// Tiny XGBoost (multi:softprob) predictor + exact TreeSHAP, for the tuned predictive-maintenance model. // Tree format: [classIndex, splitFeature[], splitThresholdOrLeaf[], left[], right[], cover[]] (function (root) { const f32 = Math.fround; function leaf(t, x) { const F = t[1], V = t[2], L = t[3], R = t[4]; let n = 0; while (L[n] !== -1) n = f32(x[F[n]]) < f32(V[n]) ? L[n] : R[n]; return V[n]; } function margins(model, x) { const m = [0, 0, 0, 0, 0].map(() => model.base); for (const t of model.trees) m[t[0]] += leaf(t, x); return m; } function softmax(m) { const mx = Math.max(...m), e = m.map((v) => Math.exp(v - mx)), s = e.reduce((a, b) => a + b, 0); return e.map((v) => v / s); } // ---- TreeSHAP (Lundberg et al. 2018, Algorithm 2) ---- function extend(p, ud, zf, of, fi) { p.f[ud] = fi; p.z[ud] = zf; p.o[ud] = of; p.w[ud] = ud === 0 ? 1 : 0; for (let i = ud - 1; i >= 0; i--) { p.w[i + 1] += (of * p.w[i] * (i + 1)) / (ud + 1); p.w[i] = (zf * p.w[i] * (ud - i)) / (ud + 1); } } function unwind(p, ud, k) { const of = p.o[k], zf = p.z[k]; let next = p.w[ud]; for (let i = ud - 1; i >= 0; i--) { if (of !== 0) { const tmp = p.w[i]; p.w[i] = (next * (ud + 1)) / ((i + 1) * of); next = tmp - (p.w[i] * zf * (ud - i)) / (ud + 1); } else p.w[i] = (p.w[i] * (ud + 1)) / (zf * (ud - i)); } for (let i = k; i < ud; i++) { p.f[i] = p.f[i + 1]; p.z[i] = p.z[i + 1]; p.o[i] = p.o[i + 1]; } } function unwoundSum(p, ud, k) { const of = p.o[k], zf = p.z[k]; let next = p.w[ud], total = 0; for (let i = ud - 1; i >= 0; i--) { if (of !== 0) { const tmp = (next * (ud + 1)) / ((i + 1) * of); total += tmp; next = p.w[i] - (tmp * zf * (ud - i)) / (ud + 1); } else total += p.w[i] / zf / ((ud - i) / (ud + 1)); } return total; } function copy(p, n) { return { f: p.f.slice(0, n), z: p.z.slice(0, n), o: p.o.slice(0, n), w: p.w.slice(0, n) }; } function recurse(t, x, phi, node, parent, ud, pz, po, pf) { const F = t[1], V = t[2], L = t[3], R = t[4], C = t[5]; const p = copy(parent, ud); extend(p, ud, pz, po, pf); if (L[node] === -1) { for (let i = 1; i <= ud; i++) phi[p.f[i]] += unwoundSum(p, ud, i) * (p.o[i] - p.z[i]) * V[node]; return; } const sf = F[node]; const hot = f32(x[sf]) < f32(V[node]) ? L[node] : R[node]; const cold = hot === L[node] ? R[node] : L[node]; let iz = 1, io = 1, k = 0; while (k <= ud && p.f[k] !== sf) k++; if (k !== ud + 1) { iz = p.z[k]; io = p.o[k]; unwind(p, ud, k); ud--; } recurse(t, x, phi, hot, p, ud + 1, (C[hot] / C[node]) * iz, io, sf); recurse(t, x, phi, cold, p, ud + 1, (C[cold] / C[node]) * iz, 0, sf); } // phi[class][feature] in log-odds (margin) space function shap(model, x) { const nf = model.features.length; const phi = [0, 1, 2, 3, 4].map(() => new Array(nf + 1).fill(0)); const empty = { f: [], z: [], o: [], w: [] }; for (const t of model.trees) recurse(t, x, phi[t[0]], 0, empty, 0, 1, 1, -1); return phi.map((r) => r.slice(0, nf)); } function features(type, air, proc, rpm, torque, wear) { const enc = { H: 0, L: 1, M: 2 }[type]; return [enc, air, proc, rpm, torque, wear, (torque * rpm * 2 * Math.PI) / 60, proc - air]; } root.PM = { margins, softmax, shap, features }; })(typeof window !== 'undefined' ? window : globalThis);