Figure 1
The illustration presents a three-stage workflow diagram for a fault diagnosis system applied to a lift door system. The process is divided into three labeled sections arranged from left to right: (1) Data preparation, (2) Feature extraction, and (3) Fault Classification. The left panel, labeled “(1) Data preparation”, shows the initial data acquisition and preprocessing stage. At the top, a label reads “Failed lift door system”. Below it, an image depicts a lift door mechanism connected to monitoring equipment, including a computer screen and a data acquisition device. An arrow points downward to the next step labeled “Data and Segmentation”. This section shows three time-series signals in different colors representing vibration or sensor measurements. An ellipsis is shown between the second and third series. Red dashed vertical boxes at the left and right on each series highlight segmented portions of the signal, indicating that the raw sensor data are divided into smaller samples for analysis. The segmented signals are labeled “Training data”. Two thick arrows from this panel point rightward to the central panel. The center panel is titled “(2) Feature extraction”. The panel is divided into two parts: the upper section shows neural network feature extraction layers, and the lower section illustrates metric learning in the feature space. At the top, several stacked neural network layers are displayed horizontally from left to right. Each layer is represented by a vertical rounded rectangle containing circular nodes. The first layers contain light yellow and light blue nodes, indicating intermediate feature maps. Solid arrows pointing right show the forward propagation of data from one layer to the next. Between some layers, dotted arrows pointing left indicate back propagation during training. The sequence of layers gradually transforms the input representation until the final layer on the far right, which contains blue circular nodes representing extracted features. Below the neural network layers, a label “Metric learning” marks the next stage. This section is enclosed by a red dashed rectangular boundary, representing the feature embedding space used to separate different health conditions. Within this feature space, several colored shapes represent samples from different health states: orange circles, blue squares, and green triangles. Each class forms clusters around center points, represented by larger outlined symbols such as a circled circle, a square outline, or a triangle outline. Gray arrows point toward these centers. Red dashed curved lines divide the feature space into regions. In the upper portion of the metric learning space, three clusters of different shapes are partially overlapping but being separated by the boundaries. In the lower portion, the clusters become more compact and clearly separated, illustrating how the metric learning process improves class separation. Two thick arrows from this central panel point rightward to the third panel. The right panel is labeled “(3) Fault Classification”. The panel depicts a simplified neural network used to classify system conditions based on extracted features. On the left side, a vertical column of green circular nodes. The circles are stacked vertically inside a rounded rectangular container, with a vertical ellipsis between them indicating additional nodes. From these input nodes, several connection lines extend to a second vertical column of rectangular output nodes, representing the classification layer. The output column is labeled “Predicted Labels” above it. The rectangles represent the predicted categories corresponding to different health states of the system. To the right of the output layer, a green rectangular box contains the symbols “L subscript m” and “L subscript c”, representing the loss functions used during training, typically metric loss and classification loss. A solid arrow pointing right connects the predicted label layer to this loss block. A dotted arrow pointing left indicates the back propagation of gradients from the loss functions back through the network. Below the classification diagram is a legend explaining the graphical symbols used in the overall framework: A solid black arrow represents forward propagation. A dotted arrow pointing left represents back propagation. Outlined symbols—a circle, square, and triangle—represent the centers of different health states. Filled symbols—blue squares, orange circles, and green triangles—represent samples belonging to different health states. A red dashed line represents a decision boundary. A gray arrow indicates distance reduction.Structure of the proposed DMSRN method. Source(s): Created by authors