting, restricted residual micro-correction, and closed-loop self-correction update mechanism. This verifies the comprehensive technical advantages of this method in complex rotation control scenarios.

[0036] 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 stepwise rotational closed-loop control method based on a fuzzy neural network, characterized in that, Includes the following steps: The system acquires the angle feedback, angular velocity feedback and drive-side feedback of the rotating system and performs time alignment and consistency verification to generate a set of rotational states, a set of disturbance characterizations and a set of sensing reliability. The set of rotational states, the set of disturbance characterizations, and the set of sensing reliability are input into the boundary self-evolution unit, and the set of progressive control stages and the set of stage boundaries are output. The current control stage is determined based on the set of progressive control stages and the set of stage boundaries. The set of rotation states and the set of sensing confidence are input into the adaptive Takagi–Sugeno fuzzy neural network, which outputs the set of rule activation intensity, the set of local model candidate outputs, and the set of model confidence. The rule activation intensity set, the sensor credibility set, and the model credibility set are input into a credibility-stability coupled rule activation weighting mechanism to generate a rule weight set. Based on the set of rule weights, the candidate output set of the local model is continuously weighted and fused, and a set of stage control instructions is generated through a constrained residual micro-correction structure. The stage control instruction set is applied to the rotating system and closed-loop feedback is collected to trigger a rotating closed-loop intelligent control system with self-correction and back-off capabilities. Parameter update information is generated and fed back to the boundary self-evolution unit and the adaptive Takagi–Sugeno fuzzy neural network.

2. The stepwise rotational closed-loop control method based on a fuzzy neural network according to claim 1, characterized in that, The generation of the rotation state set, the disturbance characterization set, and the sensing reliability set specifically includes: The original sequences of angles, angular velocities, and drive side of the rotating system are collected, and the collected time stamp information is added to obtain the angle time stamp sequence, angular velocity time stamp sequence, and drive side time stamp sequence; Perform deburring, saturated segment removal and short-term missing filling to obtain the angle cleaning sequence, angular velocity cleaning sequence and drive-side cleaning sequence, and write the saturated segment position set and the missing filling position set into the quality mark set. Using the time scale corresponding to the angle cleaning sequence as the alignment reference, the angular velocity cleaning sequence and the driving side cleaning sequence are aligned based on time offset estimation to obtain the angular velocity aligned sequence and the driving side aligned sequence; Perform unified resampling and bandpass constraint filtering on the angle cleaning sequence, angular velocity alignment sequence and drive-side alignment sequence to generate the angle alignment sequence, angular velocity unified sequence and drive-side unified sequence. Write the resampling factor and the effective segment index of the filter into the alignment configuration set. A kinematic consistency check is constructed based on the angle alignment sequence and the angular velocity unified sequence. The angle consistency residual sequence and the angular velocity consistency residual sequence are calculated. The set of consistency abnormal segments is determined based on the continuous out-of-limit segments of the residuals. The driver-side execution representation sequence is extracted based on the unified sequence of the driver side, and a perturbation representation set is jointly constructed with the angular velocity consistent residual sequence. The sensing confidence level is calculated based on the quality label set, the consistency anomaly fragment set, and the perturbation characterization set. The confidence level corresponding to the consistency anomaly fragment set is then suppressed and updated to obtain the sensing confidence level set. The angle alignment sequence, the unified angular velocity sequence, and the unified drive-side sequence are fused into a rotation state set according to the same time scale. The disturbance characterization set and the sensing reliability set are indexed and bound to the rotation state set.

3. The stepwise rotational closed-loop control method based on a fuzzy neural network according to claim 1, characterized in that, The generation of the progressive control stage set and stage boundary set specifically includes: The set of rotational states is windowed and sliced, and then time-aligned with the set of disturbance characterization and the set of sensing reliability to obtain the set of stage determination features. In the boundary self-evolution unit, the error amplitude statistics, error change rate statistics, and disturbance intensity statistics are calculated for the stage judgment feature set, and then multiplicatively weighted with the sensing credibility set to generate a credibility weighted feature set. A candidate set for progressive control stages is constructed based on a set of credible weighted features, and the set of stage membership degrees is calculated for each window. The set of progressive control stages is determined based on the set of stage membership degrees, and the initial values ​​of the stage boundaries are calculated and written into the set of stage boundaries. Perform boundary self-evolution update on the stage boundary set to generate a boundary update quantity set and obtain the updated stage boundary set; Hysteresis locking and dwell constraints are applied to the updated stage boundary set to generate a boundary stability constraint set, and the progressive control stage set is modified to obtain the final progressive control stage set. Bind the final set of progressive control stages to the updated set of stage boundaries using time-stamped indexes.

4. The stepwise rotational closed-loop control method based on a fuzzy neural network according to claim 1, characterized in that, The generation of the rule activation intensity set, the local model candidate output set, and the model credibility set specifically includes: Read the set of progressive control stages and the set of stage boundaries, determine the current control stage at the current window index, output the current control stage identifier and write it to the stage index binding set; Based on the current control stage identifier, the set of rotational states is extracted in stages and scaled uniformly to obtain the stage input set, which is then multiplicatively fused with the sensing confidence set according to the same window index to generate a reliable stage input set. In the adaptive Takagi–Sugeno fuzzy neural network, a set of antecedent conditions is constructed based on the set of credible stage inputs and the set of stage boundaries. Write the set of antecedent conditions into the rule base set to form the antecedent condition part of the rule base set, thus completing the initial construction of the rule base set; For each antecedent condition in the rule base set, a corresponding consequent local linear model set is established based on the trusted stage input set and bound to the antecedent condition; Based on the membership values ​​of the input set in the trusted stage under each antecedent condition, calculate the rule activation intensity of each rule in the current window, perform normalization processing, and generate a set of rule activation intensities; For each rule in the rule base set, perform candidate output calculations on the consequent local linear model set to generate a local model candidate output set; A set of prediction residuals is constructed based on the local model candidate output set and the rotation state set, and a set of model confidence is calculated. The set of rule activation strengths, the set of local model candidate outputs, and the set of model credibility are time-stamped and then output.

5. The stepwise rotational closed-loop control method based on a fuzzy neural network according to claim 1, characterized in that, The generation of the rule weight set specifically includes: Read the set of rule activation intensity, the set of local model candidate outputs, and the set of model confidence, and simultaneously read the set of sensor confidence. Align them according to the rule index to generate the set of rule evaluation input. A closed-loop stability evaluation feature set is constructed based on the local model candidate output set and the rotation state set. The feature set is then bound to the corresponding local model candidate output set according to the rule index to generate a rule-level stability feature set. Stability scores are calculated on the rule-level stability feature set to generate a stability score set. The first stage of weighting is performed on the set of rule activation strengths to generate an initial set of rule weights; The initial rule weight set and the sensor confidence set are modified by confidence constraint correction to generate a confidence-corrected rule weight set. Input the set of trustworthy correction rule weights and the set of stability scores into the trustworthiness-stability coupling operator to generate the set of stability coupling rule weights. Perform rule-level smoothing on the stability coupling rule weight set to generate a rule weight set, and write it into the rule weight set storage area; The rule weight set is time-stamped and bound to the local model candidate output set, and the stability score set and the reliable correction rule weight set are written into the feedback buffer.

6. The stepwise rotational closed-loop control method based on a fuzzy neural network according to claim 1, characterized in that, The generation of the stage control instruction set specifically includes: Read the set of rule weights and the set of candidate outputs of the local model, align them according to the rule index and the time index, and generate a fused and aligned input set; Perform continuous weighted fusion based on the fusion-aligned input set to generate a fusion control output set; The fusion control output set and the rotation state set are time-stamped and bound together to construct a fusion prediction residual set, which is then written into the residual buffer. Read the sensor confidence set and stability score set, align them with the fusion prediction residual set, and generate a micro-correction trigger judgment set; In the constrained residual micro-correction structure, a micro-correction configuration set is generated based on the micro-correction trigger determination set and the current control stage identifier; Based on the micro-correction configuration set, constrained residual micro-correction is performed on the fused prediction residual set to generate a micro-correction quantity set; The set of micro-correction values ​​and the set of fused control outputs are superimposed and stage consistency constraints are executed to generate a set of stage control instructions, which are then written into the stage control instruction set storage area. The set of stage control instructions is time-stamped and bound to the current control stage identifier, the set of fusion prediction residuals, and the set of micro-correction trigger judgments, and then output. The set of fusion control outputs, the set of micro-correction quantities, and the set of micro-correction configurations are written into the feedback buffer.

7. The stepwise rotational closed-loop control method based on a fuzzy neural network according to claim 1, characterized in that, The generation of the parameter update specifically includes: The set of stage control commands is sent to the drive execution interface of the rotating system and the timing of the command action is recorded. Angle feedback, angular velocity feedback and drive-side feedback are collected to obtain a closed-loop feedback set, which is indexed and bound to the set of stage control commands according to the same timing and written into the feedback buffer. Read the fusion prediction residual set in the residual buffer, the stability score set and the sensor reliability set in the feedback buffer, and calculate the closed-loop consistency evaluation set by combining the closed-loop feedback set; Generate a trigger judgment set based on the closed-loop consistency evaluation set and output the self-correction trigger identifier and the rollback trigger identifier; Under the condition that the self-calibration trigger flag is valid and the rollback trigger flag is invalid, parameter update information is generated based on the closed-loop consistency evaluation set, and indexed and bound with the current control stage flag and written into the parameter update information cache. If the rollback trigger flag is valid, generate a set of rollback actions and execute the rollback. The backoff configuration formed by the parameter update information or backoff action set is written into the feedback buffer and fed back to the boundary self-evolution unit and the adaptive Takagi-Sugeno fuzzy neural network. When the parameter update information is written, it is used to update the boundary update quantity set generation strategy, membership function set and consequent local linear model set update gating. When the backoff configuration is written, it is used to reset the input state of stage judgment and rule weighting and continue to output the stage control instruction set in the next window.

8. A tiered rotational closed-loop control system based on a fuzzy neural network, executing the tiered rotational closed-loop control method based on a fuzzy neural network as described in any one of claims 1 to 7, characterized in that, include: The rotation state acquisition module is used to acquire the angle feedback, angular velocity feedback and drive-side feedback of the rotation system, and generate the corresponding rotation state data. The stage determination module is used to generate progressive control stages and stage boundaries based on rotation state data, disturbance characterization data, and confidence information, and output the current control stage. The fuzzy neural network inference module is used to perform fuzzy rule inference on the rotation state data under the current control stage, and output the rule activation strength, local model candidate output, and model credibility. The rule weighting module is used to generate rule weights based on rule activation strength, model credibility, and sensor credibility. The control instruction generation module is used to continuously weight and fuse the candidate outputs of the local model and perform constrained residual micro-correction to generate stage control instructions. The closed-loop feedback and update module is used to apply stage control commands to the rotating system, collect closed-loop feedback, and generate parameter update or rollback control information to achieve closed-loop regulation.

Images