3.1. Servo Valve Erosion and Wear Test
The specific test design includes five parts: test hardware configuration, environmental condition setting, test typical operation, test data acquisition design, and data post-processing.
3.1.1. Test Hardware Configuration
A schematic diagram of the servo valve spool erosion and wear test system is shown in
Figure 2 below. The hydraulic pump 2 is the main pump for the entire flow control system and supplies the working medium or hydraulic oil. Thermometer 3 monitors the initial temperature of the hydraulic fluid and allows changes in the fluid temperature to be monitored at any time during the experiment. Two filters, 4 (1) and 4 (2), are used to filter large-diameter solid particles from the oil circuit. The main function of relief valve 5 is to regulate and maintain a constant pressure throughout the hydraulic system, so its main object of work is the source of the supplied fluid—the hydraulic pump 2. In addition, the four pressure gauges 6 (1)–6 (4) are used to check the pressure at the four oil ports of the servo valve and are connected to a pressure sensor that transmits the pressure signal to the computer. The throttle 8 regulates the system oil flow according to the experimental situation and can also be seen as a specific load acting on the servo valve of the measured flow. Finally, the two shut-off valves 9 can assume the on–off status of the servo valve’s return port for the measured flow. When 9 (1) pathway and 9 (2) disconnect, representing the normal working fluid state of the experiment, that is, the measured flow servo valve erosion and wear experiments are constantly carried out. When the flow servo valve 7 is in the neutral position, 9 (1) is disconnected and 9 (2) is open, the flow meter 10 can be used to detect the zero leakage of the flow servo valve, and the flow meter can be connected to the flow sensor so that the leakage signal can be input to the computer.
Figure 3 below shows the spool sleeve wear test rig, which includes a pump source, servo valve, flow meter, throttle valve, globe valve, heat exchanger, pressure gauge, thermometer, and associated piping system.
(1) Hydraulic pump source: the hydraulic pump source used in this test has a rated pressure of 20.6 MPa and a rated flow rate of 100 L/min.
(2) Servo valve under test: the servo valve used in this experiment has a maximum flow rate of 30 L/min.
(3) Flowmeter: according to the rated flow rate of the system and the requirements of the pipeline system, this experiment used the model LWGYB-25 turbine flowmeter, with a range of 1 to 30 m3/h and an accuracy level of ±0.5.
(4) Throttle valve: in this experiment, throttle valve is equivalent to a flow control valve, and its requirements are low, so this experiment used the Shandong Taifeng hydraulic production model DVP-12 throttle valve.
(5) Globe valve: according to the function of the globe valve in this experiment and the requirements of the pipeline system, we chose the model CJZQ-H20L ball core globe valve, whose nominal pressure is 20.6 MPa and nominal diameter is 20 mm.
(6) Cooler: the test system was used in the BR005 type plate radiator; its design pressure is 1.3 MPa, and the design temperature is 150 °C, with heat transfer area of 6 m2.
(7) Pressure gauge: The pressure gauge used in this experiment is a shock-resistant pressure gauge, and the accuracy level of the pressure gauge is 1.6, where the range of the pressure gauge at the P, A, and B ports is 0 to 30 MPa and the range of the pressure gauge at the T port is 0 to 2.5 MPa. The pressure sensor was placed and connected in the pressure gauge.
This test considered that the pressure gain and leakage changes are caused by the erosion and wear of particles on it, so this experiment adopted the method of real-time recording of pressure gain and leakage changes by the sensor incoming signal computer. The aim of this experiment was to test and record the changes in the performance parameters of direct-drive electro-hydraulic servo valves as the operating time increased. The performance parameters are the characteristics of an instrument and the device itself. For example, in the case of an air conditioner, its heating, cooling, power consumption, etc., are its performance. With this in mind, we selected the pressure gain and leakage of the servo valve performance as our experimental test parameters.
3.1.2. Design of Experimental Conditions
According to GJB 420B-2015 “Classification of solid contamination in aviation working fluids” [
28], the classification was based on the size, particle number, and distribution of solid contaminants in the fluid, as well as the content of solid contaminants per unit volume of working fluid. The standard denotes specific particle size ranges by the letters A, B, C, D, E, and F. The solid contamination is classified into 15 classes according to the maximum limiting number of particles in these six size ranges contained in 100 mL of working fluid.
In this test, the physical failure model was simplified and the diameter of the maximum number of solid particles in the fluid was used to reflect the different fluid contamination levels. Considering the actual application environment of DDV, four different contamination levels of hydraulic fluid were proposed in this test: GJB420B-6, GJB420B-7, GJB420B-8, and GJB420B-9.
- 2.
Opening degree
The simulation data showed four different openings of 0.1 mm, 0.2 mm, 0.3 mm, and 0.4 mm, which were selected.
- 3.
Differential pressure
As the opening and differential pressure together affect the flow rate of the valve and the flow rate of the inlet, the differential pressure also needs to be taken into account. In view of the limitations of the experimental conditions, four different differential pressure conditions of 14 MPa, 16 MPa, 18 MPa, and 20 MPa were selected.
The entire orthogonal experimental working conditions are shown in
Table 1 below. According to GJB3370-1998 “aircraft electro-hydraulic flow servo valve general specifications” [
29], the total life of the servo valve needed to have 600 Fh for its high pollution test, so for each sample test for 500 h, every 50 h between the disassembly of the spool valve sleeve, there was cleaning and debridement treatment. Because the liquid oil temperature is difficult to control during the experiment, the oil temperature was not used as one of the experimental conditions for the time being. The initial oil temperature was set at 35 °C, which is commonly used in the industry.
3.1.3. Basic Test Procedure
Based on the test system built for this test and the purpose of the test, the specific test procedures are listed below:
(1) Check whether the hardness and size of the sleeve spool of the servo valve meet the requirements and install the valve into the test bench if the conditions are met.
(2) Install the servo valve under test in the test system, throttle valve 8 is completely closed, stop valve 9 (2) is closed, and stop valve 9 (1) is opened; start hydraulic pump 2, adjust the pressure of relief valve 5 to the rated pressure of the servo valve; test whether the leakage amount inside the valve under test meets the requirements; after the test is completed, stop the hydraulic pump.
(3) Throttle valve 8 is fully open, stop valve 9 (2) is available, stop valve 9 (1) is closed, and the servo valve port under test is fully open; start hydraulic pump 2 and adjust relief valve 5 and throttle valve 8 until the flow meter is at the specified flow rate under the test profile.
(4) Apply control signal and change the direction of the servo valve every 1 h interval so that the wear on both sides of the spool is balanced and remove the spool valve sleeve for the cleaning and drying process every 50 h of operation.
(5) Repeat step 4 until the test time reaches 500 h, then stop the test of the sample. Replace the sample and start the next stress profile test.
3.1.4. Experimental Data Collection Design
(1) Collection interval and test time
The sampling interval is 50 h per cycle, and the data signal is monitored by the computer every 500 s weekly.
(2) Data collection content
(1) Recording of test environment conditions
Before each test, the test environment conditions include test environment temperature and humidity, test temperature, oil contamination, flow rate, test operator, etc. Test time includes test start time, a sampling interval of each test, etc.
(2) Determination of the initial value of the experimental material performance parameters
In this test, the change in pressure gain and leakage volume is used to express the change in erosion and wear during the work of the spool valve sleeve, which in turn characterizes the performance degradation process of the spool valve sleeve, so the pressure gain and leakage volume values at the initial moment of the spool valve sleeve need to be measured.
(3) Data recording during the test
The spool valve sleeve is dismantled at certain intervals, cleaned, and dried and then tested and measured again, as well as the sampling time, test conditions, and other parameters, and then the spool valve sleeve is installed and tested again as required.
3.1.5. Experimental Data Collection Design
The leading cause of failure of the spool sleeve is erosion and wear; therefore, the pressure gain and leakage of the spool sleeve will gradually change during the degradation process due to friction and wear, so the pressure gain and leakage were chosen as the performance degradation indicators for erosion and wear of the spool sleeve in this test. To obtain the degradation curve throughout the test, pressure and flow sensors need to be connected so that the computer can always monitor the parameter signals.
3.2. Preliminary Analysis of Data
Pressure gain is an important concept that is often used to describe the responsiveness of a particular system under different operating conditions. In some cases, pressure gain can vary with time and therefore needs to be monitored and analyzed. This article provides details on the variation of pressure gain with time under different operating conditions.
Firstly, we needed to understand the definition and meaning of pressure gain. Pressure gain is the rate of change of pressure at the output of a system for a given input signal. In some cases, the pressure gain of a system may change over time due to different operating conditions. This means that the system may have different response capabilities at different points in time. It is therefore important to understand the pattern of pressure gain over time under different operating conditions to optimize the performance and stability of the system.
In practical applications, it is important to monitor the trend of pressure gain under different operating conditions. By monitoring and analyzing the pressure gain, we can better understand the performance and stability of the system and make adjustments and optimizations accordingly. For example, in some cases, we may need to increase the responsiveness of the system to accommodate changes in external loads. In this case, we can increase the pressure gain by adjusting the system parameters to increase system responsiveness.
In summary, the pattern of pressure gain over time under different operating conditions has an important impact on the performance and stability of the system. Understanding these patterns and monitoring and analyzing them accordingly can help us optimize the system’s performance and stability and thus better suit the application’s requirements. The variation of pressure gain with time for different operating conditions is shown in
Figure 4 below:
As shown in
Figure 4, we can see the trend of pressure gain over time for different operating conditions. As can be seen from the graph, the trend of pressure gain changed differently under other operating conditions. For example, the pressure gain for condition 1 showed a clear upward trend in the entire cycle, while the pressure gain for condition 10 did not show a clear upward trend. This indicates that the servo valve aged faster in condition 1 and slower in condition 2 during this period. The reasons for these trends may have been due to dynamic changes within the system caused by different operating conditions, such as oil contamination, differential pressure, or openness; changes in external loads; or changes in other factors.
In addition, because servo valves are a commonly used control element, they are widely used in hydraulic systems. During use, leaks may occur inside the servo valve, resulting in reduced performance and increased system energy consumption. Therefore, it is important to monitor and analyze the variation pattern of the internal leakage of servo valves under different operating conditions over time. In this paper, we introduced in detail the internal leakage of servo valves under different working conditions with time. The internal leakage of a servo valve is the amount of leakage between the spool and the valve seat, which is usually used to describe the sealing performance of a servo valve. The internal servo valve leakage amount may change over time under different operating conditions. This may be due to factors such as wear of the servo valve’s internal oil seal, the material’s aging, etc. Therefore, it is important to understand how the internal leakage of servo valves varies with time under different working conditions to ensure the performance and stability of the system. As shown in the figure below, we can see the trend of the internal leakage of the servo valve under different working conditions with time. From the graph, we can see that the trend of the internal leakage of the servo valve was different under different working conditions. The variation of servo valve internal leakage over time under 16 sets of experimental conditions is shown in
Figure 5 below:
As shown in
Figure 5, we can see that some of the conditions had a faster increase in leakage, indicating that they were more affected by erosion and wear. For example, the groups of conditions with greater oil contamination showed more dramatic changes in leakage than other conditions with the same or similar conditions, but lower oil contamination. Combined with the above pressure gain and leakage data analysis graph, it is easy to see that, according to the different working conditions of the servo valve, the impact of changes in pressure gain and leakage was also different. The most noteworthy of these was that with little difference in other influencing factors, the influence of oil contamination had the greatest impact on the change in servo valve performance, followed by differential pressure and then openness. In terms of the overall trend of servo valve performance, pressure gain and leakage were generally slow to increase as the work progressed. The reason for this is that as the servo valve work is continually affected by erosion and wear, especially for the servo valve itself, the sharp edges and square holes of such structures are worn, so the internal leakage of the servo valve will increase and the pressure will be easier to build up, resulting in an increase in pressure gain.
Table 2 below lists the statistics relating to internal leakage and pressure gain.
3.3. Dynamic Prediction of Performance Indicators
In
Section 2.2.1, we explained the theory of the exponential smoothing model, and in this section, we use the pressure gain or internal leakage of the servo valve as a one-dimensional time series for dynamic prediction, wherein the main process is as follows:
Step 1: Confirmation of primary time points
As described in dynamic prediction, the time nodes to be confirmed include the training set start time, the first training set end time, the first test set end time, and the last test set end time.
For example, in this study, the training and prediction selected the pressure gain from 30,000 s to 1,400,000 s of working condition 1 as the training set, and for the prediction of the servo valve pressure gain after 1,400,000 s, the training set start time was 1,400,500 s, the first training set end time was 1,400,500 s, and the first test set end time was 1,401,000 s.
Step 2: Slice and dice the training set and test set
The data set was split according to the training set start time, the training set end time, and the test set end time.
For example, in this study, the first training and prediction in the training set start time was 30,000th s, the training set end time was 1,400,000th s, and the test set end time was 1,401,000th s. Then, the pressure gain from the 30,000th to the 1,400,000th s of condition 1 was intercepted as the training set, and the 1,400,500th s pressure gain was intercepted as the test set.
It should be noted that the time series method training set and test set are one-dimensional data because the pressure gain or and internal leakage units do not coincide.
Step 3: Parameter estimation
Based on the training set data, parameters such as the smoothing index and damping trend were estimated. Depending on the model being selected, the parameters being predicted in this step varied. For example, when simple exponential smoothing was chosen, only the smoothing index was estimated; when double exponential smoothing was chosen, only the smoothing index and the damping trend were estimated; when triple exponential smoothing was chosen, the smoothing index, the damping trend, and the seasonal parameters needed to be estimated.
Step 4: Prediction
Prediction values for the specified prediction length were calculated from the estimates of the smoothed parameters and initial values in the training set from step 2 the end values of the training set, as well as the parameter estimates obtained in step 3.
Step 5: Advancing the sliding time window
After completing a single parameter estimation and prediction, the sliding time window needs to be pushed for the next training and prediction to achieve the effect of dynamic prediction. In this study, the prediction length was fixed at 500 s, so each time, the end time of the training set and the end time of the test set were pushed forward by 500 s. At the same time, the training set start time was kept constant in the time series model.
For example, in this study, the training set started at 30,000th s, the first training set ended at 1400,000th s, and the first test set ended at 1400,500th s. After the first parameter estimation and prediction, the training set start time was kept constant, the training set end time was advanced by 500 s to 1,400,500 s, and the test set end time was also advanced by 500 s to 1,401,000 s.
Step 6: Loop
Loop through steps 2 to 5 until the end time of the test set advances to the end time of the last test set, which was 1,800,000 s at the end of the time series for this data set.
In this study, the ExponentialSmoothingPredictor class was created for the training/test set interception of exponential smoothing methods, training and prediction of simple exponential smoothing, double exponential smoothing, triple exponential smoothing, and other models, as well as for visualizing predicted/true value comparisons and calculating evaluation metrics.
The getTrainTest function was created to intercept the training set, the test set, and the training set. The main input to this function is a data frame containing the time and pressure gain/leakage data for each condition. Other parameters include condition, indexTrainStart, indexTrainEnd, and predictLen. This function calls the .loc method in the pandas module and the subTsDf (subsequence time series dataframe) function in the ExponentialSmoothingPredictor class, which intercepts the input data frame with the condition Dai, corresponding well to the training set start time and training set end time. The variable name yA (y train) is the training set, the variable name yA (y train) is the training set index, and the variable name timeA (time train) is the training set index; the test set start time is the same as the training set end time, and the test set end time is the product of the test set start time and the prediction. The amount of time to start the test set is the same as the amount of time to end the training set; the amount of time to end the test set is the sum of the product of the length of the test set and the frequency of data collection; the amount of initial velocity of emission or bore ablation between the test set start time and the test set end time in the intercepted input data frame and the condition code corresponding to the condition test set is the test set, the variable name yE (y test), the amount of time between the test set start time, and the test set end time in the intercepted input data frame, and the condition code corresponding to the condition test set is the test set index, the variable name timeE (time test). The function defaults to A, the training set start time defaults to 0, the training set end time defaults to 3000, and the prediction length defaults to 1. The final output of the function is the training set data, the training set index, the test set data, and the test assigned index.
In this study, the SimpleES (SimpleExponentialSmoothing) function and the SimpleESLoop (Simple Exponential Smoothing Loop) function of the ExponentialSmoothingPredictor class were created for the construction, parameter estimation, and prediction of simple exponential smoothing models.
The input to the SimpleES function is the training set (yA), the training set index (timeA), the test set (yE), and the test set index (timeE), and the main parameters are whether to optimize or not to optimize, the parameter estimation method, and the initialization_method. The function calls the SimpleExpSmoothing method in the stats models module, first transferring the training set and the optimized or not parameters to the SimpleExpSmoothing method in the stats models module to instantiate it, and finally calling the fit method of SimpleExpSmoothing. The smoothing index is then estimated. To obtain the best prediction results, commonly used parameter estimation methods include “L-BFGS-B”, “TNC”, “SLSQP”, “Powell”, “trust-constr”, “basinhopping”, and “last_squares”, etc. Given the time complexity and effectiveness of the algorithm, the function defaults to the L-BFGS-B method for estimating the smoothing exponent, which is a finite memory algorithm for solving large nonlinear optimization problems with simple restrictions on the variables. It is suitable for problems where information about the Hessian matrix is difficult to obtain or for large dense problems. l-BFGS-B can also be used for unconstrained problems, in which case its performance is similar to that of its predecessor algorithm, L-BFGS. The output of this function is four data frames, namely, the training set data frame (aDf, train data frame), which has two fields, yActual and yPredict, where the true and predicted values are on the training set; eDf (test data frame), which has three fields: tIndex, yActual, and yPredict, where unlike the training set data frame, the true and predicted values in the test set data frame are on the test set; paraDf (parameter data frame), which has two fields, namely, index (t) and smoothing_level; and the evaluation metrics data frame (metricsDf, metrics data frame), which has five fields, namely, index (t), mean absolute error in train set (maeA, mean absolute error in the train set), mean absolute error in the test set (maeE, mean absolute error in the test set), mapeA (mean absolute percent error in the train set), and mapeE (mean absolute percent error in the test set). The most important of the returned values from the SimpleES function was the test set data frame. The predicted values, both indicated in the dynamic prediction, were used directly to determine whether the servo valve was about to fail. Because machine learning algorithms such as support vector machines and random forests predict poorly, and exponential smoothing and ARIMA models do relatively better, the prediction results of algorithms such as support vector machines and random forests will not be elaborated on for the time being.
Figure 6,
Figure 7,
Figure 8,
Figure 9,
Figure 10,
Figure 11,
Figure 12,
Figure 13,
Figure 14,
Figure 15,
Figure 16,
Figure 17,
Figure 18,
Figure 19,
Figure 20,
Figure 21,
Figure 22,
Figure 23,
Figure 24,
Figure 25,
Figure 26,
Figure 27,
Figure 28,
Figure 29,
Figure 30,
Figure 31,
Figure 32,
Figure 33,
Figure 34,
Figure 35,
Figure 36,
Figure 37,
Figure 38,
Figure 39,
Figure 40,
Figure 41,
Figure 42,
Figure 43,
Figure 44,
Figure 45,
Figure 46,
Figure 47,
Figure 48,
Figure 49,
Figure 50,
Figure 51,
Figure 52 and
Figure 53 below show the results of the dynamic prediction of pressure gains up to 120 s in advance for different operating conditions using exponential smoothing, ARIMA, and fusion prediction. The exponential smoothing method predicts the best of the three methods, with a test set MAE of 0.0023 Mpa/mm on average and a MAPE of 1.326579691761824 × 10
−7 on average. All time units in the graph are in seconds.
Figure 6,
Figure 7,
Figure 8,
Figure 9,
Figure 10,
Figure 11,
Figure 12,
Figure 13,
Figure 14,
Figure 15,
Figure 16,
Figure 17,
Figure 18,
Figure 19,
Figure 20,
Figure 21,
Figure 22,
Figure 23,
Figure 24,
Figure 25,
Figure 26,
Figure 27,
Figure 28,
Figure 29,
Figure 30,
Figure 31,
Figure 32,
Figure 33,
Figure 34,
Figure 35,
Figure 36,
Figure 37,
Figure 38,
Figure 39,
Figure 40,
Figure 41,
Figure 42,
Figure 43,
Figure 44,
Figure 45,
Figure 46,
Figure 47,
Figure 48,
Figure 49,
Figure 50,
Figure 51,
Figure 52 and
Figure 53 show that the exponential smoothing method predicted the best of the three prediction methods. Mean absolute error (MAE) and mean absolute percentage error (MAPE) are common measures of prediction accuracy in the process of optimizing the structural parameters of a neural network using an algorithm. MAE indicates the mean of the absolute value of the difference between the predicted and observed values, and MAPE indicates the mean of the absolute value of the difference between the predicted and observed values as a proportion of the observed values. The smaller the value of these indicators, the smaller the difference between the predicted and observed values, and the higher the prediction accuracy. In the process of using the algorithm to optimize the parameters, the MAE and MAPE indicators can be considered simultaneously to assess the prediction accuracy of the model in a comprehensive manner. The MAE and MAPE are calculated using the following formulae:
where
is the mean value of the pressure gain and leakage samples,
yi is the observed value at the
ith operating time point, and
is the predicted value at the
ith operating time point.
The exponential smoothing method is better at predicting changes in performance indicators; for example, the mean absolute error (MAE) and mean absolute percentage error (MAPE) of pressure gain and leakage predicted dynamically using exponential smoothing 120 s in advance are shown in
Table 3 and
Table 4 below:
After the key performance indicators are predicted, life prediction can be made based on whether the failure threshold is reached. Because there is no international development and classification of failure standards for direct-drive electro-hydraulic servo valves or even other control valves, the research in this paper can only be based on the experience of the engineers of the institute that provides direct-drive electro-hydraulic servo valve samples and the requirements of the working environment restrictions in the aerospace industry, usually when the pressure gain is less than 2000 Mpa and the internal leakage is less than 2.5 L/min, wherein it can meet the aerospace servo Valve work requirements. The aim of this paper was therefore also to serve as a pre-requisite study and theoretical guide for the future development of servo valve failure standards in the aerospace sector. According to the long-term prediction results, the lifetime of the servo valve is about 1200 h or more, which is much higher than the 600 h required by the Chinese national standard [
29].