1. A method for predicting equipment maintenance time interval based on deep learning comprises the steps of predicting equipment maintenance time interval by establishing an equipment maintenance time prediction model RCNN of a cyclic neural network (RNN) and Convolutional Neural Network (CNN) mixed structure to obtain accurate equipment maintenance time; the method comprises the following steps:

1) extracting maintenance-related attribute information for a type of equipment from equipment repair business data, comprising: equipment maintenance service attribute data and geographic environment factor data of the equipment location corresponding to each piece of maintenance information;

2) sorting the extracted data according to maintenance time;

then, sequentially sliding according to the length and the step length of a fixed window to form a data matrix consisting of a plurality of maintenance information sequences;

simultaneously extracting geographic environment factors in the attribute information of the last equipment maintenance service in the current window, and forming a training sample by the geographic environment factors; then the next repair time interval is used as a label; thereby obtaining a plurality of training sample data and labels;

the obtained training sample data can be randomly divided into a training set and a test set according to a proportion;

3) establishing an equipment maintenance opportunity prediction model RCNN model of a hybrid structure of a Recurrent Neural Network (RNN) and a Convolutional Neural Network (CNN);

the RCNN model comprises a CNN network model and an RNN network model: the CNN network model is used for extracting the characteristics of the geographic environment factors where the equipment is located; the RNN model comprises 2 hidden layers, more information is stored through the multiple hidden layers, and the RNN model is used for learning and training the subsequent maintenance service attribute data;

4) training and verifying the established equipment maintenance opportunity prediction RCNN model; the method comprises the following steps:

correspondingly inputting all training samples and labels into an equipment maintenance opportunity prediction model RCNN network model, putting the training samples at the input end of the network, putting the labels at the output end of the network, and training the model to obtain a trained equipment maintenance opportunity prediction model;

further, the trained equipment maintenance opportunity prediction model can be tested and verified by using the test set data;

5) inputting the data of equipment repair to be predicted into the trained equipment repair opportunity prediction model, and outputting a predicted value, namely realizing the prediction of the equipment repair time interval based on deep learning.

2. The method for deep learning-based equipment repair interval prediction as claimed in claim 1, wherein the equipment is armored equipment.

3. The method for equipment repair time interval prediction based on deep learning of claim 1, wherein the equipment repair service attributes extracted from the equipment repair service data of step 1) comprise: single package number a₁Equipment type a₂Minor repair frequency a₃Number of repairs a₄Number of major repairs a₅Service time a₆And total consumption of motorcycle hours a₇Failure ofType a₈This maintenance time a₉Time interval a from last minor repair₁₀(ii) a The geographic environment factors of the equipment location corresponding to each piece of maintenance information comprise: altitude b₁Monthly average gas pressure b₂Average monthly air temperature b₃Monthly average relative humidity b₄Monthly rainfall b₅Monthly evaporation capacity b₆Monthly average wind speed b₇Average earth temperature b of the month₈And the number of days of the month b₉Average water vapor pressure b₁₀Percent oxygen content in the atmosphere b₁₁。

4. The method for predicting the equipment maintenance time interval based on the deep learning as claimed in claim 1, wherein in the step 2), the length of the fixed window is 5; the step length is 1; and sequentially sliding from top to bottom to form a data matrix consisting of a plurality of 5 multiplied by 10 maintenance information sequences.

5. The method for predicting the repair time interval of equipment based on deep learning of claim 1, wherein in the step 2), the obtained training sample data is randomly divided into a training set and a test set according to a ratio of 4: 1.

Background

In the actual maintenance guarantee of the equipment, the prior art still mainly uses a unified equipment maintenance interval standard to make an equipment repair plan. In the application of actual maintenance and guarantee, the maintenance work is difficult to be carried out aiming at the actual environment and condition of the equipment to be maintained and guaranteed, so that the equipment is not accurately repaired, and the equipment is low in use efficiency.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention provides a method for predicting the equipment maintenance time interval based on deep learning, which predicts the equipment maintenance time interval to obtain accurate equipment maintenance time and can be used for making an equipment maintenance plan.

The principle of the invention is as follows: by utilizing a deep learning method, a mixed structure model of a Recurrent Neural Network (RNN) and a Convolutional Neural Network (CNN), which is called RCNN (Recurrent and Convolutional Neural Network) model for short, is provided, and equipment repair time intervals are predicted.

The RNN is a neural network for processing time sequence data, and the RNN is selected to perform learning training on maintenance process data of the equipment, and the relevance between the front and the back of the maintenance data of the equipment is mined; meanwhile, in order to independently learn the potential relationship between the maintenance time interval and the geographic environment factors, the CNN is utilized to extract the characteristics of the geographic environment where the equipment is located. And then, combining the two models to establish an equipment maintenance opportunity prediction model with a mixed structure, and analyzing the change characteristics of the equipment maintenance time interval under the action of a training plan and a geographic environment to realize the prediction of the equipment maintenance opportunity.

The technical scheme provided by the invention is as follows:

a method for predicting equipment maintenance time interval based on deep learning comprises the steps of predicting equipment maintenance time interval by establishing an equipment maintenance time prediction model of a cyclic neural network (RNN) and Convolutional Neural Network (CNN) mixed structure to obtain accurate equipment maintenance time; the method comprises the following steps:

1) extracting maintenance-related attribute information for a type of equipment from equipment repair business data, comprising: equipment maintenance service attributes (including single-package serial numbers, equipment models, minor repair times, middle repair times, major repair times and the like), and geographic environment factor attributes of the equipment location corresponding to each piece of maintenance information;

2) sequencing the extracted data according to maintenance time, then sliding from top to bottom by a fixed window length (for example, the value is 5, and the average minor repair frequency of equipment in 2 years can be referred), and a step length (for example, 1) to form a data matrix formed by a plurality of 5 x 10 maintenance information sequences, and simultaneously extracting the geographic environment attribute in the last piece of equipment maintenance information in the current window, and combining the two to form a training sample; in addition, the next repair time interval is used as a label to obtain a plurality of training sample data and labels; the obtained training sample data can be randomly divided into a training set and a test set according to a proportion (such as a proportion of 4: 1);

3) establishing an equipment maintenance opportunity prediction model of a cyclic neural network (RNN) and Convolutional Neural Network (CNN) mixed structure, which is called as an RCNN model for short;

the RCNN model is shown in fig. 2 and includes a CNN network model and an RNN network model. Where the RNN structure includes 2 hidden layers, it is considered that more important information can be preserved through a plurality of hidden layers when the amount of information is too large. The CNN model is mainly used for extracting the characteristics of the geographic environment factors where the equipment is located. The RNN model is mainly used for learning and training basic information data of equipment and service data of equipment management and guarantee, and a correction linear unit ReLu is used as an activation function to relieve the overfitting problem. And defining the difference between the fitting value obtained by sample data training and the label as a residual error.

4) Training and verifying the established equipment maintenance opportunity prediction RCNN model;

correspondingly inputting all training samples and labels into an equipment maintenance opportunity prediction model RCNN network model (the training samples are placed at the input end of the network, and the labels are placed at the output end of the network), and training the model to obtain a trained equipment maintenance opportunity prediction model; and further testing and verifying the trained equipment maintenance opportunity prediction model by using the test set data.

5) Inputting the data of equipment repair to be predicted into the trained equipment repair opportunity prediction model, and outputting a predicted value, namely realizing equipment repair time interval prediction based on deep learning.

Compared with the prior art, the invention has the beneficial effects that:

the RNN structure in the equipment maintenance opportunity prediction RCNN model established by the method has the advantages of processing equipment maintenance data with time series characteristics, and can fully mine the influence of the historical maintenance condition of the equipment on the maintenance time interval; meanwhile, the CNN is used for analyzing the relationship between the factors of the geographic environment where the equipment is located and the maintenance interval of the equipment, the influence on the equipment caused by the difference of the environments of the area where the equipment is located is deeply excavated, and the obtained prediction result is closer to a true value. With the lapse of time, equipment maintenance information accumulates gradually, and the RCNN can independently learn more characteristics, can further improve prediction accuracy.

Drawings

FIG. 1 is an example of the composition of training samples of the predictive model RCNN established by the present invention.

FIG. 2 is a model structure of the prediction model RCNN established by the present invention;

FIG. 3 is a block flow diagram of a prediction method provided by the present invention.

Detailed Description

The invention will be further described by way of examples, without in any way limiting the scope of the invention, with reference to the accompanying drawings.

The invention provides a method for predicting equipment maintenance time interval based on deep learning, which predicts the equipment maintenance time interval by establishing an equipment maintenance time prediction model of a cyclic neural network (RNN) and Convolutional Neural Network (CNN) mixed structure to obtain accurate equipment maintenance time. Fig. 3 is a flowchart of the prediction method provided by the present invention, specifically implementing equipment maintenance time interval prediction based on deep learning for armor equipment, including the following steps:

the data is preprocessed, for example, by prediction of the armored equipment overhaul time interval. Extracting the individual serial number (a) from the armored equipment repair service information data₁) Equipment model (a)₂) Minor repair frequency (a)₃) The number of repair operations (a)₄) Number of major repairs (a)₅) Service time (a)₆) Total consumption of motorcycle hours (a)₇) Type of failure (a)₈) The maintenance time (a)₉) Interval (motocycle hours) from last minor repair (a)₁₀) Waiting for equipment maintenance business attribute and the altitude (m) (b) of the equipment location corresponding to each piece of maintenance information₁) Average qi in the moonPressure (Kpa) (b)₂) Monthly average air temperature (. degree. C.) (b)₃) Monthly average relative humidity (%) (b)₄) Monthly rainfall (mm) (b)₅) Monthly evaporation capacity (cm) (b)₆) Monthly mean wind speed (m/s) (b)₇) Monthly average earth temperature (. degree. C.) (b)₈) The number of days of the month (h) (b)₉) Monthly mean vapor pressure (hPa) (b)₁₀) Percent atmospheric oxygen content (%) (b)₁₁) And the attributes of the geographic environment factors are obtained, and the obtained data sample is shown in table 1.

Table 1 data sample table

And establishing an equipment maintenance opportunity prediction model. In the equipment maintenance timing prediction model, in order to correct the time series characteristics of the sample data, the attribute (a) in the same equipment maintenance information is used₁～a₁₀) Extracting and sorting according to maintenance time, sliding from top to bottom with fixed window length of 5 (referring to average minor repair times of 2 years of equipment) and step length of 1 to form a data matrix (shown as a frame in figure 1) formed by a plurality of maintenance information sequences of 5 multiplied by 10, and simultaneously extracting the geographic environment attribute (b) in the last piece of equipment maintenance information in the current window₁～b₁₁) (as shown in the box of fig. 1), the two are combined to form a training sample; the next repair interval is used as a label (as shown by the third box in fig. 1). In the training process, all samples and labels are put into the network one to one (the training samples are put at the input end of the network, the labels are put at the output end of the network), residual errors are defined according to the difference between fitting values obtained by sample data training and the labels, and the residual errors are advanced in the direction of reducing by continuously and automatically updating weights in the training process.

The total number of the equipment maintenance data collected in the embodiment is 72627, wherein 34278 minor repairs can obtain 11788 samples (minus the equipment with historical maintenance information less than 5 and part of maintenance information which is not utilized due to the construction of a time series data matrix) according to the above mode, sample data is randomly divided into a training set and a testing set according to the proportion of 4:1 and is input into an RCNN model, the data in the frame of each sample is input into an RNN, and the data in the frame is input into a CNN.

In order to verify the effectiveness of the RCNN model, a 59-type medium tank is taken as an object, the minor repair and repair time interval of the tank is predicted by using an exponential smoothing method, a stepwise regression method and the RCNN, and the experimental results are compared. The experimental procedures were all conducted on a Tensorflow 1.4.1 under the Ubuntu 16.10 system, the processor wasCore^TMi7-7820CPU @3.40GHz x 8, and the GPU is NVIDIAGTX 1080。

(1) Exponential smoothing method

The exponential smoothing method uses the past weighted mean of the time series to predict the future value, so that the predicted value can quickly reflect the actual change. The weight of each stage is respectively alpha, alpha (1-alpha)²…, the importance of the data decreases in steps with time. Alpha (alpha is more than 0 and less than 1) is a smooth coefficient, if the trend of the time series is more stable, a smaller alpha value is selected, and if the fluctuation is larger, the alpha value is larger, so that the influence of recent data is increased.

And performing exponential smoothing on the minor repair time interval of the 59-type medium-sized tank by using a Markov analysis system 5.0. And (4) trial calculation is carried out on the smoothing coefficient alpha from 0-1 step length to 0.1 in a network searching mode, and prediction standard errors under different alpha values are compared. The results show that when α is 0.16, R²0.825, the sum of the standard error and the squared residual error is minimal.

(2) Stepwise regression method

The regression equation obtained by stepwise regression is as follows:

wherein the content of the first and second substances,the prediction of the repair time interval of the 59 type medium tank under various geographic environment factors is carried out, and the time interval from the next repair under the specific geographic environment condition is fitted through a regression equation.

In addition to comparing the MSEs of the results predicted by the RCNN, exponential smoothing, and stepwise regression methods, the Mean Absolute Error (MAE) and Mean Relative Error (MRE) were compared, and the formula is as follows:

the MAE is the average value of absolute errors, and compared with average errors, the MAE is absolute value-converted due to dispersion, so that the situation that the positive and negative of errors are offset does not occur, and the actual situation of predicted value errors can be better reflected by the MAE. MRE is the average of the relative errors and better reflects the confidence level of the measurement. The values of the indices are shown in table 2:

TABLE 2 comparison of predicted results for three methods

As can be seen from table 2, RCNN is superior to the exponential smoothing method and the stepwise regression method in prediction accuracy, and the reasons are mainly: the exponential smoothing method can predict the value of the next period without excessive data and is more suitable for short-term prediction of a time sequence, the interdependence relationship among equipment maintenance historical data possibly spans a longer time length, the exponential smoothing can weaken the interdependence relationship among the data, and the obtained predicted value cannot well reflect the change of the maintenance time interval trend; the stepwise regression method establishes the linear relation among the variables, is convenient for analysis and can obtain better fitting effect, but simultaneously possibly neglects the interaction effect and the nonlinear relation among the variables; the RNN structure in the RCNN model has advantages in processing equipment maintenance data with time series characteristics, and can fully mine the influence of equipment historical maintenance conditions on maintenance time intervals; meanwhile, the CNN is used for analyzing the relationship between the factors of the geographic environment where the equipment is located and the maintenance interval of the equipment, the influence on the equipment caused by the difference of the environments of the area where the equipment is located is deeply excavated, and the obtained prediction result is closer to a true value. With the lapse of time, equipment maintenance information accumulates gradually, and the RCNN can independently learn more characteristics, can further improve prediction accuracy.

It is noted that the disclosed embodiments are intended to aid in further understanding of the invention, but those skilled in the art will appreciate that: various substitutions and modifications are possible without departing from the spirit and scope of the invention and appended claims. Therefore, the invention should not be limited to the embodiments disclosed, but the scope of the invention is defined by the appended claims.