Neural network learning and optimization method, system and medium for smart substation based on F5G all-optical network

Abstract

The present invention proposes a neural network learning and optimization method for a smart substation based on an F5G all-optical network, comprising the following steps: S1, using the F5G all-optical network to obtain the collected face image data entering and leaving the smart substation; S2, preprocessing the face images entering and leaving the smart substation obtained by using the F5G all-optical network; S3, constructing the preprocessed face image to realize feature extraction; S4, obtaining the face feature sample according to the face feature sample method; S5, training the face image according to the neural network to learn and optimize the face recognition model; S6, adjusting the convolutional neural network parameters according to the gradient algorithm; S7, inputting the training results into the classifier for classification prediction to realize face discrimination. The present invention can realize the identification of people entering and leaving the smart substation, and improve the accuracy and efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of smart substations, and in particular to a neural network learning and optimization method, system and medium for smart substations based on an F5G all-optical network. Background Art

[0002] As an important part of the power system, smart substations integrate modern communication and automation technologies, use advanced digital control equipment, and have the ability of remote monitoring, remote operation and remote maintenance, achieving high intelligence and autonomous decision-making. The application of F5G all-optical network in smart substations is an important part of the digital and intelligent transformation of the power industry. F5G, the fifth generation fixed network, represents the latest progress in fixed network technology and aims to build a fixed network infrastructure with ultra-wide bandwidth, all-optical connection, ultimate experience, high security and intelligent management. Summary of the invention

[0003] The present invention aims to at least solve the technical problems existing in the prior art, and particularly innovatively proposes a neural network learning and optimization method, system and medium for an intelligent substation based on an F5G all-optical network.

[0004] In order to achieve the above-mentioned object of the present invention, the present invention provides a neural network learning and optimization method for a smart substation based on an F5G all-optical network, comprising the following steps:

[0005] S1, uses F5G all-optical network to obtain the collected facial image data entering and leaving the smart substation;

[0006] S2, pre-processing the facial images entering and leaving the smart substation acquired by using the F5G all-optical network;

[0007] S3, constructing the preprocessed face image to realize feature extraction;

[0008] S4, obtaining a facial feature sample according to a facial feature sample method;

[0009] S5, training face images according to the neural network to learn and optimize the face recognition model;

[0010] S6, adjusting the convolutional neural network parameters according to the gradient algorithm;

[0011] S7, input the training results into the classifier for classification prediction to achieve face recognition.

[0012] In a preferred embodiment of the present invention, in step S2, the method for preprocessing the facial images entering and exiting the smart substation acquired by using the F5G all-optical network includes one or any combination of grayscale, denoising, and normalization to improve image quality and reduce noise interference.

[0013] In a preferred embodiment of the present invention, the method for constructing the preprocessed face image to realize feature extraction in step S3 includes:

[0014] The integer value of the value where the face is not collected is 0, and the value where the face is collected is 1. The integers 0 and 1 are used for encoding. Two integer values ​​are assigned to the dataset to set the mapping of the binary vector length 2, that is, "face not collected" is marked as 0, and "face collected" is marked as 1. If the dataset sequence collects a face, its integer code is 1.

[0015] In a preferred embodiment of the present invention, the facial feature sample method in step S4 is a convolution and pooling method;

[0016] The fully connected layer uses the softmax function to classify the image, the pooling layer uses the RELU function to remove the negative values ​​of the grouped images, and the convolutional layer uses the loss function to regularize the weight of the image.

[0017] In a preferred embodiment of the present invention, step S5 includes:

[0018]

[0019] in, Represents the bias term of the nth feature map in the mth layer;

[0020] Represents the convolution kernel weight between the nth feature map in the mth layer and the jth feature map in the m-1th layer;

[0021] Represents the output of the jth feature map in the m-1th layer;

[0022] represents the output of the nth feature map in the mth layer.

[0023] In a preferred embodiment of the present invention, the loss function is used to learn the optimization model in step S5:

[0024]

[0025] Among them, U represents the loss function;

[0026] G represents the number of samples;

[0027] xi represents the true value;

[0028] Represents the predicted value.

[0029] In a preferred embodiment of the present invention, the method for adjusting the convolutional neural network parameters according to the gradient algorithm in step S6 includes:

[0030] m t =β1m t-1 +(1-β1)g t ,

[0031] Among them, m t Represents the moving average of the gradient of the current iteration;

[0032] β1 represents the first decay rate, which is 0.95 here;

[0033] m t-1 Represents the moving average of the gradient at the t-1 iteration time;

[0034] g t Represents the gradient of the current iteration;

[0035]

[0036] v t Represents the moving average of the square of the gradient of the current iteration;

[0037] β2 represents the second decay rate, which is 0.996 here;

[0038] v t-1 Represents the moving average of the squared gradient at the t-1 iteration;

[0039] g t Represents the gradient of the current iteration;

[0040]

[0041] Represents the modified moving average of the gradient of the current iteration;

[0042] m t Represents the moving average of the gradient of the current iteration;

[0043] Indicates the gradient correction value;

[0044]

[0045] Represents the modified moving average of the squared gradient of the current iteration;

[0046] v t Represents the moving average of the squared gradient at t iterations;

[0047] Represents the squared gradient correction value;

[0048]

[0049] θ t+1 Network parameters indicating the current number of iterations;

[0050] θ t Network parameters representing the last iteration number

[0051] α represents the learning rate;

[0052] Represents the modified moving average of the gradient of the current iteration;

[0053] Represents the modified moving average of the squared gradient of the current iteration;

[0054] ε represents a very small value, which is used to prevent the denominator from being zero. Here, it is 10 -8 .

[0055] In a preferred embodiment of the present invention, step S7 includes:

[0056] The classifier used is the Bayesian classifier;

[0057] Assume that Q is a training set of samples, each sample has its class label, and there are q classes in total, namely D1, D2, D3, ..., D q , each sample is represented by an n-dimensional vector, H = {h1,h2,h3,…,h n}, describing n measured values ​​of n attributes, A1, A2, A3, ..., A n ;

[0058] Given a sample H, the classifier will predict that sample H belongs to the category with the highest posterior probability, conditional on H, that is, predict that sample H belongs to D i , and only when P(D i |H)>P(D j |H), 1≤j≤q, j≠i;

[0059] Thus we find the maximum P(D i |H) class, P(D i |H) The largest class D i It is called the maximum a posteriori hypothesis, and Bayes' theorem is:

[0060]

[0061] Since P(H) is the same for all classes, only P(H|D i )P(D i ) requires a maximum, if the class The prior probability P(D i ) is unknown, then it is usually assumed that these classes are equally likely, that is, P(D1) = P(D2) = P(D3) = ... = P(D q ), so the maximum is P(H|D i ), otherwise the maximum value is P(H|D i )P(D i ).

[0062] The present invention also discloses a neural network learning and optimization system for an intelligent substation based on an F5G all-optical network, including a face image acquisition module, a preprocessing module, a feature extraction module, a face feature sample acquisition module, a face image training optimization module, a network parameter adjustment module, and a face discrimination module;

[0063] The face image acquisition module is used to acquire the face image data of people entering and leaving the smart substation using the F5G all-optical network;

[0064] The preprocessing module is used to preprocess the facial images entering and leaving the smart substation acquired using the F5G all-optical network;

[0065] The feature extraction module is used to construct the preprocessed face image to realize feature extraction;

[0066] The facial feature sample acquisition module is used to obtain facial feature samples according to the facial feature sample method;

[0067] The face image training optimization module is used to train face images based on a neural network to learn and optimize the face recognition model;

[0068] The network parameter adjustment module is used to adjust the convolutional neural network parameters according to the gradient algorithm;

[0069] The face recognition module is used to input the training results into the classifier for classification prediction to achieve face recognition.

[0070] In a preferred embodiment of the present invention, a method for preprocessing facial images entering and exiting a smart substation acquired using an F5G all-optical network in a preprocessing module includes one or any combination of grayscale, denoising, and normalization to improve image quality and reduce noise interference.

[0071] In a preferred embodiment of the present invention, the face image training optimization module includes:

[0072]

[0073] in, Represents the bias term of the nth feature map in the mth layer;

[0074] Represents the convolution kernel weight between the nth feature map in the mth layer and the jth feature map in the m-1th layer;

[0075] Represents the output of the jth feature map in the m-1th layer;

[0076] represents the output of the nth feature map in the mth layer.

[0077] The present invention also discloses a computer system, comprising:

[0078] processor;

[0079] a memory for storing processor-executable instructions;

[0080] Among them, the processor is configured to implement the neural network learning and optimization method of the smart substation based on the F5G all-optical network when executing the executable instructions.

[0081] The present invention also discloses a computer-readable storage medium, comprising:

[0082] a memory having a computer program stored thereon;

[0083] A processor is used to execute the program in the memory to implement the neural network learning and optimization method of the smart substation based on the F5G all-optical network.

[0084] In summary, due to the adoption of the above technical solution, the present invention can identify people entering and leaving the smart substation, thereby improving accuracy and efficiency.

[0085] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0087] Figure 1 It is a schematic block diagram of the process of the present invention. DETAILED DESCRIPTION

[0088] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0089] The present invention provides a neural network learning and optimization method for intelligent substation based on F5G all-optical network, such as Figure 1 As shown, the following steps are included:

[0090] S1, uses F5G all-optical network to obtain the collected facial image data entering and leaving the smart substation;

[0091] S2, pre-processing the facial images entering and leaving the smart substation acquired by using the F5G all-optical network;

[0092] S3, constructing the preprocessed face image to realize feature extraction;

[0093] S4, obtaining a facial feature sample according to a facial feature sample method;

[0094] S5, training face images according to the neural network to learn and optimize the face recognition model;

[0095] S6, adjusting the convolutional neural network parameters according to the gradient algorithm;

[0096] S7, input the training results into the classifier for classification prediction to achieve face recognition.

[0097] In a preferred embodiment of the present invention, in step S2, the method for preprocessing the facial images entering and exiting the smart substation acquired by using the F5G all-optical network includes one or any combination of grayscale, denoising, and normalization to improve image quality and reduce noise interference.

[0098] In a preferred embodiment of the present invention, the method for constructing the preprocessed face image to realize feature extraction in step S3 includes:

[0099] The integer value of the value where the face is not collected is 0, and the value where the face is collected is 1. The integers 0 and 1 are used for encoding. Two integer values ​​are assigned to the dataset to set the mapping of the binary vector length 2, that is, "face not collected" is marked as 0, and "face collected" is marked as 1. If the dataset sequence collects a face, its integer code is 1.

[0100] In a preferred embodiment of the present invention, the facial feature sample method in step S4 is a convolution and pooling method;

[0101] The fully connected layer uses the softmax function to classify the image, the pooling layer uses the RELU function to remove the negative values ​​of the grouped images, and the convolutional layer uses the loss function to regularize the weight of the image.

[0102] In a preferred embodiment of the present invention, step S5 includes:

[0103]

[0104] in, Represents the bias term of the nth feature map in the mth layer;

[0105] Represents the convolution kernel weight between the nth feature map in the mth layer and the jth feature map in the m-1th layer;

[0106] Represents the output of the jth feature map in the m-1th layer;

[0107] represents the output of the nth feature map in the mth layer.

[0108] In a preferred embodiment of the present invention, the loss function is used to learn the optimization model in step S5:

[0109]

[0110] Among them, U represents the loss function;

[0111] G represents the number of samples;

[0112] x i represents the true value;

[0113] Represents the predicted value.

[0114] In a preferred embodiment of the present invention, the method for adjusting the convolutional neural network parameters according to the gradient algorithm in step S6 includes:

[0115] m t =β1m t-1 +(1-β1)g t ,

[0116] Among them, m t Represents the moving average of the gradient of the current iteration;

[0117] β1 represents the first decay rate, which is 0.95 here;

[0118] m t-1 Represents the moving average of the gradient at the t-1 iteration time;

[0119] gt Represents the gradient of the current iteration;

[0120]

[0121] v t Represents the moving average of the square of the gradient of the current iteration;

[0122] β2 represents the second decay rate, which is 0.996 here;

[0123] v t-1 Represents the moving average of the squared gradient at the t-1 iteration;

[0124] g t Represents the gradient of the current iteration;

[0125]

[0126] Represents the modified moving average of the gradient of the current iteration;

[0127] m t Represents the moving average of the gradient of the current iteration;

[0128] Indicates the gradient correction value;

[0129]

[0130] Represents the modified moving average of the squared gradient of the current iteration;

[0131] v t Represents the moving average of the squared gradient at t iterations;

[0132] Represents the squared gradient correction value;

[0133]

[0134] θ t+1 Network parameters indicating the current number of iterations;

[0135] θ t Network parameters representing the last iteration number

[0136] α represents the learning rate;

[0137] Represents the modified moving average of the gradient of the current iteration;

[0138] Represents the modified moving average of the squared gradient of the current iteration;

[0139] ε represents a very small value, which is used to prevent the denominator from being zero. Here, it is 10 -8 .

[0140] In a preferred embodiment of the present invention, step S7 includes:

[0141] The classifier used is the Bayesian classifier;

[0142] Assume that Q is a training set of samples, each sample has its class label, and there are q classes in total, namely D1, D2, D3, ..., D q , each sample is represented by an n-dimensional vector, H = {h1,h2,h3,…,h n}, describing n measured values ​​of n attributes, A1, A2, A3, ..., A n ; Here q is 2;

[0143] Given a sample H, the classifier will predict that sample H belongs to the category with the highest posterior probability, conditional on H, that is, predict that sample H belongs to D i , and only when P(D i |H)>P(D j |H), 1≤j≤q, j≠i;

[0144] Thus we find the maximum P(D i |H) class, P(D i |H) The largest class D i It is called the maximum a posteriori hypothesis, and Bayes' theorem is:

[0145]

[0146] Since P(H) is the same for all classes, only P(H|D i )P(D i ) requires a maximum, if the class The prior probability P(D i ) is unknown, then it is usually assumed that these classes are equally likely, that is, P(D1) = P(D2) = P(D3) = ... = P(D q ), so the maximum is P(H|D i ), otherwise the maximum value is P(H|D i )P(D i ).

[0147] The present invention also discloses a neural network learning and optimization system for an intelligent substation based on an F5G all-optical network, including a face image acquisition module, a preprocessing module, a feature extraction module, a face feature sample acquisition module, a face image training optimization module, a network parameter adjustment module, and a face discrimination module;

[0148] The face image acquisition module is used to acquire the face image data of people entering and leaving the smart substation using the F5G all-optical network;

[0149] The preprocessing module is used to preprocess the facial images entering and leaving the smart substation acquired using the F5G all-optical network;

[0150] The feature extraction module is used to construct the preprocessed face image to realize feature extraction;

[0151] The facial feature sample acquisition module is used to obtain facial feature samples according to the facial feature sample method;

[0152] The face image training optimization module is used to train face images based on a neural network to learn and optimize the face recognition model;

[0153] The network parameter adjustment module is used to adjust the convolutional neural network parameters according to the gradient algorithm;

[0154] The face recognition module is used to input the training results into the classifier for classification prediction to achieve face recognition.

[0155] In a preferred embodiment of the present invention, a method for preprocessing facial images entering and exiting a smart substation acquired using an F5G all-optical network in a preprocessing module includes one or any combination of grayscale, denoising, and normalization to improve image quality and reduce noise interference.

[0156] In a preferred embodiment of the present invention, the face image training optimization module includes:

[0157]

[0158] in, Represents the bias term of the nth feature map in the mth layer;

[0159] Represents the convolution kernel weight between the nth feature map in the mth layer and the jth feature map in the m-1th layer;

[0160] Represents the output of the jth feature map in the m-1th layer;

[0161] represents the output of the nth feature map in the mth layer.

[0162] The present invention also discloses a computer system, comprising:

[0163] processor;

[0164] a memory for storing processor-executable instructions;

[0165] Among them, the processor is configured to implement the neural network learning and optimization method of the smart substation based on the F5G all-optical network when executing the executable instructions.

[0166] The present invention also discloses a computer-readable storage medium, comprising:

[0167] a memory having a computer program stored thereon;

[0168] A processor is used to execute the program in the memory to implement the neural network learning and optimization method of the smart substation based on the F5G all-optical network.

[0169] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A neural network learning and optimization method for smart substation based on F5G all-optical network, characterized in that: The following steps are involved: S1, uses F5G all-optical network to obtain the collected facial image data entering and leaving the smart substation; S2, pre-processing the facial images entering and leaving the smart substation acquired by using the F5G all-optical network; S3, constructing the preprocessed face image to realize feature extraction; S4, obtaining a facial feature sample according to a facial feature sample method; S5, training face images according to the neural network to learn and optimize the face recognition model; S6, adjusting the convolutional neural network parameters according to the gradient algorithm; S7, input the training results into the classifier for classification prediction to achieve face recognition.

2. According to claim 1, the neural network learning and optimization method for smart substation based on F5G all-optical network is characterized in that: In step S2, the method for preprocessing the facial images entering and exiting the smart substation acquired by using the F5G all-optical network includes one or any combination of grayscale, denoising, and normalization to improve image quality and reduce noise interference.

3. According to claim 1, the neural network learning and optimization method for smart substation based on F5G all-optical network is characterized in that: Step S5 includes: in, Represents the bias term of the nth feature map in the mth layer; Represents the convolution kernel weight between the nth feature map in the mth layer and the jth feature map in the m-1th layer; Represents the output of the jth feature map in the m-1th layer; represents the output of the nth feature map in the mth layer.

4. According to claim 1, the neural network learning and optimization method for smart substation based on F5G all-optical network is characterized in that: In step S5, the loss function is used to learn the optimization model: Among them, U represents the loss function; G represents the number of samples; x i represents the true value; Represents the predicted value.

5. According to claim 1, the neural network learning and optimization method for smart substation based on F5G all-optical network is characterized in that: The method for adjusting the convolutional neural network parameters according to the gradient algorithm in step S6 includes: m t =β1m t-1 +(1-β1)g t , Among them, m t Represents the moving average of the gradient of the current iteration; β1 represents the first decay rate, which is 0.95 here; m t-1 Represents the moving average of the gradient at the t-1 iteration time; g t Represents the gradient of the current iteration; v t Represents the moving average of the square of the gradient of the current iteration; β2 represents the second decay rate, which is 0.996 here; v t-1 Represents the moving average of the squared gradient at the t-1 iteration; g t Represents the gradient of the current iteration; Represents the modified moving average of the gradient of the current iteration; m t Represents the moving average of the gradient of the current iteration; Indicates the gradient correction value; Represents the modified moving average of the squared gradient of the current iteration; v t Represents the moving average of the squared gradient at t iterations; Represents the squared gradient correction value; θ t+1 Network parameters indicating the current number of iterations; θ t Network parameters representing the last iteration number α represents the learning rate; Represents the modified moving average of the gradient of the current iteration; Represents the modified moving average of the squared gradient of the current iteration; ε represents a very small value, which is used to prevent the denominator from being zero. Here, it is 10 -8 .

6. A neural network learning and optimization system for smart substations based on F5G all-optical network, characterized in that: It includes a face image acquisition module, a preprocessing module, a feature extraction module, a face feature sample acquisition module, a face image training optimization module, a network parameter adjustment module and a face discrimination module; The face image acquisition module is used to acquire the face image data of people entering and leaving the smart substation using the F5G all-optical network; The preprocessing module is used to preprocess the facial images entering and leaving the smart substation acquired using the F5G all-optical network; The feature extraction module is used to construct the preprocessed face image to realize feature extraction; The facial feature sample acquisition module is used to obtain facial feature samples according to the facial feature sample method; The face image training optimization module is used to train face images based on a neural network to learn and optimize the face recognition model; The network parameter adjustment module is used to adjust the convolutional neural network parameters according to the gradient algorithm; The face recognition module is used to input the training results into the classifier for classification prediction to achieve face recognition.

7. According to claim 6, the neural network learning and optimization system for intelligent substation based on F5G all-optical network is characterized in that: In the preprocessing module, the method for preprocessing the facial images entering and leaving the smart substation acquired by using the F5G all-optical network includes one or any combination of grayscale, denoising, and normalization to improve image quality and reduce noise interference.

8. According to claim 6, the neural network learning and optimization system for intelligent substation based on F5G all-optical network is characterized in that: The face image training optimization module includes: in, Represents the bias term of the nth feature map in the mth layer; Represents the convolution kernel weight between the nth feature map in the mth layer and the jth feature map in the m-1th layer; Represents the output of the jth feature map in the m-1th layer; represents the output of the nth feature map in the mth layer.

9. A computer system, characterized in that: include: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to implement the neural network learning and optimization method of the smart substation based on the F5G all-optical network as described in one of claims 1 to 5 when executing the executable instructions.

10. A computer-readable storage medium, characterized in that: include: a memory having a computer program stored thereon; A processor, used to execute the program in the memory to implement the neural network learning and optimization method of the smart substation based on the F5G all-optical network as described in one of claims 1 to 5.

Images