A heavily cycled friction clutch or brake is a machine whose internal state nobody knows. After a series of engagements the drum surface is hot by an unknown amount, the slip speed between plates may be unknown, and the thin surface layer where cracks start cannot carry a sensor. Operators handle this with conservative rules, such as "wait long enough" or "never exceed a fixed ratio of absorbed energy to sliding time." This proposal asks whether we can compute the fastest open-loop command sequence that drives the clutch from any state in a bounded uncertainty set to one known state, while proving that no admissible starting state ever pushes the surface past its cracking temperature along the way. Open-loop here means the commands are fixed in advance and use no sensor feedback.
Automata theory supplies the concept. A synchronizing word (or reset word) is an input sequence that sends every state of a machine to the same state. We keep that idea and add two physical ingredients. Clutch lockup acts as a genuine discrete merge, because once the plates stick, the prior slip speed no longer matters. Heat conduction acts as a contraction, because two different temperature fields under the same inputs grow closer over time. The difficulty is that the fastest merge, a hard engagement, is also the one that deposits the most heat. The result would be a planner that trades these effects against each other with exact worst-case safety. It would also produce a minimized finite supervisor small enough to run on an embedded controller.
Core Claim
Consider an engagement system with a finite mechanical mode automaton, whose states are slip-speed bins, slipping versus locked, and lubrication regime, and a lumped linear thermal network whose heat inputs depend on mode and action. For this class, a thermally constrained reset sequence can be computed and certified exactly. The certificate avoids the intractability that general weighted-automaton safety carries, and it relies on two structural facts.
The first fact concerns the thermal state within a fixed mode $q$. The vector of node temperatures $T_k$ evolves under action $a$ as
$$T_{k+1} = A_a T_k + b_{q,a} + d_k,$$
where $A_a$ is the discretized conduction and cooling matrix, $b_{q,a}$ is the frictional heat input, and $d_k$ is a bounded disturbance covering friction-coefficient fade. Affine maps send polytopes to polytopes. The peak surface temperature is a linear function of $T$, so its maximum over the uncertainty set is attained at a vertex. Propagating vertices therefore gives an exact worst-case safety check rather than a sampled one, up to the disturbance over-approximation.
The second fact concerns the spread of the uncertainty set. The offsets cancel in differences, so for two starting states in the same mode, $e_{k+1} = A_a e_k$. The spread is governed only by products of the linear parts. This is the continuous analogue of the pair digraph, the graph of state pairs whose collapse governs synchronization in On the Interplay Between Babai and Černý's Conjectures. It also matches the matrix-product view of reset words in Primitive Sets of Nonnegative Matrices and Synchronizing Automata, where synchronization means reaching a rank-one product. Conduction never collapses the thermal spread exactly, so we define synchronization as ε-synchronization: the diameter of the belief set falls below a tolerance ε. Mechanical lockup, by contrast, collapses the discrete part exactly. What is new is a reset problem whose merges have a thermal cost, solved by search over beliefs, where a belief is a set of modes each carrying a thermal polytope, with exact vertex-based safety.
Relation to Prior Work
Several papers motivate the problem. The marine clutch study Experimental Analysis on High-Rated High-Elastic Clutch Coupling found that drum surface cracking under repeated engagement is avoided by respecting an empirical ratio of absorbed energy to sliding duration. It gives no formal model and no sequence synthesis, and the file reports no numerical value for that ratio. Our method replaces the single ratio with a certified plan that depends on the uncertainty set. The tribometer study Tribological Properties of Cylinder Block/Valve Plate Interface of Axial Piston Pump on the Tribometer found an abrupt breakdown of the oil film at 2.5 MPa, where heating and adhesive wear jump. That is exactly the kind of discontinuity we model as a guarded mode switch rather than folding it into smooth dynamics.
Other papers constrain the design. Synchronizing Words for Weighted and Timed Automata shows that synchronization of deterministic weighted automata under general upper and lower safety bounds is 3-EXPSPACE-complete, and that synchronization of nondeterministic timed automata is undecidable. Treating temperature as a generic accumulated weight is therefore a dead end. Our tractability comes from exploiting the linear, dissipative structure of conduction, which generic weights lack. Synchronizing Strategies under Partial Observability shows that partial observability raises complexity sharply for nondeterministic systems but not for deterministic ones. This supports keeping the mechanical automaton deterministic and pushing uncertainty into bounded disturbances.
A third group supplies methods. Regular-language constraints on reset words from Computational Complexity of Synchronization under Regular Constraints let us forbid actuator patterns, such as re-pressurizing within a minimum interval. Meeting a Powertrain Verification Challenge verified air-fuel invariants on a nonlinear hybrid automaton, meaning a model mixing discrete modes with continuous dynamics, using on-the-fly discrepancy bounds. It contains no friction heating, cracking thresholds, or reset synthesis. We use it as the general-purpose reachability comparator rather than as our engine. Finally, the V-belt model in Advanced Calculation Model of V-Belt Temperature Field reports thermal prediction errors of 0.15 ℃ in steady state versus 4.77 ℃ in transients. We use this as a concrete precedent for sizing the disturbance term $d_k$: our safety margin must absorb transient model error of that order, not the steady-state figure.
Research Program
The work has three parts.
The first part is modeling. We build a reduced thermal network of about four to twelve nodes for a double-cone friction clutch: drum surface, drum bulk, lining, and housing. The network comes from a finite-element model by standard model-order reduction. We couple it to a mechanical automaton over slip-speed bins, with actions drawn from a small set of clamping-pressure levels and a dwell action. Lockup is a merge transition. The heat input $b_{q,a}$ equals friction coefficient times clamping force times slip speed for that bin, with a friction-coefficient interval covering fade. Mode guards encode lubrication breakdown by pressure, in the spirit of the 2.5 MPa boundary.
The second part is synthesis. We run best-first search over beliefs, where each belief pairs a set of possible mechanical modes with one thermal polytope per mode. Safety is checked at every vertex. We prune a belief when another belief is contained in it and reachable at no greater cost. We also use a lower bound for search, taken from the slowest contraction rate among available actions. To keep vertex counts manageable, we over-approximate with zonotopes, polytopes built from a sum of line segments. We measure how much conservatism this adds against the exact vertex computation on small instances.
The third part is minimization. We discretize belief space into cells labeled "safe-reachable-to-target within budget." We then run partition refinement, Hopcroft-style state minimization, to merge cells that behave the same with respect to those labels. The result is the smallest lookup supervisor, and we report its size against the unminimized abstraction.
Baselines
The naive reference point is dwell-to-ambient. It issues no engagement until the worst-case surface temperature has cooled to within ε of ambient, then performs a single worst-case-safe engagement. It trivially synchronizes and is trivially safe. We compute it exactly on the same model with the same ε, so the only difference is the planner. Beating it means a shorter certified time-to-known-state at equal worst-case temperature margin.
The second baseline is the empirical energy/time ratio rule of Experimental Analysis on High-Rated High-Elastic Clutch Coupling, fitted to our model by choosing the largest ratio that never violates the cracking threshold on the uncertainty set. This represents current engineering practice. The file reports no numerical value for the rule, so we recalibrate it rather than quote it.
The strongest competitor is robust open-loop model predictive control with scenario sampling. This is a mixed-integer optimizer over the same action set and horizon that enforces the temperature constraint on N sampled initial states. The file reports no result on Lacuna for thermally constrained reset or clutch-engagement planning, so there is no published number to beat. It gets the same model, the same horizon, and a matched wall-clock budget, with N tuned as high as that budget allows. Beating it means either shorter certified reset time or, more likely, zero worst-case violations where scenario sampling admits vertex violations when checked afterward.
We also run a C2E2-style reachability check following Meeting a Powertrain Verification Challenge as a verifier for sequences produced by the other methods. That paper reports runtimes of 10–11 minutes for complex properties on a quad-core i7 with 8 GB of RAM, on a different benchmark. We report our certification time on comparable hardware for context only.
The key ablation removes the lockup merge, allowing thermal contraction only. This isolates whether coupling discrete merges with thermal cost is what shortens resets.
Why not just…?
Why not just wait for the clutch to cool? Cooling is a perfectly valid reset, and on many duty cycles it will be close to optimal. The proposal is only worth doing if the planner beats dwell-to-ambient substantially in regimes that matter: emergency re-engagement such as crash-astern, where waiting is not allowed, or targets that are warm operating states rather than ambient. The first experiment measures exactly this gap across uncertainty-set sizes. If the gap is small everywhere, we report that and stop.
Why not just use a temperature observer with feedback control? This is the real threat. If back-of-drum thermocouples plus a thermal model can estimate surface temperature well, closed-loop control beats any open-loop plan. The file's own evidence, the 4.77 ℃ transient error in Advanced Calculation Model of V-Belt Temperature Field, suggests estimates degrade exactly in transients. We will not assume this settles the question. We include a Kalman-observer feedback controller with a realistic subsurface sensor and measure its worst-case surface overshoot. We claim value only where observer error forces margins larger than the open-loop certificate needs. A hybrid scheme, where the reset word shrinks uncertainty before feedback takes over, is a legitimate outcome.
Why not just encode temperature as a weight and use existing weighted or timed automata tools? Synchronizing Words for Weighted and Timed Automata places general safety-bounded weighted synchronization at 3-EXPSPACE. A scalar weight also cannot represent a multi-node temperature field that relaxes between actions. Our encoding should scale where a scalar-weight encoding either fails to capture conduction or blows up, and the scaling experiment tests this.
Why not just sample finite-element simulations? Sampling finds violations but cannot rule them out. Because safety over an affine image of a polytope is decided at its vertices, exhaustive vertex checking is both cheaper and complete. The comparison against scenario-sampled control quantifies how often sampling misses the worst case.
Proposed Experiments
The first experiment is a synthetic scaling suite. We generate random reduced thermal networks, from 4 to 12 nodes, with dissipative matrices, and random deterministic mode automata with 3 to 20 states that include at least one lockup merge. We record certified reset length, computation time, supervisor size before and after minimization, and the conservatism of zonotope over-approximation. We also compare reset lengths with the quadratic growth familiar from Slowly synchronizing automata and digraphs, to see whether thermal cost changes how reset length scales with mode count.
The second experiment is a clutch case study. We reduce a finite-element model of a double-cone clutch and fit parameters to measured heating on a small inertial rig in the spirit of the 500 ps prototype. The cracking threshold is replaced by a conservative surface-temperature limit set from the drum material. We compare all baselines on time-to-known-state, worst-case margin, and violations detected by exhaustive vertex audit.
The third experiment is physical validation on a ring-on-disk tribometer of the MMU-10H type. We take a subsurface thermocouple reading as ground truth and randomize the prior thermal history. We then execute the synthesized sequences and check that measured peak temperature stays within the certified bound and that final states cluster within ε.
The decisive checks are two. First, the planner must reset at least noticeably faster than dwell-to-ambient on emergency scenarios at zero certified violations. Second, scenario-sampled control must show vertex violations that our method does not. If both fail, the automata formulation adds nothing over cooling and robust control, and we will report that.
References
- On the Interplay Between Babai and Černý's Conjectures
- Primitive Sets of Nonnegative Matrices and Synchronizing Automata
- Experimental Analysis on High-Rated High-Elastic Clutch Coupling
- Tribological Properties of Cylinder Block/Valve Plate Interface of Axial Piston Pump on the Tribometer
- Synchronizing Words for Weighted and Timed Automata
- Synchronizing Strategies under Partial Observability
- Computational Complexity of Synchronization under Regular Constraints
- Meeting a Powertrain Verification Challenge
- Advanced Calculation Model of V-Belt Temperature Field
- Slowly Synchronizing Automata and Digraphs