Neural Data Engine: Diesel Control & Optimization

• Neural Data Engine is a system that combines discretized physical dynamics with neural network models to capture diesel engine behavior under complex, multivariable and pollution constraints.
• It employs modular MISO submodules, such as speed, pressure, and opacity models, using techniques like Levenberg–Marquardt based training for real-time embedded applications.
• The system facilitates multivariable neuro-optimization, enabling precise trade-offs between engine performance and regulatory emission limits through adaptive control strategies.

The term "Neural Data Engine" refers to integrated computational systems that employ neural network models—often composed of multiple, interconnected submodules—both for modeling nonlinear system dynamics and for designing controllers that satisfy complex, multivariable criteria, including regulatory constraints. In the context of diesel engine control, neural data engines embody data-driven architectures built upon a backbone of physical equations, parameterized using empirical data, and capable of executing in real time for embedded industrial use. The following sections detail the principal attributes and technical methodology of neural data engines as manifested in the modeling and control of diesel engine systems with pollution constraints.

1. Structural Foundation: Dynamics-Informed Neural Modeling

Neural data engines are structured around the discretization of physical dynamical equations relevant to the target system. For turbocharged diesel engines, the foundational continuous-time relationships are:

• Engine speed: $\frac{dR(t)}{dt} = f_R(R(t), P(t), T(t))$
• Intake manifold pressure: $\frac{dP(t)}{dt} = f_P(R(t), P(t))$
• Exhaust gas opacity: $Op(t) = f_{op}(R(t), m(t), mf(t))$

These relationships are discretized and rewritten as nonlinear autoregressive maps, whose regressor structure is determined by selecting orders and delays corresponding to intrinsic system lags and transport phenomena. Each output variable (such as speed $R(k)$, pressure $P(k)$, opacity $Op(k)$) is predicted based on lagged sequences of itself, other state variables, and control inputs. This preserves interpretability and physical traceability within the neural architecture.

2. Modular Neural Submodel Composition

The overall model is constructed from interconnected MISO (multi-input, single-output) submodules, each realized as an output error (OE) neural model. The canonical submodules are:

• Speed Model ($NN_R$): $R(k) = NN_R(R(k-1), R(k-2), T(k-1))$
• Pressure Model ($NN_p$): $P(k) = NN_p(P(k-1), R(k-1))$
• Opacity Model ($NN_{OP}$): $Op(k) = NN_{OP}(Op(k-1), T(k-4), R(k-4), m(k-4))$

Each is implemented as a single hidden layer multilayer perceptron (MLP) with sigmoidal activation functions and a linear output neuron. The models are interconnected such that outputs from the speed module inform pressure estimation, and both speed and airflow, combined with specific delayed control signals, determine opacity—accounting for combustion and gas transport latencies.

3. System Identification and Training Procedures

Parameterization of the neural data engine leverages data collected from a real engine operating over representative regimes. The model structure—defined by regressor lags and neuron counts—is selected by minimizing the Final Prediction Error (FPE) criterion over a grid of candidate configurations, ensuring a parsimonious representation. Training employs the batch-mode Levenberg–Marquardt algorithm for nonlinear least-squares minimization, which is well-suited for MLPs and offers rapid convergence. Simulation mode training, in which the network feeds back its own predictions, is used to ensure recursivity and robustness.

4. Controller Design: Multivariable Neuro-Optimization

Control is achieved via an indirect approach, using the neural data engine as a surrogate system model within a specialized training loop. A neural controller (MLP, single hidden layer) receives as input the desired references and delayed/actual states of both speed and opacity and outputs the control signal for the next time step:

• $$U(k+1) = NNU(R_{ref}(k+1), R(k), R(k-1), Op_{ref}(k+d), Op(k+d-1))$$

The controller is trained to minimize a multivariable criterion:

• $$J(W) = \sum_{k} \left[ n_R (R_{ref}(k) - R(W, k))^2 + n_{op} (Op_{ref}(k+d-1) - Op(W, k+d-1))^2 \right]$$

The controller adapts weights using recursive Gauss–Newton updates, requiring Jacobian computations of system outputs with respect to controller input—a quantity approximated via the differentiable neural model.

5. Integration and Weighting of Pollution Constraints

Pollution constraints, specifically exhaust gas opacity, are integrated as explicit terms in both neural modeling and controller objectives. The opacity model incorporates gas transport delays and the dependence on injected fuel and airflow. During controller training, the opacity criterion weight ($n_{op}$) modulates the trade-off between speed tracking and pollution abatement. Simulations demonstrate that higher $n_{op}$ produces lower opacity peaks during transients (e.g., acceleration), albeit with increased speed tracking error—quantifying the control/constraint trade-off.

6. Advantages over Classical Approaches

Neural data engines provide several advantages:

• Universal Approximation: Capable of accurately modeling highly nonlinear and coupled thermodynamic, mechanical, and chemical phenomena with minimal prior analytical derivation.
• Black-Box Efficiency: Avoid reliance on explicit physical mapping and allow direct embedding of empirical data structure.
• Transient Performance: Outperform static mapping-based control strategies during non-steady-state regimes (e.g., load steps, accelerations).
• Parsimonious Model Complexity: Achieve requisite system accuracy with compact architectures due to rigorous structure selection.

7. Implications for Embedded Control and Regulation Compliance

The integration of neural data engines as embedded models in control systems supports direct management of performance and emissions, facilitating compliance with increasingly strict legislative standards. Their capacity for flexible multi-objective control and dynamic adaptation recommends them as alternatives to fixed-mapping controllers in future diesel engine systems. The multivariable neuro-optimization framework enables explicit trade-offs, empowering system designers to tune operational priorities per application requirements while maintaining regulatory alignment.

---

Neural data engines, as synthesized above, exemplify the intersection of empirical data-driven modeling, physically-informed neural parametric structures, and real-time embedded control optimization. Their adoption marks a paradigm shift in industrial control toward universal, adaptive, and constraint-aware neural computation platforms (0907.3342).