Fault Prediction Method for Switch Machine Based on Tunable Input LSTM Model

Abstract

The present invention discloses a method for predicting the faults of a switch machine based on an adjustable input LSTM model. Based on the current signals of phase A in the three-phase current collected under the normal state and three fault states of an AC electric switch machine for a turnout, the present invention extracts the features of the current signals to obtain the current feature sequences from the normal state to any one of the fault states, constructs the input of the belief rule fusion model from the current feature sequences, and outputs the adjustable input factors α t ; from the current feature sequences and the adjustable input factors α t construct the input of the adjustable input LSTM model, and finally predict the fault types of the AC electric switch machine. The present invention can improve the accuracy of fault prediction.

Description

Technical Field

[0001] The present invention relates to a method for predicting the faults of a switch machine based on an adjustable input LSTM model, belonging to the technical field of fault prediction and fault diagnosis of industrial equipment. Background Art

[0002] The safety of railway transportation is of utmost importance. A switch is one of the key ground signal devices for high-speed railways. Its main function is to switch the advancing direction of high-speed trains to achieve the train's line change and cross-line operation. If the switch cannot be switched to the corresponding position, the train's progress will be affected, and serious accidents may occur. The switch machine is the device that controls the switch conversion and mainly completes the function of switch conversion. Therefore, the switch machine plays an important role in the safety and efficient transportation of trains. Due to the frequent turning use of the switch, the working density of the switch machine is relatively large, the frequency of conversion actions is high, and it needs to work outdoors in the open air for a long time, so faults will inevitably occur. Its working state is directly related to the normal operation of the train and poses a safety hazard to the operation of high-speed railways. Therefore, to ensure the normal conversion of the switch, it is necessary to ensure that the switch machine is in a normal working state. Therefore, it is of great significance to predict and diagnose the faults of the switch machine.

[0003] With the rapid development of railways towards high speed and heavy load, more intelligent fault prediction and fault diagnosis methods are needed to improve the safety and reliability of switch machine equipment. Therefore, studying the fault prediction method for AC electric switch machines can realize the early prediction of the faults of AC electric switch machines, thereby avoiding greater losses caused by switch machine faults and providing guarantee for the safe operation of switch equipment. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention proposes a method for predicting the faults of a switch machine based on an adjustable input LSTM model.

[0005] The present invention includes the following steps:

[0006] S1: Collect the current signal of phase A in the three-phase current under the normal state and three fault states of the electric switch machine, extract the features of the current signal, and obtain the current feature sequence from the normal state to any one of the fault states;

[0007] S2: Construct an adjustable input LSTM fault prediction model. The input of the model is the feature vector composed of the current feature vector sequence in step S1 and the adjustable input factor, and the output of the model is the fault state of the switch machine at the future t + s moment;

[0008] S3: Construct a belief rule fusion model, and infer the value of the adjustable input factor α at time t according to the average change amount of the historical and current current feature sequences t ;

[0009] S4: After obtaining the historical sample data of various states of the electric switch machine, an optimization objective function is established to optimize the parameters in the LSTM prediction model established in S2. After obtaining the current feature sequence online, repeat steps S2 and S3 to obtain the prediction result of the switch machine failure state.

[0010] The beneficial effects of the present invention are as follows: Based on the adjustable input LSTM model, the adjustability of the input is mainly reflected in that this method can more finely mine the important historical information of the fault feature sequence, infer the adjustable input factor through the belief rule fusion model, construct the input of the adjustable input LSTM model according to the adjustable input factor, and use this information to improve the accuracy of fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. By referring to the drawings, the features and advantages of the present invention will be more clearly understood. The drawings are schematic and should not be construed as imposing any limitations on the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. Among them:

[0012] Figure 1 is the method flow block diagram of the present invention;

[0013] Figure 2 (a)-(d) are the A-phase current signal curve diagrams of the switch AC electric switch machine in the normal state and three fault states in the method embodiment of the present invention;

[0014] Figure 3 (a)-(b) are the current feature sequence diagrams obtained by extracting the features of the collected current signals in the normal state and three fault states of the switch AC electric switch machine in the method embodiment of the present invention, Figure 3 (a) and (b) respectively correspond to the kurtosis and mean square error feature sequence diagrams of the current signal;

[0015] Figure 4 is the comparison diagram of the predicted value and the true value of the switch AC electric switch machine in the normal state and the fault state in the method embodiment of the present invention;

[0016] Figure 5 is the value diagram of the adjustable input factor corresponding to the sampling moment in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] To more clearly understand the above objects, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] In the following description, many specific details are set forth in order to provide a thorough understanding of the present invention. However, the present invention may be practiced in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited by the specific embodiments disclosed below.

[0019] As Figure 1 shown, the present invention is based on the current signal of phase A in the three-phase current collected under the normal state and three fault states of the switch AC motor-driven point machine, extracts the features of the current signal, obtains the current feature sequence from the normal state to any one of the fault states, constructs the input of the belief rule fusion model from the current feature sequence, and outputs the adjustable input factor α t ; constructs the input of the adjustable input LSTM model from the current feature sequence and the adjustable input factor α t and finally predicts the fault type of the AC motor-driven point machine.

[0020] To facilitate the understanding of the above technical solution of the present invention, the above technical solution of the present invention will be described in detail below through specific embodiments.

[0021] Embodiment:

[0022] The flow block diagram of this embodiment is as Figure 1 shown and includes the following steps:

[0023] S1: Collect the current signal of phase A in the three-phase current under the normal state and three fault states of the motor-driven point machine, extract the features of the current signal, and obtain the current feature sequence from the normal state to any one of the fault states;

[0024] S2: Construct an adjustable input LSTM fault prediction model. The input of the model is the feature vector composed of the current feature vector sequence and the adjustable input factor in step S1, and the output of the model is the fault state of the point machine at the future t + s moment;

[0025] S3: Construct a belief rule fusion model, and infer the value of the adjustable input factor α t at time t according to the average change amount of the historical and current current feature sequences;

[0026] S4: After obtaining the historical sample data of various states of the motor-driven point machine, establish an optimization objective function to optimize the parameters in the LSTM prediction model established in S2. After obtaining the current feature sequence online, repeat steps S2 and S3 to obtain the prediction result of the fault state of the point machine.

[0027] Further, the state of the electric switch machine in step S1 is represented as F n , n = 0, 1, 2, 3, where F 0 represents the normal state of the electric switch machine, F 1 , F 2 , F 3 respectively represent motor abnormal failure, turnout indication circuit abnormal failure, and idling failure during the conversion process of the switch machine; the extracted characteristics of the current signal are kurtosis x 1 and mean square error x 2 as characteristic variables representing faults, and the current characteristic sequence they form is X = {[x 1 (t), x 2 (t)x 2 (t)]|t = 1, 2,..., T}, where t represents the sampling moment and T represents the length of the sampling sequence.

[0028] The action process of the AC electric switch machine includes starting, unlocking, converting, locking, and communication. The A-phase action current curve completely contains the information of each stage such as starting, unlocking, converting, locking, and communication indication during the entire turnout operation process. When a fault occurs during operation, the current representing one or several stages in the action process will be different from the current during normal operation, and the current has different responses to different faults. Therefore, by analyzing the A-phase action current curve, the fault category of the AC electric switch machine can be judged.

[0029] For the three fault categories of the AC electric switch machine, namely the normal state, motor abnormal failure, turnout indication circuit abnormal failure, and idling failure during the conversion process of the switch machine, the corresponding state labels of the AC electric switch machine are F 0 , F 1 , F 2 , F 3 . The current signal of phase A in the three-phase current is collected. Under the four states, the A-phase current signal curve diagrams of the AC electric switch machine are as shown in Figure 2 (a)-(d). The kurtosis and mean square error of the current signal are extracted as characteristic variables representing faults for each action process of the AC electric switch machine. Each of the four states of the AC electric switch machine is sampled 50 times, and a total of 200 samples of fault characteristic variables are extracted, as shown in Figure 3 (a)-(b).

[0030] Further, the input of the adjustable input LSTM model fault prediction model in step S2 is the feature vector L(t) = [x 1 (t), x 2 (t), x 1 '(t), x 2 '(t)], where x 1 ' (t) and x2 ′(t) is calculated by the following formulas (1) and (2) respectively

[0031]

[0032] where α t is called the adjustable input factor and takes a positive integer greater than or equal to 2; the output of this model is the fault state F n (t + s) of the electric switch machine at the future time t + s, s = 1, 2,... S, S ∈ N + .

[0033] Specifically, an input sample of the adjustable input LSTM model is composed of the feature vector at time t in the current feature vector sequence in step S1 and the adjustable input factor α at time t inferred by the confidence rule fusion model constructed in step S3 t The output of the adjustable input LSTM model is set to s = 1, which is the fault state of the electric switch machine at the future time t + 1. In this embodiment, the adjustable input factor α t can take values of 2, 3, or 4. To construct the first input sample of the adjustable input LSTM model, at least when t = 3, the adjustable input factor α 3 is set to 2. From formulas (1) and (2), we can obtain Therefore, the first input sample of the adjustable input LSTM model is L(3) = [x 1 (3), x 2 (3), x 1 ′(3), x 2 ′(3)], and the output of this model is the fault state of the electric switch machine at the next sampling time, that is, t = 4. In other general embodiments, when t ≥ 3, the fault of the AC electric switch machine is predicted according to this method. When 0 < t < 3, the adjustable input factor takes a fixed value of 1

[0034] Furthermore, the specific process of step S3 is as follows

[0035] (1) The input of the confidence rule fusion model is the vector where is calculated by the following formulas (3) and (4) respectively

[0036]

[0037] The output of the model is the value of the adjustable input factor α t at time t

[0038] Specifically, for the input vector of the belief rule fusion model, it is calculated from the current feature vector sequence in step S1 according to formulas (3) and (4). When t = 3, from formulas (3) and (4), we can obtain that and so the input vector of the belief rule fusion model

[0039] (2) In the belief rule fusion model described in step S3, the set of reference values for the variable is where R ≥ 2 is the number of reference values for each input, is the reference value of the r-th reference level of the p-th input, and there is The set of reference values for the output α t of this model is D = {D i | i = 1, 2,..., I}, D i takes positive integer values greater than or equal to 2, and there is D 1 < D 2 ... <... < D I , and I ≥ 2 is the number of reference values for the output;

[0040] (3) In the belief rule fusion model described in step S3, the constructed rule base consists of K rules, K = R 2 , and the k-th rule R k is described as:

[0041]

[0042] where represents the reference value corresponding to the p-th input of this model under the k-th rule, and m i,k is the confidence of D i , and it satisfies Then R k can be simplified to represent the input reference vector and the output confidence vector M k = [m 1,k , m 2,k , …, m I,k ;

[0043] Specifically, in the belief rule fusion model described in step S3, the sets of reference values for the input variables and of this model are C 1 = {3, 28, 52}, that is, C 2 = {1, 3.3, 5.6}, that is, The set of output reference values of the model is D = {2, 3, 4}, and the values in set D respectively correspond to the adjustable input factor α t The confidence levels corresponding to the output reference values are given by expert experience. From formula (5), the rule base formed by the set of input-output reference values of this model is shown in Table 1 below.

[0044] Table 1 Rule Base of the Confidence Rule Fusion Model

[0045]

[0046] (4) In the confidence rule fusion model described in step S3, calculate the Euclidean distance d between vector B(t) and vector A k Then, normalize all d using formula (6) to obtain d k ′, and then use formula (7) to obtain the activation degree vector β k of B(t) with respect to rule R k (t) k β k (t) = d

[0047]

[0048] β k (t) = d k ′·M k = [d k ′m 1,k , d k ′m 2,k , …, d k ′m I,k (7)

[0049] (5) In the confidence rule fusion model described in step S3, use formula (8) to add the corresponding elements of β k (t) of K rules to obtain the fused confidence level vector β(t)

[0050]

[0051] (6) In the confidence rule fusion model described in step S3, form vector G = [D 1 , D 2 ,..., D I from the output reference values of this model. Then, perform vector multiplication according to the following formula (9) to obtain

[0052] α = G × β T (t) (9)

[0053] Select the output reference level value with the smallest difference from α and use it as the value of the adjustable input factor α t .

[0054] Specifically, according to all the rules shown in Table 1, when an input sample B(t) = [5.6, 1.5] of the confidence rule fusion model at a certain moment, calculate the Euclidean distance d between the vector B(t) and all 9 input reference vectors in Table 1 1 ~d 9 , which are respectively d 1 = 2.6476, d 2 = 3.1623, d 3 = 4.8549, d 4 = 22.4056, d 5 = 22.4722, d 6 = 22.7721, d 7 = 46.4027, d 8 = 46.4349, d 9 = 46.5808; from formula (6), d 1 ' = 0.3441,

[0055] d 2 ' = 0.2811, d 3 ' = 0.1877, d 4 ' = 0.0407, d 5 ' = 0.0405,, d 6 ' = 0.0400, d 7 ' = 0.0196, d 8 ' = 0.0196,

[0056] d 9 ' = 0.0196; from formula (7), β 1 = d 1 '·M 1 = 0.3441×[0.1656, 0.4338,

[0057] 0.4006] = [0.0570, 0.1493, 0.1379], and similarly, β 2 ~β 9 are obtained; then from formula (8), β(t) = [0.2164, 0.4513, 0.3324] can be obtained; then from formula (9), α = G×β T (t) = [2, 3, 4]×[0.2164, 0.4513, 0.3324] T = 3.1161. Since 3.1161 is closest to the reference level value 3, the value of α t is 3, and this value will be applied to construct the input sample of the adjustable input LSTM model at time t, and the fault type of the switch AC motorized switch machine at time t + 1 will be predicted by this model.

[0058] Further, in step S4, historical samples of all states of the electric switch machine are obtained, and the number of samples for each state is at least 20, and an optimization objective function ζ is established as follows

[0059]

[0060] where, F t represents the switch machine fault state at time t, F t ∈{F 0 ,F 1 ,F 2 ,F 3}, represents the switch machine fault state given by the adjustable input LSTM fault prediction model at time t, the value set is the same as that of F t , Q>80 represents the total number of samples. The Adam algorithm is used to optimize the parameters in the adjustable input LSTM fault prediction model. The model parameters corresponding to the minimum value of the function ζ are the optimal parameters. After obtaining the current feature sequence online, steps S2 and S3 are repeated to obtain the prediction results of the switch machine fault state.

[0061] Specifically, when establishing the adjustable input LSTM fault prediction model, there are 50 samples for each state of the switch machine, a total of 200 samples. Formula (10) is used as the model optimization objective function. The Adam algorithm is used to optimize the parameters in the adjustable input LSTM fault prediction model. The model parameters corresponding to the minimum value of the function ζ are the optimal parameters. After obtaining the current feature sequence online, steps S2 and S3 are repeated to obtain the prediction results of the switch machine fault state.

[0062] Through online acquisition of the current feature sequence, the prediction result accuracy of the switch machine fault state of the AC switch machine for turnout is shown in Table 2 below. The prediction result diagram of the AC switch machine fault state is as Figure 4 shown. The prediction accuracies of the 4 fault states of the AC switch machine are 97.83%, 96%, 96%, and 97.83% respectively, and the total prediction accuracy is 96.94%. The adjustable input factor α t takes values as Figure 5 shown.

[0063] Table 2 Prediction Results of Switch Machine Fault State

[0064] Status Fault Status 0 Fault Status 1 Fault Status 2 Fault Status 3 Total Prediction Accuracy Accuracy 97.83％ 96.00％ 96.00％ 97.83％ 96.94％

Claims

1. Fault prediction method for switch machine based on adjustable input LSTM model, characterized in that, it includes the following steps: S1: Collect the current signal of phase A in the three-phase current under the normal state and three fault states of the electric switch machine, extract the features of the current signal, and obtain the current feature sequence from the normal state to any one of the fault states; S2: Construct an adjustable input LSTM fault prediction model, the input of the model is the feature vector composed of the current feature vector sequence in step S1 and the adjustable input factor, and the output of the model is the fault state of the switch machine at the future t + s moment; S3: Construct a belief rule fusion model, and infer the value of the adjustable input factor α at time t according to the average change amount of the historical and current current feature sequences. t of; S4: After obtaining the historical sample data of various states of the electric switch machine, establish an optimization objective function to optimize the parameters in the LSTM prediction model established in S2. After obtaining the current feature sequence online, repeat steps S2 and S3 to obtain the prediction result of the fault state of the switch machine; In step S2, the input of the adjustable input LSTM model fault prediction model is the feature vector L(t) = [x 1 (t), x 2 (t), x 1 ′(t), x 2 ′(t)], where x 1 ′(t) and x 2 ′(t) are calculated by the following formulas (1) and (2) respectively Among them, α t is an adjustable input factor, and its value is a positive integer greater than or equal to 2; the output of the model is the fault state F n (t + s), where s = 1, 2,... S, and S ∈ N + ; The specific process of the said step S3 is: S3-1: The input for constructing the belief rule fusion model is a vector where are calculated respectively using the following formulas (3) and (4) The output of the model is the value of the adjustable input factor α at time t t ; S3-2: In the confidence rule - based fusion model described in step S3, determine the reference value set for variable as where R≥2 is the number of reference values for each input, is the reference value of the r - th reference level of the p - th input, and there is The reference value set of the output α t of this model is D = {D i |i = 1,2,...,I}, D i takes positive integer values greater than or equal to 2, and there is D 1 <D 2 ...<...<D I , and I≥2 is the number of output reference values; S3-3: In the confidence rule fusion model described in step S3, the constructed rule base consists of K rules, where K = R 2 , and the k-th rule R k is described as: Among them, represents the reference value corresponding to the p-th input of the model under the k-th rule, and m i,k is the confidence of D i and satisfies Then R k is simplified to represent the input reference vector and the output confidence vector M k =[m 1,k , m 2,k , …, m I,k ; S3-4: In the confidence rule fusion model described in step S3, calculate the Euclidean distance d between vector B(t) and vector A k ; normalize all d using formula (6) to obtain d k ', and then use formula (7) to obtain the activation degree vector β k of B(t) for rule R k ; k (t). k (t) β k (t) = d k ′·M k = [d k ′m 1,k , d k ′m 2,k , …, d k ′m I,k (7) S3-5: In the confidence rule fusion model described in step S3, use formula (8) to add the corresponding elements of β(t) of K rules to obtain the fused confidence vector β(t). k (t) to obtain the fused confidence vector β(t). S3-6: In the confidence rule-based fusion model described in step S3, a vector G = [D 1 , D 2 ,..., D I formed by the output reference values of the model can be obtained by performing a vector multiplication according to the following formula (9) α = G × β T (t) (9) Select the output reference level value with the smallest difference from α and use it as the value of the adjustable input factor α t .

2. The fault prediction method for switch machine based on adjustable input LSTM model according to claim 1, characterized in that, The state of the electric switch machine in step S1 is represented as F n , n = 0, 1, 2, 3, where F 0 represents the normal state of the electric switch machine, F 1 , F 2 , F 3 respectively represent abnormal motor failure, abnormal failure of the switch indication circuit, and idling failure during the conversion process of the switch machine; Extract the kurtosis x of the current signal as the feature 1 and the mean square error x 2 As the characteristic variables representing faults, the current characteristic sequence they form is X = {[x 1 (t), x 2 (t)]|t = 1, 2,..., T}, where t represents the sampling time and T represents the length of the sampling sequence.

3. The fault prediction method for switch machine based on adjustable input LSTM model according to claim 1, characterized in that, in the said step S4, obtain the historical samples of all states of the electric switch machine, and the number of samples of each state is at least 20, and establish an optimization objective function ζ as shown below Among them, F t represents the switch machine fault state at time t, and F t ∈{F 0 , F 1 , F 2 , F 3}, represents the switch machine fault state given by the adjustable input LSTM fault prediction model at time t. The value set is the same as that of F t . Q > 80 represents the total number of samples. The Adam algorithm is used to optimize the parameters in the LSTM fault prediction model. The model parameters corresponding to the minimum value of the function ζ are the optimal parameters. After obtaining the current feature sequence online, steps S2 and S3 are repeated to obtain the prediction result of the switch machine fault state.

Images