Self-renewing continuous learning artificial immune system turbo pump fault diagnosis method

Abstract

The present application relates to a kind of self-renewal continuous learning artificial immune system turbine pump fault diagnosis method, comprising: data sample, detector and classification index are respectively corresponding with antigen, immune cell and affinity, establish the artificial immune system model with hyperrectangular body as memory cell;The vibration data of turbine pump operation is randomly divided into two groups, one group is marked as antibody for cell training, another group is not marked as antigen for test;Antibody is input into the model and is initialized, and preliminary memory cell is generated;The cell of the largest number of memory cells of one kind is defined as the super memory cell of this kind, and the parent cell of super memory cell is defined as totipotent stem cell, expand totipotent stem cell and super memory cell, obtain the model after training;Antigen is input into the model after training, and antigen is used to activate memory cell or empty cell in the artificial immune system model, realize fault type identification and the continuous learning of model.

Description

TECHNICAL FIELD

[0001] The application relates to a self-updating continuous learning artificial immune system turbo pump fault diagnosis method and belongs to the technical field of mechanical fault diagnosis. BACKGROUND

[0002] The safe development, successful launch and smooth recovery of a spacecraft product play a very important role in national security, social stability, economic prosperity and the safety of people's lives and property. However, due to the particularity of the spacecraft product, it usually carries flammable and explosive high-energy fuel and can generate huge energy. The spacecraft product operates at high temperature and high speed, and once an accident occurs, it will cause great destructive and killing effects, threatening the safety of related infrastructure, social resources and the natural environment, and even the safety of the staff or the people.

[0003] In recent years, spacecraft accidents have occurred frequently at home and abroad.

[0004] Therefore, improving the reliability of the spacecraft product while ensuring the safety of the equipment to protect the safety of the personnel and the equipment and avoid property loss and environmental damage is the focus of the field of spacecraft research. The new hydrogen-oxygen engine is the heart of the space rocket, providing huge power for the flight of the rocket. The turbo pump is a key component of the hydrogen-oxygen engine and can realize the pressurization of liquid hydrogen and liquid oxygen propellants, increase the pressure of the propellants to tens of thousands of kilopascals, and then send the propellants into the thrust chamber of the rocket engine for combustion to provide energy for the operation of the engine. Therefore, accurate diagnosis of the faults of the turbo pump and protection of the normal operation of the space rocket engine are of great significance to the development of the space industry.

[0005] Due to the working characteristics of the turbo pump of the space rocket engine, actual fault samples are difficult to obtain at one time but are obtained gradually over time. The traditional artificial intelligence fault diagnosis method which is good at batch learning mode only learns the newly added samples, and the catastrophic forgetting of the learned samples occurs, the performance on the old task is significantly reduced, and the accuracy of fault diagnosis is affected. If the fault recognition model is retrained using new and old data, a large amount of time and resources will be consumed, resulting in low efficiency. To solve the above problems, the application provides a continuous learning fault diagnosis method which has the ability to integrate new knowledge from new and old data and refine existing old knowledge, and can avoid significant interference of new input data on existing knowledge. SUMMARY

[0006] The technical problem solved by the application is that, in view of the low plasticity of the batch learning mode of the traditional artificial intelligence fault diagnosis method in the prior art, the catastrophic forgetting cannot be avoided, the model is easily affected by new input data, and the stability is poor, and the application provides a self-updating continuous learning artificial immune system turbo pump fault diagnosis method.

[0007] The technical scheme of the present application is as follows: a self-updating continuous learning artificial immune system turbine pump fault diagnosis method, comprising the following contents:

[0008] The data sample, detector and classification index are respectively corresponding to the antigen, immune cell and affinity, and an artificial immune system model with a hyper-rectangular body as a memory cell is established;

[0009] The vibration data of the turbine pump operation is randomly divided into two groups, one of which is marked as an antibody for cell training, and the other group is unmarked as an antigen for testing; the marking contains vibration data and corresponding fault types;

[0010] The antibody is input into the artificial immune system model and initialized to preliminarily generate memory cells;

[0011] The cells with the largest density of a type of memory cells are defined as super memory cells of the type, and the parent cells of the super memory cells are defined as totipotent stem cells, the totipotent stem cells and the super memory cells are expanded by using a double cloning cell strategy to obtain a trained artificial immune system model;

[0012] The antigen is input into the trained artificial immune system model, and the antigen is used to activate the memory cells or empty cells in the artificial immune system model to realize fault type identification and continuous learning of the model.

[0013] Preferably, the use of the antigen to activate the memory cells or empty cells in the artificial immune system model to realize fault type identification and continuous learning of the model comprises:

[0014] When the activated object of the antigen is a memory cell, the antigen is recognized as the same type as the memory cell;

[0015] When the activated object of the antigen is an empty cell, the type of the antigen is determined by the adjacent memory cell of the empty cell, the empty cell is evolved into a new memory cell through the affinity between the antigen and the adjacent memory cell, and the type of the antigen is determined by the new memory cell.

[0016] Preferably, the calculation formula of the affinity is:

[0017]

[0018] In the formula, q i is the number of antibodies in the memory cell i, g i is the generation of the memory cell i, and d is the Euclidean distance between the antigen and the memory cell i;

[0019] The memory cell corresponding to the maximum affinity is found, the type of the new memory cell is the same as that of the memory cell, and the type of the antigen is further determined.

[0020] Preferably, when the affinity cannot determine the antigen type, a weight is introduced to assist in the identification, that is, the antigen type is further comprehensively evaluated by combining the affinity and the cell weight.

[0021] Cell weight formula:

[0022]

[0023] e q is the distance between the qth original antibody in the memory cell and the empty cell center, q∈[1,2,…,u]; k∈[1,2,…,v] represents the kth cloned antibody in the memory cell; A k is the coordinate of the antibody generated by cloning in the memory cell; A k′ is the coordinate of the original antibody of the clone; g i is the generation number of the memory cell i, and d is the Euclidean distance between the antigen and the memory cell i.

[0024] Preferably, the comprehensive evaluation of the antigen type by combining the affinity and the cell weight includes:

[0025] According to the formula δ=aff×70%+W×30%, the comprehensive evaluation index δ of each adjacent memory cell of the empty cell is calculated, the memory cell corresponding to the maximum δ is found, the empty cell is evolved into the memory cell, and thus the type of the antigen is determined; the above aff is the affinity between the antigen and the adjacent memory cell.

[0026] Preferably, the affinity that cannot determine the antigen type includes different types of adjacent memory cells corresponding to the same affinity value.

[0027] Preferably, when the object activated by the antigen is an empty cell, if the generation number of the empty cell is lower than that of the super memory cell, the empty cell is first split to the same generation number as the super memory cell; if the generation number of the empty cell is higher than or equal to that of the super memory cell, a new memory cell with the same generation number as the empty cell is directly generated, and the type of the new memory cell is determined according to the voting principle of the adjacent memory cells.

[0028] Preferably, the voting principle is to determine the type of the new memory cell according to the size of the affinity aff between the antigen and the adjacent memory cells, and if the aff cannot determine, the comprehensive evaluation of the antigen type is further combined with the comprehensive evaluation index δ of the adjacent memory cells.

[0029] Preferably, the double-cloning cell strategy is an upward cloning strategy or a same-generation adjacent cloning strategy.

[0030] Compared with the prior art, the present application has the following beneficial effects:

[0031] The self-renewing continuous learning artificial immune system turbine pump fault diagnosis method provided by the application draws lessons from the automatic regulation continuous learning mechanism of omnipotent stem cells on immune response in a biological immune system, proposes a new cell generation strategy, enhances the continuous learning ability of the model, improves the recognition efficiency, and well realizes the diagnosis of turbine pump faults; the SUC is used for double cloning of memory cells, the generated omnipotent stem cells and super memory cells have strong recognition advantages; in addition, the affinity and weight combination method is used to evaluate the antigen type, the weight of the boundary cell is enhanced, the advantage of the boundary sample is strengthened, the continuous learning ability of the algorithm is enhanced, new fault types are quickly recognized on the basis of existing learning results, and the generalization ability and accuracy of the fault diagnosis method are improved. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The fault diagnosis method technical route provided for the application;

[0033] Figure 2 The self-renewing continuous learning flowchart provided for the application;

[0034] Figure 3 The cell generation and cloning schematic diagram provided for the application;

[0035] Figure 4 The upward cloning strategy of offspring to parents provided for the application;

[0036] Figure 5 The cloning strategy of same-generation adjacent cloning provided for the application; DETAILED DESCRIPTION

[0037] The application will be further described below in combination with the accompanying Figures 1-5 Embodiments further illustrate the application.

[0038] The technical scheme adopted by the application includes: a self-renewing artificial immune system fault diagnosis model based on cell density, construction of omnipotent stem cells and super memory cells with recognition advantages, affinity and cell feature weight calculation, learning and recognition of turbine pump faults. First, the immune mechanism is extracted, the data samples, detectors and classification indexes are respectively corresponding to the antigens, immune cells and affinity, and an artificial immune system model with a hypercubic body as a memory cell is established. Secondly, the double-cloned cell strategy is used to expand the omnipotent stem cells and super memory cells, effectively improving the model recognition ability. Then, the cell weight is considered to strengthen the antibody advantage, assist antigen recognition, and further improve the continuous learning effect. Finally, the model is applied to the diagnosis of turbine pump faults, and good diagnosis results are obtained. Specifically, the following steps are included:

[0039] Step 1, construction of an artificial immune system model with a hypercubic body as a memory cell

[0040] Inspired by the idea of automatic regulation of omnipotent stem cells of biological immune system, the present application takes data sample, detector and classification index as antigen, immune cell and affinity respectively, wherein the shape of the cell is a hyper-rectangular body.

[0041] In order to describe the model more clearly, the following concepts are defined:

[0042] 1. Antigen (F i ): Feature carrier of sample. All training and test samples come from antigen. The i-th antigen is denoted as:

[0043] F i = [f i1 f i2 …f in ](1)

[0044] 2. Cell (G i ): Hyper-rectangular region activated by antigen. Cell encodes the position of the cell and the type of antigen that activates the cell, denoted as:

[0045] G i = [x i1 x i2 …x in g l](2)

[0046] Where g represents the cell algebra, and l is the label of the antigen that activates this cell.

[0047] 3. Cell algebra (g): A variable that determines the size of the cell. The higher the algebra, the smaller the cell size.

[0048] 4. Memory cell density (ρ). The number of certain memory cells in a unit space, the cell density is related to the cell algebra g and the type l as follows:

[0049]

[0050]

[0051] 5. Affinity (aff): Used to evaluate the affinity between antigen and memory cell, the greater the affinity, the closer the type, and the affinity of the antigen that cannot be determined temporarily can be calculated.

[0052] 6. Memory cell weight (W): Used to measure the influence of memory cells on the category to which the cells around them belong, and to assist in identifying antigens. The greater the weight, the closer the type.

[0053]

[0054] ​The correspondence between the main processes and immune mechanisms of the artificial immune system model using hyperrectangular bodies as memory cells is shown in Table 1. In this paper, all raw data (turbo pump operation vibration data) were normalized to [0.0001, 0.9999]n (where n represents the dimension) to ensure that the cells were in the correct state space.

[0055] Table 1. Correspondence between SGCM and immune mechanisms

[0056]

[0057] Step 2: Construction of pluripotent stem cells and super memory cells

[0058] Vibration data from the turbopump operation were randomly divided into two groups. One group was labeled and used as antibodies for cell training, while the other group was unlabeled and used as antigens for testing. Here, the vibration data is labeled with its corresponding fault type.

[0059] The antibody is input into the artificial immune system model and initialized to initially generate memory cells, as follows: The state space is defined as a generation 0 giant cell. After several generations of division, a series of daughter cells of constant size are obtained. The position and size of the daughter cells are recorded by the cell generation (g) and the coordinates of the center of the hyperrectangular cell. Since the cells in the space have certain patterns, each cell in the space can be encoded.

[0060] Assume the antigen is encoded as F i =[f i1 f i2 …f in The splitting algebra is g, F i The coordinates of the activated sub-cell are C. i =[x i1 x i2 …x in ], where f ik ∈[0.0001,0.9999],x ik ∈N, k=1,2,…n. Therefore, based on the fact that the antigen is located inside the cell, we know:

[0061]

[0062] Then antigen F can be obtained. i The center coordinates of activated cells satisfy the following inequality:

[0063] 2 g ·F i -1 <C i ≤2 g ·F i

[0064] To avoid missing the information of different antigens, g is usually set large enough. Memory cells are initially evolved by cell division, and after the division, memory cells containing only one antibody and empty cells are generated. For the initially generated memory cells, the density of each generation of cells is calculated. The density formula is as follows:

[0065]

[0066] The cells of the generation with the largest density of a class of memory cells are defined as super memory cells of the class, and the parent cells of the super memory cells are defined as totipotent stem cells. Through screening, totipotent stem cells in the memory cells are determined and labeled as the basis for self-renewing continuous learning weight calculation.

[0067] Step 3, double cloning cell strategy

[0068] The double cloning strategy includes offspring to parent cloning and offspring to adjacent same generation cloning. The parent cells and offspring cells have the same DNA, so the offspring cells have the same immune genetic structure as the parent cells, and the genetic information position has stability. The same generation adjacent memory cells that meet a certain affinity also have compatible DNA. Double cloning improves the density of memory cells, greatly improving the model recognition ability.

[0069] (1) Upward cloning strategy: reverse clone from the most numerous super memory cells to the parent totipotent stem cells, and keep the spatial position unchanged, as shown in Figure 2 The yellow cells represent super memory cells, the red cells represent the parent totipotent stem cells of the super memory cells, and the dots describe the position of the antibody in the super memory cells. The cross-shaped is the position of the newly generated cloned antibody in the totipotent stem cells after cloning the offspring cells. Through upward cloning, the information contained in the offspring is more accurately restored to the parent information

[0070] (2) Same generation adjacent cloning: adjacent memory cell cloning uses the adjacent empty cells of the super sphere search cell. Take the center of the adjacent empty cell as the center and the cell side length as the radius. If there is an antibody in the range, the empty cell evolves into a memory cell. As shown in Figure 3 The center coordinates of the empty cell adjacent to the memory cell are taken as the center and the cell side length as the radius to form a super sphere as a judgment area. Since the antibody t1 is located outside the circle, the antibody is not cloned. Antibody t2 is located within the cloning range, and the antibody is cloned to the corresponding empty cell. According to the relative position of the antibody in the parent cell, the antibody is cloned to the corresponding empty cell to form antibody t′2. Antibody t3 is located in two circles at the same time, cloned to two empty cells and formed two new antibodies t′3 and t′3′.

[0071] Step 4, continuous learning and recognition of turbine pump fault diagnosis

[0072] The antigen activates the memory cell or the empty cell. In the process of continuous learning and recognition, when the antigen activates the memory cell, the antigen is recognized as the same type as the memory cell; when the antigen activates the empty cell, the type of the antigen is determined by the adjacent memory cell of the empty cell. If the size of the empty cell is greater than that of the super memory cell, the empty cell is first divided into a cell with the same cell number as the super memory cell, and then evolved into a new memory cell. The affinity between the antigen and the memory cell is calculated by formula (5), and when the affinity cannot determine the type of the antigen, the adaptive cell weight is used to evaluate the type of the antigen. The continuous learning process is shown in Figure 4 , and the detailed process is described as follows:

[0073] (1) Directly recognizing the type of the antigen, the antigen activates the memory cell, and the activated memory cell increases an antibody, that is, the weight increases. When the antigen activates the memory cell, the memory cell is of what type, and the antigen is recognized as the same type. After the recognition is completed, the cell set is updated.

[0074] (2) The type of the antigen is determined by the adjacent memory cell, and the cell set is updated after each recognition of the type of the antigen. When the antigen activates the empty cell, the type of the antigen cannot be directly determined, and the type of the antigen is determined by the adjacent memory cell. The correlation between the antigen and the memory cell is used to judge the generation of a new memory cell, and the type of the antigen is determined by the new memory cell.

[0075] In order to prevent information omission, the affinity formula designed by the application includes all parameters related to the affinity between the antigen and the cell, including the number of antibodies, the cell number and the Euclidean distance, which can improve the accuracy of antigen recognition.

[0076]

[0077] In formula (5), q i is the number of antibodies in the memory cell i, g i is the number of the memory cell i, and d is the Euclidean distance between the antigen and the memory cell i. The cell corresponding to the maximum affinity is found, the type of the new memory cell is the same as that of the cell, and thus the type of the antigen can be determined.

[0078] When the affinity cannot determine the type of the antigen, such as different types of adjacent memory cells corresponding to the same affinity value, the weight is introduced to assist in recognition, and the cell weight formula is as follows:

[0079]

[0080] In formula (6), e q is the distance between the qth original antibody in the memory cell and the center of the empty cell, q∈[1,2,…,u]; k∈[1,2,…,u] represents the kth cloned antibody in the memory cell; Ak is the antibody coordinate of the clone generated in the memory cell; A k′ is the original antibody coordinate of the clone. The most relevant memory cell is determined by the comprehensive evaluation method of affinity and cell weight:

[0081] delta = aff x 70% + W x 30% (6)

[0082] It should be noted that if the antigen-activated null cell generation is lower than the super memory cell, the null cell should be divided to the same generation as the super memory cell. If the generation of the null cell is higher than or equal to the generation of the super memory cell, the antigen will generate a new memory cell with the same generation as the null cell, and the type is determined by the adjacent memory cell voting principle. The voting principle is to determine the type of the new memory cell according to the size of the affinity aff between the antigen and the adjacent memory cell, and if the aff cannot be determined, the antigen type is comprehensively evaluated in combination with the comprehensive evaluation index delta of the adjacent memory cell.

[0083] Through the above continuous learning and identification of the turbine pump sample, the fault type of the turbine pump is diagnosed.

[0084] The application is based on a continuous learning artificial immune system fault diagnosis method, which draws lessons from the automatic adjustment and continuous learning mechanism of biological totipotent stem cells, and proposes a self-renewing artificial immune system turbine pump fault diagnosis model; by using a double clone cell strategy, totipotent stem cells and super memory cells with strong recognition ability are generated; at the same time, the weight of the boundary cell is enhanced, the advantage of the boundary sample is strengthened, and the generalization ability and accuracy of the fault diagnosis method are improved. Finally, the affinity and weight combination method is used to evaluate the antigen type, further improve the continuous learning ability, and finally show good robustness and high accuracy in the application of turbine pump fault diagnosis.

[0085] Although the application has been disclosed as above with the preferred embodiments, it is not intended to limit the application, and any person skilled in the art can make possible changes and modifications to the technical solutions of the application by using the disclosed methods and technical contents without departing from the spirit and scope of the application. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the application, which does not deviate from the technical solutions of the application, belongs to the protection scope of the technical solutions of the application.

[0086] The contents not described in detail in the specification of the application belong to the known technology of those skilled in the art.

Claims

1. A self-updating continuous learning artificial immune system turbo pump fault diagnosis method, characterized in that, The application relates to a method for identifying the type of antigen in a turbine pump, comprising the following steps: Corresponding data samples, detectors and classification indicators are respectively corresponding to antigens, immune cells and affinity, and a model of an artificial immune system with a hyper-rectangular body as a memory cell is established; Vibration data of a turbine pump is randomly divided into two groups, one of which is marked as an antibody for cell training, and the other is unmarked as an antigen for testing; the marking contains vibration data and corresponding fault types; The antibody is input into the artificial immune system model and initialized to initially generate memory cells; Memory cells with the largest density of a certain type are defined as super memory cells of the type, and parent cells of the super memory cells are defined as totipotent stem cells; the totipotent stem cells and the super memory cells are expanded by a double cloning cell strategy to obtain a trained artificial immune system model; The antigen is input into the trained artificial immune system model, and the antigen is used to activate memory cells or empty cells in the artificial immune system model to realize fault type identification and continuous learning of the model; The formula for calculating the affinity is: where q i is the number of antibodies in memory cell i, g i is the generation of memory cell i, and d is the Euclidean distance between the antigen and memory cell i. The memory cell corresponding to the maximum affinity is found, the type of a new memory cell is the same as that of the memory cell, and the type of the antigen is judged; When the affinity cannot judge the type of the antigen, a weight is introduced to assist in identification, that is, the type of the antigen is further comprehensively evaluated by combining the affinity and the cell weight with the weight setting; The formula for calculating the cell weight is: e q is the distance between the qth original antibody in the memory cell and the empty cell center, q ∈ [1, 2, …, u]; k ∈ [1, 2, …, v] represents the kth clonal antibody in the memory cell; A k is the coordinate of the clonally generated antibody in the memory cell; A k′ is the original antibody coordinate of the clone; g i is the order of memory cell i, and d is the Euclidean distance between the antigen and memory cell i.

2. The method of claim 1, wherein: The activation of the antigen on the memory cells or the empty cells in the artificial immune system model to realize the fault type identification and the continuous learning of the model comprises the following steps: When the activated object of the antigen is a memory cell, the antigen is identified as the same type as the memory cell; When the activated object of the antigen is an empty cell, the type of the antigen is determined by adjacent memory cells of the empty cell; the affinity between the antigen and the adjacent memory cells is used to evolve the empty cell into a new memory cell, and the type of the antigen is determined by the new memory cell.

3. The method of claim 1, wherein: The comprehensive evaluation of the type of the antigen by further combining the affinity and the cell weight with the weight setting comprises the following steps: According to the formula delta = aff * 70% + W * 30%, the comprehensive evaluation index delta of each adjacent memory cell of the empty cell is calculated, the memory cell corresponding to the maximum delta is found, the empty cell is evolved into the memory cell, and the type of the antigen is determined; wherein aff is the affinity between the antigen and the adjacent memory cell.

4. The method of claim 1, wherein: The affinity cannot judge the type of the antigen, which includes the same affinity value corresponding to different types of adjacent memory cells.

5. The method of claim 2, wherein: When the activated object of the antigen is an empty cell, if the generation number of the empty cell is lower than that of the super memory cell, the empty cell is split to the same generation number as that of the super memory cell; if the generation number of the empty cell is higher than or equal to that of the super memory cell, a new memory cell with the same generation number as that of the empty cell is directly generated, and the type of the new memory cell is determined by a voting principle of adjacent memory cells.

6. The method of claim 5, wherein: The voting principle is used to determine the type of the new memory cell according to the size of the affinity aff between the antigen and the adjacent memory cells, and the type of the antigen is further comprehensively evaluated by combining the comprehensive evaluation index delta of the adjacent memory cells.

7. The method of claim 1, wherein: The double cloning cell strategy is an upward cloning strategy and a same-generation adjacent cloning strategy.

Images