A method and device for predicting the occurrence time of short-wave signals

Abstract

The present invention discloses a method and device for predicting the occurrence time of short-wave signals, belonging to the technical field of time series prediction. The prediction method includes a VAR calculation step, a KDE calculation step, and a prediction step. The present invention constructs a multi-dimensional feature vector based on time mapping, and combines the results of KDE and VAR to obtain a probability score that can characterize whether a certain time point appears, solving the problem in the prior art that the result deviates from the historical behavior due to not being linked with new data or relying too much on new data, and can give a more accurate prediction result based on historical behavior while taking into account new data.

Description

Technical Field

[0001] The present invention belongs to the technical field of time series prediction, and particularly relates to a method and device for predicting the occurrence time of short-wave signals. Background Art

[0002] Short-wave communication, also known as high-frequency (HF) communication, uses a frequency range of 3 MHz - 30 MHz and mainly utilizes skywaves reflected by the ionosphere for propagation. Since it can achieve long-distance communication without the need to establish relay stations, it has an irreplaceable position in the field of communication. Despite the continuous emergence of current new radio communication systems, this ancient and traditional communication method of short waves still receives widespread attention worldwide and is still developing rapidly.

[0003] In the field of military communication, the advantages of short-wave communication in terms of anti-destruction ability and autonomous communication ability make it the last line of defense for ensuring communication and occupy an important strategic position. When faced with a large amount of data, it is impossible to continuously monitor the frequencies of interest at all times, so it is necessary to analyze its historical behavior for prediction. Existing systems all have some inevitable problems, that is, the results are not linked to new data or overly rely on new data, resulting in the problem that the results deviate from historical behavior, and it is impossible to give a more accurate prediction result based on historical behavior while taking into account new data. Summary of the Invention

[0004] The purpose of the present invention is to overcome one or more deficiencies of the prior art and provide a method and device for predicting the occurrence time of short-wave signals.

[0005] The purpose of the present invention is achieved through the following technical solutions: A method for predicting the occurrence time of short-wave signals includes a VAR calculation step, a KDE calculation step, and a prediction step;

[0006] The VAR calculation step includes:

[0007] Obtain sample data and historical data;

[0008] Generate a unique identifier based on the sample data, and aggregate the historical data based on the identifier;

[0009] Perform time mapping on the aggregated data to obtain a 1*N-dimensional time series set, where N is the number of samples;

[0010] Perform kernel density estimation on the time series set based on the Gaussian function, and perform CUT on the result of the kernel density estimation to obtain a 1*(24*K)-dimensional first vector;

[0011] The KDE calculation step includes:

[0012] Obtain sample data and historical data;

[0013] Generate a unique identifier based on the sample data, and aggregate the historical data based on the identifier;

[0014] Perform time mapping on the aggregated data to obtain a second vector of dimension D*(24*K), where D is the number of days in the sample and K is the number of minute values within one hour during time mapping;

[0015] Perform M-step prediction on the second vector based on the vector autoregressive model to obtain a third vector of dimension M*(24*K), and perform CUT-RMS on the third vector to obtain a fourth vector of dimension 1*(24*K);

[0016] The prediction step includes:

[0017] Sum the first vector and the fourth vector to obtain a fifth vector of dimension 1*(24*K);

[0018] Obtain the probability value of communication occurring at a given moment according to the fifth vector. If the probability value is greater than the first threshold, the signal will not appear at the given moment; otherwise, the signal will appear at the given moment.

[0019] Further, perform time mapping on the aggregated data to obtain a time series set of dimension 1*N, including:

[0020] Sort, perform time mapping, and construct the time series set on the aggregated data to obtain a time series set of dimension 1*N.

[0021] Further, perform CUT on the result of kernel density estimation to obtain a first vector of dimension 1*(24*K), including:

[0022] Multiply the result of kernel density estimation by a first preset value;

[0023] Scale the result of kernel density estimation multiplied by the first preset value to the interval [0,1] through normalization;

[0024] Set all values less than the second threshold in the normalized result of kernel density estimation to 0 to obtain a second vector of dimension 1*(24*K).

[0025] Further, perform time mapping on the aggregated data to obtain a second vector of dimension D*(24*K), including:

[0026] Sort, perform time mapping, perform frequency counting, and calculate the time decay weight on the aggregated data to obtain a second vector of dimension D*(24*K).

[0027] Further, the calculation formula for the time decay weight is:

[0028]

[0029] Among them, d is the data of the d-th day among D samples, and w is positively correlated with d.

[0030] Further, perform CUT-RMS on the third vector to obtain a fourth vector of 1*(24*K) dimensions, including:

[0031] Set all values less than 0 in the third vector to 0;

[0032] Calculate the root mean square error of the third vector to obtain a fourth vector of 1*(24*K) dimensions.

[0033] Further, obtain the probability value of communication occurring at a given moment according to the fifth vector, including:

[0034] Process the fifth vector to obtain a sixth vector of 24*K dimensions;

[0035] Obtain the probability value of communication occurring at a given moment according to the sixth vector.

[0036] Further, process the fifth vector to obtain a sixth vector of 24*K dimensions, including:

[0037] Perform normalization processing on the fifth vector;

[0038] Perform a Reshape operation on the normalized fifth vector to make it a sixth vector of 24*K dimensions;

[0039] Among them, 24 represents 24 hours of a day, K represents K preset minute values included after time mapping for each hour, and the value at its position represents the probability value of communication occurring at this time point.

[0040] A prediction device for the occurrence time of shortwave signals, including:

[0041] A VAR calculation module, configured to obtain sample data and historical data; generate a unique identifier based on the sample data, and aggregate the historical data based on the identifier; perform time mapping on the aggregated data to obtain a time series set of 1*N dimensions, where N is the number of samples; perform kernel density estimation on the time series set based on the Gaussian function, and perform CUT on the result of the kernel density estimation to obtain a first vector of 1*(24*K) dimensions;

[0042] A KDE calculation module is used to obtain sample data and historical data; generate a unique identifier based on the sample data, and aggregate the historical data based on the identifier; perform time mapping on the aggregated data to obtain a second vector of D*(24*K) dimension, where D is the number of days contained in the sample, and K is the number of minute values ​​within an hour during time mapping; perform M-step prediction on the second vector based on a vector autoregression model to obtain a third vector of M*(24*K) dimension, and perform CUT-RMS on the third vector to obtain a fourth vector of 1*(24*K) dimension;

[0043] The prediction module is used to sum the first vector and the fourth vector to obtain a fifth vector of 1*(24*K) dimension; obtain the probability value of communication occurring at a given time according to the fifth vector, if the probability value is greater than a first threshold value, the signal will not appear at the given time, otherwise the signal will appear at the given time.

[0044] The beneficial effects of the present invention are:

[0045] (1) The present invention constructs a multidimensional feature vector based on time mapping, and combines the results of KDE (kernel density estimation) and VAR (vector autoregression model) to obtain a probability score that can characterize whether a certain time point occurs. This solves the problem in the prior art that the results deviate from historical behavior due to lack of linkage with new data or excessive reliance on new data. It can give a more accurate prediction result based on historical behavior while taking into account new data.

[0046] (2) The present invention can effectively solve the problem that only focusing on the hours will lead to reduced prediction accuracy by mapping the minutes contained in the time in the original data, and solve the problem that the data is too discrete when accurate to each minute point and is not convenient for model reasoning and prediction;

[0047] (3) The present invention uses the time set after time mapping and removing the date as the input of KDE, and then estimates the Gaussian kernel density at each time point and cuts the values ​​below the threshold by a preset threshold. The final prediction result can be calibrated based on the historical occurrence, so that the prediction accuracy is greatly improved;

[0048] (4) The present invention aggregates historical data through identification, then performs time mapping on each group of aggregated data and distinguishes and sorts them by date, and constructs a D*144-dimensional feature vector based on the frequency of occurrence of each mapped time point, solving the problem that discrete single-feature time series data is difficult to structure input;

[0049] (5) The present invention performs a time-based weight decay calculation on the constructed D*144-dimensional feature vector, so that when the future trend is predicted by the vector autoregression algorithm, the prediction result can be linked with the new data and new trend;

[0050] (6) The present invention solves the phenomenon of floating accuracy of single-step prediction by performing negative value clipping and root mean square error calculation on the multi-step prediction results of the vector autoregressive algorithm, and generating a unique prediction result by combining the M-step prediction results.

[0051] (7) By superimposing the Gaussian kernel density of historical occurrences and the prediction results of the vector autoregressive algorithm, the present invention combines the analysis and prediction of historical occurrences from different angles, and can accurately predict the occurrence trend in the next 24 hours based on historical behaviors while taking into account new data. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flowchart of an embodiment of the prediction method in the present invention;

[0053] Figure 2 is a flowchart of an embodiment of the VAR calculation steps in the present invention;

[0054] Figure 3 is a flowchart of an embodiment of the KDE calculation steps in the present invention;

[0055] Figure 4 is a flowchart of an embodiment of the prediction steps in the present invention;

[0056] Figure 5 is a schematic diagram of the effect of the shortwave signal occurrence time prediction method based on VAR and KDE;

[0057] Figure 6 is a block diagram of an embodiment of the prediction device in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. 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 skilled in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0059] A prediction method for the occurrence time of shortwave signals, as Figure 1 shown, the method includes VAR calculation steps, KDE calculation steps, and prediction steps, which will be described in detail below.

[0060] As Figure 2 shown, the VAR calculation steps include:

[0061] S110. Obtain sample data and historical data.

[0062] Specifically, the historical data is the time data of the signal appearance within a preselected historical time period; the sample data is the historical data with signal type, signal frequency point, signal appearance time, and disappearance time.

[0063] S120. Generate a unique identifier based on the sample data, and aggregate the historical data based on the identifier.

[0064] Generally, the identifier is a combination of signal type and signal frequency point.

[0065] S130. Perform time mapping on the aggregated data to obtain a 1*N - dimensional time series set, where N is the number of samples.

[0066] In some embodiments, the method for generating the time series set is as follows: perform processing such as sorting, time mapping, and time series set construction on the aggregated data to obtain a 1*N - dimensional time series set. Sorting is arranging in the chronological order of their communication occurrence; time mapping is mapping the original time data to only six minute values of 00, 10, 20, 30, 40, and 50 per hour, and ignoring the date contained in the sample when constructing the time series set, to obtain a 1*N time series set. Specifically, map the original time data to only K minute values per hour to obtain a 1*N time series set, where N is the number of samples. For example, map the original time data to only six minute values of 00, 10, 20, 30, 40, and 50 per hour (at this time, the value of K is 6), and ignore the date contained in the sample when constructing the time series set, to obtain a 1*N time series set, where N is the number of samples.

[0067] S140. Perform kernel density estimation on the time series set based on the Gaussian function, and perform CUT on the result of the kernel density estimation to obtain a 1*(24*K) - dimensional first vector.

[0068] The meaning of CUT in this step is to perform clipping on the result.

[0069] For example, if the value of K is 6, the dimension of the first vector is 1*144 - dimensional.

[0070] In some embodiments, performing CUT on the result of the kernel density estimation to obtain a 1*(24*K) - dimensional first vector includes:

[0071] S141. Multiply the result of the kernel density estimation by a first preset value.

[0072] S142. Scale the result of the kernel density estimation multiplied by the first preset value to the interval [0, 1] through a normalization method.

[0073] S143. Set all values less than the second threshold in the result of the normalized kernel density estimation to 0 to obtain a second vector of dimension 1*(24*K).

[0074] For example, the second threshold is 0.8, that is, based on the second threshold, the result of the normalized kernel density estimation is cropped, and all values less than 0.8 in the result of the normalized kernel density estimation are set to 0.

[0075] As Figure 3 shown, the KDE calculation steps include:

[0076] S210. Obtain sample data and historical data.

[0077] Specifically, the historical data is the time data of the signal appearance within a preselected historical time period; the sample data is the historical data with signal type, signal frequency point, signal appearance time, and disappearance time.

[0078] S220. Generate a unique identifier based on the sample data and aggregate the historical data based on the identifier.

[0079] Generally, the identifier is a combination of signal type and signal frequency point.

[0080] S230. Perform time mapping on the aggregated data to obtain a second vector of dimension D*(24*K), where D is the number of days in the sample and K is the number of minute values within one hour during time mapping.

[0081] In some embodiments, the method for generating the second vector is: perform processing such as sorting, time mapping, frequency counting, and time decay weight calculation on the aggregated data to obtain a second vector of dimension D*(24*K). Sorting is to arrange in the chronological order of their communication, sort each group of sample data after aggregation based on the connection time, and map the original time data to only six minute values of 00, 10, 20, 30, 40, and 50 per hour; since each of the 24 hours after mapping has six minute values, the obtained time mapping result has a total of 144 dimensions, and a feature vector of D*144 dimensions can be obtained, where D is the number of days in the sample; the value of each dimension is the number of samples mapped to this interval; for D sample data, calculate the decay weight based on time.

[0082] Specifically, sort the aggregated data based on the occurrence time and map the original time data so that each hour only contains K minute values, obtaining a second vector of dimension D*(24*K), where D is the number of days in the sample. For example, sort the aggregated data based on the occurrence time and map the original time data so that each hour only contains six minute values of 00, 10, 20, 30, 40, and 50 (in this case, the value of K is 6), obtaining a second vector of dimension D*144, where D is the number of days in the sample, and the value of each dimension is the number of samples mapped to that interval.

[0083] The calculation formula for the time decay weight is:

[0084]

[0085] Among them, d is the data of the d-th day among the D samples, and w is positively correlated with d.

[0086] S240. Perform M-step prediction on the second vector based on the vector autoregressive model to obtain a third vector of dimension M*(24*K), and perform CUT-RMS on the third vector to obtain a fourth vector of dimension 1*(24*K).

[0087] The meaning of CUT-RMS in this step is the root mean square error after clipping.

[0088] For example, if the value of M is 3 and the value of K is 6, the dimension of the third vector is 3*144, and the dimension of the fourth vector is 1*144.

[0089] In some embodiments, the method for generating the fourth vector is: set all values less than 0 in the third vector to 0; then calculate the root mean square error of the third vector, thereby obtaining a fourth vector of dimension 1*(24*K).

[0090] As Figure 4 and Figure 5 shown, the prediction step includes:

[0091] S310. Sum the first vector and the fourth vector to obtain a fifth vector of dimension 1*(24*K).

[0092] For example, if the value is 6, the dimension of the fifth vector is 1*144.

[0093] S320. Obtain the probability value of communication occurring at a given moment according to the fifth vector. If the probability value is greater than the first threshold, the signal will not appear at the given moment, otherwise the signal will appear at the given moment.

[0094] In some embodiments, obtaining the probability value of communication occurring at a given moment according to the fifth vector includes:

[0095] Process the fifth vector to obtain a 24*K-dimensional sixth vector.

[0096] In some embodiments, perform a Reshape operation on the fifth vector to make it a 24*K-dimensional sixth vector. Specifically, perform normalization processing on the fifth vector; perform a Reshape operation on the normalized fifth vector to make it a 24*K-dimensional sixth vector; where 24 represents 24 hours of a day, and K represents K preset minute values included after time mapping for each hour, and the value at its position represents the probability value of communication connection occurring at this time point.

[0097] For example, if the value of is 6, the dimension of the sixth vector is 24*6-dimensional. At this time, processing the fourth vector to obtain a 24*6-dimensional sixth vector includes: performing normalization processing on the fourth vector; performing a Reshape operation on the normalized fourth vector to make it a 24*6-dimensional sixth vector; where 24 represents 24 hours of a day, and 6 represents 6 preset minute values included after time mapping for each hour, and the value at its position represents the probability value of communication connection occurring at this time point.

[0098] S322. Obtain the probability value of communication connection occurring at a given moment according to the sixth vector.

[0099] A prediction device for the appearance time of shortwave signals, as Figure 6 shown, includes a VAR calculation module, a KDE calculation module, and a prediction module.

[0100] The VAR calculation module is used to obtain sample data and historical data; generate a unique identifier based on the sample data, and aggregate the historical data based on the identifier; perform time mapping on the aggregated data to obtain a 1*N-dimensional time series set, where N is the number of samples; perform kernel density estimation on the time series set based on the Gaussian function, and perform CUT on the result of the kernel density estimation to obtain a 1*(24*K)-dimensional first vector.

[0101] The KDE calculation module is used to obtain sample data and historical data; generate a unique identifier based on the sample data, and aggregate the historical data based on the identifier; perform time mapping on the aggregated data to obtain a D*(24*K)-dimensional second vector, where D is the number of days included in the sample, and K is the number of minute values within an hour during time mapping; perform M-step prediction on the second vector based on the vector autoregressive model to obtain an M*(24*K)-dimensional third vector, and perform CUT-RMS on the third vector to obtain a 1*(24*K)-dimensional fourth vector.

[0102] The prediction module is used to sum the first vector and the fourth vector to obtain a fifth vector of 1*(24*K) dimensions; and obtain the probability value of communication occurring at a given moment according to the fifth vector. If the probability value is greater than the first threshold, the signal will not appear at the given moment, otherwise the signal will appear at the given moment.

[0103] The above are only the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein, should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be changed within the scope of the concept described herein through the above teachings or the techniques or knowledge in related fields. And the changes and alterations made by those skilled in the art without departing from the spirit and scope of the present invention should fall within the protection scope of the appended claims of the present invention.

Claims

1. A prediction method for the occurrence time of short-wave signals, characterized in that, It includes a VAR calculation step, a KDE calculation step, and a prediction step; The VAR calculation step includes: Obtain sample data and historical data; Generate a unique identifier based on the sample data, and aggregate the historical data based on the identifier; Perform time mapping on the aggregated data to obtain a 1*N-dimensional time series set, where N is the number of samples; Perform kernel density estimation on the time series set based on the Gaussian function, and perform CUT on the result of the kernel density estimation to obtain a 1*(24*K)-dimensional first vector; The KDE calculation step includes: Obtain sample data and historical data; Generate a unique identifier based on the sample data, and aggregate the historical data based on the identifier; Perform time mapping on the aggregated data to obtain a D*(24*K)-dimensional second vector, where D is the number of days contained in the sample, and K is the number of minute values within one hour during time mapping; Perform M-step prediction on the second vector based on the vector autoregressive model to obtain an M*(24*K)-dimensional third vector, and perform CUT-RMS on the third vector to obtain a 1*(24*K)-dimensional fourth vector; The prediction step includes: Sum the first vector and the fourth vector to obtain a 1*(24*K)-dimensional fifth vector; Obtain the probability value of communication occurring at a given moment according to the fifth vector. If the probability value is greater than the first threshold, the signal will not appear at the given moment, otherwise the signal will appear at the given moment; Performing CUT on the result of the kernel density estimation to obtain a 1*(24*K)-dimensional first vector includes: multiplying the result of the kernel density estimation by a first preset value; scaling the result of the kernel density estimation after multiplying by the first preset value to the interval [0,1] by a normalization method; setting all values less than the second threshold in the normalized result of the kernel density estimation to 0 to obtain a 1*(24*K)-dimensional second vector; Performing CUT-RMS on the third vector to obtain a 1*(24*K)-dimensional fourth vector includes: setting all values less than 0 in the third vector to 0; calculating the root mean square error of the third vector to obtain a 1*(24*K)-dimensional fourth vector; Obtaining the probability value of communication occurring at a given moment according to the fifth vector includes: processing the fifth vector to obtain a 24*K-dimensional sixth vector; obtaining the probability value of communication occurring at a given moment according to the sixth vector; Processing the fifth vector to obtain a 24*K-dimensional sixth vector includes: performing normalization processing on the fifth vector; performing a Reshape operation on the normalized fifth vector to make it a 24*K-dimensional sixth vector; where 24 represents 24 hours of a day, and K represents K preset minute values included after time mapping for each hour, and the value at its position represents the probability value of communication occurring at that moment.

2. The prediction method for the occurrence time of a short-wave signal according to claim 1, wherein Performing time mapping on the aggregated data to obtain a 1*N-dimensional time series set includes: Sort, perform time mapping, and construct a time series set on the aggregated data to obtain a 1*N-dimensional time series set.

3. The prediction method of the appearance time of a short-wave signal according to claim 1, characterized in that, Performing time mapping on the aggregated data to obtain a D*(24*K)-dimensional second vector includes: Sort, time map, frequency count, and calculate the time decay weight for the aggregated data to obtain a second vector of dimension D*(24*K).

4. A prediction method for the occurrence time of short-wave signals according to claim 3, characterized in that, The calculation formula for the time decay weight is as follows: Among them, d is the data on the d-th day among the D samples, and w is positively correlated with d.

5. A prediction device for the occurrence time of short-wave signals, characterized in that, It includes: A VAR calculation module for obtaining sample data and historical data; generating a unique identifier based on the sample data, and aggregating the historical data based on the identifier; Performing a time map on the aggregated data to obtain a time series set of dimension 1*N, where N is the number of samples; performing kernel density estimation on the time series set based on the Gaussian function, and performing CUT on the result of the kernel density estimation to obtain a first vector of dimension 1*(24*K); Among them, performing CUT on the result of the kernel density estimation to obtain a first vector of dimension 1*(24*K) includes: multiplying the result of the kernel density estimation by a first preset value; scaling the result of the kernel density estimation after multiplying by the first preset value to the range [0,1] through a normalization method; setting all values less than a second threshold in the normalized result of the kernel density estimation to 0 to obtain a second vector of dimension 1*(24*K); A KDE calculation module for obtaining sample data and historical data; generating a unique identifier based on the sample data, and aggregating the historical data based on the identifier; performing a time map on the aggregated data to obtain a second vector of dimension D*(24*K), where D is the number of days included in the sample and K is the number of minute values within one hour during the time map; performing M-step prediction on the second vector based on the vector autoregressive model to obtain a third vector of dimension M*(24*K), and performing CUT-RMS on the third vector to obtain a fourth vector of dimension 1*(24*K); Among them, performing CUT-RMS on the third vector to obtain a fourth vector of dimension 1*(24*K) includes: setting all values less than 0 in the third vector to 0; calculating the root mean square error of the third vector to obtain a fourth vector of dimension 1*(24*K); A prediction module for summing the first vector and the fourth vector to obtain a fifth vector of dimension 1*(24*K); obtaining the probability value of communication occurring at a given moment according to the fifth vector, if the probability value is greater than a first threshold, the signal will not appear at the given moment, otherwise the signal will appear at the given moment; Among them, obtaining the probability value of communication occurring at a given moment according to the fifth vector includes: processing the fifth vector to obtain a sixth vector of dimension 24*K; obtaining the probability value of communication occurring at a given moment according to the sixth vector; Among them, processing the fifth vector to obtain a sixth vector of dimension 24*K includes: performing a normalization process on the fifth vector; performing a Reshape operation on the normalized fifth vector to make it a sixth vector of dimension 24*K; where 24 represents 24 hours of a day, and K represents K preset minute values included after time mapping for each hour, and the value at that position represents the probability value of communication occurring at that moment.

Images