Joint prediction method of wireless signals and user trajectories in cellular networks based on multi-task learning
Through the multi-task LSTM model of multi-task learning, combined with user trajectory data and cellular network wireless signal data for joint prediction, the separate limitations of trajectory and signal prediction in the existing technology are solved, more accurate and timely prediction effects are achieved, and network resource management is optimized.
Patent Information
- Application Number
- CN202310005580.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-01-04
- Publication Date
- 2025-05-16
- Estimated Expiration
- 2043-01-04
AI Technical Summary
The prior art has limitations in trajectory prediction and signal strength prediction, and has failed to make full use of space-time environment information, and the scope of single-step prediction is limited, making it difficult to respond in a timely manner in an emergency situation.
Using multi-task learning method, a multi-task LSTM model is established, combined with user trajectory data and cellular network wireless signal data, and joint prediction is performed. The model includes two-layer LSTM network for position prediction and a single LSTM network for signal strength and signal quality prediction, and expand the prediction range through the local mean method.
Parallel prediction of user trajectory location and cellular network wireless signal strength is realized, the accuracy and timeliness of prediction are improved, the richness and generalization capabilities of the model are enhanced, and network resource management is optimized.
Smart Images
Figure CN116261158B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a method for jointly predicting cellular network wireless signals and user trajectories. Background Art
[0002] The trajectory data generated by mobile devices provides operators with a large amount of information about mobile users' driving behaviors and the ever-changing urban network operations, which lays the foundation for customized people-centric services. At present, the research on trajectory data mining mainly focuses on intelligent routing planning, network traffic prediction, etc. It is a time series generation task based on the historical user location to predict the future trajectory. The spatiotemporal environment information plays a key role in predicting the next location.
[0003] At present, the next location prediction using trajectory data has been widely studied. For example, the location of mobile users is predicted based on traditional Markov models and hidden Markov models, but human motion does not always satisfy the Markov assumption because it involves complex spatiotemporal interactions, and the explosive growth of data also affects the practical application capabilities of the model. In addition, there are also literature reports on the prediction of the location of mobile users based on traditional machine learning techniques such as Bayesian models, k-nearest neighbors, and decision trees. However, these methods are not suitable for trajectories composed of continuous coordinates with short sampling time intervals because the sampling points they require are discrete. In recent years, the emergence of deep learning has improved the ability to capture the spatiotemporal dependencies of large-scale trajectory data. For example, the spatiotemporal convolutional neural network (RNN) method predicts the next location by modeling local spatiotemporal environmental information, but when the time series is long, it is prone to gradient explosion or gradient disappearance problems. In order to solve the long-term dependency problem of RNN, a long short-term memory (LSTM) neural network is proposed for trajectory prediction. Although the trajectory prediction technology based on LSTM has achieved good results, several factors are still ignored:
[0004] 1) Current research mainly focuses on the prediction of location trajectory or the prediction of signal strength. The strong correlation between the two can be further explored.
[0005] 2) Making full use of spatiotemporal environmental information is crucial for location prediction. Joint modeling of multiple tasks in one model can enhance the richness and generalization ability of the model. Various cellular network indicators collected in urban traffic scenarios capture information about wireless networks and traffic conditions, and can indicate the behavior patterns of mobile users.
[0006] 3) The prediction range is limited by single-step prediction. In trajectory prediction in urban traffic scenarios, single-step prediction is more accurate but less timely than multi-step prediction, which is not conducive to timely response of users or base stations. It is difficult to effectively expand the prediction range, especially in emergency situations, to effectively avoid some accidents. Summary of the invention
[0007] The purpose of the present invention is to perform parallel prediction of the user trajectory position and the cellular network wireless signal strength in the next time period.
[0008] In order to achieve the above object, the technical solution of the present invention is to provide a method for jointly predicting cellular network wireless signals and user trajectories based on multi-task learning, which is characterized by comprising the following steps:
[0009] Step 1: Data collection:
[0010] Field testing was performed using a test mobile system to collect measurement data from an urban area served by a commercial cellular network;
[0011] Step 2: Acquire trajectory data based on the measurement data, where the trajectory data includes location information of each sampling point, a corresponding reference signal received power, and a corresponding reference signal received quality;
[0012] Step 3. Establish a multi-task LSTM model and use the multi-task LSTM model to realize the joint prediction of cellular network wireless signals and user trajectories. The multi-task LSTM model includes a two-layer LSTM network for processing the longitude and latitude sequence in the location information to realize location prediction, a single LSTM network for predicting the reference signal reception power, and a single LSTM network for predicting the reference signal reception quality.
[0013] Preferably, in step 2, the trajectory data is represented by T r ={p1, p2, ..., p n}, consisting of n sampling points, where p t Represents the location of the mobile user at time step t, including its location information x t , the corresponding reference signal received power P t and the corresponding reference signal reception quality Q t .
[0014] Preferably, in step 3, in each LSTM network: through the input gate i t Decide which data can be added to the memory cell; then, the state C of the memory cell at the previous moment t-1 With the forget gate f t The dot product determines the information to be retained; finally, the state C of the memory unit at the previous moment t-1 Combined with the retained information obtained, the state C of the memory unit at the current moment is obtained t , based on the current state of the memory unit C t Get the final output h t .
[0015] Preferably, in step 3, each LSTM network is represented as:
[0016] f t =σ(W f ·[h t-1 , X t ]+b f )
[0017] i t =σ(W i ·[h t-1 , X t ]+b i )
[0018]
[0019]
[0020] o t =σ(W o ·[h t-1 , X t ]+b0)
[0021] h t =o t *tanh(C t )
[0022] In the formula, f t represents the forget gate, i t represents the input gate, o t represents the output gate, W f and b f Represents the relevant weight matrix and bias, X t Indicates the input at the current moment, h t-1 represents the output of the LSTM network at the previous moment, σ() is the sigmoid function, is a new candidate vector created by the tanh function, W C and b C represents the weight and bias of the new candidate variable, C t Represents the current state of the memory unit, W o and b0 represent the weight and bias at the current moment.
[0023] Preferably, in step 3, when training the multi-task LSTM model, the mean square error is used as the loss function to describe the loss of the prediction task.
[0024] Preferably, in step 3, when evaluating the multi-task LSTM model, the Haversian formula is used to calculate the distance error between the expected position and the actual position, and the multi-task LSTM model is evaluated based on the distance error.
[0025] Preferably, in step 3, when a multi-task LSTM model is used to realize the joint prediction of cellular network wireless signals and user trajectories, the local mean method is used to expand the prediction range: a fixed-size local window and a sliding step are set, and the original trajectory data is processed using the fixed-size local window and the sliding step to predict the user's position at the next moment, and the average reference signal reception power and the average reference signal reception quality of the user moving to the next position within a certain sliding step are predicted.
[0026] While providing location prediction, the present invention can predict the signal strength and signal quality within the current range, which is beneficial to optimizing network resource management. The prediction input is based on real measured data, combining historical trajectories with corresponding RSRP and RSRQ to enrich positioning features and improve accuracy. Compared with baseline methods such as hidden Markov model (HMM) and Kalman filter (KF), this model has higher prediction accuracy in both positioning and signal power prediction tasks. In addition, the present invention preprocesses the raw data based on the local mean (LM) to increase the range of the next location prediction. The trade-off between prediction accuracy and prediction range is carefully evaluated at different scales to facilitate the application of multi-task prediction models in actual traffic scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 A schematic diagram of the measurement activities of the method shown in the present invention;
[0028] Figure 2 A model architecture diagram of the method shown in the present invention;
[0029] Figure 3 It is a result diagram obtained by the method shown in the present invention on the embodiment;
[0030] Figure 4 The schematic diagram of the method shown in the present invention. DETAILED DESCRIPTION
[0031] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.
[0032] The present invention uses a large number of measurement reports (MRs) of mobile user trajectories and wireless network performance collected from actual urban scenarios to construct an LSTM sequence that can predict signal power while providing location prediction, which is beneficial for optimizing network resource management. The patterns and dependencies of the location and cellular cell signal pattern sequences in MRs are captured for joint prediction to improve performance.
[0033] The present invention provides a method for jointly predicting cellular network wireless signals and user trajectories based on multi-task learning, comprising the following steps:
[0034] 1) Data collection.
[0035] A test mobile system was used to perform drive tests to collect measurement data from an urban area served by a commercial cellular network. A schematic diagram of the measurement campaign and the associated RSRP of the sampling points is shown in Figure 1. Figure 1 As shown in the figure, the TEMS log file contains the raw data collected in the network. After conversion, there are 29 columns of information in the MR sample. Key indicators include the sampling timestamp, the latitude and longitude of the sampling point, the cell identifier (ECID), the RSRP and RSRQ of the serving cell and its six neighboring cells.
[0036] 2) Obtain trajectory data.
[0037] Assume T r ={p1, p2, ..., p n} represents the urban trajectory data composed of n sampling points, where p t Represents the location of the mobile user at time step t, including its location information x t (longitude and latitude coordinates) and the corresponding reference signal received power (RSRP) P t and reference signal received quality (RSRQ)Q t .
[0038] 3) Establish a multi-task LSTM model.
[0039] Multi-task model architecture such as Figure 2 As shown in the figure, a trajectory training model of 700 sampling points is selected, a two-layer LSTM network is constructed to process the longitude and latitude sequence for location prediction, and a single LSTM is added for RSRP / RSRQ prediction. Each layer of the LSTM network has 256 units, and the dropout layer is set to 0.3 for multiple prediction outputs. The rectified linear unit (ReLU) is used to solve the gradient disappearance problem, and the adaptive moment estimation (Adam) is applied to optimize the model parameters. The relevant calculations are as follows:
[0040] f t =σ(W f ·[h t-1 , X t]+b f )
[0041] i t =σ(W i ·[h t-1 , X t ]+b i )
[0042]
[0043]
[0044] o t =σ(W o ·[h t-1 , X t ]+b0)
[0045] h t =o t *tanh(C t )
[0046] In the formula, f t represents the forget gate, i t represents the input gate, o t represents the output gate, W f and b f Represents the relevant weight matrix and bias, X t Indicates the input at the current moment, h t-1 represents the output of the LSTM network at the previous moment, σ() is the sigmoid function, is a new candidate vector created by the tanh function, W C and b C represents the weight and bias of the new candidate variable, C t Represents the current state of the memory unit, W o and b0 represent the weight and bias at the current moment. First, the input gate i t Decide which data can be added to the memory cell. Then, t-1 The state and forget gate f t The dot product determines the information retained. Finally, the two parts are combined to get the new memory cell state C t ,h t is the final output.
[0047] 4) Use mean square error (MSE) as the loss function to describe the loss of the prediction task, as shown in the following formula:
[0048]
[0049] in, represents the loss, Xi represents the i-th predicted value, represents the i-th label.
[0050] 5) Method evaluation:
[0051] The Haversing formula is used to calculate the distance error between the expected position and the actual position, as shown below:
[0052]
[0053] Among them, a is the latitude difference between the expected position and the actual position, b is the longitude difference between the expected position and the actual position, lat1 and lat2 are the latitudes of the expected position and the actual position respectively, and r is the radius of the earth, which is approximately 6371 km.
[0054] The prediction results are as follows:
[0055]
[0056] Table 1: Comparison of multi-task model training results
[0057] 6) Several widely used methods in time series prediction are selected for comparison with the method proposed in this invention, including linear regression, support vector regression (SVR), hidden Markov model (HMM) and Kalman filter (KF). The distance errors of different single-step prediction methods are shown in Figure 2. Figure 3 As shown, it can be intuitively seen that compared with the traditional model, the LSTM prediction model proposed in the present invention obtains better location prediction results.
[0058] 7) Use the local mean method to expand the prediction range, such as Figure 4 As shown, this can improve the timeliness of the prediction and make the prediction process more in line with actual application situations. The window size and sliding step size are both set to 5 to process the original trajectory data and predict the position after 1 second. In addition, the average signal power of the vehicle moving to the next position within 5 steps can also be predicted. The results are shown in Table 2 below. Although overall, compared with the single-step prediction, the local mean reduces the accuracy of the prediction of the three tasks, and the overall model accuracy decreases. However, in terms of practical applications, it makes sense to sacrifice a little prediction accuracy in exchange for an expanded prediction range.
[0059]
[0060] Table 2: Comparison between single-step prediction and local mean method
Claims
1. A method for joint prediction of cellular network wireless signals and user trajectories based on multi-task learning, characterized in that: The following steps are involved: Step 1: Data collection: Field testing was performed using a test mobile system to collect measurement data from an urban area served by a commercial cellular network; Step 2: Acquire trajectory data based on the measurement data, where the trajectory data includes location information of each sampling point, a corresponding reference signal received power, and a corresponding reference signal received quality; Step 3. Establish a multi-task LSTM model and use the multi-task LSTM model to realize the joint prediction of cellular network wireless signals and user trajectories. The multi-task LSTM model includes a two-layer LSTM network for processing the longitude and latitude sequence in the location information to realize location prediction, a single LSTM network for predicting the reference signal reception power, and a single LSTM network for predicting the reference signal reception quality.
2. The method for jointly predicting cellular network wireless signals and user trajectories based on multi-task learning as claimed in claim 1, characterized in that: In step 2, the trajectory data is represented by T r ={p1,p2,…,p n }, consisting of n sampling points, where p t Represents the location of the mobile user at time step t, including its location information x t , the corresponding reference signal received power P t and the corresponding reference signal reception quality Q t .
3. The method for jointly predicting cellular network wireless signals and user trajectories based on multi-task learning as claimed in claim 1, characterized in that: In step 3, in each LSTM network: through the input gate i t Decide which data can be added to the memory cell; then, the state C of the memory cell at the previous moment t-1 With the forget gate f t The dot product determines the information to be retained; finally, the state C of the memory unit at the previous moment t-1 Combined with the retained information obtained, the state C of the memory unit at the current moment is obtained t , based on the current state of the memory unit C t Get the final output h t .
4. The method for jointly predicting cellular network wireless signals and user trajectories based on multi-task learning as claimed in claim 3, characterized in that: In step 3, each LSTM network is represented as: f t =σ(W f ·[h t-1 ,X t ]+b f ) i t =σ(W i ·[h t-1 ,X t ]+b i ) o t =σ(W o ·[h t-1 ,X t ]+b0) h t =o t *tanh(C t ) In the formula, f t represents the forget gate, i t represents the input gate, o t represents the output gate, W f and b f Represents the relevant weight matrix and bias, X t Indicates the input at the current moment, h t-1 represents the output of the LSTM network at the previous moment, σ() is the sigmoid function, is a new candidate vector created by the tanh function, W C and b C represents the weight and bias of the new candidate variable, C t Represents the current state of the memory unit, W o and b0 represent the weight and bias at the current moment.
5. The method for jointly predicting cellular network wireless signals and user trajectories based on multi-task learning as claimed in claim 1, characterized in that: In step 3, when training the multi-task LSTM model, the mean square error is used as the loss function to describe the loss of the prediction task.
6. The method for jointly predicting cellular network wireless signals and user trajectories based on multi-task learning as claimed in claim 1, characterized in that: In step 3, when evaluating the multi-task LSTM model, the Haversine formula is used to calculate the distance error between the expected position and the actual position, and the multi-task LSTM model is evaluated based on the distance error.
7. The method for jointly predicting cellular network wireless signals and user trajectories based on multi-task learning as claimed in claim 1, characterized in that: In step 3, when the multi-task LSTM model is used to realize the joint prediction of cellular network wireless signals and user trajectories, the local mean method is used to expand the prediction range: a fixed-size local window and sliding step are set, and the fixed-size local window and sliding step are used to process the original trajectory data, and the user's position at the next moment is predicted, and the average reference signal reception power and average reference signal reception quality of the user moving to the next position within a certain sliding step are predicted.