Fault detection method based on fusion of isolation algorithm and evidence theory

Abstract

The invention discloses a fault detection method based on fusion of an isolation algorithm and an evidence theory. The method comprises the following steps: step 1, obtaining detection data of a plurality of sensors to obtain a time sequence data set; 2, constructing and training a global anomaly detection model and a local anomaly detection model; 3, obtaining a time sequence data sample x, and obtaining a global anomaly score and a local anomaly score; step 4, converting the abnormal score into basic probability assignment (BPA); 5, fusing the multi-source BPA evidences to obtain a fusion result m (A); and step 6, converting the fusion result m (A) into probability distribution so as to classify samples. According to the method, global anomalies and local anomalies can be effectively detected at the same time, anomaly scores are mapped into basic probability distribution, multi-source BPA is fused through the D-S evidence theory so as to fuse probability distribution decisions, forced classification defects of manual threshold values are avoided, and the accuracy is still maintained under the condition that sensor data are incomplete.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing based on specific calculation models, and more particularly, to a fault detection method based on isolation algorithm and evidence theory fusion. BACKGROUND

[0002] Complex industrial equipment, such as aero-engines, gas turbines, etc., is a highly integrated electromechanical system, and its operation safety and reliability are of great importance. In order to prevent major accidents, it is necessary to monitor the running state in real time so as to discover abnormalities in the early stage of fault occurrence. At present, the data-driven fault detection method has become the mainstream research direction in this field due to its advantages of not needing to establish an accurate physical model and being able to process high-dimensional nonlinear data.

[0003] The existing data-driven fault detection method relies on a single detection model in actual application, and any single model has inherent limitations, which is difficult to comprehensively capture complex and diverse fault modes, specifically: 1) Global anomaly detection method: this kind of method represented by the isolation forest iForest algorithm identifies anomalies by quickly isolating sparse points in the data space. They have high efficiency for detecting global, isolated, and large deviation mutation faults, such as the sharp jump of a certain parameter caused by foreign object damage. However, for a large number of early and weak faults, their characteristics are not globally isolated, but show a small and synchronous deviation of multiple related sensor parameters in the local high-dimensional space. Such fault points are not prominent in the global data distribution, and therefore are easily ignored by global detection methods.

[0004] 2) Local anomaly detection method: this kind of method represented by the local outlier factor LOF algorithm and the isolation-based nearest neighbor ensemble iNNE algorithm identifies anomalies by evaluating the density of data points relative to their neighboring data points, which exactly makes up for the shortcomings of global methods and is very sensitive to the above-mentioned local and dense anomalies. However, its limitations are also obvious: first, when dealing with high-dimensional data, distance or density-based calculations will face the "curse of dimensionality", resulting in a sharp rise in calculation cost; second, its performance is extremely sensitive to the integrity and quality of sensor data, and the absence of a single data point or noise interference can seriously distort the local density estimate, resulting in a large number of false positives or false negatives.

[0005] 3) Deep learning-based anomaly detection methods: Deep learning models represented by Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) have strong non-linear modeling capabilities, but their application in high-reliability fields such as aircraft engines faces two major bottlenecks: First, they rely heavily on large amounts of high-quality labeled fault samples for training, but real-world engine fault data is extremely scarce and costly to obtain, making it difficult to support effective training of deep models. Second, their "black box" nature results in a lack of transparency and explainability in the decision-making process, which is a major drawback in the highly safety-critical aerospace industry.

[0006] In addition, most existing fault detection methods rely on manually set thresholds. These methods usually output a continuous anomaly score, which must be set to a fixed threshold by technicians based on experience or through extensive offline experiments to distinguish between normal and fault states. However, a fixed threshold is difficult to adapt to complex and changing operating conditions and data distributions, and improper setting can severely affect detection accuracy. This reliance on thresholds not only increases application costs but also reduces the automation level and adaptive ability of the detection system in different scenarios.

[0007] In summary, there is an urgent need for a new automated fault detection method that can overcome the limitations of single models, effectively detect both global and local fault patterns, maintain high robustness when sensor data is incomplete, and eliminate the need for manual threshold setting. SUMMARY

[0008] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art. The present application provides a fault detection method based on isolation algorithm and evidence theory fusion, which has a simple structure and reasonable design. The present application can effectively detect both global and local anomalies, map the anomaly score to the basic probability assignment, fuse multiple source BPA through D-S evidence theory, and make a decision based on the fused probability distribution to avoid the forced classification defect of manual threshold setting. In the case of incomplete sensor data, the fault detection accuracy is still maintained.

[0009] To solve the above technical problems, the technical solution adopted by the present application is: The present application provides a fault detection method based on isolation algorithm and evidence theory fusion, characterized by the following steps: Step 1: Deploy multiple sensors on the equipment to be detected to obtain time series data set X from the detection data of multiple sensors, X = [X 1-t ,...,X i-t ,...,X n-t ], where X i-t represents the time series data of the i-th sensor, and n represents the number of sensors. Step two, constructing and training a fault detection model, the fault detection model including a global anomaly detection model using iForest global isolation algorithm and a local anomaly detection model using iNNE local isolation algorithm, the global anomaly detection model including n isolation trees; Step three, obtaining a time series data sample x, inputting the time series data sample x into the global anomaly detection model and the local anomaly detection model respectively, the global anomaly detection model outputting a global anomaly score, and the local anomaly detection model outputting a local anomaly score; Step four, converting the global anomaly score and the local anomaly score into basic probability assignments respectively through membership functions to obtain multiple BPA evidences; Step five, evidence fusion: combining multiple BPA evidences by using a D-S evidence theory combination rule formula to obtain m(A), Step six, applying decision probability conversion to the fused evidence m(A) to obtain P({N}) and P({F}), if P({N})>(F), the sample is classified as normal, otherwise, it is classified as fault.

[0010] The fault detection method based on isolation algorithm and evidence theory fusion, characterized in that: in step four, if the global anomaly score S1>0.5, the basic probability assignment is calculated based on the formula If the global anomaly score S1<0.5, the basic probability assignment is calculated based on the formula If the global anomaly score S1=0.5, =1; The membership function for the global anomaly score under the fault state is represented as The membership function for the global anomaly score under the normal state is represented as The basic probability assignment for the fault proposition under the global dimension is represented as The basic probability assignment for the normal proposition under the global dimension is represented as The basic probability assignment for the normal or fault uncertainty proposition under the global dimension is represented as

[0011] The fault detection method based on isolation algorithm and evidence theory fusion, characterized in that: in step four, if the local anomaly score S2>0.5, the basic probability assignment is calculated based on the formula If the local anomaly score S2<0.5, the basic probability assignment is calculated based on the formula If the local anomaly score S2=0.5, =1; The membership function for the local anomaly score under the fault state is represented as The membership function for the local anomaly score under the normal state is represented as The basic probability assignment for the fault proposition under the local dimension is represented as This represents the basic probability allocation for normal propositions within a local dimension. This represents the basic probability allocation for uncertain propositions regarding normal or fault conditions within a local dimension.

[0012] The aforementioned fault detection method based on the fusion of isolation algorithm and evidence theory is characterized in that: in step five, the DS evidence theory combination rule is as follows: In this context, m1(B) and m2(C) are two sets of BPA evidence to be fused, m(A) is the fused BPA evidence, and A, B, and C are the identification frameworks. A subset of, where K is the conflict factor. .

[0013] The aforementioned fault detection method based on the fusion of isolation algorithm and evidence theory is characterized in that: in step six, , where y is the recognition frame elements, ={N,F}, where A is the recognition frame. A subset of A, where |A| is the number of elements in set A, and m(A) is the basic probability assignment corresponding to focal element A in the fused evidence.

[0014] The aforementioned fault detection method based on the fusion of isolation algorithm and evidence theory is characterized by the following specific method for the global anomaly detection model to output a global anomaly score: Input the time-series data sample x into the global anomaly detection model, traverse all isolated trees, record the number of edges h(x) that x traverses from the root node to the leaf node in each tree, calculate the expected path length E(h(x)), and then use the formula... The expected path length E(h(x)) is converted into a global anomaly score S1, where c(n) represents the normalization factor and n represents the number of training samples used to construct the isolation tree.

[0015] The aforementioned fault detection method based on the fusion of isolation algorithm and evidence theory is characterized in that: the number of isolation trees n=100, and the number of subsamples used in each isolation tree is 256.

[0016] The aforementioned fault detection method based on the fusion of isolation algorithm and evidence theory is characterized by the following specific method for the local anomaly detection model to output local anomaly scores: Input the time-series data sample x into the local anomaly detection model, traverse all hypersphere groups, record the collision count cnn(x) of x in each hypersphere group, and determine the expected local cardinality η of the hypersphere group. cnn(x) Using formulas Calculate the local anomaly score for each hypersphere group. For all hyperspherical groups, the local anomaly fractions A weighted average calculation is performed to obtain a local anomaly score S2.

[0017] The fault detection method based on the isolation algorithm and the evidence theory fusion has the characteristics that the number of hyper-sphere groups p = 100, and each hyper-sphere group contains 256 hyper-spheres.

[0018] The fault detection method based on the isolation algorithm and the evidence theory fusion has the characteristics that the to-be-detected equipment is a dual-rotor turbofan engine, the detection data obtained by the plurality of sensors at least include physical parameters of a low-pressure rotor speed, a high-pressure rotor speed, a high-pressure turbine inlet, a high-pressure turbine outlet, a low-pressure turbine outlet, an outer duct outlet and an exhaust nozzle inlet, and the physical parameters at least include one of temperature, pressure and gas flow.

[0019] Compared with the prior art, the present application has the following advantages: 1. The structure of the present application is simple, reasonable in design, and easy to realize and use.

[0020] 2. The present application combines a global anomaly detection model and a local anomaly detection model, simultaneously utilizes the calculation efficiency of iForest and the advantage of identifying global isolated anomalies, and the high sensitivity of iNNE to subtle related deviations in the local neighborhood, realizes complementary detection coverage, and improves the comprehensiveness and reliability of fault detection.

[0021] 3. The present application converts the global anomaly score and the local anomaly score into basic probability assignments BPA of normal, fault and uncertainty propositions through membership functions, and this process is based on the abnormal degree of data itself, rather than artificial preset threshold to determine the basic probability assignment, thereby eliminating the inherent bias caused by artificial setting of fault threshold.

[0022] 4. In the present application, the global anomaly score and the local anomaly score are converted into basic probability assignments BPA, avoiding the one-sidedness of a single evidence source, and the complementary characteristics of the global anomaly detection model and the local anomaly detection model make the multiple BPA evidences have natural diversity, which can effectively alleviate evidence conflict.

[0023] 5. The present application fuses multiple source BPA through D-S evidence theory, and the final decision depends on the fused probability distribution, without the need for artificial setting of decision threshold, avoiding the forced classification defects of artificial threshold.

[0024] 6. In the present application, the isolation algorithm does not depend on complete data, and the evidence theory can fuse incomplete information, so that even in the case of incomplete sensor data, the classification accuracy will not decrease significantly.

[0025] In summary, the application can effectively detect global and local anomalies at the same time, map the anomaly score to the basic probability assignment, fuse the multi-source BPA through the D-S evidence theory, make the probability distribution decision through fusion, avoid the forced classification defects of the artificial threshold, and still maintain the accuracy in the case of incomplete sensor data.

[0026] The technical solutions of the application will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 The method flowchart of the application.

[0028] Figure 2 The cross-sectional symbol diagram of the dual-rotor turbofan engine.

[0029] Figure 3 The membership function curve of the global anomaly score in the fault and normal state of the application. DETAILED DESCRIPTION

[0030] The application will be further described in detail below with reference to the drawings and embodiments.

[0031] It should be noted that the embodiments and features in the embodiments in the application can be combined with each other without conflict. The application will be described in detail below with reference to the drawings and embodiments.

[0032] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit the exemplary embodiments according to the application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, there is a feature, step, operation, device, component and / or combination thereof.

[0033] It should be noted that the terms "first", "second" and the like in the specification and claims of the application and the above-described drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0034] For purposes of the description hereinafter, spatially relative terms, such as "above", "below", "up", "down", "right", "left", "vertical", "horizontal", "top", "bottom", "under", "over", "between" and the like, can be used herein for ease of description to describe one element or feature's relationship to another element(s) or feature(s) as illustrated in the figures. It will be understood that the spatially relative terms are intended to encompass different orientations of the device in use or operation in addition to the orientations depicted in the figures. For example, if a device is inverted in the figure, a spatially relative term such as "above" or "up" can indicate "below" or "down" in the inverted orientation. Thus, the exemplary term "above" can encompass both an orientation of above and below. The device can be otherwise oriented (rotated 90 degrees or at other orientations) and the spatially relative descriptors used herein interpreted accordingly. The terms "first", "second", "third", etc. can be used herein to describe various elements, regions, layers and / or sections but are not intended to be used herein to designate importance or a particular order of use.

[0035] As shown in the drawings, Figure 1 a fault detection method based on isolation algorithm and evidence theory fusion of the present application comprises the following steps: Step one, deploy multiple sensors on the equipment to be detected, respectively acquire detection data of multiple sensors, get time series data set X, X=[X 1-t ,...,X i-t ,...,X n-t ], wherein X i-t represents the time series data of the i-th sensor, and n represents the number of sensors.

[0036] In one possible embodiment, the equipment to be detected is a dual-rotor turbofan engine, and the detection data at least includes physical parameters of low-pressure rotor speed, high-pressure rotor speed, high-pressure turbine inlet, high-pressure turbine outlet, low-pressure turbine outlet, outer duct outlet and exhaust nozzle inlet; the physical parameters at least include one of temperature, pressure and gas flow.

[0037] In one possible embodiment, a simulation model is built in Matlab or Simulink environment relying on T-MATS toolbox, real representative sensor data of the gas path system is generated, and controllable experimental conditions are ensured while keeping consistency with the actual engine behavior.

[0038] The fault-free data set y healthy (t) is described by a set of nonlinear state space equations: y healthy (t)=g(x(t),u(t)), wherein x(t) represents a state vector, for a dual-rotor turbofan engine, x(t) mainly includes low-pressure rotor speed N L and high-pressure rotor speed N Hu(t) represents the control input vector, which typically includes physical parameters of the key cross-section, including at least one of temperature, pressure, and gas flow rate.

[0039] Figure 2 A symbolic diagram of a cross-section of a twin-rotor turbofan engine, such as... Figure 3 As shown, cross-section symbol S4 represents the high-pressure turbine inlet, cross-section symbol S45 represents the high-pressure turbine outlet, cross-section symbol S5 represents the low-pressure turbine outlet, and cross-section symbol S17 represents the bypass duct outlet. In one possible embodiment, y healthy (t)=[T S4 ,P S45 ,P S5 W S17 N L N H ] T Among them, T S4 P represents the temperature at the inlet of the high-pressure turbine, with cross-section symbol S4. S45 P represents the pressure at the outlet of the high-pressure turbine, with the cross-sectional symbol S45. S5 W represents the pressure at the outlet of the low-pressure turbine, with the cross-sectional symbol S5. S17 This indicates the gas flow rate at the outlet of the outer bypass duct, with the cross-section symbol S17.

[0040] Fault Dataset , where y true (t) represents the true physical value of a single sensor, i.e., y healthy (t) The final scalar component, where parameters G, D, and B represent the gain factor, drift rate, and bias, respectively. t0 represents the zero-mean Gaussian white noise component, and t0 represents the fault injection time.

[0041] The simulation model generates data samples at a sampling interval of 0.04 seconds, forming a dataset of 1150 data samples covering 46 seconds of operation. Systematic sensor failures are intentionally introduced within time intervals of 0 to 15 seconds.

[0042] Fault-free dataset y healthy (t) and fault dataset y faulty (t) constitutes the time series dataset X.

[0043] Step 2: Construct and train the fault detection model. The fault detection model includes a global anomaly detection model using the iForest global isolation algorithm and a local anomaly detection model using the iNNE local isolation algorithm. The global anomaly detection model includes n isolation trees.

[0044] Training the global anomaly detection model: input the training data set D, set the number of isolation trees n, the number of sub-samples used by each isolation tree, and the height limit of the isolation tree; build each isolation tree until the preset tree height limit is reached, and output a global anomaly detection model containing n isolation trees. The training data set D is a subset extracted from the time series data set X, containing both the fault-free data set and the fault data set, with the same data dimension as the time series data set X, used to train the global anomaly detection model.

[0045] Training the local anomaly detection model: input the training data set D' containing only normal operation data, set the number of hypersphere groups p and the number of hyperspheres contained in each hypersphere group; build each hypersphere group, with randomly selected training samples as the center and the distance from the sample to its nearest neighbor as the radius to build adaptive hyperspheres, and output a local anomaly detection model containing p groups of hypersphere groups. The training data set D' is a subset selected from the time series data set X, containing only the fault-free data set, with the same data dimension as the time series data set X, used to train the local anomaly detection model.

[0046] To effectively reduce the randomness caused by random sampling process and achieve the best balance between computational efficiency and detection accuracy, the algorithm parameters are systematically configured as follows: the number of isolation trees n = 100, and the number of sub-samples used by each isolation tree is 256. The number of hypersphere groups p = 100, and the number of hyperspheres contained in each hypersphere group is 256.

[0047] The global anomaly detection model uses the iForest global isolation algorithm. When processing aero-engine detection data, the core advantage of the iForest global isolation algorithm lies in its high-dimensional data processing efficiency and global burst anomaly recognition capability. The core mechanism of the iForest global isolation algorithm makes it good at identifying global sparse anomalies and can quickly capture severe faults such as sudden and severe deviation of key parameters caused by foreign object damage of aero-engines, providing efficient detection support for such global anomalies that may cause catastrophic consequences.

[0048] The local anomaly detection model uses the iNNE local isolation algorithm. When processing aero-engine detection data, the core advantage of the iNNE local isolation algorithm lies in its precise capture of local high-dimensional subspace anomalies. The iNNE local isolation algorithm can effectively depict the local density characteristics and neighborhood relationships of data, accurately identify small synchronous deviations of multiple sensor parameters and other such local dense anomalies, and has high sensitivity in capturing early performance degradation signals of aero-engines, which can make up for the omission of single global detection of subtle local deviations.

[0049] Step three, obtain the time series data sample x, input the time series data sample x into the global anomaly detection model and the local anomaly detection model respectively, the global anomaly detection model outputs the global anomaly score, and the local anomaly detection model outputs the local anomaly score.

[0050] In one possible embodiment, the time series data sample x is input into the global anomaly detection model, all isolation trees are traversed, the number of edges h(x) that x passes through from the root node to the leaf node in each tree is recorded, the expected path length E(h(x)) is calculated, and the expected path length E(h(x)) is converted into the global anomaly score S1 by the formula

[0051] The time series data sample x is input into the local anomaly detection model, all hyper-spherical groups are traversed, the collision count cnn(x) of x in each hyper-spherical group is recorded, and the expected local cardinality η of the hyper-spherical group is determined cnn(x) The local anomaly score S2 under each hyper-spherical group is calculated by the formula The local anomaly scores S2 under all hyper-spherical groups are weighted and averaged to obtain the local anomaly score S2.

[0052] Aero-engine failures can either be globally sparse sudden serious anomalies or locally dense early subtle deviations, and a single detection paradigm cannot comprehensively cover them. Simultaneously using the global anomaly detection model to calculate the global anomaly score S1 and using the local anomaly detection model to calculate the local anomaly score S2 can cope with the diversified manifestations of aero-engine failures, guaranteeing both rapid identification of sudden serious failures and timely capture of early subtle degradation signals, solving the detection limitations of a single model, and improving the comprehensiveness and reliability of fault detection.

[0053] Both the global anomaly detection model and the local anomaly detection model quantify the anomaly degree of data samples by calculating normalized anomaly scores. The closer the global anomaly score S1 or the local anomaly score S2 is to 1, the more likely the sample is abnormal. The closer the global anomaly score S1 or the local anomaly score S2 is to 0, the more likely the sample exhibits characteristics consistent with normal operation behavior.

[0054] Step four, the global anomaly score and the local anomaly score are respectively converted into basic probability assignments through membership functions to obtain multiple BPA evidences.

[0055] The global anomaly score S1 and the local anomaly score S2 respectively calculate membership function values based on membership functions. The membership function value of the dominant trend is assigned to the confidence of a single state, and the membership function value of the non-dominant trend is assigned to uncertainty, thereby converting the global anomaly score S1 and the local anomaly score S2 into BPA on the recognition framework Θ={N,F}. F represents failure, and N represents normal.

[0056] ​​​In one possible embodiment, if the global anomaly score S1>0.5, the basic probability assignment is calculated based on the formula If the global anomaly score S1<0.5, the basic probability assignment is calculated based on the formula If the global anomaly score S1=0.5, =1. represents the membership function for the global anomaly score under the fault state, represents the membership function for the global anomaly score under the normal state; represents the basic probability assignment for the fault proposition under the global dimension, represents the basic probability assignment for the normal proposition under the global dimension, represents the basic probability assignment for the normal or fault uncertainty proposition under the global dimension.

[0057] It should be noted that the performance degradation of the aero-engine under test device is generally not linear. and are nonlinear cosine functions, which change gently at both ends of the global anomaly score S1 close to 0 and 1, suppress the influence of small fluctuations of the score on the generation of BPA, can effectively filter system noise, and reduce the false alarm rate; while changing rapidly near 0.5, it can more subtly reflect the subtle changes in state tendency, and improve the ability to capture early weak fault signs. This feature can better simulate the nonlinear gradual change process of the device under test from normal to fault.

[0058] Figure 3 is the membership function curve of the global anomaly score under the fault and normal state, as shown in Figure 3 , and intersect at S1=0.5, so this point represents the maximum fuzzy point of the normal and fault states under the global dimension, corresponding to the maximum uncertainty state in the D-S framework. In this fuzzy case, without additional evidence, it is impossible to classify the data sample as abnormal or normal.

[0059] In one possible embodiment, the global anomaly score S1 is 0.7023, 0.7023>0.5, which is more inclined to fault, so the membership function for the global anomaly score under the fault state dominates the trend, and the formula is used to calculate , =0.7968, the membership function value of the dominant trend is assigned to the confidence of the single state, so =0.7968, the membership function value of the non-dominant trend is assigned to the uncertainty, =0.2032.

[0060] If the local anomaly score S2>0.5, the basic probability assignment is calculated based on the formula If the local anomaly score S2<0.5, the basic probability assignment is calculated based on the formula If the local anomaly score S2=0.5, =1. represents the membership function of the local anomaly score S2under the fault state, represents the membership function of the local anomaly score S2under the normal state, represents the basic probability assignment of the fault proposition under the local dimension, represents the basic probability assignment of the normal proposition under the local dimension, represents the basic probability assignment of the normal or fault uncertainty proposition under the local dimension.

[0061] It should be noted that, and intersect at S2=0.5, so this point represents the maximum fuzzy point of the normal and fault states under the local dimension, corresponding to the maximum uncertainty state in the D-S framework, in this fuzzy case, without additional evidence, it is impossible to classify the data sample as abnormal or normal.

[0062] In one possible embodiment, the local anomaly score S2 L is 0.9015, since 0.9015>0.5, it is more inclined to the fault, so the membership function of the local anomaly score S2 L under the fault state dominates the trend, and the formula is used to calculate , =0.9763, the membership function value of the dominant trend is assigned to the belief of the single state, so =0.9763, the membership function value of the non-dominant trend is assigned to the uncertainty, =0.0237.

[0063] The global anomaly score S1and the local anomaly score S2are converted into the basic probability assignments of the normal, fault and uncertainty propositions through the membership functions, avoiding the one-sidedness of a single evidence source, the complementary characteristics of the global anomaly detection model and the local anomaly detection model make the multiple BPA evidences have natural diversity, which can effectively alleviate the evidence conflict, and provide a more robust fusion basis for the D-S evidence theory combination rule.

[0064] Step five, evidence fusion: multiple BPA evidences are fused to obtain m(A) by using the D-S evidence theory combination rule formula.

[0065] The D-S evidence theory combination rule is:​ , where m1(B) and m2(C) are two sets of BPA evidence to be fused, m(A) is the fused BPA evidence, and A, B, and C are subsets of the frame of discernment , and K is the conflict factor .

[0066] In a possible embodiment, the global anomaly score S1 is 0.7023 =0.7968 =0.2032; the local anomaly score S2 is 0.9015 =0.9763 =0.0237.

[0067] , the focal elements B of the evidence m1 include {F} and {N,F}, and the focal elements C of the evidence m2 include {F} and {N,F}. The intersection of any two focal elements, such as {F} and {F}, {F} and {N,F}, is not an empty set. Therefore, there is no focal element combination that satisfies " ", so the conflict factor K = 0.

[0068] Calculate the fused evidence according to the D-S evidence theory combination rule, m({F}) and m({N,F}) , .

[0069] Step 6: Apply the decision probability transformation to the fused evidence m(A) to obtain P({N}) and P({F}). If P({N}) > P({F}), the sample is classified as normal, otherwise it is the opposite. In evidence theory, the Pignistic probability transformation decision probability transformation converts the basic probability assignment into a classical probability distribution for decision analysis. The Pignistic probability transformation decision probability transformation rule is: , where y is an element of the frame of discernment , ={N,F}, A is a subset of the frame of discernment , |A| is the number of elements in the set A, and m(A) is the basic probability assignment corresponding to the focal element A in the fused evidence.

[0070] In actual use , , since P({F}) > P({N}), it is finally determined that the device state corresponding to the time-series data sample x is a fault, and the detection result is output. Otherwise, if P({F}) < P({N}), it is determined that the device state corresponding to the time-series data sample x is normal.

[0071] The above is only an embodiment of the present application, and does not limit the present application. Any simple modification, change and equivalent structural change of the above embodiment according to the technical essence of the present application are still within the protection scope of the technical solution of the present application.

Claims

1. A fault detection method based on the fusion of isolation algorithm and evidence theory, characterized in that, Includes the following steps: Step 1: Deploy multiple sensors on the device to be tested, and acquire the detection data from each sensor to obtain a time-series dataset X, where X = [X...]. 1-t ,...,X i-t ,...,X n-t ], where X i-t This represents the time-series data of the i-th sensor, where n represents the number of sensors; Step 2: Construct and train the fault detection model. The fault detection model includes a global anomaly detection model using the iForest global isolation algorithm and a local anomaly detection model using the iNNE local isolation algorithm. The global anomaly detection model includes n isolation trees. Step 3: Obtain time series data sample x, and input the time series data sample x into the global anomaly detection model and the local anomaly detection model respectively. The global anomaly detection model outputs the global anomaly score, and the local anomaly detection model outputs the local anomaly score. Step 4: Convert the global anomaly score and the local anomaly score into basic probability assignments using membership functions to obtain multiple BPA evidences; Step 5: Evidence Fusion: Using the DS evidence theory combination rule formula, multiple BPA pieces of evidence are fused to obtain m(A). Step 6: Apply decision probability transformation to the fused evidence m(A) to obtain P({N}) and P({F}). If P({N}) > ({F}), the sample is classified as normal; otherwise, it is classified as normal.

2. The fault detection method based on the fusion of isolation algorithm and evidence theory according to claim 1, characterized in that: In step four, if the global anomaly score S1 > 0.5, based on the formula... Calculate the basic probability assignment; if the global anomaly score S1 < 0.5, based on the formula... Calculate the basic probability distribution; if the global anomaly score S1 = 0.5, =1; This represents the membership function for the global anomaly score under fault conditions. This represents the membership function for global anomaly scores under normal conditions. This represents the basic probability allocation for fault propositions at the global level. This represents the basic probability allocation for normal propositions in the global dimension. This represents the basic probability allocation for uncertain propositions regarding normal or fault conditions at the global level.

3. The fault detection method based on the fusion of isolation algorithm and evidence theory according to claim 1, characterized in that: In step four, if the local anomaly score S2 > 0.5, based on the formula... Calculate the basic probability distribution; if the local anomaly score S² < 0.5, based on the formula... Calculate the basic probability distribution; if the local anomaly score S² = 0.5, =1; This represents the membership function for local anomaly scores under fault conditions. This represents the membership function for local anomaly scores under normal conditions. This represents the basic probability allocation for fault propositions within a local dimension. This represents the basic probability allocation for normal propositions within a local dimension. This represents the basic probability allocation for uncertain propositions regarding normal or fault conditions within a local dimension.

4. A fault detection method based on the fusion of isolation algorithm and evidence theory as described in claim 1, characterized in that: In step five, the DS evidence theory combination rule is as follows: In this context, m1(B) and m2(C) are two sets of BPA evidence to be fused, m(A) is the fused BPA evidence, and A, B, and C are the identification frameworks. A subset of, where K is the conflict factor. .

5. A fault detection method based on the fusion of isolation algorithm and evidence theory as described in claim 1, characterized in that: In step six, , where y is the recognition frame elements, ={N,F}, where A is the recognition frame. A subset of A, where |A| is the number of elements in set A, and m(A) is the basic probability assignment corresponding to focal element A in the fused evidence.

6. A fault detection method based on the fusion of isolation algorithm and evidence theory according to claim 1, characterized in that: The specific method for the global anomaly detection model to output the global anomaly score is as follows: Input the time-series data sample x into the global anomaly detection model, traverse all isolated trees, record the number of edges h(x) that x traverses from the root node to the leaf node in each tree, calculate the expected path length E(h(x)), and then use the formula... The expected path length E(h(x)) is converted into a global anomaly score S. G , where c(n) represents the normalization factor and n represents the number of training samples used to construct the isolation tree.

7. A fault detection method based on the fusion of isolation algorithm and evidence theory according to claim 6, characterized in that: The number of isolation trees is n=100, and the number of subsamples used in each isolation tree is 256.

8. A fault detection method based on the fusion of isolation algorithm and evidence theory according to claim 1, characterized in that: The specific method for the local anomaly detection model to output local anomaly scores is as follows: Input the time-series data sample x into the local anomaly detection model, traverse all hypersphere groups, record the collision count cnn(x) of x in each hypersphere group, and determine the expected local cardinality η of the hypersphere group. cnn(x) Using formulas Calculate the local anomaly score for each hypersphere group. For all hyperspherical groups, the local anomaly fractions The local anomaly score S is obtained by performing a weighted average calculation. L .

9. A fault detection method based on the fusion of isolation algorithm and evidence theory according to claim 8, characterized in that: The number of hypersphere groups is p=100, and each hypersphere group contains 256 hyperspheres.

10. A fault detection method based on the fusion of isolation algorithm and evidence theory according to claim 1, characterized in that: The device to be tested is a dual-rotor turbofan engine. The detection data obtained by multiple sensors include at least the low-pressure rotor speed, the high-pressure rotor speed, and the physical parameters of the high-pressure turbine inlet, high-pressure turbine outlet, low-pressure turbine outlet, bypass duct outlet, and exhaust nozzle inlet; the physical parameters include at least one of temperature, pressure, and gas flow rate.

Images