Astronavigation energy fault early warning system and method and electronic equipment

Abstract

The invention discloses an astronavigation energy fault early warning system and method and electronic equipment. By constructing a qualitative model and a quantitative model, evolution stage analysis of different types of faults and staged change trend prediction of fault variables are realized. In practical application, the method is combined with a dynamic fault early warning model, and real-time fault early warning of the energy system is realized according to the state prediction values of the multiple key variables. Meanwhile, in order to cope with predictable and unpredictable working modes, a mode of combining a typical working mode network model and a Bayesian network learning method is adopted, and the accuracy of an evaluation result is ensured. The method has wide application value, and the safety and reliability of an aerospace energy system can be remarkably improved.

Description

TECHNICAL FIELD

[0001] The application relates to a spaceflight energy failure early warning system, method and electronic device, in particular to a spaceflight energy failure early warning system, method and electronic device based on a multi-level mechanism model topology, and belongs to the technical field. BACKGROUND

[0002] The spacecraft failure early warning system generally realizes failure avoidance through redundancy design, single machine level health management is relatively advanced, and has the rudiment of failure early warning, and the system level failure early warning is mostly safety mode design, which avoids the establishment of failure prediction ability required by failure early warning to a certain extent.

[0003] At present, the failure early warning based on signal processing, model and artificial intelligence is applied. The failure early warning method based on signal processing mainly has amplitude, time domain, frequency domain, time-frequency domain characteristics and other analysis methods. When it is difficult to establish the analytical mathematical model of the diagnosed object, the method based on signal processing is difficult to establish the analytical mathematical model of the diagnosed object, and the difficulty of extracting the mathematical model of the object is avoided. The method based on the analytical model is the earliest developed method, which requires to establish a relatively accurate mathematical model of the diagnosed object. For a nonlinear system, one method is to linearize the nonlinear system at a certain working point and use the linearized model for failure early warning, and the other is to directly use the nonlinear model for failure early warning. However, as the system becomes more and more complex, it is more and more difficult to establish an accurate mathematical model of the object. The method based on artificial intelligence is currently mostly data-driven, which has the problem of being unable to explain, and there is a big difference in model effect for different data types and application scenarios, which cannot complete the reliability requirement of the spaceflight system.

[0004] The existing intelligent management means and failure early warning method of the spacecraft energy system all rely on the lower computer to manage the single device or only complete the data acquisition function, and the upper computer part realizes the real-time monitoring, isolation and reconstruction of the power failure of the entire spacecraft energy system in orbit. In the future, in the face of massive data of different time sequences and space sequences uploaded by large-scale energy storage system devices, it is necessary to infinitely improve the data processing capacity, data mining capacity and intelligent processing capacity. This way lacks marginal constraints on satellite computing power requirements. At the same time, the failure early warning lacks analysis and application of the overall energy system state, which makes it difficult to improve the accuracy of failure early warning. At the same time, there is a lack of research on the state evaluation of the energy system, the weight distribution of the key parameters, and the insufficient accuracy of the failure early warning model. Therefore, based on the characteristics of high input dimension and obvious time sequence characteristics of multi-dimensional sensing data of complex energy systems, the application establishes a state evaluation and prediction model jointly driven by expert knowledge and failure data, extracts parameter historical time sequence characteristics, and realizes rapid early warning with high credibility for future trends. SUMMARY

[0005] The technical problem solved by the present application is to overcome the shortcomings of the prior art, provide a spaceflight energy failure early warning system, method and electronic device, and provide overall system performance.

[0006] The technical solution of the present application is:

[0007] The present application discloses a spaceflight energy failure early warning system based on a multi-level mechanism model topology, comprising: a long-term longitudinal failure data extraction model, a failure evolution path random walk decision model, and a hybrid reasoning module, wherein,

[0008] The long-term longitudinal failure data extraction model pre-processes, extracts features and reduces dimensions of component failure data to obtain failure feature data, which is sent to the failure evolution path random walk decision model;

[0009] The failure evolution path random walk decision model constructs a failure evolution model based on the physical properties and failure evolution law of the energy system, combined with the failure feature data, to generate failure depth representation;

[0010] The hybrid reasoning module predicts the failure according to the failure depth representation to obtain a prediction result.

[0011] Further, in the above system, the component failure data is pre-processed and feature-extracted to obtain failure feature data, specifically:

[0012]

[0013] Wherein, zij is the failure feature data, We is the parameter matrix, be is the bias vector, The contaminated input obtained by randomly setting some elements of xij to zero, xij is the high-dimensional vector after pre-processing of the component failure data, f(·) is a non-homogeneous random process function.

[0014] Further, in the above system, the failure evolution model is constructed based on the physical properties and failure evolution law of the energy system, combined with the failure feature data, to generate the failure depth representation, and the specific method is:

[0015] Use one-hot encoding to generate variable eij;

[0016] Take the variable eij and the failure feature data zij as inputs and input them into two independent GRU models, respectively, and introduce the conditional intensity λik(tij) to obtain the failure depth representation data hij' and hij" of the hidden layers of the GRU models, respectively.

[0017] Further, in the above system, the data failure depth representation predicts the failure to obtain a prediction result, and the specific method is:

[0018] The fault depth characterization data hij' and hij" are spliced to obtain the depth characterization at each fault time;

[0019] According to the depth characterization at each fault time, the occurrence probability of the predictable target event is obtained.

[0020] Further, in the above system, the fault depth characterization data hij' and hij" are spliced to obtain the depth characterization at each fault time, specifically:

[0021]

[0022] Wherein, hij is the depth characterization at each fault time, is a set of depth characterizations hij.

[0023] Further, in the above system, according to the depth characterization at each fault time, the occurrence probability of the predictable target event is obtained, specifically:

[0024]

[0025] Wherein, is the occurrence probability of the predictable target event, wo and bo are parameters, hij is the depth characterization at each fault time, and sigmoid is a nonlinear activation function.

[0026] The application discloses a spaceflight energy fault early warning method based on a multi-level mechanism model topology, comprising:

[0027] The component fault data is preprocessed, feature extracted and dimensionally reduced to obtain fault feature data;

[0028] Based on the physical characteristics and fault evolution law of the energy system, the fault feature data is combined to construct a fault evolution model and generate fault depth characterization;

[0029] According to the fault depth characterization, the fault is predicted to obtain a prediction result.

[0030] The application discloses an electronic device, comprising a memory and a processor:

[0031] The memory is used for storing one or more computer instructions;

[0032] The processor is used for executing the one or more computer instructions, so as to:

[0033] Complete all steps of the spaceflight energy fault early warning method based on the multi-level mechanism model topology.

[0034] The application has the beneficial effects of the prior art:

[0035] (1) Adopting the mixed reasoning method to realize the knowledge and data fusion driven interpretable time prediction, solve the problem that the pure data driven model proposed in the prior art does not use historical fault data and is not interpretable, introduce a dynamic updating optimization mechanism, update the fault features and weights in real time, and enhance the adaptability and real-time performance of the system.

[0036] (2) The fault event prediction problem is converted into a random walk problem on a knowledge graph, and the random walk path planning is regarded as a fault development trajectory and potential possibility. The knowledge graph is deeply characterized by the knowledge construction of the reliable historical fault data, so that the fault mechanism is interpretable. The requirements of the space energy system on fault reliability can be met.

[0037] (3) The neural network is used as a feature extractor, and the reinforcement learning module is trained, the jump strategy of the random walk is generated by using the reinforcement learning module, the random walk is controlled to jump to the fault to be warned as much as possible, and the fault prediction is completed. The comprehensive monitoring and early warning of the energy system fault, the efficient evaluation under different working conditions and the real-time and accurate fault evolution prediction effect are realized, and the technical problems such as insufficient real-time performance of fault early warning, incomplete coverage of typical working conditions and low prediction accuracy under complex conditions of the traditional method are solved.

[0038] (4) The model cooperative optimization is also realized, the advantages of various models are fully utilized, and the overall performance of the system is further improved. These improvements are solutions to the problems such as single model, insufficient dynamic updating mechanism and lack of model cooperative optimization ability in the prior art, effectively overcome these defects, and provide more reliable and efficient technical support for the space energy system fault early warning. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a basic model schematic diagram of the present application. DETAILED DESCRIPTION

[0040] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0041] As Figure 1 shown, the present application discloses a space energy fault early warning system based on a multi-level mechanism model topology, comprising: a long-term longitudinal fault data extraction model, a fault evolution path random walk decision model and a mixed reasoning module, wherein,

[0042] The long-term longitudinal fault data extraction model pre-processes, extracts features and reduces dimensions of the component fault data, obtains fault feature data, and sends the fault feature data to the fault evolution path random walk decision model;

[0043] The fault evolution path random walk decision model is based on the physical characteristics and fault evolution law of an energy system, combines fault feature data, constructs a fault evolution model, and generates fault depth representation.

[0044] The hybrid inference module predicts the fault according to the fault depth representation and obtains a prediction result.

[0045] Preferably, the component fault data is preprocessed and feature extraction is performed to obtain fault feature data, specifically:

[0046]

[0047] wherein zij is the fault feature data, We is a parameter matrix, be is a bias vector, is a polluted input obtained by randomly setting some elements in xij to zero, xij is a high-dimensional vector after preprocessing of the component fault data, and f(·) is a non-homogeneous random process function.

[0048] Preferably, based on the physical characteristics and fault evolution law of an energy system, combined with fault feature data, a fault evolution model is constructed to generate fault depth representation, and the specific method is:

[0049] The one-hot encoding is used to generate variable eij.

[0050] The variables eij and fault feature data zij are input as inputs into two independent GRU models, respectively, and the conditional strength λik(tij) is introduced to obtain fault depth representation data hij' and hij" of the hidden layer of the GRU model, respectively.

[0051] Preferably, the data fault depth representation is used to predict the fault to obtain a prediction result, and the specific method is:

[0052] The fault depth representation data hij' and hij" are spliced to obtain the depth representation at each fault time.

[0053] According to the depth representation at each fault time, the occurrence probability of a predictable target event is obtained.

[0054] Preferably, the fault depth representation data hij' and hij" are spliced to obtain the depth representation at each fault time, and the specific method is:

[0055]

[0056] wherein hij is the depth representation at each fault time, is a set of depth representations hij.

[0057] Preferably, according to the depth characterization at each time of failure, the occurrence probability of the predictable target event is obtained, specifically:

[0058]

[0059] wherein, is the occurrence probability of the predictable target event, wo and bo are parameters, hij is the depth characterization at each time of failure, and sigmoid is a nonlinear activation function.

[0060] The application discloses a spaceflight energy failure early warning method based on a multi-level mechanism model topology, comprising:

[0061] The component failure data is preprocessed, feature extracted and dimensionally reduced to obtain failure feature data;

[0062] Based on the physical characteristics and failure evolution law of the energy system, the failure feature data is combined to construct a failure evolution model and generate failure depth characterization;

[0063] According to the failure depth characterization, the failure is predicted to obtain a prediction result.

[0064] The application discloses an electronic device, comprising a memory and a processor:

[0065] The memory is used for storing one or more computer instructions;

[0066] The processor is used for executing one or more computer instructions, so as to:

[0067] Complete all steps of the spaceflight energy failure early warning method based on the multi-level mechanism model topology.

[0068] Embodiment

[0069] The embodiment provides a spaceflight energy failure early warning system based on a multi-level mechanism model topology, comprising the following modules:

[0070] 1. Longitudinal failure data extraction model

[0071] In view of the inherent limitations of the threshold alarm method, the data-driven method can effectively and flexibly learn the normal state information behind the big data. The key variable state evaluation or prediction model established based on the above-mentioned autoencoder and recurrent neural network models can train independent models according to the actual conditions such as the specific working conditions, load and environment of the energy system. The combination of multiple key variable real-time prediction or evaluation models can cope with the switching of different normal states of key variables in complex working environments. Multiple warning thresholds are designed under different working conditions to flexibly respond to complex working environments. Collect multi-dimensional high-coupling sensor information of the aerospace energy system, including system state parameters (such as temperature, pressure, current, energy output, etc.) and environmental parameters (such as external disturbances, working mode switching, etc.). The collected data is preprocessed, deep representation feature extraction and dimensionality reduction are carried out for subsequent analysis.

[0072] 2. Fault Evolution Path Random Walk Decision Model

[0073] Under the constraints of energy system volume and weight, an intelligent autonomous operation and early fault warning model is established for aerospace energy systems with multiple operating modes, a wide parameter perturbation range, and multi-dimensional, highly coupled sensor information. The most representative energy system state parameters for energy system fault warning are selected to construct an early fault warning model architecture that covers the system's health status. Data-driven approaches alone are not highly feasible. The fault prediction model utilizes an energy system fault evolution mechanism model and combines data-driven approaches to predict the evolution trends of fault variables. Based on the physical characteristics of the energy system and the laws of fault evolution, a multi-level mechanism model is constructed. The impact of different fault types and severities on system state parameters is analyzed, providing an explanation of the evolution trends of fault variables.

[0074] 3. Hybrid Inference Model

[0075] Hybrid reasoning model framework: Fault prediction is performed based on actual operating sensor parameter data combined with the hybrid reasoning module model. Figure 1 As shown, the present invention includes a model framework: (1) a non-homogeneous random process, (2) a deep representation zij process, and (3) a multi-sequence data processing model. The results of the mechanism model and the data-driven method are integrated to comprehensively analyze the evolution trend of the system state parameters, perform early identification and warning of potential faults, and comprehensively predict the results, including the probability of fault occurrence, time node, and key influencing factors.

[0076] Given an input high-dimensional vector xij, the model can be Output the corresponding depth representation zij. Among them, is the “contaminated input” obtained by randomly setting some elements in xij to zero, and We and be are the corresponding parameter matrices and bias vectors (bias).

[0077] A model named GRU is used in the present application to implement RNN. This model has shown excellent performance in many sequence data processing tasks (such as machine translation, speech recognition).

[0078] Specifically, in the case of a sequence σi composed of triplets (eij, tij, xij), we first extract the energy system state parameter information, and then input it into the representation learning module to obtain the corresponding deep representation vector zij.

[0079] Subsequently, eij (eij encoded using one-hot) and zij are input into two independent RNN modules to obtain hidden layers hij' and hij". hij' and hij" can be regarded as deep representations learned from the sequence (xi1, …, xij) and the sequence (ei1, …, eij), respectively.

[0080] The conditional intensity λik(tij) of the Hawkes process is introduced into the GRU to evaluate the influence of time intervals and past events. When the hidden layer of the GRU is updated, the proportion of information flow of the hidden layer is controlled using λik(tij), and the specific calculation method of λik(tij) has been described by the following formula. In the case of considering λik(tij) to control information transmission, the information update of the GRU is subject to:

[0081] hij' = GRU(zij, (α*λik(tij)*hi(j-1)')),

[0082] hij" = GRU(eij, (α*λik(tij)*hi(j-1) ").

[0083] λik(tij) is a coefficient related to past failure events and time, and the introduction of λik(tij) into the information update of the GRU can be regarded as introducing a special attention mechanism, so that the model can more finely process the influence of past failure data on event prediction. By concatenating hij' and hij", the deep representation at each failure is obtained:

[0084]

[0085] Finally, based on the obtained hij vector, an output layer is added, which can predict the occurrence probability of the target event, where wo and bo are corresponding parameters.

[0086] The long-term longitudinal fault data extraction model module acquires the operation data of the aerospace energy system through sensors, and after cleaning, conversion and feature extraction, the processed data are input into the fault evolution path random walk decision model module; the mechanism model module analyzes the system state and predicts the fault reason based on physical knowledge, and the data-driven method module uses statistical learning algorithm to identify the mode and predict the trend of the operation data, and the results of the two parts are then fused into the hybrid reasoning model framework module to generate accurate fault warning information.

[0087] Although the present application has been described in detail by the above preferred embodiments, it should be appreciated that the above description should not be considered as limiting the present application. Various modifications and alternatives will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present application should be defined by the appended claims.

[0088] The contents not described in detail in the specification of the present application are the known technology of the skilled in the art.

Claims

1. An aerospace energy failure warning system based on a multi-level mechanism model topology, characterized by: include: Long-term longitudinal fault data extraction model, fault evolution path random walk decision model, hybrid reasoning module, among which, Long-term longitudinal fault data extraction model preprocesses component fault data, extracts features, and reduces dimensionality to obtain fault feature data, which is then sent to the random walk decision model for the fault evolution path. The random walk decision model of the fault evolution path is based on the physical characteristics of the energy system and the fault evolution law, combined with the fault feature data, to build a fault evolution model and generate a deep fault representation; The hybrid reasoning module predicts the fault according to the fault depth representation and obtains the prediction result.

2. The aerospace energy failure early warning system based on a multi-level mechanism model topology according to claim 1 is characterized in that: The component fault data is preprocessed and feature extracted to obtain fault feature data, specifically: Among them, zij is the fault feature data, We is the parameter matrix, be is the bias vector, is the contaminated input obtained by randomly setting some elements in xij to zero, xij is the high-dimensional vector after preprocessing of component failure data, and f(·) is the non-homogeneous random process function.

3. The aerospace energy failure early warning system based on a multi-level mechanism model topology according to claim 1 is characterized in that: Based on the physical characteristics of the energy system and the fault evolution law, combined with the fault feature data, a fault evolution model is constructed to generate a deep fault representation. The specific method is as follows: Use one-hot encoding to generate variable eij; The variable eij and fault feature data zij are taken as input and fed into two independent GRU models respectively. The conditional strength λik(tij) is introduced to obtain the fault depth representation data hij′ and hij″ of the hidden layer of the GRU model respectively.

4. The aerospace energy failure early warning system based on a multi-level mechanism model topology according to claim 1 is characterized in that: The data fault depth characterization is used to predict the fault and obtain the prediction result. The specific method is as follows: The fault depth characterization data hij′ and hij″ are spliced ​​together to obtain the depth characterization of each fault; Based on the deep characterization of each fault, the probability of occurrence of the predictable target event is obtained.

5. The aerospace energy failure early warning system based on multi-level mechanism model topology according to claim 4 is characterized in that: The fault depth characterization data hij′ and hij″ are spliced ​​together to obtain the depth characterization of each fault, specifically: Among them, hij is the depth representation of each fault, is the set of deep representations hij.

6. The aerospace energy failure early warning system based on multi-level mechanism model topology according to claim 5 is characterized in that: According to the depth characterization of each fault, the probability of occurrence of the predictable target event is obtained, specifically: in, To predict the probability of occurrence of the target event, wo and bo are parameters, hij is the deep representation of each fault, and sigmoid is a nonlinear activation function.

7. A method for early warning of aerospace energy failure based on a multi-level mechanism model topology, characterized in that: include: Preprocess component fault data, extract features, and reduce dimension to obtain fault feature data; Based on the physical characteristics of the energy system and the fault evolution law, combined with the fault feature data, a fault evolution model is constructed to generate a deep fault representation; According to the fault depth characterization, the fault is predicted and the prediction result is obtained.

8. An electronic device, characterized in that: Including memory and processor: The memory is used to store one or more computer instructions; The processor is configured to execute the one or more computer instructions to: Complete all steps of the aerospace energy failure early warning method based on multi-level mechanism model topology as described in claim 7.

Images