Online-dcnn-based full-parameter online adaptive modeling method for hybrid electric engine

Abstract

The online self-adaptive modeling method for hybrid engine based on Online-DCNN comprises the following steps: step 1, obtaining and preprocessing multi-dimensional engine operation data; step 2, network architecture initialization and basic parameter setting; step 3, construction of dynamic convolution mechanism and design of expert convolution kernel library; step 4, offline pre-training of the framework using historical batch data, giving the model a baseline prediction capability under the initial physical state; step 5, when the system is switched to actual operation, the single sample real-time data stream generated by the physical entity is connected to the system; step 6, using the online Adam optimization algorithm to update the model parameters in real time according to the single-step prediction error; step 7, the prediction accuracy and generalization effect of the foregoing online self-adaptive updating mechanism are quantitatively evaluated in real time through multi-dimensional evaluation indexes, and an objective evaluation scale is provided for the engineering usability of the model. The present application can automatically identify and track the changes of engine performance state according to real-time monitoring data.

Description

Technical Field

[0001] This invention belongs to the field of aero-engine modeling technology, and specifically relates to an online adaptive modeling method for all parameters of a hybrid electric engine based on Online-DCNN. Background Technology

[0002] While existing mechanistic models possess clear physical meaning, the increasing complexity of hybrid electric vehicle systems and their significant nonlinear characteristics make it difficult to determine precise mathematical expressions, resulting in high modeling complexity and difficulty in adapting to complex dynamic operating conditions. This invention avoids complex physical mechanism derivations through a data-driven approach.

[0003] Existing data-driven modeling methods typically employ offline training based on historical data. However, engines operate in harsh environments for extended periods, inevitably experiencing mechanical wear and performance degradation, leading to a shift in the distribution of their real-time operational data. Offline-trained models, with their fixed parameters, cannot detect these changes, causing the models to gradually become "outdated," and their predictive accuracy to significantly decrease as engine operating time increases and performance degrades.

[0004] Since existing models cannot update parameters in real time, when the actual state of the engine deviates from the preset state of the model, control or fault diagnosis schemes designed based on outdated models have unknown safety hazards and cannot meet the requirements of high-precision real-time control.

[0005] The physical mechanism-based modeling method has the following main drawbacks. First, the modeling process requires detailed component geometric parameters, aerodynamic design parameters, and component characteristic data. This core, confidential data is difficult to obtain, limiting the modeling work. Second, the process involves a large number of equation solving and iterative calculations, resulting in a long modeling cycle and difficulty in quickly adapting to the modeling needs of different engine models. Third, the model accuracy heavily depends on the component characteristic diagrams under design conditions. In actual operation, component wear and corrosion can cause characteristics to deviate from design values, gradually reducing the accuracy of the fixed-parameter mechanistic model and failing to accurately reflect the actual performance state of the engine. Fourth, the model involves iterative solving of numerous nonlinear differential equations, resulting in long single-step simulation times and poor real-time performance, making it difficult to meet the rapid response requirements of airborne control systems.

[0006] Modeling methods based on system identification have significant limitations. Aero-engines are strongly nonlinear systems, with significant differences in input and output at different operating points. The linear or locally linearized models established by system identification are only effective near specific operating points, exhibiting poor extrapolation capabilities. When the engine operating point deviates from the identified operating condition or operates within a wide envelope, the model prediction error increases rapidly. Furthermore, obtaining the identified model often requires designing dedicated excitation signals and conducting experiments, adding extra costs.

[0007] Existing data-driven modeling methods based on fixed network structures lack adaptability. Traditional neural network models often employ offline training, requiring the collection of massive amounts of experimental data covering the entire data envelope. Due to the high cost of bench and flight testing, obtaining complete data is impractical, easily leading to limited model generalization ability. Because engine systems exhibit time-varying degradation characteristics under long-term operation, traditional networks have their parameters fixed after training, failing to perceive dynamic operating condition changes and mechanical performance drift. When encountering states outside the training set distribution, prediction accuracy drops significantly. Furthermore, traditional static feedforward networks directly map input to output, making it difficult to accurately capture the dynamic response process and temporal evolution of hybrid electric engines as complex time-varying systems. Summary of the Invention

[0008] To overcome the above technical problems, the present invention aims to provide a fully parameter-based online adaptive modeling method for hybrid electric engines based on an Online Dynamic Convolution Neural Network (Online-DCNN). By integrating physical knowledge and data-driven technology, this method significantly reduces computational complexity and dependence on detailed design parameters and large amounts of experimental data while ensuring model prediction accuracy, achieving high-precision and rapid modeling under small sample conditions. The model possesses self-learning and adaptive capabilities, enabling it to automatically identify and track changes in engine performance status based on real-time monitoring data.

[0009] To achieve the above objectives, the technical solution adopted by the casting system of the present invention is as follows:

[0010] A fully parameter-based online adaptive modeling method for hybrid electric motors based on Online-DCNN includes the following steps;

[0011] Step 1: Acquire and preprocess multi-dimensional engine operation data to provide a high-quality data foundation for subsequent modeling;

[0012] Step 2: Initialize the architecture and basic parameters of the one-dimensional convolutional neural network (1D-CNN) to provide a basic topological carrier for the embedding of dynamic convolution mechanisms;

[0013] Step 3: Construct a dynamic convolutional and expert kernel library based on a routing mechanism to form an input-adaptive feature extraction module for global model optimization;

[0014] Step 4: Use historical batch data to perform offline pre-training of the framework, giving the model the benchmark prediction capability under the initial physical state;

[0015] Step 5: Once the system is in actual operation, the real-time data stream of single samples generated by the physical entity is connected to the system to drive state prediction and trigger the learning of underlying parameters.

[0016] Step 6: Update the model parameters in real time based on the single-step prediction error using the online Adam optimization algorithm;

[0017] Step 7: As a closed-loop monitoring mechanism, the prediction accuracy and generalization effect of the aforementioned online adaptive update mechanism are evaluated in real time through multi-dimensional evaluation indicators, providing an objective evaluation scale for the engineering usability of the model.

[0018] Through the high-frequency interactive loop of steps five and six, the system is endowed with the ability to continuously learn, enabling the model to automatically identify and track changes in engine performance status based on real-time monitoring data, effectively overcoming the problem of accuracy decay caused by mechanical wear or sudden changes in operating conditions during long-term operation.

[0019] Step 1 specifically involves:

[0020] A complete data acquisition system is established during the actual operation of the hybrid electric engine. This system acquires multi-dimensional status parameters in real time through sensors distributed in key parts such as the intake manifold, compressor, turbine exhaust section, power output shaft and motor control node.

[0021] The data acquisition process covers the three power modes of the hybrid engine: pure electric, hybrid, and pure gas. It also covers a variety of typical operating conditions, including idling, acceleration, cruising, and high load, to ensure that the collected data can fully reflect the dynamic characteristics of the engine under different operating conditions.

[0022] The collected parameters include timing signals of engine speed, torque, throttle opening, intake air temperature, intake air pressure, exhaust air temperature, fuel flow rate, and power.

[0023] After data collection is completed, the raw data is preprocessed, including removing outliers, filling in missing values, filtering out high-frequency noise interference, and normalizing parameters of different dimensions to the same numerical range, thereby providing high-quality input samples for subsequent network training.

[0024] Step two specifically involves:

[0025] To address the processing requirements of time-series signals from hybrid electric motors, a one-dimensional convolutional neural network (1D-CNN) is adopted as the basic architecture; the input samples are set as time-series vectors. ,in express Engine state parameters collected at all times; the convolutional layer performs feature mapping on the input sequence through a sliding convolution kernel, and its convolution operation is represented as:

[0026]

[0027] In the formula, For the first Layer The output features of each convolutional kernel; These are the convolution kernel weight parameters; The kernel size; For bias terms; For non-linear activation functions, the ReLU function is chosen, and its expression is:

[0028]

[0029] The pooling layer uses max pooling to downsample the convolutional features, reducing the data dimensionality while preserving salient features. Its operational expression is:

[0030]

[0031] In the formula, For the first The feature region covered by each pooled window; This is the pooled output value;

[0032] After multiple convolutional-pooling operations, the extracted high-level abstract features are flattened and fed into a fully connected layer, where a linear weighted sum is used to obtain the final output.

[0033]

[0034] In the formula, This is the weight matrix of the fully connected layer; It is the bias vector; These are the estimated values ​​of the engine state parameters from the model.

[0035] Step three specifically involves:

[0036] Dynamic convolutional neural networks achieve adaptive generation of convolutional kernels through a routing mechanism. First, n basic convolutional kernels are predefined as an "expert kernel library," each adept at capturing different types of feature patterns. When input data enters the network, the system uses a routing function to determine which expert convolutional kernels should be activated and their respective activation levels.

[0037] The routing function adopts a lightweight design, which includes a global average pooling layer, a fully connected layer, and a Sigmoid activation layer in sequence;

[0038] The global average pooling layer compresses the input feature map into a compact global feature vector. The fully connected layer learns the correlation between the input features and each expert convolutional kernel. The sigmoid layer maps the correlation to weight coefficients between 0 and 1. After the routing function outputs n attention weights, the network linearly weights and combines the n basic convolutional kernels according to their corresponding weights to generate a dynamic convolutional kernel optimized for the current input. Finally, this dynamic convolutional kernel is used to perform a standard convolution operation on the input, and after passing through an activation function, the feature representation is output. The formula for the convolutional kernel used in the convolution operation is:

[0039]

[0040] Where n represents the number of expert convolutional kernels in the convolutional layer. This represents the final convolution kernel. These are the weights corresponding to the i-th expert convolutional kernel. It is the convolution kernel of the i-th expert; the output after convolution is:

[0041]

[0042] in This represents the output after convolution. Indicates the input sample. This represents the activation function.

[0043] Step four specifically involves:

[0044] After the network architecture and dynamic convolution mechanism are built, offline pre-training is first performed using the collected historical running data to provide an initial state for the network parameters that is close to the optimal solution.

[0045] The offline training process employs a batch gradient descent strategy, dividing the preprocessed historical data into training, validation, and test sets. The training set is used for parameter optimization, the validation set is used for hyperparameter tuning and preventing overfitting, and the test set is used to evaluate the model's generalization performance.

[0046] Each flight mission is recorded as one cycle. To avoid the model becoming dependent on the sample distribution, the samples in the first cycle are randomly shuffled and then divided into training and validation sets, while the test set remains unchanged to simulate real-time data acquisition. In each training batch, several time-series samples are simultaneously fed into the network. After forward propagation to calculate the predicted output, the mean squared error loss function is used to measure the difference between the predicted and actual measured values. The gradient of the loss function with respect to all trainable parameters in the network is calculated using the backpropagation algorithm, including the weights of each convolutional kernel, bias terms, expert convolutional kernel parameters in dynamic convolution, weights of the fully connected layers in the routing function, and output layer parameters. The Adam optimizer is used to update the parameters based on the gradient information. The Adam algorithm adaptively adjusts the learning rate for each parameter by maintaining the first and second moment estimates of the gradient, thereby achieving fast and stable convergence. After multiple epochs of iterative training, the network gradually learns the engine input-output mapping patterns and dynamic response characteristics contained in the historical data. Training stops when the loss on the validation set no longer decreases significantly, resulting in an initialized model with good performance under historical conditions.

[0047] Step five specifically involves:

[0048] After the hybrid engine was put into actual operation, the model switched from offline pre-training mode to online learning mode. In the online learning stage, batch data was no longer used for training. Instead, the current state measurement values ​​of the engine were received in real time as a data stream. The engine control system acquired sensor data at a fixed sampling period. Whenever a new set of observation data arrived, the online learning algorithm was immediately triggered. At this time, the operating logic of the model changed fundamentally. It no longer passively used fixed parameters for prediction, but actively learned the latest system characteristics contained in the new data while making each prediction. The real-time nature of the data stream required the online learning algorithm to have high computational efficiency and be able to complete the parameter update of a single sample within a millisecond time scale to ensure that the real-time response performance of the control system was not affected. The online learning adopted a single-sample stochastic gradient descent strategy, which used only a single observation sample at the current moment to calculate the gradient and update the parameters each time, avoiding the computational delay and memory overhead caused by batch processing.

[0049] Step six specifically involves:

[0050] For the sample received at time t First, the loss function is calculated, using the mean squared error as the loss function:

[0051]

[0052] In the formula, This represents the set of all trainable parameters of a CNN, including the kernels of each layer. Bias and fully connected layer weights ;

[0053] Calculate the gradient of the loss function with respect to the parameters:

[0054]

[0055] The Adam algorithm is used to update the first-order moment estimate (momentum) and the second-order moment estimate:

[0056]

[0057]

[0058] In the formula, and These are the exponential decay rates of the first and second moments, respectively;

[0059] because and In the initial stage, it will be biased towards zero, and deviation correction will be performed:

[0060]

[0061]

[0062] The final parameter update formula is:

[0063]

[0064] In the formula, The learning rate; To prevent small constants from being divided by zero.

[0065] Step seven specifically involves:

[0066] Select the mean absolute error (MAE), root mean square error (RMSE), and goodness of fit (R²). 2 As a core evaluation indicator, the specific calculation method and physical meaning are described below:

[0067] (1) Mean Absolute Error (MAE) is used to reflect the actual average level of prediction error. This index is calculated as the average of the absolute values ​​of the difference between the predicted value and the actual value; its calculation formula is:

[0068]

[0069] (2) Root mean square error is used to measure the sample standard deviation of the deviation between the predicted value and the true value; since the error is squared during the calculation process, its calculation formula is:

[0070]

[0071] (3) Goodness of fit is used to characterize the model's ability to explain the variability of observed data. It is a dimensionless relative index for evaluating the model's fit. The value of this index is usually between 0 and 1. The closer the value is to 1, the better. Its calculation formula is:

[0072]

[0073] Where N represents the total number of test sample sequences; For the first The actual measured value (true value) of the engine sensor at any given time; For the first The predicted output value of the Online-DCNN model at time point; This represents the arithmetic mean of the actual measurements in the test sample set. By comprehensively analyzing these three indicators, an objective evaluation of the model's steady-state accuracy and dynamic tracking performance can be achieved.

[0074] The beneficial effects of the present invention.

[0075] 1. This invention overcomes the accuracy degradation problem caused by the fixed parameters of offline models, achieving dynamic performance tracking and high-precision prediction of hybrid electric engines throughout their entire lifecycle. The proposed dynamic modeling method based on Online-DCNN introduces the online Adam optimization algorithm. This method utilizes real-time sensor data to update all trainable parameters in the network (including dynamic convolutional kernel weights, routing parameters, and bias terms) in real time. This mechanism enables the model to continuously learn and adaptively evolve, capturing and tracking changes in system characteristics caused by engine aging or failure in real time. This ensures that the model maintains a high degree of prediction accuracy matching the real system at all stages of the engine's lifecycle, even when physical parameters degrade significantly, significantly improving the system's reliability and safety.

[0076] 2. This invention addresses the challenge of single fixed models failing to adequately handle multimodal and wide-envelope operating conditions in hybrid electric systems, significantly improving the model's generalization ability and robustness under complex dynamic conditions. The invention constructs a dynamic convolutional neural network structure based on a routing mechanism and designs an expert convolutional kernel library containing multiple basic kernels. The model can adaptively calculate attention weights based on the feature distribution of the current input state through a routing function, dynamically generating the optimal composite convolutional kernel for the current operating condition. This adaptive input mechanism allows the model to flexibly and seamlessly switch feature extraction logic between different dynamic modes and operating conditions, effectively solving the problem of poor generalization ability of traditional fixed models within a wide operating envelope, and avoiding computational redundancy caused by building multiple sub-models while ensuring accuracy.

[0077] 3. This invention eliminates the reliance on detailed engine physical design parameters and large amounts of expensive experimental data, achieving low-cost, high-efficiency rapid modeling and real-time computation. Employing a data-driven Online-DCNN architecture, it directly mines the input-output mapping relationships in sensor monitoring data, without requiring prior knowledge of the engine's complex internal physical structure and aerodynamic-thermodynamic mechanisms, significantly reducing the modeling threshold and development cycle. This method can complete forward prediction and parameter updates within millisecond timescales, meeting the stringent real-time requirements of aero-engine airborne control systems and providing efficient technical support for real-time fault diagnosis and hardware-in-the-loop simulation. Attached Figure Description

[0078] Figure 1 This is a typical CNN structure.

[0079] Figure 2 Dynamic convolutional neural network structure.

[0080] Figure 3 Schematic diagram of routing function.

[0081] Figure 4 Dynamic convolutional structure framework.

[0082] Figure 5 Online-DCNN overall system architecture diagram.

[0083] Figure 6 Model output comparison.

[0084] Figure 7 Comparison of model evaluation metrics. Detailed Implementation

[0085] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0086] This invention proposes a dynamic model for a hybrid electric motor based on the Online-DCNN (Online Dynamic Convolutional Neural Network) intelligent algorithm, which essentially falls under the category of online learning. Online learning refers to the ability of a model to respond instantly and adjust its parameters synchronously as new observation data is acquired during actual system operation, thereby achieving dynamic adaptive updates. To construct the Online-DCNN dynamic model, this invention makes two key improvements to the traditional CNN network architecture.

[0087] Step 1: Data Acquisition and Preprocessing

[0088] A complete data acquisition system is established during the actual operation of the hybrid electric engine. This system acquires multi-dimensional status parameters in real time through sensors distributed in key parts such as the intake duct, compressor, turbine exhaust section, power output shaft and motor control node.

[0089] The data acquisition process covers the three power modes of the hybrid engine: pure electric, hybrid, and pure gas. It also covers a variety of typical operating conditions such as idling, acceleration, cruising, and high load, ensuring that the collected data can fully reflect the dynamic characteristics of the engine under different operating conditions.

[0090] The collected parameters mainly include timing signals such as engine speed, torque, throttle opening, intake air temperature, intake air pressure, exhaust air temperature, fuel flow rate, and power.

[0091] After data collection is completed, the raw data is preprocessed, including removing outliers, filling in missing values, filtering out high-frequency noise interference, and normalizing parameters of different dimensions to the same numerical range, thereby providing high-quality input samples for subsequent network training.

[0092] Step Two: Network Architecture Initialization and Basic Parameter Settings

[0093] Online-DCNN is designed based on the extension and optimization of the classic convolutional neural network structure, so its basic architecture follows the typical CNN network topology. A classic CNN consists of multiple functional layers connected in series, mainly including an input layer, convolutional layers, activation layers, pooling layers, fully connected layers, and an output layer. The core design concept of this network is to automatically extract local features through convolution operations, significantly reduce the model parameter size through a weight sharing mechanism, and enhance feature dimensionality reduction and anti-interference capabilities through pooling operations.

[0094] The network's input layer receives a preprocessed time-series vector of engine state parameters. The sequence length is determined based on the actual sampling frequency and the engine's dynamic response characteristics, set to cover a complete dynamic transition process within a time window. The convolutional layer design determines hyperparameters such as the number of kernels, kernel size, and stride based on the frequency characteristics and time-scale features of the engine signal. Multiple cascaded convolutional layers extract multi-scale features from local details to global patterns layer by layer. A pooling layer is configured after each convolutional layer to perform max-pooling, reducing feature dimensionality through downsampling while retaining the most significant feature responses, enhancing the model's robustness to noise and disturbances in the input signal. After multi-layer convolutional-pooling feature extraction, the high-dimensional feature map is flattened and fed into a fully connected layer, where linear weighted mapping yields the predicted output of key engine performance parameters.

[0095] Figure 1 This diagram illustrates the basic structure of a typical one-dimensional convolutional neural network (CNN). From left to right, the network consists of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The input layer receives a temporally distributed sequence of engine state parameters. The convolutional layers extract local features using fixed-stride kernels. The pooling layers perform max pooling to reduce the dimensionality of the data. Finally, the fully connected layers linearly map the high-dimensional features to the predicted output value.

[0096] Figure 2 This is a diagram of the dynamic convolutional neural network structure constructed in this invention. Unlike traditional static CNNs, the input data is divided into two computational branches here: one branch enters the routing function (Route FN) to calculate the attention weights corresponding to each of the n expert convolutional kernels based on the current input features; the other branch performs a convolution operation with the dynamic composite convolutional kernel generated by the linear combination of the above weights, and finally outputs the result after processing by the activation layer.

[0097] To address the processing requirements of time-series signals from hybrid electric vehicles, this invention employs a one-dimensional convolutional neural network (1D-CNN) as its basic architecture. Let the input sample be a time-series vector. ,in express Engine state parameters are collected at all times. The convolutional layer performs feature mapping on the input sequence using a sliding convolution kernel; its convolution operation can be represented as:

[0098]

[0099] In the formula, For the first Layer The output features of each convolutional kernel; These are the convolution kernel weight parameters; The kernel size; For bias terms; As a non-linear activation function, this invention selects the ReLU function, whose expression is:

[0100]

[0101] The pooling layer uses max pooling to downsample the convolutional features, reducing the data dimensionality while preserving salient features. Its operational expression is:

[0102]

[0103] In the formula, For the first The feature region covered by each pooled window; This is the pooled output value.

[0104] After multiple convolutional-pooling operations, the extracted high-level abstract features are flattened and fed into a fully connected layer, where a linear weighted sum is used to obtain the final output.

[0105]

[0106] In the formula, This is the weight matrix of the fully connected layer; It is the bias vector; These are the estimated values ​​of the engine state parameters from the model.

[0107] Classical convolutional neural networks use static convolutional kernels for feature extraction, meaning that all kernel parameters remain fixed after training, and the same feature extraction strategy is used for all input data. This approach has significant limitations in complex systems with multiple operating conditions and strong time-varying characteristics, such as hybrid electric vehicles (HEVs). HEVs exhibit drastically different dynamic characteristics under three power modes (pure electric, hybrid, and pure combustion) and under different operating conditions such as idling, acceleration, and high load. Fixed convolutional kernels cannot adjust the feature extraction mode for different operating conditions, leading to a significant decrease in model accuracy under certain conditions. As HEVs degrade over long-term operation due to battery performance degradation, combustion chamber carbon buildup, and turbocharger wear, the system characteristics gradually deviate from the training data distribution. Static convolutional kernels cannot adaptively track this time-varying process, causing the model accuracy to continuously decline over time. To cover multiple operating conditions and maintain model accuracy, traditional methods often require adding a large number of convolutional kernels or training multiple sub-models. This invention introduces a dynamic convolutional neural network and uses a routing mechanism to adaptively generate the input of the convolutional kernel. This enables the model to automatically match the optimal feature extraction strategy for the hybrid engine state under different power modes, operating conditions, and degradation stages. While maintaining parameter efficiency and real-time computation, it significantly improves the model's adaptability to multiple operating conditions, its ability to track time-varying characteristics, and its generalization robustness to abnormal states.

[0108] Step 3: Construction of dynamic convolution mechanism and design of expert convolution kernel library:

[0109] Dynamic Convolutional Neural Networks (CNNs) are parameter-adaptive deep learning architectures where convolutional kernel parameters are not fixed but dynamically generated based on the features of the input data. Unlike traditional CNNs that use pre-trained static convolutional kernels, dynamic convolutions use attention or gating mechanisms to linearly combine multiple basic convolutional kernels according to input-related weight coefficients, thereby generating a specific convolutional kernel for each input sample. This mechanism allows the model to automatically adjust its feature extraction strategy based on different input patterns, exhibiting stronger representation and generalization capabilities compared to static convolutions when dealing with complex systems with significant time-varying characteristics and multiple operating conditions. In engine control systems, because engines exhibit different dynamic characteristics under different operating conditions such as speed, load, and temperature, and their performance degrades over long-term operation, traditional fixed-parameter models struggle to accurately capture these time-varying nonlinear relationships. Dynamic convolutions, by adjusting the feature extraction kernel in real time, can adaptively track changes in engine state, providing a new technical path for building high-precision and robust dynamic engine models.

[0110] The core principle of dynamic convolutional neural networks is to achieve adaptive generation of convolutional kernels based on input through a routing mechanism. First, n basic convolutional kernels are predefined as an "expert convolutional kernel library," each adept at capturing different types of feature patterns. When input data enters the network, the system uses a dedicated routing function to determine which expert convolutional kernels should be activated and their activation levels. The routing function employs a lightweight design, containing a global average pooling layer, a fully connected layer, and a sigmoid activation layer. The global average pooling layer compresses the input feature map into a compact global feature vector, the fully connected layer learns the correlation between the input features and each expert convolutional kernel, and the sigmoid layer maps the correlation to weight coefficients between 0 and 1. After the routing function outputs n attention weights, the network linearly weights and combines the n basic convolutional kernels according to their corresponding weights to generate a dynamic convolutional kernel specifically optimized for the current input. Finally, this dynamic convolutional kernel is used to perform standard convolution operations on the input, and after passing through the activation function, the feature representation is output. The formula for the convolutional kernel used for convolution operations is:

[0111]

[0112] Where n represents the number of expert convolutional kernels in the convolutional layer. This represents the final convolution kernel. These are the weights corresponding to the i-th expert convolutional kernel. It is the convolution kernel of the i-th expert. The output after convolution is:

[0113]

[0114] in This represents the output after convolution. Indicates the input sample. This represents the activation function.

[0115] The improved core model structure breaks away from the traditional paradigm of using a single fixed-weight convolutional kernel, innovatively introducing an input adaptive feature extraction module composed of a lightweight routing function and an expert kernel library. In the network's forward inference structure, the input temporal signal not only enters the regular feature mapping channel but is also directed to the routing function branch. This routing branch sequentially passes through a global average pooling layer (GAP), a fully connected layer (FC), and a sigmoid activation layer, sensing and quantifying the feature distribution of the current physical condition in real time, and then adaptively outputting a set of attention weights between 0 and 1. The system then uses this weight matrix to apply attention weights to pre-set expert kernels in the expert library. The basic convolutional kernels are linearly weighted and combined to instantly generate the optimal composite convolutional kernel specific to the current input state, and then used to perform subsequent forward convolution operations.

[0116] Figure 3 This is a schematic diagram of the routing function. The routing function adopts a lightweight cascaded architecture, which includes a global average pooling layer (GAP), a fully connected layer (FC), and a sigmoid activation layer. The GAP layer is responsible for compressing the input local feature map into a global feature vector. The fully connected layer mines the non-linear correlation between this vector and each expert kernel. The sigmoid layer then converts it into n expert weights between 0 and 1 as outputs.

[0117] Dynamic convolutional neural networks (CNNs) improve the adaptability of models to various operating conditions of hybrid electric vehicles (HEVs) to some extent, but their network parameters remain fixed after optimization during the training phase. In actual operation, the state of a HEV gradually deviates from the training data distribution over time. Fixed network parameters cannot track this continuously evolving system characteristic, inevitably leading to a decrease in model accuracy over time. To address the insufficient time-varying adaptability caused by the fixed parameters of dynamic CNNs, this invention further introduces the online Adam optimization algorithm. This algorithm updates all trainable parameters of the network in real time, enabling the model to continuously learn the latest system characteristics during HEV operation. It achieves synchronous tracking of parameter dynamic evolution and system degradation, thereby maintaining stable high-precision prediction capabilities throughout the entire lifecycle.

[0118] Figure 4 This diagram illustrates the framework of a dynamic convolutional structure, showcasing the generation mechanism of the dynamic convolutional kernel and the entire feature extraction process. After the input feature map enters the network, it is processed by a routing function. It perceives the characteristics of the current working condition and outputs a set of corresponding attention weights. Subsequently, the system assigns weights to multiple pre-defined "expert cores". Perform linear weighted summation, i.e. In real time, the optimal composite convolution kernel that best matches the current working conditions is synthesized. This is used to perform the final convolution operation on the input.

[0119] Step 4: Batch learning in the offline pre-training stage:

[0120] After the network architecture and dynamic convolution mechanism are constructed, offline pre-training is first performed using collected historical operational data to provide the network parameters with an initial state close to the optimal solution. The offline training process employs a batch gradient descent strategy, dividing the preprocessed historical data into training, validation, and test sets. The training set is used for parameter optimization, the validation set for hyperparameter tuning and preventing overfitting, and the test set for evaluating the model's generalization performance. Each flight mission is recorded as one cycle. To avoid the model becoming dependent on the sample distribution, the samples in the first cycle are randomly shuffled and then divided into training and validation sets in a 4:1 ratio, while the test set retains its time series to simulate real-time data acquisition. In each training batch, several time series samples are simultaneously fed into the network. After forward propagation to obtain the predicted output, the mean squared error loss function is used to measure the difference between the predicted and actual measured values. The gradient of the loss function with respect to all trainable parameters in the network is calculated using the backpropagation algorithm, including the weights of each convolutional kernel, bias terms, expert convolutional kernel parameters in dynamic convolution, weights of the fully connected layers in the routing function, and output layer parameters. The Adam optimizer updates parameters based on gradient information. The Adam algorithm adaptively adjusts the learning rate for each parameter by maintaining the first and second moment estimates of the gradient, thereby achieving fast and stable convergence. After multiple epochs of iterative training, the network gradually learns the engine input-output mapping patterns and dynamic response characteristics contained in historical data. Training stops when the loss on the validation set no longer decreases significantly, resulting in an initialized model that performs well under historical operating conditions.

[0121] Table 1 Dataset Partitioning

[0122]

[0123] Step 5: Real-time data stream access:

[0124] After the hybrid engine was put into actual operation, the model switched from offline pre-training mode to online learning mode. In the online learning phase, training no longer uses batch data; instead, it receives the engine's current state measurements via real-time data streams. The engine control system acquires sensor data at fixed sampling periods, and the online learning algorithm is immediately triggered whenever a new set of observations arrives. At this point, the model's operating logic undergoes a fundamental change; it no longer passively uses fixed parameters for prediction but actively learns the latest system characteristics contained in the new data with each prediction. The real-time nature of the data stream requires the online learning algorithm to have high computational efficiency, capable of updating parameters for a single sample within milliseconds to ensure that the real-time response performance of the control system is not affected. To meet this requirement, online learning employs a single-sample stochastic gradient descent strategy, calculating the gradient and updating parameters using only a single observation sample at the current moment, avoiding the computational latency and memory overhead caused by batch processing.

[0125] Step Six: Parameter Update Mechanism of Online Adam Optimization Algorithm:

[0126] The dynamic model employs the online Adam algorithm to update all trainable parameters of the DCNN online. The Adam optimization algorithm, or Adaptive Moment Estimation algorithm, is a first-order gradient optimization method applied to stochastic objective functions. By calculating the first-order moment estimate and the second-order primal moment estimate of the gradient separately, and introducing a bias correction mechanism, it achieves independent and adaptive learning rate adjustment for each model parameter. This algorithm integrates historical gradient direction information from the momentum method with the adaptive scaling capability of the RMSProp algorithm, thereby dynamically generating parameter update step sizes during iteration. This algorithm can efficiently handle sparse gradient and non-stationary objective optimization problems, exhibiting fast convergence speed; it demonstrates strong robustness to initial learning rate hyperparameter settings, typically requiring no fine-tuning; and its computational memory requirements are moderate, making it suitable for training large-scale parameter models. The Adam algorithm effectively overcomes the shortcomings of traditional stochastic gradient descent methods, such as slow convergence and difficulty in learning rate selection, demonstrating good convergence performance in online learning scenarios.

[0127] After receiving a new sample at time t, the online learning model first inputs the sample into the current Online-DCNN model for forward propagation computation. When the input time-series data passes through the dynamic convolutional layer, the routing function dynamically calculates the activation weights of each expert convolutional kernel based on the current input feature distribution and generates a composite convolutional kernel optimized for that input to perform the convolution operation. After the convolutional features are extracted by the activation function and pooling layers, they are propagated forward layer by layer to the fully connected layer, ultimately outputting the predicted values ​​for key engine parameters. The predicted values ​​are compared with the actual values ​​measured by the sensors, and the mean squared error between the two is calculated as the loss function. This loss function quantifies the degree of matching between the current model parameters and the real-time state of the engine. Subsequently, the gradient of the loss function with respect to all trainable parameters in the network is calculated through an automatic differentiation mechanism. The gradient vector indicates the direction in which each parameter should be adjusted to reduce the prediction error. The online Adam optimization algorithm maintains two sets of cumulative statistics: the first-order moment estimate and the second-order moment estimate, recording the historical moving average of the gradient and the historical moving average of the squared gradient, respectively. Upon receiving new gradient information, the two sets of statistics are updated using an exponentially weighted moving average method. A decay coefficient is introduced to make the algorithm focus more on recent gradient information and gradually forget distant history, thus adapting to the time-varying nature of the system characteristics. Since the cumulative statistics initially have a zero bias, a bias correction operation is performed, dividing the first and second moments by the corresponding bias correction factors to obtain unbiased moment estimates. Based on the corrected moment estimates, the Adam algorithm calculates an adaptive update step size for each parameter, determined by the learning rate, the square root of the first moment estimate, and the second moment estimate. For parameters with large gradient variance, the algorithm automatically uses a smaller step size to ensure stability, while for parameters with stable gradients, a larger step size is used to accelerate convergence. Finally, all parameters in the network are adjusted according to the calculated update step size, including the expert convolution kernel weights in dynamic convolution, the fully connected layer parameters of the routing function, the bias terms of each layer, and the output layer weights, so that the model parameters evolve in a direction that reduces the current prediction error.

[0128] For the sample received at time t First, calculate the loss function. We use the mean squared error as the loss function:

[0129]

[0130] In the formula, This represents the set of all trainable parameters of a CNN, including the kernels of each layer. Bias and fully connected layer weights .

[0131] Calculate the gradient of the loss function with respect to the parameters:

[0132]

[0133] The Adam algorithm is used to update the first-order moment estimate (momentum) and the second-order moment estimate:

[0134]

[0135]

[0136] In the formula, and Let be the exponential decay rates of the first and second moments, respectively, typically taken as... , .

[0137] because and In the initial stage, it will be biased towards zero, and deviation correction will be performed:

[0138]

[0139]

[0140] The final parameter update formula is:

[0141]

[0142] In the formula, The learning rate; To prevent division by zero of small constants, it is usually taken as... .

[0143] Step 7: Model Evaluation Metrics

[0144] To quantitatively evaluate the prediction accuracy and generalization performance of the proposed Online-DCNN intelligent algorithm in the dynamic modeling of hybrid engines, a multi-dimensional model evaluation index system is constructed. By calculating the deviation statistics between the model's predicted output values ​​and the actual sensor measurements of the engine, the model's fitting degree and robustness under different operating conditions, including pure electric, hybrid, and pure combustion, are comprehensively measured. This invention selects the mean absolute error (MAE), root mean square error (RMSE), and goodness of fit (R²) as the evaluation metrics. 2 As a core evaluation indicator, the specific calculation method and physical meaning are described below:

[0145] (1) Mean Absolute Error (MAE):

[0146] Mean Absolute Error (MAE) reflects the actual average level of prediction error. This metric calculates the average of the absolute values ​​of the differences between predicted and actual values. Because it directly averages the absolute values ​​of the errors, it intuitively reflects the degree of deviation of the model's predictions from the actual values, avoiding the problem of positive and negative errors canceling each other out. Its calculation formula is:

[0147]

[0148] (2) Root Mean Square Error (RMSE):

[0149] The root mean square error (RMSE) measures the sample standard deviation of the deviation between predicted and actual values. Because this metric squares the error during calculation, it amplifies the weight of larger errors, making RMSE more sensitive to outliers in the data. In hybrid electric vehicle modeling, this metric effectively evaluates the model's ability to predict extreme values ​​during drastic dynamic changes or transient responses. Its calculation formula is:

[0150]

[0151] (3) Goodness of fit (R) 2 ):

[0152] Goodness of fit characterizes a model's ability to explain the variability of observed data and is a dimensionless relative index used to evaluate the quality of a model's fit. The value typically ranges from 0 to 1; the closer the value is to 1, the better the model's predicted curve matches the actual data curve, the greater the proportion of engine state variations the model can explain, and the better its approximation performance. Its calculation formula is:

[0153]

[0154] Where N represents the total number of test sample sequences; For the first The actual measured value (true value) of the engine sensor at any given time; For the first The predicted output value of the Online-DCNN model at time point; This represents the arithmetic mean of the actual measurements in the test sample set. By comprehensively analyzing these three indicators, an objective evaluation of the model's steady-state accuracy and dynamic tracking performance can be achieved.

[0155] The present invention aims to provide a dynamic model with online learning capabilities. By updating parameters such as network center, expansion coefficients and weights in real time, the model can still match the engine in real time when the engine undergoes normal degradation or changes in operating conditions, thereby maximizing the accuracy of the model and the reliability of the system.

[0156] Simulation verification

[0157] This embodiment uses a high-fidelity whole-machine simulation platform to generate an operational dataset containing mechanical degradation characteristics.

[0158] Data from the first 30 operating cycles of the engine during its normal performance degradation phase were used for validation. Each flight mission was counted as one cycle, which for this engine is approximately 5 hours. To simulate the real-world processes of "offline pre-training" and "online real-time updating," the data partitioning strategy was as follows: the samples from the first cycle were randomly sorted and divided into an offline training set and a validation set in a 4:1 ratio; the massive amounts of data from the second to the 30th cycles were used as the test set, while strictly maintaining the chronological order of the data to simulate the single-sample real-time data stream generated by airborne sensors.

[0159] Monitoring the temperature of various cross-sections of a turbofan engine is a key aspect of engine condition monitoring. Establishing corresponding cross-sectional temperature models is crucial for engine control. Therefore, taking the low-pressure turbine (LPT) outlet cross-section as an example, an Online-DCNN dynamic model for the total LPT outlet temperature is established. To comprehensively compare the accuracy and stability of the proposed Online-DCNN dynamic model, several other models are provided for comparison, namely a typical CNN model, a typical RNN model, and an LSTM model.

[0160] The hybrid electric engine dynamic model based on the Online-DCNN intelligent algorithm proposed in this invention has broad application prospects and significant engineering value in intelligent control of aerospace power systems, full life cycle health management, and digital twin technology construction.

[0161] This invention can be directly applied to the design of high-precision real-time control systems for next-generation hybrid electric engines. With the transition of aviation power towards electrification, hybrid electric engines need to frequently switch between multiple modes, including pure electric, hybrid, and pure gas, and face complex nonlinear dynamic response characteristics. The model provided by this invention, with its millisecond-level computational efficiency and excellent dynamic condition adaptability, can be embedded as a core predictive model into model predictive control (MPC) or adaptive control algorithms. It solves the problem that traditional physical mechanism models have excessively long computation times, making them difficult to meet the real-time requirements of airborne control. It also overcomes the shortcomings of traditional offline neural network models, which cannot adapt to sudden changes in engine operating conditions during actual operation. This ensures that the control system can achieve precise control of speed, torque, and power distribution across the entire engine envelope, improving flight safety and fuel economy.

[0162] Secondly, this invention is a key enabling technology for realizing fault diagnosis and predictive maintenance (PHM) of aero-engines. During the long-term service of an engine, mechanical wear, corrosion, and carbon buildup can lead to irreversible performance degradation. This invention utilizes the online Adam optimization algorithm to achieve synchronous evolution of model parameters with the engine's physical state. By comparing the model's predicted output with the actual sensor measurements, performance drift interference caused by normal aging can be effectively eliminated, thereby greatly improving the sensitivity and accuracy of detecting sudden faults. Furthermore, based on the analysis of the model parameter evolution trajectory, reliable data support can be provided for engine remaining life prediction (RUL) and condition-based maintenance (CBM) decisions, significantly reducing maintenance costs.

[0163] This invention provides a low-cost, high-fidelity modeling solution for the construction of digital twin systems for aero-engines. Existing modeling techniques often heavily rely on core design parameters from manufacturers or expensive, massive amounts of bench test data, limiting the widespread adoption and application of digital twin technology. This invention adopts a data-driven approach, enabling rapid inversion of system dynamic characteristics using limited sensor monitoring data, significantly reducing the modeling threshold and development cycle. The model's "online self-learning" capability makes it a digital avatar with growth characteristics, realistically reflecting the current state of the physical entity in virtual space. This provides a highly reliable technical platform for conducting virtual flight tests, performance evaluations, and hardware-in-the-loop (HIL) simulation tests under extreme conditions, accelerating the digital and intelligent development of aero-engine technology.

[0164] like Figure 5 The diagram shows the overall system architecture of the online adaptive modeling method based on Online-DCNN of this invention. The system is divided into two working modes: the upper one is the "offline pre-training" module, which optimizes the parameters of the initial DCNN model using batch historical data; the lower one is the core "online learning stage" module, where the system receives real-time single-sample data streams from the hybrid electric motor, outputs dynamic state estimates (serving fault diagnosis and control strategies), calculates the prediction loss in parallel, and drives the online Adam optimizer to perform real-time closed-loop updates of the entire network parameters of the DCNN model.

[0165] like Figure 6As shown, this figure compares the outputs of different models, presenting temperature prediction output comparison curves for each model in four cycles of the hybrid engine. The figure compares the actual observed values ​​with the prediction trajectories of Online-DCNN, typical CNN, RNN, and LSTM algorithms. In the second cycle, the Online-DCNN dynamic model output has a large error compared to the observed values, while in other cycles, the model output fits the observed values ​​well. The second cycle marks the beginning of online learning; the Online-DCNN dynamic model is in its growth learning phase, reaching the required accuracy and entering the mature learning phase after training on approximately 8000 samples. While other models show smaller errors compared to the observed values ​​in the second cycle, their outputs gradually deviate from the observed values ​​as engine performance degrades. Because typical neural network models rely entirely on historical data and do not update model parameters during engine operation, the larger the number of engine operation cycles, the greater the error in the model output.

[0166] like Figure 7 As shown, the comparison of different model evaluation metrics provides a more intuitive view of the prediction performance of each model. After the second cycle, the accuracy of the Online-DCNN dynamic model rapidly improved, with higher scores in MAE, RMSE, and R-value. 2 The value tends to stabilize. However, other models show varying degrees of increase in MAE and RMSE values ​​over time, while R... 2 The value is also gradually decreasing.

Claims

1. A fully parameter-based online adaptive modeling method for hybrid electric motors based on Online-DCNN, characterized in that, Includes the following steps; Step 1: Acquire and preprocess multi-dimensional engine operating data; Step 2: Initialize the one-dimensional convolutional neural network architecture and basic parameters to provide a basic topological carrier for embedding dynamic convolution mechanisms; Step 3: Construct a dynamic convolutional and expert kernel library based on a routing mechanism to form an input-adaptive feature extraction module for global model optimization; Step 4: Use historical batch data to perform offline pre-training of the framework, giving the model the benchmark prediction capability under the initial physical state; Step 5: Once the system is in actual operation, the real-time data stream of single samples generated by the physical entity is connected to the system to drive state prediction and trigger the learning of underlying parameters. Step 6: Update the model parameters in real time based on the single-step prediction error using the online Adam optimization algorithm; Step 7: As a closed-loop monitoring mechanism, the prediction accuracy and generalization effect of the aforementioned online adaptive update mechanism are evaluated in real time through multi-dimensional evaluation indicators, providing an objective evaluation scale for the engineering usability of the model.

2. The online adaptive modeling method for all parameters of a hybrid electric motor based on Online-DCNN according to claim 1, characterized in that, Step 1 specifically involves: A complete data acquisition system is established during the actual operation of the hybrid electric engine. This system acquires multi-dimensional status parameters in real time through sensors distributed in key parts such as the intake manifold, compressor, turbine exhaust section, power output shaft and motor control node. The data acquisition process covers the three power modes of the hybrid engine: pure electric, hybrid, and pure gas. It also covers a variety of typical operating conditions, including idling, acceleration, cruising, and high load, to ensure that the collected data can fully reflect the dynamic characteristics of the engine under different operating conditions. The collected parameters include timing signals of engine speed, torque, throttle opening, intake air temperature, intake air pressure, exhaust air temperature, fuel flow rate, and power. After data collection is completed, the raw data is preprocessed, including removing outliers, filling in missing values, filtering out high-frequency noise interference, and normalizing parameters of different dimensions to the same numerical range, thereby providing high-quality input samples for subsequent network training.

3. The online adaptive modeling method for all parameters of a hybrid electric motor based on Online-DCNN according to claim 1, characterized in that, Step two specifically involves: To address the processing requirements of time-series signals from hybrid electric motors, a one-dimensional convolutional neural network (1D-CNN) is adopted as the basic architecture; the input samples are set as time-series vectors. ,in express Engine state parameters collected at all times; the convolutional layer performs feature mapping on the input sequence through a sliding convolution kernel, and its convolution operation is represented as: ； In the formula, For the first Layer The output features of each convolutional kernel; These are the convolution kernel weight parameters; The kernel size; For bias terms; For non-linear activation functions, the ReLU function is chosen, and its expression is: ； The pooling layer uses max pooling to downsample the convolutional features, reducing the data dimensionality while preserving salient features. Its operational expression is: ； In the formula, For the first The feature region covered by each pooled window; This is the pooled output value; After multiple convolutional-pooling operations, the extracted high-level abstract features are flattened and fed into a fully connected layer, where a linear weighted sum is used to obtain the final output. ； In the formula, This is the weight matrix of the fully connected layer; It is the bias vector; These are the estimated values ​​of the engine state parameters from the model.

4. The online adaptive modeling method for all parameters of a hybrid electric motor based on Online-DCNN according to claim 1, characterized in that, Step three specifically involves: Dynamic convolutional neural networks achieve adaptive generation of convolutional kernels through a routing mechanism. First, n basic convolutional kernels are predefined as an "expert kernel library," each adept at capturing different types of feature patterns. When input data enters the network, the system uses a routing function to determine which expert convolutional kernels should be activated and their respective activation levels. The routing function adopts a lightweight design, which includes a global average pooling layer, a fully connected layer, and a Sigmoid activation layer in sequence; The global average pooling layer compresses the input feature map into a compact global feature vector. The fully connected layer learns the correlation between the input features and each expert convolutional kernel. The sigmoid layer maps the correlation to weight coefficients between 0 and 1. After the routing function outputs n attention weights, the network linearly weights and combines the n basic convolutional kernels according to their corresponding weights to generate a dynamic convolutional kernel optimized for the current input. Finally, this dynamic convolutional kernel is used to perform a standard convolution operation on the input, and after passing through an activation function, the feature representation is output. The formula for the convolutional kernel used in the convolution operation is: ； Where n represents the number of expert convolutional kernels in the convolutional layer. This represents the final convolution kernel. These are the weights corresponding to the i-th expert convolutional kernel. It is the convolution kernel of the i-th expert; the output after convolution is: ； in This represents the output after convolution. Indicates the input sample. This represents the activation function.

5. The online adaptive modeling method for all parameters of a hybrid electric motor based on Online-DCNN according to claim 1, characterized in that, Step four specifically involves: After the network architecture and dynamic convolution mechanism are built, offline pre-training is first performed using the collected historical running data to provide an initial state for the network parameters that is close to the optimal solution. The offline training process employs a batch gradient descent strategy, dividing the preprocessed historical data into training, validation, and test sets. The training set is used for parameter optimization, the validation set is used for hyperparameter tuning and preventing overfitting, and the test set is used to evaluate the model's generalization performance. Each flight mission is recorded as one cycle. To avoid the model becoming dependent on the sample distribution, the samples in the first cycle are randomly shuffled and then divided into training and validation sets, while the test set remains unchanged to simulate real-time data acquisition. In each training batch, several time-series samples are simultaneously fed into the network. After forward propagation to calculate the predicted output, the mean squared error loss function is used to measure the difference between the predicted and actual measured values. The gradient of the loss function with respect to all trainable parameters in the network is calculated using the backpropagation algorithm, including the weights of each convolutional kernel, bias terms, expert convolutional kernel parameters in dynamic convolution, weights of the fully connected layers in the routing function, and output layer parameters. The Adam optimizer is used to update the parameters based on the gradient information. The Adam algorithm adaptively adjusts the learning rate for each parameter by maintaining the first and second moment estimates of the gradient, thereby achieving fast and stable convergence. After multiple epochs of iterative training, the network gradually learns the engine input-output mapping patterns and dynamic response characteristics contained in the historical data. Training stops when the loss on the validation set no longer decreases significantly, resulting in an initialized model with good performance under historical conditions.

6. The online adaptive modeling method for all parameters of a hybrid electric motor based on Online-DCNN according to claim 1, characterized in that, Step five specifically involves: After the hybrid engine was put into actual operation, the model switched from offline pre-training mode to online learning mode. In the online learning stage, batch data was no longer used for training. Instead, the current state measurement values ​​of the engine were received in real time as a data stream. The engine control system acquired sensor data at a fixed sampling period. Whenever a new set of observation data arrived, the online learning algorithm was immediately triggered. At this time, the operating logic of the model changed fundamentally. It no longer passively used fixed parameters for prediction, but actively learned the latest system characteristics contained in the new data while making each prediction. The real-time nature of the data stream required the online learning algorithm to have high computational efficiency and be able to complete the parameter update of a single sample within a millisecond time scale to ensure that the real-time response performance of the control system was not affected. The online learning adopted a single-sample stochastic gradient descent strategy, which used only a single observation sample at the current moment to calculate the gradient and update the parameters each time, avoiding the computational delay and memory overhead caused by batch processing.

7. The online adaptive modeling method for all parameters of a hybrid electric motor based on Online-DCNN according to claim 1, characterized in that, Step six specifically involves: For the sample received at time t First, the loss function is calculated, using the mean squared error as the loss function: ； In the formula, This represents the set of all trainable parameters of a CNN, including the kernels of each layer. Bias and fully connected layer weights ; Calculate the gradient of the loss function with respect to the parameters: ； The Adam algorithm is used to update the first-order moment estimate (momentum) and the second-order moment estimate: ； ； In the formula, and These are the exponential decay rates of the first and second moments, respectively; because and In the initial stage, it will be biased towards zero, and deviation correction will be performed: ； ； The final parameter update formula is: ； In the formula, The learning rate; To prevent small constants from being divided by zero.

8. The online adaptive modeling method for all parameters of a hybrid electric motor based on Online-DCNN according to claim 1, characterized in that, Step seven specifically involves: Select the mean absolute error (MAE), root mean square error (RMSE), and goodness of fit (R²). 2 As a core evaluation indicator, the specific calculation method and physical meaning are described below: (1) Mean Absolute Error (MAE) is used to reflect the actual average level of prediction error. This index is calculated as the average of the absolute values ​​of the difference between the predicted value and the actual value; its calculation formula is: ； (2) Root mean square error is used to measure the sample standard deviation of the deviation between the predicted value and the true value; since the error is squared during the calculation process, its calculation formula is: ； (3) Goodness of fit is used to characterize the model's ability to explain the variability of observed data. It is a dimensionless relative index for evaluating the model's fit. The value of this index is usually between 0 and 1. The closer the value is to 1, the better. Its calculation formula is: ； Where N represents the total number of test sample sequences; For the first The actual measured value (true value) of the engine sensor at any given time; For the first The predicted output value of the Online-DCNN model at time point; It is the arithmetic mean of the actual measurements in the test sample set.

Images