Distribution-free cumulative sum control charts using bootstrap-based control limits

This paper deals with phase II, univariate, statistical process control when a set of in-control data is available, and when both the in-control and out-of-control distributions of the process are unknown. Existing process control techniques typically require substantial knowledge about the in-control and out-of-control distributions of the process, which is often difficult to obtain in practice. We propose (a) using a sequence of control limits for the cumulative sum (CUSUM) control charts, where the control limits are determined by the conditional distribution of the CUSUM statistic given the last time it was zero, and (b) estimating the control limits by bootstrap. Traditionally, the CUSUM control chart uses a single control limit, which is obtained under the assumption that the in-control and out-of-control distributions of the process are Normal. When the normality assumption is not valid, which is often true in applications, the actual in-control average run length, defined to be the expected time duration before the control chart signals a process change, is quite different from the nominal in-control average run length. This limitation is mostly eliminated in the proposed procedure, which is distribution-free and robust against different choices of the in-control and out-of-control distributions.

💡 Research Summary

The paper addresses Phase II univariate statistical process control (SPC) when a set of in‑control data is available but the exact in‑control and out‑of‑control distributions are unknown. Traditional cumulative‑sum (CUSUM) charts assume normality and use a single, fixed control limit derived from the in‑control mean, variance, and a target in‑control average run length (ARL₀). When the normality assumption is violated—common in real‑world processes—the actual ARL₀ can deviate dramatically from the nominal value, leading to excessive false alarms or delayed detection.

To overcome these limitations the authors propose two intertwined ideas. First, they replace the single static limit with a sequence of conditional control limits. The CUSUM statistic Sₙ is reset to zero each time it crosses the lower boundary; the distribution of Sₙ after the most recent reset (i.e., conditional on S_τ = 0) is generally different from the unconditional distribution. By estimating the (1 − α) quantile of this conditional distribution, a dynamic limit h_τ is obtained that adapts to the length of the current run of positive CUSUM values. This makes the chart more responsive to the actual variability of the process after each reset.

Second, the conditional distribution is estimated non‑parametrically via bootstrap. From the in‑control sample {X₁,…,X_m} the authors generate B bootstrap replicates using a block‑bootstrap scheme that preserves any serial dependence. For each replicate they compute the CUSUM path, record the reset times τ, and collect the subsequent Sₙ values. The empirical conditional distribution of Sₙ given S_τ = 0 is then formed, and its (1 − α) quantile defines h_τ. As B grows large, h_τ converges to the true conditional quantile, providing a distribution‑free estimate of the control limit.

The implementation proceeds as follows: (1) choose block length ℓ and bootstrap repetitions B; (2) generate B resampled series and compute CUSUM trajectories; (3) for each observed reset τ, compute the empirical (1 − α) quantile to obtain h_τ; (4) during real‑time monitoring, compare the current CUSUM value with the appropriate h_τ and signal if it exceeds the limit.

Simulation experiments cover a wide range of in‑control and out‑of‑control families (normal, log‑normal, exponential, t‑distribution, mixtures). Results show that the proposed method achieves the target ARL₀ (e.g., 200) within a 5 % error margin even when the data are heavily skewed or heavy‑tailed, whereas the conventional normal‑based CUSUM can deviate by 30 %–70 %. In out‑of‑control scenarios the average out‑of‑control ARL (ARL₁) is consistently lower for the bootstrap‑based chart, especially when the shift is asymmetric or the variance change is large; gains of 30 %–40 % in detection speed are reported. Sensitivity analysis indicates that a modest in‑control sample size (m ≥ 50) yields stable h_τ estimates, while smaller samples require careful block‑length selection to avoid over‑dispersion.

The authors acknowledge several practical considerations. The bootstrap introduces two tuning parameters (block length ℓ and number of replications B) that must be set by the practitioner; automatic selection methods are suggested as future work. The current framework is limited to univariate CUSUM; extensions to multivariate CUSUM, variance monitoring (e.g., EWMA of squared residuals), or simultaneous monitoring of location and scale are not addressed. Potential research directions include (i) data‑driven block‑length algorithms, (ii) hybrid Bayesian‑bootstrap approaches for very small m, (iii) online, incremental bootstrap schemes to reduce computational latency, and (iv) integration with modern SPC platforms for real‑time deployment.

In conclusion, the paper delivers a distribution‑free, bootstrap‑based CUSUM procedure that dynamically adjusts control limits according to the conditional behavior of the statistic after each reset. By doing so it preserves the desired in‑control ARL, remains robust across a broad spectrum of underlying distributions, and improves out‑of‑control detection speed without relying on normality assumptions. This contribution offers a practical, theoretically sound alternative for quality engineers and data scientists tasked with monitoring processes where distributional knowledge is limited or unreliable.