Deep learning-based airway error analysis method, device, equipment, medium and product

Abstract

The invention discloses an air route error analysis method and device based on deep learning, equipment, a medium and a product, and the method comprises the steps: carrying out the feature classification and feature normalization of an original data set according to a timestamp detection model, and obtaining a time sequence feature set and a non-time sequence feature set; wherein the original data set comprises a plurality of features of an air route; performing feature extraction and fusion on the time sequence feature set and the non-time sequence feature set according to a feature extraction fusion network to obtain a fusion feature matrix; according to a preset error prediction model and the fusion feature matrix, analyzing and quantifying contribution of each feature to an error and an association relationship between the features; wherein the error prediction model is obtained by training the fusion feature matrix through a hybrid model. According to the method, the contribution of the features to the error and the incidence relation between the features are analyzed and quantified, the source of the air route error can be comprehensively analyzed, and data-driven decision support is provided for planning and optimization of the air route.

Description

Technical Field

[0001] The present invention relates to the field of aviation technology, and particularly to a method, device, equipment, medium and product for analyzing route errors based on deep learning. Background Art

[0002] With the rapid development of the aviation industry, the prediction and analysis of route errors have become particularly important. Route errors not only directly affect the safety and efficiency of flight, but are also closely related to many factors, including environmental factors such as meteorological conditions, flight altitude, wind speed, and unpredictable events such as sudden traffic control and flight scheduling. Traditional route error analysis methods usually can only obtain the overall effect of influencing factors on errors, and cannot understand the contribution of different influencing factors to errors. In addition, traditional error analysis methods regard different influencing factors as independent existences, often ignoring complex interaction effects and correlations between features, resulting in inaccurate and insufficient comprehensiveness of error analysis. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method, device, equipment, medium and product for analyzing route errors based on deep learning, which can analyze and quantify the contribution of features to errors and the correlation relationship between features, and can comprehensively analyze the sources of route errors, providing data-driven decision support for route planning and optimization.

[0004] To achieve the above object, an embodiment of the present invention provides a method for analyzing route errors based on deep learning, including: Performing feature classification and feature normalization on the original data set according to a timestamp detection model to obtain a time series feature set and a non-time series feature set; wherein, the original data set includes several features of a route; Performing feature extraction and fusion on the time series feature set and the non-time series feature set according to a feature extraction and fusion network to obtain a fusion feature matrix; Analyzing and quantifying the contribution of each feature to errors and the correlation relationship between features according to a preset error prediction model and the fusion feature matrix; wherein, the error prediction model is obtained by training the fusion feature matrix with a hybrid model.

[0005] As an improvement of the above solution, the performing feature classification and feature normalization on the original data set according to a timestamp detection model to obtain a time series feature set and a non-time series feature set includes: Judging whether each feature in the original data set contains a time field; If not, the feature is a non-time series feature; if so, extracting the timestamp and detecting whether the timestamp is increasing; If not, the feature is a non-time series feature; if so, the feature is a candidate time series feature; Perform dynamic variability detection on the candidate time series features and calculate the change rate; If the variance of the change rate is greater than a preset threshold, the candidate time series feature is a time series feature; if the variance of the change rate is not greater than the preset threshold, the candidate time series feature is a non-time series feature; Perform normalization processing on the time series features and the non-time series features to obtain the normalized time series feature set and non-time series feature set.

[0006] As an improvement to the above solution, the feature extraction and fusion of the time series feature set and the non-time series feature set by the feature extraction fusion network to obtain a fusion feature matrix includes: Input the time series feature set into a convolutional neural network for feature extraction and output a time series feature matrix; Input the non-time series feature set into an autoencoder for feature extraction and output a non-time series feature matrix; Concatenate the time series feature matrix and the non-time series feature matrix by rows to form a high-dimensional feature matrix; Input the high-dimensional feature matrix into a fully connected layer for dimensionality reduction and output the fusion feature matrix.

[0007] As an improvement to the above solution, the hybrid model includes a multi-branch LSTM model and an XGBoost model; The multi-branch LSTM model includes a systematic error branch, a random error branch, and a merging layer; The systematic error branch is used to process the predictable features in the time series feature matrix in the fusion feature matrix and predict the systematic error; The random error branch is used to process the unpredictable features in the time series feature matrix in the fusion feature matrix and predict the random error; The merging layer is used to concatenate the outputs of the two branches and pass through a fully connected layer to predict the total error of the LSTM model; The XGBoost model is used to predict the systematic error, random error, and total error of the non-time series feature matrix in the fusion feature matrix.

[0008] As an improvement to the above solution, the analysis and quantification of the contribution of each feature to the error and the correlation relationship between features according to the preset error prediction model and the fusion feature matrix includes: Extract the feature vectors in the fusion feature matrix and calculate the SHAP value of each feature in the feature vector according to the prediction error of the error prediction model to obtain a set of SHAP values; Normalize the SHAP value set and calculate the contribution weight of each feature to obtain a feature contribution weight set; Calculate the correlation between features in the feature vector to obtain the association relationship between features.

[0009] As an improvement to the above solution, the calculation formula for the correlation includes: Pearson correlation coefficient , representing the linear relationship between features; Spearman correlation coefficient , representing the monotonic relationship between features; Mutual information , representing the non-linear relationship between features; Among them, f represents a feature; a and b both represent feature indices; represents the a th eigenvalue of the feature; represents the b th eigenvalue of the feature; Corr represents the Pearson correlation coefficient calculation function; represents the eigenvalue and covariance; and respectively represent the standard deviations of the eigenvalues and ; represents the rank obtained after sorting the eigenvalue x ; represents the joint probability distribution; and respectively represent their respective marginal probability distributions.

[0010] An embodiment of the present invention also provides a deep learning-based route error analysis device, including: A feature classification module, configured to perform feature classification and feature normalization on the original data set according to a timestamp detection model to obtain a time series feature set and a non-time series feature set; wherein, the original data set includes several features of the route; A feature fusion module, configured to perform feature extraction and fusion on the time series feature set and the non-time series feature set according to a feature extraction and fusion network to obtain a fusion feature matrix; An error analysis module, configured to analyze and quantify the contribution of each feature to the error and the association relationship between features according to a preset error prediction model and the fusion feature matrix; wherein, the error prediction model is obtained by training the fusion feature matrix with a hybrid model.

[0011] An embodiment of the present invention further provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the method for analyzing route error based on deep learning described in any one of the above is implemented.

[0012] An embodiment of the present invention further provides a computer-readable storage medium. The computer-readable storage medium includes a stored computer program. When the computer program runs, it controls the device where the computer-readable storage medium is located to execute the method for analyzing route error based on deep learning described in any one of the above.

[0013] An embodiment of the present invention further provides a computer program product. The computer program product includes a computer program or computer instructions. When the computer program or the computer instructions are executed by a processor, the method for analyzing route error based on deep learning described in any one of the above is implemented.

[0014] Compared with the prior art, the beneficial effects of a method, device, equipment, medium, and product for analyzing route error based on deep learning provided by an embodiment of the present invention are as follows: By performing feature classification and feature normalization on the original data set according to the timestamp detection model, a time-series feature set and a non-time-series feature set are obtained; wherein, the original data set includes several features of the route; According to the feature extraction and fusion network, feature extraction and fusion are performed on the time-series feature set and the non-time-series feature set to obtain a fusion feature matrix; According to the preset error prediction model and the fusion feature matrix, the contribution of each feature to the error and the correlation relationship between features are analyzed and quantified; wherein, the error prediction model is obtained by training the fusion feature matrix with a hybrid model. By analyzing and quantifying the contribution of features to the error and the correlation relationship between features, the embodiment of the present invention can comprehensively analyze the sources of route errors and provide data-driven decision support for the planning and optimization of routes. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic flowchart of a preferred embodiment of a method for analyzing route error based on deep learning provided by the present invention; Figure 2 is a schematic structural diagram of a preferred embodiment of a device for analyzing route error based on deep learning provided by the present invention; Figure 3 is a schematic structural diagram of a preferred embodiment of a terminal device provided by the present invention. DETAILED DESCRIPTION

[0016] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0017] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of a preferred embodiment of a method for analyzing route errors based on deep learning provided by the present invention. The method for analyzing route errors based on deep learning includes: S1. Classify and normalize the features of the original data set according to the timestamp detection model to obtain a time series feature set and a non-time series feature set; wherein, the original data set includes several features of the route. S2. Extract and fuse the features of the time series feature set and the non-time series feature set according to the feature extraction and fusion network to obtain a fused feature matrix. S3. Analyze and quantify the contribution of each feature to the error and the correlation relationship between features according to the preset error prediction model and the fused feature matrix; wherein, the error prediction model is obtained by training the fused feature matrix with a hybrid model.

[0018] Specifically, the embodiments of the present invention provide a method for analyzing route errors based on deep learning. First, classify and normalize the features of the original data set according to the timestamp detection model to obtain a time series feature set and a non-time series feature set. Wherein, the original data set at least includes the deviation between the planned route point and the actual route point, the deviation between the planned passing time and the actual passing time of the route point, the deviation between the actual flight speed and the planned flight speed, the deviation between the planned flight altitude and the actual flight altitude of the route point, the deviation between the predicted wind direction and wind speed and the actual wind direction and wind speed of the route point, the deviation between the predicted temperature and the actual temperature of the route point, the differences between the planned airspace restrictions, traffic flow and other conditions and the actual situation, the differences between the theoretical state and the actual state of the aircraft, and temporary or unexpected events. Then, extract the features of the time series feature set and the non-time series feature set respectively according to the feature extraction and fusion network, and fuse the extracted features to obtain a fused feature matrix. Finally, analyze and quantify the contribution of each feature to the error and the correlation relationship between features according to the preset error prediction model and the fused feature matrix. Wherein, the error prediction model is obtained by training the fused feature matrix with a hybrid model.

[0019] By analyzing and quantifying the contribution of features to the error and the correlation relationship between features, the embodiments of the present invention can comprehensively analyze the sources of route errors and provide data-driven decision support for route planning and optimization.

[0020] In another preferred embodiment, in step S1, the original data set is subjected to feature classification and feature normalization according to the timestamp detection model to obtain a time series feature set and a non-time series feature set, including: S101, determining whether each feature in the original data set contains a time field; S102, if not, then the feature is a non-time series feature; if so, extracting the timestamp and detecting whether the timestamp is increasing; S103, if not, then the feature is a non-time series feature; if so, the feature is a candidate time series feature; S104, performing dynamic variability detection on the candidate time series feature and calculating the change rate; S105, if the variance of the change rate is greater than a preset threshold, then the candidate time series feature is a time series feature; if the variance of the change rate is not greater than the preset threshold, then the candidate time series feature is a non-time series feature; S106, performing normalization processing on the time series feature and the non-time series feature to obtain the normalized time series feature set and the non-time series feature set.

[0021] Specifically, when the embodiment of the present invention performs feature classification and feature normalization on the original data set according to the timestamp detection model, for the original data set in the i th feature , if it contains the time field t, then extract the timestamp ; if it does not contain the time field t, then it is a non-time series feature. Detect whether the timestamp is increasing , if increasing, then it is a candidate time series feature; if not increasing, then it is a non-time series feature. Perform dynamic variability detection on the candidate time series feature and calculate the change rate. Among them, the calculation formula of the change rate is:

[0022] If the variance of the change rate is greater than the preset threshold, that is , then the candidate time series feature is a time series feature; if the variance of the change rate is not greater than the preset threshold, that is , then the candidate time series feature is a non-time series feature. Perform normalization processing on the time series feature and the non-time series feature, , where x is the original feature value, x represents the feature value in the feature f , which can be understood as f = x 1, x 2, …, x j ; is the normalized feature, obtaining the normalized time series feature set and the non-time series feature set , where is the i th normalized feature.

[0023] In yet another preferred embodiment, in S2, according to the feature extraction and fusion network, feature extraction and fusion are performed on the time series feature set and the non-time series feature set to obtain a fusion feature matrix, including:[[]] S201, input the time series feature set into a convolutional neural network for feature extraction, and output a time series feature matrix; S202, input the non-time series feature set into an autoencoder for feature extraction, and output a non-time series feature matrix; S203, splice the time series feature matrix and the non-time series feature matrix row by row to form a high-dimensional feature matrix; S204, input the high-dimensional feature matrix into a fully connected layer for dimensionality reduction, and output the fusion feature matrix.

[0024] It should be noted that time series data usually exhibits a time-dependent relationship. In particular, local patterns within some time windows are crucial for subsequent prediction or classification. CNN can effectively capture local features or patterns in time series data, especially when there is local correlation between different time points. CNN has a strong ability to recognize local patterns in time series data, thus better understanding the complexity of time series data. Non-time series data is usually static data without a time-dependent relationship. An autoencoder is an unsupervised learning method that compresses high-dimensional data into a low-dimensional representation through training. Through the autoencoder, the latent features of these non-time series data can be extracted, and their dimensions can be reduced while retaining their key features. For non-time series data, the autoencoder can automatically learn how to effectively encode it and extract the main features of the data. Therefore, it does not require time series modeling like CNN. Based on this, in the embodiment of the present invention, the time series feature set is input into a convolutional neural network for feature extraction, and a time series feature matrix is output, where n is the total number of samples, q is the dimension of the extracted features. The non-time series feature set is input into an autoencoder for feature extraction, and a non-time series feature matrix is output, where r is the dimension of the extracted features. The time series feature matrix and the non-time series feature matrix are spliced row by row to form a high-dimensional feature matrix . The high-dimensional feature matrix is input into a fully connected layer for dimensionality reduction, and a fusion feature matrix is output, where d is the unified feature dimension after dimensionality reduction, is the weight matrix of the fully connected layer, is the bias vector of the fully connected layer, is the activation function.

[0025] In yet another preferred embodiment, the hybrid model includes a multi-branch LSTM model and an XGBoost model; The multi-branch LSTM model includes a systematic error branch, a random error branch, and a merging layer; The systematic error branch is used to process the predictable features in the temporal feature matrix of the fusion feature matrix and predict the systematic error; The random error branch is used to process the unpredictable features in the temporal feature matrix of the fusion feature matrix and predict the random error; The merging layer is used to splice the outputs of the two branches and pass through a fully connected layer to predict the total error of the LSTM model; The XGBoost model is used to predict the systematic error, random error, and total error of the non-temporal feature matrix in the fusion feature matrix.

[0026] Specifically, the embodiment of the present invention trains the fusion feature matrix based on the hybrid model to establish an error prediction model. Among them, the hybrid model includes a multi-branch LSTM model and an XGBoost model. The multi-branch LSTM model includes a systematic error branch, a random error branch, and a merging layer. The systematic error branch is used to process the predictable features in the temporal feature matrix of the fusion feature matrix and predict the systematic error ; the random error branch is used to process the unpredictable features in the temporal feature matrix of the fusion feature matrix and predict the random error ; the merging layer is used to splice the outputs of the two branches and pass through a fully connected layer to predict the total error of the LSTM model , where, is the Sigmoid activation function, and are the weight and bias matrices. It should be noted that the systematic error specifically refers to that when measuring repeatedly under the same conditions, the error will appear in a definite direction and amplitude, showing a consistent deviation. The random error specifically refers to that when measuring repeatedly under the same conditions, the error will appear in a random direction and amplitude, showing an inconsistent fluctuation.

[0027] Exemplarily, the unit formula of the multi-branch LSTM model is: Input gate: ; Forget gate: ; Output gate: ; Update the memory cell state: ; Update the hidden state: .

[0028] Among them, t is the time step, is the hidden state of the current time step, is the memory cell state of the current time step, is the input feature vector of the current time step, W is the weight matrix input to the gated unit, U is the weight matrix from the hidden state to the gated unit, b is the bias matrix, and ⊙ is the Hadamard product.

[0029] Preferably, the mean squared error loss function is adopted in the training of the multi-branch LSTM model , where N is the number of samples, is the prediction error, is the actual error.

[0030] The XGBoost model specifically constructs XGBoost models for the non-sequential feature matrices in the fusion feature matrix to predict the systematic error and the random error respectively, and finally combines and predicts the total error of the XGBoost model .

[0031] Preferably, the objective function adopted by the XGBoost model is , where is the loss function, used to measure the deviation between the true value and the predicted value , is the regularization term, controlling the complexity of the tree, is the number of leaves of the tree, is the weight of the leaf, and are the training parameters.

[0032] Preferably, the error prediction model is specifically .

[0033] In yet another preferred embodiment, in step S3, according to the preset error prediction model and the fusion feature matrix, analyze and quantify the contribution of each feature to the error and the correlation relationship between features, including: S301, extract the feature vectors in the fusion feature matrix, and calculate the SHAP value of each feature in the feature vectors according to the prediction error of the error prediction model to obtain a set of SHAP values; S302. Normalize the SHAP value set and calculate the contribution weight of each feature to obtain a feature contribution weight set; S303. Calculate the correlation between the features in the feature vector to obtain the correlation relationship between the features.

[0034] Specifically, in the embodiment of the present invention, a feature vector in the fusion feature matrix is extracted , and the SHAP value of each feature in the feature vector is calculated according to the prediction error of the error prediction model to obtain a SHAP value set. Among them, the calculation formula of the SHAP value is:

[0035] Among them, is F any subset that does not contain the feature in S , is the number of features in the subset S , is the prediction error of the error prediction model on the feature set S , is the prediction error of the error prediction model on the feature set S after adding the feature .

[0036] It should be noted that the SHAP value (SHapley Additive exPlanations) is a tool for explaining the prediction results of machine learning models. The SHAP value generates a value for each input feature, indicating how that feature contributes to the prediction of a specified data point. Some features have a positive impact on the prediction probability, while others have a negative impact. The SHAP value can provide a detailed explanation for a single sample, indicating which features have a significant impact on the prediction result of that sample. By aggregating the SHAP values of multiple samples, the importance distribution of each feature in the entire dataset can be understood. In the prediction of a single observation, the SHAP value plot can show how the difference between the predicted value of that observation and the average of all predicted values is contributed by the values of each feature. For the entire dataset, the SHAP value can show the importance of each feature to the prediction, thus helping to understand which features have the greatest impact on the model prediction.

[0037] Normalize the SHAP value set and calculate the contribution weight of each feature to obtain a feature contribution weight set , quantifying the impact of features on the error. Among them, the calculation formula of the contribution weight is:

[0038] Calculate the correlation between features in the eigenvector to obtain the association relationship between features.

[0039] As a preferred solution, the calculation formula for the correlation includes: Pearson correlation coefficient , representing the linear relationship between features; Spearman correlation coefficient , representing the monotonic relationship between features; Mutual information , representing the non-linear relationship between features; Among them, f represents a feature; a and b both represent feature indices; represents the a th eigenvalue of the feature; represents the b th eigenvalue of the feature; Corr represents the Pearson correlation coefficient calculation function; represents the eigenvalue and covariance; and respectively represent the standard deviations of the eigenvalues and ; represents the rank obtained after sorting the eigenvalues x ; represents the joint probability distribution; and respectively represent their respective marginal probability distributions.

[0040] Exemplarily, assume there are the following three features: x 1: Time deviation of the waypoint; x 2: Wind speed deviation of the waypoint; x 3: Altitude deviation of the waypoint.

[0041] Calculate three correlation indicators to obtain the results in Table 1 below: Table 1

[0042] 1) Time deviation ( x 1) and wind speed deviation ( x 2) Pearson correlation coefficient ( r = 0.85), indicating a strong linear relationship between the two; Spearman correlation coefficient ( ρ= 0.8), indicating a strong monotonic relationship between the two, that is, whether linear or non-linear, when the wind speed deviation increases, the time deviation also tends to increase, but the value is slightly lower than the Pearson correlation coefficient because there may be some interference from non-linear relationships; Mutual information ( I = 0.95), indicating a very strong information association between the two, that is, one feature can be predicted from the other feature, including linear and non-linear information.

[0043] 2) Time deviation ( x 1) and height deviation ( x 3) Pearson correlation coefficient ( r = 0.3), indicating a weak linear relationship between the two, indicating that the linear relationship between the two is very weak and there is no significant linear positive or negative correlation pattern in terms of value; Spearman correlation coefficient ( ρ = 0.5), indicating that there may be a certain monotonic relationship between the two. For example, when the height deviation increases, the time deviation may also gradually increase, but this relationship may be non-linear; Mutual information ( I = 0.7), indicating a strong information correlation between the two, indicating that there may be complex non-linear dependencies.

[0044] 3) Wind speed deviation ( x 2) and height deviation ( x 3) Pearson correlation coefficient ( r = 0.75), indicating a strong linear positive correlation between the two. For example, under specific conditions, the wind speed deviation and the height deviation may show a consistent increasing trend; Spearman correlation coefficient ( ρ = 0.6), indicating a certain monotonic relationship between the two. Even if this relationship does not fully conform to the linear pattern, the wind speed deviation and the height deviation still tend to change synchronously; Mutual information ( I = 0.8), indicating a relatively strong information correlation between the two, and there may be complex relationship patterns.

[0045] Correspondingly, the present invention also provides a deep learning-based airway error analysis device, which can implement all the processes of the deep learning-based airway error analysis method in the above embodiments.

[0046] Please refer to Figure 2 , Figure 2 which is a schematic structural diagram of a preferred embodiment of a deep learning-based airway error analysis device provided by the present invention. The deep learning-based airway error analysis device includes: The feature classification module 201 is used to perform feature classification and feature normalization on the original data set according to the timestamp detection model, so as to obtain a time series feature set and a non-time series feature set; wherein, the original data set includes several features of the route. The feature fusion module 202 is used to perform feature extraction and fusion on the time series feature set and the non-time series feature set according to the feature extraction and fusion network, so as to obtain a fusion feature matrix. The error analysis module 203 is used to analyze and quantify the contribution of each feature to the error and the correlation relationship between features according to the preset error prediction model and the fusion feature matrix; wherein, the error prediction model is obtained by training the fusion feature matrix with a hybrid model.

[0047] Preferably, the feature classification module 201 is specifically used for: Judge whether each feature in the original data set contains a time field. If not, the feature is a non-time series feature; if so, extract the timestamp and detect whether the timestamp is increasing. If not, the feature is a non-time series feature; if so, the feature is a candidate time series feature. Perform dynamic variability detection on the candidate time series feature and calculate the change rate. If the variance of the change rate is greater than a preset threshold, the candidate time series feature is a time series feature; if the variance of the change rate is not greater than the preset threshold, the candidate time series feature is a non-time series feature. Perform normalization processing on the time series feature and the non-time series feature to obtain the normalized time series feature set and the non-time series feature set.

[0048] Preferably, the feature fusion module 202 is specifically used for: Input the time series feature set into a convolutional neural network for feature extraction, and output a time series feature matrix. Input the non-time series feature set into an autoencoder for feature extraction, and output a non-time series feature matrix. Concatenate the time series feature matrix and the non-time series feature matrix row by row to form a high-dimensional feature matrix. Input the high-dimensional feature matrix into a fully connected layer for dimensionality reduction, and output the fusion feature matrix.

[0049] Preferably, the hybrid model includes a multi-branch LSTM model and an XGBoost model. The multi-branch LSTM model includes a systematic error branch, a random error branch, and a merging layer. The systematic error branch is used to process the predictable features in the time series feature matrix in the fusion feature matrix and predict the systematic error. The random error branch is used to process the unpredictable features in the time series feature matrix in the fusion feature matrix and predict the random error; The merging layer is used to splice the outputs of the two branches and pass through a fully connected layer to predict the total error of the LSTM model; The XGBoost model is used to predict the systematic error, random error, and total error of the non-time series feature matrix in the fusion feature matrix.

[0050] Preferably, the error analysis module 203 is specifically configured to: Extract the feature vectors in the fusion feature matrix, and calculate the SHAP value of each feature in the feature vectors according to the prediction error of the error prediction model to obtain a set of SHAP values; Perform normalization processing on the set of SHAP values, and calculate the contribution weight of each feature to obtain a set of feature contribution weights; Calculate the correlation between the features in the feature vectors to obtain the association relationship between the features.

[0051] Preferably, the calculation formula for the correlation includes: Pearson correlation coefficient , representing the linear relationship between features; Spearman correlation coefficient , representing the monotonic relationship between features; Mutual information , representing the non-linear relationship between features; Among them, f represents a feature; a and b both represent feature indices; represents the a th feature value; represents the b th feature value; Corr represents the Pearson correlation coefficient calculation function; represents the feature value and covariance; and respectively represent the standard deviations of the feature values and ; represents the rank obtained after sorting the feature value x ; represents the joint probability distribution; and respectively represent their respective marginal probability distributions.

[0052] In specific implementation, the working principle, control flow, and achieved technical effects of the deep learning-based route error analysis device provided in the embodiments of the present invention are the same as those of the deep learning-based route error analysis method in the above embodiments, and will not be elaborated here.

[0053] Please refer to Figure 3 , Figure 3 which is a schematic structural diagram of a preferred embodiment of a terminal device provided by the present invention. The terminal device includes a processor 301, a memory 302, and a computer program stored in the memory 302 and configured to be executed by the processor 301. When the processor 301 executes the computer program, it implements the deep learning-based route error analysis method described in any of the above embodiments.

[0054] Preferably, the computer program can be divided into one or more modules / units (such as computer program 1, computer program 2,...). The one or more modules / units are stored in the memory 302 and executed by the processor 301 to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, and these instruction segments are used to describe the execution process of the computer program in the terminal device.

[0055] The processor 301 can be a central processing unit (CPU), or can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor 301 can also be any conventional processor. The processor 301 is the control center of the terminal device, and connects various parts of the terminal device through various interfaces and lines.

[0056] The memory 302 mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function, etc., and the data storage area can store relevant data, etc. In addition, the memory 302 can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc., or the memory 302 can also be other volatile solid-state storage devices.

[0057] It should be noted that the above terminal device may include, but is not limited to, a processor and a memory. Those skilled in the art can understand that Figure 3 the structural schematic diagram is only an example of the above terminal device, and does not constitute a limitation on the above terminal device. It may include more or fewer components than shown in the figure, or combine some components, or different components.

[0058] An embodiment of the present invention also provides a computer-readable storage medium, which includes a stored computer program. When the computer program runs, it controls the device where the computer-readable storage medium is located to execute the method for analyzing route error based on deep learning described in any one of the above embodiments.

[0059] An embodiment of the present invention also provides a computer program product, which includes a computer program or computer instructions. When the computer program or the computer instructions are executed by a processor, they implement the method for analyzing route error based on deep learning described in any one of the above embodiments.

[0060] An embodiment of the present invention provides a method, device, equipment, medium and product for analyzing route error based on deep learning. By performing feature classification and feature normalization on the original data set according to the timestamp detection model, a time series feature set and a non-time series feature set are obtained; where the original data set includes several features of the route; according to the feature extraction and fusion network, feature extraction and fusion are performed on the time series feature set and the non-time series feature set to obtain a fusion feature matrix; according to a preset error prediction model and the fusion feature matrix, the contribution of each feature to the error and the correlation relationship between features are analyzed and quantified; where the error prediction model is obtained by training the fusion feature matrix with a hybrid model. By analyzing and quantifying the contribution of features to the error and the correlation relationship between features, the embodiment of the present invention can comprehensively analyze the source of route error and provide data-driven decision support for route planning and optimization.

[0061] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the accompanying drawings of the system embodiments provided by the present invention, the connection relationships between the modules indicate that there are communication connections between them, which can be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0062] The above is the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art in this technical field, without departing from the principle of the present invention, several improvements and refinements can still be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A route error analysis method based on deep learning, characterized in that: include: Performing feature classification and feature normalization on the original data set according to the timestamp detection model to obtain a time series feature set and a non-time series feature set; wherein the original data set includes several features of the route; Extract and fuse the time series feature set and the non-time series feature set according to a feature extraction and fusion network to obtain a fusion feature matrix; According to the preset error prediction model and the fused feature matrix, the contribution of each feature to the error and the correlation between features are analyzed and quantified; wherein the error prediction model is obtained by training the fused feature matrix with a hybrid model.

2. The route error analysis method based on deep learning according to claim 1, characterized in that: The feature classification and feature normalization of the original data set are performed according to the timestamp detection model to obtain a time series feature set and a non-time series feature set, including: Determine whether each feature in the original data set includes a time field; If not, the feature is a non-time series feature; if so, extract the timestamp and detect whether the timestamp is increasing; If not, the feature is a non-time series feature; if so, the feature is a candidate time series feature; Performing dynamic change detection on the candidate time series features and calculating the change rate; If the variance of the change rate is greater than a preset threshold, the candidate time series feature is a time series feature; if the variance of the change rate is not greater than the preset threshold, the candidate time series feature is a non-time series feature; The time series features and the non-time series features are normalized to obtain the normalized time series feature set and the non-time series feature set.

3. The route error analysis method based on deep learning as claimed in claim 2, characterized in that: The extracting and fusing the time series feature set and the non-time series feature set according to the feature extraction and fusion network to obtain a fusion feature matrix includes: Inputting the time series feature set into a convolutional neural network for feature extraction, and outputting a time series feature matrix; Inputting the non-time series feature set into an automatic encoder for feature extraction, and outputting a non-time series feature matrix; Concatenate the time series feature matrix and the non-time series feature matrix row by row to form a high-dimensional feature matrix; The high-dimensional feature matrix is ​​input into a fully connected layer for dimensionality reduction, and the fused feature matrix is ​​output.

4. The route error analysis method based on deep learning as claimed in claim 3, characterized in that: The hybrid model includes a multi-branch LSTM model and an XGBoost model; The multi-branch LSTM model includes a systematic error branch, a random error branch and a merging layer; The system error branch is used to process the predictable features in the time series feature matrix in the fusion feature matrix to predict the system error; The random error branch is used to process the unpredictable features in the time series feature matrix in the fusion feature matrix and predict the random error; The merging layer is used to concatenate the outputs of the two branches and pass them through the fully connected layer to predict the total error of the LSTM model; The XGBoost model is used to predict the systematic error, random error and total error of the non-time series feature matrix in the fused feature matrix.

5. The route error analysis method based on deep learning as claimed in claim 4, characterized in that: The step of analyzing and quantifying the contribution of each feature to the error and the correlation between the features according to the preset error prediction model and the fusion feature matrix includes: Extracting a feature vector from the fused feature matrix, and calculating a SHAP value of each feature in the feature vector according to a prediction error of the error prediction model to obtain a SHAP value set; Normalizing the SHAP value set, and calculating the contribution weight of each feature to obtain a feature contribution weight set; The correlation between the features in the feature vector is calculated to obtain the correlation relationship between the features.

6. The route error analysis method based on deep learning as claimed in claim 5, characterized in that: The calculation formula of the correlation includes: Pearson correlation coefficient , characterizes the linear relationship between features; Spearman correlation coefficient , characterizes the monotonic relationship between features; Mutual Information , characterizes the nonlinear relationship between features; in, f Indicates characteristics; a and b Both represent feature indexes; Indicates a The eigenvalue of a feature; Indicates b The characteristic value of each feature; Corr represents the Pearson correlation coefficient calculation function; Represents eigenvalue and The covariance of and Represents the eigenvalues and The standard deviation of Represents the eigenvalue x The rank obtained after sorting; represents the joint probability distribution; and They represent their respective marginal probability distributions.

7. A route error analysis device based on deep learning, characterized in that: include: A feature classification module, used for performing feature classification and feature normalization on the original data set according to the timestamp detection model to obtain a time series feature set and a non-time series feature set; wherein the original data set includes several features of the route; A feature fusion module, used for extracting and fusing the time series feature set and the non-time series feature set according to a feature extraction and fusion network to obtain a fusion feature matrix; The error analysis module is used to analyze and quantify the contribution of each feature to the error and the correlation between features according to a preset error prediction model and the fused feature matrix; wherein the error prediction model is obtained by training the fused feature matrix with a hybrid model.

8. A terminal device, characterized in that: It includes a processor and a memory, wherein a computer program is stored in the memory, and the computer program is configured to be executed by the processor, and when the processor executes the computer program, the deep learning-based route error analysis method as described in any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein when the device where the computer-readable storage medium is located executes the computer program, the deep learning-based route error analysis method as described in any one of claims 1 to 6 is implemented.

10. A computer program product, characterized in that The computer program product comprises a computer program or a computer instruction, and when the computer program or the computer instruction is executed by a processor, the deep learning-based route error analysis method as described in any one of claims 1 to 6 is implemented.

Images