4.1. Experimental Design
To validate the effectiveness of FBLS in network fault diagnoses and propagation path analyses, this study designs a series of experiments based on the actual operational scenarios of the TIS system, evaluating FBLS’s performance in terms of fault feature extraction, propagation path analysis, diagnostic accuracy, and real-time performance.
The maintenance terminal of the TIS system can display the operational status of equipment in real time, replay historical information, and support fault information queries. By collecting alarm information from the network status [
19], an alarm information table for the TIS network status is compiled. The faults indicated by these alarm messages are defined as different target labels, while the status information of the corresponding logical channels and nodes for each alarm is recorded to construct sample features, as shown in
Table 3. After data cleaning and normalization, the dataset is organized into an input matrix
, where N represents the number of samples and M is the feature dimension for each sample.
Each row of the input matrix X represents an alarm sample, and each column corresponds to a feature of that sample, including the logical channel state and node state. Each alarm record is labeled with a different target, serving as the supervised learning objective of the model. The label value y represents the corresponding fault type, with a range of y ∈ {1,2,…,K}, where K is the total number of fault types. By combining the input matrix X with the label y, a complete training and testing dataset can be constructed, providing effective input data support for the FBLS model.
Data Preprocessing: The monitoring data are normalized to construct feature vectors for logical channels and node states, eliminating noise interference.
Model Training and Testing: The processed data are input into the FBLS model, with 80% of the samples used for training and 20% for testing. The model’s performance in terms of fault detection and propagation path prediction is evaluated.
Embedding Interpretable Linguistic Fuzzy Rules (ILFR): ILFR is embedded into the enhancement node layer to analyze its effect on improving the model’s interpretability.
Performance Evaluation: The performance of FBLS is compared with traditional methods (e.g., graph theory and fuzzy reasoning-based diagnostic methods) to evaluate the model’s diagnostic accuracy, real-time performance, and interpretability.
4.3. Experimental Results
To verify the effectiveness and advantages of FBLS in terms of accuracy and interpretability, we compared FBLS with several representative diagnostic algorithms, including: SVM (Support Vector Machine), a commonly used classifier in traditional machine learning; BLS (Broad Learning System), a broad learning model without a fuzzy layer; BP (Backpropagation Neural Network), a classical BP neural network; RFA (Random Forest Algorithm), an ensemble model based on decision trees; and FELM (Fuzzy Extreme Learning Machine), a fuzzy neural network model.
Table 4 summarizes the RMSE (Root Mean Squared Error), MAPE (Mean Absolute Percentage Error), training accuracy, and testing accuracy of each algorithm on the same dataset. The results demonstrate that FBLS exhibits significant advantages across all metrics:
(1) Minimal Numerical Deviation
The RMSE of the FBLS model is only 5.5497, significantly reducing the error compared to other methods. Additionally, the MAPE value is as low as 1.19%, the smallest among all methods, indicating an extremely low deviation from the true fault labels.
(2) High Accuracy and Strong Generalization Ability
FBLS achieves a training accuracy of 98.32% and a testing accuracy of 98.17%, significantly outperforming methods such as SVM (86.59%), BLS (95.12%), BP (95.12%), and FELM (95.73%). Even when compared with the relatively well-performing RFA (97.56%), FBLS still has room for further improvement.
Currently, for network fault diagnoses in the TIS system in China, methods based on graph theory and fuzzy inference are predominantly used. To better demonstrate the superior performance of FBLS, scatter plots of the actual labels and predicted labels under both the FBLS model and the fuzzy inference model have been generated, as shown in
Figure 8. These plots provide an intuitive comparison of the distribution of predicted labels from the fuzzy inference model and the FBLS model in relation to actual labels.
In terms of diagnostic accuracy, under a noisy environment, the FBLS model achieves an average diagnostic accuracy of 95.2% on the test set, which is significantly higher than that of traditional graph theory and fuzzy inference-based methods, which range between 85% and 88%.
- 2.
Real-Time Response Evaluation
To quantify the real-time performance of FBLS, the experiment defines three key time metrics:
Fault detection time (): The time required for the model to detect a logical channel fault trigger.
Diagnosis time (): The time required for the model to infer the fault type and propagation path.
Fault display time (): The time taken for the detected fault to appear on the monitoring interface.
For the optimized method proposed in this paper, the fault detection and display times remain relatively constant, while FBLS primarily optimizes the fault diagnosis time ().
(1) Collecting Fault Logical Channel Information
In the TIS (Train Control System), network faults are typically reflected through logical channels. Changes in logical channel states (such as faults, timeouts, packet loss, etc.) help detect anomalies in the system. The first step in real-time response evaluation is to collect logical channel information across the system, and then to monitor and record their state changes.
Data Collection: When a fault occurs during system operation, the network monitoring system captures the corresponding fault information. This typically includes the status of different nodes and the state of signal transmission links.
Data Preparation: Once fault information is collected, it is formatted into input data (feature vectors) that the FBLS algorithm can use. These input data include features related to the fault.
According to the technical specifications of the integrated train control and interlocking system, the minimum determination time for communication fault detection should be greater than 3000 ms. The calculation for the minimum communication fault determination time is given by:
is the minimum determination time for communication faults;
is the frame transmission time; and
is the communication request frame transmission time.
(2) Fault Type Identification Using the FBLS Model
The core task of the FBLS algorithm is to quickly determine fault types based on the collected fault feature data and predict fault propagation paths. By combining fuzzy logic and broad learning, FBLS can provide accurate classifications and inferences when dealing with complex fault patterns.
Feature Input: The collected fault feature data are fed into the FBLS model. During the training phase, FBLS processes these inputs through its fuzzy subsystems, maps them onto different fuzzy rules, and extracts useful fault information.
Inference Process: The enhancement node layer of FBLS performs nonlinear mapping, integrating outputs from multiple fuzzy subsystems. Based on these calculations, the FBLS model determines the specific fault type and predicts the affected system areas (fault propagation paths).
In this experiment, a total of 140 fault datasets were selected for testing, with a training time of 24.9560 s and a testing time of 1.1181 s.
These results indicate that the FBLS model can complete training in a relatively short time and perform fault diagnosis efficiently, significantly improving diagnostic efficiency.
To further quantify the real-time response performance of FBLS, we calculate the average diagnosis time (
) using the following formula:
is the average diagnosis time;
is the training time;
is the testing time; and
is the number of fault datasets.
For the experimental data, the FBLS model’s average diagnosis time per fault dataset is only 0.1836 s, demonstrating exceptionally high diagnostic efficiency and confirming that FBLS meets real-time response requirements in practical applications.
(3) Display of Diagnostic Results via the Network Monitoring Interface
After completing fault type identification and propagation path prediction, the FBLS algorithm displays the diagnostic results in real-time through the network monitoring interface for operation and maintenance personnel, with a display time () of 48 ms. Through the graphical interface, operators can intuitively understand the specific fault conditions and quickly locate the fault source.
For example, consider a network cable fault between TIS-B and Switch A. This fault causes all logical channels from TIS-B to Switch A to be completely interrupted. However, since Operator Display Machine A, Operator Display Machine B, and the Maintenance Machine can still communicate normally with TIS-A via Switch A, the fault is limited to a communication disruption between TIS-B and Switch A.
After diagnosis using the FBLS model, the system accurately displays the fault propagation path and the affected nodes on the interface, visually marking the fault location, as illustrated in
Figure 9. This assists maintenance personnel in quickly pinpointing and resolving the issue.
- 3.
Interpretability Analysis
In addition to its advantages in classification accuracy and structure, another key strength of FBLS is its interpretability in terms of fault analyses and explanations. Among the six models mentioned above, only FELM and FBLS integrate ILFR, enabling model interpretability. In contrast, the other four models solely perform classification tasks and cannot intuitively display the specific learning mechanisms or present the acquired knowledge in an understandable manner. This lack of interpretability is one of the major challenges in many existing machine learning-based fault diagnosis methods [
10]. However, FELM does not expand the enhancement nodes into a set of EEUs. Compared to FBLS, it can only ensure model performance by setting a large number of nodes, which results in the system containing an excessive number of rules.
By embedding Interpretable Linguistic Fuzzy Rules (ILFR), the experiment fully validated the significant advantages of the FBLS model in terms of interpretability. ILFR not only transforms complex fuzzy logic operations into natural language rules but also intuitively demonstrates the relationship between input features and fault diagnosis results. The fuzzy rules generated by ILFR cover the network fault scenarios involved in the TIS maintenance terminal, as shown in
Table 5, providing clear logical reasoning for maintenance personnel.
In this experiment, each feature is mapped to five fuzzy sets, representing Low, Slight Low, Medium, Slight High, and High.
Table 6 presents the first fuzzy rule, detailing both its conditions and conclusion. The condition section transparently indicates whether each feature is activated and, if so, the corresponding fuzzy mapping. Specifically, “DC” denotes that a feature is not activated, while the numbers 1–5 correspond to different neuro-semantic fuzzy categories. For example, if X
1 is marked as “DC”, it means X
1 is not activated in the first rule; if X
1 is assigned “2345”, it signifies that X
1 is activated and falls under one or more fuzzy conditions, i.e., Slight Low, Medium, Slight High, or High. The conclusion section consists of a four-dimensional linear output, where the condition section explicitly defines the prerequisite conditions for feature activation, and the output vector determines the specific fault type based on which conclusion is met.
In the network fault diagnosis of the TIS system, the network cable fault between TIS-B and Switch A serves as an example to analyze the impact range of the fault and the corresponding rule generation process. Specifically, when the network cable fails, logical channels 5, 6, and 14 simultaneously experience faults. Based on these conditions, ILFR (Interpretable Linguistic Fuzzy Rules) can generate a clear rule to determine the potential fault between TIS-B and Switch A and present this rule in natural language, making it easier for operation and maintenance personnel to understand and act upon. For instance, the generated rule can be expressed as “If logical channels 5, 6, and 14 experience communication interruptions, then there is a fault in the network cable between TIS-B and Switch A”. A visual representation of this rule is shown in
Figure 10.