System fault diagnosis method

Abstract

The invention relates to a system fault diagnosis method. The method comprises the following steps: constructing an anomaly diagnosis model based on principal component analysis; based on the abnormality diagnosis model, calculating a statistic control limit by using historical sample data in a normal operation state of the system; monitoring real-time sample data of system operation by using the abnormality diagnosis model and identifying an abnormal state of the system; obtaining abnormal sample data corresponding to the abnormal state of the system, and updating the statistic control limit based on the abnormal sample data; and according to the updated statistic control limit, re-determining whether the system is in an abnormal state by using the abnormality diagnosis model, and if the system is in the abnormal state, analyzing the abnormal sample data by using a fault identification model to obtain a fault classification result. The method can quickly and accurately judge the health state of the system and perform fault classification.

Description

Technical Field

[0001] The present invention relates to the technical field of system fault identification, and in particular to a system fault diagnosis method. Background Art

[0002] The hydraulic system plays an important role in the power transmission of the aircraft. The health of the hydraulic system directly affects the normal operation of the aircraft's related systems and components, and plays a very important role in the safety of the aircraft.

[0003] At present, the fault diagnosis technology for hydraulic systems is based on fault classification to determine the typical fault category. However, in actual operation, due to changes in the external environment, special interference, etc., the monitored parameters change abnormally, and there are abnormal values ​​in non-abnormal states. With the real-time accumulation of a large number of health status parameters, the constructed abnormal detection threshold will also be affected, thereby interfering with abnormal detection and increasing the system false alarm rate. In addition, as the aircraft serves for a long time, the overall performance of the hydraulic system will degrade. According to long-term statistics, the working parameters of the hydraulic system under different working conditions have an obvious degradation trend, resulting in abnormal monitoring data, which in turn affects the system fault monitoring results. At the same time, due to the high complexity of the aircraft power system, the large number of fault detection-related parameters, and the complex causes of faults, it is difficult to quickly and accurately locate the typical fault type in the event of an abnormality. Therefore, a fault diagnosis method is needed that can overcome the above problems, accurately determine the system health status and classify faults. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a system fault diagnosis method which can quickly and accurately determine the health status of the system.

[0005] The technical solution adopted by the present invention to solve the technical problem is: to provide a system fault diagnosis method, comprising the following steps:

[0006] Construct an abnormality diagnosis model based on principal component analysis;

[0007] Based on the abnormal diagnosis model, the statistical control limits are calculated using the historical sample data under the normal operating state of the system;

[0008] Using the abnormal diagnosis model to monitor the real-time sample data of the system operation and identify the abnormal state of the system;

[0009] Acquire abnormal sample data corresponding to the abnormal state of the system, and update the statistical control limit based on the abnormal sample data;

[0010] According to the updated statistical control limit, the abnormal diagnosis model is used to re-determine whether the system is in an abnormal state. If it is in an abnormal state, the fault identification model is used to analyze the abnormal sample data to obtain a fault classification result.

[0011] Furthermore, before the step of using the abnormal diagnosis model to monitor the real-time sample data of the system operation and identify the abnormal state of the system, it also includes the step of determining the system degradation stage based on the real-time sample data, and then performing abnormal data processing on the real-time sample data based on the different system degradation stages.

[0012] Further, judging the system degradation stage according to the real-time sample data includes:

[0013] Obtaining the deformation rate of the monitoring parameter curve corresponding to the real-time sample data, and then calculating the average deformation rate and tangent angle of the monitoring parameter curve at the current moment;

[0014] The current degradation stage of the system is judged according to the average deformation rate and the size of the tangent angle at the current moment.

[0015] Furthermore, before the step of obtaining the deformation rate of the monitoring parameter curve corresponding to the real-time sample data, the step of segmenting and smoothing the real-time sample data is also included.

[0016] Furthermore, judging the system degradation stage according to the average deformation rate and the tangent angle at the current moment includes:

[0017] Obtaining a first rate threshold of the system in a stable operation stage, a second rate threshold of the system in a uniform deformation stage, and a third rate threshold of the system in an accelerated deformation stage;

[0018] If the average deformation rate at the current moment is less than the first rate threshold, and the tangent angle is less than the set tangent angle threshold, it is considered that the system is in the stable operation stage;

[0019] If the average deformation rate at the current moment is greater than or equal to the first rate threshold and less than the second rate threshold, and the tangent angle belongs to the set tangent angle threshold range, it is considered that the system is in the uniform deformation stage;

[0020] If the average deformation rate at the current moment is greater than or equal to the second rate threshold, and the tangent angle is greater than the tangent angle threshold, it is considered that the system is in the accelerated deformation stage.

[0021] Furthermore, the performing abnormal data processing on the real-time sample data based on different system degradation stages includes:

[0022] Obtaining outlier thresholds for different degradation stages of the system;

[0023] It is determined whether the current real-time sample data is greater than the abnormal value threshold corresponding to the system degradation stage. If so, the current real-time sample data is updated to the average value of the real-time sample data within a set time.

[0024] Furthermore, obtaining the outlier thresholds of different system degradation stages includes:

[0025] Acquiring corresponding degradation sample data for different degradation stages of the system;

[0026] Calculating parameter standard deviations of the degradation sample data in different degradation stages of the system;

[0027] Based on the normal distribution principle, the abnormal value threshold corresponding to the system degradation stage is set according to the parameter standard deviation.

[0028] Furthermore, the abnormality diagnosis model is constructed based on the SPE statistic and the T2 statistic.

[0029] Furthermore, the statistical control limit is calculated by the following method:

[0030] Get the input data matrix;

[0031] Calculating a covariance matrix of the input data matrix;

[0032] The number of principal components is determined based on the eigenvalues ​​and eigenvectors of the covariance matrix, and then the SPE control limits and T2 control limits are calculated and obtained as the statistical control limits.

[0033] Furthermore, the SPE control limit is calculated by the following formula:

[0034]

[0035] Among them, SPE a is the SPE control limit, λ j is the eigenvalue of the covariance matrix of X, c α is a statistic with a normal distribution confidence level of α, θ i and h0 are intermediate variables.

[0036] Furthermore, the T 2 The control limits are calculated using the following formula

[0037]

[0038] in, is the T2 control limit, n is the number of samples, k is the number of principal components, F k,n-1,α is the critical value of the F distribution corresponding to the test level α, degrees of freedom k, and (nk) sample points.

[0039] Furthermore, the method also includes the step of obtaining the cause of the fault by using the knowledge graph search based on the fault classification result.

[0040] Beneficial Effects

[0041] Due to the adoption of the above-mentioned technical scheme, the present invention has the following advantages and positive effects compared with the prior art: the present invention constructs an abnormal diagnosis model based on principal component analysis, and uses historical data under normal system operation to calculate statistical control limits, and then uses the abnormal diagnosis model to monitor and identify the system status in real time. When an abnormal state is identified, the statistical control limits are updated, and then the updated statistical control limits are used to re-diagnose the abnormal state. If it is an abnormal state, the corresponding abnormal sample data is input into the fault identification model for classification, and finally the fault knowledge graph is used to find and infer the cause of the fault. It can sense and identify the system health status in real time, and then quickly and accurately determine the cause of the fault. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of a flow chart of an embodiment of the present invention;

[0043] Figure 2 It is a GRNN neural network structure diagram of an embodiment of the present invention;

[0044] Figure 3 It is a schematic diagram of a hydraulic system pressure supply oil circuit according to an embodiment of the present invention;

[0045] Figure 4 is a hydraulic system fault tree according to an embodiment of the present invention;

[0046] Figure 5 is a schematic diagram of a fault tree event list according to an embodiment of the present invention;

[0047] Figure 6 is a schematic diagram of deformation stages of monitoring parameters of a hydraulic system according to an embodiment of the present invention;

[0048] Figure 7 This is an embodiment of the present invention that constructs T based on historical sample size. 2 Schematic diagram of anomaly detection results;

[0049] Figure 8 It is a schematic diagram of anomaly detection results of constructing SPE based on historical sample size according to an embodiment of the present invention;

[0050] Fig. 9 This is an embodiment of the present invention, based on the newly added sample size to construct T 2 Schematic diagram of threshold validity verification;

[0051] Fig.10It is a schematic diagram of threshold validity verification based on the newly added sample amount constructed SPE according to an embodiment of the present invention;

[0052] Fig.11 This is an embodiment of the present invention, based on the newly added sample size to construct T 2 Schematic diagram of threshold update;

[0053] Fig.12 It is a schematic diagram of threshold value update based on newly added sample amount constructed SPE according to an embodiment of the present invention;

[0054] Fig.13 is an embodiment of the present invention based on the updated T 2 Schematic diagram of anomaly detection results using thresholds;

[0055] Fig.14 is a schematic diagram of anomaly detection results based on an updated SPE threshold according to an embodiment of the present invention;

[0056] Fig.15 is a typical fault diagnosis result diagram of an embodiment of the present invention;

[0057] Fig.16 is a typical fault diagnosis and identification result diagram of abnormal data according to an embodiment of the present invention;

[0058] Fig.17 It is a schematic diagram of a fault knowledge graph according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0060] The embodiment of the present invention relates to a fault diagnosis method for a hydraulic system based on dynamic threshold updating, such as Figure 1 As shown, including:

[0061] (i) Abnormal state judgment based on real-time abnormal value processing of monitoring parameters;

[0062] (ii) Typical fault diagnosis;

[0063] (3) Check the cause of the failure.

[0064] (I) Abnormal state judgment based on real-time abnormal value processing of monitoring parameters

[0065] Due to changes in the external environment, special interference, etc., the monitored parameters change abnormally, and there are abnormal values ​​in non-abnormal states. With the real-time accumulation of a large number of health status parameters, the constructed abnormal detection threshold will also be affected, thereby interfering with abnormal detection and causing an increase in the system false alarm rate. Therefore, it is necessary to perform real-time processing on abnormal values ​​that are difficult to capture in order to distinguish whether the data is an abnormal value or whether the state abnormality detected by the hydraulic system is actually occurring.

[0066] 1. Outlier processing

[0067] As the aircraft serves for a long time, the overall performance of the hydraulic system will degrade. According to long-term statistics, the working parameters of the hydraulic system under different working conditions have an obvious degradation trend. The hydraulic system monitoring working parameters such as hydraulic system pressure, hydraulic system accumulator pressure, pump housing return oil temperature, oil pump housing return oil filter pressure difference, high-pressure oil filter pressure difference are changing data that change with time and there is a correlation between the parameter values ​​before and after. Based on this feature of the data, abnormal value judgment and processing are carried out on the real-time hydraulic system parameter curve fluctuation, mutation amplitude, and time series relationship.

[0068] In the degradation stage of the hydraulic system, the deformation degree of different parameter curves is different. By monitoring the deformation rate of the parameter curve and the tangent angle threshold (accidental abnormality threshold), the abnormal judgment criteria are established, and the existence of abnormal values ​​is further judged based on the data statistics 3 sigma and processed.

[0069] Real-time anomaly detection performs real-time processing before abnormal status monitoring to find data in the monitoring data that is obviously inconsistent with normal data changes.

[0070] First: Due to the long operation cycle and large amount of data of the time series monitoring parameters of the hydraulic system, the monitoring parameters are segmented and processed. The data length is segmented into multiple short time series of the same length, each segment of parameter data is m, and after segmentation, it is T = (T1, T2, ..., T l ) has a total of l segments and a step length of r. Experiments have shown that the larger the sliding window, the greater the impact of abnormal points, which leads to an increase in the false alarm rate of abnormal detection. The parameter sampling frequency obtained by the hydraulic system is 1. After verification and comparison, the step length is set to 1 and the sliding window length is 6.

[0071] Secondly: for the monitoring parameter curve of hydraulic system performance degradation, calculate the parameter curve deformation rate V i and the tangent angle α i .

[0072] Deformation rate V i It is calculated by dividing the increment of two adjacent monitoring parameters by the time. The tangent angle is

[0073]

[0074] Among them, V i is the deformation rate; α i is the tangent angle corresponding to time i; the average deformation rate Where m is the window length (i.e. the number of monitoring times within the window).

[0075]

[0076] Finally: according to the historical data of the hydraulic system at different stages, set the rate threshold and tangent angle threshold, and then judge the degradation stage of the system by comparing the average deformation rate and tangent angle at the current moment with the threshold. For different degradation stages, the data should satisfy the normal distribution, so the 3σ statistical method can be used to collect and process the relevant historical data to obtain the abnormal value judgment threshold, and then the abnormal values ​​found by monitoring are corrected in real time to the mean value of the set time period. After a large amount of statistics on the historical data related to the monitoring parameters, the abnormal value thresholds established at different stages are obtained as follows:

[0077] For the stable stage, the 3σ principle is used to judge the abnormal value and the abnormal threshold X is set. TH1 satisfy is the mean value calculated for the monitoring parameter, and σ is the standard deviation of the monitoring parameter;

[0078] In the uniform deformation stage, the data showed a uniform increase or decrease trend. The monitoring data was segmented, and the 3σ principle was used to judge the abnormal value according to the sliding window length of 6 monitoring parameters. The abnormal threshold X was set. TH2 satisfy

[0079] In the accelerated deformation stage, the data showed large fluctuations. The monitoring data was segmented and the abnormal threshold X was set according to the sliding window length of 6 monitoring parameters and 4σ. TH3 satisfy The outlier is judged based on the principle.

[0080] Replace outliers with the stage mean.

[0081] 2. Abnormal status judgment

[0082] Step 1: For the hydraulic system monitoring variables, the accumulated variable samples are recorded as training samples. Assume that the sample collected once under normal operating conditions is x = [x1 x2…x n ], then the sample data set consisting of m samples is X∈R m×n , where n is the number of sensor measurement variables and m is the number of samples.

[0083] Standardization processing, for key equipment, different variables have different dimensions, which can easily cause false variation information. In order to eliminate the impact of dimensions, it is necessary to standardize the data set X. is the data matrix after standardization, then:

[0084]

[0085] in,

[0086] I m ——an m-dimensional column vector whose elements are all 1;

[0087] u=[u1,u2,…,u n ] T ——sample mean vector;

[0088] ——Sample variance matrix.

[0089] For convenience, the normalized matrix Still recorded as X. The process of establishing the principal component analysis model is essentially the process of eigenvalue decomposition of the covariance matrix of the data matrix X. m×n The covariance matrix S can be expressed as:

[0090]

[0091] Step 2: Perform eigenvalue decomposition on S and find its eigenvalue λ1≥λ2≥…≥λ n and the corresponding eigenvectors p1, p2, …, p n .

[0092] Step 3: Determine the number of pivots

[0093] Sp i =λ i p i i=1,2…n

[0094] in,

[0095] λ i ——is the i-th eigenvalue of S;

[0096] p i ——is the i-th eigenvector of S;

[0097] The data matrix can be decomposed into the following:

[0098]

[0099] in,

[0100] T=[t1 t2…t k]∈R m×k ——pivot score matrix;

[0101] ——Residual score matrix;

[0102] E——Residual term, representing the difference between X and non-pivot element p k+1 ,…,p n Changes in

[0103] P=[p1,p2,…,p k ]∈R n×k ——Principal element load matrix;

[0104] ——residual loading matrix;

[0105] K is the number of principal components, and k is less than n.

[0106] The problem of selecting the number of principal components is solved by introducing the concept of principal component cumulative variance contribution rate. The definition of principal component cumulative variance contribution rate refers to the proportion of the cumulative variance of the first k principal components to the total variance of the sample. It is calculated by the ratio of the sum of the first k eigenvalues ​​of the covariance matrix corresponding to the sample data matrix X to the sum of all eigenvalues ​​of the covariance matrix corresponding to the sample data matrix X. When the calculated ratio reaches a certain value (generally 85%), it means that the number of principal components obtained has already contained most of the original data information. In order to detect the fault information contained in the data matrix X, it is necessary to define multivariate statistics that can characterize the overall characteristics of the data.

[0107] Step 4: Determine T 2 Statistical control limits and SPE statistical control limits:

[0108] The most commonly used multivariate statistics are the SPE statistic (squared prediction error) and T 2 Statistics, the former describes the actual measurement value X of the process for the principal component model X P The degree of deviation, which reflects the change of equipment status or process operation. When the inspection level is α, the SPE control limit is calculated as follows:

[0109]

[0110] Where: λ j is the eigenvalue of the covariance matrix of X; c α The normal distribution has a reliability of α.

[0111] When the test level is α, T 2 The control limits are calculated as follows:

[0112]

[0113] Where: n is the number of samples; k is the number of principal components retained; F k,n-1,α It is the critical value of the F distribution corresponding to the test level α and the degree of freedom k, (nk) sample points.

[0114] Step 5: Test the module

[0115] The use of principal component models and multivariate statistics to monitor the operation of processes or equipment requires the calculation of SPE control limits and T from normal state data. 2 Control limits are used to identify whether the process is abnormal.

[0116]

[0117] T 2 =XP k Λ -1 P k T X T

[0118] Step 6: Dynamic Threshold Update

[0119] The anomaly detection model of principal component analysis is a data model under the normal operation of the system. It describes the relationship between the variables under the normal operation of the system. The relationship between these variables is an external reflection of the system's motion law. 2 Control limits are used for anomaly detection.

[0120] As the hydraulic system operates, the sample size gradually increases, and the anomaly threshold constructed based on historical samples will also change. At the same time, due to internal and external factors of the hydraulic system operating environment, the sample data added by real-time monitoring may experience abnormal fluctuations. The emergence of severe outliers will lead to inaccurate construction of the anomaly detection threshold, and thus increase the false alarm rate. Therefore, it is necessary to combine the incremental sample data after outlier processing to update the anomaly diagnosis threshold in real time.

[0121] The incremental sample X' is standardized and verified, and then the threshold detection is performed in step 5. When the output results SPE' and T 2 ' are all below the control limit in step 4, there is no need to update the threshold, otherwise repeat steps 1 to 4 to update the threshold.

[0122] Step 7: Dynamically update the threshold value based on the incremental samples, that is, complete the threshold value constructed with a large number of normal samples. Use the final constructed threshold value to perform abnormal diagnosis of the hydraulic system, and the diagnosis process is step 5. When any T 2If the statistical control limit and SPE statistic exceed the limit, the hydraulic system will be in an abnormal state.

[0123] Then, the samples of abnormal state are further diagnosed based on the GRNN network. The process is as follows:

[0124] 1) Sample data is represented by X, and the input of GRNN neural network is X=[x1,x2,…,x n ] T , the network output is Y = [y1,y2,…,y k ] T The GRNN neural network structure is as follows: Figure 2 Shown

[0125] (1) Input layer

[0126] The training sample data first needs to be sent to the input layer of the network. The number of neurons in the input layer is the dimension of the training sample. The training sample is directly transmitted to the pattern layer through the input layer.

[0127] (2) Model layer

[0128] The number of neurons in the pattern layer is the dimension n of the training sample, and each neuron corresponds to different training sample input data. The transfer function P of the pattern layer is i for:

[0129]

[0130] In the above formula (4.1), X is the input variable of the GRNN network, X i is the training sample data corresponding to the i-th neuron. σ is the smoothing factor.

[0131] (3) Sum layer

[0132] There are two different types of neurons that need to be summed. One is arithmetic summation and the other is weighted summation. The following is a detailed introduction to the two types of summation.

[0133] The arithmetic sum is performed on all neurons in the pattern layer. The connection weight between the pattern layer and each neuron is 1. The calculation formula used here is:

[0134]

[0135] Then the transfer function of the summation layer after arithmetic summation is:

[0136]

[0137] Another type of weighted summation is also performed on all neurons in the pattern layer, but the connection weights are different. The connection weight between the i-th neuron in the pattern layer and the j-th summing neuron is the i-th output result Y i The weighted sum is calculated as:

[0138]

[0139] Then the transfer function of the summation layer after weighted summation is:

[0140]

[0141] (4) Output layer

[0142] The output layer is the last layer of the GRNN network. The number of neurons in the output layer is the dimension k of the output vector, and the output is the result of the two types of summation in the summation layer divided by each other:

[0143]

[0144] Through nonlinear regression analysis, the output of the GRNN network can be finally obtained as:

[0145]

[0146] Here, the smoothing factor σ affects the effect of the network. When the smoothing factor σ is large, the output of the network will be very close to the mean of the sample. In contrast, when the smoothing factor σ is small and close to 0, the output will be very similar to the trained sample. Therefore, it is important to take a moderate value for the smoothing factor σ. The expansion factor SPREAD in MATLAB is equivalent to the smoothing factor σ, which is generally 1 by default. The establishment of the GRNN fault classification model is realized by training sample data through networks at different levels.

[0147] The trained network is tested for accuracy using test data, and the test sample data is entered into the created GRNN model to classify typical faults and implement typical fault diagnosis of the hydraulic system.

[0148] Finally, the hydraulic system fault tree formed based on the design characteristics of the hydraulic system, namely the structural composition principle, is established based on the neo4j knowledge graph.

[0149] Based on the design characteristics of the hydraulic system, a fault tree with hydraulic system failure as the top-level event is analyzed to establish a knowledge graph visualization model. The cause of the fault can be found according to the fault diagnosis results, guiding the troubleshooting and maintenance of the hydraulic system.

[0150] The knowledge graph visualization constructs <entity, relationship, entity> with the hydraulic system fault tree as the object, and specifically uses the graph database Neo4j to visualize the association relationship between fault events in the knowledge graph. The query and establishment of fault events based on the knowledge graph use the cypher language that comes with Neo4j.

[0151] The present embodiment will be further described below by taking an aircraft hydraulic system as an example.

[0152] 1) Hydraulic system composition and structural design principles

[0153] The aircraft hydraulic system consists of two independent left and right hydraulic systems, pumps and oil tanks. The left system is used as an example to explain the system principle. The left system pressure supply oil circuit consists of low-pressure pipelines and high-pressure pipelines. The pressure supply oil circuit mainly consists of hydraulic oil tanks, separation joints, gear pumps, check valves, hydraulic oil filters, automatic unloading valves, pressure conversion valves, brake pressure accumulator bottles, remote pressure gauge sensors, and on-board system tests and boost valves. See the attached diagram for the left hydraulic system pressure supply oil circuit principle. Figure 3 shown.

[0154] 2) Hydraulic system fault tree

[0155] The working principle, structural characteristics and functions of each component of the hydraulic system in the system. Assumptions:

[0156] ① The components exist in two states: normal and faulty;

[0157] ② Component failures are independent of each other;

[0158] ③ No external interference;

[0159] ④Failures caused by no design defects.

[0160] Fault tree Figure 4 As shown, the corresponding fault tree event table is as follows Figure 5 as shown.

[0161] 3) Real-time abnormal value processing and dynamic threshold update of hydraulic system monitoring parameters:

[0162] a) Determine abnormal values ​​of each monitoring parameter of the hydraulic system

[0163] The deformation rate and tangent angle of the monitoring parameter curve are solved for the monitoring parameters of the hydraulic system. Through a large amount of data recording and statistics, the deformation rate threshold and tangent angle threshold of each monitoring parameter curve are set. Specifically, the rate threshold and tangent angle threshold of each deformation stage of the monitoring parameter are obtained as shown in Table 1. Further, the abnormal value judgment of the parameter curve can be carried out.

[0164] Table 1 Abnormal value thresholds of monitoring parameters

[0165]

[0166]

[0167] like Figure 6 As shown, taking the hydraulic system monitoring parameter pump housing return oil temperature as an example, through a large number of hydraulic system pump housing return oil temperatures, the pump housing temperature curve is divided into three deformation stages.

[0168] After statistics based on the average rate and tangent angle of the solution, the corresponding outlier thresholds under different deformation degrees are: When α<45, the average deformation rate in this stage is calculated based on a large amount of data statistics to be 0~2℃ per small segment of data (multiple segments of data are obtained based on the sliding window and step size), so the deformation rate The corresponding outlier threshold is set to when α≈45 At this time, the parameter curve has a certain deformation growth or decrease trend and enters the uniform deformation stage. The outlier threshold is set to 13.2; when When α>45, the outlier threshold is set to 33 in the accelerated deformation stage.

[0169] By identifying the outlier threshold for each stage, the abnormal data is replaced with the stage mean to achieve abnormal value processing.

[0170] b) Constructing abnormal thresholds using historical samples

[0171] The threshold is constructed using the normal hydraulic system monitoring parameters. n is a parameter with a total of five variables, namely hydraulic system pressure, hydraulic system accumulator pressure, pump housing return oil temperature, oil pump housing return oil filter pressure difference, and high-pressure oil filter pressure difference. The number of samples is 20 for each healthy state. x = [x1 x2…x n ], 100 historical sample data, 20 test sample data when abnormality occurs, threshold value and diagnosis result of PCA model T 2 The statistical control limits and SPE statistical control limits are shown in the figure below. The threshold limit in the figure is the threshold limit obtained by training with historical normal sample data, and the detection value is the output value corresponding to the abnormal sample. The detection value of the abnormal sample is greater than the threshold limit. Figure 7 and 8 shown.

[0172] c) Added sample threshold validity test and threshold update

[0173] The threshold validity test is performed on the newly added samples. The newly added samples are not 101 to 112, and the thresholds of 101 to 111 are not exceeded. When the 112th sample exceeds the limit, the threshold is updated. Fig. 9 and 10 shown.

[0174] For T 2 Statistical control limit. As the number of samples of the healthy state of the hydraulic system increases gradually, the validity of the newly added samples is verified. The T corresponding to the 112th sample in the newly added samples 101 to 112 is 2 The statistical control limit changes, and the new sample size needs to be combined with all health history samples, that is, 1 to 112, to update the threshold; when the new sample size is 101 to 111, T 2 The control limits of the SPE statistics have not changed, so there is no need to update the thresholds. Taking the newly added sample size of 300 and 600 as an example, the updated thresholds are shown in the table below. Fig.11 and 12 shown.

[0175] In summary, as the sample size increases, any T 2 If the control limit of the SPE statistic changes, the PCA threshold needs to be updated.

[0176] d) Based on T 2 / SPE threshold for real-time abnormal state detection of hydraulic system

[0177] The threshold is updated through the continuously accumulated historical samples of health status to obtain the final anomaly detection threshold. This threshold is used to detect anomalies in the real-time operation data of the hydraulic system. The detection results are shown in the attached figure. Fig.13 and 14 As shown. Through the real-time abnormal health status detection of the hydraulic system, all of them exceed T in the healthy state. 2 If the control limit of the SPE statistic is greater than the control limit of the SPE statistic, it is judged that the hydraulic system has failed. The final threshold limit and the initial threshold limit are marked in the figure.

[0178] e) Typical fault identification of GRNN hydraulic system:

[0179] GRNN diagnostic model training, typical fault types are divided into four categories, A1 hydraulic system pressure is unstable with noise, A2 hydraulic system pressure exceeds the specified value, A3 left system unloading frequently, A4 hydraulic system pressure exceeds the specified value. Hydraulic system monitoring parameters n are five variables, namely hydraulic system pressure, hydraulic system accumulator pressure, pump housing return oil temperature, oil pump housing return oil filter pressure difference, high pressure oil filter pressure difference. The number of samples m is 20 for each health state. x = [x1 x2…x n ], then the sample data set consisting of m samples is X∈R m×n , where n is the number of sensor measurement variables and m is the number of samples.

[0180] First, 80 sets of training data are sent to the GRNN network for training, and then 80 sets of test data are sent to the GRNN network model trained with the training data to obtain the state recognition results. The state mode target outputs of {A1, A2, A3, A4} are (1, 0, 0, 0) respectively. T , (0,1,0,0) T , (0,0,1,0) T , (0,0,0,1) T The network output is Y = [y1, y2, ..., y k ] T , then k is 4.

[0181] According to the comparison between the network output results and the actual expected results, there are 80 test samples, of which 72 test samples are accurately identified. The research results show that the accuracy of the state recognition results of the GRNN neural network is 90%. The typical fault recognition results of the GRNN hydraulic system are as follows Fig.15 shown.

[0182] f) Identification of typical faults of the created GRNN hydraulic system:

[0183] The data samples judged as abnormal by the threshold abnormality of the hydraulic system are used as input for further typical fault identification. The diagnosis result is that it belongs to fault type A2, the left hydraulic system pressure exceeds the specified value. The accuracy rate is 90%, among which samples 1 and 20 are judged as state A3. The typical fault identification results of GRNN hydraulic system are as follows Fig.16 Shown

[0184] g) Hydraulic system related fault reasoning

[0185] Combined with the hydraulic system fault tree, the hydraulic system related fault events and related relationship reasoning data set are input as the data layer of the knowledge graph to mine richer semantic information in the data set, thereby constructing a hydraulic system fault reasoning knowledge graph such as Fig.17 shown.

[0186] According to the fault type diagnosed by GRNN, the knowledge graph fault reasoning is performed to find the cause of the fault, which may be caused by the spool of the x7 automatic unloading valve being stuck and blocking, causing the system to overload. The hydraulic system fault caused by the deformation or damage of the spring of the x8 automatic unloading valve is used as a guide for troubleshooting and repair.

Claims

1. A system fault diagnosis method, characterized in that: The following steps are involved: Construct an abnormality diagnosis model based on principal component analysis; Based on the abnormal diagnosis model, the statistical control limits are calculated using the historical sample data under the normal operating state of the system; Using the abnormal diagnosis model to monitor the real-time sample data of the system operation and identify the abnormal state of the system; Acquire abnormal sample data corresponding to the abnormal state of the system, and update the statistical control limit based on the abnormal sample data; According to the updated statistical control limit, the abnormal diagnosis model is used to re-determine whether the system is in an abnormal state. If it is in an abnormal state, the fault identification model is used to analyze the abnormal sample data to obtain a fault classification result.

2. The method according to claim 1, characterized in that Before the step of using the abnormal diagnosis model to monitor the real-time sample data of the system operation and identify the abnormal state of the system, it also includes the step of determining the system degradation stage based on the real-time sample data, and then performing abnormal data processing on the real-time sample data based on the different system degradation stages.

3. The method according to claim 2, characterized in that The determining the system degradation stage according to the real-time sample data includes: Obtaining the deformation rate of the monitoring parameter curve corresponding to the real-time sample data, and then calculating the average deformation rate and tangent angle of the monitoring parameter curve at the current moment; The current degradation stage of the system is judged according to the average deformation rate and the size of the tangent angle at the current moment.

4. The method according to claim 3, characterized in that Before the step of obtaining the deformation rate of the monitoring parameter curve corresponding to the real-time sample data, the method further includes the step of segmenting and smoothing the real-time sample data.

5. The method according to claim 3, characterized in that: The determining the system degradation stage according to the average deformation rate and the tangent angle at the current moment includes: Obtaining a first rate threshold of the system in a stable operation stage, a second rate threshold of the system in a uniform deformation stage, and a third rate threshold of the system in an accelerated deformation stage; If the average deformation rate at the current moment is less than the first rate threshold, and the tangent angle is less than the set tangent angle threshold, it is considered that the system is in the stable operation stage; If the average deformation rate at the current moment is greater than or equal to the first rate threshold and less than the second rate threshold, and the tangent angle belongs to the set tangent angle threshold range, it is considered that the system is in the uniform deformation stage; If the average deformation rate at the current moment is greater than or equal to the second rate threshold, and the tangent angle is greater than the tangent angle threshold, it is considered that the system is in the accelerated deformation stage.

6. The method according to claim 5, characterized in that The performing abnormal data processing on the real-time sample data based on different system degradation stages includes: Obtaining outlier thresholds for different degradation stages of the system; It is determined whether the current real-time sample data is greater than the abnormal value threshold corresponding to the system degradation stage. If so, the current real-time sample data is updated to the average value of the real-time sample data within a set time.

7. The method according to claim 6, characterized in that The obtaining of abnormal value thresholds at different system degradation stages includes: Acquiring corresponding degradation sample data for different degradation stages of the system; Calculating parameter standard deviations of the degradation sample data in different degradation stages of the system; Based on the normal distribution principle, the abnormal value threshold corresponding to the system degradation stage is set according to the parameter standard deviation.

8. The method according to claim 1, characterized in that The anomaly diagnosis model is based on the SPE statistic and T 2 Statistics are constructed.

9. The method according to claim 7, characterized in that: The statistical control limits are calculated by the following method: Get the input data matrix; Calculating a covariance matrix of the input data matrix; The number of principal components is determined based on the eigenvalues ​​and eigenvectors of the covariance matrix, and then the SPE control limits and T2 control limits are calculated and obtained as the statistical control limits.

10. The method according to claim 1, characterized in that It also includes the step of using the knowledge graph to retrieve the cause of the fault according to the fault classification result.

Images