Diesel engine heterogeneous model simulation method based on industrial mechanism model

Abstract

The application discloses a diesel engine heterogeneous model simulation method based on an industrial mechanism model, relates to the technical field of industrial software integration and heterogeneous computing, and comprises the following steps: S1, continuously outputting standardized heterogeneous model input data flow; S2, outputting an asynchronous time sequence promotion trigger signal; S3, outputting high-frequency transient state response and high-frequency transient control parameters; S4, outputting slow dynamic response characteristic data; S5, introducing a virtual resonant cavity frequency selection mechanism based on multi-path coherent interference by improving an IndRNN model, driving a residual sequence to perform non-crossing parallel recursion and dimension remodeling; S6, delimiting a dynamic convergence boundary and outputting a current iteration step residual sequence; and S7, outputting a joint performance index. The application overcomes the limitations of a traditional method in terms of dependence on static threshold scheduling and falling into redundant iteration, and provides an efficient solution for high-fidelity digital twinning of military diesel engines.

Description

Technical Field

[0001] This invention relates to the field of industrial software integration and heterogeneous computing technology, and in particular to a diesel engine heterogeneous model simulation method based on an industrial mechanism model. Background Technology

[0002] With the normalization of extreme tactical conditions in high-mobility military vehicles, real-time high-fidelity simulation of multi-source heterogeneous data for military diesel engines faces dual challenges in terms of computing power and accuracy. While existing fast-slow dual-timescale co-simulation decouples the fast dynamics of electronic control from the slow dynamics of thermodynamics, it relies on static step size thresholds and rigid convergence boundaries for scheduling. This ignores the implicit multidimensional variable time-series evolution characteristics and the deep nonlinear coupling relationship between fast and slow dynamics under extreme conditions, leading to timescale misalignment and high-frequency state truncation distortion during alternating solutions, severely limiting the accuracy of closed-loop iterative correction. Furthermore, classical convergence criteria struggle to extract local range and variance decay characteristics at the tail of the residual sequence to dynamically constrain boundaries. This easily results in non-convergence bias or redundant iterations when handling transient simulations of tactical maneuvers, significantly increasing onboard computing overhead and causing joint performance indicators to fail, making it difficult to meet the practical requirements of military power digital twins.

[0003] Therefore, how to provide a simulation method for heterogeneous diesel engine models based on industrial mechanism models is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a simulation method for heterogeneous diesel engine models based on an industrial mechanism model. A closed-loop feedback architecture is constructed by introducing an improved IndRNN model with a virtual resonant cavity frequency selection mechanism based on multi-path coherent interference. This architecture adaptively extrapolates the execution step size across both fast and slow time scales, driving the solution of the industrial mechanism model. The time-scale aligned residual sequence is injected into the improved IndRNN model. A virtual resonant cavity is constructed based on the multi-path coherent interference virtual resonant cavity frequency selection mechanism to perform multi-path shuttle propagation and phase shift modulation. The alignment correction matrix is ​​then updated via parallel recursion without crossover. Based on the alignment correction matrix, a fast and slow dual-process closed-loop iteration is driven. The dynamic convergence boundary is defined in real-time according to the residual tail characteristics until convergence is triggered, and the joint performance index is aggregated and output. This architecture effectively eliminates time-scale alignment misalignment and high-frequency state truncation distortion under extreme tactical conditions, overcoming the limitations of traditional methods that rely on static threshold scheduling and get trapped in redundant iterations. It provides an efficient solution for high-fidelity digital twins of military diesel engines.

[0005] The diesel engine heterogeneous model simulation method based on an industrial mechanism model according to an embodiment of the present invention specifically includes: S1. Input the multi-source heterogeneous state flow and control flow of the diesel engine into a unified variable dictionary to perform dynamic format conversion and unit mapping, and continuously output the standardized heterogeneous model input data flow. S2. Inject the standardized heterogeneous model input data stream into the underlying scheduling process, dynamically define the confidence judgment boundary along the temporal evolution characteristics, and output asynchronous temporal advancement trigger signals to identify and exceed the limit of the execution state fluctuation. S3. Inject the asynchronous timing advance trigger signal into the fast dynamic electronic control process, adaptively deduce the execution step size based on the historical step size and drive the forward solution, and output the high-frequency transient state response and high-frequency transient control parameters. S4. Inject the high-frequency transient control parameters and high-frequency transient state response into the industrial mechanism model to execute the combustion and heat transfer solution process. Adaptively deduce the slow dynamic solution step size based on the parameter fluctuation characteristics. Drive the in-cylinder thermodynamic process based on the step size to execute the variable step size coupled solution and output the slow dynamic response characteristic data. S5. Based on their respective physical solution time scales, the slow dynamic response characteristic data are aligned with the expected response execution time of the industrial mechanism theory corresponding to the high-frequency control parameters and the residual sequence is dynamically solved. By improving the IndRNN model, a virtual resonant cavity frequency selection mechanism based on multi-path coherent interference is introduced to drive the residual sequence through non-crossing parallel recursive update in the time dimension and matching reshaping in the control dimension, and the alignment correction matrix is ​​mapped out. S6. Based on the alignment correction matrix, the fast and slow dual-process closed-loop iteration is driven. The dynamic convergence boundary is defined in real time based on the extracted tail features of the iterative residuals, and the residual sequence of the current iteration step is output. S7. Based on the comparison result between the residual sequence of the current iteration step and the dynamic convergence boundary of the current iteration step, convergence is triggered, the simulation of the current working condition is terminated, and the joint performance index is output.

[0006] Optionally, S1 specifically includes: S11. Based on multi-source heterogeneous state flow and control flow, extract the namespace features and dimensional statistical distribution features of variable identifiers, and adaptively construct a unified variable dictionary; S12. Inject the multi-source heterogeneous state flow and control flow into a unified variable dictionary, match namespace features to perform dynamic format conversion, and output homogenized intermediate state flow and homogenized intermediate control flow. S13. Based on the dimensional statistical distribution characteristics, perform adaptive unit mapping on the homogenized intermediate state flow and the homogenized intermediate control flow, and continuously output the standardized heterogeneous model input data flow.

[0007] Optionally, S2 specifically includes: S21. Based on the input data stream of the standardized heterogeneous model, extract the time-series sliding window variance and state transition probability distribution of multidimensional variables, and adaptively define the confidence judgment boundary based on the variance extreme points and probability mutation points. S22. Inject the standardized heterogeneous model input data stream into the confidence decision boundary, calculate the boundary exceedance deviation rate of the multidimensional variables, and output the deviation rate time series. S23. Extract local gradient features and duration features based on the deviation rate time series, perform state fluctuation limit recognition based on the local gradient features and duration features, and output asynchronous time series advancement trigger signal.

[0008] Optionally, S3 specifically includes: S31. Based on the asynchronous timing advance trigger signal, extract the local variance decay characteristics and step size migration probability distribution of the high-frequency step size evolution sequence, and adaptively deduce the current execution step size based on the local variance decay characteristics and step size migration probability distribution. S32. The asynchronous timing advance trigger signal and the current execution step size are directly injected into the fast dynamic electronic control process. Based on the current execution step size, the control parameter discrete jump process is driven to perform forward deduction and output a high-frequency transient control jump sequence. S33. Inject the high-frequency transient control jump sequence into the hydraulic and pneumatic transient transmission process to perform forward dynamic evolution, and output the high-frequency transient state response and high-frequency transient control parameters based on the evolution trajectory mapping of the hydraulic and pneumatic transient transmission process.

[0009] Optionally, S4 specifically includes: S41. Based on the high-frequency transient control parameters and the high-frequency transient state response, extract the amplitude fluctuation attenuation characteristics of the high-frequency transient control parameters and the statistical distribution of the rate of change of the high-frequency transient state response, and adaptively deduce the slow dynamic solution step size based on the amplitude fluctuation attenuation characteristics and the statistical distribution of the rate of change. S42. Inject high-frequency transient control parameters and high-frequency transient state response into the industrial mechanism model to drive the fuel injection and intake swirl mixing process to perform forward flow field evolution and output in-cylinder transient flow field spatial distribution data. S43. Inject the transient flow field spatial distribution data in the cylinder into the in-cylinder thermodynamic process. Based on the slow dynamic solution step size, drive the in-cylinder thermodynamic process to perform a variable step size combustion heat release and cylinder wall heat transfer coupled solution, and output slow dynamic response characteristic data.

[0010] Optionally, the improved IndRNN model includes a time-scale aligned residual solving layer, an optical resonant cavity filtering layer, an independent recursive feature extraction layer, and a correction matrix mapping layer: The time-scale aligned residual solution layer is used to substitute the slow dynamic response feature data and the steady-state components based on high-frequency control parameters into the industrial mechanism theory expected response obtained by the mapping relationship between in-cylinder thermodynamics and fluid dynamics, and perform timestamp interpolation alignment based on their respective physical solution timescales; and perform element-by-element difference calculation between the time-scale aligned slow dynamic response feature data and the industrial mechanism theory expected response to dynamically solve the residual sequence. The optical resonant cavity filter layer is used to introduce a virtual resonant cavity frequency selection mechanism based on multipath coherent interference. The specific execution steps include: The residual sequence is mapped to an initial virtual optical wave signal tensor. A virtual resonant cavity structure containing the weight matrix of the front reflector and the weight matrix of the rear reflector is constructed. The initial virtual optical wave signal tensor is injected into the virtual resonant cavity structure to perform multi-path forward and backward shuttle propagation calculations and output the cavity shuttle feature tensor. The frequency domain transformation and energy distribution statistics are performed on the intracavity shuttle feature tensor to extract the concentration feature of the frequency domain energy distribution. An initial cavity length benchmark value is set and adaptive gradient update optimization is performed on the initial cavity length benchmark value along the iteration step size based on the concentration feature to dynamically generate the target resonant cavity length parameter. The target resonant cavity length parameter is used as a spatial propagation variable. Combined with the multi-path trajectory dimension of the cavity shuttle feature tensor, the phase offset matrix is ​​generated by solving element by element. The phase offset matrix is ​​multiplied element-wise with the cavity shuttle feature tensor. Based on the polarity distribution, the multiplication result is modulated by constructive superposition and destructive attenuation, and the output is a resonant feature tensor carrying specific physical deviation period characteristics. The independent recursive feature extraction layer is used to expand the resonance feature tensor along the time dimension into a residual time step input matrix; extract the current time step data in the residual time step input matrix sequentially along the time sequence, introduce recursive weights by multiplying element by element, perform element-wise accumulation and activation of the multiplication result with the hidden state of the previous time step, and complete the parallel refresh of the hidden state without cross-neuron interference in each time step to extract residual evolution features; The modified matrix mapping layer is used to input the residual evolution features into the fully connected mapping network to perform linear transformation and activation processing, and reshape the transformed and activated high-dimensional feature tensor into a two-dimensional matrix structure that matches the dimension of the high-frequency control parameters, and output the aligned modified matrix in real time.

[0011] Optionally, S6 specifically includes: S61. The alignment correction matrix is ​​superimposed element by element onto the high-frequency transient control parameters of the fast dynamic electronic control process to generate high-frequency correction control parameters, drive the fast dynamic electronic control process to perform forward solving, and output the high-frequency correction state response of the current iteration step. S62. Inject the high-frequency correction state response and high-frequency correction control parameters of the current iteration step into the industrial mechanism model, drive the in-cylinder thermodynamic process to perform variable step-size coupled solution, and output the slow dynamic correction response characteristic data of the current iteration step. S63. Perform element-by-element difference calculation on the slow dynamic correction response characteristic data of the current iteration step and the expected response of the industrial mechanism theory corresponding to the high-frequency correction control parameters to generate the residual sequence of the current iteration step. S64. Based on the residual sequence of the current iteration step, calculate the local variance decay rate and local range convergence rate along the time dimension. Adaptively determine the truncation position according to the inflection point of the variance decay rate and truncate the tail residual segment. Construct the iterative residual tail feature based on the global variance and global range of the residual segment. S65. Based on the relative proportion of global variance and global range in the tail features of the iterative residual, normalize and map the corresponding dynamic weights, perform weighted aggregation of global variance and global range according to the dynamic weights, and adaptively generate the dynamic convergence boundary of the current iteration step. S66. Based on the residual sequence of the current iteration step and the dynamic convergence boundary of the current iteration step, perform a spatial constraint comparison. If the constraint conditions are met, output a convergence trigger signal and simultaneously output the residual sequence of the current iteration step. Otherwise, inject the residual sequence of the current iteration step as a feedback source into the improved IndRNN model to trigger the calculation of the next iteration step.

[0012] Optionally, S7 specifically includes: S71. Based on the convergence trigger signal, terminate the simulation of the current working condition and extract the high-frequency and slow-dynamic correction responses at the end in real time; use the slow-dynamic time node as the anchor point to downsample the high-frequency response, extract the high-frequency state downsampled sequence with the timestamps completely aligned and splice it with the slow-dynamic feature data row by row to construct a dual-time-scale joint response matrix. S72. Extract the upper and lower limits of the expected response of the industrial mechanism theory corresponding to the high-frequency state downsampling sequence and slow dynamic feature data of the dual-time-scale joint response matrix, drive the dual-time-scale joint response matrix to dynamically complete the dimensionless normalization mapping according to the mechanism constraint interval, and generate a standardized joint response matrix. S73. Dynamically extract the fast and slow dynamic mean features along the rows of the standardized joint response matrix and weight them according to the dynamic fusion benchmark to aggregate and output the joint performance index.

[0013] The beneficial effects of this invention are: (1) This invention achieves adaptive extrapolation of execution step sizes on both fast and slow time scales by constructing a confidence determination boundary and an asynchronous scheduling mechanism based on temporal evolution characteristics. The sliding window variance and state transition probability distribution of multidimensional variables are extracted to dynamically define the confidence determination boundary and output asynchronous temporal advancement trigger signals. Combining local variance decay characteristics and amplitude fluctuation decay characteristics, the fast dynamic electronic control execution step size and the slow dynamic thermodynamic solution step size are extrapolated respectively. This mechanism ensures that the variable step size forward coupling solution strictly conforms to the physical evolution law of extreme transient operating conditions of military diesel engines, outputting high-frequency transient state response and slow dynamic response characteristic data.

[0014] (2) This invention introduces a virtual resonant cavity frequency selection mechanism based on multipath coherent interference by adopting an improved IndRNN model, and establishes an alignment correction matrix mapping system based on spatiotemporal reshaping of residual sequences. The time-scaled aligned residual sequence is mapped into a virtual optical wave signal tensor, and a virtual resonant cavity structure is constructed to perform multipath shuttle propagation calculations. The target resonant cavity length parameter is generated adaptively based on the frequency domain energy concentration characteristics, the phase offset matrix is ​​solved to perform constructive superposition and destructive attenuation modulation, and the resonant feature tensor is output. The resonant feature tensor is driven to perform non-crossing parallel recursive update and control dimension matching reshaping, and the alignment correction matrix is ​​mapped to the output. This system uses multipath coherent interference to extract deep physical deviation features, and combined with independent recursive parallel refresh capability, eliminates timescale mismatch and high-frequency truncation distortion caused by fast and slow dynamic alternating solutions.

[0015] (3) This invention achieves precise termination of simulation closed-loop iteration and quantification of joint performance indicators by constructing a dynamic convergence boundary delineation mechanism based on residual tail feature extraction. The alignment correction matrix is ​​superimposed onto the high-frequency control parameters to drive the dual-process closed-loop iteration. The local variance decay rate and local range convergence rate are extracted to truncate the tail residual fragments. Based on the relative proportions of global variance and global range, a weighted aggregation is performed to adaptively generate the dynamic convergence boundary. In response to the convergence trigger signal, a dual-timescale joint response matrix is ​​constructed, and dimensionless normalization mapping and weighted flow are performed according to the mechanism-constrained interval. This mechanism directly transforms the statistical features of the residual tail into spatial constraint boundaries, truncates invalid and redundant iterations, and ensures that the joint performance indicators accurately reflect the real evolution level of military diesel engines under complex tactical scenarios. 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 an overall flowchart of the diesel engine heterogeneous model simulation method based on the industrial mechanism model proposed in this invention; Figure 2 This is a flowchart illustrating the working principle of the improved IndRNN model, which is based on an industrial mechanism model for diesel engine heterogeneous model simulation. Detailed Implementation

[0017] The 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 Figure 1 and Figure 2 A simulation method for heterogeneous diesel engine models based on industrial mechanism models, specifically including: S1. Input the multi-source heterogeneous state flow and control flow of the diesel engine into a unified variable dictionary to perform dynamic format conversion and unit mapping, and continuously output the standardized heterogeneous model input data flow. S2. Inject the standardized heterogeneous model input data stream into the underlying scheduling process, dynamically define the confidence judgment boundary along the temporal evolution characteristics, and output asynchronous temporal advancement trigger signals to identify and exceed the limit of the execution state fluctuation. S3. Inject the asynchronous timing advance trigger signal into the fast dynamic electronic control process, adaptively deduce the execution step size based on the historical step size and drive the forward solution, and output the high-frequency transient state response and high-frequency transient control parameters. S4. Inject the high-frequency transient control parameters and high-frequency transient state response into the industrial mechanism model to execute the combustion and heat transfer solution process. Adaptively deduce the slow dynamic solution step size based on the parameter fluctuation characteristics. Drive the in-cylinder thermodynamic process based on the step size to execute the variable step size coupled solution and output the slow dynamic response characteristic data. S5. Based on their respective physical solution time scales, the slow dynamic response characteristic data are aligned with the expected response execution time of the industrial mechanism theory corresponding to the high-frequency control parameters and the residual sequence is dynamically solved. By improving the IndRNN model, a virtual resonant cavity frequency selection mechanism based on multi-path coherent interference is introduced to drive the residual sequence through non-crossing parallel recursive update in the time dimension and matching reshaping in the control dimension, and the alignment correction matrix is ​​mapped out. S6. Based on the alignment correction matrix, the fast and slow dual-process closed-loop iteration is driven. The dynamic convergence boundary is defined in real time based on the extracted tail features of the iterative residuals, and the residual sequence of the current iteration step is output. S7. Based on the comparison result between the residual sequence of the current iteration step and the dynamic convergence boundary of the current iteration step, convergence is triggered, the simulation of the current working condition is terminated, and the joint performance index is output.

[0019] In this embodiment, S1 specifically includes: S11. Scan the multi-source heterogeneous state flow and control flow to extract the namespace and dimensional features of variable identifiers, construct a unified variable dictionary with namespace as key and dimensional features and target data type labels as value, set the initial capacity limit to 5000 and trigger the doubling expansion mechanism as needed to complete the adaptive construction of conversion rules. S12. Inject the multi-source heterogeneous state flow and control flow into the unified variable dictionary one by one, match the namespace key, read the target data type label and dimensional characteristics, use the underlying forced type conversion to convert the variable values ​​into 64-bit double-precision floating-point numbers, bind the converted values ​​with the dimensional characteristics and rearrange them according to the timestamp, and generate the homogenized intermediate state flow and homogenized intermediate control flow. S13. Based on the dimensional features bound to the isomorphic intermediate state flow and the isomorphic intermediate control flow, look up the SI reference mapping table, set the mapping coefficient to 1.0 for standard units, extract the actual dimensional conversion ratio as the mapping coefficient for non-standard units, multiply the physical quantity value by the corresponding mapping coefficient to perform adaptive unit mapping, splice and align according to timestamp, and continuously output a standardized heterogeneous model input data stream with unified format and standardized dimensions.

[0020] In this embodiment, S2 specifically includes: S21. Read the input data stream of the standardized heterogeneous model, set a sliding window with a step size of 10 time steps to calculate the variance and construct a time-series sliding window variance sequence, and calculate the ratio of the absolute value of the difference between adjacent step size variables and the sum of their absolute values ​​to construct a state transition probability distribution sequence; calculate the absolute value of the first difference of three adjacent variance values ​​in the time-series sliding window variance sequence as the extreme value change rate, set the extreme value judgment threshold of 0.05 and the probability mutation threshold of 0.15, and extract the corresponding variable values ​​as boundary nodes when the extreme value change rate is greater than 0.05 and the probability mutation index is greater than 0.15, and connect them in time series to generate confidence judgment boundaries; S22. Inject the standardized heterogeneous model input data stream point by point into the confidence decision boundary, calculate the absolute value of the difference between the variable value at each time point and the corresponding boundary node value, divide it by the boundary node value to obtain the boundary deviation rate, and arrange them in time sequence to generate a deviation rate time series. S23. Read the deviation rate time series, calculate the difference between adjacent deviation rates and divide by the step size 10 to obtain the local gradient feature series. At the same time, set the threshold of 0.2 to extract the number of consecutive time points with values ​​greater than 0.2 as the duration feature. Input the local gradient feature into the Sigmoid activation function, set the output truncation threshold of 0.8 and the duration threshold of 5 time steps. When the Sigmoid output value is greater than 0.8 and the duration is greater than 5 time steps, output a high-level signal 1; otherwise, output a low-level signal 0. Generate an asynchronous timing advance trigger signal according to the timing combination.

[0021] In this embodiment, S3 specifically includes: S31. Read the time point corresponding to the high-level signal 1 in the asynchronous timing advance trigger signal, and extract the data points within 5 time steps before and after this time point as the high-frequency step size evolution sequence; calculate the absolute value of the difference between two adjacent data point values ​​in the high-frequency step size evolution sequence, divide the absolute value of the difference by the sum of the first data point value and the constant 0.001 to obtain the local variance decay feature; calculate the proportion of the number of times the current data point jumps to the next data point in the high-frequency step size evolution sequence, and construct the step size migration probability distribution; multiply the local variance decay feature by the step size migration probability distribution, and add the basic step size constant 1.0 to the product result to calculate the current execution step size.

[0022] S32. Inject the asynchronous timing advance trigger signal and the current execution step size directly into the fast dynamic electronic control process. Extract the discrete time node where the high-level signal 1 in the asynchronous timing advance trigger signal is located as the current node. Multiply the current execution step size by the scaling factor 0.5 to obtain the dynamic jump factor. Calculate the control parameter value of the current node and multiply it by the dynamic jump factor to obtain the jump increment. Add the jump increment to the control parameter value of the current node to obtain the updated control parameter value of the next discrete time node. Add the timestamp value of the current node to the current execution step size to obtain the timestamp of the next discrete time node. Follow this rule to deduce and arrange all the updated control parameter values ​​in chronological order to generate a high-frequency transient control jump sequence.

[0023] S33. Inject the high-frequency transient control jump sequence into the hydraulic and pneumatic transient transmission process. Set the hydraulic transmission damping coefficient to 0.85 and the pneumatic transmission delay step to 3 time steps. Multiply each control parameter value in the high-frequency transient control jump sequence by the hydraulic transmission damping coefficient 0.85, and accumulate the pneumatic transmission values ​​at 3 time steps apart. Perform forward dynamic evolution calculation to obtain the evolution trajectory of the hydraulic and pneumatic transient transmission process. Extract the state values ​​in the evolution trajectory that are completely aligned with the timestamp of the high-frequency transient control jump sequence as the high-frequency transient state response. At the same time, extract the original control parameter values ​​that are completely aligned with the timestamp of the high-frequency transient state response as the high-frequency transient control parameters. Jointly output the high-frequency transient state response and the high-frequency transient control parameters.

[0024] In this embodiment, S4 specifically includes: S41. Read the high-frequency transient control parameters and the high-frequency transient state response. Calculate the absolute value of the difference between two adjacent time points in the high-frequency transient control parameters. Divide the absolute value of the difference by the sum of the previous time point value and the constant 0.01 to obtain the amplitude fluctuation attenuation characteristic. Calculate the difference between two adjacent time points in the high-frequency transient state response and divide it by the time interval to obtain the rate of change. Arrange the rates of change in descending order and extract the top 30% of the rate of change values ​​to calculate the arithmetic mean. Use the arithmetic mean as the statistical distribution of the rate of change. Multiply the amplitude fluctuation attenuation characteristic by the statistical distribution of the rate of change and add the product to the basic slow step size constant 2.0 to calculate the slow dynamic solution step size. S42. Inject high-frequency transient control parameters and high-frequency transient state response into the industrial mechanism model, extract the fuel injection pressure value from the high-frequency transient control parameters and the intake air velocity value from the high-frequency transient state response, multiply the fuel injection pressure value by the injector orifice cross-sectional area of ​​0.5 square millimeters to obtain the fuel volumetric flow rate, multiply the intake air velocity value by the intake duct cross-sectional area of ​​10.0 square millimeters to obtain the air volumetric flow rate, and divide the fuel volumetric flow rate by the air volumetric flow rate to obtain the fuel-air mixture ratio; using the fuel-air mixture ratio as the initial condition, discretize the flow field space into a cubic grid with a side length of 1 millimeter, calculate the fluid velocity difference and pressure difference at adjacent time steps at each grid node, and update the flow velocity and pressure of each grid node by subtracting the pressure difference from the fluid velocity difference and multiplying it by a constant of 0.8. Iterate and update the values ​​of all grid nodes step by step to perform forward flow field evolution, and generate in-cylinder transient flow field spatial distribution data containing three-dimensional spatial coordinates and corresponding flow velocity and pressure values; S43. Read the transient flow field spatial distribution data in the cylinder, extract the velocity and pressure values ​​and substitute them into the in-cylinder thermodynamic process. Multiply the velocity value by the fluid density constant 1.2 to obtain the turbulent kinetic energy. Calculate the absolute value of the difference in cylinder volume between adjacent time steps as the volume change. Multiply the pressure value by the volume change to obtain the work indication. Use the slow dynamic solution step size as the time span of each iteration. Input the turbulent kinetic energy and work indication into the Tanh activation function to calculate the combustion heat release rate. At the same time, subtract the cylinder wall temperature constant 400 Kelvin from the temperature value in the transient flow field spatial distribution data to obtain the heat transfer temperature difference. Multiply the heat transfer temperature difference by the heat transfer coefficient constant 0.45 to obtain the cylinder wall heat transfer loss. Subtract the cylinder wall heat transfer loss from the combustion heat release rate and sum them up according to the slow dynamic solution step size to output the slow dynamic response characteristic data.

[0025] In this embodiment, the improved IndRNN model includes a time-scale aligned residual solving layer, an optical resonant cavity filtering layer, an independent recursive feature extraction layer, and a correction matrix mapping layer: The time-scale aligned residual solution layer is used to read slow dynamic response characteristic data and high-frequency transient control parameters. It calculates the arithmetic mean of all time-point values ​​in the high-frequency transient control parameters to obtain the steady-state component. The pressure value in the steady-state component is multiplied by the cylinder volume constant 0.005 cubic meters, and the temperature value in the steady-state component is multiplied by the gas specific heat constant 1.1. The sum of the in-cylinder gas work indication and the internal thermal energy increment is then calculated as the expected response according to the industrial mechanism theory. The timestamps of the slow dynamic response characteristic data are extracted, and the relationship between the timestamps and the working parameters is calculated. The difference between the expected response timestamps of the industrial mechanism theory and the interpolation reference step size of 0.5 milliseconds is used to obtain the interpolation index. The integer part of the interpolation index is used as the low-order index and the fractional part is used as the interpolation weight. The result of multiplying the value at the low-order index position in the slow dynamic response feature data by 1 and subtracting the interpolation weight is added to the result of multiplying the value at the low-order index position by 1 and the interpolation weight to generate the time-scale aligned slow dynamic response feature data. The element-wise difference between the time-scale aligned slow dynamic response feature data and the expected response of the industrial mechanism theory is calculated and arranged in time order to generate the residual sequence.

[0026] The optical resonant cavity filter layer is used to introduce a virtual resonant cavity frequency selection mechanism based on multipath coherent interference. The specific execution steps include: Read the residual sequence and reshape it into an initial virtual optical wave signal tensor with a total number of data points in rows of 1. Construct a front reflection surface weight matrix with a total number of data points in rows of 1 and a back reflection surface weight matrix with a total number of data points in rows of 1 and a back reflection surface weight matrix. Set the main diagonal elements of the front reflection surface weight matrix to 0.9 and the remaining elements to 0.1. Set the main diagonal elements of the back reflection surface weight matrix to 0.8 and the remaining elements to 0.2. Multiply the initial virtual optical wave signal tensor by the front reflection surface weight matrix to obtain the forward reflection tensor. Multiply the forward reflection tensor by the back reflection surface weight matrix to obtain the back reflection tensor. Add the forward reflection tensor and the back reflection tensor element by element and iterate for 3 times to generate the intracavity shuttle feature tensor.

[0027] Read the intracavity shuttle feature tensor, perform a fast Fourier transform on each row of data in the intracavity shuttle feature tensor to obtain a frequency domain complex matrix, calculate the square of the modulus of each element in the frequency domain complex matrix to obtain the frequency domain energy matrix; divide the number of elements in the frequency domain energy matrix that are greater than 0.7 times the total energy by the total number of elements to obtain the concentration feature of the frequency domain energy distribution; set the initial cavity length baseline value to 5.0, calculate the difference between the concentration feature and the target concentration threshold of 0.85 and multiply it by the learning rate of 0.01 to obtain the cavity length update gradient, subtract the cavity length update gradient from the initial cavity length baseline value to obtain the updated cavity length value, and iterate and update 10 times to output the target resonant cavity length parameter.

[0028] Read the target resonant cavity length parameter and the cavity shuttle feature tensor. Multiply the target resonant cavity length parameter by a constant 3.14 and divide by the number of rows in the cavity shuttle feature tensor to obtain the spatial fundamental phase. Construct a multipath trajectory vector with the same number of rows as the cavity shuttle feature tensor and 1 column. Set the element values ​​of each row in the multipath trajectory vector to be a sequence of natural numbers starting from 1 and increasing by 1. Multiply the spatial fundamental phase by the corresponding row element in the multipath trajectory vector and take the cosine of the result to generate a phase offset matrix with the same number of rows and columns as the cavity shuttle feature tensor. Perform element-wise multiplication of the phase offset matrix and the cavity shuttle feature tensor to obtain the interference modulation matrix. Extract the elements greater than 0 in the interference modulation matrix and keep their original values, and set the elements less than or equal to 0 to 0 to obtain the constructive modulation matrix. Extract the absolute values ​​of the elements less than 0 in the interference modulation matrix and set the elements greater than or equal to 0 to 0 to obtain the destructive modulation matrix. Subtract the destructive modulation matrix from the constructive modulation matrix to generate a resonant feature tensor with the same dimensions as the cavity shuttle feature tensor.

[0029] An independent recursive feature extraction layer is used to read the resonance feature tensor. The resonance feature tensor is expanded into a one-dimensional vector by concatenating the first and last rows. The length of the one-dimensional vector is calculated, divided by the time dimension constant 10, and rounded down to obtain the number of truncated rows in the time step. This number is then multiplied by 10 to obtain the target truncation length. The first target truncation length of data points are extracted from the one-dimensional vector and reshaped into a residual time step input matrix with 10 columns and truncation rows in the time step. An initial hidden state matrix with 10 columns and truncation rows in the time step and all elements being 0 is initialized. Data from each row of the residual time step input matrix is ​​extracted sequentially along the time series. The current row data is multiplied by the recursive weight constant 0.6 to obtain weighted input data. The weighted input data is added to the hidden state matrix of the previous time step to obtain the accumulated state matrix. The accumulated state matrix is ​​input into the Tanh activation function to calculate the output value and replace the hidden state matrix of the current time step. After parallel refreshing of all time steps, the hidden state matrix of the last time step is extracted as the residual evolution feature.

[0030] The modified matrix mapping layer is used to read the residual evolution features and high-frequency transient control parameters, and calculate the product of the number of rows and columns of the high-frequency transient control parameters to obtain the total number of target elements. The residual evolution features are input into a fully connected mapping network containing the total number of target elements and neurons. Each element in the residual evolution features is multiplied by the network weight constant 0.5 and a bias constant 0.1 is added to obtain the linear transformation result. The linear transformation result is input into the ReLU activation function to calculate the maximum feature. The number of rows and columns of the high-frequency transient control parameters is read, and the maximum feature is reshaped into a two-dimensional matrix structure with the exact same number of rows and columns as the high-frequency transient control parameters. The aligned modified matrix is ​​output.

[0031] The improved IndRNN model proposed in this step is similar to the traditional IndRNN model in that it is based on the state-space mapping theory of independent recurrent neurons. That is, it calculates the temporal dependence by projecting the multidimensional physical response sequence to the high-dimensional hidden space, uses the recursive weight mechanism of element-wise multiplication to capture the dynamic evolution of the state, and uses a fully connected mapping layer to linearly transform the final hidden state into an output matrix that matches the control dimension.

[0032] The difference lies in that this invention breaks through the limitation of traditional IndRNNs that directly input the original residual sequence into the hidden layer for linear recursive updates. Instead, it introduces a virtual resonant cavity frequency selection mechanism based on multipath coherent interference, maps the residual sequence into an initial virtual optical signal tensor, constructs a virtual resonant cavity structure containing weight matrices of the front and rear reflectors to perform multipath shuttle propagation, adaptively optimizes the target resonant cavity length parameter based on the frequency domain energy concentration characteristics, solves the phase offset matrix and performs constructive superposition and destructive attenuation modulation, and finally inputs the resonant feature tensor carrying specific physical deviation periodic characteristics into an independent recursive feature extraction layer to perform cross-free parallel recursive updates, rather than directly performing element-wise accumulation of a single weight on the original residual sequence.

[0033] The beneficial effects of the improvements are that this invention, through the cascading of coherent frequency selection in a pre-positioned virtual resonant cavity and parallel refresh in a post-positioned independent recursive system, forcibly embeds physical optical resonance constraints into the forward propagation of IndRNN. This breaks the limitation of traditional IndRNNs in extreme transient conditions where the recursive state diverges or becomes distorted due to the inability to distinguish between real physical deviations and high-frequency measurement noise. It achieves a precise conversion from linear recursive accumulation to strong constraints based on physical resonance frequency selection. This design significantly enhances the ability to target and extract deep periodic deviation features in the fast and slow dynamic coupling residuals. It can accurately filter out invalid interference in complex coherent interference spaces while retaining key physical fluctuation features. Combined with parallel recursion without cross-neuron interference, it effectively eliminates timescale alignment misalignment in fast and slow dynamic solutions, providing a high-fidelity alignment correction benchmark for closed-loop iteration of heterogeneous models of military diesel engines.

[0034] In this embodiment, S6 specifically includes: S61. Read the alignment correction matrix and the high-frequency transient control parameters of the fast dynamic control process. Calculate the sum of the values ​​of each element in the high-frequency transient control parameters and the corresponding element in the alignment correction matrix to generate the high-frequency correction control parameters. Extract the current state variable and inductance parameter constant 0.05 Henry from the fast dynamic control process at the previous time step. Calculate the high-frequency correction control parameters by subtracting the current state variable multiplied by the resistance constant 10 ohms and then dividing by the inductance parameter constant 0.05 Henry to obtain the current change rate. Multiply the current change rate by the solution step size constant 0.001 seconds and add the current state... The state variables are updated by time step to obtain the current at the current moment; the current at the current moment is calculated by multiplying the electromagnetic coupling constant of 5 Newtons per ampere and subtracting the spring reset force constant of 20 Newtons, and then dividing the result by the actuator motion mass constant of 0.1 kg to obtain the actuator acceleration. The actuator acceleration is multiplied by the square of the solution step constant of 0.001 seconds, divided by the constant 2, and then the product of the actuator velocity of the previous time step and the solution step constant of 0.001 seconds is added, along with the actuator displacement of the previous time step. The displacement-velocity-acceleration recursive evolution is performed, and the high-frequency correction state response of the current iteration step is output.

[0035] S62. Read the high-frequency correction state response and high-frequency correction control parameters of the current iteration step. Use the high-frequency correction state response of the current iteration step as the initial displacement boundary condition of the actuator and the high-frequency correction control parameters as the external excitation source to inject into the industrial mechanism model. Calculate the high-frequency correction state response of the current iteration step by multiplying it by the area constant 0.002 square meters to obtain the flow cross-sectional area of ​​the fuel valve. Combine the flow cross-sectional area of ​​the fuel valve to calculate the rate of change of the gas pressure in the cylinder. Multiply it by the time step constant 0.005 seconds and add the pressure at the previous moment to obtain the pressure at the current moment. Calculate the pressure at the current moment by multiplying it by the rate of change of the cylinder volume and adding the gas work heat dissipation rate to obtain the internal energy change. Perform variable step-size coupled solution to output the slow dynamic correction response characteristic data of the current iteration step.

[0036] S63. Read the slow dynamic correction response characteristic data and high-frequency correction control parameters of the current iteration step, calculate the arithmetic mean of the values ​​at all time points in the high-frequency correction control parameters to obtain the current steady-state component, multiply the pressure value in the current steady-state component by the cylinder volume constant 0.005 cubic meters and add the temperature value in the current steady-state component multiplied by the gas specific heat capacity constant 1.1, calculate the sum of the gas work indication in the cylinder and the internal heat energy increment as the expected response of the industrial mechanism theory corresponding to the high-frequency correction control parameters; calculate the element-by-element difference between the slow dynamic correction response characteristic data of the current iteration step and the expected response of the industrial mechanism theory, and arrange them in time sequence to generate the residual sequence of the current iteration step.

[0037] S64. Read the residual sequence of the current iteration step, set the local time window length to 5, slide the window along the time dimension and calculate the variance of the data points within the window as the local variance, calculate the difference between the maximum and minimum values ​​within the window as the local range; calculate the difference between the local variance of the current window and the local variance of the previous window, and then divide by the local variance of the previous window to obtain the local variance decay rate; calculate the difference between the local range of the current window and the local range of the previous window, and then divide by the local range of the previous window to obtain the local range convergence rate; extract the turning point where the local variance decay rate changes from negative to positive as the truncation position, extract the data segment from the truncation position to the end of the sequence in the current iteration step residual sequence as the tail residual segment; calculate the variance of all elements in the tail residual segment to obtain the global variance, calculate the difference between the maximum and minimum element values ​​in the tail residual segment to obtain the global range, and concatenate the global variance and the global range to form the iterative residual tail feature.

[0038] S65. Read the tail features of the iterative residuals, extract the global variance and global range, calculate the sum of the global variance and the constant 0.001 as the denominator, calculate the global variance divided by the denominator to obtain the relative proportion of variance, calculate the global range divided by the denominator to obtain the relative proportion of range, use the relative proportion of variance as the dynamic weight of variance, and the relative proportion of range as the dynamic weight of range; calculate the sum of global variance multiplied by the dynamic weight of variance and global range multiplied by the dynamic weight of range to generate the dynamic convergence boundary of the current iteration step.

[0039] S66. Read the residual sequence of the current iteration step and the dynamic convergence boundary of the current iteration step. Calculate the absolute value of all elements in the residual sequence of the current iteration step. Determine if there is a value in the absolute value that is greater than the dynamic convergence boundary of the current iteration step. If all absolute values ​​are less than or equal to the dynamic convergence boundary of the current iteration step, it is determined that the constraint condition is met. Output the convergence trigger signal and output the residual sequence of the current iteration step simultaneously. If any absolute value is greater than the dynamic convergence boundary of the current iteration step, it is determined that the constraint condition is not met. Inject the residual sequence of the current iteration step as a feedback source into the improved IndRNN model to trigger the calculation of the next iteration step.

[0040] The dynamic convergence determination process based on the tail characteristics of iterative residuals proposed in this step is similar to the traditional global fixed threshold termination mechanism of numerical simulation in that both are based on the error reduction and system stability determination theory. That is, by tracking the state residual sequence in the forward solution process, the convergence trend of the system towards equilibrium is evaluated by using the decreasing law of residual amplitude. Both use the comparison logic between residuals and preset boundaries to determine whether to terminate the current calculation process.

[0041] The difference lies in that this invention breaks through the limitation of traditional fixed threshold judgment ignoring the details of local temporal evolution of residuals. It adds a step of detecting the inflection point of variance decay rate to adaptively define the cutoff position, and replaces the traditional global scalar threshold with a weighted aggregation mechanism of global variance and global range. It maps the local variance decay rate and local range convergence rate to a two-dimensional dynamic weight to perform weighted space constraint construction. Finally, it performs a comparison between the residual sequence of the current iteration step and the dynamic convergence boundary to output a trigger signal, rather than a single cross-generational comparison between the absolute value of the residual and a fixed threshold.

[0042] The beneficial effects of the improvements are that, by using tail feature truncation and two-dimensional weighted aggregation, the local statistical evolution constraints of the residuals are rigidly embedded into the closed-loop iterative process. This breaks the limitations of traditional methods, which are prone to misjudgment of convergence or falling into invalid redundant iterations due to local high-frequency oscillations under extreme transient conditions. It achieves a precise conversion from global rigid truncation to local dynamic physical constraints. This design significantly enhances the defense capability against nonlinear fluctuations at the end of the fast-slow dynamic coupling solution. It can accurately capture the real convergence inflection point in the micro-scale interval of the residual tail. Combined with adaptive boundary delineation, it effectively eliminates invalid dead loops triggered by tactical maneuver commands, greatly improving the computational efficiency of complex condition simulation and the absolute reliability of termination determination.

[0043] In this embodiment, S7 specifically includes: S71. Read the convergence trigger signal, the terminal high-frequency correction state response, and the slow dynamic correction response feature data. Based on the convergence trigger signal, extract the last 50 data points of the terminal high-frequency correction state response and the last 10 data points of the slow dynamic correction response feature data. Extract the timestamps corresponding to the last 10 data points of the slow dynamic correction response feature data as the slow dynamic time anchor set. Calculate the difference between each timestamp in the slow dynamic time anchor set and the starting timestamp of the last 50 data points of the high-frequency correction state response, and divide it by the high-frequency sampling step size of 0.1 milliseconds to obtain the downsampling interpolation index. Use the integer part of the downsampling interpolation index as the low-order index and the fractional part as the downsampling weight. Calculate the result of multiplying the low-order index position value by 1 and subtracting the downsampling weight, plus the result of multiplying the low-order index position value by 1 and the downsampling weight. Sequentially extract and generate a high-frequency state downsampling sequence containing 10 values. Concatenate the high-frequency state downsampling sequence as the first row and the last 10 data points of the slow dynamic correction response feature data as the second row to construct a 2-row, 10-column dual-time-scale joint response matrix.

[0044] S72. Read the dual-timescale joint response matrix, extract the theoretical expected upper limit of high-frequency pressure (15 MPa) and lower limit of 5 MPa corresponding to the high-frequency downsampling sequence in the first row, and the theoretical expected upper limit of slow-dynamic temperature (600 Kelvin) and lower limit of 400 Kelvin corresponding to the slow-dynamic feature data in the second row; calculate the difference between subtracting the lower limit of 5 MPa from the value of each element in the first row of the dual-timescale joint response matrix and the difference between the upper limit of 15 MPa and the lower limit of 5 MPa; calculate the difference between subtracting the lower limit of 400 Kelvin from the value of each element in the second row of the dual-timescale joint response matrix and the difference between the upper limit of 600 Kelvin and the lower limit of 400 Kelvin; arrange all the calculated mapping values ​​according to the original matrix row and column positions, and generate a standardized joint response matrix.

[0045] S73. Read the standardized joint response matrix, calculate the arithmetic mean of all elements in the first row of the standardized joint response matrix to obtain the high-frequency mean feature, calculate the arithmetic mean of all elements in the second row of the standardized joint response matrix to obtain the slow dynamic mean feature; set the dynamic fusion benchmark to 0.6, calculate the result of multiplying the high-frequency mean feature by the dynamic fusion benchmark 0.6, adding the slow dynamic mean feature multiplied by the constant 1 and subtracting the dynamic fusion benchmark 0.6, and aggregate and output the joint performance index.

[0046] Example 1: To verify the feasibility of this invention in digital twins of military vehicle power systems, the method of this invention was applied to the extreme operating condition simulation platform of a high-mobility military diesel engine at a military research institute (hereinafter referred to as "Institute A"). In traditional joint simulation systems for military diesel engines, fast and slow dual-timescale collaborative solution methods based on fixed step size or conventional variable step size are commonly used. These methods not only struggle to accurately match the temporal evolution laws of the fast dynamics of electronic control and the slow dynamics of thermodynamics under extreme tactical maneuvering conditions, but also fail to accurately extract the local range and variance decay characteristics at the tail of the iterative residual sequence to dynamically constrain the convergence boundary, easily leading to high-frequency truncation distortion or getting trapped in invalid redundant iterations in the simulation state. To solve the above problems, Institute A decided to adopt the diesel engine heterogeneous model simulation method based on an industrial mechanism model proposed in this invention.

[0047] During implementation, Institute A first utilized a real vehicle-mounted sensor network and bench-based measurement and control system to acquire multi-source heterogeneous state and control flows of a certain type of high-power military diesel engine under complex tactical scenarios such as extreme off-road driving, high-altitude assault, and high-power sprints. After preprocessing operations such as dynamic format conversion and adaptive unit mapping based on a unified variable dictionary, a standardized heterogeneous model input data stream was continuously constructed and output. Simultaneously, Institute A's overall evaluation experts conducted precise physical dimension verification and mechanistic boundary calibration of the collected multi-source data, serving as the benchmark for model simulation and combat target evaluation.

[0048] Institute A extracts the temporal sliding window variance and state transition probability distribution of multidimensional variables through a low-level scheduling process, dynamically defines confidence judgment boundaries, and performs state fluctuation limit identification, outputting asynchronous temporal advancement trigger signals. Next, utilizing the local variance decay characteristics and amplitude fluctuation decay characteristics of the high-frequency step-size evolution sequence, it adaptively extrapolates the fast dynamic electronic control execution step size and the slow dynamic thermodynamic solution step size, effectively eliminating the rigid truncation interference caused by fixed step sizes. Subsequently, the trigger signal and execution step size are injected into the fast dynamic electronic control process and the industrial mechanism model, driving the hydraulic and pneumatic transient transmission processes and the in-cylinder fuel injection and intake swirl mixing processes to perform variable-step forward evolution solutions, outputting high-frequency transient state response and slow dynamic response characteristic data.

[0049] In the core correction and convergence phase, this invention introduces a virtual resonant cavity frequency selection mechanism based on multipath coherent interference by improving the IndRNN model. This maps the time-scaled aligned residual sequence into a virtual optical signal tensor, constructing a virtual resonant cavity structure to perform multipath shuttle propagation calculations. Based on the frequency domain energy concentration characteristics, the target resonant cavity length parameter is adaptively optimized, and the phase offset matrix is ​​solved to perform constructive superposition and destructive attenuation modulation, outputting a resonant feature tensor. The resonant feature tensor is driven to perform non-crossing parallel recursive updates and control dimension matching reshaping, mapping and outputting an alignment correction matrix. Subsequently, the dynamic convergence boundary delineation module delineates the dynamic convergence boundary in real time based on the residual tail characteristics until convergence is triggered, and the terminal response is extracted to construct a dual-timescale joint response matrix. Dimensionless normalization mapping is completed according to the mechanism-constrained interval, and the joint performance index is aggregated and output, achieving a closed-loop leap from state awareness to residual correction and then to precise convergence.

[0050] During implementation, the technical team at Institute A discovered that, compared to traditional fixed-threshold scheduling methods, the method of this invention significantly improves the fidelity and real-time performance of simulations of extreme transient conditions for military diesel engines. Traditional methods cannot quantify high-frequency state truncation distortion and have poor termination control for transient simulations triggered by tactical maneuver commands. In contrast, the method of this invention, through multi-path coherent interferometric frequency selection filtering, independent recursive parallel refreshing, and dynamic convergence boundary delineation, effectively achieves precise correction of fast and slow dynamic deep nonlinear coupling deviations and effective truncation of ineffective redundant iterations.

[0051] To further verify the actual performance of the method of the present invention, Institute A conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1: Table 1. Comparison of Simulation Performance of Military Diesel Engine A under Extreme Operating Conditions in Institute A

[0052] As shown in Table 1, the performance of the military diesel engine extreme condition simulation platform was comprehensively improved after applying the method of this invention. The timing alignment mismatch rate decreased from 15.6% in the traditional method to 1.2%, and the high-frequency state truncation distortion decreased from 22.4% to 2.1%, significantly improving the accuracy of solving the state evolution of the heterogeneous model and providing a reliable basis for subsequent closed-loop iterative correction. The number of invalid redundant iterations decreased from 35.0 times / condition to 3.5 times / condition, effectively avoiding serious waste of computing resources. The response latency of the combat target evaluation was significantly shortened from 85.0 seconds to 8.5 seconds, significantly enhancing the system's timeliness. In addition, the accuracy of extreme transient condition simulation increased from 76.0% to 97.5%, and the resource utilization rate of the vehicle-mounted computing platform decreased from 92.0% to 54.0%, significantly reducing the computing power consumption of the airborne computing platform. The synchronization lag of the digital twin was reduced from 450 milliseconds to 45 milliseconds, and the satisfaction rate of the actual vehicle tactical matching was also significantly improved, from the original 82.0% to 98.0%.

[0053] Through the method of this invention, Institute A successfully achieved high-fidelity digital simulation and rapid combat target evaluation of military diesel engines using multi-source heterogeneous data in complex tactical scenarios. This effectively eliminated simulation deviations caused by extreme transient loading and severe thermal shock, ensured the accurate operation of the digital twin base of the military power system, significantly improved the robustness and adaptability of the core twin system of military vehicle power, significantly reduced the cycle burden of military bench testing, enhanced the stability and real-time performance of the simulation platform, and provided strong technical support for the intelligent upgrading of highly mobile military equipment.

[0054] 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 simulation method for heterogeneous diesel engine models based on industrial mechanism models, characterized in that, Includes the following steps: S1. Input the multi-source heterogeneous state flow and control flow of the diesel engine into a unified variable dictionary to perform dynamic format conversion and unit mapping, and continuously output the standardized heterogeneous model input data flow. S2. Inject the standardized heterogeneous model input data stream into the underlying scheduling process, dynamically define the confidence judgment boundary along the temporal evolution characteristics, and output asynchronous temporal advancement trigger signals to identify and resolve execution state fluctuations that exceed limits. S3. Inject the asynchronous timing advance trigger signal into the fast dynamic electronic control process, adaptively deduce the execution step size based on the historical step size and drive the forward solution, and output the high-frequency transient state response and high-frequency transient control parameters. S4. Inject the high-frequency transient control parameters and high-frequency transient state response into the industrial mechanism model to execute the combustion and heat transfer solution process. Adaptively deduce the slow dynamic solution step size based on the parameter fluctuation characteristics. Based on the step size, drive the in-cylinder thermodynamic process to execute the variable step size coupled solution and output the slow dynamic response characteristic data. S5. Based on their respective physical solution time scales, the slow dynamic response characteristic data are aligned with the expected response execution time of the industrial mechanism theory corresponding to the high-frequency transient control parameters and the residual sequence is dynamically solved. By improving the IndRNN model, a virtual resonant cavity frequency selection mechanism based on multi-path coherent interference is introduced to drive the residual sequence through non-crossing parallel recursive update in the time dimension and matching reshaping in the control dimension, and the alignment correction matrix is ​​mapped out. S6. Based on the alignment correction matrix, the fast and slow dual-process closed-loop iteration is driven. The dynamic convergence boundary is defined in real time based on the extracted tail features of the iterative residuals, and the residual sequence of the current iteration step is output. S7. Based on the comparison result between the residual sequence of the current iteration step and the dynamic convergence boundary of the current iteration step, convergence is triggered, the simulation of the current working condition is terminated, and the joint performance index is output. The improved IndRNN model includes a time-scale aligned residual solving layer, an optical resonant cavity filtering layer, an independent recursive feature extraction layer, and a correction matrix mapping layer: The time-scale aligned residual solution layer is used to substitute the slow dynamic response feature data and the steady-state components based on the high-frequency transient control parameters into the industrial mechanism theory expected response obtained by the mapping relationship between in-cylinder thermodynamics and fluid dynamics, and perform timestamp interpolation alignment based on their respective physical solution timescales; then, it performs element-by-element difference calculation between the time-scale aligned slow response feature data and the industrial mechanism theory expected response to dynamically solve the residual sequence; The optical resonant cavity filter layer is used to introduce a virtual resonant cavity frequency selection mechanism based on multipath coherent interference. The specific execution steps include: The residual sequence is mapped to an initial virtual optical wave signal tensor. A virtual resonant cavity structure containing the weight matrix of the front reflector and the weight matrix of the rear reflector is constructed. The initial virtual optical wave signal tensor is injected into the virtual resonant cavity structure to perform multi-path forward and backward shuttle propagation calculations and output the cavity shuttle feature tensor. The frequency domain transformation and energy distribution statistics are performed on the intracavity shuttle feature tensor to extract the concentration feature of the frequency domain energy distribution. An initial cavity length benchmark value is set and adaptive gradient update optimization is performed on the initial cavity length benchmark value along the iteration step size based on the concentration feature to dynamically generate the target resonant cavity length parameter. The target resonant cavity length parameter is used as a spatial propagation variable. Combined with the multi-path trajectory dimension of the cavity shuttle feature tensor, the phase offset matrix is ​​generated by solving element by element. The phase offset matrix is ​​multiplied element-wise with the cavity shuttle feature tensor. Based on the polarity distribution, the multiplication result is modulated by constructive superposition and destructive attenuation, and the output is a resonant feature tensor carrying specific physical deviation period characteristics. The independent recursive feature extraction layer is used to expand the resonance feature tensor along the time dimension into a residual time step input matrix; extract the current time step data in the residual time step input matrix sequentially along the time sequence, introduce recursive weights by multiplying element by element, perform element-wise accumulation and activation of the multiplication result with the hidden state of the previous time step, and complete the parallel refresh of the hidden state without cross-neuron interference in each time step to extract residual evolution features; The correction matrix mapping layer is used to input the residual evolution features into the fully connected mapping network to perform linear transformation and activation processing, and reshape the transformed and activated high-dimensional feature tensor into a two-dimensional matrix structure that matches the dimension of the high-frequency transient control parameters, and output the aligned correction matrix in real time.

2. The diesel engine heterogeneous model simulation method based on an industrial mechanism model according to claim 1, characterized in that, S1 specifically includes: S11. Based on multi-source heterogeneous state flow and control flow, extract the namespace features and dimensional statistical distribution features of variable identifiers, and adaptively construct a unified variable dictionary; S12. Inject the multi-source heterogeneous state flow and control flow into a unified variable dictionary, match namespace features to perform dynamic format conversion, and output homogenized intermediate state flow and homogenized intermediate control flow. S13. Based on the dimensional statistical distribution characteristics, perform adaptive unit mapping on the homogenized intermediate state flow and the homogenized intermediate control flow, and continuously output the standardized heterogeneous model input data flow.

3. The diesel engine heterogeneous model simulation method based on an industrial mechanism model according to claim 1, characterized in that, S2 specifically includes: S21. Based on the input data stream of the standardized heterogeneous model, extract the time-series sliding window variance and state transition probability distribution of multidimensional variables, and adaptively define the confidence judgment boundary based on the variance extreme points and probability mutation points. S22. Inject the standardized heterogeneous model input data stream into the confidence decision boundary, calculate the boundary exceedance deviation rate of the multidimensional variables, and output the deviation rate time series. S23. Extract local gradient features and duration features based on the deviation rate time series, perform state fluctuation limit recognition based on the local gradient features and duration features, and output asynchronous time series advancement trigger signal.

4. The diesel engine heterogeneous model simulation method based on an industrial mechanism model according to claim 1, characterized in that, S3 specifically includes: S31. Based on the asynchronous timing advance trigger signal, extract the local variance decay characteristics and step size migration probability distribution of the high-frequency step size evolution sequence, and adaptively deduce the current execution step size based on the local variance decay characteristics and step size migration probability distribution. S32. The asynchronous timing advance trigger signal and the current execution step size are directly injected into the fast dynamic electronic control process. Based on the current execution step size, the discrete jump process of the control parameters is driven to perform forward deduction and output a high-frequency transient control jump sequence. S33. Inject the high-frequency transient control jump sequence into the hydraulic and pneumatic transient transmission process to perform forward dynamic evolution, and output the high-frequency transient state response and high-frequency transient control parameters based on the evolution trajectory mapping of the hydraulic and pneumatic transient transmission process.

5. The diesel engine heterogeneous model simulation method based on an industrial mechanism model according to claim 1, characterized in that, S4 specifically includes: S41. Based on the high-frequency transient control parameters and the high-frequency transient state response, extract the amplitude fluctuation attenuation characteristics of the high-frequency transient control parameters and the statistical distribution of the rate of change of the high-frequency transient state response, and adaptively deduce the slow dynamic solution step size based on the amplitude fluctuation attenuation characteristics and the statistical distribution of the rate of change. S42. Inject high-frequency transient control parameters and high-frequency transient state response into the industrial mechanism model to drive the fuel injection and intake swirl mixing process to perform forward flow field evolution and output in-cylinder transient flow field spatial distribution data. S43. Inject the transient flow field spatial distribution data in the cylinder into the in-cylinder thermodynamic process. Based on the slow dynamic solution step size, drive the in-cylinder thermodynamic process to perform a variable step size combustion heat release and cylinder wall heat transfer coupled solution, and output slow dynamic response characteristic data.

6. The diesel engine heterogeneous model simulation method based on an industrial mechanism model according to claim 1, characterized in that, S6 specifically includes: S61. The alignment correction matrix is ​​superimposed element by element onto the high-frequency transient control parameters of the fast dynamic electronic control process to generate high-frequency correction control parameters, drive the fast dynamic electronic control process to perform forward solving, and output the high-frequency correction state response of the current iteration step. S62. Inject the high-frequency correction state response and high-frequency correction control parameters of the current iteration step into the industrial mechanism model, drive the in-cylinder thermodynamic process to perform variable step-size coupled solution, and output the slow dynamic correction response characteristic data of the current iteration step. S63. Perform element-by-element difference calculation on the slow dynamic correction response characteristic data of the current iteration step and the expected response of the industrial mechanism theory corresponding to the high-frequency correction control parameters to generate the residual sequence of the current iteration step. S64. Based on the residual sequence of the current iteration step, calculate the local variance decay rate and local range convergence rate along the time dimension. Adaptively determine the cutoff position according to the inflection point of the change in variance decay rate and cut off the tail residual segment. Construct the iterative residual tail feature based on the global variance and global range of the residual segment. S65. Based on the relative proportion of global variance and global range in the tail features of the iterative residual, normalize and map the corresponding dynamic weights, perform weighted aggregation of global variance and global range according to the dynamic weights, and adaptively generate the dynamic convergence boundary of the current iteration step. S66. Based on the residual sequence of the current iteration step and the dynamic convergence boundary of the current iteration step, perform a spatial constraint comparison. If the constraint conditions are met, output a convergence trigger signal and simultaneously output the residual sequence of the current iteration step. Otherwise, inject the residual sequence of the current iteration step as a feedback source into the improved IndRNN model to trigger the calculation of the next iteration step.

7. The diesel engine heterogeneous model simulation method based on an industrial mechanism model according to claim 1, characterized in that, Specifically, S7 includes: S71. Based on the convergence trigger signal, terminate the simulation of the current working condition and extract the high-frequency corrected state response and slow dynamic corrected response at the end in real time; use the slow dynamic time node as the anchor point to downsample the high-frequency transient state response, extract the high-frequency transient state response downsampled sequence with the timestamps completely aligned and splice it with the slow dynamic response feature data row by row to construct a dual time scale joint response matrix. S72. Extract the upper and lower limits of the expected response of the industrial mechanism theory corresponding to the high-frequency transient state response downsampling sequence and slow dynamic response feature data of the dual-time-scale joint response matrix, drive the dual-time-scale joint response matrix to dynamically complete the dimensionless normalization mapping according to the mechanism constraint interval, and generate a standardized joint response matrix. S73. Dynamically extract the fast and slow dynamic mean features along the rows of the standardized joint response matrix and weight them according to the dynamic fusion benchmark to aggregate and output the joint performance index.

Images