CN114445252A - Data completion method, device, electronic device and storage medium - Google Patents

Abstract

The embodiment of the invention provides a data completion method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence. The method comprises the following steps: acquiring an original data set; carrying out data preprocessing on an original data set and classifying the original data set into a incomplete data set and a historical data set; extracting time characteristics of the incomplete data set to obtain a first time sequence; extracting time characteristics of the historical data set to obtain a second time sequence; performing multi-head attention mechanism calculation on the first time sequence to obtain a first output matrix; performing multi-head attention mechanism calculation on the second time sequence to obtain a second output matrix; and performing fusion processing on the first output matrix and the second output matrix to obtain the completion data. The method can complete the missing traffic data, and has good data completing effect in the face of different data missing conditions.

Description

Data completion method, device, electronic device and storage medium

technical field

The present invention relates to the technical field of artificial intelligence, and in particular, to a data completion method, device, electronic device and storage medium.

Background technique

Traffic data is a collection of road characteristics over a period of time, such as flow and speed, and is an important part of Intelligent Transportation System (ITS). On the basis of road characteristic data, traffic departments can carry out reasonable and effective traffic control, and enterprises can also provide more accurate and reliable services. But in practice, traffic datasets are often missing due to sensor failures, regional power outages, extreme weather, etc. Therefore, how to provide a data completion method to complete the missing traffic data has become an urgent technical problem to be solved.

SUMMARY OF THE INVENTION

The present invention aims to solve one of the technical problems in the related art at least to a certain extent. Therefore, the present invention provides a data completion method, device, electronic device and storage medium, which can complete the missing traffic data on the basis of historical traffic data, and have good data completion effect in different data missing situations.

To achieve the above object, a first aspect of the embodiments of the present invention provides a data completion method, including:

get the original dataset;

performing data preprocessing on the original dataset to classify the original dataset into incomplete datasets and historical datasets;

Extracting temporal features on the incomplete data set to obtain a first time series;

performing time feature extraction on the historical data set to obtain a second time series;

Perform multi-head attention mechanism calculation on the first time series to obtain a first output matrix;

Perform multi-head attention mechanism calculation on the second time series to obtain a second output matrix;

Fusion processing is performed on the first output matrix and the second output matrix to obtain complementary data.

In some embodiments of the present invention, performing temporal feature extraction on the incomplete data set to obtain a first time series, including:

Inputting the incomplete dataset into a preset long short-term memory completion network;

extracting the observation data in the preset time period in the incomplete data set;

Calculate the missing data of the preset period according to the observation data and the prediction data of the previous period of the preset period;

According to the missing data and the observed data, the first time series is obtained.

In some embodiments of the present invention, performing temporal feature extraction on the historical data set to obtain a second time series, including:

inputting the historical data set into a preset historical data processing network;

Calculate the historical average data of the historical data set through the historical data processing network, and use the historical average data as the historical data of the preset period;

The second time series is obtained according to the historical average data.

In some embodiments of the present invention, the multi-head attention mechanism calculation is performed on the first time series to obtain a first output matrix, including:

inputting the first time series to a first multi-head attention layer;

converting the first time series into a first attention matrix through the first multi-head attention layer;

converting the first attention matrix into a first probability matrix by using a preset function;

Perform an attention mechanism calculation on the first probability matrix to obtain a first feature matrix;

Dimension reduction processing is performed on the first feature matrix to obtain a first output matrix.

In some embodiments of the present invention, the multi-head attention mechanism calculation is performed on the second time series to obtain a second output matrix, including:

inputting the second time series to a second multi-head attention layer;

converting the second time series into a second attention matrix by the second multi-head attention layer;

converting the second attention matrix into a second probability matrix through a preset function;

Perform attention mechanism calculation on the second probability matrix to obtain a second feature matrix;

A dimensionality reduction process is performed on the second feature matrix to obtain a second output matrix.

In some embodiments of the present invention, performing fusion processing on the first output matrix and the second output matrix to obtain complementary data includes:

performing splicing processing on the first output matrix and the second output matrix to obtain a splicing matrix;

The attention mechanism is calculated on the splicing matrix through the single-head attention layer, and the target matrix is obtained;

Feature extraction is performed on the target matrix through a linear layer to obtain complementary data.

In some embodiments of the present invention, the calculation of the attention mechanism is performed on the splicing matrix through the single-head attention layer, and the obtained target matrix includes:

inputting the splicing matrix into a single-head attention layer;

converting the splicing matrix into an attention matrix through the single-head attention layer;

converting the attention matrix into a probability matrix by a preset function;

Perform attention mechanism calculation on the probability matrix to obtain the target matrix.

To achieve the above purpose, a second aspect of the embodiments of the present invention provides a data completion device, including:

The original data set acquisition module is used to obtain the original data set;

an original data set preprocessing module, configured to perform data preprocessing on the original data set, so as to classify the original data set into incomplete data sets and historical data sets;

a first time series extraction module, configured to perform time feature extraction on the incomplete data set to obtain a first time series;

A second time series extraction module, configured to perform time feature extraction on the historical data set to obtain a second time series;

a first output matrix calculation module, configured to perform multi-head attention mechanism calculation on the first time series to obtain a first output matrix;

A second output matrix calculation module, configured to perform multi-head attention mechanism calculation on the second time series to obtain a second output matrix;

A data fusion module, configured to perform fusion processing on the first output matrix and the second output matrix to obtain complementary data.

To achieve the above purpose, a third aspect of the embodiments of the present invention provides an electronic device, including:

at least one memory;

at least one processor;

at least one program;

The program is stored in the memory, and the processor executes the at least one program to implement the data completion method according to the first aspect of the present invention.

In order to achieve the above object, a fourth aspect of the present invention provides a storage medium, which is a computer-readable storage medium, and the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used for Make the computer execute:

The data completion method as described in the first aspect above.

The data completion method, device, electronic device, and storage medium proposed by the embodiments of the present invention obtain the original data set, perform data preprocessing on the original data set, classify and process the original data set into incomplete data sets and historical data sets, and perform data processing on the incomplete data sets. Temporal feature extraction, obtain the first time series, perform temporal feature extraction on the historical data set, obtain the second time series, perform multi-head attention mechanism calculation on the first time series, obtain the first output matrix, and perform multi-head on the second time series. The attention mechanism is calculated to obtain the second output matrix, and finally the first output matrix and the second output matrix are fused to obtain the completion data. The technical solution provided by the embodiments of the present invention can solve the problem of data loss caused by the fact that the collection of traffic data is often affected by communication errors, sensor failures, storage loss, etc. Make full use of historical data to effectively improve the effect of traffic data completion.

Description of drawings

1 is a schematic diagram of complete random missing traffic data provided by an embodiment of the present invention;

2 is a schematic diagram of non-random missing traffic data provided by an embodiment of the present invention;

Fig. 3 is the traffic speed diagram of every 10 minutes on different dates in Dongsanlihe Street, Beijing, China provided by the embodiment of the present invention;

4 is a flowchart of a data completion method provided by an embodiment of the present invention;

Fig. 5 is the flow chart of step S130 in Fig. 4;

Fig. 6 is the flow chart of step S140 in Fig. 4;

Fig. 7 is the flow chart of step S150 in Fig. 4;

Fig. 8 is the flow chart of step S160 in Fig. 4;

Fig. 9 is the flow chart of step S170 in Fig. 4;

Fig. 10 is the flow chart of step S620 in Fig. 9;

11 is a schematic diagram of an overall framework of an attention recurrent neural network for traffic data completion provided by an embodiment of the present invention;

12 is a schematic diagram of a data set in Beijing, China provided by an embodiment of the present invention;

13 is a schematic diagram of a data set in the fifth district of California provided by an embodiment of the present invention;

14 is a schematic diagram of a data set in Hong Kong, China provided by an embodiment of the present invention;

FIG. 15 is a schematic diagram of a hardware structure of an electronic device provided by an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

In the description of the present invention, it should be understood that the orientations related to the description, such as up, down, front, rear, left, right, etc., are based on the orientation or positional relationship shown in the drawings, and are only In order to facilitate the description of the present invention and simplify the description, it is not indicated or implied that the indicated device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore should not be construed as limiting the present invention.

In the description of the present invention, the meaning of several is one or more, the meaning of multiple is two or more, greater than, less than, exceeding, etc. are understood as not including this number, above, below, within, etc. are understood as including this number. If it is described that the first and the second are only for the purpose of distinguishing technical features, it cannot be understood as indicating or implying relative importance, or indicating the number of the indicated technical features or the order of the indicated technical features. relationships, nor are they necessarily used to describe a particular order or sequence.

In the description of the present invention, unless otherwise clearly defined, words such as setting, installation, connection should be understood in a broad sense, and those skilled in the art can reasonably determine the specific meaning of the above words in the present invention in combination with the specific content of the technical solution.

First, some terms involved in this application are analyzed:

Reinforcement learning (Reinforcement Learning): Reinforcement learning, also known as reinforcement learning, evaluation learning or reinforcement learning, is one of the paradigms and methodologies of machine learning. The problem of strategies to maximize rewards or achieve specific goals; learn the mapping from environmental states to behaviors, so that the behavior selected by the agent can obtain the maximum reward of the environment, so that the external environment can evaluate the learning system in a certain sense (or The operating performance of the whole system) is the best. The basic principle is: if a certain behavior strategy of the agent leads to a positive reward (strengthening signal) in the environment, then the tendency of the agent to generate this behavior strategy in the future will be strengthened. The agent's goal is to discover the optimal policy in each discrete state to maximize the expected discounted reward sum. Reinforcement learning regards learning as a tentative evaluation process. Agent selects an action to use in the environment. After the environment accepts the action, the state changes, and at the same time, a reinforcement signal (reward or punishment) is generated to feed back to the agent. Choose the next action. The principle of selection is to make reinforcement learning different from supervised learning in connectionist learning. It is mainly reflected in the teacher's signal. In reinforcement learning, the reinforcement signal provided by the environment is a kind of agent's decision on the quality of the generated action. Evaluate (usually a scalar signal) rather than tell the agent how to generate the correct action. Since the external environment provides little information, the agent must learn from its own experience. In this way, the agent acquires knowledge in the environment where actions are evaluated one by one, and the probability of improving the action plan to adapt to the environment is increased by positive reinforcement (reward). The selected action not only affects the immediate enhancement value, but also affects the state of the environment at the next moment and the final enhancement value. The goal of reinforcement learning system learning is to dynamically adjust parameters to maximize the reinforcement signal. Commonly referred to as reinforcement learning, the agent, as a learning system, obtains the current state information s of the external environment, adopts a tentative behavior u to the environment, and obtains the evaluation r of the action and the new environment state feedback from the environment. If an action u of the agent leads to a positive reward (immediate reward) in the environment, then the tendency of the agent to produce this action in the future will be strengthened; otherwise, the tendency of the agent to produce this action will be weakened. In the repeated interaction between the control behavior of the learning system and the state and evaluation of the environmental feedback, the mapping strategy from the state to the action is continuously modified in a learning manner to achieve the purpose of optimizing the system performance. Reinforcement learning includes two categories: Value-based (value-based) and Policy-based (policy-based), where Value-based is to learn a value function, take a strategy from the value function, and determine a strategy at _t , which is an indirect strategy. method; the action-value estimates in Value-Base will eventually converge to the corresponding true values (usually different finite numbers that can be converted into probabilities between 0 and 1), so a deterministic strategy is usually obtained. policy); Policy-based is a learning policy function, a method of directly generating policies, which will generate the probability of each action πθ(a∣s); Policy-Based usually does not converge to a deterministic value; Policy-Based is suitable for continuous The action space of , in the continuous action space, instead of calculating the probability of each action, the action can be selected through the Gaussian distribution (normal distribution).

RNN (Recurrent Neural Networks): RNN consists of an input layer, a hidden layer and an output layer. A recurrent neural network neuron not only predicts, but also passes a time step s _t-1 to the next neuron. Among them, the output layer Ot is a fully connected layer, Ot=g(Vs _t ), g is the activation function, V is the network weight matrix of the output layer, and s _t is the hidden layer; that is, each of its nodes is related to the hidden layer. connected to each node. The current output is calculated by the hidden layer, and the calculation of the current hidden layer is not only related to the input, but also related to the output of the previous hidden layer; s _t =f(Ux _t +Ws _t-1 ), U is the network weight of the input x Matrix, W is the network weight matrix of the previous value as the input of this time, f is the activation function, and the hidden layer is the recurrent layer.

Long short-term memory network (Long short-Term Memory): the full name of the long short-term memory network (LongShort Term Memory networks), is a special RNN. LSTMs work very well on many problems and are now widely used. LSTMs are designed to explicitly avoid the problem of long-term dependencies. All RNNs have the form of chains of repeating neural network modules. In ordinary RNNs, the repeating module structure is very simple, such as only one tanh layer. LSTMs use a "gate" structure to remove or add information to the cell's ability to state. A gate is a way of letting information through selectively. Contains a sigmoid neural network layer and a pointwise multiplication operation. The sigmoid layer outputs a number between 0 and 1 that describes how much of each part can pass through. 0 means "don't allow any amount to pass" and 1 means "allow any amount to pass". LSTM has three gates to protect and control the cell state: input gate, output gate, and forget gate.

Attention Mechanism originated from the study of human vision. In cognitive science, due to bottlenecks in information processing, humans selectively focus on a portion of all information while ignoring other visible information. The above mechanism is often referred to as the attention mechanism. Neural attention mechanisms can give a neural network the ability to focus on a subset of its inputs: select specific inputs. Attention can be applied to any type of input regardless of its shape. In the case of limited computing power, attention mechanism is a resource allocation scheme that is the main means to solve the problem of information overload, allocating computing resources to more important tasks. The Multi-head Attention Mechanism uses multiple queries to compute and select multiple pieces of information from the input information in parallel. Each attention focuses on a different part of the input information.

Statistical learning-based methods: Traditional statistical learning applies statistical operations to perform simple completion of incomplete data, such as linear interpolation, historical averages and last observations to supplement missing values, but such methods can only solve the case where the data distribution is relatively simple . Some researchers propose to use the data around the missing position in the feature matrix to calculate the missing value, such as the classic K-Nearest Neighbors (KNN) algorithm, which fills the missing value by calculating the average of K adjacent positions around the missing position. Autoregressive Autoregressive Models with Integral Moving Average (ARIMA) and its variants calculate missing values by making predictions on missing data based on historical data, however, these methods cannot effectively utilize features collected after the occurrence of missing data. Constrained methods establish completion rules based on the overall data characteristics in the dataset; however, this method is only suitable for univariate data and does not work well in most practical scenarios involving multivariate data.

Method based on matrix factorization: Matrix factorization (Matrix Factorization) finds the correlation in the data by decomposing and reconstructing the data matrix, and completes the missing values. Temporal Regularized Matrix Factorization (TRMF) is a time series completion method that introduces regularization and scalable matrix factorization based on matrix factorization. In addition, there is a technical solution that introduces Probabilistic Principal Component Analysis (PPCA) into matrix decomposition, and this method assumes that the latent features of the observed data conform to a Gaussian distribution, so as to perform data completion. There is also a more sophisticated low-rank tensor complementation algorithm to recover missing data. A Bayesian Gaussia Candecomp/Parafac (BGCP) tensor decomposition method is proposed to convert the original data matrix into a high-dimensional tensor, and then describe and restore the incomplete matrix. However, matrix factorization-based methods require inputs with specific shapes, which limits their applications.

Generative Adversarial Networks (GAN, Generative Adversarial Networks): It is a deep learning model and one of the most promising methods for unsupervised learning on complex distributions in recent years. The model produces fairly good outputs through the mutual game learning of at least two modules in the framework: Generative Model and Discriminative Model. In the original GAN theory, it is not required that G and D are both neural networks, but only functions that can fit the corresponding generation and discrimination. But in practice, deep neural networks are generally used as G and D. An excellent GAN application needs to have a good training method, otherwise the output may be unsatisfactory due to the freedom of the neural network model.

Historical Average (HA): HA uses the complete data of the last five days and calculates the average value for the corresponding time period of the day for each road segment to fill in the missing values.

K-Nearest Neighbors Algorithm (KNN): KNN is a non-parametric interpolation method that calculates the missing by the average of the k nearest neighbors. In the embodiment of the invention, 4 nearest neighbors are considered for calculation .

Bayesian Gaussian CP decomposition (BGCP): BGCP is a Bayesian tensor decomposition model that uses Markov Chain Monte Carlo to model latent factors, ie low-rank structures.

Denoising Stacked Autoencoders (DSAE): DSAE receives the time series data feature matrix containing observations and noise, and then extracts the hidden features of the data by reducing and increasing the dimension, and outputs the completion result.

Bayesian Temporal Matrix Factorization (BTMF): BTMF uses a Gaussian vector autoregressive process to model temporal dependencies to complement time series data.

Parallel Data and Generative Adversarial Networks for Imputation (PGAN): PGAN is a GAN-based data completion method for generating missing traffic data. Both the generator and the discriminator consist of linear layers.

Bidirectional Recurrent Imputation (BRITS): BRITS inputs the missing time data into the forward and reverse RNNs respectively, and combines the outputs of the two RNNs to estimate the missing time series values.

Completely Missing Data and Non-Random Missing: The present invention defines two common data missing modes, namely completely random missing (MCAR) and non-random missing (MNAR), as shown in Figure 1 and Figure 2, respectively, where the gray cells are missing data, white cells are non-missing data. In MCAR, the distribution of missing data is scattered and random across the time series, while in MNAR, missing data points occur at consecutive time points. Relatively speaking, MNAR is a more challenging problem due to the lack of adjacent information necessary to recover a single missing point.

Traffic data is a collection of road features (such as flow and speed) over a period of time, and is an important part of Intelligent Transportation System (ITS). On the basis of these data, traffic departments can carry out reasonable and effective traffic control, and enterprises can also provide more accurate and reliable services. Recently, deep learning-based algorithms

However, most deep learning based methods are highly dependent on the quality of the data. In practice, traffic datasets are often missing due to sensor failures, regional power outages, extreme weather, etc. Therefore, the problem of missing traffic data needs to be solved urgently.

To address the missingness problem, the most straightforward approach is to delete as and unobserved data. However, all such operations result in the loss of temporal or spatial information. To solve this problem, an appropriate data completion method can be used to estimate the incomplete data. Data completion methods estimate missing values by analyzing the dependence or distribution of the data. Appropriate data completion methods can accurately recover missing data, thereby avoiding the performance degradation of various downstream data mining algorithms in intelligent transportation systems.

Furthermore, compared to regular time series data, such as stock index and medical device data, traffic data is strongly cyclical and volatile. Figure 3 shows the road speed of Dongsanlihe Street in Beijing, China over five days. Overall, traffic speeds have similar patterns for consecutive days. For example, during the morning and evening rush hours every day, road speeds are low. Furthermore, due to various factors such as weather, accident and date, traffic speeds are completely different from day to day, and the speed at each point in time is highly correlated with the road speed at the time before and after. Therefore, how to combine the observational data before and after different time points and the overall historical data is crucial for dealing with missing data.

Some technical solutions design algorithms based on statistical methods such as KNN and ARIMA to complete incomplete data, but these methods are only effective for data with relatively simple distributions. Based on matrix decomposition, some technical solutions introduce probability models and Gaussian distributions, but these methods are only suitable for low-rank data and are difficult to deal with traffic data affected by many factors. In addition, these methods have specific formats for input data. requirements, which limit its application scenarios. In recent years, deep learning has been widely used in data completion problems. For example, some technical solutions propose BRITS with a bidirectional recurrent neural network structure to combine data before and after the missing position for completion. Some technical solutions propose E2GAN based on adversarial generative networks to generate missing data. However, these studies often suffer from slow computation or convergence problems on large datasets. Although existing methods have produced acceptable results on the incomplete traffic data completion problem, some challenges remain. For example, some existing methods are too complex and many models are difficult to train. Existing methods only consider the temporal dependencies in incomplete time series data, and do not take advantage of the inherent periodicity of traffic data. In addition, the question of how to better extract temporal features from mutilated data remains thought-provoking.

In order to solve the above problems and fill the research gap, the present invention proposes a brand-new data completion model, called Attention-Driven Recurrent Imputation Network (ADRIN). Unlike previous models, the present invention extracts features from mutilated inputs and historical averages to account for the volatility and periodicity of traffic data. ADRIN uses a Long Short-Term Memory for Imputation (LSTM-I) network to receive incomplete temporal inputs, and a multi-head attention network to model complete temporal features. Furthermore, the present invention applies a multi-head attention network to historical data to extract features related to historical information. The outputs of these two modules are then passed through a fusion module, which consists of an attention layer and a linear layer, to obtain the completion result.

Based on this, the embodiments of the present invention provide a data completion method, apparatus, electronic device, and storage medium, which can realize the completion of traffic data.

The embodiments of the present invention provide a data completion method, an apparatus, an electronic device, and a storage medium, which are specifically described by the following embodiments. First, the data completion method in the embodiments of the present invention is described.

The embodiments of the present invention can acquire and process related data based on artificial intelligence technology. Among them, artificial intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. .

The basic technologies of artificial intelligence generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, and mechatronics. Artificial intelligence software technology mainly includes computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning/deep learning.

The data completion method provided by the embodiment of the present invention relates to the technical field of artificial intelligence, and in particular, to a data completion method, an apparatus, an electronic device, and a storage medium. The data completion method provided by the embodiment of the present invention can be applied to a terminal, also can be applied to a server, and can also be software running in a terminal or a server. In some embodiments, the terminal may be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart watch, etc.; the server may be an independent server, or may provide cloud services, cloud databases, cloud computing, cloud functions, and cloud storage , network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and cloud servers for basic cloud computing services such as big data and artificial intelligence platforms; The application of the whole method, etc., but not limited to the above forms.

The present application may be used in numerous general purpose or special purpose computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, including A distributed computing environment for any of the above systems or devices, and the like. The application may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The application may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including storage devices.

FIG. 4 is an optional flowchart of a data completion method provided by an embodiment of the present invention. The method in FIG. 4 may include, but is not limited to, steps S110 to S170.

Step S110, obtaining the original data set;

Step S120, performing data preprocessing on the original data set to classify the original data set into incomplete data sets and historical data sets;

Step S130, performing time feature extraction on the incomplete data set to obtain a first time series;

Step S140, performing time feature extraction on the historical data set to obtain a second time series;

Step S150, performing multi-head attention mechanism calculation on the first time series to obtain a first output matrix;

Step S160, performing multi-head attention mechanism calculation on the second time series to obtain a second output matrix;

Step S170, performing fusion processing on the first output matrix and the second output matrix to obtain complementary data;

In practice, the collection of traffic data is often affected by factors such as communication errors, sensor failures, and storage loss, resulting in missing data, resulting in damage to the collected data, which seriously affects the effectiveness of downstream applications. However, existing completion methods only estimate missing data from observational data with incompleteness, ignoring the utilization of historical data. In the present invention, the attention cycle neural network is used to solve the problem of traffic data loss, which can effectively improve the traffic data completion effect.

In step S110 of some embodiments, an original data set is obtained, and the original data set can be used to collect traffic speed data in different regions according to the data sources by setting different data sources. For example, the data adopts three real-world traffic speed datasets, including but not limited to: China Beijing traffic data, California traffic data, and Hong Kong traffic data. Beijing, China traffic data includes the average speed of 1368 roads in Beijing, China from 00:00 on January 1, 2019 to 23:55 on June 30, 2019. The California traffic data includes 144 sensor stations in California's Fifth District, recorded from 00:00 on January 1, 2013 to 23:55 on June 30, 2013. Traffic data for Hong Kong, China includes average road speeds for major roads in Hong Kong, China from 00:00 on March 10, 2021 to 23:50 on July 31, 2021.

In step S120 of some embodiments, data preprocessing is performed on the original data set to classify the original data set into incomplete data sets and historical data sets. The collected data is sampled at different time intervals such as 5 minutes or 10 minutes, and the complete data is regarded as the historical data set, and the incomplete data is regarded as the incomplete data set.

In step S130 of some embodiments, time feature extraction is performed on the incomplete data set to obtain a first time series. For example, temporal feature extraction can be performed on incomplete datasets through an improved long short-term memory completion network. It is also possible to perform temporal feature extraction on incomplete datasets through parallel generative adversarial completion networks, recurrent neural networks, bidirectional recurrent neural networks, etc., but not limited to this.

In step S140 of some embodiments, time feature extraction is performed on the historical data set to obtain a second time series. For example, temporal feature extraction can be performed on historical datasets through a historical averaging network. It is also possible to perform temporal feature extraction on the historical data set through the K-proximity algorithm, linear interpolation, matrix decomposition-based method, etc., but not limited to this.

In step S150 of some embodiments, multi-head attention mechanism calculation is performed on the first time series to obtain a first output matrix. The multi-head attention mechanism calculation is performed on the first time series obtained in step S130 to obtain a first output matrix. Although the improved long-short-term memory completion network can gradually estimate missing data, it has limited ability to capture the temporal dependence of longer time series data, especially the traffic data often has hundreds of time periods in a day. By applying the multi-head attention mechanism, multi-head operations can be learned in several subspaces respectively, and time series features can be captured for the time series of long-term series data.

In step S160 of some embodiments, multi-head attention mechanism calculation is performed on the second time series to obtain a second output matrix. The multi-head attention mechanism calculation is performed on the second time series obtained in step S140 to obtain a second output matrix. The historical average network also cannot capture the time series of long time series data. In order to use the historical data more effectively, the multi-head attention mechanism is applied to calculate the historical data, and the second output matrix is obtained.

In step S170 of some embodiments, a fusion process is performed on the first output matrix and the second output matrix to obtain complementary data. The incomplete data set and the historical data set are outputted by the previous multi-head attention mechanism, and the two output matrices are fused in the fusion processing module to calculate the final completion result.

The data completion method provided by the embodiments of the above steps of the present invention can solve the problem of the lack of traffic data, and proposes an Attention Recurrent Neural Network (ADRIN) network structure, which is similar to the existing deep learning-based attribution methods. Ratio, attention recurrent neural network exploits the unique periodicity and volatility of traffic data, extracts features from incomplete input and historical average and fills missing values, attention recurrent neural network has long short-term memory completion network (LSTM-I ) to accept inputs with missings, while applying a multi-head attention mechanism to extract temporal features from historical average and LSTM-I outputs, respectively. The output of the multi-head attention mechanism is input to the fusion module containing the attention layer and the fully connected layer, and the completed result is obtained. The problem that the prior art solution cannot use historical data to complement missing data is solved, the embodiment of the present invention can solve the problem of traffic data loss, and can effectively improve the traffic data complementing effect.

Please refer to FIG. 5, in some embodiments, step S130 may include but is not limited to including steps S210 to S240;

Step S210, input the incomplete data set into the preset long-term and short-term memory completion network;

Step S220, extracting the observation data in the preset time period in the incomplete data set;

Step S230, according to the observation data and the prediction data of the previous period of the preset period, calculate the missing data of the preset period;

Step S240, obtaining a first time series according to the missing data and the observed data.

Specifically, in step S210 of some embodiments, the incomplete data set is input into the long-term and short-term memory completion network; in step S220 of some embodiments, observation data in a preset time period in the incomplete data set is extracted, and in the incomplete data set The degree of data integrity in different time periods in the collection is different, for example, including but not limited to: a data set from 8:00 to 20:00, the data from 10:00 to 12:00 is incomplete, while other time periods are complete data, then 10:00 to 12:00 is the observation data of the preset time period. In step S230 of some embodiments, the missing data of the preset period is calculated according to the observation data and the prediction data of the previous period of the preset period; according to the time period of the current observation data, for example, including but not limited to: August From 10:00 to 12:00 on the 23rd, due to the periodicity of traffic data, the missing data for the preset period can be calculated according to the time period from 10:00 to 12:00 on August 22nd. At the same time, according to the fitting of various data, the missing data of the preset period is calculated in the long short-term memory completion network. In step S240 of some embodiments, the first time series is obtained according to the missing data and the observed data, and the missing data of the preset period can be obtained by completing the missing data set to obtain the first time series. With the data completion method provided by the embodiments of the above steps of the present invention, the reconstructed LSTM-I uses the predicted value of the previous time period to estimate the missing value of the current time period. For each time period, LSTM-I uses the estimated and observed values to restore the features, resulting in the first time series.

Please refer to FIG. 6 , in some embodiments, step S140 may include but is not limited to including steps S310 to S330;

Step S310, input the historical data set into the preset historical data processing network;

Step S320, calculating the historical average data of the historical data set through the historical data processing network, and using the historical average data as the historical data of the preset period;

Step S330, obtaining a second time series according to the historical average data.

Specifically, in step S310 of some embodiments, the historical data set is input into a preset historical data processing network; in step S320 of some embodiments, the historical average data of the historical data set is calculated through the historical data processing network , take the historical average data as the historical data of the preset period, and use the complete data of the last five days to calculate the average value for the corresponding time period of each road section in one day to fill the missing value; step S330, according to the historical average data, obtain Second time series. Through the data completion method provided by the embodiments of the above steps of the present invention, due to the periodicity of traffic data, historical data plays an important role in the completion of traffic data, and calculating the historical average data of the historical data set can reduce errors and reduce interference. Prepare for the one-step fusion process.

Please refer to FIG. 7 , in some embodiments, step S150 may include but is not limited to including steps S410 to S450;

Step S410, inputting the first time series to the first multi-head attention layer;

Step S420, converting the first time series into a first attention matrix through the first multi-head attention layer;

Step S430, converting the first attention matrix into a first probability matrix by a preset function;

Step S440, performing attention mechanism calculation on the first probability matrix to obtain a first feature matrix;

Step S450, performing dimension reduction processing on the first feature matrix to obtain a first output matrix.

Specifically, in step S410 of some embodiments, the first time series is input to the first multi-head attention layer; in step S420 of some embodiments, the first time series is converted by the first multi-head attention layer is the first attention matrix, performing matrix operations in the multi-head attention mechanism to convert the first time series into a first attention matrix; in step S430 of some embodiments, the first attention matrix is converted by a preset function is the first probability matrix, using the softmax function to convert the first attention matrix into a first probability matrix, and the sum of the probabilities of all columns is 1; in step S440 of some embodiments, the attention mechanism calculation is performed on the first probability matrix , to obtain the first feature matrix, and through the calculation of the multi-head attention mechanism, richer feature relationships can be captured to obtain the first feature matrix; in step S450 in some embodiments, the first feature matrix is subjected to dimensionality reduction processing to obtain The first output matrix is calculated by the multi-head attention mechanism. Compared with the single-head attention mechanism, the output matrix Z of multiple attention mechanisms can be obtained. All the obtained output matrices Z are combined, and the dimensionality reduction of the fully connected layer is carried out. Output a first output matrix of the same shape as the input time series. Through the data completion method provided by the embodiments of the above steps of the present invention, the first time series is input to the first multi-head attention layer, and the first output matrix is finally obtained through matrix operation, attention mechanism calculation and dimensionality reduction processing. LSTM structure Time series data can only be modeled in a single direction, which makes it difficult to combine observations after missing data when performing completion tasks. The multi-head attention mechanism can be used to complete the incomplete data based on the entire time series data, combined with the observations before and after the missing position.

Please refer to FIG. 8 , in some embodiments, step S160 may include but is not limited to including steps S510 to S550;

Step S510, input the second time series to the second multi-head attention layer;

Step S520, converting the second time series into a second attention matrix through the second multi-head attention layer;

Step S530, converting the second attention matrix into a second probability matrix through a preset function;

Step S540, performing attention mechanism calculation on the second probability matrix to obtain a second feature matrix;

Step S550, performing dimension reduction processing on the second feature matrix to obtain a second output matrix.

Specifically, in step S510 of some embodiments, the second time series is input to the second multi-head attention layer; in step S520 of some embodiments, the second time series is converted by the second multi-head attention layer For the second attention matrix, the second time series matrix is converted into a second attention matrix; in step S530 of some embodiments, the second attention matrix is converted into a second probability matrix by a preset function, and a softmax function is used Convert the second attention matrix into a second probability matrix, and the sum of the probabilities of all columns is 1; in step S540 of some embodiments, the attention mechanism calculation is performed on the second probability matrix to obtain a second feature matrix. The calculation of the attention mechanism can capture richer feature relationships and obtain a second feature matrix; in step S550 in some embodiments, a dimensionality reduction process is performed on the second feature matrix to obtain a second output matrix. Through the data completion method provided by the embodiment of the above steps of the present invention, the second time series is input into the second multi-head attention layer, and the second output matrix is finally obtained through matrix operation, attention mechanism calculation and dimension reduction processing, LSTM The output result of -I can be regarded as a preliminary incomplete result, but it is limited to the ability of the model, and the completion effect is limited, so the embodiment of the present invention hopes to combine the multi-head attention mechanism with stronger time series modeling ability to combine before and after the incomplete position. The observation value of the data is used to complete, and considering the periodicity of traffic data, a historical data module (HA) is added. In order to make more effective use of historical data, and at the same time to ensure that the distribution of the output of the left and right parts of the model is similar at the fusion module, if they are inconsistent and input to the fusion module after direct splicing, the distribution difference will be too large and it will be difficult to perform parameter learning. The approximate distribution Makes the model parameters easier to learn.

Please refer to FIG. 9 , in some embodiments, step S170 may include but is not limited to including steps S610 to S630;

Step S610, performing splicing processing on the first output matrix and the second output matrix to obtain a splicing matrix;

Step S620, calculating the attention mechanism of the splicing matrix through the single-head attention layer to obtain the target matrix;

Step S630, perform feature extraction on the target matrix through the linear layer to obtain complementary data.

In step S610 of some embodiments, splicing processing is performed on the first output matrix and the second output matrix to obtain a splicing matrix. The incomplete data set and the historical data set are outputted through the previous multi-head attention mechanism, and the output matrix is splicing processing. Get the splice matrix.

In step S620 of some embodiments, an attention mechanism is calculated on the splicing matrix through a single-head attention layer to obtain a target matrix. The splicing matrix after splicing processing is input into the single-head attention layer, and the output matrix of the single-head attention layer is obtained.

In step S630 of some embodiments, feature extraction is performed on the target matrix through a linear layer to obtain complementary data. After the single-head attention layer is calculated, it is input to the linear layer for feature extraction, and a fully connected neural network is used in the linear layer to obtain the completion data. As shown in formula (1):

是注意力循环神经网络的最终补全结果。H and H ^* represent the concatenation of implicit states, W _l and b _l are the parameters of the linear layer,

is the final completion result of the attention recurrent neural network.

Through the data completion method provided by the embodiments of the above steps of the present invention, the completion data can be finally calculated by the fusion processing module, and the upper half of the attention recurrent neural network has a first output matrix for processing the incomplete data set, There are second output matrices for historical data set processing. These output matrices contain the characteristics of incomplete data sets and historical data sets. The fusion processing module needs to process the first output matrix and the second output matrix to obtain the final completion. data.

Please refer to FIG. 10 , in some embodiments, step S620 may include but is not limited to including steps S710 to S730

Step S710, input the splicing matrix to the single-head attention layer;

Step S720, converting the splicing matrix into an attention matrix through a single-head attention layer;

Step S730, converting the attention matrix into a probability matrix through a preset function;

Step S740, perform attention mechanism calculation on the probability matrix to obtain a target matrix.

In step S710 of some embodiments, the splicing matrix is input to the single-head attention layer, and the splicing matrix of the first output matrix and the second output matrix is input to the single-head attention layer; in step S720 of some embodiments, The splicing matrix is converted into an attention matrix through a single-head attention layer, and the splicing matrix is processed by the single-head attention layer to form an attention matrix; in step S730 of some embodiments, the attention matrix is converted into an attention matrix through a preset function probability matrix, use the softmax function to convert the attention matrix into a probability matrix, and the sum of the probabilities of all columns is 1; in step S740 of some embodiments, perform attention mechanism calculation on the probability matrix to obtain a target matrix, and use single-head attention Calculate the force mechanism to get the target matrix.

本发明实施例有带有缺失数据点的输入特征，用

表示，其中n代表节点如传感器站或道路段的数量，T代表一天中的时间步数，x_ij代表节点i在第j个时间步数的观察数据点。本发明实施例另外定义了一个掩码矩阵(也称为标志矩阵)

如公式(2)所示：In a specific application scenario, the goal of traffic speed data completion is to use the known incomplete traffic speed data to estimate the missing data. Taking into account real road speed data

Embodiments of the present invention have input features with missing data points, using

where n represents the number of nodes such as sensor stations or road segments, T represents the number of time steps in a day, and x _ij represents the observed data points for node i at the jth time step. The embodiment of the present invention additionally defines a mask matrix (also called a flag matrix)

As shown in formula (2):

For ease of understanding, the following is an example of the matrix form of the incomplete traffic data and the corresponding mask matrix:

和Y之间的误差应该最小化，其中

是归因结果。It can be seen that in the feature matrix, the data values of positions (1,3), (2,5), (3,2) are missing, and the missing data are represented by question marks. The corresponding binary value is 1, while the other observed data points have the value 0 in their respective mask matrices. The purpose of missing data incompleteness is to recover missing data points from existing observed data values.

and Y should be minimized, where

is the attribution result.

考虑到交通数据具有很强的周期相关性，本发明实施例另外构建了右边的模块，它接受输入前最近五天的历史平均数据矩阵

作为输入。这两个部分的输出是两个隐式特征矩阵，分别表示为

和

前者包含提取的残缺时间段之间的时序信息，而后者则包含了周期性的时间序列信息。最后，通过一个融合模块来处理两个模块的输出，并得到的补全后的数据

Figure 11 depicts the framework of the Attention Recurrent Neural Network (ADRIN). The embodiment of the present invention constructs two main data processing flows according to the time series characteristics of traffic speed data, as shown in the left and right parts of FIG. 11 . The module on the left focuses on extracting temporal features from the mutilated input, i.e.

Considering that the traffic data has a strong cyclical correlation, the embodiment of the present invention additionally constructs a module on the right, which accepts the historical average data matrix of the last five days before the input

as input. The outputs of these two parts are two implicit feature matrices, denoted as

and

The former contains time series information between the extracted incomplete time periods, while the latter contains periodic time series information. Finally, a fusion module is used to process the output of the two modules, and the completed data is obtained

此外，本发明实施例在时序数据处理中采用了多头注意力机制，分别提取X^和历史特征x^(a)的时间依赖性，并获得隐式特征矩阵H和H^*。Specifically, the embodiments of the present invention customize some advanced neural network methods according to the time series traffic speed data, and integrate them into ADRIN. In particular, the embodiments of the present invention first improve the ordinary LSTM and propose a Long Short-Term Memory Completion Network (LSTM-I), which can accept inputs with missing data. LSTM-I estimates missing values through forward prediction to generate predicted feature maps

In addition, the embodiment of the present invention adopts a multi-head attention mechanism in time series data processing, extracts the time dependencies of X^ and historical features x ^(a) respectively, and obtains implicit feature matrices H and H ^* .

In a specific application scenario, LSTM has achieved many achievements in time series data modeling tasks, especially in time series prediction. Compared with ordinary RNN, LSTM can prevent gradient explosion during learning. However, for existing LSTM networks, the input time series data must be complete. However, this requirement is not realistic in traffic data scenarios. Therefore, the embodiment of the present invention reconstructs the existing LSTM network, and proposes LSTM-I, which is specially used for processing the input with missing data points.

表示不完整的观测，

是预测值。当输入的

包含缺失数据时，本发明实施例将

和

拼接为当前输入。具体来说，重构后的LSTM-I使用前一个时间段的预测值来估算当前时间段的缺失值。对于每个时间段，LSTM-I采用估计值和观察值来复原特征，如公式(3)所示：As shown in Figure 11, at time point t,

represents an incomplete observation,

is the predicted value. when input

When missing data is included, this embodiment of the present invention will

and

Concatenate as the current input. Specifically, the reconstructed LSTM-I uses the predicted values of the previous time period to estimate the missing values of the current time period. For each time period, LSTM-I uses the estimated and observed values to restore the features, as shown in Equation (3):

表示第t个时间步的缺失位置，⊙表示Hadamard积。in

denotes the missing position at the t-th time step, and ⊙ denotes the Hadamard product.

其中d是LSTM单元输出的隐藏特征的尺寸。LSTM单元的整个计算过程如公式(4)所示:For each layer of LSTM-I, LSTM units with shared parameters for each layer are used for computation. For the t-th LSTMC cell, the input includes the cell state c _t-1 from the previous time point, the hidden feature h _t-1 , and the input h _t-1 . There are three kinds of gating units in the _LSTM cell, namely the input gate it, the forget gate ft, and the output gate _{ot ,} which are used to decide whether to add/remove information to the cell state. These gates adaptively save the input information to the current memory state and compute hidden features

where d is the dimension of the hidden features output by the LSTM unit. The entire calculation process of the LSTM unit is shown in formula (4):

为单元输入激活向量，

和

分别为单元的权重矩阵和偏置参数；

分别表示输入、遗忘和输出门的权重矩阵。

是相应门的偏置参数；

和

分别是记忆单元的权重矩阵和偏置矩阵；σ表示sigmoid激活函数。此外，在下一时间步的预测值

是根据隐藏特征h_t计算的，即

其中

和

分别为权重和偏置矩阵。in

input activation vector for the cell,

and

are the weight matrix and bias parameter of the unit, respectively;

represent the weight matrices of the input, forget, and output gates, respectively.

is the bias parameter of the corresponding gate;

and

are the weight matrix and bias matrix of the memory unit, respectively; σ represents the sigmoid activation function. Additionally, the predicted value at the next time step

is calculated from the hidden feature _ht , i.e.

in

and

are the weight and bias matrices, respectively.

和历史平均X^(a)的特征。多头操作可以分别在几个子空间进行学习，再将所有得到的结果结合起来，通过全连接层的降维，并输出与输入时间序列相同形状的输出矩阵。In a specific application scenario, although LSTM-I can gradually estimate missing data, its ability to capture time-series dependencies for longer time-series data is limited, especially when traffic data often has hundreds of time periods in a day. Therefore, the embodiment of the present invention applies a multi-head attention mechanism to further extract the output of LSTM-I

and the characteristics of the historical average X ^(a) . The multi-head operation can be learned separately in several subspaces, and then all the obtained results are combined, through the dimensionality reduction of the fully connected layer, and output an output matrix with the same shape as the input time series.

(对应于

和X^(a)的转置)。在计算过程中，有三个定义的组成部分，即Query Q、Key K和Value V，它们的定义如公式(5)所示:For the calculation of the attention mechanism, the embodiment of the present invention defines the time series input as

(corresponds to

and the transpose of X ^(a) ). In the calculation process, there are three defined components, namely Query Q, Key K and Value V, and their definitions are shown in formula (5):

和

分别为相应部分的参数矩阵。注意力矩阵E的计算过程如公式(6)所示：in

and

are the parameter matrices of the corresponding parts, respectively. The calculation process of the attention matrix E is shown in formula (6):

Here, the attention matrix is converted into a probability matrix using the softmax function, and the sum of the probabilities of all columns is 1. Therefore, _Eij represents the influence of the ith time point on the jth time point. The output of the attention mechanism is denoted by Z, and the calculation is shown in Equation (7):

For the computation of the multi-head attention mechanism, following the paradigm, let multiple attention mechanisms calculate their respective outputs Z _i ,i..H, where H is the number of attention heads. This allows separate learning in different attention subspaces derived from the above equation, which can capture richer feature relationships. Finally, the embodiment of the present invention outputs all Z _i to the linear layer to obtain the final result, which is expressed as formula (8):

Output=concat(Z ₁ ,...,Z _h )W _c formula (8)

where Output is the final output of MSL, corresponding to hidden states H and H ^* , and _Wc is the parameter weight of the linear layer.

In a specific application scenario, in order to combine the outputs of the two modules and calculate the final completion result, an embodiment of the present invention designs a fusion module, which includes an attention layer and a linear layer. The attention layer is a single-headed version of MSL, whose input is the concatenation of two implicit states. After the attention layer is calculated, the embodiment of the present invention uses a fully connected neural network in the linear layer to obtain the completion result, and the calculation is as shown in formula (9):

是ADRIN的最终补全结果。where attention( ) represents the formulas (3)-(5) introduced in the multi-head attention mechanism, W _l and b _l are the parameters of the linear layer,

is the final completion result of ADRIN.

和估算的输出

本发明实施例定义了掩蔽损失函数

其表述如公式(10)所示：In a specific application scenario, in order to better train different modules of ADRIN, the embodiment of the present invention defines respective loss functions for the LSTM-I, the fusion module and the final output. Taking into account ground truth road speed data

and the estimated output

The embodiment of the present invention defines a masking loss function

Its expression is shown in formula (10):

是上述实施例定义的掩码矩阵，用于表示丢失的数据点。in,

is the mask matrix defined in the above embodiment, used to represent the missing data points.

(即LSTM-I的输出)与X^(？)相似，并加快LSTM-I的收敛速度，本发明实施例将损失函数

定义如公式(11)所示:to make sure

(that is, the output of LSTM-I) is similar to X ^(?) , and to speed up the convergence speed of LSTM-I, the embodiment of the present invention uses the loss function

The definition is shown in formula (11):

它的表述如公式(12) 所示：Combining the above two sub-loss functions, the embodiment of the present invention obtains the final loss function

It is expressed as Equation (12):

In a specific application scenario, the embodiment of the present invention adopts three real-world traffic speed data sets, namely NavInfo-Beijing: Beijing, China (BJ), PeMS: California Highway Performance Evaluation System (PeMSD5), and Hong Kong Traffic Speed (HK). Specifically, as shown in Figure 12, the Beijing China dataset is provided by the 4D Traffic Index platform, which contains the average of 1368 roads in Beijing, China during the period from 00:00 on January 1, 2019 to 23:55 on June 30, 2019 speed. The recorded sampling interval was 5 minutes. In order to minimize the impact of missing data while preserving the complexity of the dataset, the embodiment of the present invention only uses road speed data with an overall data missing rate less than 5% for experiments, that is, a total of 168 roads. Furthermore, embodiments of the present invention apply linear interpolation to supplement missing data and record their locations, which are removed during the evaluation phase. As shown in Figure 13, the PeMSD5 dataset contains traffic speed data collected from the California Department of Transportation measurement system. The dataset includes 144 sensor stations in the Fifth District of California, USA, recorded from 00:00 on January 1, 2013 to 23:55 on June 30, 2013. It is worth mentioning that the PeMSD5 records are collected by sensors on highways, which are different from the average traffic speeds of urban roads in the Beijing, China and Hong Kong, China datasets. As shown in Figure 14, the Hong Kong, China dataset includes the average road speeds of major roads in Hong Kong, China from 00:00 on March 10, 2021 to 23:50 on July 31, 2021. Records from remote areas such as Tuen Mun and Sha Tin have remained unchanged for a long time. Therefore, the embodiment of the present invention only adopts the road speed in Hong Kong Island, China, which is composed of 84 roads, and the sampling interval is 10 minutes. A summary of these datasets is presented in Table 1 below. The collection locations of some roads or sensors in these three datasets are described.

Table 1

BJ PeMSD5 HK Number of road sections 168 144 84 days 181 181 144 time interval 5min 5min 10min average speed 36.70km/h 54.47mi/h 51.24km/h standard deviation 11.52km/h 7.32mi/h 20.19km/h

In the present invention, Z-score normalization is used to preprocess the data. In terms of cross-validation, the present invention follows the previous work method and divides the three data sets into two non-overlapping subsets in time order, namely training set and test set: the first 80% of all samples are training data, and the remaining 20% for test data. In addition, the method of data augmentation is adopted in the training phase. For each sample Y in the training set, the present invention randomly generates 10 missing inputs X ^(?) . Use Adam as the optimizer with an initial learning rate of 0.001. The number of training epochs is set to 200 and the batch size is 20. The number of heads H in the multi-head attention mechanism is set to 8, and the feature dimension d in LSTM-I is set to 168. PyTorch used for experiments, hardware configuration includes nVidia RTX 2080Ti GPU and Xeon Silver 4210 CPU.

In a specific application scenario, in order to fully evaluate the model in this case study, the embodiment of the present invention conducts experiments on two missing modes, namely MCAR and MNAR, respectively. In addition, in order to verify the validity of the model in different situations, the embodiments of the present invention compare ADRIN with existing methods under different data missing rates ranging from 10% to 90%. In addition, the adopted methods come from various completion methods of the aforementioned noun parsing, some of which are considered to be the current advanced missing data completion methods.

Among these methods, matrix factorization-based methods (such as TRMF and BTMF) must strictly guarantee the input format shape is day × time × road. In addition, the embodiments of the present invention uniformly set the input of deep learning methods such as DSAE, PGAN, BRITS, and ADRIN as missing data of one day, that is, time×road. In addition, since it is difficult to guarantee the convergence of PGAN, the embodiment of the present invention sets the learning rate of the generator and the discriminator in {0.00001, 0.0001, 0.001, 0.01}, and applies grid search to obtain the best results. The hyperparameters of the other models remain the same.

Table 1 is the completion accuracy rate (MAE/MAPE (%)) of the Beijing China (BJ) dataset, and Table 2 is the completion accuracy rate (MAE/MAPE (%)) of the California Fifth District (PeMSD) dataset. 3 is the completion accuracy rate (MAE/MAPE (%)) of the Hong Kong (HK) data set, the embodiment of the present invention selects the root mean square error (RMSE) and the mean absolute percentage error (MAPE) as the indicators of the embodiment of the present invention . The above three tables summarize the completion results for MCAR and MNAR, respectively, and the missing rates for these three datasets are between 10% and 90%. Experimental results show that the proposed ADRIN achieves the lowest RMSE and MAPE values in most cases, outperforming all other methods. On complex datasets, such as urban road speed data Beijing, China and Hong Kong, China, the results are outstanding, while on the simpler highway speed California District 5 dataset, ADRIN can achieve results close to SOTA on MCAR . This is due to the fact that ADRIN can extract temporal dependencies from crippled inputs and historical averaging. In particular, statistical learning-based methods such as HA and KNN produce significantly worse completion results than ADRIN. This is because such methods only rely on simple sample distributions within the dataset to supplement missing values, and temporal correlations cannot be extracted, resulting in poor performance. Furthermore, when there are large-scale missing data, some missing points make the KNN method difficult to apply efficiently.

Table 2

table 3

Table 4

Considering deep learning based methods, the proposed ADRIN outperforms state-of-the-art DSAE, PGAN and BRITS. Due to its superior ability to separately model the time-series dependencies of missing input and historical data, ADRIN achieves significant improvements over other methods, even with high missing rates, regardless of whether the missing data is completely random or not. For example, compared to other deep learning-based methods on the Beijing, China dataset, such as BRITS, in both cases, ADRIN reduces MAPE by 7.16% and 7.77% at 90% miss rate, while at 10% The MAPE results were reduced by 6.72% and 7.64%, respectively, under the missing rate of . This shows that ADRIN can achieve a more significant improvement in the case of high missing rate than in the case of low due to the combination of historical features. The same phenomenon can also be observed on the Hong Kong dataset. Furthermore, since DASE and PGAN only use linear modules to model features, they can only extrapolate acceptable results on the California District 5 dataset with a simple data distribution. However, the proposed ADRIN takes into account the temporal dependencies and can deal with the missing data of urban roads and highways simultaneously.

Different from methods based on matrix factorization, deep learning methods have more advantages for modeling complex data such as urban road speed, and these models achieve better completion effects on the whole. However, deep learning models are highly dependent on data quality, and high missing rates make it difficult to capture the correlation of temporal features. Taking MCAR as an example, BTMF achieves the best performance on 90% of the California District 5 dataset because the data distribution of highway speeds is simpler and easier to learn than urban roads. In addition, matrix factorization-based methods have stable results across all missing rates. This is because these methods estimate missing values from the overall data distribution. For example, on the Beijing, China dataset, the absolute difference in the MAPE results for ADRIN is 0.60% (MCAR) and 0.43% (MNAR) for all missing rate cases, but 0.21% (MCAR) and 0.26% (MNAR) for BGCP. ).

Finally, comparing the completion results of MCAR and MNAR, it can be seen that all methods have lower MAPE under MCAR than MNAR. Modeling the feature dependencies of the data in MNAR is challenging due to the continuous large batch of missing blocks, resulting in the corresponding performance of MNAR being worse than MCAR. That is, when the missing rates are the same, all models perform better under MCAR because the distribution of observations is more spread out and thus more accurately represents the overall data. However, in the case of MNAR, ADRIN can achieve more pronounced attribution results in the case of high missing rates compared to DSAE and BRITS because ADRIN uses historical average data.

The embodiment of the present invention also provides a data completion device, which can implement the above data completion method, and the device includes:

The original data set acquisition module is used to obtain the original data set;

The original dataset preprocessing module is used to perform data preprocessing on the original dataset to classify the original dataset into incomplete datasets and historical datasets;

The first time series extraction module is used to extract the time feature of the incomplete data set to obtain the first time series;

The second time series extraction module is used to extract the time feature of the historical data set to obtain the second time series;

a first output matrix calculation module, configured to perform multi-head attention mechanism calculation on the first time series to obtain a first output matrix;

The second output matrix calculation module is used to perform multi-head attention mechanism calculation on the second time series to obtain a second output matrix;

The data fusion module is used to perform fusion processing on the first output matrix and the second output matrix to obtain complementary data.

The specific implementation of the data completion apparatus in this embodiment is basically the same as the specific implementation of the above-mentioned data completion method, and will not be repeated here.

Embodiments of the present disclosure also provide an electronic device, including:

at least one memory;

at least one processor;

at least one program;

The program is stored in the memory, and the processor executes at least one program to implement the above-mentioned data completion method of the present invention. The electronic device may be any intelligent terminal including a mobile phone, a tablet computer, a personal digital assistant (Personal Digital Assistant, PDA for short), a vehicle-mounted computer, and the like.

Please refer to FIG. 15. FIG. 15 illustrates a hardware structure of an electronic device according to another embodiment. The electronic device includes:

The processor 1501 can be implemented by a general-purpose CPU (Central Processing Unit, central processing unit), a microprocessor, an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), or one or more integrated circuits, etc., for executing relevant programs to achieve The technical solutions provided by the embodiments of the present invention;

The memory 1502 may be implemented in the form of a ROM (ReadOnly Memory, read only memory), a static storage device, a dynamic storage device, or a RAM (Random Access Memory, random access memory). The memory 1502 may store an operating system and other application programs. When implementing the technical solutions provided by the embodiments of the present specification through software or firmware, the relevant program codes are stored in the memory 1502 and invoked by the processor 1501 to execute the implementation of the present disclosure. Example data completion method;

Input/output interface 1503, used to realize information input and output;

The communication interface 1504 is used to realize the communication interaction between the device and other devices, and the communication can be realized by wired means (such as USB, network cable, etc.), or by wireless means (such as mobile network, WIFI, Bluetooth, etc.);

A bus 1505 that transfers information between the various components of the device (eg, processor 1501, memory 1502, input/output interface 1503, and communication interface 1504);

The processor 1501 , the memory 1502 , the input/output interface 1503 and the communication interface 1504 realize the communication connection among each other within the device through the bus 1505 .

Embodiments of the present disclosure further provide a storage medium, where the storage medium is a computer-readable storage medium, and the computer-readable storage medium stores computer-executable instructions, where the computer-executable instructions are used to cause a computer to execute the above data completion method .

As a non-transitory computer-readable storage medium, the memory can be used to store non-transitory software programs and non-transitory computer-executable programs. Additionally, the memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid state storage device. In some embodiments, the memory may optionally include memory located remotely from the processor, which may be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

The embodiments described in the embodiments of the present invention are for the purpose of illustrating the technical solutions of the embodiments of the present invention more clearly, and do not constitute a limitation on the technical solutions provided by the embodiments of the present invention. With the emergence of application scenarios, the technical solutions provided by the embodiments of the present invention are also applicable to similar technical problems.

Those skilled in the art can understand that the technical solutions shown in FIG. 4 to FIG. 10 do not constitute a limitation on the embodiments of the present invention, and may include more or less steps than those shown in the drawings, or combine certain steps, or different steps.

The apparatus embodiments described above are only illustrative, and the units described as separate components may or may not be physically separated, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment.

Those of ordinary skill in the art can understand that all or some of the steps in the methods disclosed above, functional modules/units in the systems, and devices can be implemented as software, firmware, hardware, and appropriate combinations thereof.

The terms "first", "second", "third", "fourth", etc. (if any) in the description of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It is to be understood that data so used may be interchanged under appropriate circumstances so that the embodiments of the application described herein can be practiced in sequences other than those illustrated or described herein. Furthermore, the terms "comprising" and "having", and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those expressly listed Rather, those steps or units may include other steps or units not expressly listed or inherent to these processes, methods, products or devices.

It should be understood that, in this application, "at least one (item)" refers to one or more, and "a plurality" refers to two or more. "And/or" is used to describe the relationship between related objects, indicating that there can be three kinds of relationships, for example, "A and/or B" can mean: only A, only B, and both A and B exist , where A and B can be singular or plural. The character "/" generally indicates that the associated objects are an "or" relationship. "At least one item(s) below" or similar expressions thereof refer to any combination of these items, including any combination of single item(s) or plural items(s). For example, at least one (a) of a, b or c, can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c" ", where a, b, c can be single or multiple.

In the several embodiments provided in this application, it should be understood that the disclosed apparatus and method may be implemented in other manners. For example, the apparatus embodiments described above are only illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components may be combined or Can be integrated into another system, or some features can be ignored, or not implemented. On the other hand, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, and may be in electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution in this embodiment.

In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware, or may be implemented in the form of software functional units.

The integrated unit, if implemented in the form of a software functional unit and sold or used as an independent product, may be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of software products in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, and the computer software products are stored in a storage medium , including multiple instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, Read-Only Memory (ROM for short), Random Access Memory (RAM for short), magnetic disk or CD, etc. that can store programs medium.

The preferred embodiments of the embodiments of the present invention have been described above with reference to the accompanying drawings, which are not intended to limit the scope of the rights of the embodiments of the present invention. Any modifications, equivalent replacements, and improvements made by those skilled in the art without departing from the scope and spirit of the embodiments of the present invention shall fall within the scope of the rights of the embodiments of the present invention.

Claims (10)

1. A data completion method, comprising:

acquiring an original data set;

performing data preprocessing on the original data set to classify the original data set into a incomplete data set and a historical data set;

extracting time characteristics of the incomplete data set to obtain a first time sequence;

extracting time characteristics of the historical data set to obtain a second time sequence;

performing multi-head attention mechanism calculation on the first time sequence to obtain a first output matrix;

performing multi-head attention mechanism calculation on the second time sequence to obtain a second output matrix;

and performing fusion processing on the first output matrix and the second output matrix to obtain the completion data.

2. The data completion method according to claim 1, wherein said performing temporal feature extraction on the incomplete data set to obtain a first time series comprises:

inputting the incomplete data set into a preset long-term and short-term memory completion network;

extracting observation data in the incomplete data set at a preset time period;

calculating missing data of the preset time interval according to the observation data and the prediction data of the previous time interval of the preset time interval;

and obtaining the first time sequence according to the missing data and the observation data.

3. The data completion method according to claim 1, wherein said performing temporal feature extraction on said historical data set to obtain a second time series comprises:

inputting the historical data set into a preset historical data processing network;

calculating historical average data of the historical data set through the historical data processing network, and taking the historical average data as the historical data of the preset time period;

and obtaining the second time sequence according to the historical average data.

4. The data completion method according to claim 1, wherein said performing a multi-point attention mechanism calculation on the first time series to obtain a first output matrix comprises:

inputting the first temporal sequence into a first multi-headed attention layer;

converting the first time series into a first attention matrix by the first multi-head attention layer;

converting the first attention moment matrix into a first probability matrix through a preset function;

performing attention mechanism calculation on the first probability matrix to obtain a first characteristic matrix;

and performing dimensionality reduction on the first feature matrix to obtain a first output matrix.

5. The data completion method according to claim 1, wherein said performing a multi-point attention mechanism calculation on the second time series to obtain a second output matrix comprises:

inputting the second time series to a second multi-headed attention layer;

converting, by the second multi-head attention layer, the second time series into a second attention matrix;

converting the second attention matrix into a second probability matrix through a preset function;

performing attention mechanism calculation on the second probability matrix to obtain a second feature matrix;

and performing dimensionality reduction on the second feature matrix to obtain a second output matrix.

6. The data completion method according to any one of claims 1 to 5, wherein the fusing the first output matrix and the second output matrix to obtain the completion data comprises:

splicing the first output matrix and the second output matrix to obtain a spliced matrix;

calculating an attention mechanism of the spliced matrix through a single-head attention layer to obtain a target matrix;

and performing feature extraction on the target matrix through a linear layer to obtain the completion data.

7. The data completion method according to claim 6, wherein the calculating an attention mechanism of the mosaic matrix through a single-head attention layer to obtain a target matrix comprises:

inputting the stitching matrix into a single-headed attention layer;

converting the mosaic matrix into an attention matrix by the single-headed attention layer;

converting the attention moment array into a probability matrix through a preset function;

and carrying out attention mechanism calculation on the probability matrix to obtain the target matrix.

8. A data complementing device, comprising:

the original data set acquisition module is used for acquiring an original data set;

the system comprises an original data set preprocessing module, a data preprocessing module and a data processing module, wherein the original data set preprocessing module is used for performing data preprocessing on the original data set so as to classify the original data set into a defective data set and a historical data set;

the first time sequence extraction module is used for extracting time characteristics of the incomplete data set to obtain a first time sequence;

the second time sequence extraction module is used for extracting time characteristics of the historical data set to obtain a second time sequence;

the first output matrix calculation module is used for performing multi-head attention mechanism calculation on the first time sequence to obtain a first output matrix;

the second output matrix calculation module is used for performing multi-head attention mechanism calculation on the second time sequence to obtain a second output matrix;

and the data fusion module is used for carrying out fusion processing on the first output matrix and the second output matrix to obtain the completion data.

9. An electronic device, comprising:

at least one memory;

at least one processor;

at least one program;

the programs are stored in a memory, and a processor executes the at least one program to implement:

the data completion method according to any one of claims 1 to 7.

10. A storage medium that is a computer-readable storage medium having stored thereon computer-executable instructions for causing a computer to perform:

the data completion method according to any one of claims 1 to 7.

Images