ControlGen

ControlGen is a mechanistic dataset generator for structure-to-time-series tasks in control systems. The current version focuses on realistic SISO control chains built from a modular control graph: reference, error summing junction, controller, actuator, plant, disturbance, sensor, and measurement noise.

What is implemented

• Modular graph DSL with loop, reference, sum, controller, actuator, plant, disturbance, sensor, noise, and tap
• Low-level dynamic block support for tf, gain, pid, delay, ss, series, parallel, and feedback
• Parameterized module library for controller, actuator, plant, disturbance path, and sensor dynamics
• Realistic scenario sampling for reference signals, load disturbances, and measurement noise
• Multi-signal simulation producing r, d, n, e, u_cmd, u_act, y, and y_m
• Dataset schema with graph_dsl, module_params, scenario, signals, state-space module views, and mechanism-oriented metrics

Quick start

python3 -m pip install -e .[dev]
python3 -m controlgen.cli --count 3 --seed 42
pytest

Sample Graph DSL

loop(
  reference=reference(kind='multistep', name='reference'),
  sum=sum(signs=[1, -1], name='error_sum'),
  controller=controller(kind='pid', name='controller'),
  actuator=actuator(kind='lag_saturation', name='actuator'),
  plant=plant(kind='delay_plus_lag', name='plant'),
  disturbance=disturbance(kind='load_step', injection='output', name='disturbance'),
  sensor=sensor(kind='lag', name='sensor'),
  noise=noise(kind='white_noise', name='measurement_noise'),
  taps=[tap(signal='r', name='tap_r'), tap(signal='y', name='tap_y')]
)

Output shape

Each sample contains:

• graph_dsl: serialized modular control graph
• graph_ast: JSON-friendly graph tree
• module_params: instantiated controller, actuator, plant, disturbance path, and sensor parameters
• scenario: reference, disturbance, noise, and simulation settings
• t: time axis
• signals: r, d, n, e, u_cmd, u_act, y, y_m
• system_view: per-module linear state-space cores and nonlinear wrappers such as actuator saturation
• metrics: tracking, control effort, disturbance rejection, and measurement-chain metrics
• tags, split_tags: structure family, controller family, plant family, disturbance family, and reference family

Notes

• The current release targets mechanistic SISO control loops.
• The linear core is simulated through discretized state-space blocks; actuator saturation is handled in the time-domain loop.
• Delays use a Padé approximation inside low-level dynamic modules.