CN107786958A - A kind of data fusion method based on deep learning model - Google Patents

Abstract

The present invention relates to a data fusion method based on a deep learning model, comprising: firstly training the constructed feature extraction model at the sink node, the network structure contains 3 convolutional layers, 1 pooling layer and 2 Fully connected layer. Complete the training of the model before using the feature extraction model to fuse the node data; each terminal node extracts the original data features through the model; sends the fused data to the Sink node. In the present invention, the constructed feature extraction model is firstly trained at the converging node, then each terminal node extracts the original data features through the model, and finally sends the fused data to the converging node, thereby reducing data transmission volume and prolonging network life. Compared with similar data fusion methods, the present invention can greatly reduce network energy consumption under the same amount of data, and effectively improve the efficiency and accuracy of data fusion.

Description

A Data Fusion Method Based on Deep Learning Model

technical field

The invention relates to the technical field of data fusion, in particular to a data fusion method based on a deep learning model.

Background technique

With the rapid development of the Internet of Things technology, wireless sensor networks (WSNs), as the core component of the Internet of Things perception layer, have been widely used in various environmental monitoring fields. However, in practice, most of the sensor nodes are powered by batteries, which leads to very limited resources in the network. Due to the uneven geographical distribution of a large number of nodes, there is too much redundant information in the data, which increases energy consumption and transmission delay. In addition, due to the widespread interference in the application environment of the Internet of Things, it will directly weaken the data communication transmission capability and reduce the accuracy of data collection, which affects the overall performance of the Internet of Things system.

Contents of the invention

The purpose of the present invention is to provide a data fusion method based on a deep learning model that can eliminate redundancy and reduce data transmission volume, thereby improving network performance, prolonging network life and reducing energy consumption.

In order to achieve the above object, the present invention adopts the following technical solutions: a data fusion method based on a deep learning model, the method includes the steps in the following order:

(1) First, train the constructed feature extraction model at the sink node, the network structure contains 3 convolutional layers, 1 pooling layer and 2 fully connected layers. After using the feature extraction model to perform node data Complete the training of the model before fusion;

(2) Each terminal node extracts the original data features through the model;

(3) Send the fused data to the Sink node.

In the step (1), the loss function of the model training is:

The training objective is given by:

Constantly iteratively update the parameters to minimize the loss function J(θ), where θ is a trainable parameter, including the weight and bias of the convolution kernel, and α is the learning rate.

In order to find the partial derivative For convolutional layers there are:

In the formula, is the sensitivity of the jth feature map of the l-th layer, is the parameter of the jth feature map of the l+1th layer, the Substituting the following formula can get the derivative of the convolution kernel weight ω and bias b;

In the formula, It is the result of the convolution operation between the feature map of the l-1 layer and the convolution kernel of the l layer, combined with Complete the parameter update of a convolution.

In the step (1), for the pooling layer:

In the formula, Represents the j-th feature map of the l-th layer, down means to perform a pooling operation, and substitute the result into Complete a parameter update of the pooling layer.

In the step (1), for the fully connected layer, the backpropagation algorithm is used for training. As a result, the forward propagation process completes the training of the model, and finally obtains the model parameters. The specific steps are as follows:

(5a) The Sink node extracts data containing label information from the corresponding database according to the type of data to be processed;

(5b) Input the training data to the built model, start training, and then the Sink node sends the trained parameters to each terminal node through the cluster head;

(5c) Each terminal node uses the pre-trained model to perform multi-layer convolution feature extraction and pooling on the collected sensor data, and then sends the fused feature data to the corresponding cluster head node, where convolution and pooling The process is the process of data fusion;

(5d) The cluster head node uses the Logistic regression classifier to classify the fusion data generated in step (5c), obtains the classification result, and sends the fusion data to the Sink node;

(5e) The network completes a round of data collection fusion and transmission process, the Sink node re-clusters and selects the cluster head node, and then jumps to step (5c).

It can be known from the above technical solution that the present invention first trains the constructed feature extraction model at the converging node, then each terminal node extracts the original data features through the model, and finally sends the fused data to the converging node, thereby reducing the amount of data transmission and prolonging the Network longevity. Compared with similar data fusion methods, the present invention can greatly reduce network energy consumption under the same amount of data, and effectively improve the efficiency and accuracy of data fusion.

Description of drawings

Fig. 1 is a node routing diagram among the present invention;

Fig. 2 is a flow chart of the method in the present invention.

Detailed ways

As shown in Figures 1 and 2, a data fusion method based on a deep learning model includes the following sequential steps:

(1) First, train the constructed feature extraction model at the sink node, the network structure contains 3 convolutional layers, 1 pooling layer and 2 fully connected layers. After using the feature extraction model to perform node data Complete the training of the model before fusion;

(2) Each terminal node extracts the original data features through the model;

(3) Send the fused data to the Sink node.

In the step (1), the loss function of the model training is:

The training objective is given by:

Constantly iteratively update the parameters to minimize the loss function J(θ), where θ is a trainable parameter, including the weight and bias of the convolution kernel, and α is the learning rate.

In order to find the partial derivative For convolutional layers there are:

In the formula, is the sensitivity of the jth feature map of the l-th layer, is the parameter of the jth feature map of the l+1th layer, the Substituting the following formula can get the derivative of the convolution kernel weight ω and bias b;

In the formula, It is the result of the convolution operation between the feature map of the l-1 layer and the convolution kernel of the l layer, combined with Complete the parameter update of a convolution.

In the step (1), for the pooling layer:

In the formula, Represents the j-th feature map of the l-th layer, down means to perform a pooling operation, and substitute the result into Complete a parameter update of the pooling layer.

As shown in Figure 1, in the step (1), for the fully connected layer, the backpropagation algorithm is used for training, and as a result, the forward propagation process completes the training of the model, and finally obtains the model parameters. The specific steps are as follows:

(5a) The Sink node extracts data containing label information from the corresponding database according to the type of data to be processed;

(5b) Input the training data to the built model, start training, and then the Sink node sends the trained parameters to each terminal node through the cluster head;

(5c) Each terminal node uses the pre-trained model to perform multi-layer convolution feature extraction and pooling on the collected sensor data, and then sends the fused feature data to the corresponding cluster head node, where convolution and pooling The process is the process of data fusion;

(5d) The cluster head node uses the Logistic regression classifier to classify the fusion data generated in step (5c), obtains the classification result, and sends the fusion data to the Sink node;

(5e) The network completes a round of data collection fusion and transmission process, the Sink node re-clusters and selects the cluster head node, and then jumps to step (5c).

In summary, the present invention firstly trains the constructed feature extraction model at the converging node, then each terminal node extracts the original data features through the model, and finally sends the fused data to the converging node, thereby reducing the amount of data transmission and extending the network life. Compared with similar data fusion methods, the present invention can greatly reduce network energy consumption under the same amount of data, and effectively improve the efficiency and accuracy of data fusion.

Claims (5)

1. A kind of 1. data fusion method based on deep learning model, it is characterised in that：This method includes the step of following order：

   (1) it is that Sink node is trained to the Feature Selection Model of structure first in aggregation node, network structure contains 3 altogether Convolutional layer, 1 pond layer and 2 full articulamentum are completed before being merged using Feature Selection Model to node data should The training of model；

   (2) each terminal node passes through the model extraction initial data feature；

   (3) data after fusion are sent to Sink node.
2. 2. the data fusion method according to claim 1 based on deep learning model, it is characterised in that：In the step (1) in, the loss function of the model training is：

   <mrow> <mi>J</mi> <mrow> <mo>(</mo> <mi>&amp;theta;</mi> <mo>)</mo> </mrow> <mo>=</mo> <mo>-</mo> <mfrac> <mn>1</mn> <mi>m</mi> </mfrac> <mo>&amp;lsqb;</mo> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>m</mi> </munderover> <msup> <mi>y</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>)</mo> </mrow> </msup> <msub> <mi>lnh</mi> <mi>&amp;theta;</mi> </msub> <mrow> <mo>(</mo> <msup> <mi>x</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>)</mo> </mrow> </msup> <mo>)</mo> </mrow> <mo>+</mo> <mrow> <mo>(</mo> <mn>1</mn> <mo>-</mo> <msup> <mi>y</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>)</mo> </mrow> </msup> <mo>)</mo> </mrow> <mi>l</mi> <mi>n</mi> <mrow> <mo>(</mo> <mn>1</mn> <mo>-</mo> <msub> <mi>h</mi> <mi>&amp;theta;</mi> </msub> <mo>(</mo> <msup> <mi>x</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>)</mo> </mrow> </msup> <mo>)</mo> <mo>)</mo> </mrow> <mo>&amp;rsqb;</mo> </mrow>

   The target of training is given by the following formula：

   <mrow> <msub> <mi>&amp;theta;</mi> <mi>i</mi> </msub> <mo>=</mo> <msub> <mi>&amp;theta;</mi> <mi>i</mi> </msub> <mo>-</mo> <mi>&amp;alpha;</mi> <mfrac> <mo>&amp;part;</mo> <mrow> <mo>&amp;part;</mo> <mi>&amp;theta;</mi> </mrow> </mfrac> <mi>J</mi> <mrow> <mo>(</mo> <mi>&amp;theta;</mi> <mo>)</mo> </mrow> </mrow>

   Continuous iteration undated parameter to minimize loss function J (θ), wherein, θ is trainable parameter, includes the power of convolution kernel Weight and biasing, α is learning rate.
3. 3. the convolutional neural networks structure according to claim 2 based on deep learning model realizes radio sensing network number According to the method for fusion, it is characterised in that：To obtain partial derivativeHave for convolutional layer：

   <mrow> <msubsup> <mi>&amp;delta;</mi> <mi>j</mi> <mi>l</mi> </msubsup> <mo>=</mo> <msubsup> <mi>&amp;beta;</mi> <mi>j</mi> <mrow> <mi>l</mi> <mo>+</mo> <mn>1</mn> </mrow> </msubsup> <mrow> <mo>(</mo> <msup> <mi>f</mi> <mo>&amp;prime;</mo> </msup> <mo>(</mo> <msubsup> <mi>u</mi> <mi>j</mi> <mi>l</mi> </msubsup> <mo>)</mo> <mo>&amp;CenterDot;</mo> <mi>u</mi> <mi>p</mi> <mo>(</mo> <msubsup> <mi>&amp;delta;</mi> <mi>j</mi> <mrow> <mi>l</mi> <mo>+</mo> <mn>1</mn> </mrow> </msubsup> <mo>)</mo> <mo>)</mo> </mrow> </mrow>

   In formula,For the sensitivity of j-th of characteristic pattern of l layers,, will for the parameter of j-th of characteristic pattern of l+1 layersUnder substitution Formula can obtain convolution kernel weights omega and bias b derivative；

   <mrow> <mfrac> <mrow> <mo>&amp;part;</mo> <mi>J</mi> </mrow> <mrow> <mo>&amp;part;</mo> <msub> <mi>&amp;omega;</mi> <mrow> <mi>i</mi> <mi>j</mi> </mrow> </msub> </mrow> </mfrac> <mo>=</mo> <munder> <mo>&amp;Sigma;</mo> <mrow> <mi>u</mi> <mo>,</mo> <mi>v</mi> </mrow> </munder> <msub> <mrow> <mo>(</mo> <msubsup> <mi>&amp;delta;</mi> <mi>j</mi> <mi>l</mi> </msubsup> <mo>)</mo> </mrow> <mrow> <mi>u</mi> <mi>v</mi> </mrow> </msub> <msub> <mrow> <mo>(</mo> <msubsup> <mi>p</mi> <mi>i</mi> <mrow> <mi>l</mi> <mo>-</mo> <mn>1</mn> </mrow> </msubsup> <mo>)</mo> </mrow> <mrow> <mi>u</mi> <mi>v</mi> </mrow> </msub> </mrow>

   <mrow> <mfrac> <mrow> <mo>&amp;part;</mo> <mi>J</mi> </mrow> <mrow> <mo>&amp;part;</mo> <msub> <mi>b</mi> <mi>j</mi> </msub> </mrow> </mfrac> <mo>=</mo> <munder> <mo>&amp;Sigma;</mo> <mrow> <mi>u</mi> <mo>,</mo> <mi>v</mi> </mrow> </munder> <msub> <mrow> <mo>(</mo> <msubsup> <mi>&amp;delta;</mi> <mi>j</mi> <mi>l</mi> </msubsup> <mo>)</mo> </mrow> <mrow> <mi>u</mi> <mi>v</mi> </mrow> </msub> </mrow>

   In formula,The result of convolution operation is carried out for l-1 layers characteristic pattern and l layers convolution kernel, with reference toComplete a convolution into parameter renewal.
4. 4. the data fusion method according to claim 1 based on deep learning model, it is characterised in that：In the step (1) in, have for pond layer：

   <mrow> <msubsup> <mi>z</mi> <mi>j</mi> <mi>l</mi> </msubsup> <mo>=</mo> <mi>f</mi> <mrow> <mo>(</mo> <msubsup> <mi>&amp;beta;</mi> <mi>j</mi> <mi>l</mi> </msubsup> <mi>d</mi> <mi>o</mi> <mi>w</mi> <mi>n</mi> <mo>(</mo> <mrow> <msubsup> <mi>z</mi> <mi>j</mi> <mrow> <mo>(</mo> <mi>l</mi> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> </msubsup> <mo>+</mo> <msubsup> <mi>b</mi> <mi>j</mi> <mi>l</mi> </msubsup> </mrow> <mo>)</mo> <mo>)</mo> </mrow> </mrow>

   <mrow> <msubsup> <mi>&amp;delta;</mi> <mi>l</mi> <mi>l</mi> </msubsup> <mo>=</mo> <msubsup> <mi>&amp;Sigma;</mi> <mrow> <mi>j</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>M</mi> </msubsup> <msubsup> <mi>&amp;beta;</mi> <mi>l</mi> <mrow> <mo>(</mo> <mi>l</mi> <mo>+</mo> <mn>1</mn> <mo>)</mo> </mrow> </msubsup> <mo>*</mo> <msub> <mi>k</mi> <mrow> <mi>l</mi> <mi>j</mi> </mrow> </msub> </mrow>

   In formula,J-th of characteristic pattern of l layers is represented, down is represented to perform a pondization operation, result is substituted intoComplete the parameter renewal of a pond layer.
5. 5. the data fusion method according to claim 1 based on deep learning model, it is characterised in that：In the step (1) in, for full articulamentum, it is trained using back-propagation algorithm, as a result propagated forward process completes the training of model, most After obtain model parameter, specific step is as follows：

   The data type that (5a) Sink node is handled as needed, the data containing label information are extracted from associated databases；

   (5b) inputs training data to the model of structure, starts to train, and then the parameter trained is passed through cluster by Sink node Hair delivers to each terminal node；

   (5c) each terminal node uses the model of pre-training, and the feature extraction of multilayer convolution and pond are carried out to the sensing data of collection Change, then send the characteristic that fusion obtains to corresponding leader cluster node, wherein, convolution and the process in pond are exactly data The process of fusion；

   (5d) leader cluster node returns grader using Logistic and fused data caused by step (5c) is classified, and is divided Class result, and send fused data to Sink node；

   (5e) network completes a wheel data acquisition fusion and transmitting procedure, and Sink node sub-clustering and chooses leader cluster node again, so After jump to step (5c).