A situation assessment method based on deep learning parameter anchoring

Abstract

A situation assessment method based on deep learning parameter anchoring belongs to the field of aerospace measurement and control technology. The present invention includes the following steps: constructing a convolutional neural network model; the input of the convolutional neural network model is a two-dimensional matrix; making a training data set, and using the data set to train the neural network; deploying the trained convolutional neural network model to the carrier rocket intelligent assisted autonomous situation assessment decision system. When the carrier rocket is flying, the carrier rocket intelligent assisted autonomous situation assessment decision system is sensitive to the parameters collected by sensors of various sections of the carrier rocket in real time, and inputs them into the neural network model, obtains the output of the convolutional neural network model, detects and identifies the abnormal state of the current carrier rocket parameters, and anchors its parameter position. The problem of aerospace flight situation assessment detection is solved, especially the problem of sudden abnormalities of key parameters during the operation of the target system.

Description

Technical Field

[0001] The present invention relates to a situation assessment method based on deep learning parameter anchoring, and belongs to the technical field of aerospace measurement and control. Background Art

[0002] At this stage, the research on space vehicle situation assessment methods is concentrated on low-level methods such as manual interpretation and threshold interpretation based on prior knowledge. There is a lack of artificial intelligence deep mining, modeling and analysis capabilities for big data. The utilization rate of existing historical data of space flight tests is not high, and there is insufficient mining of data correlation and regularity between launches. The large amount of useful information hidden behind the surface data cannot be effectively developed and utilized. Summary of the invention

[0003] The technical problem solved by the present invention is: to overcome the shortcomings of the prior art and provide a situation assessment method based on deep learning parameter anchoring, which is used to solve the problem of aerospace flight situation assessment detection, especially for sudden abnormalities of key parameters during the operation of the target system, through data normalization preprocessing, pre-convolutional network, anchor area network, anchor area pooling and screening, ultimately achieving effective detection, reliable prediction and accurate warning of potential abnormalities of the target system.

[0004] The technical solution of the present invention is: a situation assessment method based on deep learning parameter anchoring, comprising the following steps:

[0005] Constructing a convolutional neural network model; the input of the convolutional neural network model is a two-dimensional matrix;

[0006] Create a training data set and use it to train the neural network;

[0007] The trained convolutional neural network model is deployed to the launch vehicle intelligent assisted autonomous situation assessment and decision-making system. When the launch vehicle is flying, the launch vehicle intelligent assisted autonomous situation assessment and decision-making system collects parameters from sensors of various sections of the launch vehicle in real time and inputs them into the neural network model to obtain the output of the convolutional neural network model, detect and identify the abnormal status of the current launch vehicle parameters, and anchor its parameter position.

[0008] Furthermore, the aerospace status parameter data sequence includes current, voltage, overload, flow, pulse, heat flow, equipment status, external safety, temperature and humidity, pressure, liquid level, and rotation speed that are sensitive to sensors in various sections of the rocket.

[0009] Furthermore, the preprocessing comprises the following steps:

[0010] (1) Generate a two-dimensional matrix: The aerospace state parameter data sequence is converted into two dimensions to generate a data matrix for deep learning and parameter anchoring;

[0011] (2) Label the data matrix;

[0012] (3) The labeled data matrix is ​​read in and randomly divided into a training and validation data set and a test data set in a ratio of 80% and 20%, and the data information of the training data set, the validation data set and the test data set is saved.

[0013] Furthermore, the data matrix is ​​an S*S square matrix; wherein, The operator [·] indicates rounding up; at the same time, the state parameters in the sequence X are filled in the matrix, and the remaining insufficient digits are filled with zeros.

[0014] Furthermore, the labeling of the data matrix includes labeling of data classification labels and labeling of dimensional components of abnormal state parameters as anchoring targets; labeling the classification labels of the generated matrix; anchoring the dimensional components of the abnormal state parameters as targets, and the rest as background.

[0015] Furthermore, the situation assessment deep neural network model includes: a convolutional neural network module, an anchor region generation module, an anchor region pooling module, an anchor regression and classification detection module;

[0016] The convolutional neural network module is used to obtain parameter features in a two-dimensional feature space, including 5 to 13 convolutional layers, 5 to 13 relu layers, and 2 to 8 pooling layers;

[0017] The anchor region generation module is used to realize the generation of candidate anchor regions for parameter features, using the convolution kernel matrix to slide the window and adapt the anchor region;

[0018] The anchor region pooling module is used to fix the number of parameter features, including anchor region mapping, anchor region division, and regional maximum pooling;

[0019] The anchor regression and classification detection module is used to locate the abnormal parameter position and classify the parameter health status. The anchor position offset calculation performs predictive regression calculations on the coordinates of the upper left and lower right corners of the target box, and through positive and negative labeled data learning, it realizes the classification and discrimination of the foreground and background data of abnormal / normal parameters.

[0020] Furthermore, the loss function when training the situation assessment deep neural network model is:

[0021]

[0022] Among them, p i is the probability of anchor prediction as target, p * i is the label function. When the label is negative, it is 1; otherwise, it is 0. iis the predicted bounding box parameter coordinate, t i * is the coordinate vector of the ground truth bounding box corresponding to the positive anchor, L cls (p i ,p i * ) is a binary cross entropy loss, L reg (t i ,t i * ) is the regression loss, N cls is the total number of classification samples, N reg is the total number of regression samples, λ is the proportional weight coefficient, and its value range is 0.5~2.

[0023] A situation assessment system based on deep learning parameter anchoring, comprising:

[0024] The first module is used to construct a convolutional neural network model; the input of the convolutional neural network model is a two-dimensional matrix;

[0025] The second module is used to create a training data set and use the data set to train the neural network;

[0026] The third module is located in the launch vehicle intelligent assisted autonomous situation assessment and decision-making system, which is used to deploy the trained convolutional neural network model. When the launch vehicle is flying, the launch vehicle intelligent assisted autonomous situation assessment and decision-making system collects parameters from sensors of various sections of the launch vehicle in real time and inputs them into the neural network model to obtain the output of the convolutional neural network model, detect and identify the abnormal state of the current launch vehicle parameters, and anchor its parameter position;

[0027] The aerospace state parameter data sequence includes current, voltage, overload, flow, pulse, heat flow, equipment status, external safety, temperature and humidity, pressure, liquid level, and rotation speed that are sensitive to sensors in various sections of the rocket;

[0028] The pre-processing comprises the following steps:

[0029] (1) Generate a two-dimensional matrix: The aerospace state parameter data sequence is converted into two dimensions to generate a data matrix for deep learning and parameter anchoring;

[0030] (2) Label the data matrix;

[0031] (3) Read the labeled data matrix and randomly divide it into a training and validation data set and a test data set in a ratio of 80% and 20%, and save the data information of the training data set, validation data set, and test data set;

[0032] The data matrix is ​​an S*S square matrix; wherein, The operator [·] indicates rounding up; at the same time, the state parameters in the sequence X are filled in the matrix, and the remaining insufficient bits are filled with zeros;

[0033] The labeling of the data matrix includes labeling the data classification labels and labeling the dimension components of the abnormal state parameters as anchor targets; labeling the classification labels of the generated matrix; anchoring the dimension components of the abnormal state parameters as targets, and the rest as background;

[0034] The situation assessment deep neural network model includes: a convolutional neural network module, an anchor region generation module, an anchor region pooling module, an anchor regression and classification detection module;

[0035] The convolutional neural network module is used to obtain parameter features in a two-dimensional feature space, including 5 to 13 convolutional layers, 5 to 13 relu layers, and 2 to 8 pooling layers;

[0036] The anchor region generation module is used to realize the generation of candidate anchor regions for parameter features, using the convolution kernel matrix to slide the window and adapt the anchor region;

[0037] The anchor region pooling module is used to fix the number of parameter features, including anchor region mapping, anchor region division, and regional maximum pooling;

[0038] The anchor regression and classification detection module is used to locate the abnormal parameter position and classify the parameter health status. The anchor position offset calculation performs predictive regression calculation on the coordinates of the upper left corner and lower right corner of the target box, and realizes the classification and discrimination of the foreground and background data of abnormal / normal parameters through positive and negative labeled data learning;

[0039] The loss function when training the situation assessment deep neural network model is

[0040]

[0041] Among them, p i is the probability of anchor prediction as target, p * i is the label function. When the label is negative, it is 1; otherwise, it is 0. i is the predicted bounding box parameter coordinate, t i * is the coordinate vector of the ground truth bounding box corresponding to the positive anchor, L cls (p i ,p i * ) is a binary cross entropy loss, L reg (t i ,ti * ) is the regression loss, N cls is the total number of classification samples, N reg is the total number of regression samples, λ is the proportional weight coefficient, and its value range is 0.5~2.

[0042] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of a situation assessment method based on deep learning parameter anchoring are implemented.

[0043] A situation assessment device based on deep learning parameter anchoring includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the situation assessment method based on deep learning parameter anchoring are implemented.

[0044] The advantages of the present invention compared with the prior art are:

[0045] The present invention proposes a deep learning network model based on parameter anchoring regions. On the basis of the convolutional neural network model, a parameter anchoring region network is added to realize the positioning function of the state abnormal parameter target. Compared with the traditional convolutional neural network detection, the deep learning network constructed by the present invention realizes a network training mode of classification and positioning in parallel, so that the positioning convolutional neural network for generating anchoring windows and the classification convolutional neural network for target detection share the operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is an overall block diagram of the situation assessment method based on deep learning parameter anchoring of the present invention. DETAILED DESCRIPTION

[0047] In order to better understand the above technical scheme, the technical scheme of the present application is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical scheme of the present application, rather than limitations on the technical scheme of the present application. In the absence of conflict, the embodiments of the present application and the technical features in the embodiments can be combined with each other.

[0048] The following is a further detailed description of a situation assessment method based on deep learning parameter anchoring provided by an embodiment of the present application in conjunction with the accompanying drawings of the specification. The specific implementation method may include (such as Figure 1 shown):

[0049] 1) Data normalization preprocessing. Preprocess the existing aerospace state parameter data sequence, including but not limited to: current, voltage, overload, flow, pulse, heat flow, equipment status, external safety, temperature and humidity, pressure, liquid level, speed, etc. The main steps include:

[0050] (1) Generate a two-dimensional matrix: The aerospace state parameter data sequence is converted into two dimensions to generate a data matrix that is convenient for deep learning and parameter anchoring. Suppose the aerospace state parameter sequence is X = {x1, x2, …, x N}, N is the sequence dimension, the generated data matrix A is an S*S square matrix, and the number of rows and columns of the matrix is ​​S

[0051]

[0052] The operator [·] indicates rounding up. At the same time, fill the state parameters in the sequence X in the matrix.

[0053] a 1,1 =x1;a 1,2 =x2; ...; a 1,s =x s ;…;a 2,s =x 2s ;… (2)

[0054] The remaining insufficient digits are filled with zeros.

[0055] (2) Generate matrix data for annotation. This includes annotation of data classification labels and annotation of the dimension components of abnormal state parameters as anchor targets. Annotate the classification labels of the generated matrix; anchor the dimension components of abnormal state parameters as targets, and use the rest as background.

[0056] (3) The labeled generation matrix is ​​read in and randomly divided into a training validation data set and a test data set in a ratio of 80% and 20%, and the data information of the training data set, the validation data set and the test data set is saved.

[0057] 2). Pre-convolutional network modeling

[0058] (1) Convolutional layer

[0059] The convolution layer extracts the features of the data through the convolution kernel. The convolution kernel can also be said to be a matrix. Starting from the upper left corner, the size of the convolution kernel corresponds to the range of the generated matrix, and then multiply and add to get a value. In this order, a convolution operation is performed every other pixel of a step length, and then a value obtained by the convolution operation is obtained. Sliding the convolution kernel window of the entire matrix will form a convolution and matrix. The elements in the matrix are output through a convolution layer, and then the output is processed by the activation function to obtain the data feature map.

[0060] (2) Activation function layer

[0061] When a convolutional neural network is training a network model, an activation function needs to be introduced to increase the nonlinearity of the neural network. The present invention uses the ReLU function as the activation function, also known as a rectifier or corrector.

[0062] The ReLU activation function is expressed as formula (3).

[0063] Relu=max(0,x) (3)

[0064] The ReLU function is a piecewise linear function that can perform this unilateral inhibition operation. The network will not tend to be saturated during back propagation, and there will be no particularly small gradients, which will cause the gradient to disappear. After passing through the ReLU function, all negative numbers are set to 0 and positive values ​​remain unchanged. This will cause the sparsity of the network, reduce the large dependencies between parameters, and alleviate the occurrence of overfitting problems. It is precisely because of the ReLU function that the convergence speed of the model during training can be maintained in a stable state.

[0065] (3) Pooling layer

[0066] The present invention adopts the maximum value pooling method, and selects the largest number in the specified area to represent the entire area. The dimensional output result is the same as the convolution layer, but the calculation method is different. The pooling layer can reduce the size of the feature map, thereby achieving the purpose of reducing the training network parameters. In the feature map, the features extracted from the same area are similar. By using the maximum value pooling method, the feature map size is effectively reduced and the network parameters are reduced.

[0067] 3). Anchor area network modeling

[0068] The present invention introduces an anchored region network to extract the target area of ​​data features, selects a convolutional neural network to extract the detection area frame, and the anchored region network detection network shares the parameters of the network convolution layer. This parameter sharing method directly accelerates the generation speed of the anchored region network.

[0069] The principle of generating target candidate regions by anchor region network. First, the two-dimensional matrix is ​​generated as input, and the data feature map is obtained through the pre-convolutional neural network. The feature map is used as the input of the anchor region network. The corresponding convolution kernel is used in the anchor region network to slide the sample on the feature matrix. Each sliding window predicts k anchor parameters at the same time. The convolution kernel is used to slide the window on the feature matrix to ensure that the entire feature space of the data can be associated. Finally, anchor targets of various scales and aspect ratios on the original image are obtained. After the convolution layer, two sub-fully connected layers are connected for classification and regression respectively. The final anchor region is obtained by classification and regression of the anchor parameters. The classification sublayer is used to determine whether the anchor region belongs to the abnormal state parameter region or the non-abnormal state parameter region. It is a binary classification problem and requires a vector dimension of 2k; the regression sublayer is used to calculate the offset and size scaling of the target candidate box.

[0070] The classification part uses the obtained feature vector to perform calculations and classification through a fully connected layer and two fully connected sub-layers. One sub-layer is connected to a SoftmaxWithloss classifier for category output, and the other sub-layer is connected to a SmoothL1loss classifier to calculate the position offset of the target box and perform target detection by anchor region regression positioning.

[0071] The loss function of the model is the multi-task loss, which is the sum of the prediction classification loss of the target box and the regression loss of the anchor target box.

[0072]

[0073] Among them, pi is the probability that the anchor is predicted as the target;

[0074]

[0075] t i ={t x ,t y ,t w ,t h} is a vector representing the predicted bounding box parameter coordinates;

[0076] t i * is the coordinate vector of the ground truth bounding box corresponding to the positive anchor;

[0077] L cls (p i ,p i * ) is a cross entropy loss for two categories (target & non-target):

[0078] Lcls (p i ,p i * )=-log[p i * p i +(1-p i * )(1-p i )] (6)

[0079] L reg (t i ,t i * ) is the regression loss, using L reg (t i ,t i * )=R(t i -t i * ) is used to calculate, R is the smooth L1 function:

[0080]

[0081] 4) Anchor region pooling

[0082] The function of anchor region pooling is to combine the data feature map obtained by the previous convolution and the target area information obtained by the anchor region network to obtain the coordinates and scales of each target area after generalization. The principle is to map the coordinates of the target candidate area to a feature matrix of a specific size, and divide the target candidate area into n equal parts in the horizontal and vertical directions of the corresponding area on the feature map. In order to obtain an output result of a fixed size, each part after equal division must be subjected to maximum pooling (Max Pooling) to finally achieve a fixed-length output.

[0083] The solution provided in the embodiment of the present application includes the following steps:

[0084] 1. Data normalization preprocessing: Assume that the state parameter data sequence dimension is 796 and the number of samples is 1,000,000, including: current, voltage, overload, flow, pulse, heat flow, equipment status, external safety, temperature and humidity, pressure, liquid level, speed, etc.

[0085] (1) Generate matrix A. Calculate the number of rows and columns S of the matrix, but,

[0086] a 1，1 =x1;a 1，2 =x2; ...; a 1，6 =x6; ...; a 2,6 =x 12; ...a 5,5 =x 29 ; a 5,6 =0;…

[0087] Then, the generator matrix A is expressed as,

[0088]

[0089] Pad the matrix elements that are larger than the dimension of the state parameter data sequence with zeros.

[0090] (2) Data labeling, including labeling of data classification labels and labeling of dimensional component anchor targets of abnormal parameters.

[0091] The data classification is represented as:

[0092] {(A1,c1),(A2,c2),…,(A i ,c j )} (6)

[0093] Where cj represents the classification label.

[0094] Dimensional components of abnormal parameters anchor targets:

[0095]

[0096] For example, in formula (7), x30 is the voltage parameter anomaly, and x31 is the overload parameter anomaly. They are anchored in the matrix and the border is located unknown, so that the parameter anomaly and the classification label are linked in the subsequent learning, and the correlation is learned through the convolutional network, and the mapping relationship is established by iterating the network parameter weights.

[0097] (3) Read the labeled generation matrix and randomly divide it into a training validation data set and a test data set in a ratio of 80% and 20%,

[0098] Training set:

[0099] T={(A1,c1),(A2,c2),…,(A 800000 ,c j )|j=1,2,…,n} (8)

[0100] Test set:

[0101] V={(A1,c1),(A2,c2),…,(A 200000 ,c j )|j=1,2,…,n}

[0102] 2. Pre-convolutional network modeling

[0103] The pre-convolution network mainly uses convolutional neural network to extract data features and obtain the feature map of the matrix. The pre-convolution network is designed for the state parameter data sequence with a dimension of 796 and a sample number of 10,000. In the present invention, the pre-convolution has 5 convolution layers, 5 activation function layers and 2 pooling layers. It mainly includes: convolution layer, pooling layer and activation function layer.

[0104] Input 29*29 data, the number of channels is 256, the convolution kernel size is 3*3, and the step size is 1. When all data points are covered at least once, the output of a convolution layer can be generated, which becomes 29*29*256.

[0105] 3. Anchor Area Network Modeling

[0106] The output of the previous convolutional network is used as the input of the anchor region network. In the anchor region network, a 3*3 convolution kernel is used to slide the sample on the matrix. Each sliding window predicts k (k=9) anchor targets at the same time, and a four-dimensional variable (x, y, w, h) is used to represent an anchor region, which respectively represents the center point coordinates and width and height of the window.

[0107] The specific steps are as follows:

[0108] Step 1 generates the basic anchor area, sets three aspect ratios (1:1, 1:2, 2:1), and sets three scaling scales (2, 4, 8). Therefore, the number of basic anchor areas k = 9;

[0109] Step 2: Based on the basic anchor region, for each element on the feature matrix, 9 (k=9) anchor region bounding boxes of different scales are generated in the corresponding receptive area of ​​the original matrix, with it as the center, generating a total of 29*29*9=7569 anchor regions;

[0110] 4. Anchored Region Pooling

[0111] The anchor region pooling is for 29*29*256 data, and the category is defined as 2*9. Next, the structure of the corresponding training and test layers in the network model needs to be modified. Modify the num_classes and num_output of the input layer and output layer in the train.prototxt and test.prototxt files. The prediction of the target box needs to correspond to the two coordinates of the upper left corner and the lower right corner. The bbox_pred layer corresponds to 4 positions, and a total of 36 coordinate values ​​are predicted and regressed. In order to adapt to the data set, modify the size of Scales in the anchor_target_layer.py and proposal_layer.py under the lib / rpn file to make the anchor target box candidate size match the inspection network.

[0112] Finally, 1*1 convolution is realized, that is, the fully connected layer. The parameter configuration of the fully connected layer is shown in Table 1.

[0113] Table 1 Parameter configuration of the fully connected layer

[0114]

[0115] The fully connected layer integrates the features at different positions and outputs a value, which is equivalent to a classifier. The input is 29*29*256 data, the classification output is 29*29*18, and the positioning output is 29*29*36.

[0116] The dropout rate of the dropout layer is 0.5. The shape of the output data remains unchanged, still 256*1. The parameter configuration of the model training is shown in Table 2.

[0117] Table 2 Parameter configuration of model training

[0118]

[0119] Epoch represents the number of iterations. When a complete data set passes through the neural network once and returns once, this process is called an epoch. The batch_size will determine the number of samples for one training. If Batch_Size is increased appropriately, the accuracy of the gradient descent direction will increase and the amplitude of training shock will decrease. The parameter configuration of the model optimization is shown in Table 3.

[0120] Table 3 Parameter configuration of model optimization

[0121]

[0122] Based on Figure 1 With the same inventive concept, the present invention also provides a situation assessment system based on deep learning parameter anchoring, comprising:

[0123] The first module is used to construct a convolutional neural network model; the input of the convolutional neural network model is a two-dimensional matrix;

[0124] The second module is used to create a training data set and use the data set to train the neural network;

[0125] The third module is located in the launch vehicle intelligent assisted autonomous situation assessment and decision-making system, which is used to deploy the trained convolutional neural network model. When the launch vehicle is flying, the launch vehicle intelligent assisted autonomous situation assessment and decision-making system collects parameters from sensors of various sections of the launch vehicle in real time and inputs them into the neural network model to obtain the output of the convolutional neural network model, detect and identify the abnormal state of the current launch vehicle parameters, and anchor its parameter position;

[0126] The aerospace state parameter data sequence includes current, voltage, overload, flow, pulse, heat flow, equipment status, external safety, temperature and humidity, pressure, liquid level, and rotation speed that are sensitive to sensors in various sections of the rocket;

[0127] The pre-processing comprises the following steps:

[0128] (1) Generate a two-dimensional matrix: The aerospace state parameter data sequence is converted into two dimensions to generate a data matrix for deep learning and parameter anchoring;

[0129] (2) Label the data matrix;

[0130] (3) Read the labeled data matrix and randomly divide it into a training and validation data set and a test data set in a ratio of 80% and 20%, and save the data information of the training data set, validation data set, and test data set;

[0131] The data matrix is ​​an S*S square matrix; wherein, The operator [·] indicates rounding up; at the same time, the state parameters in the sequence X are filled in the matrix, and the remaining insufficient bits are filled with zeros;

[0132] The labeling of the data matrix includes labeling the data classification labels and labeling the dimension components of the abnormal state parameters as anchor targets; labeling the classification labels of the generated matrix; anchoring the dimension components of the abnormal state parameters as targets, and the rest as background;

[0133] The situation assessment deep neural network model includes: a convolutional neural network module, an anchor region generation module, an anchor region pooling module, an anchor regression and classification detection module;

[0134] The convolutional neural network module is used to obtain parameter features in a two-dimensional feature space, including 5 to 13 convolutional layers, 5 to 13 relu layers, and 2 to 8 pooling layers;

[0135] The anchor region generation module is used to realize the generation of candidate anchor regions for parameter features, using the convolution kernel matrix to slide the window and adapt the anchor region;

[0136] The anchor region pooling module is used to fix the number of parameter features, including anchor region mapping, anchor region division, and regional maximum pooling;

[0137] The anchor regression and classification detection module is used to locate the abnormal parameter position and classify the parameter health status. The anchor position offset calculation performs predictive regression calculation on the coordinates of the upper left corner and lower right corner of the target box, and realizes the classification and discrimination of the foreground and background data of abnormal / normal parameters through positive and negative labeled data learning;

[0138] The loss function when training the situation assessment deep neural network model is

[0139]

[0140] Among them, p i is the probability of anchor prediction as target, p * i is the label function. When the label is negative, it is 1; otherwise, it is 0. i is the predicted bounding box parameter coordinate, t i * is the coordinate vector of the ground truth bounding box corresponding to the positive anchor, L cls (p i ,p i * ) is a binary cross entropy loss, L reg (t i ,t i * ) is the regression loss, N cls is the total number of classification samples, N reg is the total number of regression samples, λ is the proportional weight coefficient, and its value range is 0.5~2.

[0141] The present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and when the computer instructions are executed on a computer, the computer executes Figure 1 The method described.

[0142] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) that contain computer-usable program code.

[0143] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1A device that provides the functions specified in a block or multiple blocks.

[0144] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0145] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0146] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

[0147] The contents not described in detail in the specification of the present invention belong to the common knowledge of those skilled in the art.

Claims

1. A situation assessment method based on deep learning parameter anchoring, characterized in that: The steps include: Constructing a situation assessment deep neural network model; the input of the situation assessment deep neural network model is a two-dimensional matrix; Create a training dataset and use it to train a situation assessment deep neural network model; The trained situation assessment deep neural network model is deployed to the launch vehicle intelligent assisted autonomous situation assessment decision system. When the launch vehicle is flying, the launch vehicle intelligent assisted autonomous situation assessment decision system collects parameters from sensors of various sections of the launch vehicle in real time and inputs them into the situation assessment deep neural network model to obtain the output of the situation assessment deep neural network model, detect and identify the abnormal state of the current launch vehicle parameters, and anchor its parameter position; The situation assessment deep neural network model includes: a convolutional neural network module, an anchor region generation module, an anchor region pooling module, an anchor regression and classification detection module; The convolutional neural network module is used to obtain parameter features in a two-dimensional feature space, including 5 to 13 convolutional layers, 5 to 13 relu layers, and 2 to 8 pooling layers; The anchor region generation module is used to realize the generation of candidate anchor regions for parameter features, using the convolution kernel matrix to slide the window and adapt the anchor region; The anchor region pooling module is used to fix the number of parameter features, including anchor region mapping, anchor region division, and regional maximum pooling; The anchor regression and classification detection module is used to locate the abnormal parameter position and classify the parameter health status. The anchor position offset calculation performs predictive regression calculations on the coordinates of the upper left and lower right corners of the target box, and through positive and negative labeled data learning, it realizes the classification and discrimination of the foreground and background data of abnormal / normal parameters.

2. A situation assessment method based on deep learning parameter anchoring according to claim 1, characterized in that: The construction of the situation assessment deep neural network model includes preprocessing the existing aerospace state parameter data sequence; the aerospace state parameter data sequence includes the current, voltage, overload, flow, pulse, heat flow, equipment status, external safety, temperature and humidity, pressure, liquid level, and rotation speed that are sensitive to sensors in various sections of the rocket.

3. A situation assessment method based on deep learning parameter anchoring according to claim 2, characterized in that: The pre-processing comprises the following steps: (1) Generate a two-dimensional matrix: The aerospace state parameter data sequence is converted into two dimensions to generate a data matrix for deep learning and parameter anchoring; (2) Label the data matrix; (3) The labeled data matrix is ​​read in and randomly divided into a training and validation data set and a test data set in a ratio of 80% and 20%, and the data information of the training data set, the validation data set and the test data set is saved.

4. A situation assessment method based on deep learning parameter anchoring according to claim 3, characterized in that: The data matrix is ​​an S*S square matrix; wherein, The operator [·] indicates rounding up; at the same time, the state parameters in the sequence X are filled in the matrix, and the remaining insufficient digits are filled with zeros.

5. The situation assessment method based on deep learning parameter anchoring according to claim 3 is characterized in that: The labeling of the data matrix includes labeling the data classification labels and labeling the dimensional components of the abnormal state parameters as anchor targets; labeling the classification labels of the generated matrix; anchoring the dimensional components of the abnormal state parameters as targets, and the rest as background.

6. A situation assessment method based on deep learning parameter anchoring according to claim 1, characterized in that: The loss function when training the situation assessment deep neural network model is Among them, p i is the probability of anchor prediction as target, p * i is the label function. When the label is negative, it is 1; otherwise, it is 0. i is the predicted bounding box parameter coordinate, t i * is the coordinate vector of the ground truth bounding box corresponding to the positive anchor, L cls (p i ,p i * ) is a binary cross entropy loss, L reg (t i ,t i * ) is the regression loss, N cls is the total number of classification samples, N reg is the total number of regression samples, λ is the proportional weight coefficient, and its value range is 0.5~2.

7. A situation assessment system based on deep learning parameter anchoring, characterized in that: include: The first module builds a deep neural network model for situation assessment; The input of the situation assessment deep neural network model is a two-dimensional matrix; The second module is to create a training data set and use the data set to train the situation assessment deep neural network model; The third module deploys the trained situation assessment deep neural network model to the launch vehicle intelligent assisted autonomous situation assessment decision system. When the launch vehicle is flying, the launch vehicle intelligent assisted autonomous situation assessment decision system collects parameters from sensors of various sections of the launch vehicle in real time and inputs them into the situation assessment deep neural network model to obtain the output of the situation assessment deep neural network model, detect and identify the abnormal state of the current launch vehicle parameters, and anchor its parameter position. The situation assessment deep neural network model includes: a convolutional neural network module, an anchor region generation module, an anchor region pooling module, an anchor regression and classification detection module; The convolutional neural network module is used to obtain parameter features in a two-dimensional feature space, including 5 to 13 convolutional layers, 5 to 13 relu layers, and 2 to 8 pooling layers; The anchor region generation module is used to realize the generation of candidate anchor regions for parameter features, using the convolution kernel matrix to slide the window and adapt the anchor region; The anchor region pooling module is used to fix the number of parameter features, including anchor region mapping, anchor region division, and regional maximum pooling; The anchor regression and classification detection module is used to locate the abnormal parameter position and classify the parameter health status. The anchor position offset calculation performs predictive regression calculation on the coordinates of the upper left corner and lower right corner of the target box, and realizes the classification and discrimination of the foreground and background data of abnormal / normal parameters through positive and negative labeled data learning; The construction of the situation assessment deep neural network model includes preprocessing the existing aerospace state parameter data sequence; the aerospace state parameter data sequence includes the current, voltage, overload, flow, pulse, heat flow, equipment status, external safety, temperature and humidity, pressure, liquid level, and rotation speed sensitive to sensors of various sections of the rocket; The pre-processing comprises the following steps: (1) Generate a two-dimensional matrix: The aerospace state parameter data sequence is converted into two dimensions to generate a data matrix for deep learning and parameter anchoring; (2) Label the data matrix; (3) Read the labeled data matrix and randomly divide it into a training and validation data set and a test data set in a ratio of 80% and 20%, and save the data information of the training data set, validation data set, and test data set; The data matrix is ​​an S*S square matrix; wherein, The operator [·] indicates rounding up; at the same time, the state parameters in the sequence X are filled in the matrix, and the remaining insufficient bits are filled with zeros; The labeling of the data matrix includes labeling the data classification labels and labeling the dimension components of the abnormal state parameters as anchor targets; labeling the classification labels of the generated matrix; anchoring the dimension components of the abnormal state parameters as targets, and the rest as background; The loss function when training the situation assessment deep neural network model is Among them, p i is the probability of anchor prediction as target, p * i is the label function. When the label is negative, it is 1; otherwise, it is 0. i is the predicted bounding box parameter coordinate, t i * is the coordinate vector of the ground truth bounding box corresponding to the positive anchor, L cls (p i ,p i * ) is a binary cross entropy loss, L reg (t i ,t i * ) is the regression loss, N cls is the total number of classification samples, N reg is the total number of regression samples, λ is the proportional weight coefficient, and its value range is 0.5~2.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A situation assessment device based on deep learning parameter anchoring, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

Images