A control communication interface optimization method in a semi-physical simulation

Abstract

The application relates to a control communication interface optimization method in semi-physical simulation, in particular to the field of ship gas turbine semi-physical simulation, which effectively improves the timing accuracy and anti-interference of semi-physical simulation control through multi-technology fusion, realizes high-precision error collection by using an atomic clock marker, dynamically optimizes control instructions in combination with an intelligent algorithm, compensates for current transmission delay, prospectively suppresses future disturbance through a neural network, and finally outputs accurate frequency compensation through an adaptive fusion mechanism, thereby realizing signal alignment and closed-loop feedback, and significantly enhancing the stability and response speed of the system under complex working conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of ship gas turbine semi-physical simulation, more particularly, the present application relates to a control communication interface optimization method in semi-physical simulation. BACKGROUND

[0002] The ship gas turbine semi-physical simulation system is a key platform for propulsion power plant research and verification. It combines real-time simulation models with physical execution mechanisms to simulate dynamic control processes under real working conditions. In this system, the control platform generates adjustment instructions based on the calculation results of the simulation model, which are transmitted to the throttle execution mechanism via a communication interface, thereby driving the stepper motor to adjust the fuel valve opening degree and achieving precise control of the gas turbine speed. Typical application scenarios include ship sudden load increase or sudden load decrease conditions, such as the dynamic process of rapidly increasing the power turbine speed from 60% of the rated load to 80% of the rated load. In such scenarios, the real-time performance of the control instructions and the accuracy of the signal transmission directly determine the control quality of the system. However, the ship engine room environment is complex, with high-intensity electromagnetic interference, mechanical vibration, and temperature and humidity fluctuations, which pose serious challenges to the transmission reliability of control signals.

[0003] Currently, semi-physical simulation systems generally use analog signal transmission schemes, such as communication interfaces based on four to twenty milliampere current signals cooperating with independent signal conversion modules to complete instruction transmission. These modules usually include Schmidt trigger circuits, delay circuits, and voltage-controlled oscillation circuits, aiming to improve anti-interference ability and convert analog instructions into pulse signals recognizable by stepper motors. Although this scheme meets basic control requirements to some extent, it has inherent defects in failing to achieve collaborative optimization of control algorithms and signal conversion circuits. On the one hand, analog signal transmission itself is susceptible to electromagnetic interference, leading to instruction delay or distortion. On the other hand, the fixed parameter design of each functional circuit in the signal conversion module cannot adapt to changes in dynamic working conditions, especially when the active disturbance rejection control algorithm is used as the core controller, which has fast response characteristics, requiring the rise time to be controlled within a few seconds, which poses high requirements on the real-time performance of the communication interface. In existing technologies, the optimization of control algorithms and the design of communication interfaces are mutually exclusive. For example, the grey wolf optimization algorithm only adjusts the controller parameters without considering the impact of signal transmission delay on system dynamic performance, and the signal conversion module does not introduce a real-time feedback mechanism to compensate for instruction asynchrony. This design deficiency leads to significant timing deviations between control instructions and execution mechanism driving signals under high dynamic working conditions, causing power turbine speed overshoot, execution mechanism step loss, and other control quality deterioration phenomena, which may even trigger gas turbine protection shutdown, severely restricting the reliability and effectiveness of semi-physical simulation systems. SUMMARY

[0004] The present application aims at the technical problems existing in the prior art, and provides a control communication interface optimization method in semi-physical simulation, to solve the problems in the background art.

[0005] The technical solution of the present application to solve the above technical problems is as follows: specifically comprising the following steps:

[0006] Step S1, when the semi-physical simulation platform executes the load mutation working condition, deploying an atomic clock unit at the pulse width modulation instruction output end of the active disturbance rejection controller and the photoelectric isolation input end of the signal conversion module respectively, generating a control instruction generation time tag and a signal receiving time tag, and calculating the difference between the two as asynchronous error data;

[0007] Step S2, inputting the asynchronous error data into the grey wolf optimization algorithm, constructing an optimization model with the sum of asynchronous error squares and the sum of frequency deviation absolute values as the objective function, and outputting the frequency adjustment amount of the voltage-controlled oscillator circuit;

[0008] Step S3, inputting the historical electromagnetic interference intensity data and the historical asynchronous error data into the long short-term memory neural network, calculating the predicted delay amount through forward propagation, and generating the predicted frequency compensation amount based on the Sigmoid function;

[0009] Step S4, fusing the frequency adjustment amount of step S2 and the predicted frequency compensation amount of step S3, calculating the final frequency compensation value through a dynamic weight formula, and inputting the result into the synchronization circuit of the signal conversion module to realize instruction-signal timing alignment, and feeding back the fused timing error to the extended state observer of the active disturbance rejection controller;

[0010] In one preferred embodiment, in step S1, the specific process of generating the control instruction generation time tag and the signal receiving time tag is as follows:

[0011] Deploying an atomic clock unit at the pulse width modulation instruction output end of the active disturbance rejection controller and the photoelectric isolation input end of the signal conversion module, the synchronization accuracy of the clock unit is plus or minus ten nanoseconds; the control instruction generation time tag records the instruction sending time corresponding to the refresh period of the active disturbance rejection controller output, and the signal receiving time tag records the actual receiving time of the instruction at the signal conversion module drive board end; the two time tags are transmitted to the central processing unit through shielded twisted pair, and the sampling time interval is kept synchronous with the zero point one second step length of the semi-physical simulation platform.

[0012] In one preferred embodiment, the specific process of calculating the difference between the two as asynchronous error data comprises:

[0013] Firstly, the original asynchronous error value is calculated, which is the original time difference obtained by subtracting the control command generation time from the signal receiving time; then the original time difference is processed by a dynamic weight decay function, which uses a sliding window of five groups of data, assigns a weight coefficient that decays over time to each data point in the window, assigns a higher weight to recent data and a lower weight to long-term data, and introduces a decay factor of 0.8 to control the influence of historical data; finally, the compensated asynchronous error data is output, which will mark the corresponding working condition identifier and form a data group with the original timestamp, and is stored in the shared memory ring buffer, the buffer depth is fifty groups of data, and each group of data contains four fields of compensated asynchronous error value, working condition identifier, control command generation time and signal receiving time.

[0014] In a preferred embodiment, the specific operation of constructing an optimization model with the sum of asynchronous error squares and the sum of frequency deviation absolute values as the objective function in step S2 is as follows:

[0015] Firstly, the objective function is defined to consist of two main parts, the first part is the sum of asynchronous error squares, which squares and sums the compensated asynchronous error data calculated in step S1, to quantify the timing deviation between the control command and the execution signal; the second part is the sum of frequency deviation absolute values, which calculates the absolute difference between the current output frequency of the voltage-controlled oscillator circuit and the ideal frequency center value, to describe the degree of deviation of the hardware output frequency from the ideal state; the objective function combines the two parts to form the overall optimization target through a dynamic regularization coefficient, which is dynamically adjusted by the grey wolf optimization algorithm during optimization to achieve the balance of different optimization targets.

[0016] In a preferred embodiment, the specific optimization process of the grey wolf optimization algorithm is as follows: the frequency adjustment amount and the dynamic regularization coefficient together form a two-dimensional position vector as the optimization variable; an improved grey wolf search strategy is used, wherein The position update of the wolf, beta wolf and delta wolf introduces an exponential decay weight mechanism based on the difference in objective function values; after each position update, it checks whether the frequency adjustment amount exceeds the allowed output range of the voltage-controlled oscillator circuit, and applies a maximum penalty value to the out-of-range solution to eliminate it; finally, the frequency adjustment amount that minimizes the objective function value is selected as the optimization result output, and the corresponding optimal dynamic regularization coefficient value at this time is recorded.

[0017] In a preferred embodiment, the specific process of inputting historical electromagnetic interference intensity data and historical asynchronous error data into the long short-term memory neural network in step S3 is as follows: First, a fused input vector containing frequency domain features and time domain features is constructed. The frequency domain features are obtained by applying a fast Fourier transform to the historical electromagnetic interference intensity data to extract the amplitude values ​​of the three frequency points with the highest energy in the 0 to 100 kHz frequency band. The time domain features are obtained by calculating the first-order difference of the historical asynchronous error data to describe the delay change trend. At the same time, the optimal regularization coefficient obtained from the gray wolf optimization algorithm is added to the input vector as prior knowledge to describe the dynamics of the environment. All input data come from historical data records stored in the shared memory circular buffer.

[0018] In a preferred embodiment, the specific process of calculating the predicted delay through forward propagation and generating the frequency compensation amount based on the Sigmoid function is as follows: a dual-channel long short-term memory neural network architecture is used to process the frequency domain features and the time domain features respectively; the outputs of the two channels are fused through an adaptive attention weight mechanism to generate the final hidden state; the final hidden state is input into a fully connected layer to calculate the predicted delay; and finally, the predicted delay is converted into the frequency compensation amount through a combination transformation function containing the hyperbolic tangent function and the Sigmoid function.

[0019] In a preferred embodiment, the specific process of calculating the final frequency compensation value using the dynamic weighting formula in step S4 is as follows:

[0020] First, based on the difference between the absolute value of the predicted delay obtained in step S3 and the preset delay perception threshold, an initial dynamic weight value is calculated using an exponential growth function. Then, a cognitive decay function is introduced to perform a nonlinear transformation on the initial dynamic weight value to obtain the final dynamic weight value. The final dynamic weight value is used to adjust the contribution of the frequency adjustment in step S2. At the same time, 1 is subtracted from the final dynamic weight value to obtain another weight value, which is used to adjust the contribution of the predicted frequency compensation in step S3.

[0021] In a preferred embodiment, the specific process by which the synchronization circuit of the result input signal conversion module achieves command-signal timing alignment is as follows:

[0022] The reference frequency value is superimposed with the frequency adjustment amount after dynamic weighting and the predicted frequency compensation amount to generate the final frequency compensation value. This final frequency compensation value is converted into a corresponding square wave pulse signal by the voltage-controlled oscillator circuit of the input signal conversion module. It is then processed by the Schmitt trigger and delay circuit in the synchronization circuit, ensuring that the rise time of the output square wave pulse deviates from the time of the active disturbance rejection controller command generation time by no more than ten microseconds.

[0023] In a preferred embodiment, the specific implementation process of feeding the fusion timing error to the extended state observer of the active disturbance rejection controller is as follows:

[0024] The difference between the final frequency compensation value and the reference frequency value is mapped into a fusion timing error amount in the time domain through an inverse hyperbolic tangent transformation; the fusion timing error amount is multiplied by a learning rate coefficient and superimposed with the current total disturbance estimation value of the extended state observer, while introducing the time gradient of the frequency compensation value as an adjustment factor, to jointly update the output of the extended state observer, forming a closed-loop anti-disturbance control loop.

[0025] The method effectively improves the timing accuracy and anti-interference of the semi-physical simulation control through multi-technology fusion, realizes high-precision error collection using an atomic clock marker, dynamically optimizes the control instruction using an intelligent algorithm, compensates for the current transmission delay, prospectively suppresses future disturbances through a neural network prediction, and finally outputs precise frequency compensation through an adaptive fusion mechanism, thereby achieving signal alignment and closed-loop feedback, and significantly enhancing the stability and response speed of the system under complex working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 The method flowchart of the application is shown in the figure. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the application.

[0028] In the description of the application, the terms "first", "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0029] In the description of the present application, the term "for example" is used to mean "serving as an example, instance, or illustration." Any embodiment described as "for example" in this application is not necessarily to be construed as preferred or advantageous over other embodiments. The following description is presented to enable any person skilled in the art to make and use the application. In the following description, for the purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present application. It will be apparent, however, to one skilled in the art that the present application can be practiced without using these specific details. In other instances, well-known structures and processes are not elaborated upon in order to avoid unnecessary detail, which can obscure the description of the present application. Thus, the present application is not intended to be limited by the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

[0030] Embodiment 1

[0031] The embodiment provides a control communication interface optimization method in a semi-physical simulation as shown in the technical scheme, and specifically comprises the following steps: Figure 1

[0032] Step S1, when the semi-physical simulation platform executes the load mutation working condition, deploying an atomic clock unit at the pulse width modulation instruction output end of the active disturbance rejection controller and the photoelectric isolation input end of the signal conversion module respectively, generating a control instruction generation time tag and a signal receiving time tag, and calculating the difference value as asynchronous error data;

[0033] Step S2, inputting the asynchronous error data into the grey wolf optimization algorithm, constructing an optimization model with the sum of the asynchronous error squares and the absolute value of the frequency deviation as the objective function, and outputting the frequency adjustment amount of the voltage-controlled oscillator circuit;

[0034] Step S3, inputting the historical electromagnetic interference intensity data and the historical asynchronous error data into the long short-term memory neural network, calculating the predicted delay amount through forward propagation, and generating the predicted frequency compensation amount based on the Sigmoid function;

[0035] Step S4, fusing the frequency adjustment amount of step S2 and the predicted frequency compensation amount of step S3, calculating the final frequency compensation value through a dynamic weight formula, inputting the result into the synchronization circuit of the signal conversion module to realize instruction-signal timing alignment, and feeding back the fused timing error to the extended state observer of the active disturbance rejection controller.

[0036] In the embodiment, specifically, in step S1, the specific process of generating the control instruction generation time tag and the signal receiving time tag is as follows:

[0037] ​Atomic clock units are deployed at the pulse width modulation command output terminal of the active disturbance rejection controller (ADRC) and the opto-isolated input terminal of the signal conversion module, respectively. The synchronization accuracy of these clock units is ±10 nanoseconds; the control command generation time... The tag records the command issuance time and signal reception time corresponding to the refresh cycle of the active disturbance rejection controller output. The tag records the actual reception time of the instruction at the signal conversion module driver board; the tags at two moments are transmitted to the central processing unit via shielded twisted-pair cable, and the sampling time interval is strictly synchronized with the zero-one-second step of the semi-physical simulation platform. );

[0038] The specific process of calculating the difference between the two as asynchronous error data includes:

[0039] First, calculate the original asynchronous error value. This value is the original time difference obtained by subtracting the control command generation time from the signal reception time. The formula for calculating the original asynchronous error value is:

[0040] ;

[0041] in, Indicates the index of the current point (target point). This represents the original asynchronous error value, in microseconds, with a range of values. , Indicates the signal reception time (the first...) Each instruction is received at a timestamp in the signal conversion module (unit: seconds), with a value range of... , Indicates the instruction generation time (the first...) (Timestamp of each control command's generation), in seconds, with a value range of... Subsequently, a dynamic weight decay function was used to process the original time difference. This function employs a sliding window with five data sets, assigning a weight coefficient that decays over time to each data point within the window. More recent data is assigned a higher weight, while older data is assigned a lower weight. A decay factor of 0.8 is introduced to control the influence of historical data. The formula for the dynamic weight decay function is:

[0042] ;

[0043] in, This represents the width of the sliding window, with a fixed value. , Indicates the index of the point within the sliding window (used for summation). This represents the time decay weight, used to assign different weights to different sampling points within the window. The calculation method is as follows: ,in, Indicates the first in the window Data points, weights decrease as the data age increases (i.e. more recent data has higher weight), for example: (the most recent point): (the earliest point): In fact, the denominator is a constant (31), so the weight distribution emphasizes recent data, denotes the time decay weight a virtual summation index (for normalization) in the calculation, which is irrelevant to , denotes the decay factor, fixed at 0.8, which controls the degree of exponential decay function , denotes the absolute distance between the current point index and the window point index , The greater the value, the faster the historical data decays, for example: , if This indicates that data far from the current point has a small weight, denotes the compensated asynchronous error data, in microseconds, which is smoother and more reliable than the original asynchronous error value , and is used to eliminate high-frequency interference;

[0044] The function of this formula is to suppress high-frequency jitter caused by electromagnetic interference through exponential decay weighting, while retaining the trend delay caused by load mutation; the final output is the compensated asynchronous error data will mark the corresponding working condition identifier (such as 60%→80% load mutation is identified as OP03), and form a data group with the original timestamp and store it in the shared memory ring buffer, with a buffer depth of fifty groups of data, each group of data containing four fields: compensated asynchronous error value, working condition identifier, control instruction generation time and signal reception time .

[0045] In this embodiment, the specific operation of constructing an optimization model with the sum of asynchronous error squares and the sum of frequency deviation absolute values as the objective function in step S2 is as follows:

[0046] ​​​​Firstly, the objective function is defined by two main parts. The first part is the sum of squares of asynchronous error, which is used to quantify the timing deviation between the control command and the execution signal. The second part is the sum of absolute values of frequency deviation, which is used to describe the degree of deviation of the hardware output frequency from the ideal state. The two parts are combined by a dynamic regularization coefficient to form the overall optimization target. The regularization coefficient is dynamically adjusted by the grey wolf optimization algorithm during the optimization process to achieve the balance of different optimization targets. The expression of the objective function is:

[0047] ;

[0048] wherein, represents the objective function, wherein represents the parameter vector to be optimized, represents the number of valid data points in the sliding window (the total number of valid data points in the current target sliding window for calculation), which is 5, consistent with the sliding window width of step S1 (i.e. ), represents the error sensitivity enhancement function (a mathematical function used to amplify the influence of larger error values), which is calculated as follows: wherein represents the sensitivity coefficient, which is used to control the sensitivity of the function to error, and its value is set to 0.05, represents the current output frequency of the voltage-controlled oscillator (i.e. the actual output frequency value of the hardware circuit (voltage-controlled oscillator)), which ranges from 250 to 2500 Hz, represents the ideal frequency center value (the ideal target frequency that the voltage-controlled oscillator is expected to reach and maintain), which is set to a fixed value of 1375 Hz, represents the frequency deviation tolerance threshold, which is set to a fixed value of 100 Hz, when and the absolute value of the deviation is within this threshold (100 Hz), no penalty will be triggered, represents the dynamic regularization coefficient, which ranges from 0.1 to 5.0, and its value is adaptively adjusted by the external grey wolf optimization algorithm, represents the maximum output frequency limit allowed by the voltage-controlled oscillator (VCO) hardware, which defines an absolute safety boundary for calculating the penalty degree;

[0049] The specific optimization process of the grey wolf optimization algorithm is as follows: the frequency adjustment amount and the dynamic regularization coefficient are combined to form a two-dimensional position vector as the optimization variable. The expression of the two-dimensional position vector is: ​

[0050] ;

[0051] wherein, the parameter vector to be optimized (representing the position of a gray wolf (a candidate solution) in the search space) is a two-dimensional vector, represents the frequency adjustment amount (i.e., the first optimization variable) optimized by the gray wolf algorithm, which will be superimposed on the current frequency to form a new control instruction for adjusting the frequency of the voltage-controlled oscillator circuit, represents the dynamic regularization coefficient (i.e., the second optimization variable) for dynamically balancing the weights of the "timing error" ( ) and "frequency stability" ( ) in the objective function, The larger the value is, the more conservative the algorithm is, and the priority is to ensure the frequency stability, at the expense of some synchronization accuracy. The smaller the value is, the more aggressive the algorithm is, and the priority is to pursue synchronization accuracy, and it is more willing to risk the frequency deviation from the center value; an improved gray wolf search strategy is adopted, wherein The position updates of the wolf, beta wolf, and delta wolf introduce an exponential decay weight mechanism based on the difference in objective function values, expressed as:

[0052] ;

[0053] wherein, represents the two-dimensional position vector of the current gray wolf at the th iteration, represents the updated two-dimensional position vector of the gray wolf at the th iteration (i.e., the next iteration), represents the learning rate (value 0.8) for controlling the overall size of each update step, represents a random number uniformly distributed in the range [0, 1] for introducing randomness and avoiding the algorithm from falling into local optimum too early, represents the two-dimensional position vector of the th wolf (the alpha wolf) in the current wolf pack, which is the best solution found so far, represents the convergence factor (value 1.2) for controlling the strength of the exponential decay weight and affecting the ability of local exploration of the algorithm in the later iterations, represents the objective function value corresponding to the current two-dimensional position vector , represents the objective function value corresponding to the two-dimensional position vector of the current th wolf (i.e., the best objective function value found so far), represents the exponential decay weight, and the current solution is the optimal solution​ The greater the difference in the objective function value, the smaller the weight value, and thus the smaller the update step size. This is an adaptive fine-grained search mechanism. After each position update, it is checked whether the frequency adjustment exceeds the allowable output range of the voltage-controlled oscillator circuit (i.e., ...). ,in, This represents the new frequency after applying the adjustment given by the Grey Wolf algorithm. This indicates the allowable output frequency range (in Hz) of the voltage-controlled oscillator circuit; this is a hardware constraint. (This indicates that the new frequency is not within the allowed range), and a maximum penalty value is applied to the out-of-bounds solution to eliminate it (i.e., ... ,in, This represents imposing an infinite penalty on the objective function to ensure that any solution that violates the frequency constraint will have a very poor objective function value, and will therefore be automatically eliminated during algorithm iterations. Finally, the frequency adjustment that minimizes the objective function value is selected as the optimization result output, expressed as:

[0054] At the same time, record the corresponding optimal dynamic regularization system. The values ​​are used in the subsequent step S4.

[0055] In this embodiment, the specific process of inputting historical electromagnetic interference intensity data and historical asynchronous error data into the long short-term memory neural network in step S3 is as follows: First, a fused input vector containing frequency domain features and time domain features is constructed. The frequency domain features are obtained by applying a fast Fourier transform to the historical electromagnetic interference intensity data to extract the amplitude values ​​of the three highest energy points in the 0 to 100 kHz frequency band. The time domain features are obtained by calculating the first-order difference of the historical asynchronous error data to describe the delay change trend. At the same time, the optimal regularization coefficient obtained from the gray wolf optimization algorithm is added to the input vector as prior knowledge describing the dynamics of the environment. The expression is:

[0056] ;

[0057] in, This indicates the sample index (i.e., which data sample is currently being processed).

[0058] Indicates the first The fused input vector of each sample, This represents the Fast Fourier Transform (FFT). Indicates the first The historical electromagnetic interference intensity data collected synchronously by the group (is a signal sequence that varies over time). This indicates the historical electromagnetic interference intensity data. The frequency domain features obtained after fast Fourier transform, specifically, the amplitude of the three frequency points with the highest energy in the 0-100 kHz frequency band, denotes a differential operator, denotes the historical asynchronous error data of the sample, which is the past time deviation record read from the shared memory ring buffer, denotes the first-order difference of the historical asynchronous error data , which characterizes the trend of delay change (for example, whether the error is accelerating or decelerating), denotes the optimal regularization coefficient, which is a scalar value passed from the previous optimization algorithm (such as the grey wolf algorithm in step S2), containing prior knowledge of environmental dynamics, as a feature input network to assist decision-making; all input data come from historical data records stored in the shared memory ring buffer;

[0059] The specific process of calculating the predicted delay and generating the frequency compensation amount based on the Sigmoid function is as follows: a two-channel long short-term memory neural network architecture is used to process the frequency domain features and the time domain features respectively. The expression of the frequency domain feature processing channel is:

[0060] ;

[0061] wherein, denotes the hidden state, denotes the current time step, i.e. the data at the current time in sequence processing, denotes the previous time step, i.e. the data at the previous time in sequence processing, used to calculate the current state (reflecting the "memory" function), denotes the historical electromagnetic interference intensity data at the current time step , the output hidden state of the frequency domain feature processing channel LSTM (dimension 64), denotes the LSTM unit dedicated to processing frequency domain features, and the subscript represents the frequency domain, denotes fast Fourier transform (FFT), denotes the historical electromagnetic interference intensity data at the current time step , and the input data comes from the shared memory ring buffer;

[0062] The expression of the time domain feature processing channel is:

[0063] ;

[0064] wherein, denotes the historical electromagnetic interference intensity data at the current time step , the output hidden state of the time-domain feature processing channel LSTM (dimension 64), represents an LSTM unit dedicated to processing time-domain features, subscript represents the time domain, represents the differential operator, represents the history asynchronous error data at time step , the input data comes from the shared memory ring buffer; the outputs of the two channels are fused to generate the final hidden state through an adaptive attention weight mechanism, and the expression of the adaptive attention weight mechanism is

[0065] ;

[0066] wherein, represents the adaptive attention weight, which is a value automatically learned by the network, ranging from 0 to 1 (guaranteed by the Sigmoid function), used to measure the importance of frequency domain and time domain features at the current time, represents the Sigmoid activation function, used to compress the input to the (0, 1) interval, thereby generating an effective weight value, represents the weight matrix in the adaptive attention weight mechanism (trainable parameter), represents the bias vector in the adaptive attention weight mechanism (trainable parameter);

[0067] The expression of the final hidden state is:

[0068] ;

[0069] wherein, represents the final hidden state that integrates frequency domain and time domain information at time step , which is the direct input for the subsequent full connection layer to calculate the "predicted delay amount"; the final hidden state is input into the full connection layer to calculate the predicted delay amount, and the calculation formula of the predicted delay amount is:

[0070] ;

[0071] wherein, represents the predicted delay amount (the future possible timing error (delay) predicted by the double-channel long short-term memory neural network), with the unit of microsecond (μs) or nanosecond (ns), represents the final hidden state that integrates frequency domain and time domain information at time step , represents the weight matrix (learnable parameter of the full connection layer), denotes the bias vector (learnable parameters of the fully connected layer); finally, the predicted delay is converted into a frequency compensation through a combined conversion function consisting of a hyperbolic tangent function and a sigmoid function, and the expression of the frequency compensation is:

[0072] ;

[0073] wherein, denotes the predicted frequency compensation, which is a frequency adjustment value calculated from the predicted delay, and needs to be applied to the voltage-controlled oscillator circuit, with the unit of hertz (Hz), denotes the maximum frequency compensation amplitude (value 200), which is a scaling factor that directly limits the output range of the function within [-200, 200] Hz, consistent with the hardware limit, denotes the compensation scaling factor (value 0.01), which is a tuning factor that, after being multiplied by the predicted delay, serves as the input of the tanh function, used to control the sensitivity of the tanh function to the delay, denotes the hyperbolic tangent function, which is an S-shaped function with an output range of (-1, 1), and its role here is to map the input to a smooth, normalized range, denotes the sigmoid function, whose output range is (0, 1), and its role here is similar to a "gate" or "weight" that determines the strength of the output signal according to the size of the delay, denotes the time constant of the sigmoid function (value 100 μs), which is a tuning factor that controls the steepness and center point of the sigmoid function, thereby determining the delay at which the output reaches 0.5 (half-open state), and setting determines the direction (positive or negative) and approximate proportion of the compensation, with the output ranging from -1 to 1, and setting determines the confidence or intensity of the compensation, with the output ranging from 0 to 1, and when the predicted delay is very small, this value is close to 0, making the final compensation also close to 0; when the predicted delay is very large, this value is close to 1, allowing the output of the tanh function to be fully output.

[0074] In this embodiment, it is specifically necessary to explain that in step S4, the specific process of calculating the final frequency compensation value through the dynamic weight formula is as follows:

[0075] First, based on the difference between the absolute value of the predicted delay obtained in step S3 and the preset delay perception threshold, an initial dynamic weight value is calculated through an exponential growth function, and the formula of the initial dynamic weight value is:

[0076] ;

[0077] wherein, represents an initial dynamic weight value, which is a value between 0 and 1, and reflects the initial dynamic weight value obtained only according to the delay amount, represents a predicted delay amount, represents a weight sensitivity coefficient (value 0.02, unit ), which controls the steepness of the growth of the Sigmoid function, The greater the value, the faster and more sensitive the transition of the exponential growth function from 0 to 1 near the delay perception threshold , represents a delay perception threshold (value 300 ), which is a key threshold value, when the absolute value of the predicted delay amount is greater than this threshold, the initial dynamic weight value will tend to 1; when it is much smaller than this threshold, will tend to 0; then a cognitive decay function is introduced to nonlinearly transform the initial dynamic weight value to obtain the final dynamic weight value, which can prevent the initial dynamic weight value from being oversaturated, and the formula of the final dynamic weight value is:

[0078] ;

[0079] wherein, represents a final dynamic weight value, which is the final result after processing by the cognitive decay function, and is also a value between 0 and 1, which will be used to adjust the contribution degree of different compensation amounts, represents a cognitive decay function, which is a self-defined function, and its specific form is Its role is to prevent the initial dynamic weight value from being too high (oversaturated), and to smooth the growth of the weight by introducing a decay term , represents a weight decay coefficient (its value is set to 2.5), which controls the decay strength of the cognitive decay function , The greater the value, the stronger the decay effect, and the final dynamic weight value will be smaller than the initial dynamic weight value ; the final dynamic weight value is used to adjust the contribution degree of the frequency adjustment amount in step S2, and another weight value obtained by subtracting the final dynamic weight value from 1 is used to adjust the contribution degree of the predicted frequency compensation amount in step S3;

[0080] The specific process of realizing instruction-signal timing alignment by inputting the result into the synchronous circuit of the signal conversion module is as follows:

[0081] The reference frequency value is superimposed with the frequency adjustment amount after dynamic weight adjustment and the predicted frequency compensation amount to generate a final frequency compensation value, and the formula is:

[0082] ;

[0083] Wherein, represents the final frequency compensation value, which is a final frequency control instruction (unit: Hz) containing the reference value and all compensation amounts to be sent to the voltage-controlled oscillator, represents the reference frequency (value is 1375Hz), which is the center working frequency of the voltage-controlled oscillator and the reference point of all compensation and adjustment, and is consistent with in the previous step S2, represents the final dynamic weight value, which is used to adjust the contribution size of the compensation amount from the grey wolf optimization algorithm (step S2), and the value is generated by the "cognitive weight dynamic generation mechanism" discussed earlier, represents the frequency adjustment amount after grey wolf optimization, which is the frequency adjustment value (unit: Hz) calculated by the grey wolf optimization algorithm (step S2) to compensate for the current timing error, represents the sign function, which is used to extract the sign (positive or negative) of its internal variable, and its role is to ensure the consistency of the adjustment direction, for example, if the predicted delay is positive (indicating that the clock is slow), then is also positive, so that the adjustment direction of is positive (speed up the clock), represents another weight value, which is used to adjust the contribution size of the predicted frequency compensation amount from the neural network prediction (step S3), represents the predicted frequency compensation amount, represents the prediction compensation confidence function, which is a self-defined function that calculates the confidence of this prediction result according to the size of the predicted delay amount , represents the predicted delay amount; the final frequency compensation value is input into the voltage-controlled oscillator of the signal conversion module to convert into a corresponding square wave pulse signal; through the Schmidt trigger and delay circuit in the synchronization circuit, the time deviation between the rising edge time of the output square wave pulse and the time when the active disturbance rejection controller instruction is generated is not more than ten microseconds, so as to realize accurate timing alignment;

[0084] The specific implementation process of feeding the fused timing error back to the extended state observer of the active disturbance rejection controller is as follows:

[0085] The final frequency compensation value is transformed by inverse hyperbolic tangent transformation with the reference frequency value The difference of the two is mapped to the fusion timing error in time domain, which is given by:

[0086] ;

[0087] wherein, denotes the fusion timing error (unit: μs or ns), denotes the final frequency compensation value, denotes the reference frequency, denotes the inverse hyperbolic tangent function, which is the inverse function of hyperbolic tangent function, and is used here to perform inverse transformation from frequency domain to time domain, to restore the difference of frequency back to the error in time, denotes the frequency-time transformation coefficient (value is usually 200), denotes the normalized scaling factor, which is set to 0.01; after multiplying the fusion timing error by the learning rate coefficient, it is superimposed with the current total disturbance estimation value of the extended state observer, and the time gradient of the frequency compensation value is introduced as an adjustment factor, and the disturbance estimation update formula is:

[0088] ;

[0089] wherein, denotes the updated total disturbance estimation value at the next iteration (i.e. the next step), denotes the total disturbance estimation value of the extended state observer at the current iteration, which represents the total of all internal unmodeled dynamics and external disturbances of the system, and is the core of the active disturbance rejection controller, denotes the observer learning rate (value is 0.8), which is a coefficient between 0 and 1, controlling the speed and stability of the observer updating according to the fusion timing error, the greater the value, the faster the update, but it may cause oscillation; the smaller the value, the smoother the update, but the response is slower, denotes the time gradient of the frequency compensation value, which represents the speed or trend of the change of the control command, this is a very important adjustment factor: if the command is changing rapidly, it means that the system is in a dynamic adjustment process, and the feedback may contain more transient information, and its weight should be adjusted by this gradient to update the output of the extended state observer together, forming a closed-loop anti-disturbance control loop.

[0090] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0091] Those skilled in the art will appreciate that embodiments of the present application can be devised for a variety of applications. It is intended that the present application be limited only by the scope of the appended claims, and it is intended that various modifications and alterations made by those skilled in the art be considered as within the scope of the present application. The embodiments of the present application will be described with reference to the attached drawings identified below.

[0092] The present application is described in reference to the drawings using a flowchart and / or a block diagram of the method, apparatus (system) and computer program product according to embodiments of the application. It will be understood that each block of the flowchart and / or block diagram, and combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0093] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0094] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0095] While the preferred embodiments of the application have been described, additional variations and modifications can be employed by those skilled in the art. Therefore, the appended claims are intended to cover all such variations and modifications as falling within the scope of the present application.

[0096] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A method for control communication interface optimization in a semi-physical simulation, characterized in that, Specifically comprising the following steps: Step S1, when the semi-physical simulation platform executes the load mutation working condition, an atomic clock unit is arranged at the pulse width modulation instruction output end of the active disturbance rejection controller and the photoelectric isolation input end of the signal conversion module, control instruction generation time tags and signal receiving time tags are generated, and the difference between the two is calculated as asynchronous error data; Step S2, the asynchronous error data is input into the grey wolf optimization algorithm, an optimization model with the sum of the asynchronous error square sum and the frequency deviation absolute value as the objective function is constructed, and the frequency adjustment amount of the voltage-controlled oscillator circuit is output; Step S3, historical electromagnetic interference intensity data and historical asynchronous error data are input into the long short-term memory neural network, a prediction delay amount is calculated through forward propagation, and a prediction frequency compensation amount is generated based on the Sigmoid function; Step S4, the frequency adjustment amount of step S2 and the prediction frequency compensation amount of step S3 are fused, a final frequency compensation value is calculated through a dynamic weight formula, and the result is input into the synchronization circuit of the signal conversion module to realize instruction-signal timing alignment, and the fused timing error is fed back to the extended state observer of the active disturbance rejection controller.

2. The method of claim 1, wherein: In the step S1, the specific process of generating the control instruction generation time tag and the signal receiving time tag is as follows: An atomic clock unit is arranged at the pulse width modulation instruction output end of the active disturbance rejection controller and the photoelectric isolation input end of the signal conversion module, the synchronization accuracy of the clock unit is plus or minus ten nanoseconds, the control instruction generation time tag records the instruction sending time corresponding to the refresh period of the active disturbance rejection controller, and the signal receiving time tag records the actual receiving time of the instruction at the signal conversion module drive board end; The two time tags are transmitted to the central processing unit through shielded twisted pair lines, and the sampling time interval is kept synchronous with the zero point one second step length of the semi-physical simulation platform.

3. The method of claim 2, wherein: The specific process of calculating the difference between the two as asynchronous error data includes: First, calculate the original asynchronous error value, which is the original time difference obtained by subtracting the control instruction generation time from the signal receiving time; then, process the original time difference by using a dynamic weight decay function, the dynamic weight decay function uses a sliding window of five groups of data, assigns a weight coefficient that decays over time to each data point in the window, and introduces a decay factor of zero point eight to control the influence degree of historical data; the formula of the dynamic weight decay function is: ； wherein, N represents the width of the sliding window, fixed value N = 5, k represents the point index in the sliding window, represents the time decay weight, used to give different weights to different sampling points in the window, and the calculation method is: wherein, the weight decreases with the increase of k, and j represents the time decay weight the virtual summation index in the calculation is irrelevant to i, represents the decay factor, fixed value 0.8, controls the degree of exponential decay function , represents the absolute distance between the current point index i and the window point index k, represents the compensated asynchronous error data, represents the original asynchronous error value collected by the kth sampling point; The final output of the compensated asynchronous error data will mark the corresponding working condition identifier, and form a data group with the original timestamp and store it in the shared memory ring buffer, the buffer depth is fifty groups of data, and each group of data contains four fields of compensated asynchronous error value, working condition identifier, control instruction generation time and signal receiving time.

4. The method of claim 3, wherein: In the step S2, the specific operation of constructing the optimization model with the sum of the asynchronous error square sum and the frequency deviation absolute value as the objective function is as follows: Firstly, the objective function is defined by two parts, the first part is the sum of squares of asynchronous error, which is used to quantify the time sequence deviation between the control command and the execution signal, and the second part is the sum of absolute values of frequency deviation, which is used to describe the degree of deviation of the hardware output frequency from the ideal state. The objective function is combined by the two parts through a dynamic regularization coefficient, which is dynamically adjusted by the grey wolf optimization algorithm during the optimization process to balance different optimization objectives.

5. The method of claim 4, wherein: The specific optimization process of the grey wolf optimization algorithm is as follows: a two-dimensional position vector composed of the frequency adjustment amount and the dynamic regularization coefficient is taken as an optimization variable; an improved grey wolf search strategy is adopted, wherein The position updating of the wolf, the beta wolf and the delta wolf introduces an exponential decay weight mechanism based on the difference of the objective function values; after each position updating, it is checked whether the frequency adjustment amount exceeds the allowed output range of the voltage-controlled oscillator circuit, and a maximum penalty value is applied to the out-of-range solution so that it is eliminated; finally, the frequency adjustment amount that makes the objective function value minimum is selected as the optimization result output, and the optimal dynamic regularization coefficient value at this time is recorded.

6. The method of claim 5, wherein: In step S3, the specific process of inputting historical electromagnetic interference intensity data and historical asynchronous error data into the long short-term memory neural network is as follows: firstly, a fusion input vector containing frequency domain features and time domain features is constructed, the frequency domain features are obtained by applying fast Fourier transform to the historical electromagnetic interference intensity data to extract the amplitude values of the three highest energy points in the frequency band of zero to one hundred kilohertz, the time domain features are obtained by calculating the first-order difference of the historical asynchronous error data to describe the delay change trend, and the optimal regularization coefficient obtained from the grey wolf optimization algorithm is added to the input vector as prior knowledge describing environmental dynamics; all input data come from the historical data records stored in the shared memory ring buffer.

7. The method of claim 6, wherein: The specific process of calculating the predicted delay amount by forward propagation and generating the predicted frequency compensation amount based on the Sigmoid function is as follows: a dual-channel long short-term memory neural network architecture is used to process the frequency domain features and the time domain features respectively; the outputs of the two channels are fused to generate the final hidden state through an adaptive attention weight mechanism; the final hidden state is input into a fully connected layer to calculate the predicted delay amount; finally, the predicted delay amount is converted into a frequency compensation amount through a combined conversion function containing a hyperbolic tangent function and a Sigmoid function.

8. The method of claim 7, wherein: In step S4, the specific process of calculating the final frequency compensation value through a dynamic weight formula is as follows: Firstly, based on the difference between the absolute value of the predicted delay amount obtained in step S3 and the preset delay perception threshold, an initial dynamic weight value is calculated through an exponential growth function; then, a cognitive decay function is introduced to nonlinearly transform the initial dynamic weight value to obtain a final dynamic weight value; the final dynamic weight value is used to adjust the contribution degree of the frequency adjustment amount in step S2, and another weight value obtained by subtracting the final dynamic weight value from 1 is used to adjust the contribution degree of the predicted frequency compensation amount in step S3.

9. The method of claim 8, wherein: The specific process of realizing instruction-signal time sequence alignment by the synchronous circuit of the result input signal conversion module is as follows: The reference frequency value is superimposed with the frequency adjustment amount and the predicted frequency compensation amount adjusted by the dynamic weight to generate a final frequency compensation value; the final frequency compensation value is input into the voltage-controlled oscillator circuit of the signal conversion module to convert it into a corresponding square wave pulse signal; the square wave pulse is processed through the Schmidt trigger and delay circuit in the synchronous circuit, and the time deviation between the rising edge of the output square wave pulse and the time when the active disturbance rejection controller instruction is generated is not more than ten microseconds.

10. The method of claim 9, wherein: The specific implementation process of feeding the fused timing error to the extended state observer of the active disturbance rejection controller is as follows: The difference between the final frequency compensation value and the reference frequency value is mapped into a fused timing error in the time domain through an inverse hyperbolic tangent transformation; the fused timing error is multiplied by a learning rate coefficient and superimposed with the current total disturbance estimation value of the extended state observer, and the time gradient of the frequency compensation value is introduced as an adjustment factor to jointly update the output of the extended state observer, thereby forming a closed-loop anti-disturbance control loop.

Images