A new industrial quality control methodology combines supervised machine learning with phase-segmented Shapley Additive Explanations to deliver transparent, automated monitoring of safety-critical crimp connections. Developed by engineering researchers Bernd Hofmann, Patrick Bründl, Huong Giang Nguyen, and Jörg Franke, the data-driven framework resolves the long-standing trade-off between the high diagnostic accuracy of algorithmic classifiers and the interpretability required on factory production floors. In rigorous validation trials conducted on physical crimp force curves, the system achieved a multi-class fault classification accuracy of 95.9 percent while accurately pinpointing the exact mechanical deformation stages responsible for detected defects. By mapping post-hoc game-theoretic attribution scores directly onto four established physical phases of the crimping cycle, the platform allows manufacturing operators to inspect intuitive, color-graded visual indicators rather than opaque numeric predictions, ensuring that automated defect detection satisfies the strict traceability demands of serial production.
Crimping represents an indispensable, solderless joining technique extensively utilized across modern manufacturing sectors, ranging from electrical control cabinet assembly to complex automotive wiring harnesses. The procedure establishes an electrically conductive and mechanically stable bond by permanently deforming a metallic terminal casing around an inserted stranded wire conductor. Because these joints frequently transmit vital electrical power or high-speed communication signals in safety-critical environments, loose strands, severed cores, or trapped cable sheaths can precipitate disastrous functional failures. Traditional in-line verification relies on commercial crimp force monitoring instruments that track the mechanical force exerted by the press ram throughout its downward working stroke. To separate acceptable parts from defective units, these conventional instruments compare the real-time force curve against a prerecorded reference profile using static tolerance corridors or integrated curve areas established through preliminary destructive testing.
Despite their widespread industrial adoption, conventional threshold-based crimp force monitoring systems exhibit severe operational limitations that reduce manufacturing efficiency. Because standard commercial systems rely heavily on manual feature selection, such as peak force measurements or coarse regional integrals under the curve, they frequently fail to detect subtle, high-dimensional anomalies buried within complex signal trajectories. Furthermore, establishing reliable tolerance boundaries demands extensive trial-and-error sampling and deep domain expertise, leaving setups vulnerable to material fluctuations, ambient environmental variations, and human setup inconsistencies. When batches change, manufacturers must engage in time-consuming and expensive recalibration procedures to avoid false alarms. Although deterministic monitoring affords direct visual feedback by showing where a curve violates a tolerance envelope, its inability to adapt dynamically to operational noise severely constrains continuous quality assurance in modern high-throughput manufacturing facilities.
To overcome the rigidity of manual feature engineering, researchers have increasingly deployed data-driven artificial intelligence to uncover subtle, multidimensional patterns that elude human visual inspection. Supervised learning algorithms, including deep convolutional networks and ensemble classifiers, have repeatedly demonstrated superior fault detection rates in mechanical assembly tasks. However, these complex architectures operate predominantly as black-box models, obscuring the mathematical rationale governing their final predictions. In safety-critical sectors such as automotive engineering and aerospace manufacturing, this missing transparency prevents plant engineers from conducting root cause analyses and impedes regulatory validation. When an uninterpretable model produces a false alarm or misclassifies a defective component, factory technicians cannot deduce the underlying physical malfunction to implement rapid countermeasures, conflicting directly with foundational manufacturing philosophies such as Jidoka that require immediate comprehension of production anomalies.
To establish a fully transparent yet highly capable monitoring architecture, the investigative team gathered and curated an empirical experimental dataset encompassing 1,617 individual crimping operations. Data collection was performed using a semi-automatic benchtop crimping press equipped with an automated terminal feed and a piezoelectric force sensor positioned directly upon the press ram connection rod. Two distinct copper conductor configurations were evaluated: a 16-strand conductor with a cross-sectional area of 0.50 square millimeters, accounting for 845 recorded operations, and a 12-strand conductor with a cross-sectional area of 0.35 square millimeters, contributing 772 curves. Alongside standard acceptable joints, the researchers systematically induced two prevalent industrial defect categories by manipulating wire preparation: missing strand defects, simulating incorrect stripping blade depths or cable reel crushing, and crimped insulation faults, simulating misaligned axial stop positions.
The data preparation pipeline implemented a raw-sensing methodology designed to preserve the complete morphological dynamics of the crimping stroke without discarding information through arbitrary feature extraction. Raw sensor readouts, initially comprising 3,567 discrete sampling points per stroke, were mathematically inverted to compensate for the upside-down physical mounting of the load cell and subsequently zeroed to ensure baseline cross-machine comparability. The continuous signals were then trimmed to an exact region of interest spanning 500 sequential data points, bounded strictly by the initial contact of the crimping stamp against the terminal flanks and its final detachment. Following min-max normalization into a uniform numerical range using standard software libraries, the curated dataset was partitioned, dedicating 80 percent, or 1,293 curves, to model training and hyperparameter optimization, while reserving the remaining 20 percent, consisting of 324 unseen curves, for unbiased performance evaluation.
For the core classification engine, the researchers deployed a Random Forest algorithm implemented via scikit-learn, selecting the ensemble technique for its proven resistance to overfitting and robust performance across tabular time-series representations. Hyperparameters were systematically optimized through grid search combined with integrated five-fold cross-validation over the training partition, ultimately establishing an optimal configuration of 150 decision trees operating with a maximum tree depth of 20 splits. Across the five cross-validation training iterations, the classifier demonstrated stable learning trajectories, achieving fold accuracy scores of 0.934, 0.953, 0.957, 0.937, and 0.976. When deployed against the independent test set of 324 force curves, the optimized model achieved an overall classification accuracy of 0.959, alongside a macro-averaged precision of 0.958, a macro-averaged recall of 0.952, and a macro-averaged F1-score of 0.955.
Detailed examination of the test confusion matrix confirmed that the ensemble classifier misclassified only 4 percent of all evaluated components, proving highly robust for serial manufacturing deployment. Out of 161 acceptable joints in the test cohort, only six were erroneously labeled as defective false negatives, an industrial scenario that incurs economic scrap costs without compromising end-user safety. Crucially, only four out of 163 genuinely defective joints were misclassified as acceptable false positives. Physical inspection revealed that all four missed anomalies stemmed from marginal, single-strand manipulations so minor that current commercial monitoring systems likewise failed to register them. Additionally, only three defective instances experienced cross-classification errors between the two fault types, confirming that the trained algorithm reliably distinguishes between healthy production and distinct defect signatures across variable wire cross-sections.
To demystify the internal decision logic of the Random Forest classifier, the researchers integrated the TreeExplainer algorithm from the Shapley Additive Explanations framework. Rooted in cooperative game theory, this post-hoc explainer computes local Shapley values that quantify the positive or negative contribution of each individual input point toward the final quality prediction. Because individual discrete time-series points lack standalone physical meaning for human machine operators, the team grouped the 500 points into four physically grounded process phases: centring, spanning points 0 to 75; rolling-in, spanning points 76 to 150; compression, spanning points 151 to 345; and springback, spanning points 346 to 500. By calculating the mean Shapley value across each discrete segment, the system converts raw mathematical attributions into actionable phase importance metrics that correspond directly to mechanical tool interactions.
To communicate these diagnostic metrics effectively on the factory floor, the architecture translates the normalized phase importance values into an intuitive, color-coded visual interface. Rather than presenting abstract statistical bar plots that disconnect importance weights from temporal sequences, the software renders the processed crimp force curve enclosed within an overlaid, segmented graphic pipe. The local importance scores determine a normalized color gradient that shades each pipe segment, immediately drawing operator attention toward the specific temporal regions that dictated the classifier’s verdict. Positioned above the line plot, a concise header displays the predicted multi-class outcome. This human-centric visualization bridges the gap between algorithmic calculations and operator understanding, empowering shop-floor technicians to diagnose emerging mechanical faults without requiring specialized data science training.
The physical relevance of the generated explanations was rigorously validated through a comprehensive quantitative perturbation study based on feature selectivity metrics. The underlying principle dictated that if an identified curve phase is genuinely vital to classification, corrupting its data points must trigger a substantial drop in model accuracy. The investigators trained and evaluated 42 distinct Random Forest models across 14 combinatorial phase modifications, testing three distinct data corruption strategies: replacing target values with zeros, substituting uniform random noise, or excising the segments entirely. While manipulating non-critical sections such as phase one or phase four caused negligible performance degradation, corrupting phase three caused test accuracy to plummet from 0.959 down to 0.861 across all substitution strategies. Furthermore, when three phases were corrupted simultaneously, models retaining solely the original phase three data maintained accuracies between 0.904 and 0.917, proving that the compression phase provides the primary mathematical foundation for automated fault identification.
Complementing the algorithmic sensitivity benchmarks, an empirical qualitative assessment was conducted with independent manufacturing process experts who scored the physical significance of each curve phase on a scale from zero to three. For missing strand defects, both human specialists identified phase three as the dominant diagnostic zone because fewer wire strands reduce total compaction resistance, which closely mirrored the model’s visual emphasis on the compression phase supported by phase two. For crimped insulation defects, the experts noted that foreign plastic sheath material introduces premature resistance during rolling-in before softening under compaction pressure, creating significant deviations across phases two and three. The visual output generated by the explainable system mirrored these domain dynamics, demonstrating strong alignment with expert engineering observations while uncovering subtle phase-level dependencies that human observers might otherwise overlook.
Despite its proven diagnostic efficacy, the researchers identified specific structural constraints that define opportunities for future technological refinement. Notably, the four standardized physical phases encompass an unequal distribution of raw sensor readings, with the compression and springback segments containing considerably more sampling points than the initial centring stage, an asymmetry that could modestly skew perturbation analyses. Future investigations will evaluate whether dynamic, data-driven segmentation algorithms can enhance feature balance without compromising the intuitive physical meaning of the curve intervals. Additionally, the research team aims to transfer this transparent monitoring paradigm beyond wire termination to other safety-critical industrial joining processes that generate continuous time-series data, including automated threaded fastening, structural press-fitting, and resistance spot welding operations.
Subject of Research: Explainable machine learning for industrial crimp force monitoring and fault detection
Article Title: Transparent machine learning for crimp force monitoring using phase-based SHAP explanations
Article References: Hofmann, B., Bründl, P., Nguyen, H. G., & Franke, J. (2026). Transparent machine learning for crimp force monitoring using phase-based SHAP explanations. Industrial Artificial Intelligence, 4(1), Article 2. https://doi.org/10.1007/s44244-026-00028-6
Image Credits: AI Generated
DOI: 10.1007/s44244-026-00028-6
Keywords: Crimp Force Monitoring, Explainable Artificial Intelligence, Shapley Additive Explanations, Random Forest Classifier, Industrial Quality Control, Wiring Harness Manufacturing, Time Series Classification, Manufacturing Process Monitoring, Mechanical Joining, Fault Detection, Perturbation Analysis, Feature Selectivity, Industrial Automation, Solderless Connections, Multi-Class Classification, TreeExplainer, In-Line Quality Assurance
Cite Scienmag News
Blake Davidson. (November 17, 2026). Phase-Based Explainable Machine Learning Identifies Industrial Crimping Faults With High Accuracy. Scienmag. https://scienmag.com/phase-based-explainable-machine-learning-identifies-industrial-crimping-faults-with-high-accuracy/
Blake Davidson. "Phase-Based Explainable Machine Learning Identifies Industrial Crimping Faults With High Accuracy." Scienmag, 17 November 2026, https://scienmag.com/phase-based-explainable-machine-learning-identifies-industrial-crimping-faults-with-high-accuracy/. Accessed 18 November 2026.
Blake Davidson. "Phase-Based Explainable Machine Learning Identifies Industrial Crimping Faults With High Accuracy." Scienmag. November 17, 2026. https://scienmag.com/phase-based-explainable-machine-learning-identifies-industrial-crimping-faults-with-high-accuracy/








