A flight guarantee process residual execution time prediction method based on deep learning

Abstract

A flight support process residual execution time prediction method based on deep learning, the prediction model is composed of an input layer, a bidirectional LSTM layer, an Attention layer, a fusion layer and a prediction layer, the model input contains non-sequential global features and sequential execution state of each time step process. The Attention-BiLSTM model extracts the time sequence information of the process execution state, generates the intermediate state code, and is fused with the non-sequential global features, and then passes through the multilayer perceptron to obtain the residual execution time of the process. At the beginning of the process, a non-sequential global feature vector is generated, the process execution state is collected every fixed time, and a sliding window method is used to generate sequence feature data, and the feature data is calculated by the model to obtain the dynamic prediction value of the residual time of the flight support process.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of process mining and deep learning, and particularly relates to a flight support process residual execution time prediction method based on deep learning. BACKGROUND

[0002] The problem of predicting the residual time of a business process instance is as follows: given an event log containing multiple historical records of the execution of the business process instance, for an ongoing process instance, predict how much residual time there is from now to the end of the entire process. The residual time prediction of the business process is an important problem in business process monitoring, and can be used as an important means of business process optimization.

[0003] With the continuous growth of air transportation demand, high-density operation of airport flights has become the norm, and airports and airlines are facing huge operation pressure. The ground support process is the core part of the transit flight operation, and the dynamic prediction of its delay becomes a key problem for realizing fine management of flights. Excluding weather conditions and other uncontrollable reasons, whether the support process is completed on time has a direct impact on whether the flight is delayed, and the prediction of the residual execution time of the support process can enable the operation management personnel to dynamically grasp the flight support process and take relevant emergency measures in advance to reduce adverse effects.

[0004] Traditional residual execution time prediction methods, such as transition systems, stochastic Petri nets, and Bayesian networks, usually take activities in the business process as the unit, and predict the delay risk of the support process by mining the activity execution information from the event log of the support process. For an ongoing support process instance, only the trajectory prefix generated at the end of each activity can be used for prediction, and the real-time performance is poor.

[0005] Unlike general process prediction problems, each flight support process instance generally has the same trajectory, but serial and parallel activities coexist during process execution, and the coupling dependency relationship of each activity is relatively complex. Existing researches usually use sequential encoding or random initialization methods to process the input of the model, which cannot effectively utilize the mutual relationship between activities. On the other hand, existing process prediction methods usually only focus on the context of the activity and the information of the activity itself, and inevitably ignore the global features of the entire process, such as the executor, the executed, the time period, the urgency requirement, etc. For the flight ground support process, such feature selection is one-sided. SUMMARY

[0006] The purpose of the present application is to provide a flight support process residual execution time prediction method based on deep learning, which solves the problems of the existing prediction methods that cannot effectively utilize the features of the process itself, and low prediction accuracy and poor timeliness.

[0007] The technical solution proposed in the present invention includes data processing, model construction and training, and remaining time prediction.

[0008] Data processing stage:

[0009] An execution instance of the flight support process corresponds to a record in the flight support process event log, which records the names, start times, and end times of all activities. In addition, each execution instance also contains some global information relative to all activities in the process. This feature information is divided into numerical features and categorical features. The non-sequential global feature vector x of the process is g It is obtained by concatenating the vectors after one-hot encoding of the category features and normalization of the numerical features.

[0010] Assume that the total number of activities included in each flight support process is L, and the execution status of each activity in the process is extracted in units of α minutes. Then, at the t-th time step (the t-th time step corresponds to the αt-th minute of the process, and the beginning of the process is recorded as the 0th time step), the execution status information of each activity is x t =[s1,s2,...,s L ],s x ≥0,s x Represents the elapsed execution time of the xth activity in the process.

[0011] The above features x g and x t The normalization formula is:

[0012]

[0013] in, Represents x g or x t The normalized value of the j-th feature, x max,j and x min,j Represents x in the dataset g or x t The maximum and minimum values ​​of the j-th feature.

[0014] Each record gets the execution status information x of each activity t The process execution state sequence data X s =[x0,x1,...,x n ],x t ∈R L , where R L represents an L-dimensional vector in the real number field, where L is the total number of activities. y finishIt represents the execution time of the flight support process instance represented by this record, that is, the time difference between the end time of all activities and the start time of the first activity. α represents the time interval for feature extraction.

[0015] In order to solve the problem of feature data redundancy and inconsistent time step length, a sliding window method is used to obtain the process execution state sequence data X for each record. s Assume that the size of the selected window is K. For the tth (t≥K) time step of each record, the current time step and the previous K-1 time steps, a total of K execution state vectors, are used as a window, that is, X st =[x t-K+1 ,x t-K+2 ,...,x t ],x t ∈R L At each time step after the window is filled, delete the first state vector x in the window t-K+1 , and the current state vector x t Put it at the end. Each state sequence X st The corresponding target value is the remaining execution time of the process at the corresponding time step. The calculation method for the remaining execution time of the process at the t-th time step is:

[0016] y actual,t =y finish -αt,t=0,1,...,n

[0017] Model building and training phase:

[0018] The neural network model of the Attention-BiLSTM neural network model constructed by the present invention includes an input layer, a bidirectional LSTM layer, an Attention layer, a fusion layer, and a prediction layer.

[0019] Since each neuron in LSTM depends on the previous neuron, information is transmitted in one direction. Therefore, each neuron in this one-way LSTM can only learn the information before it, but not the information after it. Therefore, a bidirectional LSTM is used to model the process execution state, and the output state is obtained by simultaneous forward propagation. and backpropagation to obtain the output state To update the current state h t ,Right now:

[0020]

[0021] in and Represents the forward LSTM and reverse LSTM outputs, h t Represents the output of the bidirectional LSTM layer at the t-th time step.

[0022] The data of the input layer is the process execution state sequence X st =[x t-K+1 ,x t-K+2 ,...,x t ],x t ∈R L The bidirectional LSTM layer outputs the intermediate code, which serves as the input to the Attention layer. The model introduces an attention mechanism to capture the long-range interdependencies in the execution state representation sequence. The specific calculation method is as follows:

[0023]

[0024] where att t represents the weight coefficient of the tth time step, l represents the output of the Attention layer, and K is the selected window size.

[0025] The output of the Attention layer is used as the input of the fusion layer, combined with the non-sequential process global feature x g , use the concat function to splice into a one-dimensional vector to meet the input requirements of the prediction layer in the prediction layer, that is:

[0026] h * =concat(l,x g )=[l,x g ]

[0027] The prediction layer is a multi-layer perceptron with ReLU as the activation function. The prediction layer predicts the remaining execution time, and the output of the prediction layer is:

[0028] predictTime=MLP(h * )

[0029] Each record in the flight support process event log is processed according to the method described above to obtain a dataset, which is divided into a training set, a validation set, and a test set.

[0030] The parameters of the neural network model are trained by the time series back propagation algorithm, and the mean square error function is used as the loss function of the model:

[0031]

[0032] Among them, m is the total number of samples, y predict,i is the predicted value of the remaining execution time of the process of the i-th sample, y actual,i is the true value of the remaining execution time of the process of the i-th sample.

[0033] The weights and biases of each neuron in the network model are continuously updated iteratively to converge the loss function. To prevent overfitting of the trained neural network model, hyperparameters such as the learning rate and number of iterations are adjusted based on the prediction results of the validation set data before training is continued until the model with the best prediction results is obtained.

[0034] Remaining time prediction stage:

[0035] In order to dynamically monitor the execution of flight processes and ensure the timeliness of process prediction, a global feature x is generated at the beginning of each process. g , collect the execution status of each activity in the guarantee process every K minutes, and calculate the execution status feature vector x of the time step t When the K-th time step window is filled, the remaining execution time of the process is predicted, and all the process execution state vectors in the window are taken to generate the sequence feature X st . st and x g The data is input into the input layer and fusion layer of the Attention-BiLSTM neural network model respectively, and the predicted value of the remaining time of the process at the current time point is obtained through the forward propagation algorithm.

[0036] Compared with the existing invention, the beneficial effects of the present invention are:

[0037] (1) For processes with complex activity coupling relationships, such as flight support, simply representing individual activities cannot effectively utilize the relationships between them. This invention uses a method to collect the execution status of each step at fixed intervals and generates sample data using a sliding window approach, thus extracting sequence features simply and effectively.

[0038] (2) In addition to using sequential features such as the flight support process execution status, the present invention also combines the non-sequential global features of the process itself to extract features more comprehensively. Comprehensive and high-quality features can improve the accuracy of the remaining execution time prediction.

[0039] (3) The present invention adopts a neural network model based on the attention mechanism and the bidirectional long short-term memory network, which has a high sequence feature extraction capability and can effectively capture the reverse and long-distance interdependencies in the sequence. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a structural diagram of the Attention-BiLSTM neural network model in an embodiment of the present invention.

[0041] Figure 2 It is a schematic diagram of a flow chart in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0043] See Figure 1 , Figure 2 The present invention provides a method for predicting the remaining execution time of a flight support process based on deep learning. The detailed steps are as follows:

[0044] Step 1: Process the flight support process log to generate a data set.

[0045] For an execution instance of the flight support process, six types of data are taken as the global features of the support process: aircraft type, flight type, stand type, number of flights in the support process, and airline. The categorical features of aircraft type, flight type, stand type, weather conditions, and airline are one-hot encoded, and the numerical feature of the number of flights in the support process is normalized. The non-sequential global feature vector x of the process is obtained by splicing. g .

[0046] The time step interval α is set to 1 minute. The execution status information of 13 activities, including block loading, jet bridge docking, passenger disembarkation, refueling, cleaning, water filling, catering, boarding, cargo and mail unloading, cargo and mail loading, jet bridge evacuation, and block removal, is extracted from the process event log. The elapsed time of each activity is calculated. If the activity has ended, the elapsed time is the difference between the activity completion time and the start time; if the activity has not started, the elapsed time is 0. The execution status information of each time step is obtained as x t =[s1,s2,...,s 13 ], for each time step x t According to the above method, the execution status sequence data X is generated for each record. s =[x0,x1,...,x n ].

[0047] The time window K is set to 15. In order to ensure that there is enough data for prediction, starting from the 14th time step (the 14th time step of the process, the process starts at the 0th time step), the execution state sequence X of each record is calculated. s Extract the execution state sequence X in the window corresponding to each time step st =[x t-14 ,x t-13 ,...,x t ], and calculate the remaining execution time y corresponding to this time step actual,t =y finish-αt,t=0,1,...,n. If the process execution time is 60 minutes, then the remaining execution time corresponding to the 14th time step is 46 minutes. st , the global feature x of the corresponding process instance g and the target value remaining execution time y actual,t Assemble into a sample data.

[0048] All flight support process execution instances are processed according to the above method to obtain the entire data set for subsequent model training.

[0049] Step 2: Build Figure 1 The Attention-LSTM neural network shown in the figure includes an input layer, a bidirectional LSTM layer, an Attention layer, a fusion layer, and a prediction layer. st As the input of the input layer, the high-level features of the sequence data are extracted through the bidirectional LSTM layer and the Attention layer, and the non-sequence feature vector x g The high-level features output by the fusion layer and the attention layer are combined, and the multi-layer perceptron in the prediction layer predicts the remaining execution time.

[0050] Step 3: Train the prediction model. All data processed in Step 1 is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1. The constructed Attention-BiLSTM neural network model is iteratively trained using the training set data, using the Adam optimizer to achieve convergence of the loss function. Training is terminated when the model's loss on the training set data and the loss on the validation set data become constant. The hyperparameters and structure of the prediction model are adjusted until the optimal prediction model is obtained. Finally, the standard mean square error (NMSE) between the actual values ​​of the test set and the model's predictions is used to evaluate the model's final prediction performance.

[0051] Step 4: At the beginning of the flight support process, the trained prediction model is initialized, and the global feature vector x that needs to be predicted is calculated according to the method in step 1. g Perform calculations.

[0052] Step 5: From the beginning of the process, that is, at the 0th minute of the process, collect the execution status of each activity in the guarantee process every 1 minute, and calculate the characteristic vector x of the execution status of the time step t Starting from the 14th time step (when 15 sequence data have been extracted), the remaining execution time of the process is predicted, and X is generated in the same way as in step 1. st , X stAnd the global feature x obtained in step 4 g This information is fed into the input and fusion layers of the Attention-BiLSTM neural network model. The prediction layer then generates a predicted value for the remaining time of the current process. Repeat step 5 until the flight support process requiring prediction is complete.

[0053] A simulation experiment was conducted to test the above method. The dataset used was real flight support process data from a major domestic airport from January to November 2021. After cleaning the abnormal data and processing it using the method in Step 1, 201,103 data items were obtained. The NMSE between the actual and predicted values ​​of the test set was used as a metric to evaluate the model's predictive performance. The smaller the NMSE, the better the predictive performance. The simulation experiments used Python version 3.6.5, TensorFlow version 2.4.0, and Keras version 2.4.3.

[0054] To verify the effectiveness of the Attention-BiLSTM neural network model, the prediction performance of the Attention-BiLSTM model was compared with that of LSTM, Attention-LSTM, and Attention-BiRNN. Table 1 shows the average NMSE of the four models from 10 experiments on the dataset. These four models do not incorporate global features.

[0055]

[0056] Table 1

[0057] The LSTM, Attention-LSTM, and Attention-BiLSTM models all have good prediction effects on the dataset. Among them, the Attention-BiLSTM model has the lowest error between the predicted value and the true value on the test set.

[0058] To verify the effectiveness of the Attention-BiLSTM neural network model in fusing global features, Table 2 statistically compares the average NMSE of 10 experiments conducted on the dataset with and without fusing global features.

[0059] Model Fusion of global features No integration of global features NMSE 0.00087033 0.00125833

[0060] Table 2

[0061] Experimental results show that fusing global features can significantly reduce the prediction error of the Attention-BiLSTM model on the remaining execution time of the process on the dataset.

[0062] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed in the present invention should be included in the protection scope recorded in the claims.

Claims

1. A method for predicting the remaining execution time of a flight support process based on deep learning, characterized by: The following steps are involved: Step 1: Process the flight support process event log, extract the sequence features and non-sequence features of each record, and generate a data set using the sliding window method; In step 1, the event logs for the support process are processed. Specifically, for each execution instance, the global features of the support process are taken as the aircraft type, slot type, number of flights currently being supported, weather conditions, and airline. The categorical features are one-hot encoded, the numerical features are normalized, and the two are concatenated to obtain the non-sequential global feature vector x of the process g ; The execution time of each activity in each record in the process event log is calculated and normalized in units of α minutes to obtain the execution status x at each time step. t =[s1,s2,...,s L ]; extract the execution state sequence X with the window value K st =[x t-K+1 ,x t-K+2 ,...,x t ],x t ∈R L ; Calculate the remaining execution time y corresponding to each time step of the process actual,t =y finish -αt,t=0,1,...,n;the state sequence X st , global characteristics x of the process instance g and the target value remaining execution time y actual,t As a sample data; Step 2: Build the Attention-BiLSTM neural network model; In step 2, the Attention-BiLSTM neural network model includes the input layer, bidirectional LSTM layer, attention layer, fusion layer, and prediction layer; the process execution state sequence X st As the input of the input layer, the high-level features of the sequence data and the global features x are extracted through the bidirectional LSTM layer and the Attention layer. g The fusion layer uses the concat function to connect it into a one-dimensional vector, and then the multi-layer perceptron of the prediction layer calculates the remaining execution time y actual,t Make predictions; Step 3: Train the constructed neural network model; Step 4: Initialize the prediction model at the beginning of the flight support process that needs to be predicted, and calculate the global feature vector of the process; Step 5: Collect the execution status of each activity in the assurance process every α minutes, calculate the elapsed execution time of each activity at that moment, and use the Attention-BiLSTM neural network model to predict the remaining execution time until the entire assurance process is completed.

2. The method for predicting the remaining execution time of a flight support process based on deep learning according to claim 1, characterized in that: Each record in the process event log includes 12 activities, including loading chocks, berthing at the jet bridge, disembarking passengers, refueling, cleaning, water filling, catering, boarding, unloading cargo and mail, loading cargo and mail, jet bridge evacuation and unloading chocks.

3. The method for predicting the remaining execution time of a flight support process based on deep learning according to claim 1, characterized in that: In step three, all the data processed in step one are randomly divided into training set, validation set, and test set in a ratio of 8:1:

1. The constructed Attention-BiLSTM neural network model is iteratively trained using the training set data, and the Adam optimizer is used to converge the loss function. The learning rate and number of iterations of the prediction model are adjusted using the validation set data to obtain the prediction model. The final effect of the model is evaluated using the test set.

4. The method for predicting the remaining execution time of a flight support process based on deep learning according to claim 1, characterized in that: In step 5, the dynamic prediction method of the remaining time of the process based on Attention-BiLSTM is as follows: when dynamically predicting the remaining execution time of an executing process, the execution status of each activity in the process is collected every α minutes, and the execution state sequence X is generated in a sliding window manner. st , and the global feature x g Input the Attention-BiLSTM neural network model to obtain the predicted value of the remaining time of the process at the current moment.

Images