Decision-level fusion method for full system health assessment of liquid rocket engines

Abstract

The present invention discloses a decision-level fusion method for the health assessment of the whole system of a liquid rocket engine, which includes the following steps: first, a multi-variable, multi-system time evolution sequence state parameterization model is constructed based on a conditional random field as a basic method for system-level state assessment; second, a time tracking conditional random field algorithm for key and non-key subsystems of the whole system is constructed to achieve secondary marking of the health state of the subsystem on a continuous time series; accurate discrimination of the health state of the subsystem under a strong interference state is achieved through the optimization solution of the normalization factor and parameters of the conditional random field; finally, a health decision analysis of the whole system state of the liquid rocket engine is achieved through an integrated discrimination algorithm for key and non-key subsystems. The present invention can effectively reduce the occurrence of misdiagnosis and missed diagnosis of complex mechanical equipment such as liquid rocket engines, and achieve accurate assessment of the health state of the system and decision-making suggestions.

Description

Technical Field

[0001] The invention belongs to the field of liquid rocket engine full system health assessment and relates to a decision-level fusion method for liquid rocket engine full system health assessment. Background Art

[0002] Liquid rocket engine test is aimed at evaluating the engine system performance, component performance and overall structure, and providing the most effective and direct test verification means for system solution optimization and rocket flight performance evaluation. Analyzing the real working process of the engine and comprehensively judging the health status based on the actual engine test or flight signal is an important way to ensure the normal service of the system level and control the spread of faults. The fault extension time of the rich fuel cycle engine and the oxygen rich cycle engine is short and belongs to different orders of magnitude. Therefore, when the system-level measurement parameters and fault diagnosis system have not yet performed fault judgment or transmitted to the rocket body control system, an irreversible major fault may have occurred; currently, there are about 200 engine test signals, and there are only about 10 engine status monitoring signals in rocket flight tests, which makes it difficult to make accurate decisions based on measured signals. At present, missed diagnosis and misdiagnosis occur frequently in the health decision-making of liquid rocket engine test and rocket flight stages in my country. Therefore, it is very important to study the system-level health decision-making method based on engine test and flight signals to provide a basis for the continued service of the engine, the normal flight of the rocket and fault control.

[0003] At present, the more common decision-level fusion methods are mainly divided into the following five categories: Bayes inference, expert system, DS evidence reasoning, fuzzy theory and conditional random field. These methods have the advantages of low communication volume, strong anti-interference ability and low sensor dependence. However, research shows that the above methods are the final results of method fusion. As a fusion for decision-making goals, its fusion results will directly affect the decision level. Among them, conditional random field (CRF), as one of the most advanced feature recognition and fusion methods, has the function of time analysis and prediction of physical information, but the marking effect of CRF under strong interference and weak signal is not ideal, and it is still difficult to meet the decision fusion and high-precision health assessment requirements of high-end complex equipment such as rocket engines, resulting in the need to improve the accuracy of state discrimination and marking. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a decision-level fusion method for full system health assessment of liquid rocket engines, which can meet the requirements of high-precision, high-accuracy and low-misdiagnosis rate decision fusion and health assessment.

[0005] To achieve the above object, the time tracking decision level fusion method TTCRF for liquid rocket engine full system health assessment described in the present invention comprises the following steps:

[0006] 1) According to the weights of each subsystem in the liquid rocket engine system and engineering requirements, several subsystems with high importance are selected for identification to construct a mathematical model of time tracking conditional random fields;

[0007] 2) Extract the time domain index features and perform bilateral second-order synchronous compression transformation on the rapid-varying signals of liquid rocket engines to obtain the time domain index features and instantaneous frequency of the rapid-varying signals of liquid rocket engines, perform trend analysis on the slow-varying signals of liquid rocket engines, and use them for decision-level fusion and health assessment to obtain the time series feature vector;

[0008] 3) The whole system is constructed by using the independent l key systems and O non-key subsystems in the liquid rocket engine, and the health matrix of the whole system is constructed by using the time series feature vector obtained in step 2), and the signal corresponding to the health matrix of the whole system is fused at the time tracking feature level to obtain the conditional probability model of TTCRF;

[0009] 4) Based on the conditional random field theory, the observation sequence and the label sequence are modeled to obtain the conditional probability model, and then the rectangular window function for health status discrimination is introduced;

[0010] 5) Based on the conditional probability model established in step 4), mark the health status of each subsystem and its time point, and obtain the probability deviation by giving the qualified parameters in the conditional probability model to determine the final qualified parameters of the conditional probability model;

[0011] 6) Based on the conditional probability model obtained in step 5), a sub-state sequence set is constructed, and the health value space over a continuous time period is represented by the sub-state sequence set, and a health model is constructed for the system level represented by all subsystems, the system level represented by key subsystems, and non-key subsystems;

[0012] 7) Setting the number of continuous points for system status determination and the health threshold, and determining the health status of the entire liquid rocket engine system through the health model established in step 6).

[0013] The rapid-changing signal of the liquid rocket engine in step 1) is a vibration or fluid pulsation signal; the slow-changing signal of the liquid rocket engine includes the temperature, fluid pressure, flow rate and rotation speed of the liquid rocket engine.

[0014] Assume that there are l+O independent subsystems, then their health function vector is:

[0015] G t o tal(t)=[G1(t),G2(t)……G l (t),G l+1 (t),……G l+O (t)] T

[0016]

[0017] Where l is the key component / subsystem, O is the number of non-key subsystems, G s-total The state matrix formed for key components / subsystems, G o-total State matrix formed for non-critical subsystems.

[0018] In step 3), for the discrete signal acquisition system, the sampling period is T. In the time period [0, kT], the system-level health indicator matrix and the state model of the discrete sequence composed of l+O subsystems are:

[0019]

[0020]

[0021] Among them, G mk (kT) is the health of the mth subsystem at time kT, G *k (kT) is the health of the whole system at time kT, G Lk (kT) is the health of the system composed of L key subsystems at time kT, G Ok (kT) is the health of the system consisting of O non-critical subsystems at time kT.

[0022] In step 4), based on the CRF theory, the observation sequence G o-total and marker sequence Modeling, CRF conditional probability model of the i-th subsystem on the time series [0, ΔT, Δ2T, … ΔkT] for:

[0023]

[0024] The normalization factor in the conditional probability model in step 5 is:

[0025]

[0026] Among them, α is the unknown coefficient, [r i,j ] is the health label matrix of the ith subsystem at time jT, r i，j～k For r i,j The value of .

[0027] The health model constructed at the system level of all subsystems, the system level of key subsystems, and any discrete time period [jT, (j+h)T] of non-key subsystems is:

[0028]

[0029] in, is the health decision function of the whole system consisting of l+O subsystems in the time series [jT,(j+h)T], is the system health decision function consisting of l key subsystems in the time series [jT, (j+h)T], is the system health decision function consisting of O non-critical subsystems on the time series [jT, (j+h)T].

[0030] In step 7),

[0031] When satisfied

[0032]

[0033] All critical subsystems and non-critical systems are normal, and the liquid rocket engine is in full system condition;

[0034] When satisfied

[0035]

[0036] If the non-critical system is normal, but the critical system fails, the entire liquid rocket engine system will be abnormal;

[0037] When satisfied

[0038]

[0039] If the non-critical system fails, the critical system is normal, and the entire liquid rocket engine system is normal;

[0040]

[0041] Then there are faults in both critical subsystems and non-critical non-systems, and the entire liquid rocket engine system fails.

[0042] The present invention has the following beneficial effects:

[0043] The decision-level fusion method for the whole system health assessment of liquid rocket engines described in the present invention adopts the time tracking conditional random field (TTCRF) method to analyze the discrete time series composed of multiple system states, adopts the multi-source information tracking and fusion method to obtain the marking characteristics of the time series, performs neighborhood feature fusion and secondary marking of the time series, obtains a marking state with higher accuracy and credibility, avoids the interference of strong signal interference on the accuracy of the time marking state, and dynamically characterizes the changing law of the system health state, improves the accuracy and credibility of system-level health judgment of complex mechanical equipment such as liquid rocket engines, effectively improves the long-term misdiagnosis and missed diagnosis problems in the health decision-making of liquid rocket engine test and flight stages, and meets the requirements of decision fusion and high-precision health assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flowchart of step 1);

[0045] Figure 2 It is a schematic diagram of the mathematical model of the time tracking conditional random field;

[0046] Figure 3a Distribution plot of normalization factors;

[0047] Figure 3b It is a schematic diagram of the probability model;

[0048] Figure 4 This is a structural diagram of the main system of the liquid engine;

[0049] Figure 5a This is the characteristic level fusion result diagram of the gas generator;

[0050] Figure 5b Feature-level fusion result diagram of the fuel pump;

[0051] Figure 5c Characteristic level fusion result diagram of fuel pre-pressure pump;

[0052] Figure 5d Characteristic level fusion result diagram of oxygen pre-pressure pump;

[0053] Figure 5e This is the characteristic level fusion result diagram of the oxidizer pump;

[0054] Figure 5f This is the characteristic level fusion result diagram of the thrust chamber;

[0055] Figure 6a This is a magnified view of the characteristic level fusion result of the gas generator;

[0056] Figure 6b A zoomed-in view of the feature-level fusion results for the fuel pump;

[0057] Figure 6c A zoomed-in view of the characteristic level fusion results for the fuel pre-pressure pump;

[0058] Figure 6d A zoomed-in view of the characteristic level fusion results of the oxygen pre-pressure pump;

[0059] Figure 6e This is a zoomed-in view of the characteristic stage fusion result of the oxidizer pump;

[0060] Figure 6f This is a magnified view of the characteristic level fusion result of the thrust chamber;

[0061] Figure 7a for y thre-h Schematic diagram of subsystem health judgment results using CRF when =0.85;

[0062] Figure 7b for y thre-h Schematic diagram of subsystem health judgment results using CRF when =0.75;

[0063] Figure 7c for y thre-h Schematic diagram of subsystem health discrimination results using CRF when =0.65;

[0064] Figure 7d for y thre-h Schematic diagram of subsystem health judgment results using TTCRF when =0.85;

[0065] Figure 8a for y thre-h =0.5 (threshold 1) Schematic diagram of subsystem health decision indicators using TTCRF;

[0066] Figure 8b for y thre-h =0.85 (threshold 4) Schematic diagram of subsystem health decision indicators using TTCRF;

[0067] Figure 9a for y thre-h = 0.5 (threshold 1) System-level health decision using TTCRF Schematic diagram;

[0068] Figure 9b for y thre-h = 0.85 (threshold 4) System-level health decision using TTCRF Schematic diagram. DETAILED DESCRIPTION

[0069] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only an embodiment of a part of the present invention, not all embodiments, and is not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concepts disclosed in the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of the present invention.

[0070] The accompanying drawings show schematic diagrams of structures according to embodiments disclosed in the present invention. These figures are not drawn to scale, and some details are magnified and some details may be omitted for the purpose of clear expression. The shapes of various regions and layers shown in the figures and the relative sizes and positional relationships therebetween are only exemplary, and may deviate in practice due to manufacturing tolerances or technical limitations, and those skilled in the art may additionally design regions / layers with different shapes, sizes, and relative positions according to actual needs.

[0071] The decision-level fusion method for liquid rocket engine full system health assessment described in the present invention comprises the following steps:

[0072] 1) According to the weights of each subsystem in the liquid rocket engine and engineering requirements, several subsystems with high importance are selected for state discrimination, with the local replacing the whole and the important representing the general, so as to establish the basic model of time tracking conditional random field (TTCRF).

[0073] Assume a whole system consisting of l key subsystems and O non-key subsystems, then the health function matrix represented in the time domain is:

[0074] G t o tal (t)=[G1(t),G2(t)……G l (t),G l+1 (t),……G l+O (t)] T

[0075]

[0076] Where l is the key component / subsystem, O is the number of non-key subsystems, G s-total The state matrix formed for key components / subsystems, G o-total State matrix formed for non-critical subsystems.

[0077] 2) Perform time domain index feature extraction and bilateral second-order synchronous compression transform (BFSST2) on the rapid-changing signal of the liquid rocket engine to obtain the time domain index features and instantaneous frequency of the rapid-changing signal of the liquid rocket engine, wherein the rapid-changing signal of the liquid rocket engine is a vibration and fluid pulsation signal; perform trend analysis on the slow-changing signal of the liquid rocket engine, wherein the slow-changing signal of the liquid rocket engine includes temperature, fluid pressure, flow rate and speed, and perform decision-level fusion and health assessment on this to obtain a time series feature vector;

[0078] 3) For the independent l key systems and O non-key subsystems constituting the whole system, the time series feature vectors obtained in step 2) are used to form the health matrix of the whole system, and the corresponding signals are fused at the time tracking feature level to obtain the conditional probability model of TTCRF.

[0079] For the discrete signal acquisition system, the sampling period is T, the sensor signal is a discrete time series, and the state models of the whole system, health index matrix and discrete sequence composed of l+O subsystems in the time period [0, kT] are:

[0080]

[0081]

[0082] Among them, G mk (kT) is the health of the mth subsystem at time kT, G *k (kT) is the health of the whole system at time kT, G Lk (kT) is the health of the system composed of L key subsystems at time kT, G Ok (kT) is the health of the system consisting of O non-critical subsystems at time kT.

[0083] Specifically, in step 3), the observation sequence G o-total and marker sequence Modeling, CRF conditional probability model of the i-th subsystem on the time series [0, ΔT, Δ2T, … ΔkT] The expression is:

[0084]

[0085] 4) Based on the conditional random field (CRF) theory, the observation sequence and the label sequence are modeled to obtain the conditional probability model, and the rectangular window function is introduced for discrimination. Then, the qualified parameters of the normalization factor are obtained, and the health status is accurately discriminated through the qualified parameters of the normalization factor;

[0086]

[0087] Among them, α is the unknown coefficient, [r i,j ] is the health label matrix of the ith subsystem at time jT, r i，j～k For r i,j The value of is 0 or 1, that is:

[0088]

[0089] r i，j～k ={0, 1};

[0090] In addition, the discriminant function introduced in step 4) is a rectangular window function, and the expression of the discriminant function is:

[0091]

[0092]

[0093] Among them, sign(·) is the sign function, y thre-h is the health threshold, r ij is the health mark of the i-th subsystem at time jT.

[0094] 5) Based on the conditional probability model established in step 4), mark the health status of each subsystem and its time point, and obtain the probability deviation through the qualified parameters in the given probability model, so as to finally determine the qualified parameters and probability deviation that meet the conditions; thus, it is necessary to select a reasonable parameter a so that the health mark probability is 1 within the health threshold, and when it is not satisfied, the health mark probability is 0 within the health threshold, so as to accurately judge the subsequent health status. In order to further obtain the qualified parameters of the normalization factor, select a reasonable parameter a so that the health mark probability is 1 within the health threshold, and when it is not satisfied, the health mark probability is 0 within the health threshold, so as to accurately judge the subsequent health status, that is:

[0095]

[0096] The conditional probability model and normalization factor in step 4) are:

[0097]

[0098]

[0099] To satisfy the parameter r i，j～k = 0 (unhealthy state) when the conditional probability is 0, and when the condition is met, that is, r i，j～k =1 (healthy state), as close to 1 as possible, where assuming 1-ε, ε is the probability error value formed by the softplus function, then:

[0100]

[0101] Considering that the health status of each subsystem and its time point has been marked as 0 or 1, suppose that the i-th subsystem has N points in normal state in the discrete sampling time period [jT, (j+h)T], then the health status value corresponding to each point is 1, and when all points exceed the threshold, the corresponding health value of each point and the whole is 0, then N∈[0,h], then:

[0102]

[0103] Since the above formula has no analytical solution, by giving a different value, the probability deviation ε is obtained to finally determine the parameters a and ε that meet the condition requirements, as follows: Figure 2 As shown, when the coefficient a is 1 to 10, the standard factor Z is close to v to varying degrees, and as a increases, the value of Z(av) is closer to v. In addition, although the parameter a uses different values, the probability value P(av) is 0 when v = 0, which meets the requirement that the probability is 0 when multiple consecutive time series are in an unhealthy state. When v ≥ 1, only when a is large enough (for example, a ≥ 5), the probability value P(av) is closer to 1, and the difference in the probability value corresponding to different parameter a values ​​is very small, as shown in Table 1.

[0104] Table 1

[0105]

[0106] 6) According to the conditional probability model established in step 5), a sub-state sequence set is constructed, and the health value space in the continuous time period is represented by the sub-state sequence set. The system level represented by all subsystems, the system level represented by key subsystems, and the health model of non-key subsystems are modeled separately, that is, the large complex system is secondary labeled and information fused to improve the reliability of test signals and state evaluation; wherein, the health model of the system represented by subsystems, the system level represented by key subsystems, and the health model of non-key subsystems in any time period [jT, (j+h)T] is:

[0107]

[0108] in, is the health decision function of the whole system consisting of l+O subsystems in the time series [jT,(j+h)T], is the system health decision function consisting of l key subsystems in the time series [jT, (j+h)T], is the system health decision function consisting of O non-critical subsystems on the time series [jT, (j+h)T].

[0109] 7) Setting the number of continuous points for system status determination and the health threshold, and determining the health status of the entire liquid rocket engine system through the health model established in step 6).

[0110] 1) When satisfied

[0111]

[0112] All critical subsystems and non-critical systems of the engine are normal, the whole system is in a completely normal state, and the whole system can continue to work;

[0113] 2) When satisfied

[0114]

[0115] The non-critical systems of the engine are normal, but due to the failure of the critical systems, the whole machine is abnormal and needs to be stopped;

[0116] 3) When satisfied

[0117]

[0118] Although the non-critical system of the engine fails, the critical system is normal, so the whole engine system is normal and can continue to operate according to fault-tolerant control. However, it is necessary to continue to observe the status of the key components of the whole system to prevent the non-critical system failure from spreading to the critical system.

[0119] 4) When satisfied

[0120]

[0121] In view of the system-level health assessment decision composed of all subsystems and the health decision requirements supported by key subsystems, the present invention designs two parameters to be determined, namely, the number of continuous points h for system state judgment and the health threshold y. thre-h , so that the actual application needs can be met through parameter adjustment, and the accurate and timely judgment of the system-level health status can be achieved.

[0122] Take the health decision of a certain type of liquid rocket engine as an example to illustrate. Figures 5a to 6f The subsystem feature-level fusion result and its enlarged diagram show that the feature-level fusion still exceeds the preset threshold, making it difficult to accurately express the health degree. Figures 7a to 7d The evaluation results of the health of key subsystems by the traditional CRF method and the present invention are as follows: Figures 7a to 7d It can be seen that TTCRF can effectively reduce the interference effect of state evaluation and improve the accuracy of discrimination; Figure 8a and Figure 8bThe TTCRF health evaluation results under different thresholds are shown; Figure 9a and Figure 9b This is the discrimination result of the whole system using TTCRF.

[0123] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A decision-level fusion method for full system health assessment of liquid rocket engines, characterized in that: The following steps are involved: 1) According to the weights of each subsystem in the liquid rocket engine system and engineering requirements, several subsystems with high importance are selected for identification to construct a mathematical model of time tracking conditional random fields; 2) Extract the time domain index features and perform bilateral second-order synchronous compression transformation on the rapid-varying signals of liquid rocket engines to obtain the time domain index features and instantaneous frequency of the rapid-varying signals of liquid rocket engines, perform trend analysis on the slow-varying signals of liquid rocket engines, and use them for decision-level fusion and health assessment to obtain the time series feature vector; 3) The whole system is constructed by using the independent l key systems and O non-key subsystems in the liquid rocket engine, and the health matrix of the whole system is constructed by using the time series feature vector obtained in step 2), and the signal corresponding to the health matrix of the whole system is fused at the time tracking feature level to obtain the conditional probability model of TTCRF; 4) Based on the conditional random field theory, the observation sequence and the label sequence are modeled to obtain the conditional probability model, and then the rectangular window function for health status discrimination is introduced; 5) Based on the conditional probability model established in step 4), mark the health status of each subsystem and its time point, and obtain the probability deviation by giving the qualified parameters in the conditional probability model to determine the final qualified parameters of the conditional probability model; 6) Based on the conditional probability model obtained in step 5), a sub-state sequence set is constructed, and the health value space over a continuous time period is represented by the sub-state sequence set, and a health model is constructed for the system level represented by all subsystems, the system level represented by key subsystems, and non-key subsystems; 7) Setting the number of continuous points for system status determination and the health threshold, and determining the health status of the entire liquid rocket engine system through the health model established in step 6).

2. The decision-level fusion method for liquid rocket engine full system health assessment according to claim 1, characterized in that: The rapid-changing signal of the liquid rocket engine in step 1) is a vibration or fluid pulsation signal; the slow-changing signal of the liquid rocket engine includes the temperature, fluid pressure, flow rate and rotation speed of the liquid rocket engine.

3. The decision-level fusion method for liquid rocket engine full system health assessment according to claim 1 is characterized in that: There are l+O independent subsystems, and their health function vector is: G total (t)=[G1(t),G2(t)……G l (t),G l+1 (t),……G l+O (t)] T Where l is the key component / subsystem, O is the number of non-key subsystems, G s-total The state matrix formed for key components / subsystems, G o-total State matrix formed for non-critical subsystems.

4. The decision-level fusion method for liquid rocket engine full system health assessment according to claim 1, characterized in that: In step 3), for the discrete signal acquisition system, the sampling period is T. In the time period [0, kT], the system-level health indicator matrix and the state model of the discrete sequence composed of l+O subsystems are: Among them, G mk (kT) is the health of the mth subsystem at time kT, G *k (kT) is the health of the whole system at time kT, G Lk (kT) is the health of the system composed of L key subsystems at time kT, G Ok (kT) is the health of the system consisting of O non-critical subsystems at time kT.

5. The decision-level fusion method for liquid rocket engine full system health assessment according to claim 1, characterized in that: In step 4), based on the CRF theory, the observation sequence G o-total and marker sequence Modeling is performed on the CRF conditional probability model P(r i* |G i,j→j+h (jΔT)) is:

6. The decision-level fusion method for liquid rocket engine full system health assessment according to claim 1, characterized in that: The normalization factor in the conditional probability model in step 5) is: Among them, α is the unknown coefficient, [r i,j ] is the health label matrix of the ith subsystem at time jT, r i，j～k For r i,j The value of .

7. The decision-level fusion method for liquid rocket engine full system health assessment according to claim 1, characterized in that: The health model constructed at the system level of all subsystems, the system level of key subsystems and non-key subsystems in any discrete time period [jT, (j+h)T] is: in, is the health decision function of the whole system consisting of l+O subsystems in the time series [jT,(j+h)T], is the system health decision function consisting of l key subsystems in the time series [jT, (j+h)T], is the system health decision function consisting of O non-critical subsystems on the time series [jT, (j+h)T].

8. The decision-level fusion method for liquid rocket engine full system health assessment according to claim 7, characterized in that: In step 7), When satisfied All critical subsystems and non-critical systems are normal, and the liquid rocket engine is in full system condition; When satisfied If the non-critical system is normal, but the critical system fails, the entire liquid rocket engine system will be abnormal; When satisfied Then the non-critical system fails, the critical system is normal, and the entire liquid rocket engine system is normal; When satisfied Then there are faults in both critical subsystems and non-critical non-systems, and the entire liquid rocket engine system fails.

Images