Abstract In modern manufacturing systems, especially assembly lines, human input is a critical resource to provide dexterity and flexibility. However, the repetitive precision tasks common in assembly lines can have adverse effects on workers and overall system performance. We present a data-driven approach to evaluating task performance using wearable sensor data (kinematics, electromyography and heart rate). Eighteen participants (gender-balanced) completed repeated cycles of maze tracking and assembly/disassembly. Various combinations of input data types and classification algorithms were used to model task performance. The use of the linear discriminant analysis (LDA) algorithm and kinematic data provided the most promising classification performance. The highest model accuracy was found using the LDA algorithm and all data types, with respective levels of 62.4, 88.6, 85.8 and 94.1% for predicting maze errors, maze speed, assembly/disassembly errors and assembly/disassembly speed. The presented approach provides the possibility for real-time, on-line and comprehensive monitoring of system performance in assembly-lines or similar industries. Practitioner summary: This paper proposed models the repetitive precision task performance using data collected from wearable sensors. The use of the LDA algorithm and kinematic data provided the most promising classification performance. The presented approach provides the possibility for real-time, on-line and comprehensive monitoring of system performance in assembly lines or similar industries. Abbreviations: AD: anterior deltoid; BB: biceps brachii; ECR: extensor carpi radialis; ECU: extensor carpi ulnaris; FCR: flexor carpi radialis; FCU: flexor carpi ulnaris; FN: false negatives; FP: false positives; HR: heart rate; HRR: heart rate reserve; IMUs: inertial measurement units; kNN: k-nearest neighbors; LDA: linear discriminant analysis; MD: medial deltoid; MF: median power frequency; MNF: mean power frequency; MVIC: maximum voluntary isometric contraction; nRMS: normalized root-mean-square amplitudes; PD: posterior deltoid; RandFor: random forests; RHR: resting heart rate; RMS: root-mean-square amplitudes; sEMG: surface electromyographic; SVM: support vector machines; TB: triceps brachii medial; TN: true negatives; TP: true positives; t-SNE: t-distributed Stochastic Neighbor Embedding; UT: upper trapezius

中文翻译：

---

使用可穿戴传感器在重复性精密任务中建模性能：数据驱动的方法

摘要 在现代制造系统中，尤其是装配线，人力输入是提供灵巧性和灵活性的关键资源。然而，装配线上常见的重复性精密任务会对工人和整体系统性能产生不利影响。我们提出了一种使用可穿戴传感器数据（运动学、肌电图和心率）评估任务性能的数据驱动方法。十八名参与者（性别平衡）完成了迷宫跟踪和组装/拆卸的重复循环。输入数据类型和分类算法的各种组合用于对任务性能进行建模。线性判别分析 (LDA) 算法和运动学数据的使用提供了最有希望的分类性能。使用 LDA 算法和所有数据类型发现了最高的模型精度，预测迷宫错误、迷宫速度、组装/拆卸错误和组装/拆卸速度的水平分别为 62.4、88.6、85.8 和 94.1%。所提出的方法为在装配线或类似行业中实时、在线和全面监控系统性能提供了可能性。从业者总结：本文提出了使用从可穿戴传感器收集的数据对重复性精确任务性能进行建模。LDA 算法和运动学数据的使用提供了最有希望的分类性能。所提出的方法为在装配线或类似行业中实时、在线和全面监控系统性能提供了可能性。缩写： AD：三角肌前束；BB：肱二头肌；ECR：桡侧腕伸肌；ECU：尺侧腕伸肌；FCR：桡侧腕屈肌；FCU：尺侧腕屈肌；FN：假阴性；FP：误报；HR：心率；HRR：心率储备；IMUs：惯性测量单元；kNN：k-最近邻；LDA：线性判别分析；MD：内侧三角肌；MF：中值电源频率；MNF：平均电源频率；MVIC：最大自愿等长收缩；nRMS：归一化均方根幅度；PD：三角肌后束；RandFor：随机森林；RHR：静息心率；RMS：均方根幅度；sEMG：表面肌电图；SVM：支持向量机；TB：肱三头肌内侧；TN：真阴性；TP：真阳性；t-SNE：t-distributed Stochastic Neighbor Embedding；UT：上斜方肌 k-最近邻；LDA：线性判别分析；MD：内侧三角肌；MF：中值电源频率；MNF：平均电源频率；MVIC：最大自愿等长收缩；nRMS：归一化均方根幅度；PD：三角肌后束；RandFor：随机森林；RHR：静息心率；RMS：均方根幅度；sEMG：表面肌电图；SVM：支持向量机；TB：肱三头肌内侧；TN：真阴性；TP：真阳性；t-SNE：t-distributed Stochastic Neighbor Embedding；UT：上斜方肌 k-最近邻；LDA：线性判别分析；MD：内侧三角肌；MF：中值电源频率；MNF：平均电源频率；MVIC：最大自愿等长收缩；nRMS：归一化均方根幅度；PD：三角肌后束；RandFor：随机森林；RHR：静息心率；RMS：均方根幅度；sEMG：表面肌电图；SVM：支持向量机；TB：肱三头肌内侧；TN：真阴性；TP：真阳性；t-SNE：t-distributed Stochastic Neighbor Embedding；UT：上斜方肌 均方根振幅；sEMG：表面肌电图；SVM：支持向量机；TB：肱三头肌内侧；TN：真阴性；TP：真阳性；t-SNE：t-distributed Stochastic Neighbor Embedding；UT：上斜方肌 均方根振幅；sEMG：表面肌电图；SVM：支持向量机；TB：肱三头肌内侧；TN：真阴性；TP：真阳性；t-SNE：t-distributed Stochastic Neighbor Embedding；UT：上斜方肌