Complex system fault twinborn deduction method based on virtual-real dynamic fusion

Abstract

The invention discloses a complex system fault twinborn deduction method based on virtual-real dynamic fusion, and the method comprises the steps: firstly designing a preprocessing scheme of multi-modal data fusion, and carrying out the spatial-temporal feature unification and standardized state vector construction of heterogeneous sensing data; secondly, establishing a real-to-virtual digital mapping model, deducing a multilevel system state based on a dynamic Bayesian network, and introducing a weighted adaptive gradient correction mechanism to improve the fitting precision of the model; and finally, proposing a fault evolution evaluation method of reflecting real from virtual, and realizing hierarchical identification and early warning triggering of the system operation state. By introducing a digital twinning thought, establishing a virtual model synchronously evolved with an entity system, and constructing a whole-process twinning deduction framework of perception, modeling and evaluation, the method can significantly improve the accuracy, response efficiency and robustness of fault prediction of a complex system.

Description

Technical Field

[0001] The present invention belongs to the field of digital twin and intelligent manufacturing system fault prediction, and specifically relates to a fault twin deduction method and system for complex unmanned flexible systems based on the dynamic fusion of virtual and real based on digital twin technology. Background Art

[0002] Digital twins are a modeling and deduction technology that integrates the physical and digital worlds. They can dynamically simulate and optimize physical systems through high-precision digital models. In complex unmanned systems, digital twins not only accurately represent system structure but also provide real-time perception of operating status and deduction of future behavior, demonstrating significant value in intelligent operations and maintenance, such as fault prediction and health management.

[0003] Traditional methods for predicting faults in complex systems often rely on static rules or empirical models, primarily using fixed thresholds, historical statistics, or expert knowledge for anomaly detection and trend analysis. These methods have significant limitations in practical applications: First, they lack the ability to integrate multi-source, heterogeneous sensory data, making it difficult to uniformly express system states. Second, the modeling structure is relatively static, lacking the ability to dynamically depict the evolution of system states. Third, when faced with the state coupling and sudden disturbances of complex, multi-level systems, the prediction response is delayed and the accuracy is low, making it difficult to meet the requirements of high-reliability scenarios such as flexible manufacturing.

[0004] In contrast, the complex system fault twin deduction method based on the dynamic fusion of virtual and real fully embodies the core concept of digital twins and can more effectively address the shortcomings of traditional methods in data fusion, state modeling, and fault prediction. By constructing a digital mapping body in virtual space that evolves synchronously with the physical system, this method achieves real-time perception, evolutionary deduction, and risk assessment of the system's operating status, with greater dynamic adaptability and modeling accuracy. Relying on the feedback mechanism of virtual-real linkage, this method significantly improves the accuracy of complex system fault identification and response efficiency, providing reliable intelligent prediction and operation and maintenance support for unmanned systems. Summary of the Invention

[0005] Purpose of the Invention: To address the issues with traditional complex system fault prediction methods, this invention proposes a complex system fault twin deduction method based on the dynamic fusion of virtual and real, based on the concept of digital twins. By constructing a digital model that evolves synchronously with the physical system, this method achieves unified representation of multi-source heterogeneous data, dynamic deduction of system states, and real-time assessment of fault evolution. This method can better reflect the multi-level structure and dynamic change characteristics of complex, unmanned, flexible systems, effectively improve the accuracy and response efficiency of fault identification, and enhance the system's intelligent perception and autonomous operation and maintenance capabilities.

[0006] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is:

[0007] A complex system fault twin deduction method based on dynamic fusion of virtual and real, including the following steps:

[0008] Step 1: Fusion preprocess the multi-source heterogeneous sensor data of the complex unmanned system of intelligent manufacturing in the flexible assembly line of automotive parts, and achieve time alignment through a unified time reference to generate a standardized multimodal state vector.

[0009] Step 2: Based on the standardized multimodal state vector, a dynamic Bayesian network is used to construct a multi-level state structure to model the operating state of the complex unmanned system. By combining causal dependence and time evolution, dynamic deduction of the complex system state is achieved.

[0010] Step 3: The model of the complex unmanned system operation state model is deduced using a deduction method based on weighted adaptive gradient correction. By comparing the weighted state error between the deduced state and the actual state, the state transfer parameters in the deduction model are dynamically adjusted. The adaptive gradient descent strategy is used to optimize the parameters according to the state importance, and the iteration termination is controlled by the error convergence condition to achieve continuous approximation and efficient fitting of the virtual state to the entity evolution process.

[0011] Step 4: Using a virtual-to-real system fault evolution assessment method, the simulated state information is used to perform qualitative change stratification screening, identify key out-of-limit behaviors, and determine the fault level. If the qualitative change conditions are not met, a quantitative change multi-dimensional fusion mechanism is used to assess degradation trends, enabling dynamic identification and graded early warning of system faults.

[0012] Preferably: the weighted state error function in step 3 is:

[0013]

[0014] in, represents the weighted state error function, represents the state variables of each component at time t predicted by the deduction model, represents the state variable of the j-th subsystem at time t predicted by the deduction model, represents the state variables of the overall system at time t predicted by the deduction model, represents the state variables of each component of the real system at time t, represents the state variable of the jth subsystem of the real system at time t, X sys (t) represents the state variable of the real system as a whole at time t, represents the weight coefficient of the component layer, represents the weight coefficient of the subsystem layer, ω sys Represents the weighting coefficient of the system layer.

[0015] Preferably, the gradient update and convergence determination in step 3 are as follows: the state transition probability of the basic component is defined as follows:

[0016]

[0017] in, represents the probability that component i transitions from state m to state m+1 at time t, represents the component degradation intensity coefficient, ρ i (t) represents the external load factor that the component is subjected to at time t, represents the basic degradation rate, m represents the mth state of component i, Represents the state space size of the i-th component.

[0018] The parameter update formula for step t is as follows:

[0019]

[0020] Among them, η (t) represents the learning rate at the current time t, and uses attenuation to control the update amplitude to avoid late oscillation or too fast convergence. The gradient term can be expanded according to the chain rule:

[0021]

[0022] When the change amplitude of the error function is less than the set threshold ε within Q consecutive time steps, it satisfies:

[0023]

[0024] The model is considered to have reached the fitting accuracy requirement and the parameter update process is stopped.

[0025] Preferably, the state classification in step 4 includes the following steps:

[0026] In the fault evolution assessment mechanism, the deduction model output includes the basic component layer status Subsystem level status and system-level status The three constitute a multi-level, time-series state set. To achieve effective judgment and evaluation, it is necessary to classify the state variables:

[0027] A type of state variable will directly cause system operation interruption or safety failure after crossing the set threshold, with obvious critical trigger characteristics, and is defined as a qualitative change state. This type of state mainly includes the following two categories:

[0028]

[0029] in It represents the maximum degradation level of the i-th basic component. When this level is reached, the component is considered to have failed. It represents the safe operation threshold of the jth subsystem. If the threshold is exceeded, it is considered that the subsystem loses stability.

[0030] Another type of state variable does not directly lead to failure, but its continuous deviation or multi-dimensional superposition can indicate system performance degradation and pose potential risks. It is defined as a quantitative change state:

[0031]

[0032] When all qualitative change states are not triggered, the quantitative change state will serve as the input for subsequent multi-dimensional fusion calculations.

[0033] Preferably, the qualitative change stratification screening in step 4 includes the following steps:

[0034] All qualitative change states are judged hierarchically according to their action levels, and a qualitative change response fusion function is introduced.

[0035]

[0036] Among them, N cpt Indicates the number of basic components, N sub Indicates the number of functional subsystems, represents the state variables of each component of the real system at time t, represents the state variable of the jth subsystem of the real system at time t, represents the state space size of the i-th component, Indicates the division level of the subsystem state space, X sys (t) represents the state variable of the real system as a whole at time t, K sys Indicates the highest operating status level.

[0037] The qualitative change response fusion function characterizes the local qualitative change risk by using the collaborative degradation strength between components and subsystems. It then uses the system-level state as a regulation factor to comprehensively reflect the overall qualitative change evolution trend of the system.

[0038] On this basis, the hierarchical fault judgment function of the system qualitative change state is constructed as follows:

[0039]

[0040] in, represents the layered fault determination function, L represents the highest layer of fault stratification, represents the qualitative change response fusion function, η L represents the comprehensive judgment threshold of the fault level of the Lth layer, η lIndicates the comprehensive judgment threshold of the fault level of layer l, satisfying η1<η2<…<η L .when When it falls into the corresponding interval, it is determined that the system is in the first The qualitative change risk level of the level. The system is considered to have entered the fault state domain.

[0041] Preferably, the quantitative multi-dimensional fusion in step 4 includes the following steps:

[0042] like Then enter the fusion evaluation process of the quantitative change level, and construct the quantitative change multidimensional fusion function Calculate the comprehensive response intensity of multidimensional states in the quantitative change domain:

[0043]

[0044] Obtaining the quantitative fusion evaluation value of the system at the current moment Finally, a judgment function for trend fault identification is constructed. Considering the continuous characteristics of behavioral changes during system operation, the time window length is set to T, and the continuous over-limit criterion is introduced. The expression is as follows:

[0045]

[0046] Among them, ζ represents a certain moment of system operation, Θ warn The threshold value set by the system. This expression means that if the system has a time window length of T, there are consecutive moments such that Always not lower than the preset threshold Θ warn , it is considered that the system has entered the quantitative failure range.

[0047] Preferably, step 2 comprises the following steps:

[0048] Step 21: Based on the different functions and physical deployment locations of the devices, the overall system structure can be abstractly divided into three levels: component layer, subsystem layer, and system layer.

[0049] Step 211: The system includes N cpt basic components, the operating state of each component at time t can be represented by the state variable Represents, where i=1,2,...,N cpt The state variable takes the value of a finite state set in Indicates the state space size of the i-th component, which is used to characterize the health level or functional status of the component. sub functional subsystems, the operating state of the jth subsystem at time t is represented by the state variable Represents, where j=1,2,...,N sub .variable The value range is defined as in Indicates the division level of the subsystem state space. At the system level, the state variable of the overall system at time t is defined as X sys (t), whose value space is {1,2,...,K sys}, used to describe the global operating status level of the flexible assembly line.

[0050] In step 212, the degradation process of each component is an irreparable state evolution, which satisfies the first-order discrete-time homogeneous Markov process. The state space size of component i is The state transfer process consists of a The upper triangular transfer probability matrix is composed of the following forms:

[0051]

[0052] Among them, the elements represents the probability of component i transitioning from state m to state n, and only n ≥ m is allowed, that is, the state can only remain or evolve in a worse direction and satisfy the normalization constraint:

[0053] Step 213: System state variable X sys The state space of (t) is {1,2,...,K sys}, K sys Represents the highest operating state level, where the first h states constitute the reliable state space. When the system is within this interval, it is judged to be reliable. The reliability of the system at time t is defined as:

[0054]

[0055] At the same time, the system state probability distribution satisfies the normalization condition, where R(t) represents the reliability.

[0056] Step 22: The multi-state hierarchical system is divided into basic component states Functional subsystem status and the overall system state X sys (t). At any time t, define an acyclic graph in:

[0057] The node collection is:

[0058]

[0059] The set of directed edges is:

[0060]

[0061] Among them, Sub j Represents the set of components that constitute subsystem j.

[0062] The joint probability distribution of the system at time t can be expanded by the chain rule as:

[0063]

[0064] Among them, P sub (t), P cpt (t) are the conditional probabilities of the subsystem layer and the state probabilities of the basic component layer, respectively. Each conditional probability is determined by the parent node set of the corresponding node.

[0065] For subsystem status nodes The parent node set is the state of the basic components under it At the same time, the subsystem status is determined by the status of its worst component:

[0066]

[0067] System status node X sys The parent node set of (t) is all subsystem states Also using the minimum dominant strategy, the system state depends on the most serious subsystem state:

[0068]

[0069] Step 23: The system operation process is discretized into T time slices, and each time t∈{1,2,...,T} corresponds to a complete static Bayesian network structure In the time dimension, the following cross-time dependency edges are introduced:

[0070]

[0071] Thus forming a dynamic Bayesian network structure The node set and edge set are represented as follows:

[0072]

[0073] where ε static (t) is the hierarchical structure of the static Bayesian network at time t, ε time (t, t+1) is the evolution connection between time slices.

[0074] Under the dynamic Bayesian structure, the joint state probability of the system in T time slices can be expressed as:

[0075]

[0076] The conditional probability of each time slice can be decomposed into the following three layers of causal relationships:

[0077]

[0078] Combined with the joint distribution of the static structure within each time slice, it can be expressed as:

[0079]

[0080] After combination, the complete dynamic joint probability of the system is uniformly expressed as:

[0081]

[0082] in:

[0083]

[0084] The state transition probability is defined by the upper triangular matrix P i cpt Control, further introduce the parameter variables of the control matrix:

[0085]

[0086] in, Represents the component degradation intensity coefficient. Represents the basic degradation rate. ρ i (t) represents the external load factor borne by the component at time t.

[0087] Preferably, step 1 comprises the following steps:

[0088] Step 11: State vector normalization: Assume that there are M types of sensors in the system, and the i-th type of sensor contains N i Data collection node, the raw data collected by the j-th sensor of the i-th category at time t is expressed as:

[0089]

[0090] in, Represents the state value of the jth sensor of the i-th category at time t.

[0091] Define a unified system fusion state vector S t , which is composed as follows:

[0092]

[0093] in, Represents the state value after standardization, alignment, denoising and semantic encoding, forming the state fusion vector S at the unified time tt As input to the virtual-reality deduction model.

[0094] Source heterogeneous sensors include temperature sensors, pressure sensors, humidity sensors, vibration sensors, displacement sensors, and current / voltage sensors.

[0095] Step 12, time series alignment: Use a fixed time step Δt to construct a unified global time series:

[0096] T={t0,t1,...,t k ,...,t n},t k =t0+k·Δt

[0097] Where T represents the local time series, t k represents the kth time point, t0 represents the initial moment, k represents the index of the time step, and Δt represents the time step length. This time series serves as a unified time frame for the evolution of the system state and is used to align the original data under different modes.

[0098] Step 13: Anomaly identification and signal denoising: For time series The sliding window method is used for local statistical analysis. Assume that the sliding window length is W and the samples contained in the window are for:

[0099]

[0100] Calculate the mean within the window and standard deviation It is calculated as follows:

[0101]

[0102] At time t k The observed value deviates too much from the statistical center of the window, that is, it meets the following abnormal judgment conditions:

[0103]

[0104] Then the point is regarded as an outlier, where λ is the adjustable outlier determination coefficient.

[0105] The sliding weighted average filter is used for smoothing. Assuming the weighted window size is K, the state value after filtering is calculated as follows:

[0106]

[0107] Among them, ω m is the weighted coefficient of the mth position, satisfying ω m> 0, and gradually decreases as time moves away from the current point, and finally a smooth state sequence is obtained after abnormal elimination and filtering.

[0108] Preferably, the specific alignment process in step 12 is as follows:

[0109] Step 121: Set a unified sampling time step Δt.

[0110] Step 122: for each type of sensor, let its original sampling time point sequence be {τ1,τ2,...,τ T}, since its sampling moment is the same as some moment t in the unified time series k The data do not completely overlap, and interpolation estimation is required on this time base.

[0111] Step 123, for a state variable X ij , if the time t k Between the original sampling time τ p With τ p+1 If the value between , the estimated state value at that moment can be calculated by linear interpolation method. The interpolation formula is as follows:

[0112]

[0113] Among them, τ p ≤t k ≤τ p+1 , indicating t k The time interval, At time t k The estimated value of At time τ p The actual value of At time τ p+1 The actual value of .

[0114] Another object of the present invention is to provide a complex system fault twin deduction system based on virtual-real dynamic fusion, which is used to implement a complex system fault twin deduction method based on virtual-real dynamic fusion, including an acquisition unit, a preprocessing unit, a dynamic Bayesian network construction unit, a deduction unit, and an evaluation unit, wherein:

[0115] The acquisition unit is used to collect multi-source heterogeneous sensor data of a complex unmanned system in an intelligent manufacturing in a flexible assembly line for automobile parts.

[0116] The preprocessing unit is used to perform fusion preprocessing on multi-source heterogeneous sensor data, and realize time sequence alignment through a unified time reference to generate a standardized multimodal state vector.

[0117] The dynamic Bayesian network construction unit is used to construct a multi-level state structure using a dynamic Bayesian network based on a standardized multimodal state vector, model the operating state of a complex unmanned system, and combine causal dependence with time evolution to achieve dynamic deduction of the complex system state.

[0118] The deduction unit is used to deduce the complex unmanned system operation state model by using a deduction method based on weighted adaptive gradient correction. By comparing the weighted state error between the deduced state and the actual state, the state transfer parameters in the deduction model are dynamically adjusted. The adaptive gradient descent strategy is used to optimize the parameters according to the state importance, and the iteration termination is controlled by the error convergence condition to achieve continuous approximation and efficient fitting of the virtual state to the entity evolution process.

[0119] The assessment unit uses a virtual-to-real system fault evolution assessment method to perform qualitative change stratification screening using deduced state information, identify key out-of-limit behaviors, and determine the fault level. If the qualitative change conditions are not met, a quantitative change multi-dimensional fusion mechanism is used to assess degradation trends, enabling dynamic identification and graded early warning of system faults.

[0120] Compared with the prior art, the present invention has the following beneficial effects:

[0121] 1. Based on the concept of digital twins, this invention constructs a virtual-reality fusion deduction model that evolves synchronously with the physical system. It can achieve accurate modeling and dynamic deduction of the state of complex unmanned flexible systems, effectively improving the perception ability and prediction accuracy of the system operation status.

[0122] 2. The present invention realizes dynamic identification and graded warning of system faults through a method combining multimodal data fusion, dynamic Bayesian network modeling and weighted gradient correction mechanism, which significantly improves the timeliness and accuracy of fault prediction and the intelligent level of system operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0123] Figure 1 This is a flowchart for twin deduction of complex system faults based on dynamic fusion of virtual and real;

[0124] Figure 2 It is the convergence trend diagram of the stratification error;

[0125] Figure 3 It is a three-dimensional optimization path diagram during the parameter gradient correction process;

[0126] Figure 4 It is the time series curve for comparing virtual and real states;

[0127] Figure 5 It is a process of hierarchical screening of qualitative changes and multi-dimensional fusion judgment of quantitative changes. DETAILED DESCRIPTION

[0128] The present invention is further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0129] like Figure 1 As shown, the present invention designs a complex system fault twin deduction method based on dynamic fusion of virtual and real, and constructs a virtual-real integrated deduction and evaluation mechanism to address the difficulties in modeling equipment state evolution and the lag in fault prediction in the unmanned intelligent manufacturing system of flexible assembly lines for automotive parts. The method first designs a preprocessing scheme for multimodal data fusion, unifies the spatiotemporal features of heterogeneous sensor data and constructs a standardized state vector; secondly, a digital mapping model that adjusts the real to the virtual is established, and the multi-level system state is deduced based on a dynamic Bayesian network, and a weighted adaptive gradient correction mechanism is introduced to improve the model fitting accuracy; finally, a fault evolution evaluation method that maps the virtual to the real is proposed to achieve hierarchical identification and early warning triggering of the system operation state, which specifically includes the following steps:

[0130] Step 1: Fusion preprocess the multi-source heterogeneous sensor data of complex unmanned systems, achieve time series alignment through a unified time base, and combine interpolation, anomaly removal and weighted smoothing to generate a standardized multimodal state vector to provide consistent input for subsequent state deduction models.

[0131] Step 2: Use dynamic Bayesian networks to construct a multi-level state structure to model the operating state of complex unmanned systems, combine causal dependence and time evolution, and realize dynamic deduction of system states.

[0132] Step 3: Based on the deduction method of weighted adaptive gradient correction, by comparing the weighted error between the deduced state and the actual state, the state transfer parameters in the deduction model are dynamically adjusted, the adaptive gradient descent strategy is used to optimize the parameters according to the state importance, and the error convergence condition is used to control the iteration termination, so as to achieve continuous approximation and efficient fitting of the virtual state to the entity evolution process.

[0133] Step 4: Through the virtual-to-real system fault evolution assessment method, the deduced state information is used to perform qualitative change hierarchical screening, determine key out-of-limit behaviors and determine the fault level; if the qualitative change conditions are not met, the quantitative change multi-dimensional fusion mechanism is used to assess the degradation trend, realizing dynamic identification and graded warning of system faults.

[0134] In step 1, the process of fusion preprocessing of multi-source heterogeneous sensor data in complex unmanned systems includes the following steps:

[0135] The first step is state vector standardization: In the intelligent manufacturing scenario of flexible assembly lines for automotive parts, the system usually deploys multiple types of sensors to collect real-time operating status information of key equipment and processes. Sensor types include temperature sensors, pressure sensors, humidity sensors, vibration sensors, displacement sensors, and current / voltage sensors, etc. They are used to sense key dimensions such as environmental conditions, equipment operating parameters, and processing stability during the assembly process. Suppose there are M types of sensors in the system, and the i-th type of sensor contains N i Data collection node, the raw data collected by the j-th sensor of the i-th category at time t is expressed as:

[0136]

[0137] in, Represents the state value of the jth sensor of the i-th category at time t.

[0138] Define a unified system fusion state vector S t , which is composed as follows:

[0139]

[0140] in, Represents the state value after standardization, alignment, denoising and semantic encoding, forming the state fusion vector S at the unified time t t As input to the virtual-reality deduction model.

[0141] The second step is time series alignment: a fixed time step Δt is used to construct a unified global time series:

[0142] T={t0,t1,...,t n},t k =t0+k·Δt

[0143] This time series serves as a unified time frame for the system state evolution and is used to align the raw data under different modes. The specific alignment process is as follows:

[0144] ① Set a unified sampling time interval Δt, for example, Δt = 0.1 seconds, to meet the sampling accuracy requirements of system state changes;

[0145] ② For each type of sensor, let its original sampling time point sequence be {τ1,τ2,...,τ T}, since its sampling moment is the same as some moment t in the unified time series k If the data do not completely overlap, it is necessary to perform interpolation estimation on the time base;

[0146] ③For a certain state variable X ij , if the time tk Between the original sampling time τ p With τ p+1 If the value between , the estimated state value at that moment can be calculated using linear interpolation. The interpolation formula is as follows:

[0147]

[0148] Among them, τ p ≤t k ≤τ p+1 , indicating t k The time interval, At time t k The estimated value of At time τ p The actual value of .

[0149] The third step is anomaly identification and signal denoising: for time series The sliding window method is used for local statistical analysis; let the sliding window length be W, and the samples contained in the window be:

[0150]

[0151] Calculate the mean within the window and standard deviation It is calculated as follows:

[0152]

[0153] At time t k The observed value deviates too much from the statistical center of the window, that is, it meets the following abnormal judgment conditions:

[0154]

[0155] Then the point is regarded as an outlier, where λ is the adjustable outlier determination coefficient.

[0156] In order to further improve the continuity and availability of the signal, a sliding weighted average filter is used for smoothing. Assuming the weighted window size is K, the state value after filtering is calculated as follows:

[0157]

[0158] Among them, ω m is the weighted coefficient of the mth position, satisfying ω m > 0, and gradually decreases as time moves away from the current point, and finally a smooth state sequence is obtained after abnormal elimination and filtering.

[0159] In step 2, the main process of modeling the operating status of a complex unmanned system using a digital mapping model based on a dynamic Bayesian network is as follows:

[0160] Step 1: Based on the different device functions and physical deployment locations, the overall system structure can be abstractly divided into three levels: component level, subsystem level, and system level. To simplify the analysis, this method makes the following assumptions:

[0161] ①Assume that the system contains N cpt basic components, the operating state of each component at time t can be represented by the state variable Represents, where i=1,2,...,N cpt The state variable takes the value of a finite state set in Represents the state space size of the i-th component, which is used to characterize the health level or functional status of the component; further, assuming that there are N sub functional subsystems, the operating state of the jth subsystem at time t is represented by the state variable Represents, where j=1,2,...,N sub .variable The value range is defined as in Represents the division level of the subsystem state space; at the system level, the state variable of the overall system at time t is defined as X sys (t), whose value space is {1,2,...,K sys}, used to describe the global operating status level of the flexible assembly line.

[0162] ② The degradation process of each component is an irreparable state evolution, which satisfies the first-order discrete-time homogeneous Markov process. Assume that the state space size of component i is The state transfer process consists of a The upper triangular transfer probability matrix is composed of the following forms:

[0163]

[0164] Among them, the elements represents the probability of component i transitioning from state m to state n, and only n ≥ m is allowed, that is, the state can only remain or evolve in a worse direction and satisfy the normalization constraint:

[0165] ③Assume that the system state variable X sys The state space of (t) is {1,2,...,K sys}, where the first h states constitute the reliable state space. When the system is within this interval, it is judged to be reliable. The reliability of the system at time t is defined as:

[0166]

[0167] At the same time, the system state probability distribution satisfies the normalization condition:

[0168]

[0169] Step 2: Divide the multi-state hierarchical system into basic component states Functional subsystem status and the overall system state X sys (t). At any time t, define an acyclic graph in:

[0170] The node collection is:

[0171]

[0172] The set of directed edges is:

[0173]

[0174] Among them, Sub j Represents the set of components that constitute subsystem j.

[0175] The joint probability distribution of the system at time t can be expanded by the chain rule as:

[0176]

[0177]

[0178] Among them, P sub (t), P cpt (t) are the conditional probabilities of the subsystem layer and the state probabilities of the basic component layer, respectively. Each conditional probability is determined by the parent node set of the corresponding node.

[0179] For subsystem status nodes The parent node set is the state of the basic components under it At the same time, the subsystem status is determined by the status of its worst component:

[0180]

[0181] System status node X sys The parent node set of (t) is all subsystem states Also using the minimum dominant strategy, the system state depends on the most serious subsystem state:

[0182]

[0183] Therefore, the conditional probability table at the system level is defined as:

[0184]

[0185] in Represents the specific state value of subsystem j at time t.

[0186] Step 3: Assume that the system operation process is discretized into T time slices, and each moment t∈{1,2,...,T} corresponds to a complete static Bayesian network structure In the time dimension, the following cross-time dependency edges are introduced:

[0187]

[0188] Thus forming a dynamic Bayesian network structure The node set and edge set are represented as follows:

[0189]

[0190] where ε static (t) is the hierarchical structure of the static Bayesian network at time t, ε time (t, t+1) is the evolution connection between time slices.

[0191] Under the dynamic Bayesian structure, the joint state probability of the system in T time slices can be expressed as:

[0192]

[0193] The conditional probability of each time slice can be decomposed into the following three layers of causal relationships:

[0194]

[0195] Further combined with the joint distribution of the static structure within each time slice, it can be expressed as:

[0196]

[0197] After combination, the complete dynamic joint probability of the system can be uniformly expressed as:

[0198]

[0199] in:

[0200]

[0201] The state transition probability is defined by the upper triangular matrix P i cpt Control, further introduce the parameter variables of the control matrix:

[0202]

[0203] in, represents the component degradation intensity coefficient; represents the basic degradation rate; ρ i (t) represents the external load factor borne by the component at time t.

[0204] In step 3, the main process of the deduction method based on weighted adaptive gradient correction is as follows:

[0205] (1) Construction of error function

[0206] At each moment t, the virtual system obtains the virtual state set through dynamic Bayesian network deduction based on the state and model parameters of the previous moment:

[0207]

[0208] The physical system obtains the actual state set through real-time perception:

[0209]

[0210] In order to reflect the importance of state variables at different levels in system deduction, a weighted vector is introduced:

[0211]

[0212] Among them, the larger the weight, the more significant the impact of the state on system stability or fault evolution.

[0213] Based on the above structure, the weighted state error function at time t is constructed:

[0214]

[0215] (2) Gradient update and convergence judgment

[0216] The state transition probability of the basic components is defined as follows:

[0217]

[0218] The parameter update formula for step t is as follows:

[0219]

[0220] Among them, η (t) represents the learning rate at the current time t, and uses a decaying form to control the update amplitude. This is to avoid oscillation or too fast convergence in the later stage. The gradient term can be expanded according to the chain rule:

[0221]

[0222] When the change amplitude of the error function is less than the set threshold ε within Q consecutive time steps, it satisfies:

[0223]

[0224] The model is considered to have reached the fitting accuracy requirement and the parameter update process is stopped.

[0225] In step 4, this method adopts a fault assessment mechanism that combines qualitative change hierarchical screening with quantitative change multi-dimensional fusion to achieve dynamic identification and graded early warning of virtual simulation status. The main process is as follows:

[0226] (1) Status classification

[0227] In the fault evolution assessment mechanism, the deduction model output includes the basic component layer status Subsystem level status and system-level status The three constitute a multi-level, time-series state set. State variables need to be classified as follows:

[0228] A type of state variable will directly cause system operation interruption or safety failure after crossing the set threshold, with obvious critical trigger characteristics, and is defined as a qualitative change state. This type of state mainly includes the following two categories:

[0229]

[0230] in represents the maximum degradation level of the i-th basic component. When this level is reached, the component is considered to have failed. It represents the safe operation threshold of the jth subsystem. If the threshold is exceeded, it is considered that the subsystem loses stability.

[0231] Another type of state variable does not directly lead to failure, but its continuous deviation or multi-dimensional superposition can indicate system performance degradation and pose potential risks. It is defined as a quantitative change state:

[0232]

[0233] When all qualitative change states are not triggered, the quantitative change state will serve as the input for subsequent multi-dimensional fusion calculations.

[0234] (2) Qualitative change stratification screening

[0235] In order to achieve a more differentiated risk response strategy, all qualitative change states are hierarchically determined according to their action levels, and a qualitative change response fusion function is introduced.

[0236]

[0237] This function characterizes the local qualitative change risk through the collaborative degradation intensity between components and subsystems; and then combines the system layer state as a regulation item to comprehensively reflect the overall qualitative change evolution trend of the system.

[0238] On this basis, the hierarchical fault judgment function of the system qualitative change state is constructed as follows:

[0239]

[0240] Among them, η l Indicates the comprehensive judgment threshold of the fault level of layer l, satisfying η1<η2<…<η L .when When it falls into the corresponding interval, it is determined that the system is in the first The qualitative change risk level of the level. The system is considered to have entered the fault state domain.

[0241] (3) Quantitative change and multi-dimensional integration

[0242] like Then enter the fusion evaluation process of the quantitative change level, and construct the quantitative change multidimensional fusion function Calculate the comprehensive response intensity of multidimensional states in the quantitative change domain:

[0243]

[0244] This formula highlights the change items with larger state amplitudes in the multidimensional quantitative change process by accumulating the normalized squares of the states of components, subsystems, and system layers; it can be used to comprehensively measure the quantitative evolution intensity of the system at time t and the concentration of multi-layer structural changes.

[0245] Obtaining the quantitative fusion evaluation value of the system at the current moment Finally, a judgment function for trend fault identification is constructed. Considering the continuous characteristics of behavioral changes during system operation, the time window length is set to T, and the continuous over-limit judgment criterion is introduced: if there are ζ consecutive moments within T that make the fusion value always exceed the warning threshold Θ warn , then the system is judged to have entered the quantitative change fault range. The expression is as follows:

[0246]

[0247] The qualitative change part of this method is centered on the over-limit triggering of critical states, and quickly identifies the direct failure risk of the system through hierarchical screening; the quantitative change part integrates multi-dimensional degradation state information and combines continuous over-threshold behavior within a time window to achieve dynamic assessment and early warning of trend faults.

[0248] In another embodiment, a complex system fault twin deduction system based on virtual-real dynamic fusion is provided, which is used to implement a complex system fault twin deduction method based on virtual-real dynamic fusion, including an acquisition unit, a preprocessing unit, a dynamic Bayesian network construction unit, a deduction unit, and an evaluation unit, wherein:

[0249] The acquisition unit is used to collect multi-source heterogeneous sensor data of a complex unmanned system in an intelligent manufacturing in a flexible assembly line for automobile parts.

[0250] The preprocessing unit is used to perform fusion preprocessing on multi-source heterogeneous sensor data, and realize time sequence alignment through a unified time reference to generate a standardized multimodal state vector.

[0251] The dynamic Bayesian network construction unit is used to construct a multi-level state structure using a dynamic Bayesian network based on a standardized multimodal state vector, model the operating state of a complex unmanned system, and combine causal dependence with time evolution to achieve dynamic deduction of the complex system state.

[0252] The deduction unit is used to deduce the complex unmanned system operation state model by using a deduction method based on weighted adaptive gradient correction. By comparing the weighted state error between the deduced state and the actual state, the state transfer parameters in the deduction model are dynamically adjusted. The adaptive gradient descent strategy is used to optimize the parameters according to the state importance, and the iteration termination is controlled by the error convergence condition to achieve continuous approximation and efficient fitting of the virtual state to the entity evolution process.

[0253] The assessment unit uses a virtual-to-real system fault evolution assessment method to perform qualitative change stratification screening using deduced state information, identify key out-of-limit behaviors, and determine the fault level. If the qualitative change conditions are not met, a quantitative change multi-dimensional fusion mechanism is used to assess degradation trends, enabling dynamic identification and graded early warning of system faults.

[0254] The convergence trend of the layering error in this embodiment is as follows Figure 2 As shown, the three-dimensional optimization path in the parameter gradient correction process is as follows Figure 3 As shown in the figure; the dynamic process of hierarchical multi-dimensional fusion fault assessment is as follows Figure 4 As shown, the qualitative change hierarchical screening and quantitative change multi-dimensional fusion judgment process is as follows Figure 5 shown.

[0255] By introducing the concept of digital twins, establishing a virtual model that evolves synchronously with the physical system, and constructing a full-process twin deduction framework for perception, modeling, and evaluation, the present invention can significantly improve the accuracy, response efficiency, and robustness of complex system fault prediction. It has broad application prospects in the equipment status monitoring and fault prediction scenarios of flexible assembly lines for automotive parts in the field of intelligent manufacturing.

[0256] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A complex system fault twin deduction method based on dynamic fusion of virtual and real, characterized by: The following steps are involved: Step 1: Fusion preprocessing of multi-source heterogeneous sensor data from complex unmanned systems in intelligent manufacturing of automotive parts flexible assembly lines is performed, and time series alignment is achieved through a unified time reference to generate a standardized multimodal state vector. Step 2: Based on the standardized multimodal state vector, a dynamic Bayesian network is used to construct a multi-level state structure to model the operating state of the complex unmanned system. By combining causal dependence and time evolution, dynamic deduction of the complex system state is achieved. Step 3: The model of the complex unmanned system's operational state is deduced using a deduction method based on weighted adaptive gradient correction. By comparing the weighted state error between the deduced state and the actual state, the state transition parameters in the deduction model are dynamically adjusted. The adaptive gradient descent strategy is used to optimize the parameters according to the state importance, and the error convergence condition is used to control the termination of the iteration, achieving continuous approximation and efficient fitting of the virtual state to the physical evolution process. Step 4: Using a virtual-to-real system fault evolution assessment method, the simulated state information is used to perform qualitative change stratification screening, determine key out-of-limit behaviors, and determine the fault level. If the qualitative change conditions are not met, a quantitative change multi-dimensional fusion mechanism is used to evaluate the degradation trend and realize dynamic identification and graded early warning of system failures.

2. The complex system fault twin deduction method based on dynamic fusion of virtual and real according to claim 1 is characterized by: The weighted state error function in step 3 is: in, represents the weighted state error function, represents the state variables of each component at time t predicted by the deduction model, represents the state variable of the j-th subsystem at time t predicted by the deduction model, represents the state variables of the overall system at time t predicted by the deduction model, represents the state variables of each component of the real system at time t, represents the state variable of the jth subsystem of the real system at time t, X sys (t) represents the state variable of the real system as a whole at time t, represents the weight coefficient of the component layer, represents the weight coefficient of the subsystem layer, ω sys Represents the weighting coefficient of the system layer.

3. The complex system fault twin deduction method based on dynamic fusion of virtual and real according to claim 2 is characterized by: The gradient update and convergence determination in step 3 are as follows: The state transition probability of the basic component is defined as follows: in, represents the probability that component i transitions from state m to state m+1 at time t, represents the component degradation intensity coefficient, ρ i (t) represents the external load factor that the component is subjected to at time t, represents the basic degradation rate, m represents the mth state of component i, Indicates the state space size of the i-th component; The parameter update formula for step t is as follows: Among them, η (t) represents the learning rate at the current time t, and uses attenuation to control the update amplitude to avoid late oscillation or too fast convergence. The gradient term can be expanded according to the chain rule: When the change amplitude of the error function is less than the set threshold ε within Q consecutive time steps, it satisfies: The model is considered to have reached the fitting accuracy requirement and the parameter update process is stopped.

4. The complex system fault twin deduction method based on dynamic fusion of virtual and real according to claim 3 is characterized by: The status classification in step 4 includes the following steps: In the fault evolution assessment mechanism, the deduction model output includes the basic component layer status Subsystem level status and system-level status The three constitute a multi-level, time-series state set; in order to achieve effective judgment and evaluation, the state variables need to be classified: One type of state variable will directly cause system operation interruption or safety failure after crossing the set threshold. It has obvious critical trigger characteristics and is defined as a qualitative change state. This type of state mainly includes the following two categories: in represents the maximum degradation level of the i-th basic component. When this level is reached, the component is considered to have failed. represents the safe operation threshold of the jth subsystem. Exceeding this threshold is considered as a loss of subsystem stability; Another type of state variable does not directly lead to failure, but its continuous deviation or multi-dimensional superposition can indicate system performance degradation and pose potential risks. It is defined as a quantitative change state: When all qualitative change states are not triggered, the quantitative change state will serve as the input for subsequent multi-dimensional fusion calculations.

5. The complex system fault twin deduction method based on dynamic fusion of virtual and real according to claim 4 is characterized by: The qualitative change stratification screening in step 4 includes the following steps: All qualitative change states are judged hierarchically according to their action levels, and a qualitative change response fusion function is introduced. Among them, N cpt Indicates the number of basic components, N sub Indicates the number of functional subsystems, represents the state variables of each component of the real system at time t, represents the state variable of the jth subsystem of the real system at time t, represents the state space size of the i-th component, Indicates the division level of the subsystem state space, X sys (t) represents the state variable of the real system as a whole at time t, K sys Indicates the highest operating status level; The qualitative change response fusion function characterizes the local qualitative change risk through the collaborative degradation intensity between components and subsystems. It then combines the system-level state as a regulation item to comprehensively reflect the overall qualitative change evolution trend of the system. On this basis, the hierarchical fault judgment function of the system qualitative change state is constructed as follows: in, represents the layered fault determination function, L represents the highest layer of fault stratification, represents the qualitative change response fusion function, η L represents the comprehensive judgment threshold of the fault level of the Lth layer, η l Indicates the comprehensive judgment threshold of the fault level of layer l, satisfying η1<η2<…<η L ;when When it falls into the corresponding interval, it is determined that the system is in the first The qualitative change risk level of the level; if The system is considered to have entered the fault state domain.

6. The complex system fault twin deduction method based on dynamic fusion of virtual and real according to claim 5 is characterized by: The quantitative multi-dimensional fusion in step 4 includes the following steps: like Then enter the fusion evaluation process of the quantitative change level, and construct the quantitative change multidimensional fusion function Calculate the comprehensive response intensity of multidimensional states in the quantitative change domain: Obtaining the quantitative fusion evaluation value of the system at the current moment Finally, a judgment function for trend fault identification is constructed; considering the continuous characteristics of behavioral changes during system operation, the time window length is set to T, and the continuous over-limit criterion is introduced, which is expressed as follows: Among them, ζ represents a certain moment of system operation, Θ warn The threshold value set by the system, which means that if the system has a time window length of T, there are consecutive moments such that Always not lower than the preset threshold Θ warn , it is considered that the system has entered the quantitative failure range.

7. The complex system fault twin deduction method based on dynamic fusion of virtual and real according to claim 6 is characterized by: Step 2 includes the following steps: Step 21: Based on the different functions and physical deployment locations of the devices, the overall system structure can be abstractly divided into three levels: component layer, subsystem layer, and system layer; Step 211: The system includes N cpt basic components, the operating state of each component at time t can be represented by the state variable Represents, where i=1,2,...,N cpt ; The state variable takes the value of a finite state set in Represents the state space size of the i-th component, which is used to characterize the health level or functional status of the component; there are N sub functional subsystems, the operating state of the jth subsystem at time t is represented by the state variable Represents, where j=1,2,...,N sub ;variable The value range is defined as in Represents the division level of the subsystem state space; at the system level, the state variable of the overall system at time t is defined as X sys (t), whose value space is {1,2,...,K sys }, used to describe the global operating status level of the flexible assembly line; In step 212, the degradation process of each component is an irreparable state evolution, which satisfies the first-order discrete-time homogeneous Markov process; the state space size of component i is The state transfer process consists of a The upper triangular transfer probability matrix is composed of the following forms: Among them, the elements represents the probability of component i transitioning from state m to state n, and only n ≥ m is allowed, that is, the state can only remain or evolve in a worse direction and satisfy the normalization constraint: Step 213: System state variable X sys The state space of (t) is {1,2,...,K sys }, K sys Represents the highest operating state level, where the first h states constitute the reliable state space. When the system is within this interval, it is judged to be reliable. The reliability of the system at time t is defined as: At the same time, the probability distribution of the system state satisfies the normalization condition, where R(t) represents the reliability; Step 22: The multi-state hierarchical system is divided into basic component states Functional subsystem status and the overall system state X sys (t); At any time t, define an acyclic graph in: The node collection is: The set of directed edges is: Among them, Sub j represents the set of components that constitute subsystem j; The joint probability distribution of the system at time t can be expanded by the chain rule as: Among them, P sub (t), P cpt (t) are the conditional probability of the subsystem layer and the state probability of the basic component layer, respectively. Each conditional probability is determined by the parent node set of the corresponding node; For subsystem status nodes The parent node set is the state of the basic components under it At the same time, the subsystem status is determined by the status of its worst component: System status node X sys The parent node set of (t) is all subsystem states Also using the minimum dominant strategy, the system state depends on the most serious subsystem state: Step 23: The system operation process is discretized into T time slices, and each time t∈{1,2,...,T} corresponds to a complete static Bayesian network structure In the time dimension, the following cross-time dependency edges are introduced: Thus forming a dynamic Bayesian network structure The node set and edge set are represented as follows: where ε static (t) is the hierarchical structure of the static Bayesian network at time t, ε time (t, t+1) is the evolution connection between time slices; Under the dynamic Bayesian structure, the joint state probability of the system in T time slices can be expressed as: The conditional probability of each time slice can be decomposed into the following three layers of causal relationships: Combined with the joint distribution of the static structure within each time slice, it can be expressed as: After combination, the complete dynamic joint probability of the system is uniformly expressed as: in: The state transition probability is defined by the upper triangular matrix P i cpt Control, further introduce the parameter variables of the control matrix: Among them, θ i deg represents the component degradation intensity coefficient; θ i base represents the basic degradation rate; ρ i (t) represents the external load factor borne by the component at time t.

8. The complex system fault twin deduction method based on dynamic fusion of virtual and real according to claim 7 is characterized by: Step 1 includes the following steps: Step 11: State vector normalization: Assume that there are M types of sensors in the system, and the i-th type of sensor contains N i Data collection node, the raw data collected by the j-th sensor of the i-th category at time t is expressed as: in, represents the state value of the jth sensor of the i-th category at time t; Define a unified system fusion state vector S t , which is composed as follows: in, Represents the state value after standardization, alignment, denoising and semantic encoding, forming the state fusion vector S at the unified time t t As input to the virtual reality model; Source heterogeneous sensors include temperature sensors, pressure sensors, humidity sensors, vibration sensors, displacement sensors, and current / voltage sensors; Step 12, time series alignment: Use a fixed time step Δt to construct a unified global time series: T={t0,t1,...,t k ,...,t n },t k =t0+k·Δt Where T represents the time series, t k Denotes the kth time point, t0 denotes the initial moment, k denotes the index of the time step, Δt denotes the time step length. This time series serves as a unified time frame for the evolution of the system state and is used to align the raw data under different modes. Step 13: Anomaly identification and signal denoising: For time series Use the sliding window method to perform local statistical analysis; let the sliding window length be W, and the samples contained in the window for: Calculate the mean within the window and standard deviation It is calculated as follows: At time t k The observed value deviates too much from the statistical center of the window, that is, it meets the following abnormal judgment conditions: Then the point is regarded as an outlier, where λ is the adjustable outlier determination coefficient; The sliding weighted average filter is used for smoothing. Assuming the weighted window size is K, the state value after filtering is calculated as follows: Among them, ω m is the weighted coefficient of the mth position, satisfying ω m > 0, and gradually decreases as time moves away from the current point, and finally a smooth state sequence is obtained after abnormal elimination and filtering.

9. The complex system fault twin deduction method based on dynamic fusion of virtual and real according to claim 8 is characterized by: The specific alignment process in step 12 is as follows: Step 121, setting a unified sampling time step Δt; Step 122: for each type of sensor, let its original sampling time point sequence be {τ1,τ2,...,τ T }, since its sampling moment is the same as some moment t in the unified time series k If the data do not completely overlap, it is necessary to perform interpolation estimation on the time base; Step 123, for a state variable X ij , if the time t k Between the original sampling time τ p With τ p+1 If the value between , the estimated state value at that moment can be calculated by linear interpolation method. The interpolation formula is as follows: Among them, τ p ≤t k ≤τ p+1 , indicating t k The time interval, At time t k The estimated value of At time τ p The actual value of At time τ p+1 The actual value of .

10. A system for implementing the complex system fault twin deduction method based on dynamic fusion of virtual and real as described in claim 1, characterized in that: It includes an acquisition unit, a preprocessing unit, a dynamic Bayesian network construction unit, a deduction unit, and an evaluation unit, among which: The acquisition unit is used to collect multi-source heterogeneous sensor data of the intelligent manufacturing complex unmanned system in the flexible assembly line of automobile parts; The preprocessing unit is used to perform fusion preprocessing on multi-source heterogeneous sensor data, and realize time sequence alignment through a unified time reference to generate a standardized multimodal state vector; The dynamic Bayesian network construction unit is used to construct a multi-level state structure using a dynamic Bayesian network based on the standardized multimodal state vector, model the operating state of the complex unmanned system, and realize the dynamic deduction of the complex system state by combining causal dependence and time evolution; The deduction unit is used to deduce the complex unmanned system operation state model using a deduction method based on weighted adaptive gradient correction. By comparing the weighted state error between the deduced state and the actual state, the state transition parameters in the deduction model are dynamically adjusted. The parameters are optimized according to the state importance using an adaptive gradient descent strategy, and the iteration termination is controlled by the error convergence condition, so as to achieve continuous approximation and efficient fitting of the virtual state to the physical evolution process. The evaluation unit is used to use a virtual-to-real system fault evolution evaluation method to perform qualitative change hierarchical screening using deduced state information, determine key out-of-limit behaviors and determine the fault level; if the qualitative change conditions are not met, a quantitative change multi-dimensional fusion mechanism is used to evaluate the degradation trend, thereby realizing dynamic identification and graded warning of system faults.

Images