A deep-sea tsunami wave detection method and device based on an RBF neural network

Abstract

The application relates to the technical field of ocean monitoring and disaster prevention and reduction, and discloses a deep-sea tsunami wave detection method and device based on an RBF neural network, which comprises the following steps: S1, collecting original data by using a deep-sea seabed pressure sensor, selecting data of specific days from the original data as input data, wherein a single sample of the input data is in a vector form, that is, n sampling points are selected from a fixed time sequence window as an input data vector of a neural network, a time sequence window is slid by a step of 1, sampling point vectors are extracted in sequence to form a training set of the RBF neural network, actual measurement values of a data vector of a sampling point to be predicted corresponding to each input data vector are taken as supervision vector values of each group of input data vectors, all the supervision vector values are taken to form a target set, and the like; the method can improve tidal wave prediction accuracy, reduce the probability of large errors in prediction, thereby reducing false alarm rate and false report rate, and increasing algorithm robustness and intelligent level.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ocean monitoring and disaster prevention and mitigation, and particularly relates to a deep-sea tsunami wave detection method and device based on a RBF neural network. BACKGROUND

[0002] At present, the international and domestic deep-sea tsunami monitoring buoy tsunami wave detection algorithm mostly adopts a polynomial interpolation method. When fitting the tidal wave, the polynomial interpolation method has a residual term. If too many difference orders are used in the polynomial interpolation method, overfitting is easy to occur. If too few difference orders are used in the polynomial interpolation method, the residual term is easy to be truncated and is not easy to be eliminated. Especially, if there is a seabed interference wave, the original smooth tidal wave shallow water wave mode is destroyed, and a large number of prediction errors exceeding 1 cm are generated. Relative to the 3 cm detection threshold of the Pacific tsunami wave, the influence on false alarm and false report is large. Sometimes, manual determination is needed, labor cost is increased, intelligent level is reduced, and misjudgment may affect the disaster response time, causing economic losses and casualties. SUMMARY

[0003] In view of the deficiencies in the prior art, the purpose of the present application is to provide a deep-sea tsunami wave detection method and device based on a RBF neural network, which can solve the problems of improving early warning accuracy and reducing false alarm and false report rates.

[0004] In order to achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0005] A deep-sea tsunami wave detection method based on RBF neural network, comprising: S1, collecting original data by using a deep-sea seabed pressure sensor, selecting data of specific days as input data from the original data, the single sample of the input data is in the form of a vector, that is, n sampling points are selected from a fixed time sequence window as the input data vector of the data neural network, the time sequence window is slid with a step of 1, and sampling point vectors are extracted in turn to form a training set of the RBF neural network, the actual measurement value of the data vector to be predicted corresponding to each input data vector is taken as the supervision vector value of each group of input data vectors, and all the supervision vector values are combined to form a target set; S2, normalizing the data of the training set and the target set; S3, performing Euclidean distance calculation on the normalized training set data and the four-dimensional center vector of the Gaussian kernel function of the hidden layer of the RBF neural network, so as to realize the projection of the input data kernel function to the high-dimensional space of the number of nodes of the hidden layer, and obtain the prediction value of the target value after weighted summation of the output values of the nodes of the hidden layer; S4, using the error sum of squares as the neural network training target function, performing error back propagation, and then adjusting the output layer weight value and the center value and the expansion function value of the Gaussian kernel function of the hidden layer, increasing the number of nodes of the hidden layer, reducing the root mean square error of the neural network prediction value set, and the number of nodes of the hidden layer in the RBF neural network is proportional to the root mean square error of the neural network prediction value set, and the number of nodes of the hidden layer is selected according to the prediction error accuracy; the error back propagation is performed on the target function of the neural network training, and then the output layer weight value and the center value and the expansion function value of the Gaussian kernel function of the hidden layer are adjusted, and the minimum value of the error gradient descent or the number of cyclic iterations of the entire training sample is taken as the external loop stop training condition of the neural network; the batch processing or random processing of all training set samples is taken as the internal loop stop condition; S5, predicting the tidal wave data according to the trained neural network parameters, and comparing the difference between the prediction value of the target value and the target value; S6, comparing at least one of the difference values with a threshold value to determine whether a tsunami wave occurs.

[0006] In the application, preferably, the vector form of the RBF neural network structure is:

[0007] o=f o {B 2 +W 2 radbas(P-XB 1 )};

[0008] Wherein, B 1 represents the inverse vector of the square of the hidden layer expansion constant, B 2 represents the output layer bias vector, W 2 represents the weight vector from the hidden layer to the output layer, and radbas represents the radial basis function.

[0009] The general function form of the RBF neural network structure is:

[0010]

[0011] wherein, H p (t') is a single output value, H * is a small time sampling average value, j is the number of hidden layer nodes, T=[t,t,...,t] T is the input data time reference vector at the current moment, t' is t plus the increase time value of the predicted value time point relative to the start point of the input sampling data, U=[0,1,...n] T is the input vector and the time difference Δt multiple of the current time, c j is the center point vector of the jth radial basis function, σ j is the expansion constant of the jth radial basis function, is the weight coefficient of the ith input to the jth node of the next layer neuron in the lth layer, is the output bias, the target vector dimension is 1, h is the number of hidden layer nodes when the accuracy meets the requirements, f0 is a pureline activation function.

[0012] In the application, preferably, in S4, the loss formula used in the error gradient descent weight and parameter adjustment calculation is:

[0013]

[0014] wherein, E is the error sum of squares, l is the output serial number, d is the supervision value, o is the output value, k is the number of neuron nodes related to the weight;

[0015] The weight adjustment formula is:

[0016]

[0017] wherein, α is the learning rate, ita is the iteration number, j is the layer number, k is the number of neuron nodes related to the weight, is the partial derivative of the prediction error of the lth output node value to the weight coefficient in the ita th iteration.

[0018] In the application, preferably, the specific number of days is at least 59 days, the single data in the original data is the average value of each sampling value within 15 minutes, and the n is taken within 3 to 12.

[0019] In the application, preferably, the deep-sea tsunami wave detection method based on the RBF neural network is realized by C language programming.

[0020] The application discloses a deep-sea tsunami wave detection device based on an RBF neural network, which comprises a collection module, a normalization module, a training module, a prediction comparison module and a judgment and early warning module.The collection module is used for collecting original data by using a deep-sea seabed pressure sensor, selecting data of specific days as input data from the original data, taking n sampling points selected from a fixed time sequence window as an input data vector of a data neural network, sliding the time sequence window by a step of 1, extracting sampling point vectors in turn to form a training set of the RBF neural network, taking actual measurement values of a data vector of a sampling point to be predicted corresponding to each input data vector as a supervision vector value of each group of input data vectors, and forming a target set by using all the supervision vector values.The normalization module is used for performing normalization processing on data of the training set and the target set.The training module is used for performing Euclidean distance calculation on the normalized training set data and a four-dimensional center vector of a Gaussian kernel function of an implicit layer of the RBF neural network, so as to realize projection of input data kernel functions to a high-dimensional space of the number of nodes of the implicit layer, obtain a prediction value of a target value by performing weighted summation on output values of each node of the implicit layer, take an error square sum as a neural network training target function, perform error back propagation, and then adjust output layer weight values and center values and spread function values of the Gaussian kernel function of the implicit layer, increase the number of nodes of the implicit layer, reduce a root mean square error of a neural network prediction value set, and make the number of nodes of the implicit layer of the RBF neural network and the root mean square error of the neural network prediction value set be proportional, so as to select a proper number of nodes of the implicit layer according to prediction error accuracy.The error back propagation is performed on the target function of the neural network training, and then the output layer weight values and the center values and the spread function values of the Gaussian kernel function of the implicit layer are adjusted, so as to take a minimum value of an error gradient or a loop iteration number of the whole training sample as a neural network external loop stop training condition, take batch processing or random processing of all training set samples as an internal loop stop condition, and use the prediction comparison module to predict tide wave data according to trained neural network parameters, compare a difference between a prediction value of the target value and the target value, and use the judgment and early warning module to compare at least one of the differences with a threshold value, and determine whether a tsunami wave occurs.

[0021] In the application, preferably, the vector form of the RBF neural network structure is as follows:

[0022] o=f o {B 2 +W 2 radbas(P-XB 1 )};

[0023] Wherein, B 1 represents an inverse vector of an implicit layer spread constant square, B 2 represents an output layer bias vector, W 2 represents an implicit layer to output layer weight vector, and radbas represents a radial basis function.

[0024] The general function form of the RBF neural network structure is:

[0025]

[0026] Wherein, H p (t') is a single output value, H * is a small time sampling average value, j is the number of hidden layer nodes, T = [t, t, … t] T is the time reference vector of input data at the current time, t' is t plus the increase time value of the predicted value time point relative to the start point of the input sampling data, U = [0, 1, … n] T is the multiple of the time difference Δt between the input vector and the current time, c j is the center point vector of the jth radial basis function, σ j is the expansion constant of the jth radial basis function, is the weight coefficient of the ith input of the lth layer to the jth node of the next layer neuron, is the output bias (the subscript 1 indicates that the target value is 1), h is the number of hidden layer nodes when the accuracy meets the requirements, f0 is a pureline activation function.

[0027] In the present application, preferably, in the training module, the loss formula used for error gradient descent weight and parameter adjustment calculation is:

[0028]

[0029] Wherein, E is the error sum of squares, l is the output serial number, d is the supervision value, o is the output value, k is the number of neuron nodes related to the weight, the target of neural network training is to reduce the target function of the training result as much as possible under the condition that the time and other overheads are not too large, so as to improve the prediction accuracy of the neural network;

[0030] The weight adjustment formula is:

[0031]

[0032] Wherein, α is the learning rate, ita is the iteration number, j is the layer number, k is the number of neuron nodes related to the weight, is the partial derivative of the prediction error of the lth output node value of the ita iteration to the weight coefficient.

[0033] In the present application, preferably, the specific number of days is at least 59 days, the single data in the original data is the average value of each sampling value within 15 minutes, and the n is taken within 3 to 12.

[0034] A computer readable storage medium comprising instructions which, when run on a computer, cause the computer to perform the method of deep-sea tsunami wave detection based on an RBF neural network according to any one of the preceding.

[0035] Compared with the prior art, the present application has the following beneficial effects:

[0036] The present method and device can improve the tidal wave prediction accuracy and reduce the probability of large error in prediction, thereby reducing the false alarm rate and the false report rate, increasing the algorithm robustness and the intelligent level, and embedding the single-chip microcomputer in the buoy by using the computer C language programming to realize the neural network parameter training. The present method and device have higher tsunami wave detection accuracy, and can reduce the probability of more than 1cm error by about 89.7% compared with the prior art, effectively reducing the tsunami wave false alarm and missed report probability. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The flowchart of the method of deep-sea tsunami wave detection based on the RBF neural network according to an embodiment of the present application.

[0038] Figure 2 The structure diagram of the RBF neural network according to an embodiment of the present application.

[0039] Figure 3 The schematic diagram of the seabed measurement tidal wave time axis sampling point and prediction point according to an embodiment of the present application.

[0040] Figure 4 The schematic diagram of the function approximation RBF neural network according to an embodiment of the present application.

[0041] Figure 5 The comparison diagram of the root mean square error of the RBF neural network and the polynomial interpolation method prediction value set according to an embodiment of the present application.

[0042] Figure 6 The effect statistical diagram of the long-term (one year) prediction prediction value set error mean of the RBF neural network and the polynomial interpolation method according to an embodiment of the present application.

[0043] Figure 7 The comparison diagram of the root mean square error of the RBF neural network and the polynomial interpolation method long-term (one year) prediction error value set mean according to an embodiment of the present application.

[0044] Figure 8 The computer C language program screenshot of the method of deep-sea tsunami wave detection based on the RBF neural network according to an embodiment of the present application.

[0045] Figure 9Fig. 1 is a structural schematic diagram of a deep-sea tsunami wave detection device based on an RBF neural network according to another embodiment of the present application.

[0046] In the drawings: 1, acquisition module; 2, normalization module; 3, training module; 4, prediction comparison module; 5, judgment and early warning module. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0048] It should be noted that when a component is referred to as being "fixed" to another component, it can be directly on the other component or there can be intervening components. When a component is referred to as being "connected" to another component, it can be directly connected to the other component or there can be intervening components. When a component is referred to as being "disposed" on another component, it can be directly on the other component or there can be intervening components. The terms "vertical", "horizontal", "left", "right", and similar terms as used herein are for purposes of description only.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description of the application herein is for describing particular embodiments only and is not intended to be limiting of the application. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0050] Please refer to Figure 1 A deep-sea tsunami wave detection method based on an RBF neural network is provided in a preferred embodiment of the present application, comprising:

[0051] S1, collecting original data by using a deep-sea seabed pressure sensor, selecting data of specific days from the original data as input data, a single sample of the input data being in a vector form, i.e. selecting n sampling points from a fixed time sequence window as an input data vector of a data neural network, sliding the time sequence window by a step of 1, extracting sampling point vectors in turn to form a training set of an RBF neural network, an actual measurement value of a data vector of a sampling point to be predicted corresponding to each input data vector being a supervision vector value of each group of input data vectors, and all the supervision vector values forming a target set;

[0052] The deep-sea bottom pressure sensor is used to collect the bottom pressure data for several months or even several years, that is, to obtain the tidal wave data in the local continuous time period. A fixed interval of n sampling points in a period of bottom pressure data is selected as an input data vector, and the input data vector is extracted by sliding the time series window with a step of 1 from the sampling data to form an RBF neural network training set. Each group of input data corresponds to the tidal wave data to be predicted as the prediction object, that is, the target set. According to the propagation law of shallow water waves and deep water waves, it can be known that the deep sea has a filtering effect on high-frequency waves, and the main component of the collected data is the tidal wave, as well as some low-amplitude environmental clutter caused by animal activities. Preferably, the data (i.e. the original data) of the deep-sea bottom pressure sensor is at least 2 months (the shortest continuous 2 natural months is 59 days, and the longest is 62 days). The single data in the original data can use the average value of each sampling value within 15 minutes. The time period sliding window is selected, which can be set to three hours or longer, and the value of the sampling point number n can be greater than or equal to 3. Four sampling points (1 hour interval) are selected from the three hours (calculated by taking one sampling point every 15 minutes, a total of 13 sampling points) as the input data of the RBF neural network training set, and each group of input corresponds to the next sampling point as the prediction object, that is, the target set.

[0053] As shown in the structure shown in Figure 2 , the RBF neural network used here is a three-layer neural network, n, h and k represent the number of input layer, hidden layer and output layer neurons respectively, which correspond to the number of sampling data, the number of hidden layer nodes and the number of target values (supervision vector value dimension).

[0054] S2, normalize the data of the training set and the target set.

[0055] The collected tidal wave data is arranged, and a two-month data segment is taken as a training set, and the tidal wave data of the next prediction time point is taken as a target set. The data is normalized.

[0056] S3, calculate the Euclidean distance between the normalized training set data and the four-dimensional center vector of the Gaussian kernel function of the RBF neural network hidden layer, and then realize the projection of the input data kernel function to the high-dimensional space of the number of hidden layer nodes. The output value of each node in the hidden layer is summed to obtain the prediction value of the target value.

[0057] Specifically, the vector form of the RBF neural network structure is:

[0058] o=f o {B 2 +W 2 radbas(P-XB 1 )};

[0059] where B 1 represents the inverse of the square of the spread constant of the hidden layer, B 2 represents the bias vector of the output layer, W 2 represents the weight vector of the hidden layer to the output layer, radbas represents a radial basis function, as shown in Figure 4

[0060] The general function form of the RBF neural network structure is:

[0061]

[0062] where H p (t') is a single output value, H * is a small time sampling average value, j is the number of hidden layer nodes, T = [t, t,..., t] T is the time reference vector of the input data at the current time, t' is the increase time value of the prediction value time point relative to the start of the input sampling data, U = [0, 1,..., n] T is the multiple of the time difference Δt between the input vector and the current time, Δt can be 1 hour here, c j is the center point vector of the jth radial basis function, σ j is the spread constant of the jth radial basis function, is the weight coefficient of the ith input of the lth layer to the jth node of the next layer neuron, is the output bias (the subscript indicates that the target value is 1, or the target vector dimension is 1), h is the number of hidden layer nodes when the accuracy meets the requirements (the root mean square error changes less than a certain value during training, and the training stops), f0 is a pureline activation function, as shown in Figure 3

[0063] S4, using the error sum of squares as the neural network training objective function (satisfying the prediction accuracy requirement), performing error back propagation, and then adjusting the output layer weight and the center value and spread function value of the hidden layer Gaussian kernel function, increasing the number of hidden layer nodes, the root mean square error of the neural network prediction value set is reduced, the number of hidden layer nodes in the RBF neural network and the root mean square error of the neural network prediction value set are proportional, and the number of hidden layer neuron nodes is selected according to the prediction error accuracy; the error back propagation is performed on the objective function of the neural network training, and then the output layer weight and the center value and spread function value of the hidden layer Gaussian kernel function are adjusted, and the minimum value of the error gradient or the number of loop iterations of the entire training sample is taken as the external loop stop training condition of the neural network; the batch processing or random processing of all training set samples is taken as the internal loop stop condition.

[0064] ​​The four-dimensional samples collected and normalized in S2 are input into the RBF neural network of S3 to calculate the output. Then, the batch processing or single sample updating method is used to perform error gradient descent weight and parameter adjustment calculation.

[0065] Specifically, the target function (loss formula) used in the error gradient descent weight and parameter adjustment calculation is as follows:

[0066]

[0067] wherein E is the error sum of squares, I is the output serial number, d is the supervision value, o is the output value, k is the number of neuron nodes related to the weight, and the target of the neural network training is to reduce the target function of the training result as much as possible under the condition that the time and other overheads are not too large, so as to improve the prediction accuracy of the neural network.

[0068] The weight adjustment formula is as follows:

[0069]

[0070] wherein a is the learning rate, ita is the iteration number, j is the layer number, k is the number of neuron nodes related to the weight, is the partial derivative of the prediction error of the lth output node value in the ita iteration with respect to the weight coefficient.

[0071] The neural network parameters are adjusted so that the number of hidden layer nodes of the neural network is increased in the error gradient descent process, and the target function (error mean square root) is reduced. The above two are directly proportional, the prediction accuracy (error mean square root and ) is measured after the number of hidden layer neurons is increased, and the appropriate network structure, i.e. the number of hidden layer neurons, is selected. The adjustment of the neural network weight and other parameters can be in the form of batch training data or single training data sample training, and the double loop iteration condition is: the training of all samples and the iteration of the training number, until the minimum value of the error mean square root iteration change of the error gradient descent or the maximum value of the iteration number is reached. When both conditions are met, the adjustment cycle of the gradient descent method is exited, and the training of the neural network weight and parameters is completed. The key of the training algorithm technology lies in the adjustment of the training learning rate. A smaller learning rate a is selected for the weight coefficient or parameter that has a greater impact on the training result, and a larger learning rate a is selected for the opposite. The value of a is between 0.08 and 0.15.

[0072] As shown in Figure 5 , Figure 6 and Figure 7 , the error mean square roots of the prediction values of the RBF neural network method and the polynomial interpolation method are compared, the prediction error mean effect statistics of the long-term (one year) prediction are given, and the prediction effect of the RBF neural network is obviously better.

[0073] S5, according to the trained neural network parameters, predicting the tidal wave data, comparing the difference between the predicted value of the target value and the target value.

[0074] That is, the RBF neural network of S3 uses the data normalized by S2 to predict the tidal wave, and the difference is obtained by subtracting the actual measured tidal wave value target value (supervision vector value) from the predicted value.

[0075] S6, compare at least one difference value with a threshold value, and determine whether a tsunami wave occurs.

[0076] The advantages and effects of the method compared with the interpolation polynomial method for tsunami wave detection can be verified by comparing the root mean square error of tidal wave prediction and the occurrence probability of large error greater than 1cm of the same data set, as shown in Tables 1 and 2.

[0077] Root mean square error Polynomial (mm) RBF (mm) 3072 sampling points 3.2 2.8

[0078] Table 1 Comparison of prediction results in January 2010

[0079]

[0080] Table 2 Statistical count of 1cm large error interval of three methods

[0081] Preferably, the deep-sea tsunami wave detection method based on RBF neural network of the embodiment can be realized by C language programming, as shown in Figure 8 Using computer C language, it has the advantage of wide single-chip microcomputer embeddability.

[0082] As shown in Figure 9 The embodiment of the application also provides a deep-sea tsunami wave detection device based on RBF neural network, comprising:

[0083] The acquisition module is used to collect raw data by using a deep-sea seabed pressure sensor, select data of specific days from the raw data as input data, and input data single sample is in the form of a vector, that is, n sampling point data selected from a fixed time sliding window is used as an input data vector of a data neural network, a time series window is slid with a step of 1, and sampling point vectors are extracted in turn to form a training set of the RBF neural network, the actual measurement value of the predicted sampling point data vector corresponding to each input data vector is used as a supervision vector value of each group of input data vectors, and all the supervision vector values form a target set; preferably, the specific days are at least 59 days, the single data in the raw data is the average value of each sampling value within 15 minutes, the time window length of the time sliding window can be 3 hours, and the value of the sampling point number n is 4 (n can be within 3 to 12).

[0084] A normalization module is configured to normalize data of the training set and the target set.

[0085] A training module is configured to perform Euclidean distance calculation on the normalized training set data and four-dimensional center vectors of Gaussian kernel functions of the hidden layer of the RBF neural network, so as to realize projection of input data kernel functions to a high-dimensional space of the number of nodes of the hidden layer, weighted summation of output values of each node of the hidden layer to obtain a predicted value of a target value, increase of the number of nodes of the hidden layer, and reduction of root mean square error of the neural network predicted value set, in which the number of nodes of the hidden layer of the neural network is proportional to the root mean square error of the neural network predicted value set, and a proper number of nodes of the hidden layer of the neural network is selected according to a prediction accuracy requirement and a calculation cost requirement; error square sum is used as a neural network training target function to perform error back propagation, and then center values and spread function values of the output layer weight and the hidden layer Gaussian kernel function are adjusted, a double loop of neural network training is set, a minimum value of error gradient or a loop iteration number of the entire training sample is used as a neural network external loop stop training condition, and batch processing or random processing of all training set samples is used as an internal loop stop condition. The key of the algorithm technology of the training lies in adjustment of a training learning rate, a weight coefficient or a parameter having a greater influence on the training result is selected to have a smaller learning rate α, and a larger learning rate α is selected in the opposite case, and the value of α is between 0.08 and 0.15. A prediction comparison module is configured to predict the tidal wave data according to the trained neural network parameters, and compare the difference between the predicted value of the target value and the target value.

[0086] A judgment and early warning module is configured to compare at least one of the differences with a threshold value to determine whether a tsunami wave occurs.

[0087] In the embodiment, preferably, the vector form of the RBF neural network structure is as follows:

[0088] o=f o {B 2 +W 2 radbas(P-XB 1 )};

[0089] In the formula, B 1 represents an inverse vector of the square of the hidden layer spread constant, B 2 represents an output layer bias vector, W 2 represents a weight vector from the hidden layer to the output layer, and radbas represents a radial basis function.

[0090] The general function form of the RBF neural network structure is as follows:

[0091]

[0092] In the formula, H p (t′) is a single output value, H* is a sampling average value for a short time, j is the number of nodes of the hidden layer, and T = [t, t, …, t] T is a time reference vector of the input data at the current time, t' is t plus an increase time value of the time point of the prediction value relative to the start point of the input sampling data, and U = [0, 1, …, n] T is a multiple of the time difference Δt between the input vector and the current time, and Δt can be 1 hour here, c j is a center point vector of the jth radial basis function, σ j is an expansion constant of the jth radial basis function, is a weight coefficient of the ith input to the jth node of the next layer neuron of the lth layer, is an output bias, h is the number of hidden layer nodes when the accuracy meets the requirements, and f0 is a pureline activation function.

[0093] In this embodiment, preferably, in the training module, the objective function (loss formula) used for error gradient descent weight and parameter adjustment calculation is:

[0094]

[0095] wherein E is the error sum of squares, l is the output serial number, d is the supervision value, o is the output value, k is the number of neuron nodes related to the weight, and the goal of the neural network training is to reduce the objective function of the training result as much as possible under the condition that the calculation time and other overheads are not too large, so as to improve the prediction accuracy of the neural network;

[0096] The weight adjustment formula is:

[0097]

[0098] wherein α is the learning rate, ita is the iteration number, j is the layer number, and k is the number of neuron nodes related to the weight, is the partial derivative of the error of the lth output node value of the ita iteration with respect to the weight coefficient.

[0099] The limitation condition also includes selecting a suitable number of hidden layer neural network nodes based on the prediction accuracy.

[0100] The embodiment of the present application also provides a computer readable storage medium, and the computer readable storage medium stores a computer program. The computer program is executed by a processor to realize each process of the above-mentioned deep-sea tsunami wave detection method based on the RBF neural network, and the same technical effects can be achieved. The computer readable storage medium can be, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, an optical disk, or the like.

[0101] The above description is a detailed description of the preferred embodiment of the present application, but the embodiment is not used to limit the patent application range of the present application. Any equivalent changes or modifications made under the technical spirit of the present application should belong to the patent range covered by the present application.

Claims

1. A deep-sea tsunami wave detection method based on RBF neural network, characterized in that, The method comprises the following steps: S1, collecting original data by using a deep-sea seabed pressure sensor, selecting data of specific days from the original data as input data, a single sample of the input data being in a vector form, that is, n sampling points are selected from a fixed time sequence window as an input data vector of a data neural network, a time sequence window is slid with a step of 1, sampling point vectors are sequentially extracted to form a training set of the RBF neural network, and actual measurement values of a data vector of a sampling point to be predicted corresponding to each input data vector are taken as a supervision vector value of each group of input data vectors, and all the supervision vector values are combined to form a target set; S2, performing normalization processing on the data of the training set and the target set; S3, performing Euclidean distance calculation on the normalized training set data and a four-dimensional center vector of a Gaussian kernel function of an implicit layer of the RBF neural network, so as to realize projection of input data kernel functions to a high-dimensional space of the number of nodes of the implicit layer, and obtain a predicted value of a target value through weighted summation of output values of nodes of the implicit layer; S4, taking a sum of squares of errors as a neural network training target function, performing error back propagation, and then adjusting output layer weights and center values and spread function values of the Gaussian kernel function of the implicit layer, increasing the number of nodes of the implicit layer, reducing a root mean square error of a neural network predicted value set, and being proportional between the number of nodes of the implicit layer in the RBF neural network and the root mean square error of the neural network predicted value set, so as to select the number of nodes of the implicit layer according to a prediction error accuracy; performing error back propagation on the target function of the neural network training, and then adjusting the output layer weights and the center values and the spread function values of the Gaussian kernel function of the implicit layer, taking a minimum value of error gradient descent or a number of cyclic iterations of the entire training sample as a neural network external loop stop training condition; taking batch processing or random processing of all training set samples as an inner loop stop condition; S5, predicting tide wave data according to trained neural network parameters, and comparing a difference between a predicted value of a target value and the target value; S6, comparing at least one of the differences with a threshold value, and determining whether a tsunami wave occurs; a vector form of the RBF neural network structure is as follows: ； wherein, represents an inverse vector of the square of the hidden layer spread constant, represents an output layer bias vector, represents a weight vector from the hidden layer to the output layer, represents a radial basis function; a general function form of the RBF neural network structure is as follows: ； wherein, is a single output value, is a small time sample average value, is a number of hidden layer nodes, is a current time input data time reference vector, is plus a prediction value time point relative to the start of the input sample data increase time value, is a multiple of the time difference between the input vector and the current time , is a center point vector of the th radial basis function, is a spread constant of the th radial basis function, is a weight coefficient of the th input of the layer to the th node of the next layer neuron, is an output bias, the target vector dimension is 1, is a number of hidden layer nodes when the accuracy meets the requirements, is a pureline activation function; in S4, a loss formula used for error gradient descent weight and parameter adjustment calculation is as follows: ； wherein, is the sum of squared errors, is the output sequence number, is the supervisory value, is the output value, is the number of neuron nodes related to the weight value; a weight adjustment formula is as follows: ； wherein, is the learning rate, is the number of iterations, is the layer number, is the number of neuron nodes related to the weight, is the weight coefficient of the th output node value in the th iteration, and is the partial derivative of the prediction error of the 2.The deep-sea tsunami wave detection method based on RBF neural network according to claim 1, wherein, the specific days are at least 59 days, and a single data in the original data is an average value of sampling values in 15 minutes. 3.The deep-sea tsunami wave detection method based on RBF neural network according to claim 1, wherein, The deep-sea tsunami wave detection method based on the RBF neural network is realized by using C language programming.

4. A deep-sea tsunami wave detecting apparatus based on an RBF neural network, characterized by, The method comprises the following steps: a collection module is configured to collect original data by using a deep-sea seabed pressure sensor, select data of specific days from the original data as input data, take a single sample of the input data in a vector form, that is, n sampling points are selected from a fixed time sequence window as an input data vector of a data neural network, a time sequence window is slid with a step of 1, sampling point vectors are sequentially extracted to form a training set of the RBF neural network, actual measurement values of a data vector of a sampling point to be predicted corresponding to each input data vector are taken as a supervision vector value of each group of input data vectors, and all the supervision vector values are combined to form a target set; The normalization module is configured to normalize data of the training set and the target set; The training module is configured to perform Euclidean distance calculation on the normalized training set data and four-dimensional center vectors of Gaussian kernel functions of an RBF neural network hidden layer, so as to realize projection of input data kernel functions to a high-dimensional space of the number of hidden layer nodes, weighted summation of output values of the hidden layer nodes to obtain a predicted value of a target value, and error back propagation using a sum of squared errors as a neural network training objective function, so as to adjust output layer weights and center values and spread function values of the Gaussian kernel functions of the hidden layer, increase the number of hidden layer nodes, reduce a root mean square error of a neural network predicted value set, and make the number of hidden layer nodes in the RBF neural network and the root mean square error of the neural network predicted value set be proportional, so as to select the number of hidden layer neuron nodes according to prediction error accuracy; the error back propagation is performed on the objective function of the neural network training, so as to adjust the output layer weights and the center values and the spread function values of the Gaussian kernel functions of the hidden layer, and a minimum value of error gradient descent or a number of loop iterations of the entire training sample is taken as a neural network external loop stop training condition; Batch processing or random processing of all training set samples is taken as an inner loop stop condition; The prediction comparison module is configured to predict tidal wave data according to trained neural network parameters, and compare a difference between a predicted value of a target value and the target value; The judgment and early warning module is configured to compare at least one of the differences with a threshold value, and determine whether a tsunami wave occurs; The vector form of the RBF neural network structure is: ； wherein, represents an inverse vector of the square of the hidden layer spread constant, represents an output layer bias vector, represents a weight vector from the hidden layer to the output layer, represents a radial basis function; The general function form of the RBF neural network structure is: ； wherein, is a single output value, is a small time sample average value, is the number of hidden layer nodes, is the current time input data time reference vector, is plus the increase time value of the prediction value time point relative to the start of the input sample data, is the multiple of the time difference between the input vector and the current time is the center point vector of the th radial basis function, is the spread constant of the th radial basis function, is the weight coefficient of the th input of the layer to the th node of the next layer neuron, is the output bias, the target vector dimension is 1, is the number of hidden layer nodes when the accuracy meets the requirements, is a pureline activation function;​​ In the training module, A loss formula used for error gradient descent weight and parameter adjustment calculation is: ； wherein, is the sum of squared errors, is the output sequence number, is the supervisory value, is the output value, is the number of neuron nodes related to the weight value; A weight adjustment formula is: ； wherein, is the learning rate, is the number of iterations, is the layer number, is the number of neuron nodes related to the weight, is the weight coefficient of the th iteration of the th output node value.

5. The RBF neural network-based deep-sea tsunami wave detection apparatus of claim 4, wherein, The specific number of days is at least 59 days, each single data in the original data is an average value of sampling values in 15 minutes, and n is selected from 3 to 12.

6. A computer-readable storage medium, characterized in that, The computer readable storage medium includes instructions, when the instructions are executed on the computer, the computer executes the deep-sea tsunami wave detection method based on the RBF neural network as claimed in any one of claims 1 to 3.

Images