CN104698839B - A kind of multiple agent fault detect based on information interaction and compensating control method - Google Patents

Abstract

The present invention is directed to current distributed multi agent system easily to break down, and without simple and feasible this problem of real time fail processing scheme, propose a kind of distributed real-time fault detection based on information interaction and compensating control method.Step one, system and fault modeling: described modeling comprises nodes dynamics model, Information Interaction Model, typical fault model; Step 2, multiple agent real-time fault detection based on information interaction; Step 3, based on the information integration of Gossip algorithm and process; The compensation rate of step 4, Control-oriented amount calculates and applies; Step 5, the connectedness designed based on two hop-informations keep: from Information Interaction Model, Content of Communication between malfunctioning node is analyzed, by utilizing two hop-informations wherein, setting up virtual information transfer path, ensureing that fault handling scheme can not the normal work of influential system.

Description

A multi-agent fault detection and compensation control method based on information interaction

technical field

The invention relates to a multi-agent fault detection and compensation control method based on information interaction, which belongs to the technical field of multi-agent control.

Background technique

In recent years, with the rapid development of computer and network technology, the scale of multi-agent systems is also growing rapidly. The traditional centralized control scheme has become more and more difficult to meet the needs of practical problems due to the limitation of the computing speed and sensing range of the central node. The distributed control scheme has gradually become the mainstream of multi-agent control research because of its low requirements for a single agent itself and good scalability. However, it is worth noting that in the distributed control scheme, there is no central node to coordinate and plan the behavior of all nodes, which makes the system vulnerable to attacks from faulty nodes and malicious nodes, which may lead to the failure of the entire system in severe cases. paralysis. Therefore, for a distributed multi-agent system, designing a safe and efficient fault detection scheme so that the system can automatically complete the detection and repair of faulty nodes is an urgent task with broad application prospects.

For the fault detection of distributed multi-agent systems, the existing solutions mainly include the following:

Scenario 1: Literature (I. Shames, A.M.H. Teixeira, H. Sandberg, and K.H. Johansson. Distributed fault detection for interconnected second order system [J]. Automatica, Oct. 2011, to appear.) and literature (S. Sundaramand C.N. Hadjicostis. :Attackingthenetwork[C].AmericanControlConference, june2008.) proposed to use the unknown input observer (UIO), through long-term observation, accumulate enough data to estimate the initial state of the system, and then calculate the final state of the system, and based on this Determine whether the movement of the current node meets the expected requirements. Using observers to monitor faults in real time is the mainstream solution for fault diagnosis of multi-agent systems. The fault signal acts as an unknown input in the system, drives the observer to generate an error output, and diagnoses and compensates the fault by using the error signal.

Fault diagnosis using the unknown input observer (UIO) has its own advantages, such as clear physical meaning, easy to understand; does not depend on physical models, and has a wide range of applications. In addition, the literature (Chung, W.H., Speyer, J.L., & Chen, R.H. Adecentralized fault detection filter [J]. Journal of Dynamic Systems, Measurement, and Control, 123(2), 237–247, 2001) pointed out that with other observers, such as Beard-Jones fault detection Compared with the filter (Beard-JonesFaultDetectionFilter), the unknown input observer (UIO) structure is relatively simple, and it is easy to apply the optimization algorithm to obtain an approximate optimal solution. But on the other hand, this scheme also has some shortcomings: for a node with N adjacent nodes, the number of unknown input observers (UIO) required to detect the failure of all its adjacent nodes is N+1. When the system topology is complex or the number of nodes is large, the amount of data and computation required by this solution will be very large, which will place high demands on the hardware of the nodes. At the same time, the operation of this scheme will occupy a large amount of computing resources, which will have adverse effects on other control tasks of the system.

Scheme 2: The literature (M.Franceschelli, M.Egerstedt, and A.Giua.Motion probes for fault detection and recovery in networked control systems [C]. American Control Conference, pages 4358–4363, June 2008.) proposed to use motion detectors to stimulate networked control systems by applying additional excitation signals , according to the response of the system to judge the current operating status of the system, so as to detect the faulty node. Different from the first scheme, this scheme adopts the method of active detection. For this scheme, the main problem is that it is difficult to operate in practice, the selection of excitation signal, the time of signal application, etc. will be restricted by many conditions.

Scheme 3: Document (Guo, M., Dimarogonas, D.V., & Johansson, K.H. (2012, June). Distributed real-time fault detection and isolation for cooperative multi-agent systems [C]. In American Control Conference (ACC), 2012 (pp.5270-5275). IEEE.) proposed Using the data information interaction between nodes, by receiving the control quantity information of adjacent nodes, the motion state is simulated and reproduced, and compared with the detected actual motion state of adjacent nodes, as a basis for fault detection. The biggest advantage of this scheme is that it is easy to calculate and has strong operability, but there are also some obvious shortcomings, such as too strict restrictive conditions, narrow scope of application, and high misoperation rate of the system.

Inspired by the above scheme 3, the present invention proposes a distributed fault detection and compensation control scheme based on information interaction while fully absorbing its advantages and aiming at its own shortcomings. This scheme improves the fault discrimination mechanism of nodes, and effectively improves the problem of high misoperation rate of the system by adopting the gossip propagation (Gossip) algorithm. At the same time, the scheme resets the content of information interaction between nodes, so that nodes can use the received information more effectively. In addition, considering the complexity of the information exchange protocol, the present invention proposes a control quantity-oriented fault recovery scheme, which greatly relaxes the system's restrictions on the node information exchange protocol and expands the scope of application of the scheme.

Contents of the invention

Aiming at the problem that the current distributed multi-agent system is prone to failure and there is no simple and feasible real-time fault handling scheme, the present invention proposes a distributed real-time fault detection and compensation control method based on information interaction, through information interaction between nodes , to complete the detection, isolation and repair of faulty nodes, so as to realize the timely processing of system faults and reduce the losses caused by them.

A distributed real-time fault detection and compensation control method based on information interaction of the present invention comprises the following steps:

Step 1. System and fault modeling: the modeling includes a node dynamics model, an information interaction model, and a typical fault model; wherein the node dynamics model adopts a single integrator model, and the motion state of the node is described by a first-order differential equation; the information The interaction model is described by an undirected graph, that is, two-way communication between nodes is possible, and each independent agent uses this to perform information interaction and complete system control tasks; the typical fault model includes the fault types that often occur in real agents;

Step 2. Multi-agent real-time fault detection based on information interaction: select relevant state variables from the expression of the node dynamic model described in step 1 as the description of the node's operating state; by setting the threshold function, the node's The running status is divided to distinguish between normal nodes and faulty nodes; at the same time, a single node obtains the status information of its adjacent nodes by means of the information interaction model described in step 1, and detects whether it is faulty through a detection algorithm to form a detection result of a single node;

Step 3. Information integration and processing based on Gossip algorithm: Due to the existence of problems such as communication packet loss and time lag, the single-node detection results in step 2 are greatly affected by the environment, and the reliability is not high; therefore, using the Gossip algorithm, the single-node The node detection results are integrated to obtain comprehensive detection results with higher reliability, and this is used as the final judgment basis for the node operating status to distinguish normal nodes from faulty nodes;

Step 4. Calculation and application of compensation for control quantities: If a faulty node is detected, the faulty node will be isolated through corresponding operations, and at the same time, based on the influence of the faulty node on the control quantity of its adjacent nodes, a relevant calculation scheme will be designed to obtain compensation The value of the quantity is added to the original control quantity to offset the impact of the faulty node on the system;

Step 5. Design connectivity maintenance based on two-hop information: starting from the information interaction model, analyze the communication content between faulty nodes, and establish a virtual information transmission path by using the two-hop information to ensure that the fault handling scheme will not affect the normal operation of the system.

The fault types mentioned therein include destructive faults, runaway faults and disturbing faults.

Compared with existing solutions, the advantages and innovations of the present invention mainly include the following points:

(1) In view of the problem that most of the existing schemes have high requirements on system hardware and need to occupy a large amount of computing resources (as shown in schemes 1 and 2), the present invention starts from the basic control rules of the multi-agent system, and makes full use of its existing In some calculation results, the real-time monitoring of adjacent nodes is realized under the condition of occupying very few calculation resources, which greatly reduces the application cost of the present invention. Simultaneously, at the expense of a small increase in communication content, the present invention effectively overcomes the interference problem of random signals to fault detection results (as shown in scheme 3) by utilizing the gossip algorithm, and this innovation ensures the reliability of fault detection results, It also makes the present invention have practical application value.

(2) There are not many studies on the isolation and repair of faulty nodes in the existing schemes, and most of them adopt a simple way of directly terminating communication, and the fault repair scheme is only applicable to linear control protocols (as shown in scheme 3), The scope of application is limited. The present invention starts from the control results of the system, and designs a fault isolation and compensation algorithm based on the control quantity. This algorithm fully considers the most common saturation characteristics of the system, and can effectively solve the problem of system fault repair under the nonlinear control protocol. The scope of application of the present invention is greatly expanded.

(3) For the problem of how to maintain the connectivity of the system after the fault node is isolated, none of the existing solutions has done in-depth research on this. The present invention designs a system topology structure maintenance scheme based on two-hop information by utilizing the existing communication content and reliable detection information obtained by using the gossip algorithm. This scheme can ensure that if the fault node does not leave the communication range of the normal node, that is The virtual information transmission path can be established by means of the information of the adjacent nodes transmitted by it to ensure that the system is connected at all times, and its normal functions will not be completely destroyed due to the isolation of faulty nodes.

Description of drawings

Figure 1—Topological structure diagram of multi-agent system;

Figure 2—Schematic diagram of fault detection scheme;

Figure 3—Schematic diagram of an information processing scheme based on the Gossip algorithm;

Figure 4—The relationship between node expected output and actual output;

Figure 5—the system topology structure diagram after using the two-hop information;

Figure 6—Comparison of results with or without a treatment plan for fault 1;

Figure 7—Comparison of results with or without a treatment plan for fault 2;

Figure 8—Comparison of results with or without a treatment plan for fault 3;

Fig. 9 - Flow chart of distributed real-time fault detection and compensation control method based on information interaction.

detailed description

Below in conjunction with accompanying drawing and example the present invention will be further described:

First give the system and detection model:

In the actual multi-agent system, the node's motion information is usually controlled as the target, in order to realize the node's motion state or position distribution to meet the control requirements. For a node using a single integrator model, its dynamic model satisfies the following form:

${\stackrel{<span\; class="notranslate">\; \&Center\; Dot;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

This formula shows that the control quantity of the node depends on the derivative of the node state, generally speaking, it is the speed information of the node. Equation (1) gives the dynamic model of the node in the continuous time state, but in the actual control system, because the node needs to sample the state information, and the sampling period cannot be infinitely small, therefore, it is necessary to establish Kinetic models in discrete-time regimes. According to the knowledge of computer control system and other related disciplines, the typical single integrator model in the discrete time state obtained after discretizing the above model has the following form:

z _i ((k+1)T)=z _i (kT)+u _i (kT)T,i=1,...,N(2)

where T is the sampling time. For simplicity, write z _i ^k = z _i (kT), u _i ^k = u _i (kT), and satisfy where z _i ^k is the position coordinate of the node in two-dimensional space, u _i ^k ∈ R ² is the control quantity of node i in each time step k. The actual physical meaning represented by this model is: take the position of the nodes in the system as the control target, adjust the node position by controlling the speed of the nodes in each time step k, and finally make the distribution of the nodes meet the control requirements. Regardless of information transmission via wired or wireless networks, multi-agent systems are based on information exchange between nodes to achieve collaborative control. The entire information interaction network can be described by a graph G={V,E}, where V={1,...,N} is a vertex in the graph, and also represents each node in the system, is the road in the figure. We define: if node i can transmit its own information to j, then i is called the adjacent node of j, that is (i, j)∈E. Denote N _i ^k = {i ₁ ,...,i _p } as the set of adjacent nodes of node i within time step k, and |N _i ^k | is its base. In addition, we assume that G is an undirected graph, that is, $(i,j)\&Element;\mathrm{E.}\&DoubleLeftRightArrow;(j,i)\&Element;\mathrm{E.}.$

In addition, it is assumed that the control rule of the node has the following structure:

u _i ^k ＝P _i (z _i ^k ,I _i ^k )(3)

Among them, P _i : R ² → R ² is the control protocol, which is determined by the control target of the node; is the state of adjacent nodes of node i within time step k, where N _i ^k ={i ₁ ,…,i _p }, p=|N _i ^k |. The structure shown in formula (3) is a commonly used structure in multi-agent cooperative control, that is, the control amount of a node is determined by its current state and the states of all adjacent nodes. P={P ₁ ,...P _N } is a preset control protocol, if it satisfies Then P is called a homogeneous control protocol, otherwise it is called a non-homogeneous control protocol. The present invention only considers the situation that the information exchange protocol is homogeneous.

Denote the set of all faulty nodes in the system as F, and the detection result of node i to node j within time step k is satisfy:

${\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\left\{\begin{array}{c}\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; \&NotElement;</span>\mathrm{<span\; class="notranslate">\; f</span>}\\ \mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; f</span>}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

In each time step k, the node will perform fault detection on all its adjacent nodes and obtain the detection results at the same time In addition, the detection result of the system for a node is defined as the synthesis of the detection results of all adjacent nodes of the node, and its form is as follows:

${\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}{\overset{{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

Where N _i ^k ={i ₁ ,...,i _p }, p=|N _i ^k |. Through data information interaction, each node will obtain the evaluation result of the system itself, and the specific acquisition method will be described in detail below.

The following is an analysis of the information interaction model between nodes:

In a multi-agent system, each node controls itself by sensing the surrounding environment. If the node's perception of the environment is based on the mutual information interaction between nodes, this model is called a model based on information interaction. In the system information interaction model discussed in the present invention, the content of information interaction between nodes consists of the following parts:

Content 1: Node i∈V transmits the control quantity u _i ^k obtained by formula (3) and its current state z _i ^k to all its adjacent nodes j∈N _i k within time step ^k .

Content 2: Node i∈V changes the state of its adjacent nodes within time step k And the detection result {j∈N _i ^k |Q _j ^k } obtained by the system to its adjacent nodes obtained by formula (5) is transmitted to all its adjacent nodes j∈N _i ^k .

Content 3: Node i∈V compares its detection results to the corresponding adjacent nodes within time step k And the detection results of adjacent nodes for i It is transmitted to all its adjacent nodes j∈N _i ^k .

The description of the value is the detection result of the adjacent node pair i transmitted in content 3 is not the form of Q _i ^k , although the two are completely equivalent in meaning. Here we mainly consider the case that node i is a malicious node. If Q _i ^k is directly transmitted, the data may be intentionally modified by the malicious node so that the system cannot detect the malicious node. use Since the data contains the detection information of the node itself to the malicious node, it can be used to check the information, or confirm the detection result by checking the data with the adjacent nodes of the malicious node. This belongs to the research category of information confrontation, which is not discussed in detail in the present invention.

Typical failure types are given below:

When defining system faults, due to the different composition and operation modes of different systems, the definition of faults is also different. For a system whose functional division is relatively independent and whose structure is fully known, it can be defined starting from the cause of the fault. For example, for a car, faults such as engine damage and brake failure can be specifically defined from the aspects of power system and braking system. The advantage of this is that it is highly targeted, and it can minimize damage to the system and ensure that its functions are not affected. Most systems in the real world use this fault definition method. However, for a multi-agent system, due to its complex and diverse operating modes and different topological structures, it is often difficult to know the specific cause of the fault, so it is impossible to define the fault from the source. Considering that a multi-agent system is a multi-agent system that cooperates to complete the control task, a single agent cannot have a decisive impact on the system. Therefore, it can be considered not to analyze the cause of the fault in detail, but to start from the actual operation results of the nodes Define it, that is, as long as the operation result of a certain node does not meet the system requirements, it will be deemed to be faulty and removed from the system. This way of fault definition will cause certain damage to the system, but compared with stopping the whole system to repair a node failure, the loss is still relatively small. In addition, this definition method can greatly simplify the difficulty of detecting faults in the system, and can handle faults in real time, ensuring that the system can still complete the expected tasks to the maximum extent in the presence of faults. Now combined with the actual multi-agent system, the following typical fault forms are given:

Fault 1: Destructive fault. The specific performance is that the node stops abnormally during operation, or although there is a movement trend, the actual state does not change as expected. The cause of this kind of failure may be that the node is attacked by the outside, causing the power system to be damaged, or the energy of the node is exhausted, and the source of power is lost. In addition, consider a special situation in the process of system operation, that is, the node cannot continue to move due to the influence of the surrounding environment or its own program. In this case, although the node itself is not damaged, it cannot move normally, so it is still classified as a destructive fault.

Fault 2: Out-of-control fault. The specific performance is that the node motion is uncontrolled, the speed remains unchanged, or changes unconventionally, so that the control effect cannot meet the system requirements. The reason may be that the control system fails to generate control information normally, or the power system of the node loses contact with the control system, and the actuator cannot obtain the correct control amount. In addition, this kind of fault is most likely to occur when the system is under malicious attack. It can be used as one of the signs to detect whether there are malicious nodes in the system. If this fault is detected, preventive measures should be taken in time to prevent the further spread of malicious information .

Fault 3: Interference type fault. The specific performance is that there are a large number of irregular movements of the nodes, and the deviation between the actual operating state and the theoretical operating state is too large, which has caused harm to the normal operation of the system. There are many reasons for this failure, and it is the most common in the actual multi-agent system. The specific reason may be that the node is affected by strong external random interference, such as the terrain is too rough, or the precision of the node actuator is insufficient, and the random error generated is too large. For such faults, one should be cautious when dealing with them, because the occurrence of errors is inevitable. If the detection procedure is too strict, a large number of nodes may be identified as faulty, which will bring unnecessary losses to the system. In order to solve this kind of problem, it can be considered to use filtering algorithm to compensate it.

In addition, for the model based on information interaction, the node's perception of the surrounding environment and the acquisition of its own control quantity completely depend on the data information interaction with adjacent nodes. Therefore, the data information interaction is the bridge between the node and the system. Running plays a vital role. For the above three fault types, if the node is only the controller or the power system fails, but still retains the normal information interaction function, it is called a type I fault; if the node information interaction system is destroyed, the normal data information exchange cannot be performed , it is called a Type II fault.

The specific implementation method of the fault detection scheme based on information interaction is given below:

As can be seen from the above discussion, the present invention is based on the fault definition of the control effect of the node, that is, once the actual operating state of the node does not meet the control requirements, it is considered to be faulty. From this, it is natural to think of a fault detection scheme: detect the operating state of the system, if there is an error r between the theoretical operating state and the actual operating state of a certain node, and the error exceeds a certain range, it is concluded that the node is faulty.

Since the node model considered in this invention is a single integrator model, the output is reflected in the continuous displacement z _i ^k+1 -z _i ^k of the node, or the velocity output u _i ^k of the node, so we use u _i ^k As the performance index for calculating the residual signal of the system: r _i ^k =u _i ^r,k -u _i ^a,k , where u _i ^r,k ∈R ² is the system obtained by the control protocol P within the time step k The theoretical motion state, u _i ^a,k ∈ R ² is the actual motion state of the system obtained through real-time measurement, satisfying:

u _i ^r,k ＝u _i ^k ＝P _i (z _i ^k ,I _i ^k )(6)

u _i ^a,k ＝h(z _i ^k+1 ,z _i ^k )(7)

If the state of node i is continuous, then z _i ^k+1 and z _i ^k can be measured by the built-in sensor of the node, and h can be obtained using a simple first-order differential equation (z _i ^{k+1 -} z _i ^k )/[(k+1)T-kT].

The fault node is now defined as follows:

Definition 1: For a node i using a single integrator model, if it satisfies the conditions described in formula (8), it is called a faulty node.

||r _i ^k ||＝||u _i ^r,k -u _i ^a,k ||＞χ(||u _i ^r,k ||,δ)(8)

Among them, χ(||u _i ^r,k ||,δ) is called the threshold function, and its value depends on the size of the input signal ||u _i ^r,k || and the disturbance δ. Generally, χ(||u _i ^r,k ||,δ)=γ ₁ +γ ₂ ||u _i ^r,k ||, where the constant γ ₁ depends on the disturbance δ, and the time variable γ ₂ ||u _i ^r,k || depends on the instantaneous input of the node.

For systems that may contain faulty nodes, our control goal is: the system can complete the original task, while detecting and isolating the faulty nodes. Since the faulty node cannot participate in the original task, it is stipulated that if all the nodes that have not failed have completed the original task, it is considered that the entire system has completed the expected goal.

As shown in Figure 2, the present invention proposes the following fault detection scheme:

Assuming that node j is performing fault detection on node i at this time, within time step k, through information interaction content 1 and 2 described above, node j can obtain the state z _i ^k of node i at this moment and all its adjacent nodes information Since the information exchange protocol is homogeneous, node j can use its own control protocol and I _i ^k to obtain the theoretical control quantity u _i ^r,k of node i. In the next time step k+1, similarly, node j can obtain z _i ^k+1 and u _i ^r,k+1 , and use formula (7) to obtain u _i ^a,k . At this point, node j can use formula (8) to judge whether node i fails within time step k.

Intuitively speaking, the fault detection scheme is to get the theoretical motion state of the target node by obtaining the information of the adjacent nodes of the target node and using the homogeneous information interaction protocol, and compare it with the detected actual motion state. If the error exceeds A certain amplitude, it is determined that the node is faulty.

The fault detection scheme mentioned above obtains the detection result of a single node, which is greatly affected by random factors, and the reliability of the result is not high. For example, there are often delays and data loss in the process of information interaction between nodes. If a node does not receive the information of an adjacent node in time, or the received information is incomplete, it is very likely that the adjacent node will be misjudged. Taking actions such as isolating the faulty node, these operations will be regarded as faulty by its adjacent nodes, causing the node itself to be detected as a faulty node. If this continues, a large number of normal nodes in the system will be isolated due to misoperation, resulting in a serious waste of resources, and may even lead to failure to achieve the global goal. Considering these situations, the present invention proposes to use a Gossip algorithm to process information on the detection results of each node, thereby improving the accuracy of the detection results. The schematic diagram of the scheme is shown in Figure 3, and the specific implementation scheme is as follows:

Taking node i as an example, first of all, within the time step k, node i performs fault diagnosis alone, and uses formula (8) and formula (4) to obtain its diagnosis results for all adjacent nodes At the same time, all adjacent nodes of i are doing the same operation. Next, as shown in information interaction content 3, node i transmits the diagnostic results of the adjacent nodes to its adjacent nodes respectively, and at the same time receives the diagnostic results of the adjacent nodes for i Finally, the node i takes the comprehensive detection results of the adjacent nodes to itself Send it to all its adjacent nodes, and receive the comprehensive detection results of its adjacent nodes at the same time. In this way, using formula (5), node i can calculate the detection result of the system to its adjacent nodes By setting up the parameter Q _con , when , it can be judged that node j is faulty. Generally speaking, the parameter Q _con is a constant on the (0,1] interval, and its value is affected by factors such as the accuracy of node execution, the intensity of environmental interference, and the quality of information interaction between nodes. The larger the value of Q _con , the higher the reliability of the system's detection results for faults, but the greater the probability of missed detection. Therefore, the value of Q _con should be selected appropriately according to the actual system.

The specific implementation methods of the fault isolation and repair plan are given below:

After the fault detection is completed, the system often needs to isolate the faulty node to eliminate its influence on the remaining normal nodes. In addition, in a distributed multi-agent system, since the fault detection task of each node is carried out independently, it is very likely that the fault node is detected by its adjacent nodes at different times. Moreover, because the nodes need to apply the gossip propagation (Gossip) algorithm to process the detection results, this will also bring a certain degree of delay. Therefore, the time when a node fails is usually inconsistent with the time when the node is diagnosed as a faulty node by the system. During this time, the faulty nodes still act on the system, causing deviations in the final control results. In order to eliminate this effect, this chapter will propose a fault recovery algorithm that compensates the control amount of the system by applying an external signal.

From the above discussion, it can be found that the time when each node in the system detects the fault node may be inconsistent. If each node isolates and repairs the fault at the moment when it detects the fault node, the isolation repair operation is very likely Take the same action for it that is diagnosed as a failure by its neighbors. This situation will spread step by step, eventually leading to the collapse of the entire system. Therefore, it is necessary to specify a time for each node to perform operations on faults uniformly. We introduce a new parameter: fault detection and repair period, denoted as T _p =p*T. where the constant p*∈Z ⁺ , T is the sampling time. In each cycle T _p , the node performs fault detection and information processing in the k∈[k ^* T _p +T,(k ^* +1)T _p -T] time period, and k=(k ^* +1) T _p , k∈Z ⁺ time period to isolate and repair the faulty nodes. It is worth noting that since fault isolation and repair is an unconventional operation, it is likely to be detected as a fault by its adjacent nodes. Therefore, within the time period k=(k*+1)T _p , k∈Z ⁺ , The fault detection function of each node should be shielded temporarily.

The isolation scheme for the faulty node is given below:

Fault isolation refers to removing the control quantity of the faulty node from its adjacent nodes, and blocking the information exchange channel for the faulty node to receive the information of the adjacent nodes, so as to eliminate the influence of the faulty node. It is easy to think that when a node detects that its adjacent nodes have failed, it only needs to remove the node from the set of adjacent nodes and stop sending its own status information to the node to complete the isolation work. Note that there is no interruption of sending the information of its own adjacent nodes to the faulty node, because the two are not directly connected, and sending this information has little impact on the system. Termination of self-information transmission is mainly due to information security considerations, because the cause of the failure is unknown, if the node has been controlled by the hostile party, continuing to send data may be maliciously used by the node, thereby affecting its own control. However, the termination of information transmission will also bring about a problem, that is, the faulty node cannot receive the information of the adjacent node, it will judge its adjacent node as a faulty node and also stop the transmission of information, so that when the faulty node terminates the transmission to all adjacent nodes information, it will be completely invisible to the system, and the harm caused by its movement to the system will also be completely unavoidable, which will lead to many situations that are unfavorable to the system, such as collisions between nodes, and system topology changes. Destructive damage, etc. In order to avoid the above situation, the following operations are defined for the node:

Operation 1: When a node cannot receive the information of a certain adjacent node, use the received evaluation of the adjacent node to obtain the self-diagnosis result of the system by using formula (5). If the result exceeds a certain amplitude Q _con , it can be judged that it has a fault. At this time, its own fault detection function is blocked, but data information is still sent to adjacent nodes, but the information will no longer contain the parts listed in content 3. .

Through the above operations, the faulty node will not evaluate the remaining normal nodes, but its movement is still visible to the system, so that the system can react to its destructive activities early. In addition, the retained information exchange content 1 and 2 will make the faulty node become a relay node for information transmission, avoiding the destructive damage to the system topology caused by the fault isolation operation. The specific discussion of this part will be given below.

The fault repair plan is given below:

The purpose of fault repair is: if the faulty node fails to be isolated in time after the fault occurs, and still has a certain impact on the system, fault repair is adopted to eliminate this part of the impact. For fault recovery, an intuitive idea is to separate the control quantity of the faulty node, invert it and add it back to the original control quantity, so as to offset the influence of the faulty node. However, this scheme requires the control protocol of the nodes to have a linearly superimposed form, so that the control quantities of the faulty nodes can be separated. But for complex multi-agent systems, it often happens that the control protocol is non-linear and cannot be superimposed. Therefore, the present invention will start from the actual control effect of the node, define a new compensation amount calculation and fault repair scheme, and at the same time provide proof of the feasibility of the scheme.

For any node i in the multi-agent system, if it does not fail, the relationship between the expected output u _i and the actual output y _i of the node must satisfy the relationship shown in Figure 4. Here, the execution error of the actuator itself is ignored.

In the figure, u _imax represents the maximum expected output of node i, and y _imax represents the actual output upper limit of the node. For the actual node, when the expected output exceeds the actual output upper limit that the node can achieve, the node can only operate under the maximum output y _imax , which will cause some control variables to fail to appear in the output. Therefore, in order to prevent nodes from diagnosing this saturation characteristic as a fault, it is necessary to limit the control quantity, that is, In this way, there will be the following relationship between the expected output and the actual output:

y _i =a·u _i (9)

The constant a∈R ⁺ is the output gain of the system, and it is assumed that a=1 in the present invention.

It should be noted that u _i here is only a mathematical description of the expected output of the node, not the real output of the controller. For the actual actuator, its nonlinearity is complex and diverse, not just the saturation characteristic, the controller also needs to adopt corresponding control algorithms, such as PID control, fuzzy control, etc. to ensure that the nodes can perform output tasks normally. In addition, the actual node output y _i mentioned above refers to the idealized steady-state response result of the node, and its dynamic response characteristics are not within the scope of the present invention.

The following operations are defined for the calculation of the compensation amount:

Operation 2: When node i detects that its adjacent node j has failed at k=T _i , use formula (3) to calculate its own control quantity without the influence of j Simultaneous computing The value of , and its inversion and accumulation, until the next fault isolation and repair time k=T _ip is reached.

From operation 2, we can see that for node i, the amount of compensation that needs to be applied to eliminate the influence of faulty node j is:

${\mathrm{<span\; class="notranslate">\; u</span>}}_{{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; comp</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\overset{{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; ip</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; (</span>{{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \backslash </span>\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 10</span>}<span\; class="notranslate">\; )</span>$

Since there is a maximum output value in the actual system, the compensation value obtained by formula (10) cannot be simply added to the original control value, and it is necessary to consider whether the addition of the compensation value will cause the original control value to exceed the limit value and appear Circumstances of Inadequate Compensation. To do this define the following operations:

Operation 3: When k=(k*+1)T _p , k∈Z ⁺ , if node i has confirmed that node j has failed, add the compensation amount calculated by formula (10) to the original control amount, and at the same time Detect whether the control amount at this time exceeds the limit value, if so, reassign the part exceeding the limit value to the compensation amount, and continue to compensate at the next isolation and repair time; if not, clear the compensation amount and repair job done.

It can be seen from operation 3 that after several fault detection and repair cycles, the compensation amount will be completely added to the control amount, and the repair work on the faulty node is completed at this time.

The network connectivity maintenance scheme based on two-hop information is given below:

From the above analysis, it can be seen that for the nodes where the type II failure occurs, that is, the information exchange system is destroyed, its impact on network connectivity will be irreparable under the existing topology. However, if the node still retains the normal information interaction function, it can be regarded as a relay node for information transmission, and a two-hop information transmission path can be established to repair the network topology that may be destroyed. The specific implementation plan is as follows:

As shown in Figure 1, assuming that node 3 fails, if only isolation and repair operations are taken on it, there will be no information interaction between nodes 1, 2 and nodes 4, 5, 6, and 7, and the connectivity of the original graph will be destroyed , the system will not be able to cooperate to complete the control objectives. Considering the regulations on the faulty node in operation 1, it can be seen that the adjacent nodes of the faulty node can still receive the information of content 1 and 2 from the faulty node, and the content 2 will contain the complete information of its adjacent nodes. Therefore, it can be imagined that the faulty node is used as a relay node for information transmission, and a virtual information transmission path is established between its two non-adjacent adjacent nodes, so as to maintain the connectivity of the original graph. Define the following operations:

Operation 4: If node i detects that its adjacent node j is faulty, after the fault isolation operation is completed, check the information of its adjacent nodes in the information interaction content 2 of node j and like and order $z_l^k\&Element;I_i^k,Q_l^k\&Element;\{m\&Element;N_i^k|Q_m^k\}.$

After operation 4, Figure 1 will become the topology shown in Figure 5, where node 3 fails. The dotted line represents the two-hop information transmission path with node 3 as the relay node.

It can be seen from operation 4 that if there is no direct information interaction between the two adjacent nodes of the faulty node, after the above operations, a virtual information transmission channel will be established between the two nodes, making the two nodes a theoretically significant The adjacent nodes on the graph can ensure that the connectivity of the original graph is not destroyed.

Now aiming at the consistency problem in multi-agent control, the feasibility of the fault detection, isolation and repair scheme proposed by the present invention is verified.

First assume that for all roads The values of a and _ij are all equal, then the control protocol P is homogeneous at this time, that is, the control quantity generation methods of all nodes are exactly the same. In this way, by using information exchange content 1-3, the node receives the information of its adjacent nodes, and uses its own control protocol to diagnose the adjacent nodes, and then uses the gossip propagation (Gossip) algorithm for information processing to complete the fault detection . Afterwards, by using operations 1-4, the tasks of isolating and repairing faults and maintaining network connectivity can be normally completed. There are no special conditions in the whole process to limit the application of the fault handling scheme, so it can be proved that if the weight values of the roads in the multi-agent network are equal, the fault handling scheme can be applied to the consistency problem.

However, there may be such a situation in the actual multi-agent system: in different information interaction networks, the weight values of different roads are also different, that is to say, the control protocol P is no longer homogeneous. At this time, some modifications need to be made to the content of information interaction 2. Node i no longer transmits the state information I _i ^k of its adjacent nodes, but transmits new information where {i ₁ , . . . , i _p }=N _i ^k . In this way, the control protocol between nodes no longer contains non-homogeneous items a _ij , and it can still be regarded as a homogeneous control protocol. Therefore, referring to the above analysis, we can see that the original scheme is still applicable.

The software simulation results are given below:

As shown in Figure 6-8, these three figures show the results of the consensus control simulation of 8 multi-agents using MATLAB. Figure 6 (left), Figure 7 (left), and Figure 8 (left) are the results of fault detection, isolation and repair of fault 1, fault 2, and fault 3, respectively, while the corresponding graphs 6 (right), 7 ( Right) and Figure 8 (right) are the corresponding results when no fault handling operation is taken. It can be seen from the figure that if the fault is not dealt with, the remaining nodes will be taken away from the expected target by the faulty node over time, which will lead to the failure of the entire system control target to be realized. After adopting the fault handling scheme, the faulty node will no longer affect the remaining nodes, and the remaining normal nodes can still complete the consistency control as expected.

What has been described above is only a preferred embodiment of the present invention, and the present invention is not limited to the above-mentioned embodiment, and all local changes, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in within the protection scope of the present invention.

Claims (2)

1. A distributed real-time fault detection and compensation control method based on information interaction, characterized in that, comprising the steps: Step 1. System and fault modeling: the modeling includes a node dynamics model, an information interaction model, and a typical fault model; wherein the node dynamics model adopts a single integrator model, and the motion state of the node is described by a first-order differential equation; the information The interaction model is described by an undirected graph, that is, two-way communication between nodes is possible, and each independent agent uses this to perform information interaction and complete system control tasks; the typical fault model includes the fault types that often occur in real agents; Step 2. Multi-agent real-time fault detection based on information interaction: select relevant state variables from the expression of the node dynamic model described in step 1 as the description of the node's operating state; by setting the threshold function, the node's The running status is divided to distinguish between normal nodes and faulty nodes; at the same time, a single node obtains the status information of its adjacent nodes by means of the information interaction model described in step 1, and detects whether it is faulty through a detection algorithm to form a detection result of a single node; Step 3. Information integration and processing based on Gossip algorithm: Due to the existence of problems such as communication packet loss and time lag, the single-node detection results in step 2 are greatly affected by the environment, and the reliability is not high; therefore, using the Gossip algorithm, the single-node The node detection results are integrated to obtain comprehensive detection results with higher reliability, and this is used as the final judgment basis for the node operating status to distinguish normal nodes from faulty nodes; Step 4. Calculation and application of compensation for control quantities: If a faulty node is detected, the faulty node will be isolated through corresponding operations, and at the same time, based on the influence of the faulty node on the control quantity of its adjacent nodes, a relevant calculation scheme will be designed to obtain compensation The value of the quantity is added to the original control quantity to offset the impact of the faulty node on the system; Step 5. Design connectivity maintenance based on two-hop information: starting from the information interaction model, analyze the communication content between faulty nodes, and establish a virtual information transmission path by using the two-hop information to ensure that the fault handling scheme will not affect the normal operation of the system. 2. A distributed real-time fault detection and compensation control method based on information interaction according to claim 1, wherein said fault types include destructive faults, out-of-control faults and interference faults.