I.e. the probability that a sample is not drawn at all times is 36.8%. Therefore, each sample subset contains 63.2% of original data, so that the mutually overlapped sample subsets are obtained, and on the premise that the individual learners have better learning performance, the individual learners have larger difference, and the overall generalization performance is improved.

2. Training an individual decision tree classifier.

And the random forest adopts a CART decision tree as an individual classifier. And training the decision tree classifier by using a CART algorithm by using a sample subset. Each sample subset trains a classifier. Meanwhile, when the decision tree nodes are split, random selection of attributes is introduced, a plurality of attributes are randomly selected from all the attributes of the nodes to form a subset, the best attribute is found in the dependency subset by using the Gini index as a splitting index to split, no pruning is freely generated, and the individual decision tree classifier is generated.

The splitting attribute selection of the random forest is different from a common decision tree algorithm, in the training process, for a common decision tree, all attributes of a current node need to be traversed, the optimal attributes are found and divided, for the random forest, only a plurality of attributes are randomly selected to form an attribute subset each time the random forest is split downwards, and the optimal attributes in the attribute subset are found, namely attribute disturbance. Therefore, the difference between the trained individual decision trees is further increased, the overfitting phenomenon possibly occurring in the decision trees is reduced, and the generalization capability is effectively improved.

The single decision tree classifier training process is as follows:

inputting: training subset D₁。

Wherein D is₁The random attribute number is M, and the total attribute number of each sample is M.

1) Randomly selecting m attributes from all attributes of the current node of the decision tree to form an attribute subset;

2) calculating the possible value-taking kini index of each attribute in the attribute subset;

the kini index is defined as:

where D is the sample set, a is the attribute that divides the sample set D, and a has V possible values { a }¹，a²，...，a^vGet V sub-nodes after division, D^vI.e. the sample contained in the v-th child node, Gini (D)^v) Is a data set D^vThe value of (a).

The definition of the kini value is as follows:

where D is the sample set, p_k(k-1, 2, …, | y |) is the proportion of the kth sample in the sample set D.

3) Finding out the attribute with the minimum kini index for splitting;

4) and repeating the steps for the obtained child nodes, and freely growing until the division stops.

And (3) outputting: a trained CART decision tree.

3. And obtaining a final classification result by a majority voting method.

The majority voting method is one of integrated learning and combining strategies, and the final result is integrated and output by using the method, namely if the number of votes obtained by a certain type of mark is the maximum, the mark is the final classification result. And if a plurality of class marks all obtain the highest vote number, randomly selecting one of the class marks as a final classification result.

The parameter selection in the random forest comprises the following steps: 1) the number of decision trees in the forest is set as x; 2) and when the decision tree node is split, selecting the number of the features in the feature subset, and setting the number as y. The selection of random forest parameters is optimized by a grid search method. The selection steps of the grid search method for optimizing the random forest parameters are as follows:

1. for x and y, respectively setting a certain step length (set as a, b) to carry out iterative search, calculating classification precision through a random forest algorithm, and obtaining a parameter (x) with an optimal value for one-time search₁,y₁)；

2. For x₁And y₁At x₁And y₁Vicinity (x)₁-a，x₁+a),(y₁-b，y₁+ b), setting a certain step length (c, d) again for iterative search to obtain the optimal parameter (x) of the secondary search₂,y₂)；

3. For x₂And y₂Reducing the search range and the step length again to carry out iteration until the step length is set to be 1 to obtain the final search optimal parameter (x)₀,y₀)。

To this end, the number of decision trees in the forest is set to x₀When the decision tree node is split, the number of the features in the feature subset is selected to be y₀And optimizing parameter selection in the random forest classification model.

And step three, carrying out fault classification prediction on the test sample set by using a random forest model.

And classifying the test sample set of the turnout action data by using the random forest classification model trained in the step two, voting the results output by all decision trees by using a majority voting method, and obtaining the classification label with the maximum vote number which is the final classification result of the turnout fault, thereby realizing the fault diagnosis of the turnout.

The example adopts the electric power data of the periodic action of the S700K type turnout switch machine collected by the Guangzhou subway turnout experiment table for verification. The data of 14 states including normal, single fault (abnormal friction, close contact, foreign matter inclusion) and multiple faults are collected. The experimental data are shown in table 1.

Table 1 experimental data description

Numbering	Status type	Number of samples	Description of the state
				1	Normal	9	The plaster is 0
2	Nfault5	50	Close contact 0.5mm
				3	Nfault15	43	Close contact 1.5mm
4	Ffault0	223	Abnormal frictional force
				5	Ffault3	127	Abnormal friction force and close contact of 0.3mm
6	Ffault10	127	Abnormal friction force and close contact of 1.0mm
				7	Ffault13	104	Abnormal friction force and close contact of 1.3mm
8	Ffault20	109	Abnormal friction force and 2.0mm close contact
				9	Ffault25	61	Abnormal friction force and 2.5mm close contact
10	Ffault30	61	Abnormal friction force and close contact of 3.0mm
				11	Fforeign2	12	Abnormal frictional force + 2.0mm foreign matter
12	Fforeign4	10	Abnormal frictional force + foreign matter 4.0mm
				13	locked25	22	Friction force abnormity + sealing 2.5mm (unlocking fault)
14	locked30	4	Friction force abnormity + sealing 3.0mm (unlocking fault)

The data is used for detecting and verifying the turnout fault diagnosis method based on the random forest, and meanwhile, the effectiveness and the efficiency of the method are verified by comparing the operation results of different classification methods. The specific experiment is as follows:

there were 962 sets of data, each set having a characteristic number of 296. Firstly, a data set is randomly divided into a training set and a testing set.

1) And (4) diagnosing and analyzing turnout faults by using a random forest algorithm.

Applying a random forest algorithm, according to experience, setting random forest parameters as follows:

number of decision trees in random forest: 1,11, 21, …, 901;

selecting the number of features in the feature subset when each tree is split: less than log₂The largest integer of M +1 (M is the number of sample features, here 296).

Training sample data is used as input training random forest models, fault diagnosis and classification are carried out on the test samples, and the obtained results are shown in fig. 2 and table 2. In the graph, the horizontal axis is the number of decision tree classifiers, and the vertical axis is the fault classification accuracy of the random forest. It can be seen from the figure that when the number of the decision trees reaches 21, a better fault diagnosis recognition rate (87.99%) is obtained, and when the number of the decision trees exceeds 141, the fault classification precision is close to stable. When the number of the decision trees reaches 901, the decision trees still have stable classification precision, and overfitting cannot occur.

To better illustrate the fault diagnosis results of the proposed algorithm, we record the mean, maximum and model build times (at 901 decision trees) of classification accuracy after 21 decision trees as in table 2. Therefore, the random forest is used for diagnosing the turnout fault, and has high fault identification accuracy, good stability and high operation speed.

TABLE 2 Fault diagnosis and classification of random forest switches

2) And (4) selecting and optimizing random forest parameters.

One iteration:

setting the iteration step size of the decision tree number to be 20, namely, setting the tree _ num to be 1:20: 301; and setting the iteration step size of the feature number in the selected feature subset at the time of splitting each tree to be 5, namely feature _ num is 1:5: 296. And (4) utilizing a random forest algorithm to diagnose faults, and combining the first four parameters with the highest fault classification accuracy to perform the next iteration. The parameter combinations are respectively as follows: (181,6), (141,11), (161,16), (41,31), the classification accuracy is: 97.03 percent

And (3) secondary iteration:

the search range is: (max (x-20,1), x +20), (max (y-5,1), min (y1+5,296)), where x, y represent the currently taken parameter value; the iteration step size of tree _ num and feature _ num is 1. And (3) utilizing a random forest algorithm to carry out fault diagnosis, and obtaining the parameter combination with the highest fault classification accuracy in each search range as follows: (161, 8), (160, 17), (174, 19), (51, 36), the classification accuracy rates are respectively: 97.03%, 97.53%, 97.88% and 97.53%. Thus, an optimal set of parameters for the random forest is determined (174, 19), with an optimal classification accuracy of: 97.88 percent. Therefore, the random forest classification accuracy after parameter selection is greatly improved.

3) And carrying out fault diagnosis comparison analysis on the random forest algorithm, the CART decision tree and the Bagging algorithm.

The number of decision trees is set to 174 and the number of features in the split feature subset is set to 19. The model build times and classification accuracy for the different classifiers are shown in table 3.

TABLE 3 diagnosis and classification of turnout faults of different classifiers

Algorithm	Accuracy (%)	Model build time(s)
			Random forest	97.88	14.74
Decision tree Bagging	93.58	394.04
			CART decision tree	88.69	2.40

According to experimental results, the turnout fault diagnosis method based on the random forest is superior to decision trees and Bagging methods. The method can realize fault detection and classification of the turnout, has high fault accuracy and has obvious practical application value.

Claims

1. A turnout fault diagnosis method based on a random forest is characterized by comprising the following steps:

2. The turnout fault diagnosis method based on the random forest as claimed in claim 1, wherein the step of selecting the optimized random forest parameters by the grid search method in the second step is as follows:

2) For x₁And y₁At x₁And y₁Vicinity (x)₁-a，x₁+a),(y₁-b，y₁+ b), setting a certain step length (c, d) again for iterative search to obtain the optimal parameter (x) of the secondary iterative search₂，y₂)；

3) For x₂And y₂Reducing the search range and the step length again to carry out iteration until the step length is set to be 1 to obtain the final search optimal parameter (x)₀，y₀)；