Fault monitoring method for complex mechanical systems with functional cross-linking based on heterogeneous attribute networks
By constructing a heterogeneous attribute network model and combining a generalized linear model, state space model and Kalman filtering algorithm, the fault diagnosis problems caused by functional crosslinking and state fluctuations in complex mechanical systems are solved, and efficient fault monitoring and diagnosis are achieved. Improves the accuracy and timeliness of fault monitoring.
Patent Information
- Application Number
- CN202111197980.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-14
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2041-10-14
AI Technical Summary
Existing mechanical fault diagnosis methods are difficult to deal with fault diagnosis when facing complex mechanical systems, especially due to the difficulty of diagnosis caused by random fluctuations in functional crosslinking and state parameters, as well as poor interpretability of intelligent diagnostic systems and difficulty in obtaining data.
Using a method based on heterogeneous attribute network, a heterogeneous attribute network model is constructed for functional crosslinking systems, functional crosslinking degree and part importance are calculated, and the generalized linear model and state space model are combined, and Kalman filtering algorithm and statistical process control monitoring are used. Pearson residuals distinguish normal and abnormal fluctuations, improving the timeliness and accuracy of fault monitoring.
It effectively improves the accuracy and timeliness of fault monitoring of complex mechanical systems, can accurately judge abnormalities in the random fluctuations of system status, supports fault diagnosis of functional crosslinking complex mechanical systems, and provides the ability to measure functional crosslinking and position fault locations. Reduces failure loss.
Smart Images

Figure CN113962038B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a method for monitoring faults of a functionally cross-linked complex mechanical system based on a heterogeneous attribute network, and belongs to the field of mechanical manufacturing. Background Art
[0002] Mechanical fault diagnosis technology is a scientific technology that monitors, diagnoses and predicts the status of continuously running mechanical equipment and possible faults, and ensures the smooth and safe operation of mechanical equipment. It is of great significance to ensure the smooth operation of equipment, ensure the safety of personnel and avoid economic losses. Existing mechanical system fault diagnosis methods can be roughly summarized into three categories: (1) Fault diagnosis based on mechanism model, that is, judging the operating status of the equipment by comparing the measured data of equipment operation with the data calculated by the mechanism model; (2) Fault diagnosis based on signal processing, which is more common in rotating mechanical systems, that is, by collecting and analyzing dynamic operation signals such as vibration signals, possible signs of system failure are identified; (3) Fault diagnosis based on intelligent data algorithms, that is, using a large amount of equipment operation history data, with the help of intelligent algorithms such as machine learning and deep learning, to judge the operating status of the system.
[0003] It is undeniable that the above fault diagnosis methods have been applied and promoted in many fields and have achieved good results. However, with the continuous improvement of the complexity of mechanical systems, many fault diagnosis problems that cannot be handled by existing methods have emerged, including: (1) It is difficult to construct an accurate mechanism model of complex mechanical systems, making it difficult to carry out fault diagnosis methods based on mechanism models; (2) The intelligent diagnosis system is weak, and the intelligent data algorithm is a "black box" between the fault influencing factors and the fault signal, so the diagnosis results cannot be explained, and the historical data required by the algorithm is difficult to obtain in some systems; (3) Complex mechanical systems have diverse functions, and there are cross-links and couplings between different functions. The abnormality of a function will cause a series of fluctuations in the status of other functions. Most existing methods do not consider the cross-linking of functions; (4) The state parameters of complex mechanical systems during normal operation are not stable constant values, but will fluctuate randomly within a certain range. How to judge the abnormality of the system state from the continuously randomly fluctuating state monitoring values is a new challenge for the current complex mechanical system fault monitoring and diagnosis. Summary of the invention
[0004] In view of the limitations of existing fault monitoring and diagnosis methods when facing functionally cross-linked complex mechanical systems, the purpose of the present invention is to provide a method for monitoring faults in functionally cross-linked complex mechanical systems based on heterogeneous attribute networks, taking into account the cross-linking of complex mechanical system functions, effectively distinguishing normal fluctuations from abnormal fluctuations in the random fluctuations of system states, thereby improving the timeliness and accuracy of fault monitoring. The monitoring results of the present invention can be applied to the fault diagnosis of functionally cross-linked complex mechanical systems, thereby solving related engineering and technical problems.
[0005] The present invention discloses a method for monitoring faults of a complex mechanical system with functional cross-linking based on a heterogeneous attribute network, comprising the following steps:
[0006] Step 1: Analyze the complex mechanical system and construct a heterogeneous attribute network model of the functional cross-linking system; in the heterogeneous attribute network model of the functional cross-linking system, calculate the functional cross-linking degree and part importance according to the number of nodes and edges; according to the characteristics of the complex mechanical system, determine the dynamic attributes and static attributes in the heterogeneous attribute network model.
[0007] Step 1-1: Analyze complex mechanical systems and construct a functionally cross-linked heterogeneous property network model.
[0008] The parts or components of complex mechanical systems often have coupling or cross-linking relationships, which can be represented by attribute information networks with the help of graph theory. The information network is defined as a mapping function with object type V→A and relationship type mapping function ψ: E→R directed or undirected graph G=(V,E). If the number of object types |A|>1 or the number of relationship types |R|>1 in the information network, it is called a heterogeneous information network, otherwise it is a homogeneous information network. The node types in complex mechanical systems are not the same, so the use of heterogeneous attribute information networks can more comprehensively characterize system information.
[0009] Most complex mechanical systems have multiple functions, and each function may use different parts or units. Therefore, for each function of the mechanical system, a heterogeneous attribute network model can be constructed as shown in formula (1), where k is the number of functions of the mechanical system:
[0010] G=(G 1 ,G 2 …,G k ) (1)
[0011] Step 1-2: Calculate the functional cross-linking degree and part importance according to the number of nodes and edges in the heterogeneous attribute network model of the functional cross-linking system.
[0012] According to the number of shared nodes and edges in the heterogeneous attribute network model of each function, the functional cross-linking degree that represents the degree of dependence between two functions is calculated. The total number of nodes in the system is N, the total number of edges is M, and the number of shared nodes between two functions is n ij , the number of shared edges is m ij , then the cross-linking degree of any two functions can be expressed as:
[0013]
[0014] The importance of each part or node in a complex mechanical system is also different. According to the number of connecting edges of each node, the importance of the part is obtained as follows, where mi is the number of connected edges of a node.
[0015]
[0016] Step 1-3: Determine the dynamic and static properties of the heterogeneous attribute network model based on the characteristics of the complex mechanical system.
[0017] The edge E in the network model constructed in step 1-1 expresses the connection properties between nodes, which can be divided into dynamic properties and static properties. Static properties are the inherent property relationships between two nodes, including the node type of components and the service life of components; dynamic properties are the property relationships between two nodes that change over time, including the flow value between hydraulic components in the hydraulic system and the voltage value between components in the electrical system. ij,t and w ij,t represents the static and dynamic properties of the network model at time t, and for the convenience of calculation, w ij,t Vectorized to w t , x ij,t Matrix X t .
[0018] Step 2: Construct a generalized linear model GLM to connect the static and dynamic attribute distribution parameters of the heterogeneous network; use the state space model SSM to transform the network parameter β t The changes in the network model are incorporated into GLM, making GLM's description of the network model more accurate.
[0019] Step 2-1: Construct a generalized linear model GLM to connect the static and dynamic attribute distribution parameters of the heterogeneous network.
[0020] The static attribute matrix X of the mechanical system heterogeneous attribute network model obtained in step 1 is t and the dynamic attribute vector w t , due to w t Each element in is often a random variable that obeys an exponential family distribution, so the generalized linear model GLM shown in formula (4) is used to connect the dynamic and static properties of the network.
[0021]
[0022] In the above formula, β t It is used to connect dynamic attributes and static attributes. It is the model parameter of GLM, which is called the network parameter at time t. The function g is called the connection function. ij,t The type that obeys the exponential family distribution, the function g has many forms, such as when w ij,t When it obeys the binomial distribution, the function g is in the form of:
[0023]
[0024] When ij,t When it obeys the normal distribution, the function g is in the form of:
[0025] g(x)=x (6)
[0026] In formula (4), θ t w ij,t The distribution parameter of the function f is w ij,t The probability density function of .
[0027] Step 2-2: Using the state space model SSM, transform β t The changes in the network model are incorporated into GLM, making GLM's description of the network model more accurate.
[0028] In the GLM model constructed in step 2-1, the parameter β t It is the only quantitative representation of the network model at time t. When the external conditions remain unchanged, β t It should be stable at a certain value or within a certain range. However, in the operation of complex mechanical systems, as working conditions, environment and other conditions change, β t will develop and change according to a certain rule, and the network parameters at time t+1 will be affected by the system state at time t. t The changes of β are incorporated into GLM, making it a dynamic network flow model. t The SSM representation is as follows:
[0029] β t =Fβ t-1 +ε t (7)
[0030] Where F is the state transfer matrix of network parameters, ε t is the SSM model error at time t, which conforms to the Gaussian distribution with a mean of 0.
[0031] Step 3: Use historical data to determine the state space model SSM parameters; in order to improve the evaluation accuracy, as a preferred method, the least squares and Kalman filter algorithms are combined to determine the dynamic parameters β in the network model t .
[0032] Step 3-1: Determination of SSM model parameters.
[0033] The state transfer matrix F in the SSM model will be used under normal operation of the system, β t The historical data are used to estimate the
[0034] Step 3-2: Combine the least squares and Kalman filter algorithms to determine the dynamic parameter β in the network model t .
[0035] From step 2, we can know that the network parameter β t The accuracy of will directly affect the efficiency and accuracy of fault diagnosis of complex mechanical systems. The state transfer matrix F in the state transfer model SSM determined in step 3-1 is equivalent to β t The prediction model will obtain w during the operation of the system t Therefore, with the help of Kalman filtering principle, using the measured and predicted values of network parameters, more accurate network parameter estimates can be obtained.
[0036] The prediction model of the Kalman filter algorithm is as shown in equations (8) to (9), β t The initial value of is estimated by the least squares method, where β t|t-1 β t The prior estimate of β t-1t-1 β t-1 The posterior estimate of t|t-1 is the covariance of the prior errors, P t-1t-1 is the error covariance at time t-1, Q t is t The covariance of .
[0037] β t|t-1 =Fβ t-1t-1 (8)
[0038] P t|t-1 =FP t-1t-1 F T +Q t (9)
[0039] The correction model of the Kalman filter algorithm is as shown in equations (10) to (12), where β t|t β t The posterior estimate of t|t-1 w t The prior estimate of K t is the Kalman gain, P t|t is the error covariance at time t, G t is the observation matrix, R t To measure noise.
[0040] β t|t =β t|t-1 +K t (w t -w t|t-1 ) (10)
[0041] P t|t =(IKt G t ) t|t-1 (11)
[0042]
[0043] Step 4: The network model dynamic parameters β determined by step 3 t and w t The Pearson residual is calculated based on the measured value, and the statistical process control (SPC) is used to monitor the Pearson residual. According to the change of the Pearson residual, the normal fluctuation and abnormal fluctuation can be effectively distinguished in the random fluctuation of the system state, so as to realize the fault monitoring of complex mechanical system and improve the timeliness and accuracy of fault monitoring.
[0044] Step 4-1: The network model dynamic parameters β determined by step 3 t and w t Pearson residuals are calculated for the measured values.
[0045] Step 3 calculates a more accurate estimate of the network parameters. With the help of formula (8), we can obtain the prior estimate of the network parameters and then obtain w t The prior estimate of is as follows:
[0046] w t|t-1 =g(X t β t|t-1 ) (13)
[0047] Thus, we can obtain the i-th edge w in the network model at time t+1 t The prior estimate of w t The Pearson residual between the measured values of is as follows, where MSE is the mean square error between the two.
[0048]
[0049] Take the average of the Pearson residuals of all edges and obtain the Pearson residual value of the entire network model:
[0050]
[0051] The state of complex mechanical systems can be monitored by real-time monitoring of the Pearson residual of the network model.
[0052] Step 4-2: Construct a cumulative sum control chart CUSUM and use the cumulative sum control chart CUSUM to monitor large fluctuations in the Pearson residuals.
[0053] The CUSUM control chart is a control chart that is more effective in controlling mean changes, and is often more capable of detecting larger fluctuations greater than one standard deviation σ. For the Pearson residuals of the network model at each moment, the cumulative sums of the upper and lower sides of the CUSUM control chart are:
[0054]
[0055] in is the cumulative sum on one side, is the cumulative sum of the lower side, and μ 0 is the mean of the monitored object, the mean of the Pearson residual is 0, and Y is the allowable deviation. or When the determination value H is exceeded, the system is determined to be abnormal.
[0056] Step 4-3: Construct an exponentially weighted moving average (EWMA) control chart and use the exponentially weighted moving average (EWMA) to monitor smaller fluctuations in the Pearson residuals.
[0057] The EWMA control chart is a control chart that has a good control effect on small fluctuations less than one standard deviation σ. The statistics of the EWMA control chart are:
[0058]
[0059] where z t is the control chart statistic, and z 0 =0, λ is the weight parameter.
[0060] The upper control limit (UCL) and lower control limit (LCL) of the EWMA control chart are:
[0061]
[0062] Where l is the control limit parameter and σ is the standard deviation of the control chart statistic.
[0063] Step 5: Use the CUSUM control chart and EWMA control chart constructed in step 4 to monitor the heterogeneous network Pearson residuals of each function of the complex mechanical system. When the Pearson residual exceeds the control limit of the control chart, it is determined that a certain function of the system at that moment is abnormal. Then, the functional cross-linking degree calculated in step 1 is used to determine the probability that the current functional abnormality will affect other functions of the complex mechanical system.
[0064] Step 6: Locate the fault location based on the monitoring results and Pearson residual value of step 4. When a function in the system is abnormal at time t, the Pearson residual of the function at time t exceeds the control limit of the control chart. The abnormality of the function is caused by the excessive Pearson residual of each edge in the heterogeneous network. By finding the edge with the largest Pearson residual in the heterogeneous network model, the possible fault location is located, and then control or maintenance measures are taken to avoid or reduce the fault loss.
[0065] Beneficial effects:
[0066] 1. The present invention discloses a method for monitoring faults of complex mechanical systems with functional cross-linking based on heterogeneous attribute networks. The method analyzes complex mechanical systems and constructs a heterogeneous attribute network model of the functional cross-linking system. In the heterogeneous attribute network model of the functional cross-linking system, the functional cross-linking degree and the importance of parts are calculated according to the number of nodes and edges. The dynamic attributes and static attributes in the heterogeneous attribute network model are determined according to the characteristics of the complex mechanical system. A generalized linear model GLM is constructed to connect the static attribute and dynamic attribute distribution parameters of the heterogeneous network. With the help of the state space model SSM, the network parameter β t The changes of are incorporated into GLM, making GLM more accurate in describing the network model. The state space model SSM parameters are determined using historical data; in order to improve the evaluation accuracy, the least squares and Kalman filter algorithms are combined to determine the dynamic parameters β in the network model as an optimal choice. t The dynamic parameters of the network model determined in step 3 are t Calculate the Pearson residual and use statistical process control (SPC) to monitor the Pearson residual. According to the change of the Pearson residual, effectively distinguish normal fluctuations from abnormal fluctuations in the random fluctuations of the system state, realize fault monitoring of complex mechanical systems, and improve the timeliness and accuracy of fault monitoring.
[0067] 2. The method for monitoring faults of functionally cross-linked complex mechanical systems based on heterogeneous attribute networks disclosed in the present invention utilizes the functional cross-linking degree of the complex mechanical system established in step 2 to achieve the measurement of the multifunctional cross-linking in the complex mechanical system, thereby providing support for fault diagnosis of the functionally cross-linked mechanical system.
[0068] 3. The present invention discloses a method for fault monitoring and diagnosis of a complex mechanical system with functional cross-linking based on a heterogeneous attribute network. On the basis of beneficial effects 1 and 2, the CUSUM control chart and EWMA control chart constructed in step 4 are used to monitor the heterogeneous network Pearson residuals of various functions of the complex mechanical system. When the Pearson residual exceeds the control limit of the control chart, it is determined that a certain function of the system at that moment is abnormal, and then the functional cross-linking degree calculated in step 1 is used to determine the probability that the current functional abnormality will affect other functions of the complex mechanical system.
[0069] 4. The present invention discloses a method for monitoring and diagnosing faults of a complex mechanical system with functional cross-linking based on a heterogeneous attribute network. Based on the beneficial effects 1 and 2, the fault location is located by the monitoring result and Pearson residual of step 4. When a function in the system is abnormal at time t, the possible fault location is located by finding the edge with the largest Pearson residual in the heterogeneous network model, and then control or maintenance measures are taken to avoid or reduce the fault loss. BRIEF DESCRIPTION OF THE DRAWINGS
[0070] Figure 1It is an overall flow chart of the method for monitoring faults of a complex mechanical system with functional cross-linking based on a heterogeneous attribute network disclosed in the present invention;
[0071] Figure 2 A hydraulic system diagram in an embodiment of the present invention;
[0072] Figure 3 This is a network model diagram of the heterogeneous attributes of various functions of the hydraulic system in an embodiment of the present invention, where: Figure 3 (a) is the overall topological diagram of the hydraulic system. Figure 3 (b) is the fast-forward functional network model diagram. Figure 3 (c) is a functional network model diagram. Figure 3 (d) is the diagram of the two-step functional network model. Figure 3 (e) is the network model diagram of the stay function. Figure 3 (f) is the network model diagram of fast rewind function;
[0073] Figure 4 This is an SPC control diagram of a process function in an embodiment of the present invention, wherein: Figure 4 (a) is the CUSUM control chart, Figure 4 (b) is the EWMA control chart. DETAILED DESCRIPTION
[0074] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:
[0075] Example 1
[0076] Mechanical hydraulic systems are widely used in the fields of engineering machinery, aerospace, weapons, ships, etc., and with the increasing number of hydraulic components, mechanical hydraulic systems can become very large and complex, which brings challenges to the fault diagnosis of the system. In the hydraulic system, various hydraulic components are connected to each other in a network-like layout. The liquid pressure value or flow value between the components fluctuates randomly within a certain range. A hydraulic system often has multiple working conditions (functions) and is interconnected with each other. The present invention uses the YT4543 power slide hydraulic system to further illustrate the above method by simulation.
[0077] Step 1: Analyze the complex mechanical system and construct a heterogeneous attribute network model of the functional cross-linking system; in the heterogeneous attribute network model of the functional cross-linking system, calculate the functional cross-linking degree and part importance according to the number of nodes and edges; according to the characteristics of the complex mechanical system, determine the dynamic attributes and static attributes in the heterogeneous attribute network model.
[0078] Step 1-1: Analyze complex mechanical systems and construct a functionally cross-linked heterogeneous property network model.
[0079] The hydraulic system diagram of YT4543 power slide is as attached. Figure 2The system has five working conditions, namely fast forward, one working forward, two working forwards, stop and fast reverse. The hydraulic components used in each function are different. The heterogeneous attribute network model diagram of each function is constructed as shown in the attached figure. Figure 3 , each node shape represents a type of hydraulic component. This embodiment will take a fault monitoring and diagnosis process of a working function as an example to further illustrate the invention method.
[0080] Step 1-2: Calculate the functional cross-linking degree and part importance according to the number of nodes and edges in the heterogeneous attribute network model of the functional cross-linking system.
[0081] In the attached Figure 3 Based on the above, the functional cross-linking degree and the component importance are calculated using formula (2) and formula (3) respectively. The cross-linking degrees of the five functions calculated using formula (2) are shown in Table 1:
[0082] Table 1 Functional cross-linking degree of hydraulic system of YT4543 power slide
[0083]
[0084] Step 1-3: Determine the dynamic and static properties of the heterogeneous attribute network model based on the characteristics of the complex mechanical system.
[0085] In this embodiment, the dynamic attribute w of the hydraulic system network model t is the liquid pressure value between node elements, and this value obeys normal distribution. The static property X t For each hydraulic component type.
[0086] Step 2: Construct a generalized linear model GLM to connect the static and dynamic attribute distribution parameters of the heterogeneous network; use the state space model SSM to transform the network parameter β t The changes in the network model are incorporated into GLM, making GLM's description of the network model more accurate.
[0087] Step 2-1: Construct a generalized linear model GLM to connect the static and dynamic attribute distribution parameters of the heterogeneous network.
[0088] According to formula (4) and the dynamic properties w of the functional heterogeneous network of the hydraulic system determined in step 3 t and static property X t , determine the generalized linear model GLM of the hydraulic system. Since w ij,t It obeys the normal distribution, so the connection function g is calculated using formula (6).
[0089] Step 2-2: Using the state space model SSM, transform β t The changes in the network model are incorporated into GLM, making GLM's description of the network model more accurate.
[0090] Use formula (7) to construct the network parameter β t The state space model SSM makes the heterogeneous attribute network model in step 1 become a dynamic network flow model, and the parameters of SSM will be calculated in step 3.
[0091] Step 3: Use historical data to determine the state space model SSM parameters; in order to improve the evaluation accuracy, as a preferred method, the least squares and Kalman filter algorithms are combined to determine the dynamic parameters β in the network model t .
[0092] Step 3-1: Determination of SSM model parameters.
[0093] In the state space model SSM of formula (7), let F be the unit matrix to simulate the random walk process of parameter β, and the error ε t The mean of the covariance matrix is 0, t For a diagonal matrix, take Q 11 =1×10 -7 , other elements are 1×10 -9 , thus determining the state transition model of the hydraulic system.
[0094] Step 3-2: Combine the least squares and Kalman filter algorithms to determine the dynamic parameter β in the network model t .
[0095] Formulas (8) to (12) are used to construct the prediction model and correction model of the Kalman filter algorithm, and the network parameters β of each function at each time are recursively calculated. t Estimated value. Assume that at t = 0, the initial pressure value between the components in the network model of the working function system is w 0 =[3.00; 2.98; 2.97; 0.43; 0.42; 2.64; 2.63; 0.41], in MPa, the GLM at this moment is fitted by the least squares method and β is obtained 0 Since this embodiment illustrates the invention content by means of simulation, the subsequent dynamic network parameter β t It will be directly obtained through the state transition model of the hydraulic system.
[0096] Step 4: The network model dynamic parameters β determined by step 3 t and w t The Pearson residual is calculated based on the measured value, and the statistical process control (SPC) is used to monitor the Pearson residual. According to the change of the Pearson residual, the normal fluctuation and abnormal fluctuation can be effectively distinguished in the random fluctuation of the system state, so as to realize the fault monitoring of complex mechanical system and improve the timeliness and accuracy of fault monitoring.
[0097] Step 4-1: The network model dynamic parameters β determined by step 3 t Calculate Pearson residuals
[0098] When obtaining the measured value w of the dynamic attribute at each moment t and the network parameter prediction value β t Then, equations (13) to (15) are used to calculate the Pearson residuals at each moment of the functional network model.
[0099] Step 4-2: Construct a cumulative sum control chart CUSUM and use the cumulative sum control chart CUSUM to monitor large fluctuations in the Pearson residuals.
[0100] Construct a CUSUM control chart, where the allowable deviation Y = 0.5σ and the judgment value H = 5σ.
[0101] Step 4-3: Construct an exponentially weighted moving average (EWMA) control chart and use the exponentially weighted moving average (EWMA) to monitor smaller fluctuations in the Pearson residuals.
[0102] Construct the EWMA control chart and take l=2.814, λ=0.1.
[0103] Step 5: Use the CUSUM control chart and EWMA control chart constructed in step 4 to monitor the heterogeneous network Pearson residuals of each function of the complex mechanical system. When the Pearson residual exceeds the control limit of the control chart, it is determined that a certain function of the system at that moment is abnormal. Then, the functional cross-linking degree calculated in step 1 is used to determine the probability that the current functional abnormality will affect other functions of the complex mechanical system.
[0104] The constructed SSM model is used to obtain a continuous network parameter sequence. In the present invention, the 15th element of β is given 0.5σ at time t=120. 15 The small disturbance of the fault is injected into the SPC control diagram as shown in the attached figure. Figure 4 , including Figure 4 (a) is the CUSUM control chart, attached Figure 4 (b) is the EWMA control chart. It can be seen that at t=138, the EWMA control chart identified a system anomaly, while the CUSUM control chart did not find any anomalies. This reflects the EWMA control chart's ability to detect and control small fluctuations.
[0105] Step 6: Locate the fault location based on the monitoring results and Pearson residuals of step 4. When a function in the system is abnormal at time t, the Pearson residual of the function at time t exceeds the control limit of the control chart. This is because the Pearson residuals of each edge in the heterogeneous network are too large. By finding the edge with the largest Pearson residual in the heterogeneous network model, the possible fault location can be located, and then control or maintenance measures can be taken to avoid or reduce fault losses.
[0106] Finally, by finding the edge with the largest Pearson residual in the functional network model at time t=138, the possible fault location is located, and other functions that share this edge and have a large degree of interconnection with this function are most likely to also fail.
[0107] The specific description above further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for monitoring faults in complex mechanical systems based on functional cross-linking of heterogeneous attribute networks, characterized by: The following steps are included: Step 1: Analyze the complex mechanical system and construct a heterogeneous attribute network model of the functional cross-linking system; in the heterogeneous attribute network model of the functional cross-linking system, calculate the functional cross-linking degree and the importance of parts according to the number of nodes and edges; Determine the dynamic and static properties in heterogeneous attribute network models based on the characteristics of complex mechanical systems; Step 2: Construct a generalized linear model GLM to connect the static and dynamic attribute distribution parameters of the heterogeneous network; With the help of the state space model SSM, the network parameter β t The changes of are incorporated into GLM, making GLM's description of the network model more accurate; Step 2 is implemented as follows: Step 2-1: Construct a generalized linear model GLM to connect the static and dynamic attribute distribution parameters of the heterogeneous network; The static attribute matrix X of the mechanical system heterogeneous attribute network model obtained in step 1 is t and the dynamic attribute vector w t , due to w t Each element in is often a random variable that obeys an exponential family distribution, so the generalized linear model GLM shown in formula (4) is used to connect the dynamic and static properties of the network; In the above formula, β t It is used to connect dynamic attributes and static attributes. It is the model parameter of GLM, which is called the network parameter at time t. The function g is called the connection function. ij,t The type that obeys the exponential family distribution, the function g has many forms, such as when w ij,t When it obeys the binomial distribution, the function g is in the form of: When ij,t When it obeys the normal distribution, the function g is in the form of: g(x)=x(6) In formula (4), θ t w ij,t The distribution parameter of the function f is w ij,t The probability density function of Step 2-2: Using the state space model SSM, transform β t The changes of are incorporated into GLM, making GLM's description of the network model more accurate; In the GLM model constructed in step 2-1, the parameter β t It is the only quantitative representation of the network model at time t. When the external conditions remain unchanged, β t It should be stable at a certain value or within a certain range; however, in the operation of complex mechanical systems, as working conditions, environment and other conditions change, β t will develop and change according to a certain rule, and the network parameters at time t+1 will be affected by the system state at time t; with the help of the state space model SSM, β t The changes of β are incorporated into GLM, making it a dynamic network flow model; t The SSM representation is as follows: b t =Fβ t-1 +e t (7) Where F is the state transfer matrix of network parameters, ε t is the SSM model error at time t, which conforms to the Gaussian distribution with a mean of 0; Step 3: Determine the state space model SSM parameter F using historical data; Step 4: The network model dynamic parameters β determined by step 3 t and w t The measured value is used to calculate the Pearson residual, and the statistical process control (SPC) is used to monitor the Pearson residual. According to the change of the Pearson residual, the normal fluctuation and abnormal fluctuation can be effectively distinguished in the random fluctuation of the system state, so as to realize the fault monitoring of complex mechanical systems and improve the timeliness and accuracy of fault monitoring. The method for monitoring faults of a complex mechanical system with functional cross-linking based on a heterogeneous attribute network also includes the following steps: Step 5: Use the CUSUM control chart and EWMA control chart constructed in step 4 to monitor the heterogeneous network Pearson residuals of various functions of the complex mechanical system. When the Pearson residual exceeds the control limit of the control chart, it is determined that a certain function of the system at that moment is abnormal. Then, the functional cross-linking degree calculated in step 1 is used to determine the probability that the current functional abnormality will affect other functions of the complex mechanical system. Step 6: Locate the fault location based on the monitoring results of step 4 and the Pearson residual value; when a function in the system is abnormal at time t, the Pearson residual of the function at time t exceeds the control limit of the control chart. The abnormality of the function is caused by the excessive Pearson residuals of each edge in the heterogeneous network. By finding the edge with the largest Pearson residual in the heterogeneous network model, the possible fault location is located, and then control or maintenance measures are taken to avoid or reduce fault losses.
2. The method for monitoring faults of a complex mechanical system with functional cross-linking based on a heterogeneous attribute network as claimed in claim 1, characterized in that: Step 1 is implemented as follows: Step 1-1: Analyze complex mechanical systems and construct a functional cross-linked heterogeneous property network model; The parts or components of complex mechanical systems are often coupled or cross-linked. With the help of graph theory, they are represented by attribute information networks. The information network is defined as a mapping function with object types. :V→A and relationship type mapping function ψ:E→R of a directed or undirected graph G=(V,E); If the number of object types A>1 or the number of relationship types R>1 in the information network, it is called a heterogeneous information network, otherwise it is a homogeneous information network; the node types in complex mechanical systems are not the same, so the use of heterogeneous attribute information networks can more comprehensively represent system information; Most complex mechanical systems have multiple functions, and each function may use different parts or units. Therefore, for each function of the mechanical system, a heterogeneous attribute network model can be constructed as shown in formula (1), where k is the number of functions of the mechanical system: G=(G1,G2...,G k )(1) Step 1-2: Calculate the functional cross-linking degree and the importance of parts according to the number of nodes and edges in the heterogeneous attribute network model of the functional cross-linking system; According to the number of shared nodes and edges in the heterogeneous attribute network model of each function, the functional cross-linking degree representing the degree of dependence between two functions is calculated; the total number of nodes in the system is N, the total number of edges is M, and the number of shared nodes between two functions is n. ij , the number of shared edges is m ij , then the cross-linking degree of any two functions can be expressed as: The importance of each part or node in a complex mechanical system is also different. According to the number of connecting edges of each node, the importance of the part is obtained as follows, where m i is the number of connected edges of a node; Step 1-3: Determine the dynamic and static properties of the heterogeneous attribute network model based on the characteristics of the complex mechanical system; The edge E in the network model constructed in step 1-1 expresses the connection attributes between nodes, which can be divided into dynamic attributes and static attributes. Static attributes are the inherent attribute relationships between two nodes, including the node type of components and the service life of components. Dynamic attributes are the attribute relationships between two nodes that change over time, including the flow value between hydraulic components in the hydraulic system and the voltage value between components in the electrical system. ij,t and w ij,t represents the static and dynamic properties of the network model at time t, and for the convenience of calculation, w ij,t Vectorized to w t , x ij,t Matrix X t .
3. The method for monitoring faults of a complex mechanical system with functional cross-linking based on a heterogeneous attribute network as claimed in claim 1, characterized in that: In order to improve the evaluation accuracy, the least squares and Kalman filter algorithms are combined to determine the dynamic parameter β in the network model. t , the implementation method of step 3 is, Step 3-1: Determination of SSM model parameters; The state transfer matrix F in the SSM model will be used under normal operation of the system, β t It is estimated by using historical data; Step 3-2: Combine the least squares and Kalman filter algorithms to determine the dynamic parameter β in the network model t ; From step 2, we can know that the network parameter β t The accuracy of will directly affect the efficiency and accuracy of fault diagnosis of complex mechanical systems. The state transfer matrix F in the state transfer model SSM determined in step 3-1 is equivalent to β t The prediction model will obtain w during the operation of the system t Therefore, with the help of Kalman filtering principle, using the measured and predicted values of network parameters, more accurate network parameter estimates can be obtained; The prediction model of the Kalman filter algorithm is as shown in equations (8) to (9), β t The initial value of is estimated by the least squares method, where β tt-1 β t The prior estimate of β t-1t-1 β t-1 The posterior estimate of tt-1 is the covariance of the prior errors, P t-1t-1 is the error covariance at time t-1, Q t is t The covariance of β tt-1 =Fβ t-1t-1 (8) P tt-1 =FP t-1t-1 F T +Q t (9) The correction model of the Kalman filter algorithm is as shown in equations (10) to (12), where β tt β t The posterior estimate of tt-1 w t A priori estimate of K t is the Kalman gain, P tt is the error covariance at time t, G t is the observation matrix, R t To measure noise; b tt =b tt-1 +K t (w t -w tt-1 (10) P tt =(I-K t G t )P tt-1 (11) 4. The method for monitoring faults of a complex mechanical system with functional cross-linking based on a heterogeneous attribute network as claimed in claim 3, characterized in that: Step 4 is implemented as follows: Step 4-1: The network model dynamic parameters β determined by step 3 t and w t The measured values were used to calculate the Pearson residuals; Step 3 calculates a more accurate estimate of the network parameters. With the help of formula (8), we can obtain the prior estimate of the network parameters and then obtain w t The prior estimate of is as follows: w tt-1 =g(X t b tt-1 (13) Thus, we can obtain the i-th edge w in the network model at time t+1 t The prior estimate of w t The Pearson residual between the measured values of is as follows, where MSE is the mean square error between the two; Take the average of the Pearson residuals of all edges and obtain the Pearson residual value of the entire network model: By monitoring the Pearson residual of the network model in real time, the state of the complex mechanical system can be monitored; Step 4-2: Construct a cumulative sum control chart CUSUM and use the cumulative sum control chart CUSUM to monitor large fluctuations in the Pearson residuals; The CUSUM control chart is a control chart that is more effective in controlling mean changes. It is often more capable of detecting larger fluctuations that are greater than one standard deviation σ. For the Pearson residuals of the network model at each moment, the cumulative sums of the upper and lower sides of the CUSUM control chart are: in is the cumulative sum on one side, is the cumulative sum of the lower side, and μ0 is the mean of the monitored object, the mean of the Pearson residual is 0, and Y is the allowable deviation. or When the judgment value H is exceeded, the system is judged to be abnormal; Step 4-3: Construct an exponentially weighted moving average (EWMA) control chart and use the exponentially weighted moving average (EWMA) to monitor smaller fluctuations in the Pearson residuals; The EWMA control chart is a control chart that has a good control effect on small fluctuations less than one standard deviation σ; the statistics of the EWMA control chart are: where z t is the control chart statistic, and z0=0, λ is the weight parameter; The upper control limit (UCL) and lower control limit (LCL) of the EWMA control chart are: Where l is the control limit parameter and σ is the standard deviation of the control chart statistic.
