A fault warning method for tracked vehicle power unit based on oil spectral analysis

Abstract

The present invention belongs to the technical field of fault warning of tracked vehicle power plant, and specifically relates to a fault warning method of tracked vehicle power plant based on oil spectral analysis, including: (1) using particle swarm algorithm-support vector machine to estimate the probability density of oil spectral data, integrating the probability density function to obtain the probability distribution of the element concentration, and combining dynamic weights to determine the warning value and danger value of spectral characteristic parameters; (2) adding the oil replenishment amount and the detection time interval to the data to achieve data reconstruction, and using the gated recurrent unit GRU prediction model to predict the spectral characteristic parameters after normalizing the data, and comparing the prediction curve with the warning value and danger value, so as to achieve fault graded warning. The present invention has strong engineering applicability, realizes the purpose of threshold classification and graded warning of spectral characteristic parameters, and the method is reasonable and simple, providing a methodological basis for similar fault warning problems.

Description

Technical Field

[0001] The invention belongs to the technical field of fault warning of a tracked vehicle power unit, and in particular relates to a fault warning method of a tracked vehicle power unit based on oil spectral analysis. Background Art

[0002] Oil spectral analysis is a commonly used method for fault warning of tracked vehicle power units. The core work is to determine the failure thresholds such as warning values ​​and danger values ​​of the spectral characteristic parameters in the tracked vehicle power unit oil, and predict the spectral characteristic parameters based on multiple oil test results. Then, by comparing the predicted value with the failure threshold, it is determined whether to issue a fault alarm.

[0003] The traditional method for determining the failure threshold assumes that the spectral characteristic parameters obey the normal distribution, and uses the results of multiple oil spectral analysis to calculate the mean and standard deviation of the spectral characteristic parameters, obtain the normal distribution function of the spectral characteristic parameters, and further deduce the failure threshold. However, the measured oil spectral characteristic parameters of the tracked vehicle power unit are greatly affected by the environment and working conditions, and the distribution form is complex, generally not subject to the normal distribution. The failure threshold of the spectral characteristic parameters determined by the traditional method is low in accuracy. A new method is needed to directly deduce the failure threshold of the spectral characteristic parameters based on the probability distribution function and dynamic weight of the spectral characteristic parameters.

[0004] The traditional spectral characteristic parameter prediction method uses historical detection data to estimate the next detection result, and the accuracy is higher when the time intervals of each detection are close. However, due to the constraints of engineering operability, the intervals between each detection of the tracked vehicle power unit are generally quite different. The traditional method cannot consider the influence of the detection time interval, so the prediction accuracy is low. It is necessary to adopt a new method that can consider the difference in the time interval of the oil detection of the tracked vehicle power unit, reconstruct the data of the spectral characteristic parameter detection result sequence, and thus improve the accuracy of the spectral characteristic parameter prediction results. Summary of the invention

[0005] (I) Purpose of the invention

[0006] In view of the problems in the existing tracked vehicle power unit fault warning method based on oil spectral analysis, such as the failure threshold accuracy of spectral characteristic parameters is greatly affected by the environment and working conditions, and the predicted values ​​of spectral characteristic parameters have too high requirements on the consistency of detection time intervals, the present invention proposes a tracked vehicle power unit fault warning method based on dynamic weights and time series reconstruction of oil spectral data, thereby improving the accuracy of tracked vehicle power unit fault warning results.

[0007] (II) Technical solution

[0008] The present invention provides a method for early warning of a power unit fault of a tracked vehicle based on oil spectral analysis, the method comprising the following steps:

[0009] Step 1, determine the failure threshold of the oil spectral characteristic parameters: use particle swarm algorithm-support vector machine to estimate the probability density of the oil spectral data, obtain the probability density function, integrate the probability density function, obtain the probability distribution of the spectral characteristic parameters, and then determine the failure threshold of the spectral characteristic parameters, that is, the warning value and the danger value, through dynamic weights;

[0010] Step 2, spectral feature parameter prediction: add the detection time interval parameter to the tracked vehicle power unit oil detection data for normalization, and then reconstruct the time series. Then, use the gated recurrent unit GRU model to predict the spectral feature parameters. Finally, compare the prediction result with the failure threshold. If the prediction result is greater than the failure threshold, a fault warning is issued.

[0011] Among them, the failure threshold is divided into two types according to the severity and hazard of the power unit failure, namely, the warning value and the danger value;

[0012] Among them, the warning value refers to the spectral characteristic parameter value corresponding to a general fault, and the danger value refers to the spectral characteristic parameter value corresponding to a serious fault.

[0013] Among them, in the step 1, the dynamic weight is determined using the normal distribution 3σ principle, so that the failure threshold changes dynamically with the sample size of the spectral characteristic parameters, and the failure threshold of the spectral characteristic parameters is determined in combination with the probability distribution of the spectral characteristic parameters.

[0014] Wherein, in the step 1, the probability distribution of the spectral characteristic parameters is a probability distribution function constructed by using a particle swarm algorithm-support vector machine based on the measured spectral data of the oil.

[0015] Among them, in the step 1, in the particle swarm algorithm-support vector machine, the particle swarm algorithm PSO is used to optimize the kernel parameter σ and the penalty parameter C, the PSO population size is set in the range of 40 to 50, the algorithm termination condition is to reach the maximum number of iterations, the number of iterations is set to 1000 to 1500, and the value ranges of the kernel parameter σ and the penalty parameter C are 0.5 to 2 and 0.1 to 0.9 respectively.

[0016] Among them, in step 2, the input of the gated recurrent unit GRU model is 5 groups of normalized continuous spectral feature parameter detection values, detection time intervals and oil replenishment amounts.

[0017] Wherein, the mathematical model of the normalization process is:

[0018]

[0019] z i represents the normalized input vector, z i′ represents the input vector before normalization, For z i The average value of ′, σ z For z i The standard deviation of ′.

[0020] The gated recurrent unit (GRU) model is used to iteratively calculate input vectors containing five consecutive groups in a step of 10 motor hours according to the measured spectral characteristic parameter values, so as to obtain the predicted values ​​of the spectral characteristic parameters at the subsequent detection time.

[0021] (III) Beneficial effects

[0022] The method for early warning of a tracked vehicle power unit fault based on oil analysis proposed in the present invention includes two steps: failure threshold determination and spectral characteristic parameter prediction. Failure threshold determination: using particle swarm algorithm-support vector machine to estimate the probability density of the spectral characteristic parameters of the tracked vehicle power unit oil, obtain the probability distribution function of the spectral characteristic parameters, and then use dynamic weights to determine the early warning value, danger value and other failure thresholds of the spectral characteristic parameters; Spectral characteristic parameter prediction: adding the detection time interval parameter to the tracked vehicle power unit oil detection data for normalization processing, and then using the gated recurrent unit GRU model to predict the spectral characteristic parameters, and finally comparing the prediction results with the early warning value, danger value and other failure thresholds to determine whether to perform a fault early warning.

[0023] Compared with the prior art, the tracked vehicle power unit fault warning method based on oil analysis proposed in the present invention can calculate the failure threshold of the spectral characteristic parameters according to the probability distribution function and dynamic weight of the spectral characteristic parameters, and consider the difference in oil detection time intervals to predict the spectral characteristic parameters. This solves the problems that the traditional fault warning method is greatly affected by the environment and working conditions and has high requirements on the time intervals between each detection, and improves the accuracy of tracked vehicle power unit fault warning based on oil analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is the overall logic block diagram of the present invention.

[0025] Figure 2 It is a basic flow chart of the calculation of the failure threshold of the spectral characteristic parameters of the present invention.

[0026] Figure 3 It is a schematic diagram of the empirical distribution function of the Fe element measured by the spectrum of a tracked vehicle power unit and the probability distribution curve estimated by PSO-SVM.

[0027] Figure 4 It is a basic flow chart for predicting the spectral characteristic parameters of the example of the present invention.

[0028] Figure 5It is an algorithm structure diagram of the GRU prediction model of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, content, and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below in conjunction with the accompanying drawings and examples.

[0030] In order to solve the problems of the prior art, the present invention provides a method for early warning of a tracked vehicle power unit fault based on oil spectral analysis, such as Figures 1 to 5 As shown, the method comprises the following steps:

[0031] Step 1, determine the failure threshold of the oil spectral characteristic parameters: use particle swarm algorithm-support vector machine to estimate the probability density of the oil spectral data, obtain the probability density function, integrate the probability density function, obtain the probability distribution of the spectral characteristic parameters, and then determine the failure threshold of the spectral characteristic parameters, i.e., the warning value and the danger value, through dynamic weights;

[0032] Step 2, spectral feature parameter prediction: add the detection time interval parameter to the tracked vehicle power unit oil detection data for normalization processing, and then reconstruct the time series. Then, use the gated recurrent unit GRU model to predict the spectral feature parameters. Finally, compare the prediction results with failure thresholds such as warning values ​​and danger values ​​to determine whether to issue a fault warning.

[0033] Example

[0034] The oil spectral characteristic parameters refer to the element composition and concentration values ​​obtained by multiple spectral detection of the oil of the tracked vehicle power unit. The following takes the analysis of the Fe element in the oil spectrum of a tracked vehicle power unit as an example. The detected Fe element concentration data {x1, x2, …, x i ,…,x N}, where x i Represents the Fe element concentration detected for the i-th time.

[0035] like Figure 2 As shown, the steps of calculating the failure threshold in the present invention include:

[0036] Step 1.1, calculate the empirical distribution function F of Fe element concentration based on the above data N (x):

[0037]

[0038] N is the measurement sample length, x i represents the Fe element concentration detected for the i-th time, and x is the independent variable of the function.

[0039] Substituting the Fe element concentration value into the above empirical distribution function, we get the following data pairs: {(x1,F1(x)),(x2,F2(x)),…,(x N ,F N (x))}.

[0040] Step 1.2, solve the probability density function through support vector machine, first define the kernel function and cross kernel function.

[0041] Kernel function:

[0042]

[0043] x i ,x j represents the coordinates of the kernel function point, Ψ(x) is a linear transformation, σ represents the kernel parameter, and t1, t2 are integral variables.

[0044] Cross kernel function:

[0045]

[0046] x i ,x represents the coordinates of the cross kernel function point, is another linear transformation, t is the integral variable. The kernel function and the cross kernel function use the same kernel parameter.

[0047] Using support vector machine in F N Regression estimation is performed in the (x, w) space, and a linear support vector machine with an ε-insensitive loss function is used to transform the problem of estimating the probability density function into the following optimization problem:

[0048]

[0049] The constraints are:

[0050]

[0051]

[0052]

[0053] ξ i ≥0,i=1,…,N

[0054] in, is a non-negative slack variable; α i is the Lagrange multiplier; C is the penalty parameter; N is the length of the measurement sample; x i ,x j is the concentration value detected by spectrum; F N (x i) is the empirical distribution function; λ is a constant greater than 0; K(.) is the kernel function; the penalty parameter C and the kernel parameter σ are the parameters to be optimized.

[0055] Step 1.3 Kernel parameter and penalty parameter optimization

[0056] Step 1.3.1, initialize PSO-SVM parameters. Initialize the population size to 40 (usually set to 40-50), the algorithm terminates when the maximum number of iterations reaches 1000 (usually set to 1000-2000), and randomly initialize the speed of each particle (k represents the kth particle, the same below; The value range is -1 to 1) and position (Corresponding to the kernel parameter σ and the penalty parameter C, the value range of σ is 0.5~2, and the value range of C is 0.1~0.9).

[0057] Step 1.3.2, calculate the fitness function value: N represents the sample length, F N (.) represents the empirical distribution function, F(.) represents the estimated probability distribution function, x i Represents the i-th sample value in the sample. Adjust the optimal position pBest of the particle history according to the size of the fitness value k (that is, the particle position corresponding to the minimum fitness value of the particle in all historical iterations) and the global optimal position gBest of the group (the particle position corresponding to the minimum fitness value of all particles in this iteration).

[0058] Step 1.3.3, update the particle's speed and position according to the particle's historical optimal position and the global optimal position. The update formula is as follows:

[0059] v′ k =ω×v k +c1×r1×(pBest k -s k )-c2×r2×(gBest-s k )

[0060] s′ k =s k +v′ k

[0061] In the formula, v k represents the velocity of the kth particle before update, v′ k represents the updated speed of the kth particle, ω is the inertia weight, which is generally 0 to 1, c1, c2 are acceleration coefficients, which are generally 2, r1, r2 are random numbers between 0 and 1, s krepresents the position of the kth particle before update, s′ k Represents the updated position of the kth particle.

[0062] Step 1.3.4, re-evaluate the fitness function value according to the updated particle speed and position, and update the particle's historical optimal position and the group's global optimal position. If the termination condition is met, output the global optimal position and exit, otherwise go to step 1.3.3 to continue execution.

[0063] Step 1.4, estimate the probability distribution

[0064] Solve the optimization problem in step 1.2 by using the kernel function to obtain the support vector and the corresponding coefficients The probability density function of Fe element concentration is calculated using the following formula:

[0065]

[0066] k(x i ,x) is the cross kernel function, x is the independent variable, and N is the measurement sample length.

[0067] Integrating the probability density function, we can get the probability distribution of Fe element concentration:

[0068]

[0069] Figure 3 The empirical distribution function of Fe element in the spectrum measurement of a tracked vehicle power unit and the probability distribution curve estimated by PSO-SVM.

[0070] Step 1.5, determine the failure threshold

[0071] The failure threshold of the concentration of abrasive elements is determined by a dynamic weight method. The present invention uses the dynamic weight γ in combination with the estimated probability distribution to determine the limit value of the concentration.

[0072] Step 1.5.1, first determine the Fe element concentration y1 when F(y1) = φ(1) = 0.8413, the Fe element concentration y2 when F(y2) = φ(2) = 0.9772, and the Fe element concentration y3 when F(y3) = φ(3) = 0.9987, where F(.) is the estimated probability distribution function and φ(.) represents the standard normal distribution.

[0073] Step 1.5.2, determine the dynamic weights γ1, γ2. Use the following formula: Indicates the number of samples whose Fe element concentration value in the measured sample is less than y1; It indicates the number of samples whose Fe element concentration value is greater than y2 in the measured sample; N indicates the length of the measured sample.

[0074] Step 1.5.3, determine the concentration warning value, Y w =γ1y1+(1-γ1)y2; determine the concentration risk value, Y r =γ2y3+(1-γ2)y2.

[0075] like Figure 4 As shown, the steps of predicting the concentration of Fe element in the present invention include:

[0076] Step 2.1, determine the model input

[0077] Step 2.1.1, the concentration of Fe in the oil of the tracked vehicle power unit is not only related to the degree of wear of the parts, but also to the amount of oil replenished before the spectral detection. Therefore, the amount of oil replenishment needs to be taken into account when predicting the concentration of Fe in the oil of the tracked vehicle power unit. In addition, since the time interval of spectral detection ranges from a few hours to dozens of hours, it belongs to non-uniform sampling. The longer the time interval, the greater the wear of the parts of the tracked vehicle power unit, and the higher the detected Fe concentration. The prediction of the Fe concentration in the oil of the tracked vehicle power unit needs to take the measurement time interval into account.

[0078] Taking the analysis of Fe element in the oil spectrum of a tracked vehicle power unit as an example, the Fe element concentration data {x1, x2, …, x N}, spectral detection time {t1,t2,…,t N}, the amount of lubricating oil added after each spectrum detection {b1,b2,…,b N}, N represents the number of spectrum detection times. The time interval of spectrum detection is expressed as {Δt1, Δt2, …, Δt N}, where Δt1 = 0, Δt i =t i -t i-1 (i=2,3,…,N).

[0079] Therefore, the Fe element concentration of the oil spectrum of a tracked vehicle power unit, the spectrum detection time interval, and the amount of oil replenishment are used as the input vector Z of the prediction model. i ′=[x i ,b i ,Δt i ] T , the data set is a set of N input vectors: Z′={Z1′,Z′2,…Z′ N}.

[0080] Step 2.2, data preprocessing

[0081] 1. In order to enhance the anti-interference ability of the prediction model, a Gaussian low-pass filter is used to filter the data in the data set;

[0082] 2. The prediction model requires that the input data have the same scale, so the parameters in the data set are normalized to the maximum and minimum. The normalization formula is as follows:

[0083]

[0084] Z i represents the normalized input vector, Z i ′ represents the input vector before normalization, Z i The average value of ′, σ Z Z i The standard deviation of ′.

[0085] Step 2.3, Fe element concentration prediction

[0086] The predicted value of Fe element concentration in the tracked vehicle power unit oil is related to the historical detection data, and one of the characteristics of the GRU model is that it needs historical detection data as one of the input quantities, so GRU is selected as the prediction model for spectral element concentration. To ensure the accuracy of the prediction, the number of prediction steps in this prediction model is set to 1, and the prediction time for each step is 10 motor hours.

[0087] The basic structure of the GRU prediction model is as follows Figure 3 shown.

[0088] 1. Take 5 continuous vectors {Z1, Z2, Z3, Z4, Z5} as the historical input data of the model, and the concentration value x6 of the next item as the predicted value of the model;

[0089] 2. Continue to select the model's historical input data and predicted values ​​in steps of 1 until the predicted value reaches x N For example, after step 1, {Z2, Z3, Z4, Z5, Z6} is used as the historical input data of the model, and the concentration value x7 of the next item is used as the predicted value of the model.

[0090] 3. Train the model until the loss function meets the requirements.

[0091] 4. N-4 ,Z N-3 ,Z N-2 ,Z N-1 ,Z N} as model input to predict the next concentration value x N+1 , and for x N+1 After reverse normalization, the predicted value of concentration can be obtained.

[0092] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for early warning of a tracked vehicle power unit fault based on oil spectral analysis, characterized in that: The method comprises the following steps: Step 1, determine the failure threshold of the oil spectral characteristic parameters: use the particle swarm algorithm-support vector machine to estimate the probability density of the oil spectrum, obtain the probability density function, integrate the probability density function, obtain the probability distribution of the spectral characteristic parameters, and then determine the failure threshold of the spectral characteristic parameters, that is, the warning value and the danger value, through the dynamic weight; The process of determining the failure threshold of the spectral characteristic parameters through dynamic weights is as follows: Step 1.5, determine the failure threshold The failure threshold of wear element concentration is determined by the dynamic weight method, and the limit value of concentration is determined by the dynamic weight γ combined with the estimated probability distribution. Step 1.5.1, first determine the concentration y1 in the oil spectral data when F(y1) = φ(1) = 0.8413, the concentration y2 in the oil spectral data when F(y2) = φ(2) = 0.9772, and the concentration y3 in the oil spectral data when F(y3) = φ(3) = 0.9987, where F(.) is the estimated probability distribution function and φ(.) represents the standard normal distribution; Step 1.5.2, determine the dynamic weights γ1, γ2; use the following formula: Indicates the number of samples whose concentration values ​​in the oil spectrum data of the measured sample are less than y1; It indicates the number of samples whose concentration value in the oil spectrum data in the measured sample is greater than y2; N indicates the length of the measured sample; Step 1.5.3, determine the concentration warning value, Y w =γ1y1+(1-γ1)y2; determine the concentration risk value, Y r =γ2y3+(1-γ2)y2; Step 2, spectral feature parameter prediction: add the detection time interval parameter to the tracked vehicle power unit oil detection data for normalization, and then reconstruct the time series. Then, use the gated recurrent unit GRU model to predict the spectral feature parameters. Finally, compare the prediction result with the failure threshold. If the prediction result is greater than the failure threshold, a fault warning is issued.

2. The method for early warning of a tracked vehicle power unit fault based on oil spectral analysis according to claim 1, characterized in that: According to the severity and hazard of the power unit failure, the failure threshold is divided into two types, namely, warning value and danger value; Among them, the warning value refers to the spectral characteristic parameter value corresponding to a general fault, and the danger value refers to the spectral characteristic parameter value corresponding to a serious fault.

3. The method for early warning of a tracked vehicle power unit fault based on oil spectral analysis according to claim 2, characterized in that: In step 1, the dynamic weight is determined using the normal distribution 3σ principle, so that the failure threshold changes dynamically with the sample size of the spectral characteristic parameter, and the failure threshold of the spectral characteristic parameter is determined in combination with the probability distribution of the spectral characteristic parameter.

4. The method for early warning of a tracked vehicle power unit fault based on oil spectral analysis according to claim 3, characterized in that: In the step 1, the probability distribution of the spectral characteristic parameters is a probability distribution function constructed by using a particle swarm algorithm-support vector machine based on the measured spectral data of the oil.

5. The method for early warning of a tracked vehicle power unit fault based on oil spectral analysis according to claim 4, characterized in that: In the step 1, in the particle swarm algorithm-support vector machine, the particle swarm algorithm PSO is used to optimize the kernel parameter σ and the penalty parameter C, the PSO population size is set in the range of 40 to 50, and the algorithm termination condition is to reach the maximum number of iterations.

6. The method for early warning of a tracked vehicle power unit fault based on oil spectral analysis according to claim 5, characterized in that: The number of iterations is set to 1000-1500.

7. The method for early warning of a tracked vehicle power unit fault based on oil spectral analysis according to claim 5, characterized in that: The value ranges of the kernel parameter σ and the penalty parameter C are 0.5 to 2 and 0.1 to 0.9 respectively.

8. The method for early warning of a tracked vehicle power unit fault based on oil spectral analysis according to claim 1, characterized in that: In step 2, the input of the gated recurrent unit GRU model is 5 groups of normalized continuous spectral feature parameter detection values, detection time intervals and oil replenishment amounts.

9. The method for early warning of a tracked vehicle power unit fault based on oil spectral analysis according to claim 8, characterized in that: The mathematical model of the normalization process is: z i represents the normalized input vector, z i ′ represents the input vector before normalization, For z i The average value of ′, σ z For z i The standard deviation of ′.

10. The method for early warning of a tracked vehicle power unit fault based on oil spectral analysis according to claim 8, characterized in that: The gated recurrent unit (GRU) model is used to iteratively calculate input vectors containing five consecutive groups in sequence according to the measured spectral characteristic parameter values ​​with a step length of 10 motor hours to obtain the predicted values ​​of the spectral characteristic parameters at the subsequent detection time.

Images