A link outage prediction method and system based on multi-model fusion

Abstract

The present invention proposes a link interruption prediction method and system based on multi-model fusion. The present invention uses the Mallat algorithm and Symlet basic wavelet function to decompose the trend component and detail component of the collected reference signal power; uses the autoregressive moving average model to predict the decomposed trend component; uses the LSSVM model optimized by the improved whale optimization algorithm to predict the decomposed detail component; uses the prediction results of the trend component and the detail component to obtain the final prediction result in a linear weighted manner; uses the prediction result to calculate the predicted signal-to-noise ratio, and then determines whether to trigger a link interruption declaration. The present invention solves the problem of using a single prediction model and improves the RSRP prediction accuracy and RLF prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the field of communication network technology, and in particular to a link outage prediction method and system based on multi-model fusion. Background Art

[0002] In the context of the integration of space and earth networks, enabling mobile phones to obtain communication services through direct satellite connections is gradually becoming a new development direction and focus. Mobile phones that can directly connect to satellites can communicate with the outside world through both ground mobile communication systems and satellite mobile communication systems, and can send and receive signals to and from satellites without the need for signal relay equipment [1].

[0003] In the scenario of mobile phones directly connected to satellites, mobile phones need to be able to access not only multiple heterogeneous networks on the ground, but also satellite communication networks. The mobile phone direct connection to satellite function enables on-demand switching between satellite and terrestrial communication networks using a unified terminal, that is, heterogeneous network fusion, while having the advantages of both satellite and terrestrial communication networks. The process of switching between different types of communication networks is called vertical switching [2]. Since vertical switching technology is an important technical guarantee for realizing the fusion of satellite and terrestrial heterogeneous networks, it is of great significance to study vertical switching technology for mobile phone direct connection to satellite scenarios.

[0004] Vertical handoff algorithms are crucial in vertical handoffs, determining the target network for users when performing vertical handoffs. They impact the seamlessness of the entire vertical handoff process, as well as the Quality of Service (QoS) and user experience. Currently, a wealth of research has been conducted on vertical handoff algorithms, both domestically and internationally, and they can be broadly categorized into the following types.

[0005] The vertical handover algorithm originally proposed makes decisions based on the received signal strength (RSS) from each base station [3]. The user equipment (UE) measures the RSS of each candidate network. If the RSS of a candidate network is greater than the RSS of the currently accessed network, the handover process is executed. This algorithm has low complexity and is simple to implement, but it has obvious disadvantages. When the RSS fluctuates near the handover threshold, a serious ping-pong effect will occur, resulting in a large consumption of handover resources. To address this problem, Liu M et al. [4] introduced a dwell time to the RSS-based handover algorithm, that is, after deciding to perform the handover, they wait for a period of time. If the RSS is still higher than the threshold, the vertical handover process is executed. Ahuja K et al. [5] improved the RSS-based handover algorithm using user distance and interruption probability. The algorithm first estimates the distance of the overlapping area and then selects the best candidate network based on the average RSS. The RSS-based algorithm only considers a single performance and ignores other network parameters and user subjective needs, so it has significant limitations.

[0006] The vertical handover algorithm based on multiple attribute decision making (MADM) evaluates candidate networks by comprehensively considering multiple attributes, such as RSS, delay, bandwidth, cost and terminal energy consumption, and then selects the best candidate network as the target network for vertical handover through the MADM method. In order to determine the impact of weight calculation methods on vertical handover performance, Yadav AK et al. [6] studied the impact of various weight calculation methods on the approximate ideal solution sorting method, multi-criteria optimization and compromise solution method and grey correlation analysis. The results show that the approximate ideal solution sorting method works better when combined with the entropy weight method, while the latter two methods work better when combined with the improved CRITIC method. Chen H et al. [7] proposed a vertical handover algorithm that combines the hierarchical analysis method and the entropy weight method, constructing a utility function based on the combination of subjective weights and objective weights, and determining the optimal access network based on the function value. Almutairi AF et al. [8] used particle swarm optimization technology to improve the multi-attribute weight calculation method, generating the optimal weight through two different optimization functions, thereby ensuring vertical handover performance. Zhong Y et al. [9] proposed a network selection algorithm based on fuzzy analysis hierarchical process, standard deviation method, and gray relational analysis method. The fuzzy analysis hierarchical process and standard deviation method are used to calculate subjective weights and objective weights respectively, while the gray relational analysis method is used to sort candidate networks. Satapathy P et al.

[10] divided the vertical switching algorithm into two steps. First, a multi-objective optimization problem of network quality and switching cost is constructed to determine whether vertical switching is required. Then, the fuzzy analysis hierarchical process and the approximate ideal solution sorting method are used to determine the target network. Ding Qi

[11] proposed an improved MADM algorithm to improve the robustness and robustness of the algorithm by citing fuzzy theory when selecting network parameters in the case of unstable signals in the Internet of Vehicles environment.

[0007] The vertical switching algorithm based on artificial intelligence uses existing data to train mathematical models so that it can make reasonable vertical switching decisions and adapt to environmental changes and change its own performance. Zhao P et al.

[12] designed a vertical switching algorithm based on hierarchical fuzzy reasoning for different attributes and fuzzy requirements described in natural language, providing differentiated services for different users, and reducing the signaling overhead of collecting the values ​​of each decision attribute through software-defined network architecture. Hussain SM et al.

[13] proposed a vertical switching algorithm based on dynamic Q learning and fuzzy convolutional neural network for the vehicle communication environment. The algorithm determines fuzzy rules by considering parameters such as signal strength, distance, and vehicle density. Pramod Kumar P et al.

[14] proposed a vertical switching algorithm based on fuzzy graph neural network, which reduces the impact of non-line-of-sight conditions on target network decisions. Wu Liping et al.

[15] used self-organizing network technology to obtain network indicator data in real time, and introduced a dynamic fuzzy neural network with a dynamically adjusted rule base for vertical switching decisions, effectively reducing the switching failure rate in ultra-dense heterogeneous wireless networks. Yang Xiaoping et al.

[16] proposed a vertical handover algorithm based on double-delay deep deterministic policy gradient, which determines the best candidate network by considering four indicators: packet loss rate, delay, delay jitter, and bit error rate, and solving the optimal solution problem of reinforcement learning. Verma J et al.

[17] constructed a vertical handover decision model based on fuzzy logic decision-making, making decisions based on attributes such as bandwidth and cost. Wu J et al.

[18] proposed a vertical handover strategy based on the Sarsa-λ reinforcement learning algorithm for the mixed visible light and infrared communication scenario in the cabin, which effectively improved the downlink data transmission rate when the user is in a mobile state.

[0008] In scenarios where mobile phones are directly connected to satellites, satellite-to-ground links and terrestrial links are prone to sudden RLFs due to environmental factors. Existing vertical handover strategies mostly make handover decisions based on network metrics such as latency and bandwidth, directly executing the handover after calculating the target network. This approach does not fully consider the impact of sudden RLFs on the handover process. Therefore, the present invention improves the standard DAPS handover process to address the issue of sudden RLFs in scenarios where mobile phones are directly connected to satellites. To address the problem that sudden RLFs in links in scenarios where mobile phones are directly connected to satellites can easily lead to DAPS handover failures, the present invention designs a DAPS handover based on link interruption prediction, reducing the handover failure rate. This prediction method uses multiple prediction models for independent modeling and prediction, and improves the model parameter optimization algorithm through two strategies: population variation and low-discrepancy sequence initialization. Finally, the predicted values ​​are weighted and summed to obtain the final prediction result. Furthermore, to address the problem that standard A3 measurement events in DAPS handovers tend to trigger a large number of measurement reports, by integrating location information into the A3 measurement event, mobile phones in different locations have different A3 measurement event triggering conditions, thereby reducing unnecessary handovers and ping-pong handovers.

[0009] [1] Chen Hong. Development and thinking of global mobile phone direct satellite application [J]. Satellite Applications, 2023, (11): 37-43.

[0010] [2] Stanic I, Drajic D, Cica Z. Overview of network selection and vertical handover approaches and simulation tools in heterogeneous wireless networks [C] / / 2023 16th International Conference on Advanced Technologies, Systems and Services in Telecommunications (TELSIKS). IEEE, 2023: 133-142.

[0011] [3]Pahlavan K,Krishnamurthy P,Hatami A,et al.Handoff in hybridmobiledata networks[J].IEEE Personal Communications,2000,7(2):34-47.

[0012] [4]Liu M, Li Z, Guo X, et al.Performance analysis and optimization ofhandoff algorithms in heterogeneous wireless networks[J].IEEE Transactions onMobile Computing,2008,7(7):846-857.

[0013] [5]Ahuja K,Singh B,Khanna R.Network selection algorithm based on linkquality parameters for heterogeneous wireless networks[J].Optik,2014,125(14):3657-3662.

[0014] [6]Yadav A K,Singh K.The influence of different weighting methods onMADM ranking techniques and its impact on network selection for handover inhetnet[C] / / 2023International Conference on Artificial Intelligence and SmartCommunication(AISC).IEEE,2023:959-963.

[0015] [7]Chen H,Wang X,Zhou C.A vertical handoff algorithm for maritimecommunication in heterogeneous networks[C] / / 2023 15th InternationalConference on Communication Software and Networks(ICCSN).IEEE,2023:74-78.

[0016] [8]Almutairi A F,Al-Gharabally M,Salman A A.Particle swarmoptimization application for multiple attribute decision making in verticalhandover in heterogenous wireless networks[J].Journal of EngineeringResearch,2021,9(1):176.

[0017] [9]Zhong Y,Wang H,Lv H,et al.A vertical handoff decision scheme usingsubjective-objective weighting and grey relational analysis in cognitiveheterogeneous networks[J].Ad Hoc Networks,2022,134:102924.

[0018]

[10] Satapathy P,Mahapatro J.An efficient multicriteria-based verticalhandover decision-making algorithm for heterogeneous networks[J].Transactionson Emerging Teleco-mmunications Technologies,2022,33(4):e4409.

[0019]

[11] Ding Qi. Network switching technology and its application in the Internet of Vehicles environment[D]. Jiangsu: Nanjing University of Posts and Telecommunications, 2020.

[0020]

[12] Zhao P, Yu W, Yang

[0021]

[13] Hussain SM, Yusof KM, Yusof K M. Dynamic Q-learning and fuzzy CNNbased vertical handover decision for integration of DSRC, mmwave 5G and LTE ininternet of vehicles (IoV) [J]. J. Commun., 2021, 16(5): 155-166.

[0022]

[14] Pramod Kumar P,Sagar K.Reinforcement learning and neuro-fuzzyGNN-based vertical handover decision on internet of vehicles[J].Concurrencyand Computation:Practice and Experience, 2023,35(12):e7688.

[0023]

[15] Wu Liping, Wang Shuangshuang, Ma Bin. Vertical switching algorithm to improve user experience[J]. Journal of Electronics & Information Technology, 2022, 44(8): 2824-2832.

[0024]

[16] Yang Xiaoping, Liu Shui, Wang Xue, et al. Vertical handover method for heterogeneous wireless networks based on reinforcement learning TD3 algorithm [P]. Jilin Province: CN202111120444.5, 2022-09-23.

[0025]

[17] Verma J, Manhas P, Thakral S. Novel Fuzzy model for networks selection in heterogeneous network [C] / / 2023International Conference on Communication, Circuits, and Systems (IC3S). IEEE, 2023: 1-7.

[0026]

[18] Wu J,Han D,Zhang M,et al.Reinforcement learning adaptive verticalhandover scheme for hybrid VLC-IR networks in ship cabins[C] / / 2021 17thInternational Symposium on Wireless Communication Systems(ISWCS).IEEE,2021:1-5. Summary of the Invention

[0027] In order to solve the problem of using a single prediction model, the present invention proposes a link outage prediction method and system based on multi-model fusion to improve the RSRP prediction accuracy and RLF prediction accuracy.

[0028] The technical solution of the method of the present invention is a link outage prediction method based on multi-model fusion, and the specific steps are as follows:

[0029] Step 1: Use Mallat algorithm and Symlet basic wavelet function to decompose the collected reference signal power into trend component and detail component.

[0030] Step 2: Use the autoregressive moving average model to predict the decomposed trend component.

[0031] Step 3: Use the LSSVM model optimized by the improved whale optimization algorithm to predict the decomposed detail components.

[0032] Step 4: Using the prediction results of trend component and detail component, the final prediction result is obtained by linear weighting.

[0033] Step 5: Use the prediction result to calculate the predicted signal-to-noise ratio, and then determine whether to trigger a link interruption declaration.

[0034] Wherein, the maximum decomposition level J of the Mallat algorithm in step 1 is 3;

[0035] Wherein, the vanishing moment order of the Symlet basic wavelet function in step 1 is N=5;

[0036] The autoregressive moving average model in step 2 is:

[0037]

[0038] Among them, X(t) represents the time series value at time t, ε(t) represents the random error at time t, L represents the lag operator, and L m represents the lag operator of the mth term, α m represents the coefficient of the lag term m in the autoregressive model, β m It represents the coefficient of the lag term m in the moving average model. After X(t) is transformed into a stationary series through d-time difference, it is modeled through the p-order autoregressive model and the q-order moving average model. The error term ε(t) in AR(p) is expressed as an MA(q) model, and the ARIMA(d,p,q) model is obtained. m ,m=1,2,...,p and β m ,m=1,2,...,q is the order of the ARIMA model, that is, the autoregressive coefficient and the moving average coefficient;

[0039] Wherein, the LSSVM model in step 3 is:

[0040]

[0041] Among them, J(w, e) represents the objective function, w represents the weight coefficient vector, γ is the regularization coefficient, e represents the error term, and e i represents the error term of the i-th solution, N represents the number of historical data that make up the training set, The input value x of the i-th solution i Nonlinear transformation, here we take N = 100;

[0042] In the constraints, the training set S(t)={(x i ,y i ):i=1,2,...,N} consists of N historical data, xi represents the input value of the i-th solution, y i Represents the output value of the i-th solution, and the conversion function f makes y i =f(x i );

[0043] in, Indicates that the nonlinear transformation satisfies the formula, w is the weight coefficient vector, b is a constant, e i is the error term of the ith solution;

[0044]

[0045] Among them, K(x i ,x j ) kernel function, used to measure the solution x i and x j The similarity between them, for Gaussian kernel, the value of kernel function is between 0 and 1. i and x j denote the i-th and j-th solutions in the original input space, respectively. Indicates that x i and x j The feature map is mapped to a higher dimensional space, and the similarity is calculated by the inner product, where σ is the width of the radial basis function;

[0046] Among them, the whale optimization algorithm is improved in step 3, and the specific execution steps are as follows:

[0047] Set the algorithm parameters, set the search ranges of γ and σ to [0.001, 100] and [0.001, 10] respectively, the population size N, and the maximum number of iterations T max , fitness variance threshold The number of mutation individuals is M, and the initial number of iterations is k=0;

[0048] Input sample dataset

[0049] Sobol sequence generates N initial feasible solutions in, represents the solution of the i-th initial iteration, and the superscript represents the number of iterations. and They correspond to the regularization coefficient and radial basis function width parameter of the i-th solution respectively;

[0050] Start iteration and check whether the number of iterations k is less than the maximum number of iterations T max ;

[0051] For each k-th iteration, the i-th solution Do the following:

[0052] Step 3.1: Update parameters r,p,l, among them, is the position vector of the i-th solution in the k-th iteration, is an update parameter of the position vector of the i-th solution in the k-th iteration, which is used to calculate the position of the new solution to ensure that the solution moves in the search space, is another update parameter for the i-th solution in the k-th iteration, r is a random number ranging from [0,1], p is a probability value, and l is a step size that controls the distance the solution moves in the search space;

[0053] Step 3.2: Exploitation and Construct the LSSVM model, where represents the i-th solution in the k-th iteration; and They correspond to the regularization coefficient and radial basis function width parameter of the i-th solution in the k-th iteration respectively;

[0054] Step 3.3: Calculate the sample The predicted value of in, represents the predicted value of the jth model at the kth iteration, and L represents a total of L independent prediction models;

[0055] Step 3.4: Calculation Fitness Among them, f i k The fitness of the i-th solution in the k-th iteration, At the kth iteration, the true value of the jth model;

[0056] Step 3.5: Update the feasible solution position;

[0057] Step 3.6: Calculate the fitness variance σ of the feasible solution 2 ,If it is less than the fitness variance threshold, M feasible solutions are randomly selected and the mutation operation is performed;

[0058] Step 3.7: Select the particle with the highest fitness and use it as the candidate for the next generation. Compare whether the new optimal solution is better than the current optimal solution.

[0059] If f i k+1 >f * k The new solution is better, and the particle position and fitness are updated, where f i k+1 represents the fitness of the i-th solution at the k+1-th iteration, f * krepresents the optimal solution at the kth iteration;

[0060] Update the number of iterations k = k + 1; return to step 3.1;

[0061] When the maximum number of iterations is reached, the algorithm terminates and outputs the optimal solutions γ and σ.

[0062] The specific formula for updating the parameters in step 3.1 is:

[0063] In the kth iteration cycle, is the solution closest to the global optimum or the global optimum solution. Other solutions update their positions using the following formula:

[0064]

[0065]

[0066]

[0067]

[0068] in, represents the position of the i-th solution in the k-th generation, represents the optimal solution of the kth iteration, is the coefficient of the i-th solution in the k-th generation, represents the distance between the i-th solution of the k-th generation and the optimal position, a k The control parameter of the kth iteration, r i k is the random parameter of the i-th solution between [0,1] in the k-th generation, is the coefficient of the i-th solution k generations, and a is the convergence factor, which decays linearly to 0 as the number of iterations increases:

[0069] a k =2(1-k / T max )

[0070] Among them, a k The control parameter of the kth iteration, T max is the maximum number of iterations;

[0071] After all feasible solutions update their positions, the fitness of each feasible solution is recalculated, and the feasible solution with the largest fitness is selected as the global optimal solution for the next iteration;

[0072] The specific formula for updating the feasible solution in step 3.5 is:

[0073] At each iteration, a method is randomly selected between shrinking and spiral updating to update the feasible solution position:

[0074]

[0075] in, represents the optimal solution of the kth iteration, is the coefficient of the i-th solution k generations, represents the distance between the i-th solution of the k-th generation and the optimal position, and p is a random number in the range [0,1];

[0076] When p<0.5 and |A|>1, the feasible solution position is updated according to the following formula:

[0077]

[0078]

[0079] in, represents the position of the kth generation of the ith solution, is the coefficient of the i-th solution in the k-th generation, represents the distance between the i-th solution of the k-th generation and the optimal position, is the coefficient of the i-th solution k generations, It is a feasible solution randomly selected to replace the optimal solution at this time Forces a global optimal solution search.

[0080] The specific formula for performing the mutation operation in step 3.6 is:

[0081] Use a random variable ξ that follows a standard Cauchy distribution to perturb the position of the feasible solution:

[0082]

[0083] Among them, X k represents the position of the kth generation, Represents the position of the kth generation after adding the disturbance.

[0084] The linear weighting method in step S4 is:

[0085]

[0086] Among them, w0 is the weight of the trend component, w1 is the weight of detail component 1, w2 is the weight of detail component 2, w3 is the weight of detail component 3, and b is a constant. is the trend component, is detail component 1, It is detail component 2, It is detail component 3;

[0087] Normalize the predicted value of the k+1th generation Denormalization can get the denormalized predicted value of RSRP k+1 generation

[0088]

[0089] Here, max(X(k)) represents the maximum value in the kth generation, and min(X(k)) represents the minimum value in the kth generation.

[0090] The technical solution of the system of the present invention is a link outage prediction system based on multi-model fusion, comprising:

[0091] The wavelet function decomposition module is used to decompose the collected reference signal power into trend components and detail components using the Mallat algorithm and the Symlet basic wavelet function;

[0092] Average model prediction module, used to predict the decomposed trend component using the autoregressive moving average model;

[0093] The LSSVM model prediction module is used to predict the decomposed detail components using the LSSVM model optimized by the improved whale optimization algorithm;

[0094] The linear weighted calculation module is used to obtain the final prediction result by using the prediction results of the trend component and the detail component in a linear weighted manner;

[0095] The link judgment module is used to calculate the predicted signal-to-noise ratio using the prediction results, and then determine whether to trigger a link interruption declaration.

[0096] The advantage of the present invention is that it overcomes the shortcomings of using only a single prediction model. The present invention adopts multiple prediction models to separately model and predict different characteristics of RSRP, and finally obtains the final RSRP prediction value by weighted summation of the prediction results of multiple models. Since the changing trend characteristics of RSRP can be obtained from a large time scale and the fluctuation characteristics of RSRP can be obtained from a small time scale, the trend component and detail component of the large time scale and the small time scale can be extracted respectively by wavelet decomposition, wherein the trend component has the characteristics of periodicity and non-stationarity, while the detail component has weak predictability and nonlinear characteristics. Taking into account the different characteristics of the trend component and the detail component and the demand for lower computational complexity, the present invention adopts an autoregressive moving average model and a least squares support vector machine model to model and predict the trend component and the detail component respectively. Since the regularization coefficient and radial basis function width of the LSSVM model greatly affect its prediction performance, the present invention adopts a whale optimization algorithm to determine the optimal parameters of the model. In response to the shortcomings of the standard WOA algorithm, the present invention designs an improved whale optimization algorithm through the strategy of low-difference sequence and population variation, thereby improving the prediction performance of the LSSVM model. Since the prediction accuracy of each model varies, the final RSRP prediction value is obtained through weighted summation. This multi-model fusion approach can fully utilize the advantages of multiple prediction models to improve RSRP prediction accuracy, thereby improving RLF prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 : A flow chart of a method according to an embodiment of the present invention.

[0098] Figure 2 : Prediction result diagram of multi-model fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0099] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0100] In specific implementation, the method proposed in the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. System devices that implement the method, such as computer-readable storage media that store the corresponding computer program of the technical solution of the present invention and computer equipment that runs the corresponding computer program, should also be within the scope of protection of the present invention.

[0101] The following combination Figure 1-2The specific embodiment of the present invention is a link outage prediction method based on multi-model fusion to improve the RSRP prediction accuracy and RLF prediction accuracy.

[0102] like Figure 1 Shown is a flow chart of a method according to an embodiment of the present invention.

[0103] The technical solution of the method of the present invention is a link outage prediction method based on multi-model fusion, and the specific steps are as follows:

[0104] Step 1: Use Mallat algorithm and Symlet basic wavelet function to decompose the collected reference signal power into trend component and detail component.

[0105] Step 2: Use the autoregressive moving average model to predict the decomposed trend component.

[0106] Step 3: Use the LSSVM model optimized by the improved whale optimization algorithm to predict the decomposed detail components.

[0107] Step 4: Using the prediction results of trend component and detail component, the final prediction result is obtained by linear weighting.

[0108] Step 5: Use the prediction result to calculate the predicted signal-to-noise ratio, and then determine whether to trigger a link interruption declaration.

[0109] Wherein, the maximum decomposition level J of the Mallat algorithm in step 1 is 3;

[0110] Wherein, the vanishing moment order of the Symlet basic wavelet function in step 1 is N=5;

[0111] The autoregressive moving average model in step 2 is:

[0112]

[0113] Among them, X(t) represents the time series value at time t, ε(t) represents the random error at time t, L represents the lag operator, and L m represents the lag operator of the mth term, α m represents the coefficient of the lag term m in the autoregressive model, β m It represents the coefficient of the lag term m in the moving average model. After X(t) is transformed into a stationary series through d-time difference, it is modeled through the p-order autoregressive model and the q-order moving average model. The error term ε(t) in AR(p) is expressed as an MA(q) model, and the ARIMA(d,p,q) model is obtained. m ,m=1,2,...,p and β m,m=1,2,...,q is the order of the ARIMA model, that is, the autoregressive coefficient and the moving average coefficient;

[0114] The LSSVM model in step 3 is:

[0115]

[0116] Among them, J(w, e) represents the objective function, w represents the weight coefficient vector, γ is the regularization coefficient, e represents the error term, and e i represents the error term of the i-th solution, N represents the number of historical data that make up the training set, The input value x of the i-th solution i Nonlinear transformation, here we take N = 100;

[0117] In the constraints, the training set S(t)={(x i ,y i ):i=1,2,...,N} consists of N historical data, x i represents the input value of the i-th solution, y i Represents the output value of the i-th solution, and the conversion function f makes y i =f(x i );

[0118] in, Indicates that the nonlinear transformation satisfies the formula, w is the weight coefficient vector, b is a constant, e i is the error term of the ith solution;

[0119]

[0120] Among them, K(x i ,x j ) kernel function, used to measure the solution x i and x j The similarity between them, for Gaussian kernel, the value of kernel function is between 0 and 1. i and x j denote the i-th and j-th solutions in the original input space, respectively. Indicates that x i and x j The feature map is mapped to a higher dimensional space, and the similarity is calculated by the inner product, where σ is the width of the radial basis function;

[0121] Among them, the whale optimization algorithm is improved in step 3, and the specific execution steps are as follows:

[0122] Set the algorithm parameters, set the search ranges of γ and σ to [0.001, 100] and [0.001, 10] respectively, the population size N, and the maximum number of iterations T max, fitness variance threshold The number of mutation individuals is M, and the initial number of iterations is k=0;

[0123] Input sample dataset

[0124] Sobol sequence generates N initial feasible solutions in, represents the solution of the i-th initial iteration, and the superscript represents the number of iterations. and They correspond to the regularization coefficient and radial basis function width parameter of the i-th solution respectively;

[0125] Start iteration and check whether the number of iterations k is less than the maximum number of iterations T max ;

[0126] For each k-th iteration, the i-th solution Do the following:

[0127] Step 3.1: Update parameters r,p,l, among them, is the position vector of the i-th solution in the k-th iteration, is an update parameter of the position vector of the i-th solution in the k-th iteration, which is used to calculate the position of the new solution to ensure that the solution moves in the search space, is another update parameter for the i-th solution in the k-th iteration, r is a random number ranging from [0,1], p is a probability value, and l is a step size that controls the distance the solution moves in the search space;

[0128] Step 3.2: Exploitation and Construct the LSSVM model, where represents the i-th solution in the k-th iteration; and They correspond to the regularization coefficient and radial basis function width parameter of the i-th solution in the k-th iteration respectively;

[0129] Step 3.3: Calculate the sample The predicted value of in, represents the predicted value of the jth model at the kth iteration, and L represents a total of L independent prediction models;

[0130] Step 3.4: Calculation Fitness Among them, f i k The fitness of the i-th solution in the k-th iteration, At the kth iteration, the true value of the jth model;

[0131] Step 3.5: Update the feasible solution position;

[0132] Step 3.6: Calculate the fitness variance σ of the feasible solution 2 ,If it is less than the fitness variance threshold, M feasible solutions are randomly selected and the mutation operation is performed;

[0133] Step 3.7: Select the particle with the highest fitness and use it as the candidate for the next generation. Compare whether the new optimal solution is better than the current optimal solution.

[0134] If f i k+1 >f * k The new solution is better, and the particle position and fitness are updated, where f i k+1 represents the fitness of the i-th solution at the k+1-th iteration, f * k represents the optimal solution at the kth iteration;

[0135] Update the number of iterations k = k + 1; return to step 3.1;

[0136] When the maximum number of iterations is reached, the algorithm terminates and outputs the optimal solutions γ and σ.

[0137] The specific formula for updating the parameters in step 3.1 is:

[0138] In the kth iteration cycle, is the solution closest to the global optimum or the global optimum solution. Other solutions update their positions using the following formula:

[0139]

[0140] in, represents the position of the i-th solution in the k-th generation, represents the optimal solution of the kth iteration, is the coefficient of the i-th solution in the k-th generation, represents the distance between the i-th solution of the k-th generation and the optimal position, a k The control parameter of the kth iteration, r i k is the random parameter of the i-th solution between [0,1] in the k-th generation, is the coefficient of the i-th solution k generations, and a is the convergence factor, which decays linearly to 0 as the number of iterations increases:

[0141] a k =2(1-k / T max ) (5)

[0142] Among them, a k The control parameter of the kth iteration, T max is the maximum number of iterations;

[0143] After all feasible solutions update their positions, the fitness of each feasible solution is recalculated, and the feasible solution with the largest fitness is selected as the global optimal solution for the next iteration;

[0144] The specific formula for updating the feasible solution in step 3.5 is:

[0145] At each iteration, a method is randomly selected between shrinking and spiral updating to update the feasible solution position:

[0146]

[0147] in, represents the optimal solution of the kth iteration, is the coefficient of the i-th solution k generations, represents the distance between the i-th solution of the k-th generation and the optimal position, and p is a random number in the range [0,1];

[0148] When p<0.5 and |A|>1, the feasible solution position is updated according to the following formula:

[0149]

[0150] in, represents the position of the kth generation of the ith solution, is the coefficient of the i-th solution in the k-th generation, represents the distance between the i-th solution of the k-th generation and the optimal position, is the coefficient of the i-th solution k generations, It is a feasible solution randomly selected to replace the optimal solution at this time Forces a global optimal solution search.

[0151] The specific formula for performing the mutation operation in step 3.6 is:

[0152] Use a random variable ξ that follows a standard Cauchy distribution to perturb the position of the feasible solution:

[0153]

[0154] Among them, X k represents the position of the kth generation, Represents the position of the kth generation after adding the disturbance.

[0155] The linear weighting method in step S4 is:

[0156]

[0157] Among them, w0 is the weight of the trend component, w1 is the weight of detail component 1, w2 is the weight of detail component 2, w3 is the weight of detail component 3, and b is a constant. is the trend component, is detail component 1, It is detail component 2, It is detail component 3;

[0158] Normalize the predicted value of the k+1th generation Denormalization can get the denormalized predicted value of RSRP k+1 generation

[0159]

[0160] Here, max(X(k)) represents the maximum value in the kth generation, and min(X(k)) represents the minimum value in the kth generation.

[0161] An embodiment of the system of the present invention is a link outage prediction system based on multi-model fusion, comprising:

[0162] The wavelet function decomposition module is used to decompose the collected reference signal power into trend components and detail components using the Mallat algorithm and the Symlet basic wavelet function;

[0163] The average model prediction module is used to predict the decomposed trend component using the autoregressive moving average model.

[0164] The LSSVM model prediction module is used to predict the decomposed detail components using the LSSVM model optimized by the improved whale optimization algorithm.

[0165] The linear weighted calculation module is used to use the prediction results of the trend component and the detail component to obtain the final prediction result in a linear weighted manner.

[0166] The link judgment module is used to calculate the predicted signal-to-noise ratio using the prediction results, and then determine whether to trigger a link interruption declaration.

[0167] The key to the RLF prediction method based on multi-model fusion lies in RSRP prediction. Therefore, this experiment first uses an open-source RSRP data archive to verify the performance of the prediction method based on multi-model fusion. The data archive contains RSRP data from 415,244 signal measurement points in a certain urban area, with a duration of 800 seconds. For the ARIMA model predicting the trend component, the grid search method can be used to solve the (p, q) value combination that minimizes the BIC. The search range is set to the interval [1, 10], and the search result is p = 1, q = 4. Since the trend component can pass the ADF test after two difference operations, d = 2. Therefore, the ARIMA (2, 1, 4) model is used to predict the trend component. For the LSSVM model used to predict the detail component, the improved WOA algorithm designed by the present invention is used to determine the optimal values ​​of the regularization coefficient γ and the radial basis function width σ. The search ranges of the model parameters γ and σ are set to [0.001, 100] and [0.001, 10] respectively, the population size N = 40, and the maximum number of iterations T. max =150, fitness variance threshold The number of mutant individuals M = 6. The results of solving the optimal values ​​of the LSSVM model parameters are shown in the table.

[0168] Table 1 Parameter optimal value solution results

[0169]

[0170] The prediction accuracy of the model is measured using the root mean square error (RMSE), which is calculated as follows:

[0171]

[0172] Among them, x i is the true value, is the predicted value. The smaller the RMSE, the smaller the difference between the predicted value and the true value, and the higher the prediction accuracy. The first 600s of the trend component and the detail component data are taken as the training set, and the last 200s of the data are taken as the test set. The prediction results of each component on the training set and the test set are shown as follows:

[0173] As shown in the table. Because the ARIMA model structure lacks long-term forecasting capabilities for time series with a certain degree of randomness and chaos, it can only provide a good fit to the test data set over a short period of time. This results in large long-term forecast errors and a high RMSE. For detail components D1 and D2, the LSSVM model achieved good forecasting results, with relatively low RMSEs for both the training and test sets. Because detail component D3 exhibits the greatest volatility, with the most spikes and mutations, the LSSVM model also exhibits the largest forecast error.

[0174] Table 2 RMSE of prediction results of each component

[0175]

[0176] From the above analysis, it can be seen that the ARIMA model and the LSSVM model have different prediction accuracy when predicting each component. Direct addition will bring about a large error. Therefore, the present invention chooses the weighted summation method to solve the final prediction value. The weights of the trend component T, detail components D1, D2, and D3 are obtained by the least squares method as 1.0203, 1.2160, 0.9422, and 0.8354 respectively. Since the higher the level of detail components, the greater the prediction error, the weights of D1, D2, and D3 decrease in turn, and the smaller the contribution to the final prediction result. The final prediction result of the entire data set is as follows: Figure 2 As shown in the figure, the RMSE is 0.110dB. If the prediction values ​​of each component are fused using the same weight, the calculated RMSE is 0.337dB. The prediction error is greater than the weighted summation method, which shows that the weighted summation method is appropriate for fusing the prediction results of each component.

[0177] After the source base station receives the measurement report from the mobile phone, it obtains the RSRP prediction value of each adjacent base station through the multi-model fusion prediction method. It can use the T310 timer to determine whether the target base station will trigger the RLF declaration and adjust the handover signaling sending time.

[0178] It should be understood that parts not elaborated in detail in this specification belong to the prior art.

[0179] It should be understood that the above description of the embodiments is relatively detailed and cannot be regarded as limiting the scope of protection of the patent of the present invention. Under the guidance of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested by the present invention shall be based on the attached claims.

Claims

1. A link outage prediction method based on multi-model fusion, characterized in that: The following steps are involved: Step 1: The collected reference signal power is decomposed into trend component and detail component using Mallat algorithm and Symlet basic wavelet function; Step 2: Use the autoregressive moving average model to predict the decomposed trend component; Step 3: Use the LSSVM model optimized by the improved whale optimization algorithm to predict the decomposed detail components; Step 4: Using the prediction results of the trend component and the detail component, the final prediction result is obtained by linear weighting; Step 5: Use the prediction result to calculate the predicted signal-to-noise ratio, and then determine whether to trigger a link interruption declaration.

2. The link outage prediction method based on multi-model fusion according to claim 1 is characterized by: in, The maximum decomposition level J of the Mallat algorithm in step 1; Wherein, the vanishing moment order of the Symlet basic wavelet function in step 1 is N.

3. The link outage prediction method based on multi-model fusion according to claim 2, characterized in that: The autoregressive moving average model in step 2 is: Among them, X(t) represents the time series value at time t, ε(t) represents the random error at time t, L represents the lag operator, and L m represents the lag operator of the mth term, α m represents the coefficient of the lag term m in the autoregressive model, β m represents the coefficient of the lag term m in the moving average model. After X(t) is transformed into a stationary series through d-time difference, it is modeled through the p-order autoregressive model and the q-order moving average model. The error term ε(t) in AR(p) is expressed as an MA(q) model to obtain the ARIMA(d,p,q) model. m ,m=1,2,...,p and β m ,m=1,2,…,q is the order of the ARIMA model, that is, the autoregressive coefficient and the moving average coefficient.

4. The link outage prediction method based on multi-model fusion according to claim 3 is characterized by: The LSSVM model in step 3 is: Among them, J(w, e) represents the objective function, w represents the weight coefficient vector, γ is the regularization coefficient, e represents the error term, and e i represents the error term of the i-th solution, N represents the number of historical data that make up the training set, The input value x of the i-th solution i Nonlinear transformation, here we take N = 100; In the constraints, the training set S(t)={(x i ,y i ):i=1,2,...,N} consists of N historical data, x i represents the input value of the i-th solution, y i Represents the output value of the i-th solution, and the conversion function f makes y i =f(x i ); in, Indicates that the nonlinear transformation satisfies the formula, w is the weight coefficient vector, b is a constant, e i is the error term of the ith solution; Among them, K(x i ,x j ) kernel function, used to measure the solution x i and x j The similarity between them, for Gaussian kernel, the value of kernel function is between 0 and 1; i and x j denote the i-th and j-th solutions in the original input space respectively; Indicates that x i and x j The feature map is mapped to a higher dimensional space, and the similarity is calculated by the inner product, where σ is the width of the radial basis function.

5. The link outage prediction method based on multi-model fusion according to claim 4 is characterized in that: In step 3, the whale optimization algorithm is improved, and the specific execution steps are as follows: Set the algorithm parameters, set the search ranges of γ and σ to [0.001, 100] and [0.001, 10] respectively, the population size N, and the maximum number of iterations T max , fitness variance threshold The number of mutation individuals is M, and the initial number of iterations is k=0; Input sample dataset Sobol sequence generates N initial feasible solutions in, represents the solution of the i-th initial iteration, and the superscript represents the number of iterations. and They correspond to the regularization coefficient and radial basis function width parameter of the i-th solution respectively; Start iteration and check whether the number of iterations k is less than the maximum number of iterations T max .

6. The link outage prediction method based on multi-model fusion according to claim 5, characterized in that: For each k-th iteration, the i-th solution Do the following: Step 3.1: Update parameters in, is the position vector of the i-th solution in the k-th iteration, is an update parameter of the position vector of the i-th solution in the k-th iteration, which is used to calculate the position of the new solution to ensure that the solution moves in the search space, is another update parameter for the i-th solution in the k-th iteration, r is a random number ranging from [0,1], p is a probability value, and l is a step size that controls the distance the solution moves in the search space; Step 3.2: Exploitation and Construct the LSSVM model, where represents the i-th solution in the k-th iteration; and They correspond to the regularization coefficient and radial basis function width parameter of the i-th solution in the k-th iteration respectively; Step 3.3: Calculate the sample The predicted value of in, represents the predicted value of the jth model at the kth iteration, and L represents a total of L independent prediction models; Step 3.4: Calculation Fitness in, The fitness of the i-th solution in the k-th iteration, At the kth iteration, the true value of the jth model; Step 3.5: Update the feasible solution position; Step 3.6: Calculate the fitness variance σ of the feasible solution 2 ,If it is less than the fitness variance threshold, M feasible solutions are randomly selected and the mutation operation is performed; Step 3.7: Select the particle with the highest fitness and use it as the candidate for the next generation. Compare whether the new optimal solution is better than the current optimal solution. if The new solution is better, and the particle position and fitness are updated, where represents the fitness of the i-th solution at the k+1-th iteration, f * k represents the optimal solution at the kth iteration; Update the number of iterations k = k + 1; return to step 3.1; When the maximum number of iterations is reached, the algorithm terminates and outputs the optimal solutions γ and σ.

7. The link outage prediction method based on multi-model fusion according to claim 6, characterized in that: The specific formula for updating the parameters in step 3.1 is: In the kth iteration cycle, is the solution closest to the global optimum or the global optimum solution. Other solutions update their positions using the following formula: in, represents the position of the i-th solution in the k-th generation, represents the optimal solution of the kth iteration, is the coefficient of the i-th solution in the k-th generation, represents the distance between the i-th solution of the k-th generation and the optimal position, a k The control parameters for the kth iteration, is the random parameter of the i-th solution between [0,1] in the k-th generation, is the coefficient of the i-th solution k generations, and a is the convergence factor, which decays linearly to 0 as the number of iterations increases: a k =2(1-k / T max ) Among them, a k The control parameter of the kth iteration, T max is the maximum number of iterations; After all feasible solutions update their positions, the fitness of each feasible solution is recalculated, and the feasible solution with the largest fitness is selected as the global optimal solution for the next iteration.

8. The link outage prediction method based on multi-model fusion according to claim 7, characterized in that: The specific formula for updating the feasible solution in step 3.5 is: At each iteration, a method is randomly selected between shrinking and spiral updating to update the feasible solution position: in, represents the optimal solution of the kth iteration, is the coefficient of the i-th solution k generations, represents the distance between the i-th solution of the k-th generation and the optimal position, and p is a random number in the range [0,1]; When p<0.5 and |A|>1, the feasible solution position is updated according to the following formula: in, represents the position of the kth generation of the ith solution, is the coefficient of the i-th solution in the k-th generation, represents the distance between the i-th solution of the k-th generation and the optimal position, is the coefficient of the i-th solution k generations, It is a feasible solution randomly selected to replace the optimal solution at this time Forces a global optimal solution search.

9. The link outage prediction method based on multi-model fusion according to claim 8, characterized in that: The specific formula for performing the mutation operation in step 3.6 is: Use a random variable ξ that follows a standard Cauchy distribution to perturb the position of the feasible solution: Among them, X k represents the position of the kth generation, represents the position of the kth generation after adding the disturbance; The linear weighting method in step S4 is: Where w0 is the weight of the trend component, w1 is the weight of detail component 1, w2 is the weight of detail component 2, w3 is the weight of detail component 3, and b is a constant; is the trend component, is detail component 1, It is detail component 2, It is detail component 3; Normalize the predicted value of the k+1th generation Denormalization can get the denormalized predicted value of RSRP k+1 generation Here, max(X(k)) represents the maximum value in the kth generation, and min(X(k)) represents the minimum value in the kth generation.

10. A link outage prediction system based on multi-model fusion, characterized in that: include: The wavelet function decomposition module is used to decompose the collected reference signal power into trend components and detail components using the Mallat algorithm and the Symlet basic wavelet function; Average model prediction module, used to predict the decomposed trend component using the autoregressive moving average model; The LSSVM model prediction module is used to predict the decomposed detail components using the LSSVM model optimized by the improved whale optimization algorithm; The linear weighted calculation module is used to obtain the final prediction result by using the prediction results of the trend component and the detail component in a linear weighted manner; The link judgment module is used to calculate the predicted signal-to-noise ratio using the prediction results, and then determine whether to trigger a link interruption declaration.

Images