Result-selective combined model frequency modulation sequence prediction method and related equipment

Abstract

The invention relates to the technical field of frequency modulation signal prediction, and discloses a result-selective combined model frequency modulation sequence prediction method and related equipment, and the method comprises the steps: predicting a frequency modulation instruction sequence, and correspondingly generating a plurality of groups of initial prediction results; calculating a deviation degree between the multiple groups of initial prediction results and the average prediction sequence, and performing a first-level elimination strategy on the multiple groups of initial preset neural network models based on the deviation degree to obtain a first-level preset neural network model set; calculating the distinction degree between every two in the first-level preset neural network model set, and performing a second-level elimination strategy on the first-level preset neural network model set based on the distinction degree to obtain a second-level preset neural network model set; the deviation degree weight of the second-level preset neural network model set is calculated, weighted fusion is carried out on the second-level preset neural network model set based on the deviation degree weight, a final prediction sequence is generated, and the overall prediction accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of frequency modulation signal prediction, and particularly to a method for predicting a frequency modulation sequence with a result-selective combined model and related devices. Background Art

[0002] Frequency modulation signals widely exist in scenarios such as communication, radar, and power system frequency control. Their frequency change patterns usually exhibit highly nonlinear characteristics (such as exponential modulation and burst modulation, etc.) and long-period time series dependence characteristics (such as periodic frequency modulation or gradual frequency offset, etc.). Traditional prediction methods (such as autoregressive integrated moving average model, hidden Markov model, and Kalman filter, etc.) build models based on linear or weakly nonlinear assumptions, and it is difficult to accurately fit complex frequency modulation patterns, resulting in significant prediction errors. Specifically, the prior art has the following defects: 1. Insufficient ability to capture non-linear features Traditional statistical models (such as autoregressive integrated moving average model, etc.) and shallow machine learning methods (such as hidden Markov model, support vector machine, etc.) usually assume that the data follows a linear or stationary distribution, while the frequency change of frequency modulation signals often contains strong non-linear features such as mutations and oscillations. For example, the frequency of an exponential frequency modulation signal increases exponentially with time. Due to the limitations of the model structure (such as linear difference equations or fixed state transition probabilities), traditional methods cannot effectively fit such patterns, resulting in a large deviation between the prediction result and the actual signal.

[0003] 2. Difficulty in modeling long-term time series dependence The frequency change of a frequency modulation sequence may span dozens of seconds or even longer time periods (such as periodic pulse frequency modulation), while traditional methods (such as Markov model, recurrent neural network RNN) have limitations in modeling long-distance time series correlations. The Markov model cannot capture cross-period dependence relationships due to the "memoryless" assumption; although the RNN can process sequence data, its vanishing gradient problem leads to lag or divergence in the prediction of long-period frequency modulation rules, making it difficult to meet the requirements of high-precision prediction.

[0004] To overcome the limitations of a single model, the prior art has proposed a combined prediction model (such as multi-model weighted fusion), which improves the overall performance by superimposing the prediction results of different models. However, although such a scheme partially alleviates the defects of a single model, it introduces new technical problems: 1. Lack of dynamic weight allocation mechanism Combined models usually adopt fixed weights or simple weighting strategies (such as equal-weight averaging), and cannot adapt to the non-stationarity and randomness of frequency modulation sequences. For example, in the case of a frequency mutation scenario, static weights based on historical performance may cause the model to respond lag; if dynamic weight optimization is introduced (such as adjusting weights based on real-time error feedback), additional computing resources are required and it may not meet the real-time prediction requirements due to algorithm delays.

[0005] 2. Conflict between Overfitting and Generalization Ability The combined model improves the expression ability by integrating multiple sub - models. However, when the amount of training data is insufficient, the model may overfit the noise in the training set (such as accidental fluctuations in specific FM patterns), resulting in a decline in the generalization ability for unknown data. Especially in the FM signal scenario, the complex structure of the combined model (such as the fusion of multi - layer neural networks) under small - sample conditions is prone to exacerbate the overfitting risk and limit its practical application value.

[0006] Traditional methods cannot adapt to complex FM patterns due to model assumptions. Although the combined model attempts to improve performance through multi - model fusion, it is difficult to meet the requirements of high - precision and real - time prediction due to weight optimization and generalization ability defects. Therefore, there is an urgent need for a prediction method that can dynamically adapt to the characteristics of FM signals and balance model complexity and generalization ability. Summary of the Invention

[0007] In order to overcome the defects of the above - mentioned existing technologies, the purpose of the present invention is to provide a result - selective combined model FM sequence prediction method and related devices to solve the technical problems of non - linear modeling, long - period dependence capture, and dynamic adaptation of the combined model in FM signal prediction.

[0008] The present invention is realized through the following technical solutions: In the first aspect, the present invention provides a result - selective combined model FM sequence prediction method, including: Using multiple groups of initially preset neural network models to predict the FM command sequence, and correspondingly generating multiple groups of initial prediction results; Calculating an average prediction sequence based on multiple groups of initial prediction results; Calculating the deviation degree between multiple groups of initial prediction results and the average prediction sequence, and performing a first - level elimination strategy on multiple groups of initially preset neural network models based on the deviation degree to obtain a first - level preset neural network model set; Calculating the difference degree between any two models in the first - level preset neural network model set, and performing a second - level elimination strategy on the first - level preset neural network model set based on the difference degree to obtain a second - level preset neural network model set; Calculating the deviation - degree weight of the second - level preset neural network model set, and performing weighted fusion on the second - level preset neural network model set based on the deviation - degree weight to generate a final prediction sequence.

[0009] Preferably, the multiple groups of initial preset neural network models include at least 5 groups of models among the Informer model, Autoformer model, N-HiTS model, PatchTST model, TimeGrad model, DLinear model, TFT model, Graph WaveNet model, TimesNet model, or TiDE model.

[0010] Preferably, the calculation formula for the average prediction sequence includes:

[0011] Among them, M is the number of groups of initial preset neural network models currently participating in the calculation; is the i initial prediction result of the

[0012] Preferably, calculate the deviation degree between the multiple groups of initial prediction results and the average prediction sequence, and perform a first-level elimination strategy on the multiple groups of initial preset neural network models based on the deviation degree to obtain the first-level preset neural network model set. The calculation formula for the deviation degree includes:

[0013] Among them, L is the total length of the prediction sequence; is the i initial prediction result of the initial preset neural network model; is the i average prediction sequence; h is the first prediction result of the prediction sequence; N is the number of initial prediction results; is the deviation degree of the

[0014] Preferably, calculate the difference degree between any two in the first-level preset neural network model set, and perform a second-level elimination strategy on the first-level preset neural network model set based on the difference degree to obtain the second-level preset neural network model set. The calculation formula for the difference degree includes:

[0015] Among them, is the iThe initial prediction result of an initial preset neural network model; For the j initial prediction result of an initial preset neural network model; h is the first prediction result of the prediction sequence; N is the number of initial prediction results; For the i th and the j th difference degree between the initial preset neural network models; ∣∣ represents the absolute value; The secondary elimination strategy includes: Compare the difference degrees between every two in the set of primary preset neural network models, and determine the two groups of primary preset neural network models corresponding to the maximum difference degree among all the difference degrees; Based on the two groups of primary preset neural network models corresponding to the maximum difference degree, calculate the average value of the difference degrees associated with the initial prediction results of the two groups of primary preset neural network models respectively, and delete the group of primary preset neural network models with the larger average value to obtain the set of secondary preset neural network models.

[0016] Preferably, the calculation formula of the deviation degree weight includes:

[0017] where S is the set of secondary preset neural network models; For the i deviation degree of the th neural network model; k is any neural network model;

[0018] Preferably, the calculation formula of the weighted fusion includes:

[0019] where is the prediction result of the remaining neural networks; is the weight; is the weighted fusion result; i represents the i th initial preset neural network model in the set of neural network models.

[0020] In a second aspect, the present invention also provides a result-selective combined model frequency modulation sequence prediction system, including: An initial prediction module, configured to use multiple groups of initial preset neural network models to predict a frequency modulation instruction sequence and correspondingly generate multiple groups of initial prediction results; An average value calculation and processing module, configured to calculate an average prediction sequence based on the multiple groups of initial prediction results; The first-level elimination processing module is used to calculate the deviation degrees of multiple groups of initial prediction results from the average prediction sequence, and perform a first-level elimination strategy on multiple groups of initial preset neural network models based on the deviation degrees to obtain a first-level preset neural network model set; The second-level elimination processing module is used to calculate the difference degrees between every two of the first-level preset neural network model set, and perform a second-level elimination strategy on the first-level preset neural network model set based on the difference degrees to obtain a second-level preset neural network model set; The prediction sequence generation module is used to calculate the deviation degree weights of the second-level preset neural network model set, and perform weighted fusion on the second-level preset neural network model set based on the deviation degree weights to generate a final prediction sequence.

[0021] In a third aspect, the present invention further provides a mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the result-selective combined model frequency modulation sequence prediction method as described above is implemented.

[0022] In a fourth aspect, the present invention further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the result-selective combined model frequency modulation sequence prediction method as described above is implemented.

[0023] Compared with the prior art, the present invention has the following beneficial technical effects: The present invention provides a result-selective combined model frequency modulation sequence prediction method. By using multiple groups of preset neural network models to predict the frequency modulation instruction sequence, this method can integrate the non-linear modeling capabilities of different models; calculating the average prediction sequence of multiple groups of initial prediction results helps to smooth short-term fluctuations and highlight long-term trends. By calculating the deviation degrees of the initial prediction results from the average prediction sequence and performing first-level elimination based on the deviation degrees, models with inaccurate long-term trend predictions can be eliminated and those that can better capture long-cycle dependencies can be retained. On the basis of the first-level elimination, calculating the difference degrees between the initial prediction results of the remaining models and performing second-level elimination based on the difference degrees helps to eliminate those models with highly similar prediction results and lack of complementarity, so as to retain a more diverse and complementary model combination, calculating the deviation degree weights of the remaining models after the second-level elimination, and performing weighted fusion on the initial prediction results based on these weights to avoid the limitations of a single model and achieve the dynamic adaptation of the combined model. Furthermore, by calculating the average prediction sequence from multiple groups of initial prediction results, the prediction requirements of both short-term and long-term can be taken into account to a certain extent, thereby improving the overall prediction accuracy.

[0024] Furthermore, the deviation measures the difference between the model prediction results and the average prediction sequence. The larger the deviation, the worse the model's ability to capture the data distribution or trend. By eliminating the model with the largest deviation, models that contribute little or even interfere with the overall prediction can be removed, thereby enhancing the stability of the combined model.

[0025] Furthermore, by calculating the distinctiveness between the initial prediction results of the models, the secondary elimination strategy can identify models with highly similar prediction results. When determining the elimination targets, the average values of the distinctiveness of two groups of models are compared. A group of models with a larger average value means that its prediction results are more different from those of other models. The secondary elimination strategy reduces the negative impact of low-quality models on the weighted fusion result by removing redundant models, thereby improving the accuracy of the final prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flowchart of the combined model frequency modulation sequence prediction method in an embodiment of the present invention; Figure 2 is a flowchart of the primary elimination strategy in an embodiment of the present invention; Figure 3 is a flowchart of the secondary elimination strategy in an embodiment of the present invention; Figure 4 is a schematic diagram of the principle structure of the combined model frequency modulation sequence prediction system in an embodiment of the present invention; In the figure: 1. Initial prediction module; 2. Average value calculation and processing module; 3. Primary elimination processing module; 4. Secondary elimination processing module; 5. Prediction sequence generation module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0028] The purpose of the present invention is to provide a combined model frequency modulation sequence prediction method and related devices with result selection to solve the technical problems of non-linear modeling, long-cycle dependence capture, and dynamic adaptation of the combined model in frequency modulation signal prediction.

[0029] The following further describes the present invention in detail with reference to the accompanying drawings: Embodiment 1 Refer to Figure 1, in an embodiment of the present invention, a method for predicting a frequency modulation sequence with result selection combination model is provided, including: Step 1, predicting a frequency modulation instruction sequence by using multiple groups of initial preset neural network models, and correspondingly generating multiple groups of initial prediction results; Specifically, the multiple groups of initial preset neural network models include at least 5 groups of models among the Informer model, Autoformer model, N-HiTS model, PatchTST model, TimeGrad model, DLinear model, TFT model, Graph WaveNet model, TimesNet model or TiDE model.

[0030] The Informer model is an improved model based on Transformer, which proposes a "ProbSparse" attention mechanism, significantly reducing the computational complexity and being suitable for long sequence prediction.

[0031] The Autoformer model is used to introduce an auto-correlation mechanism to replace the traditional attention, and improve the prediction accuracy through sequence periodic decomposition.

[0032] The N-HiTS model is used for a multi-layer hierarchical architecture, combined with multi-scale interpolation technology to improve the long-term prediction efficiency.

[0033] The PatchTST model is used to divide the time series into local segments (Patches), and draw on the idea of ViT to enhance the local feature capture ability.

[0034] The TimeGrad model is based on a diffusion model, and generates probability predictions by gradually denoising.

[0035] The DLinear model is used for a simple linear layer architecture, challenging complex models through sequence decomposition (trend + residual), emphasizing the "simplicity advantage".

[0036] The TFT model (Temporal Fusion Transformer model) is used to fuse LSTM and Transformer, supporting multi-modal modeling of static features, known future inputs and dynamic features.

[0037] The Graph WaveNet model is used to combine a graph neural network with dilated causal convolution to model spatio-temporal dependencies.

[0038] The TimesNet model is used to transform the time series into a two-dimensional space, and use CNN to capture multi-period changes. The TiDE model is used for an encoder-decoder architecture based on MLP to achieve high-speed inference through feature densification.

[0039] Step 2: Calculate the average prediction sequence based on multiple groups of initial prediction results. Specifically, the calculation formula for the average prediction sequence includes:

[0040] Among them, M is the number of groups of the initial preset neural network models currently participating in the calculation; is the i initial prediction result of the

[0041] Step 3: Calculate the deviation degree between multiple groups of initial prediction results and the average prediction sequence, and perform a first-level elimination strategy on multiple groups of initial preset neural network models based on the deviation degree to obtain a first-level preset neural network model set. Specifically, the calculation formula for the deviation degree includes:

[0042] Among them, L is the total length of the prediction sequence; is the i initial prediction result of the average prediction sequence; h is the first prediction result of the prediction sequence; N is the number of initial prediction results; is the i deviation degree of the According to Figure 2 shown, the first-level elimination strategy includes: Step 31: Compare the deviation degrees of multiple groups of initial preset neural network models, delete the neural network model corresponding to the maximum deviation degree, and obtain a first-level preset neural network model set. Step 32: When the maximum deviation degrees in multiple groups of initial preset neural network models are the same, randomly delete any one of the neural network models corresponding to multiple groups of the same maximum deviation degree to obtain a first-level preset neural network model set.

[0043] Step 4: Calculate the difference degrees between any two of the first-level preset neural network model set, and perform a second-level elimination strategy on the first-level preset neural network model set based on the difference degrees to obtain a second-level preset neural network model set. Specifically, the calculation formula for the difference degree includes:

[0044] Among them, is the iThe initial prediction result of an initial preset neural network model; is the j initial prediction result of the th initial preset neural network model; h is the first prediction result of the prediction sequence; N is the number of initial prediction results; i is the j th and th difference degree between the Figure 3 initial preset neural network models; || is the absolute value; As shown in step 41, compare the difference degrees between every two in the set of first-level preset neural network models, and determine the two groups of first-level preset neural network models corresponding to the maximum difference degree among all the difference degrees;

[0045] step 5, calculate the deviation degree weights of the set of second-level preset neural network models, and perform weighted fusion on the set of second-level preset neural network models based on the deviation degree weights to generate a final prediction sequence.

[0046] Specifically, the calculation formula of the deviation degree weight includes:

[0047] where S is the set of second-level preset neural network models; is the i th deviation degree of the k-th neural network model; k is any neural network model;

[0048] Specifically, the calculation formula of the weighted fusion includes:

[0049] where is the prediction result of the remaining neural networks; is the weight; is the weighted fusion result; i represents the i th initial preset neural network model in the neural network model set.

[0050] According to this embodiment, the frequency modulation instruction is set as Pt = [X1, X2, X3,... X i .., X N (N = 1000); With [X1, X2, X3,... Xi ..,X N as the input, predict the sequence with a future length of 0.1N. That is, when N = 1000, put [X1, X2, X3, .X i ..,X 1000 into the neural network to predict the future [X 1001 ,X 1002 ,X 1003 ,...,X 1100 values.

[0051] The prediction result of the first neural network model, Informer, is [X 1001 （1） ,X 1002 （1） ,X 1003 （1） ,...,X 1100 （1） ; The prediction result of the second neural network model, Autoformer, is [X 1001 （2） ,X 1002 （2） ,X 1003 （2） ,...,X 1100 （2） ; The prediction result of the third neural network model, N-HiTS, is [X 1001 （3） ,X 1002 （3） ,X 1003 （3） ,...,X 1100 （3） ; The prediction result of the i-th neural network model is [X 1001 （i） ,X 1002 （i） ,X 1003 （i） ,...,X 1100 （i） ; The prediction result of the 10th neural network model, TiDE, is [X 1001 （10） ,X 1002 （10） ,X 1003 （10） ,...,X 1100 （10） .

[0052] Let the average prediction results of 10 neural network models be: , , ,..., .

[0053] Taking as an example, the specific calculation method is: = . The other numerical calculations are carried out in the same way.

[0054] Taking the prediction result of the first neural network as an example, the calculation method of the deviation degree D1 of the first neural network is:

[0055] Calculate [D1, D2, D3,..., D 10 in turn according to this method; After calculating the deviation degree, the frequency modulation sequence with the largest deviation degree indicates the worst prediction effect, so it is eliminated and deleted. That is, for example, if the value of D5 in [D1, D2, D3,..., D 10 is the largest, then the prediction result [X 1001 （5） , X 1002 （5） , X 1003 （5） ,..., X 1100 （5） of the 5th neural network TimeGrad is discarded in the follow-up and no longer participates in the follow-up.

[0056] Calculate the difference degree Z between every two of the remaining 9 initial prediction results. The greater the difference degree, the greater the difference between these two prediction results, which means that the prediction accuracy of at least one party is poor. Taking prediction result 1 and prediction result 2 as an example to illustrate the difference degree Z 12 The calculation formula is as follows:

[0057] Calculate the difference degrees between every two of the remaining 9 initial prediction results according to the above formula, and there are 36 in total.

[0058] The 36 difference degrees calculated are [Z 12 , Z 13 , Z 14 ,..., Z 89 ; find the maximum value in [Z 12 , Z 13 , Z 14 ,..., Z 89 , and assume the maximum value is Z 67, it indicates that the prediction accuracy of one of the initial prediction results 6 and initial prediction result 7 is poor. Among them, the discrimination degrees related to the initial prediction result 6 are [Z 16 , Z 26 , Z 36 , Z 46 , Z 67， Z 68 , Z 69 , and the discrimination degrees related to the initial prediction result 7 are [Z 17 , Z 27 , Z 37 , Z 47 , Z 67， Z 78 , Z 79 . Compare the magnitudes of the averages of these two groups. If the average of the initial prediction result 6 is greater than the average of the initial prediction result 7, then the initial prediction result 6 is eliminated and deleted.

[0059] According to the above, in the first-level elimination strategy, the initial prediction result of the 5th preset neural network model is deleted, and in the second-level elimination strategy, the initial prediction result of the 6th preset neural network model is deleted. The final prediction result is [X 1001 , X 1002 , X 1003 ,..., X 1100 Taking X 1001 as an example to illustrate the generation method of the final sequence.

[0060]

[0061] Other results are deduced by analogy.

[0062] A method for predicting a frequency modulation sequence with a result selection-based combined model provided by the present invention predicts a frequency modulation command sequence by using multiple groups of preset neural network models. This method can integrate the non-linear modeling capabilities of different models; calculating the average prediction sequence of multiple groups of initial prediction results helps to smooth short-term fluctuations and highlight long-term trends. By calculating the deviation degree between the initial prediction result and the average prediction sequence and performing first-level elimination based on the deviation degree, models with inaccurate long-term trend predictions can be eliminated and those models that can better capture long-period dependencies can be retained. On the basis of the first-level elimination, calculate the discrimination degree between the initial prediction results of the remaining models and perform second-level elimination based on the discrimination degree, which helps to eliminate those models with highly similar prediction results and lack of complementarity, so as to retain a more diverse and complementary model combination. Calculate the deviation degree weights of the remaining models after the second-level elimination and perform weighted fusion on the initial prediction results based on these weights to avoid the limitations of a single model and achieve the dynamic adaptation of the combined model. Embodiment 2 ​According to Figure 4 As shown, the present invention also provides a result - selective combined model FM sequence prediction system, including: An initial prediction module 1, configured to use multiple groups of initial preset neural network models to predict a frequency - modulation instruction sequence, and correspondingly generate multiple groups of initial prediction results; An average - value calculation and processing module 2, configured to calculate an average prediction sequence based on multiple groups of initial prediction results; A first - level elimination processing module 3, configured to calculate the deviation degree between multiple groups of initial prediction results and the average prediction sequence, and perform a first - level elimination strategy on multiple groups of initial preset neural network models based on the deviation degree to obtain a first - level preset neural network model set; A second - level elimination processing module 4, configured to calculate the difference degree between every two of the first - level preset neural network model set, and perform a second - level elimination strategy on the first - level preset neural network model set based on the difference degree to obtain a second - level preset neural network model set; A prediction sequence generation module 5, configured to calculate the deviation - degree weight of the second - level preset neural network model set, and perform weighted fusion on the second - level preset neural network model set based on the deviation - degree weight to generate a final prediction sequence.

[0063] Embodiment 3 The present invention also provides a mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, such as a result - selective combined model FM sequence prediction program.

[0064] When the processor executes the computer program, it implements the above - mentioned result - selective combined model FM sequence prediction method, for example: Use multiple groups of initial preset neural network models to predict a frequency - modulation instruction sequence, and correspondingly generate multiple groups of initial prediction results; Calculate an average prediction sequence based on multiple groups of initial prediction results; Calculate the deviation degree between multiple groups of initial prediction results and the average prediction sequence, and perform a first - level elimination strategy on multiple groups of initial preset neural network models based on the deviation degree to obtain a first - level preset neural network model set; Calculate the difference degree between every two of the first - level preset neural network model set, and perform a second - level elimination strategy on the first - level preset neural network model set based on the difference degree to obtain a second - level preset neural network model set; Calculate the deviation - degree weight of the second - level preset neural network model set, and perform weighted fusion on the second - level preset neural network model set based on the deviation - degree weight to generate a final prediction sequence.

[0065] Or, when the processor executes the computer program, it implements the functions of each module in the above - mentioned system, for example: An initial prediction module 1, configured to use multiple groups of initial preset neural network models to predict a frequency modulation instruction sequence, and correspondingly generate multiple groups of initial prediction results; An average value calculation and processing module 2, configured to calculate an average prediction sequence based on multiple groups of initial prediction results; A first-level elimination processing module 3, configured to calculate the deviation degrees of multiple groups of initial prediction results from the average prediction sequence, and perform a first-level elimination strategy on multiple groups of initial preset neural network models based on the deviation degrees to obtain a first-level preset neural network model set; A second-level elimination processing module 4, configured to calculate the difference degrees between any two of the first-level preset neural network model set, and perform a second-level elimination strategy on the first-level preset neural network model set based on the difference degrees to obtain a second-level preset neural network model set; A prediction sequence generation module 5, configured to calculate the deviation degree weights of the second-level preset neural network model set, and perform weighted fusion on the second-level preset neural network model set based on the deviation degree weights to generate a final prediction sequence.

[0066] Exemplarily, the computer program may be divided into one or more modules / units. The one or more modules / units are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, and these instruction segments are used to describe the execution process of the computer program in the mobile terminal.

[0067] For example, the computer program may be divided into an initial prediction module 1, an average value calculation and processing module 2, a first-level elimination processing module 3, a second-level elimination processing module 4, and a prediction sequence generation module 5; The specific functions of each module are as follows: An initial prediction module 1, configured to use multiple groups of initial preset neural network models to predict a frequency modulation instruction sequence, and correspondingly generate multiple groups of initial prediction results; An average value calculation and processing module 2, configured to calculate an average prediction sequence based on multiple groups of initial prediction results; A first-level elimination processing module 3, configured to calculate the deviation degrees of multiple groups of initial prediction results from the average prediction sequence, and perform a first-level elimination strategy on multiple groups of initial preset neural network models based on the deviation degrees to obtain a first-level preset neural network model set; A second-level elimination processing module 4, configured to calculate the difference degrees between any two of the first-level preset neural network model set, and perform a second-level elimination strategy on the first-level preset neural network model set based on the difference degrees to obtain a second-level preset neural network model set; The prediction sequence generation module 5 is used to calculate the deviation weight of the second-level preset neural network model set, perform weighted fusion on the second-level preset neural network model set based on the deviation weight, and generate a final prediction sequence.

[0068] The mobile terminal can be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The mobile terminal may include, but is not limited to, a processor and a memory.

[0069] The processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor, or the processor may also be any conventional processor, etc. The processor is the control center of the mobile terminal, and connects various parts of the entire mobile terminal through various interfaces and lines.

[0070] The memory can be used to store the computer program and / or module. The processor realizes various functions of the mobile terminal by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory.

[0071] The memory mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0072] Embodiment 4 The present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, it implements the method for predicting a frequency modulation sequence of a result-selective combination model.

[0073] If the modules / units integrated in the mobile terminal are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0074] Based on such an understanding, to implement all or part of the processes in the above method, the present invention can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the above-mentioned aggregated reinforcement learning resource scheduling method can be realized. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file or some intermediate form, etc.

[0075] The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0076] It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0077] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: the specific implementation manners of the present invention can still be modified or equivalently replaced, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for predicting a combined model frequency modulation sequence with result selection, characterized in that Including: Predicting a frequency modulation instruction sequence by using multiple groups of initial preset neural network models, and correspondingly generating multiple groups of initial prediction results; Calculating an average prediction sequence based on the multiple groups of initial prediction results; Calculating the deviation degrees between the multiple groups of initial prediction results and the average prediction sequence, and performing a first-level elimination strategy on the multiple groups of initial preset neural network models based on the deviation degrees to obtain a first-level preset neural network model set; Calculating the difference degrees between every two of the first-level preset neural network model set, and performing a second-level elimination strategy on the first-level preset neural network model set based on the difference degrees to obtain a second-level preset neural network model set; Calculating the deviation degree weights of the second-level preset neural network model set, and performing weighted fusion on the second-level preset neural network model set based on the deviation degree weights to generate a final prediction sequence.

2. The method for predicting a frequency modulation sequence of a result-selective combined model according to claim 1, characterized in that The multiple groups of initial preset neural network models include at least 5 groups of models among Informer model, Autoformer model, N-HiTS model, PatchTST model, TimeGrad model, DLinear model, TFT model, Graph WaveNet model, TimesNet model or TiDE model.

3. A method for predicting a frequency modulation sequence of a result-selective combined model according to claim 1, characterized in that, The calculation formula of the average prediction sequence includes: Among them, M is the number of groups of the initial preset neural network models currently participating in the calculation; is the i initial prediction result of the i th initial preset neural network model.

4. A method for predicting a frequency modulation sequence of a result-selective combined model according to claim 1, characterized in that, In the step of calculating the deviation degrees between the multiple groups of initial prediction results and the average prediction sequence, and performing a first-level elimination strategy on the multiple groups of initial preset neural network models based on the deviation degrees to obtain a first-level preset neural network model set, the calculation formula of the deviation degree Including: Among them, L is the total length of the prediction sequence; is the initial prediction result of the i th initial preset neural network model; is the average prediction sequence; h is the first prediction result of the prediction sequence; N is the number of initial prediction results; is the deviation degree of the i th neural network model; || is the absolute value; The first-level elimination strategy includes: Comparing the deviation degrees of the multiple groups of initial preset neural network models, deleting the neural network model corresponding to the maximum deviation degree to obtain a first-level preset neural network model set; When the maximum deviation degrees in the multiple groups of initial preset neural network models are the same, randomly deleting any one of the neural network models corresponding to the multiple same maximum deviation degrees to obtain a first-level preset neural network model set.

5. A method for predicting a frequency modulation sequence of a result-selective combined model according to claim 1, characterized in that, In the step of calculating the difference degrees between every two of the first-level preset neural network model set, and performing a second-level elimination strategy on the first-level preset neural network model set based on the difference degrees to obtain a second-level preset neural network model set, the calculation formula of the difference degree Including: Among them, is the initial prediction result of the i th initial preset neural network model; is the initial prediction result of the j th initial preset neural network model; h is the first prediction result of the prediction sequence; N is the number of initial prediction results; is the difference degree between the i th and the j th initial preset neural network models; ∣∣ represents the absolute value; The second-level elimination strategy includes: Comparing the difference degrees between every two of the first-level preset neural network model set, and determining two groups of first-level preset neural network models corresponding to the maximum difference degree among all the difference degrees; Based on the two groups of first-level preset neural network models corresponding to the maximum difference degree, respectively calculating the average value of the difference degrees associated with the initial prediction results of the two groups of first-level preset neural network models, and deleting the group of first-level preset neural network models with the larger average value to obtain a second-level preset neural network model set.

6. The method for predicting a frequency modulation sequence of a result-selective combined model according to claim 1, wherein The calculation formula of the deviation degree weights includes: Among them, S is a set of second-level preset neural network models; is the deviation degree of the i th neural network model; k is any neural network model; is the weight.

7. A method for predicting a frequency modulation sequence of a result-selective combined model according to claim 1, characterized in that The calculation formula of the weighted fusion includes: Among them, is the prediction result of the remaining neural networks; is the weight; is the weighted fusion result; i represents the i th initial preset neural network model in the neural network model set.

8. A result-selective combined model frequency modulation sequence prediction system, characterized in that, Including: An initial prediction module, configured to predict a frequency modulation instruction sequence by using multiple groups of initial preset neural network models, and correspondingly generate multiple groups of initial prediction results; An average value calculation and processing module, configured to calculate an average prediction sequence based on the multiple groups of initial prediction results; The first-level elimination processing module is used to calculate the deviation degrees of multiple groups of initial prediction results from the average prediction sequence, and perform a first-level elimination strategy on multiple groups of initial preset neural network models based on the deviation degrees to obtain a first-level preset neural network model set; The second-level elimination processing module is used to calculate the difference degrees between every two of the first-level preset neural network model set, and perform a second-level elimination strategy on the first-level preset neural network model set based on the difference degrees to obtain a second-level preset neural network model set; The prediction sequence generation module is used to calculate the deviation degree weights of the second-level preset neural network model set, and perform weighted fusion on the second-level preset neural network model set based on the deviation degree weights to generate a final prediction sequence.

9. A mobile terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the result-selective combined model frequency modulation sequence prediction method according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the result-selective combined model frequency modulation sequence prediction method according to any one of claims 1-7.

Images