Event-Triggered Collaborative Fault Diagnosis for UAV–UGV Systems

Abstract

The heterogeneous unmanned system, which is composed of unmanned aerial vehicles (UAV) and unmanned ground vehicles (UGV), has been broadly applied in many domains. Collaborative fault diagnosis (CFD) among UAVs and UGVs has become a key technology in these unmanned systems. However, collaborative fault diagnosis in unmanned systems faces the challenges of the dynamic environment and limited communication bandwidth. This paper proposes an event-triggered collaborative fault diagnosis framework for the UAV–UGV system. The framework aims to achieve autonomous fault monitoring and cooperative diagnosis among unmanned systems, thus enhancing system security and reliability. Firstly, we propose a fault trigger mechanism based on broad learning systems (BLS), which utilizes sensor data to accurately detect and identify faults. Then, under the dynamic event triggering mechanism, the network communication topology between the UAV–UGV system and BLS is used to achieve cooperative fault diagnosis. To validate the effectiveness of our proposed scheme, we conduct experiments on a software-in-the-loop (SIL) simulation platform. The experimental results demonstrate that our method achieves high diagnosis accuracy for the UAV–UGV system.

1. Introduction

Heterogeneous unmanned systems, which consist of Unmanned Aerial Vehicles (UAVs) and Unmanned Ground Vehicles (UGVs), have been used in many applications, such as logistics, transportation, and industrial production [1,2,3,4]. For example, in logistics tasks, UGVs can be used for long-distance cargo transportation, while UAVs are designed for rapid distribution. This unique capability provided using UAVs and UGVs enables efficient logistics, which is often missing in homogeneous unmanned systems [5,6,7,8]. Meanwhile, the increase in the complexity and scale of unmanned systems also poses challenges to the security and reliability of the system. Thus, several Cooperative Fault Diagnosis (CFD) techniques in heterogeneous unmanned systems have become a popular topic [9,10,11].

Over the past few decades, many model-based fault diagnosis methods have been proposed for unmanned systems [12,13,14,15,16]. For example, ref. [12] offers analytical redundancy using the differential flatness property of flat systems. This approach can provide the required residuals for fault diagnosis on sensors and actuators of multi-rotor vehicles. Ref. [14] designed a robust Linear Parameter Varying (LPV) observer to diagnose sensor faults of a UAV. The LPV observer produces a set of residuals, ensuring that each residual is influenced by only one fault. Ref. [16] implements anomaly detection for sensors in a fixed-wing UAV using the maximum likelihood and particle filter methods. Ref. [17] introduced an adaptive fault estimation observer with adjustable parameters for unmanned systems with directed topology. Ref. [18] proposed a distributed fault estimation method for unmanned systems, in which each agent employs an observer to simultaneously estimate both states and faults. The observers in [19,20] can estimate both the original state and newly defined variables. Based on these state and variable measurements, fault estimation can then be obtained.

The above methods improve the reliability of unmanned systems by using model-based fault detection and diagnosis solutions. However, these algorithms are mainly used for homogeneous unmanned systems. To the best of our knowledge, there are only a few studies on fault diagnosis for heterogeneous unmanned systems [11]. In practice, the above model-based methods may suffer from the time-varying parameters of unmanned systems, resulting in inaccurate fault detection and diagnosis results.

Recently, deep learning has made significant progress and has also been applied in the field of fault diagnosis [21,22,23,24,25,26,27,28,29]. Deep learning technology is crucial in fault diagnosis since it possesses better feature extraction capability. This allows for the extraction of many fault characteristics from data, resulting in improved diagnosis accuracy and efficiency.

Ref. [30] proposed a hybrid method that combines discrete wavelet transform and deep neural networks for UAV fault diagnosis, achieving high classification accuracy. Ref. [31] implemented various techniques, such as the sliding window method, deep residual shrinkage networks, and wide convolution layers, to diagnose micro-damage faults in quadrotor propellers. Ref. [32] introduced the stacked-informer network with gradient-focused optimizers for power line trip fault sequence prediction. Based on generative adversarial networks and minimum singular values, ref. [33] developed a deep transfer learning method for bearing fault diagnosis.

However, the aforementioned methods are developed on deep neural networks, resulting in inefficient inference and complex architectures. To address the above issues, ref. [34] proposed a FD method based on a broad learning system (BLS), which demonstrates high efficiency and enables FD applications in heterogeneous Multi-Agent Systems (MAS). Ref. [35] proposed the broad learning system (BLS) algorithm for homogeneous MAS consisting only of UAVs. The end-to-end fault diagnosis approach with a periodic communication scheme is used in RFBLS, leading to increased communication resource usage. Thus, in this work, we propose an event-triggered collaborative fault diagnosis method for UAV–UGV systems. The main contributions of this study are summarized as follows:

(1) In this work, we propose an event-triggered BLS, a two-stage fault diagnosis approach, to reduce wireless communication overhead. In the first stage, each agent uses a binary classifier to determine whether a fault has occurred. If a fault is detected, the data from each agent are wirelessly exchanged to facilitate collaborative fault diagnosis. If no fault is detected, no further fault diagnosis process occurs.

(2) This work simulates the proposed event-triggered collaborative fault diagnosis in large-scale UAV–UGV systems. Several other state-of-the-art fault diagnosis methods are also implemented, and the comprehensive experimental results demonstrate the effectiveness and efficiency of the proposed method.

The rest of this paper is structured as follows: Section 2 presents an overview of the UAV–UGV system, BLS and the event-triggered mechanism. Section 3 proposes the event-triggered collaborative fault diagnosis framework. In Section 4, we present simulation results for large-scale heterogeneous UAV–UGV systems. Finally, we conclude this work in Section 5.

2. Preliminaries

In this section, we mainly introduce the UAV–UGV system, broad learning system and event-triggered mechanism.

2.1. Overview of UAV–UGV System

The UAV–UGV system offers an efficient approach to executing complex missions by leveraging the strengths of both UAVs and UGVs. Even though UAVs and UGVs differ in form and function, they still share similar characteristics, such as failures in actuators, sensors, and wireless communication.

In the real world, the UAV–UGV systems often consists of several subsystems. Each subsystem includes input signals $u_{*}$, output signals $y_{*}$, parameters ${\phi }_{*}$, dynamic states $x_{*}$, dynamic state functions $f_{*}(·)$ and output functions $h_{*}(·)$, expressed as the following:

$$\left\{\begin{array}{c}{\stackrel{•}{x}}_{*}=f_{*}(x_{*},{\phi }_{*},u_{*})\\ y_{*}=h_{*}(x_{*},{\phi }_{*},u_{*})\end{array}\right.$$

(1)

2.2. Overview of BLS

The broad learning system (BLS) model for UGV–UAV systems was proposed by [34]. This method has demonstrated generalization capabilities and computational efficiency, allowing it to quickly adapt to the variability of complex environments. Different from the traditional deep neural network, the BLS model employs a flattened architecture composed of feature nodes and enhancement node layers

The input data $\mathit{X}$ are mapped to feature nodes by the nonlinear projection function. The output $\mathit{Y}$ of BLS is the probability of which one class is included. It can be calculated as

$$\underset{W}{\min }{‖Y-QW‖}_2^2+\lambda {‖W‖}_1$$

(2)

where $\lambda$ there is the constraint on the sum of the squared output weight. The weights $\mathit{W}$ of the BLS model are determined through ridge regression, enabling a rapid solution compared to the time-consuming back-propagation method used in traditional deep learning models.

2.3. Event-Triggered Mechanism

The event-triggered mechanism applied to the fault diagnosis of heterogeneous unmanned systems ensures that the fault diagnosis process and data exchange occur only when the system detects a potential failure (for example, a sudden decrease in the flight altitude of a drone or the abnormal speed of an unmanned vehicle). This approach reduces wireless communication overhead, making the event-triggered fault diagnosis method more efficient in saving communication resources compared to the periodic fault diagnosis solution.

3. The Event-Triggered Collaborative Fault Diagnosis Method

In this section, the entire architecture of the proposed CFD method is shown in Figure 1. The proposed CFD framework consists of three phases: ZCA whitens the data to minimize the correlation between features, and this pre-processing step is applied in the first phase. The ZCA method not only preserves all the information in the data set but also minimizes redundancy between features, thereby improving the efficiency and accuracy of subsequent processing.

For the second phase, BLS is used to for an event-triggered mechanism, which activates diagnosis and communication when a failure occurs in the UAV–UGV system. This ensures that the system responds to faults and improves the real-time accuracy of fault diagnosis. Once the fault is detected, we proceed to the third phase. Cooperative fault diagnosis is conducted using the communication network and BLS. The wireless communication allows all unmanned systems to share fault information. The BLS further analyzes the fault information from each unmanned system for a comprehensive diagnosis.

Thus, the CFD framework is designed to effectively diagnose faults within the UAV–UGV system by utilizing an event-triggered mechanism. The input data, consisting of sensor data and state estimation information, are sent for fault detection using a fault detector, and the system communication network and fault diagnostics are used for cooperative fault diagnosis when the fault occurs. Thus, the proposed method can enhance the overall reliability and performance of the system. The input data consist of two parts, namely, the sensor data and state estimation information (such as position, velocity, and acceleration, etc.), denoted by the data set ${\mathit{X}}_n$ and the fault data set ${\mathit{X}}_f$, where $f=$ 1, 2, …, F.

3.1. Data Pre-Processing

First, we normalize the collected data from the heterogeneous UAV–UGV system, including position, velocity, and acceleration, denoted as $\mathit{X}$ (including normal data set ${\mathit{X}}_n$ and fault data set ${\mathit{X}}_f$). Diagnosing a fault involves identifying the fault type of the corresponding UAVs or UGVs in the UAV–UGV system. However, directly using raw data may yield poor performance. To address this issue, ZCA is introduced to reduce feature redundancy and enhance fault characterization. The process of ZCA is detailed here:

The data set $\mathit{X}$ is pre-processed using ZCA. The covariance matrix is constructed by using the following equation:

$$C=\frac{1}{m}X{X}^T$$

(3)

Next, Single Value Decomposition (SVD) is performed on the above covariance matrix:

$$[U,T,V]=\mathrm{svd}(X)$$

(4)

where $\mathit{U}$ represents the eigenvector matrix of $\mathit{C}$. $\mathit{T}$ represents the eigenvalue matrix of $\mathit{C}$. The data rotation process is performed using the following equation:

$$X_{\mathrm{rot}}=U^{T}X$$

(5)

PCA whitening is as follows:

$$X_{PCAwhite}=X^{-\frac{1}{2}}\cdot X_{\mathrm{rot}}$$

(6)

where ${\mathit{X}}^{-\frac{1}{2}}$ represents the scaling factor and the scaling factor is obtained by dividing the square root of the corresponding eigenvalue.

$$X\prime =U\cdot X_{fPCAwhite}$$

(7)

3.2. Event-Triggered Mechanism of UAV–UGV System

The BLS model is then applied for dynamic event triggering. Specifically, the fault detector is introduced to monitor the occurrence of failures. The learning mechanism of the fault detector model follows

$${T_i}^1={\varphi }_i(X^{\prime }{W}_{ei}+{\beta }_{ei}^T),i=1,\cdots ,n$$

(8)

where ${\varphi }_{\mathrm{i}}$ represents the activation function for the $i$-th feature node. The variables ${\mathit{W}}_{\mathit{e}\mathit{i}}$ and ${\mathit{\beta }}_{\mathit{e}\mathit{i}}$ are the randomly set weight and bias for the $i$-th feature node, respectively. It is worth noting that all the feature nodes can be referred to as the feature nodes layer ${\mathit{T}}^{n1}$.

$$T^{n1}=[{T_1}^1,{T_2}^1,\cdots ,{T_n}^1]$$

(9)

The feature nodes layer ${\mathit{T}}^{n1}$ is used to map the enhancement nodes (composed of $m$ sets). This mapping is represented as follows:

$${E_j}^1={\phi }_j(T^{n1}{W}_{h_j}+{\beta }_{h_j}^T),j=1,\cdots ,m$$

(10)

where ${\mathit{W}}_{{\mathit{h}}_{\mathit{j}}}$ is a weight matrix and ${\mathit{\beta }}_{{\mathit{h}}_{\mathit{j}}}$ denotes a bias matrix. Both the weight and bias matrices are randomly generated. Finally, all the enhancement nodes are used to form the enhancement node layer ${\mathit{E}}^{m1}$, following

$$E^{m1}=[{E_1}^1,{E_2}^1,\cdots ,{E_m}^1]$$

(11)

Therefore, the output of the fault detection result is ${\mathit{F}}^1$ and can be calculated as

$$F^1=[T^{\mathrm{n}1}|E^{m1}]W^{m1}$$

(12)

where the weights ${\mathit{W}}^{m1}$ is solved using pseudo-inverse method.

3.3. Collaborative Diagnosis of UAV–UGV System

Wireless communication and the fault diagnoser are used to realize the collaborative diagnosis. The input of the fault diagnoser is denoted as ${\mathit{X}}_{\mathit{f}}^{\prime }$, and the feature node layer can be obtained by using Equation (8). These equations are defined as follows:

$$T^{n2}=[{T_1}^2,{T_2}^2,\cdots ,{T_n}^2]$$

(13)

where ${\mathit{T}}_n^2$ represents the feature nodes of fault diagnostics. The enhancement nodes layer ${\mathit{H}}^{m^{\prime }}$ of fault diagnostics are mapped from Equation (10):

$$E^{m2}=[{E_1}^2,{E_2}^2,\cdots ,{E_m}^2]$$

(14)

The output from fault diagnosis follows

$$F^2=[T^{\mathrm{n}2}|E^{m2}]W^{m2}$$

(15)

The diagnosis result ${\mathit{F}}^2$ from the fault diagnostics is used for distributed collaborative diagnosis among UAVs and UGVs. For example, in UAV 2, the collaborative fault diagnosis results are achieved by wirelessly exchanging its own fault diagnosis results with those from neighboring UAVs and UGVs.

4. Result and Discussion

In this section, the proposed fault detection and diagnosis methods are verified on the software-in-the-loop (SIL) UAV–UGV system platform.

4.1. Experimental Settings and Datasets

In this study, various types of sensor faults, including position, velocity and acceleration, are introduced at different locations in the heterogeneous UGV–UAV system. The rate of sensor failure is characterized by a random variable ranging from 0% to 3%, 3% to 6% and from 6% to 9%. As a result, a total of nine different failure scenarios and one normal working state are generated. The data set possesses 20,000 samples, where 70% are allocated for training and the remaining 30% are used for testing.

4.2. Analysis of Event-Triggered Detection Results

We present the performance evaluation of different hyperparameters of the fault detector during the fault detection phase. Figure 2 illustrates the event-triggered results obtained through varying the parameters for both UAVs and UGVs. The number of feature nodes is set at a fixed value of 20. We then determine the suitable number of feature nodes. Figure 2a indicates that UAVs and UGVs achieve the best performance when utilizing 60 enhancement nodes. Figure 2b shows that for UAVs and UGVs, the optimal parameter settings for the number of feature nodes are 26 and 28, respectively.

The selection of these parameter settings is crucial as they impact the performance of the event-triggered process. By selecting the optimal number of feature nodes, the detection phase can effectively identify the faults.

For UAV–UGV systems, continuous signal transmission will cause large network communication overhead and energy consumption. We propose an event-triggered mechanism to realize the self-monitoring of UAVs and UGVs. When the fault occurs, UAVs and UGVs will trigger the alarm to communicate with other nearby UAVs and UGVs. Therefore, the event-triggered mechanism can effectively reduce communication overhead. In the SIL simulation platform, the fault detection results of the UAV–UGV system are shown in Figure 3. The event-triggered accuracy of UAVs, UAVs with ZCA, UGVs and UGVs with ZCA are 97.98%, 98.96%, 94.53% and 99%, respectively.

4.3. Analysis of Fault Diagnosis Results

Figure 4 shows the parameters used for fault diagnosis. In the cooperative fault diagnosis stage of fault diagnostics, compared with the event-triggered stage of the fault detector, the number of enhancement nodes and feature nodes are reduced for UAVs and UGVs. Specifically, on UAVs, there are 3440 fewer enhancement nodes and 324 fewer feature nodes. In UGVs, the number of enhanced nodes decreased by 2940 and the number of feature nodes decreased by 222. The event-triggered collaborative fault diagnosis is more efficient. Compared with the periodic fault diagnosis method, this improvement can effectively reduce the computational complexity of the network, thus saving computing resources and improving the algorithm calculation speed.

Figure 5 shows the results of the collaborative fault diagnosis. Specifically, Figure 5a,c demonstrate the diagnostic accuracy of the UAVs and UGVs without ZCA, and we achieve accuracies of 90.33% and 95.46%, respectively. In contrast, Figure 5b,d present the FD accuracy of the UAVs and UGVs with ZCA, and we achieve accuracies of 94.85% and 97.62%, respectively. These results indicate that the diagnostic performance significantly improves after applying ZCA for data processing.

4.4. Fault Diagnosis Comparison Results

To further validate the effectiveness of the proposed method, we select several methods, such as the Sparse Auto Encoder (SAE) [36], Support Vector Machines (SVM) [37], and the broad learning method for comparison. The optimal settings for the enhancement node and feature node of BLS are determined to be 3500 and 350, respectively, for UAV. For UGV, the optimal settings are 3000 and 250, respectively. For SVM, the standard deviation value and penalty coefficient are set to 3.5 and 5.5, respectively. The SAE neural network configuration consists of 50-30-20 neurons.

Table 1. Performance comparison using three methods.

Method	SAE	SVM	BLS	The Proposed Method
UAVs	86.93%	93.62%	90.33%	94.85%
UGVs	88.89%	93.03%	95.46%	97.62%
UAV–UGV systems	87.91%	93.32%	92.89%	96.23%

summarizes the comparison results among the four methods. For UAVs, compared with SVE, SVM and BLS models, the diagnostic accuracy of the proposed method is improved by 7.92%, 1.23% and 4.52%, respectively. For UGVs, compared with SVE, SVM and BLS models, the diagnostic accuracy of the proposed method is improved by 8.73%, 4.59% and 2.16%, respectively. For UAV–UGV systems, compared with SVE, SVM and BLS models, the diagnostic accuracy of the proposed method is improved by 8.32%, 2.91% and 3.34%, respectively. In summary, by comparing with SVE, SVM and BLS models, the proposed method outperforms the other three fault diagnosis methods.

5. Conclusions

This paper presents a collaborative fault diagnosis scheme for UAV–UGV systems, utilizing an event-triggered mechanism and wireless communication. This approach enables both the self-monitoring and collaborative diagnosis of heterogeneous UAV–UGV systems while effectively reducing wireless communication overhead. The proposed scheme is validated on the SIL platform, and comprehensive experimental results demonstrate its effectiveness.

In the future, further research could explore the integration of advanced machine learning techniques to improve the system’s diagnostic capabilities. In addition, the test platform is extended to include more diverse and complex fault scenarios to verify the robustness and adaptability of the algorithm.