AI explanations for crimp force monitoring: evidence for accuracy and understanding

A “missing strands” alarm appears, and the screen highlights part of the crimp force curve in red. This can help an engineer decide where to start checking. It has not yet answered two questions: whether the product really has missing strands, and whether the red region helps someone take the right action.

This review is for quality and process engineers evaluating AI crimp force monitoring. Using a public study by Hofmann and colleagues, we read classification results, model explanations and effects on users separately. We identify the denominators and comparison conditions that can be used directly in a technical review. The paper was published on March 31, 2026; its version of record is dated April 10. This is a focused review of an existing study. Original paper

Separate the three things to evaluate

Connections between these three can be established step by step, but none substitutes for another. Our recommendation for a review is to examine the direction of errors first, then how the explanation was tested, and finally which practical tasks users completed. The following sections work through the paper’s available evidence in that order.

From one curve to four colored phases: what the model does

The study uses a random forest to classify preprocessed crimp force curves into three categories: normal, missing strands and insulation crimped into the conductor crimp. A random forest consists of multiple decision trees. Each tree makes a judgment from the input features, and these judgments are combined into the model output. In this study, the inputs are consecutive sample points along the curve, rather than a single peak selected by a person. SHAP’s TreeExplainer then calculates each sample point’s contribution to the output for a particular class. Paper, Section 3; Official random forest documentation

The processing chain can be read as follows:

1. Acquisition and preprocessing→
2. Random forest outputs a class→
3. TreeSHAP explains the output for that class→
4. Contributions are aggregated by process phase→
5. Colors are displayed on the original curve

This is a schematic of the relationships within the method. Classification happens before the colors are displayed; the colors do not add an independent quality inspection.

Existing crimp force monitoring should not be treated as universally limited to comparing a single peak. The manufacturer’s BB07i4 description cited in the paper includes comparison with reference curves, drift handling and multiple reference settings. The contribution examined here connects learned classification with phase explanations. The paper does not provide a common benchmark sufficient to quantify “better than all conventional monitoring.” Manufacturer’s BB07i4 description

The paper divides the curve into four phases: alignment, rolling, compression and springback. These correspond to the terminal wings beginning to contact and align with the crimper, the wings rolling into the conductor, compaction of the conductor and terminal, and elastic recovery after the crimper retracts. The study uses boundary indices of 75, 150, 345 and 500. These are positions within this data-processing method. They cannot be used directly as times, travel distances or universal segmentation parameters for another machine. Paper, Sections 2.1 and 3.5, Equation 4

Phase aggregation contains an easily overlooked detail: Equation 6 in the paper takes the mean of the SHAP values within each phase, then normalizes the result to form the color bands. It does not turn each of the four phases into a “failure probability,” nor does it simply sum all absolute contributions within each phase.

Consider a purely arithmetic example. Two points in the same phase have contributions of +0.02 and -0.02. Their signed mean is 0; the mean of their absolute values is 0.02. The former preserves the relationship between contributions that push and cancel, while the latter describes contribution magnitude. These two aggregations answer different questions. These are not measurements from the paper. When reading or reproducing a color map, retain the class being explained, the signs, the aggregation method and the color normalization rules together. Otherwise, the meaning of “the reddest region” changes.

TreeSHAP explains how the model output differs from its baseline value. It also requires a choice about how to handle dependent input features. Consecutive curve points are correlated, so the software version, background data and feature-dependence options are also conditions of the explanation. Saving a color map alone does not make the same color an invariant indicator across versions. Official TreeExplainer documentation

Which errors make up the 95.9% result?

The study recorded 1,617 curves from two wire–terminal pairings involving 0.50 mm² and 0.35 mm² wires. The researchers deliberately prepared the missing-strand and crimped-in-insulation samples. They randomly split the data into 1,293 training samples and 324 test samples. This is not a set of shipment records sampled at the defect prevalence of a real production line. Paper, Section 3.2

The authors report a three-class accuracy of 95.9%. Correctness here is relative to the authors’ labels. A label for a deliberately prepared defect must still be distinguished from nonconformity determined against product requirements. Start by separating the directions of error:

The percentages in the table are descriptive values calculated from the counts in the paper and rounded. Their denominators are the number of samples labeled normal and the number labeled defective, respectively. They are neither the overall error rate nor the defect proportion among samples the model classifies as normal. A confusion matrix should clearly identify which axis represents reference labels and which represents predicted classes. Otherwise, the same cell can easily be read as answering a different question. Paper, Section 4.1; Confusion matrix definition

There is also a small discrepancy between the reported value and the counts given in the text. The counts total 6 + 4 + 3 = 13 misclassifications. Recalculating with a total of 324 gives a three-class accuracy of approximately 96.0%, which does not exactly match the authors’ 95.9%. We therefore retain the authors’ reported value and list the error counts and denominators separately, without claiming they produce an exactly matching recalculation.

For a quality review, the more important distinction is between confusing the two defect classes and classifying a defective sample as normal. The former affects the direction of diagnosis; the latter concerns the model’s ability to detect defects. If the task is reduced to two classes, normal and defective, the 3 errors between defect classes no longer count as errors, and the metric changes. This also shows why accuracies from different classification tasks cannot be compared directly.

What the perturbation tests support

To examine whether the phase explanations were useful, the researchers set selected phases to zero, replaced them with random values or removed them, then retrained the models using the same training method. Combinations of one, two and three phases produced 42 perturbed models, which were compared with the unperturbed baseline. This compares what a model retrained without a particular region’s information can still achieve. It is not simply a one-off masking of the original model’s input. Paper, Sections 3.7 and 4.2

In Table 4, perturbing the compression phase alone produces an accuracy of 0.861 with each of the three treatments, below the authors’ reported baseline of 0.959. Perturbing the rolling or springback phase alone gives corresponding results of 0.954 in every case. This supports a specific finding: with these data and this comparison method, the compression phase contains classification information that the other regions have greater difficulty compensating for.

This test does not directly establish that the red region is the physical cause of the defect. The loss of performance after retraining is affected by how much information is removed, whether other phases can compensate, the model and the data-splitting method. The paper also acknowledges that the phases contain different numbers of sample points. Removing a longer phase may cause a greater loss; the size of that loss alone does not establish that every point is more important.

If the next step is to compare two explanation methods, a more targeted additional test can be proposed. Let each select the same number of points or intervals of the same length, then compare them using the same replacement method, training settings and test set. This suggestion aims to reduce the confounding effect of “removing more.” It is not a result already obtained in the paper.

Engineers agreeing with a phase explanation does not establish better task performance

For the human-factors part of the paper, two process experts assessed the influence of each phase on curves whose quality classes were already identified. Their assessments were then compared qualitatively with the visualizations. This can check whether the explanations broadly align with professional knowledge. It does not provide blind-test accuracy for operators across multiple shifts, alarm response times or changes in incorrect actions before and after adding the color bands. Paper, Section 4.3

The two questions can be tested separately:

• To check whether a phase explanation is plausible, ask engineers which region of the curve a particular defect is expected to affect, then compare that with the region highlighted by the model.
• To check whether the interface helps people respond, ask users to handle records for which they do not know the answers. Observe their judgments when viewing the original curve, then a class indication, then the phase explanation. Recording “unable to decide” is meaningful too; users should not be forced to give a cause every time.

The latter comparison should ideally hold sample difficulty and available information constant and account for learning effects caused by presentation order. Otherwise, a faster second response may simply mean that the person has already seen the same curve. Retain explanation displays for missed or misclassified samples as well. A clear explanation may still be explaining an incorrect prediction.

What is available now, and what is missing for transfer?

The paper provides open full text and links to the Crimp Force Curve Dataset on Harvard Dataverse and a Scientific Data paper describing acquisition, labeling and technical validation. The current dataset release is V3.0. The full dataset contains 2,439 curves, which differs from the subset of 1,617 used in this study. The data retain both manual labels and CFM decisions, including records where they disagree. Before evaluation, specify which reference labels will be used and determine actual conformity with methods applicable to the project. These public materials support offline evaluation of methods; a dataset access point is not validated online monitoring software. Dataset DOI; Data description paper

The authors also provide a public cfc-analysis code repository. The link here is fixed to the August 22, 2025 commit cf20535, which contains data preprocessing, validation and plotting files. It offers a way to understand the data. It does not establish that the complete random forest, TreeSHAP and interface code for this 2026 study has been released.

For this review, we did not download and run the full dataset, or obtain and run the complete training, splitting and visualization code package for the SHAP study. Claiming exact reproduction would still require checking the following together:

• the data version and sample IDs actually used;
• preprocessing and scaling ranges;
• training/test assignments and random seeds;
• dependency versions and SHAP options;
• phase boundaries and display rules.

Training once from the parameters in the paper cannot be assumed to produce the same result.

If the evaluation moves to new terminals, wires or equipment, evidence can be prepared in the following order. These are evaluation suggestions, not production release criteria.

1. First establish how the reference labels are obtained. Record how missing strands and crimped-in insulation are prepared and confirmed, and how minor anomalies are distinguished from acceptable conditions. Keep this information with the sample IDs. When labels are confused, improving the model’s score does not resolve the definition of quality.
2. Reserve new conditions for a genuinely external test. To demonstrate use across batches, hold out entire batches. To demonstrate use after changing equipment or terminals, hold out those conditions. Randomly splitting data from the same source answers only part of these questions. Unknown defects are not automatically covered by a three-class result either.
3. Compare with existing monitoring on the same set of samples. Count false alarms on normal samples, missed defects and incorrect defect-type indications separately against the reference labels, retaining the actual denominators. Do not compare the paper’s accuracy with the alarm rate on a different set of workpieces. When product defect prevalence changes, also examine the groups predicted as defective and normal separately to determine how many anomalies in each are confirmed by the project’s methods.
4. Finally assess whether the explanation changes actions. Within the existing response procedure, check whether the color bands reduce checks in the wrong direction or help engineers recognize when the model is unreliable. If the only finding is that phase explanations feel more intuitive, limit the conclusion to explanation plausibility and retain the need to validate people’s responses.

The most useful first page for a technical review may contain just three lines: error directions and denominators under the new conditions; the perturbation comparison used to test the explanation; and the judgment tasks users have actually completed. The study’s random forest, TreeSHAP and phase color bands offer a combination worth investigating for the first two lines. The third still needs evidence from the intended working environment.

To continue preparing a crimp quality review, read what crimp height, pull force and cross-sections each reveal to distinguish the uses of process curves and finished-product inspection. Find more Frontier Reviews and practical guides in the resource library.

Sources and reading notes

Main study: Hofmann, B., Bründl, P., Nguyen, H. G. and Franke, J., Transparent machine learning for crimp force monitoring using phase-based SHAP explanations, Industrial Artificial Intelligence 4, 2 (2026), DOI: 10.1007/s44244-026-00028-6. The Chinese review and evaluation suggestions were independently written from the publicly available full text; none of the original figures is reproduced. The original paper is licensed under CC BY 4.0. The arithmetic example and proposed additional tests in this review are not the authors’ measured results, nor do they represent reproduction or production-line validation by MTTJ.

Sources accessed: October 7, 2026. If the authors subsequently release code, revise results, or independent validation across equipment becomes available, the implementation conditions and assessment of transfer should be updated first.

AI assisted with organizing sources and drafting this article.