CN109779791A - An intelligent diagnosis method for abnormal data in solid rocket motor - Google Patents

Abstract

The present invention proposes abnormal data intelligent diagnosing method in a kind of solid propellant rocket, abnormal point value is checked using t first, secondly according to BP-ML algorithm, the preparation method of weight in traditional artificial neural network is improved, is changed to that each parameter value { x of engine will be inputted_iIt is assumed that a Gaussian process, meets AR (P) model, acquires weighted value by Maximum Likelihood Estimation Method, verify anomaly parameter, and carry out parameter value reconstruct.The present invention solves the problems, such as method deficiency under the conditions of the prior art, reduces the error of artificial micro-judgment, establishes abnormal data method for diagnosing faults.

Description

An intelligent diagnosis method for abnormal data in solid rocket motor

technical field

The invention belongs to the technical field of solid rocket motor test measurement and control, and mainly relates to intelligent automatic diagnosis of abnormal data in solid rocket motor tests, reduces errors in judgment by human experience, and establishes an intelligent diagnosis method for abnormal data.

Background technique

The solid rocket motor ground test is a high-risk, high-investment, high-energy-consumption and irreversible test project. With the development of test technology, the process reliability of the test measurement and control system has been continuously improved, and the main parameters can be obtained without errors. However, during the test, the working conditions are relatively complex, and it is inevitably affected by factors such as engine design, tooling, sensor damage, high temperature ablation of cables, and falling measuring blocks, resulting in abnormal test data. The abnormal situation is divided into abnormal point value and abnormal parameter. Judgment methods for outlier values include t test, Laida criterion and Chauville criterion, etc. Among them, t test is a classic statistical judgment method, which belongs to a method in the hypothesis test of normal population mean and variance. , the premise is that the data as a whole conforms to a normal distribution, which can effectively and strictly detect outliers.

During the test, the solid rocket motor is easily affected by factors such as impact, temperature, and structural deformation. The parameter signals measured by the test have nonlinear and non-stationary characteristics. At present, when judging abnormal parameters in the domestic solid rocket motor field, the traditional data processing method is still used, which is relatively simple. When the test data is abnormal, it is impossible to judge whether the cause of the failure belongs to the failure of the measurement and control system or the failure of the engine itself. Experts are required to conduct experience interpretation. Therefore, it is necessary to apply the current more advanced data mining methods in this field.

During the test, the solid rocket motor is easily affected by factors such as impact, temperature, and structural deformation. The parameter signals measured in the test have nonlinear and non-stationary characteristics. For the engine in the finalization stage, the performance parameters are in a relatively stable state, and the test environment , statistical characteristics do not change with time. Therefore, in practical engineering applications, the solid motor ground test process is regarded as a stationary random process, and the collected parameter signals conform to the characteristics of stationary random signals. Logistic regression models, time series and other widely used machine learning, artificial intelligence, and statistical methods.

The relationship between the parameters of the solid rocket motor test is non-stationary and nonlinear, and the data obtained are different due to the different factors in the specific test site. It is difficult to obtain the specific functional relationship between the parameters, and the artificial neural network model With nonlinear mapping and generalization capabilities, a 3-layer BP neural network can approximate any nonlinear function (according to the Kolrnogorov theorem), as shown in Figure 1 below. In the analysis and processing of experimental data, it can be used to express the relationship between parameters. Artificial neural network consists of input layer, hidden layer and output layer. The learning process consists of two processes, the forward propagation of the signal and the back propagation of the error. By adjusting the weights many times, the error of the network output is reduced to an acceptable level, or the number of learning times set in advance is carried out. The invention designs a new weight estimation algorithm to learn a nonlinear relationship between the dependent variable and the independent variable. When there are some abnormal parameters, they can be detected according to this nonlinear relationship. At the same time, the abnormal parameters can also be reconstructed and restored.

SUMMARY OF THE INVENTION

The post-analysis and processing of the existing solid rocket motor test data does not involve intelligent diagnosis of abnormal data. Basically, the traditional data processing method is still used, which is relatively simple. When the test data is abnormal, it is impossible to determine the cause of the failure. System failure or engine failure requires expert interpretation, and there is a certain error problem. The invention proposes an intelligent diagnosis method for abnormal data in a solid rocket motor, which judges the abnormal point value in the engine test through the classical t test method; and designs a new method for testing abnormal parameters of the engine based on an artificial neural network model, which improves the The weight estimation algorithm is used, and the implicit fault rule is found from the historical data of the engine test, which can effectively detect the abnormal parameters, reduce the judgment error of human experience, and realize the intelligent diagnosis of abnormal data.

The technical scheme of the present invention is:

The method for intelligently diagnosing abnormal data in a solid rocket motor is characterized by comprising the following steps:

Step 1: Collect the test data of the solid rocket motor, and check whether the selected test data satisfies the normal distribution, if so, proceed to Step 2, otherwise select other test data for re-testing;

Step 2: Use the t test method to identify the abnormal points in the selected test data and eliminate them;

Step 3: Using the BP-ML algorithm and using the neural network to test the abnormal parameters of the test data processed by the t test in step 2, if there are abnormal parameters, reconstruct the parameter values of the abnormal parameters.

In a further preferred solution, the method for intelligently diagnosing abnormal data in a solid rocket motor is characterized in that: in step 1, it is judged whether the test data satisfies the normal distribution by calculating the skewness and kurtosis of the test data.

A further preferred solution, the method for intelligently diagnosing abnormal data in a solid rocket motor, is characterized in that: in step 3, the BP-ML algorithm adopts the maximum likelihood estimation method for the weight vector estimation method in the artificial neural network BP algorithm. achieved by obtaining.

A further preferred solution, the method for intelligently diagnosing abnormal data in a solid rocket motor, is characterized in that: the specific process of performing abnormal parameter inspection in step 3 is:

Step 3.1: Use the data without abnormal parameters as sample data to train the neural network to obtain a trained neural network;

Step 3.2: Input the actual test data into the neural network to repair the neural network. If the correlation between the repaired output value and the target value is poor, it is inferred that there are abnormal parameters in the input value of the neural network; if there are abnormal parameters in the input value of the neural network , then the parameter values in the input value of the neural network are judged separately to obtain specific abnormal parameters;

Step 3.3: Use the abnormal parameters obtained in step 3.2 as the output parameters of the neural network, remove the abnormal parameters from the input parameters of the neural network, and then reconstruct the neural network to obtain the repaired abnormal parameters.

beneficial effect

Compared with the prior art, the present invention has the following features: (1) conforming to the characteristics of solid rocket motor test data; (2) widely used; (3) constructing a complete abnormal data inspection method; (4) completing abnormal data reconstruction process.

Additional aspects and advantages of the present invention will be set forth, in part, from the following description, and in part will be apparent from the following description, or may be learned by practice of the invention.

Description of drawings

The above and/or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of embodiments taken in conjunction with the accompanying drawings, wherein:

Figure 1: Schematic diagram of a three-layer neural network;

Figure 2: Schematic diagram of target value and repair value;

Figure 3: Schematic diagram of the correlation between the target value and the repair value;

Figure 4: Schematic diagram of CCDD correlation coefficient;

Figure 5: Schematic diagram of the correlation between the target value and the repair value;

Figure 6: Schematic diagram of target value and repair value;

Figure 7: Schematic diagram of the correlation between the target value and the repair value;

Figure 8: Schematic diagram of CCDD correlation coefficient;

Figure 9: The weight w of the neural network;

Figure 10: Comparison between the repaired value of the abnormal parameter wP9 and the target value;

Figure 11: Abnormal parameter wP9.

Detailed ways

The abnormal data test method involved in the present invention uses t test, the premise of which is established on the theoretical basis that the measured value follows a normal distribution. Although it is theoretically understood that in the steady state stage of the solid rocket motor ground test, the characteristics of the measured performance parameters do not change with time and conform to a normal distribution, but during the test, it will be disturbed by different situations, and its distribution will be different. Therefore, before using the t test, the test data must be tested for normal distribution.

According to the formula of skewness and kurtosis, when the random variable {x _n } obeys the normal distribution, its skewness k ₁ =0, kurtosis k ₂ =3,

Calculated as follows:

In the t-test, when the variance is unknown, the test method for a single normal population mean is:

Assuming that the population follows a normal distribution N(μ,σ ² ), μ,σ ² are unknown parameters, (X ₁ ,...,X _N ) is a sample with a population size of n. To test the hypothesis: The statistic is usually constructed with the sample-corrected variance instead of the population variance.

in, When the hypothesis H ₀ holds, T obeys the t distribution with n-1 degrees of freedom. When the value of |T| is large, the hypothesis is unlikely to hold, and H ₀ should be rejected. Therefore, for a given 0<α<1, The critical value t _α/2 (n-1) of the test can be obtained from the t distribution table, so that

p{|T|≥t _α/2 (n-1)}=α (4)

So check

If |T|≥t _α/2 (n-1), reject the hypothesis H ₀ , that is, consider that the population mean is significantly different from μ ₀ ; if |T|＜t _α/2 (n-1), accept H ₀ , that is, the overall mean is considered to be not significantly different from μ ₀ . This test method that uses a statistic that obeys the t-distribution as a test statistic is called the t-test method.

The steady-state data of the pressure value p, one of the most important parameters measured in the solid rocket motor test, is selected as the test object. Select the data point value in 10.5s ~ 17.5s of a hair styling test, the sampling rate is 5000Sa/s, the time interval is 0.0002s, and the skewness k ₁ =0.2279 of the measured data is calculated according to formulas (1) and (2) respectively, The kurtosis k ₂ =2.7308, the normal distribution characteristics are good, and the t test can be used.

The neural network model has a strong generalization ability and can non-linearly map the relationship between various parameters. Therefore, this paper selects the thrust data that has a greater impact on the engine design as the output neuron node, and the selection of the input layer neuron node is related to the output layer. parameters of pressure, displacement, strain, and temperature. Suppose this system equation has an m-dimensional input vector and an n-dimensional output vector, {x _i } is the value of each input parameter point, {f _i } is the value of the output parameter point, x=(x ₁ ,x ₂ ,...,x _m ) ^T ,f=(f ₁ ,f ₂ ,...,f _n ) ^T , the network corresponds to m input nodes and n output nodes, _zi is the system identification function, also called activation function, you can choose linear function, slope function , s-shaped function, etc. The system equations are as follows:

Suppose the neural network has M layers, the number of nodes in the lth layer is n _l , and f _i ^l is the output of the lth layer node i, then

in, is the state of node i at layer l, Coefficient row vector, f ^l layer l-1 output column vector.

At this time, according to the existing artificial neural network BP algorithm, the weight coefficient in the system equation Using the tutored learning method, the given objective function is g(w) _min :

That is, after the input parameters {X, F} are given, the objective function is minimized

is the output value, According to the method in the optimization - the fastest gradient descent method, the weight vector can be obtained

In order to improve the accuracy and efficiency of the output value of the system equation, the present invention improves the artificial neural network BP algorithm, changes the estimation method of the weight vector to maximum likelihood estimation (ML), and the new algorithm is named as the BP-ML algorithm,

According to the test of formulas (1) and (2) in this paper, the input engine parameter values {x _i } conform to the normal distribution and satisfy the independent and identical distribution, because each parameter value {x _i } is a process that changes with time. It is assumed that {x _i } is a Gaussian process and satisfies the AR(P) model, that is, the formula (11) at this time is used to replace the formula (9) in the BP algorithm,

X _k+p+1 =a ₁ X _k+1 +a ₂ X _k+2 +...+a _p X _k+p +ε _k+p+1 . (11)

Among them, it is assumed that white noise ε _m ~ iidN(0,σ ² ), (X ₁ , X ₂ ,...,X _M ) satisfies a normal distribution with mean μ and covariance Σ. Let (x _p+1 , x _p+2 ,…x _p+m ) ^T be an input sample of the AR(P) sequence. but:

(x _p+1 ,x _p+2 ,…x _p+m )～N(μ,Σ)

where Σ is the covariance matrix:

Our goal is to estimate the value of the unknown parameters σ ² , a ₁ , a ₂ ,..., a _p based on the observed values x _p+1 , x _p+2 ,…x _p+m , denoted θ=(σ ² ,a ₁ ,a ₂ ,...,a _p ), let x ₁ ,x ₂ ,...,x _p be the unknown parameters. Equation (11) can be changed to:

So based on the observations x _p+1 ,x _p+2 ,…x _p+m , we can get

Then the joint probability density of (ε _p+1 ,ε _p+2 ,…,ε _p+m ) is

Then the likelihood function is

Then the likelihood equation is

where 1≤i≤p, 1≤l≤p.

First of all, since we assume that the model is a p-order autoregressive model, a ₁ , a ₂ ,..., a _p ≠ 0, according to 1≤i≤p can obtain ε _p+1 =ε _p+2 =ε _2p =0. Then the maximum likelihood estimate of σ ² can be solved as:

Secondly, according to 1≤l≤p can be obtained

When l=1,2,...,p

where ε _j =x _j -a ₁ x _jp -a ₂ x _j-p+1 -...-a _p x _j - ₁ , substitute this formula into (19) to get

From (20), the values of parameters a ₁ , a ₂ ,..., a _p can be solved, and then the values of parameters μ and Σ can be obtained.

That is, in the objective function g(w) _min , the equation of formula (9) has been established by the AR(P) model, and the unknown parameters have been solved by the maximum likelihood estimation method. At this time, the value of a _i is the value of _wi , which is substituted into In equation (10), the output sample value can be obtained when g(w) _min is the smallest achieve the purpose of signal reconstruction. This BP-ML method estimates more parameters and is more complicated, but it improves the accuracy and error of the BP algorithm.

The embodiments of the present invention are described in detail below, and the embodiments are exemplary and intended to explain the present invention, but should not be construed as a limitation of the present invention.

When training the neural network, 12 rounds of data from a certain batch of sampling tests are selected as training samples, which can typically represent the test conditions of solid rocket motors. In a specific application, because the BP-ML algorithm of the above invention is relatively complex in estimating the weight value process, it is applicable in the case of high precision requirements. Therefore, the BP method is still implemented in the following example.

judgement standard

The error judgment of the restored value is based on the root mean square of the signal as the judgment standard,

The signal root mean square (SMSE) formula is as follows:

According to the effect of the recovery value obtained by the neural network, the correlation coefficients of different delays (Different Delay Correlation Coefficients, DDCC) can be used to evaluate, and the degree of correlation between the recovery value and the target value can be judged. The higher the accuracy obtained, the more accurate the neural network established. Given that the correlation coefficient threshold is 5%, that is, |1-C|≤0.05, the correlation is considered to be high.

Where: s _T is the target output, is the recovery value, and τ is the delay value.

The following is a detailed description of the abnormal parameter test of the neural network:

Step 1: Establish a neural network and train to obtain the network weight value; apply a three-layer BP artificial neural network, select the thrust F and pressure P, the main parameters of the twelve batches of sampled test data, as the research objects, and the working time is 0s ~ 18s , the time interval is selected as 0.2s, that is, the number of each waypoint value is 91, the input layer neuron node selects the relevant parameters affecting the output layer node, respectively selects the thrust F value and the pressure P value of the 8 data, and the output layer selects the reflection The key parameter of the engine working condition is the pressure value P ₁ as the target output, and P ₁ is the pressure value of the first data, that is, the input parameters are P ₁ , P ₂ , P ₄ , P ₅ , P ₆ , P ₇ , P ₉ , P ₁₀ , F ₁ , F ₂ , F ₄ , F ₅ , F ₆ , F ₇ , F ₉ , F ₁₀ , 8 channels of pressure, 8 channels of thrust, a total of 16 channels of parameters, the output target value is P ₁ , hidden layer neural The selection of the number of meta nodes needs to comprehensively consider the training efficiency and training effect. The bigger the better, the larger the number of nodes. If the number of nodes is too large, the amount of computation will increase, resulting in slow training, or even over-training, resulting in divergent output values; if the number of nodes is too small, It will make the training results poor and the output unsatisfactory. Therefore, after data analysis, the number of nodes is selected as 50, the training effect is the most ideal, and the error value is the smallest; the activation function selects a bipolar s-shaped function, because its value range is in [- 1,1], and the maximum pressure value reaches 10 ⁵ kPa, so all input values and target values can be reduced by ^10-5 .

The initial weight matrix is assigned random numbers to start training the network. When the error reaches the predetermined requirement, the weight at this time is saved, and the actual engine test data can be input as a test sample, which is calculated on behalf of the human value for parameter detection.

Figure 3 shows the correlation between the target value and the repair value. It can be seen that the source signal and the separated signal have a high degree of correlation, regardless of the trend or the value range, they are all consistent; and under different delays, the correlation coefficients are different. When τ is 7, the value of CCDD is the largest, indicating the highest degree of correlation. SMSE=2.8687. C=1.0008, that is, the error accuracy between the target value and the repair system is 0.08%.

In order to verify the reliability of this recovery value, the pressure P ₂ of another data is selected as the output. When τ is 21, the correlation coefficient is C=1.0005, that is, the error accuracy between the target value and the repair system is 0.05%, and the degree of correlation is 0.05%. Highest.

Step 2: Discrimination of abnormal parameters; in order to avoid misjudging normal data as abnormal, the error threshold during detection should be greater than the training error threshold. By saving the weight at this time, the actual engine test data can be input as a test sample, which can be calculated on behalf of the human value for parameter detection. Because the pressure and thrust data of a certain batch of sampling test data collected are used as the main parameters, the acquisition rate is 100%, and there is no abnormal situation. Therefore, we artificially expand the 40 point values in _P9 by ten times to construct abnormality. For parameters _wP9 , input values P1, P2, P4, _P5 , _P6 , _P7 , _wP9 , _P10 , _F1 , _F2 , _F4 , _F5 , _{F6 , F7 , F 9} , F ₁₀ , 8 channels of pressure, 8 channels of thrust, a total of 16 parameters, the output target value is P ₁ , and the neural network is repaired to obtain the repaired P ₁ , no matter when the delay value τ takes any value, |1- C|≥0.05, that is, the correlation between the repaired P1 and the target value P1 is poor, so it can be deduced that there are abnormal parameters in the input value. In the same way, under the premise of the neural network model we constructed, if we want to determine whether the thrust F and pressure P of a certain engine are abnormal, then replace the thrust and pressure value in the input value with the parameter to be determined, and the output target value remains unchanged. , let it be P ₁ , get the repaired value and compare it to observe whether the correlation coefficient is within the specified threshold, that is |1-C|≤0.05, if it is satisfied, there is no abnormal parameter, if not, there is abnormal parameters, and further replace the detected values in turn to determine specific abnormal parameters.

Step 3: Repair of abnormal parameters; according to the neural network model we trained, after it is determined that wP ₉ is an abnormal parameter, abnormal data needs to be repaired, then set the output layer neuron node in Figure 1 as the target parameter to be restored wP ₉ , remove the target parameter from the input layer neuron node, establish the data of each key parameter separately, reconstruct the neural network, and the training method is the same as before.

Application steps of the present invention:

The first step: determine the abnormal point value; select the steady-state data of the pressure value P, the most important parameter in a certain solid rocket motor test. First check the normality of the data, and calculate the skewness and kurtosis of the measured data according to (1) and (2) above. Apply t-test; secondly, according to formulas (3), (4), (5), complete the t-test process, identify abnormal points and eliminate them.

Step 2: Train the neural network, detect abnormal parameters, and reconstruct the parameter values.

Although the embodiments of the present invention have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Variations, modifications, substitutions, and alterations to the above-described embodiments are possible within the scope of the present invention without departing from the scope of the present invention.

Claims (4)

1. an abnormal data intelligent diagnosis method in a solid rocket motor, is characterized in that: comprise the following steps: Step 1: Collect the test data of the solid rocket motor, and check whether the selected test data satisfies the normal distribution, if so, proceed to Step 2, otherwise select other test data for re-testing; Step 2: Use the t test method to identify the abnormal points in the selected test data and eliminate them; Step 3: Using the BP-ML algorithm and using the neural network to test the abnormal parameters of the test data processed by the t test in step 2, if there are abnormal parameters, reconstruct the parameter values of the abnormal parameters. 2. The method for intelligently diagnosing abnormal data in a solid rocket motor according to claim 1, wherein in step 1, it is judged whether the test data satisfies normal distribution by calculating the skewness and kurtosis of the test data. 3. the abnormal data intelligent diagnosis method in a kind of solid rocket motor according to claim 1, is characterized in that: in step 3, BP-ML algorithm is to adopt maximum likelihood estimation by the weight vector estimation method in artificial neural network BP algorithm method is achieved. 4. the abnormal data intelligent diagnosis method in a kind of solid rocket motor according to claim 3, is characterized in that: the concrete process that carries out abnormal parameter inspection in step 3 is: Step 3.1: Use the data without abnormal parameters as sample data to train the neural network to obtain a trained neural network; Step 3.2: Input the actual test data into the neural network to repair the neural network. If the correlation between the repaired output value and the target value is poor, it is inferred that there are abnormal parameters in the input value of the neural network; if there are abnormal parameters in the input value of the neural network , then the parameter values in the input value of the neural network are judged separately to obtain specific abnormal parameters; Step 3.3: Use the abnormal parameters obtained in step 3.2 as the output parameters of the neural network, remove the abnormal parameters from the input parameters of the neural network, and then reconstruct the neural network to obtain the repaired abnormal parameters.