CN113453153B - Wireless positioning fusion method and system based on deep learning - Google Patents
Wireless positioning fusion method and system based on deep learning Download PDFInfo
- Publication number
- CN113453153B CN113453153B CN202010229645.8A CN202010229645A CN113453153B CN 113453153 B CN113453153 B CN 113453153B CN 202010229645 A CN202010229645 A CN 202010229645A CN 113453153 B CN113453153 B CN 113453153B
- Authority
- CN
- China
- Prior art keywords
- model
- data
- value
- neural network
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
- H04W4/02—Services making use of location information
- H04W4/021—Services related to particular areas, e.g. point of interest [POI] services, venue services or geofences
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S11/00—Systems for determining distance or velocity not using reflection or reradiation
- G01S11/02—Systems for determining distance or velocity not using reflection or reradiation using radio waves
- G01S11/06—Systems for determining distance or velocity not using reflection or reradiation using radio waves using intensity measurements
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
- H04W4/30—Services specially adapted for particular environments, situations or purposes
- H04W4/33—Services specially adapted for particular environments, situations or purposes for indoor environments, e.g. buildings
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Position Fixing By Use Of Radio Waves (AREA)
Abstract
The invention relates to a wireless positioning method, in particular to a wireless positioning fusion method and a wireless positioning fusion system based on deep learning, which can effectively improve the accuracy of wireless positioning, and comprises a model building and training stage and a model application stage, wherein the model building and training stage executes the following steps: acquiring original data, determining a preprocessing mode according to the data, and training and generating a convolutional neural network model and a K neighbor model by respectively utilizing the data processed by the preprocessing mode A and the preprocessing mode B and the corresponding position data thereof; determining an optimal threshold value through a data experiment; compared with the existing algorithm, the method can improve the accuracy of wireless positioning and perform more accurate identification; the device does not need to be additionally added, and has higher practical value.
Description
The technical field is as follows:
the invention relates to a wireless positioning method, in particular to a wireless positioning fusion method and system based on deep learning, which can effectively improve the wireless positioning accuracy.
Background art:
in recent years, intelligent mobile terminals and wireless sensors are increasingly popularized and developed, wireless technology plays an important role in the fields of medical treatment, security protection and the like, and accurate positioning of indoor personnel becomes a key point of research due to promotion of huge practical value of the wireless technology.
Wireless positioning can be classified into a geometric method and a position fingerprint method according to a positioning method, and classified into positioning using received signal strength and positioning using time (or angle) according to a data type. The geometric method can also adopt the received signal strength to position, and a signal attenuation formula is used for estimating the distance between the target and the signal transmitter so as to determine the position of the target; however, in practical application, the error is large by using a geometric method, mainly because the electromagnetic wave propagation speed is too high, and the received signal strength is influenced by complex environments such as walls, floors and the like; therefore, the distance estimated by using the time or the received signal strength is often not accurate enough, and the positioning error is large. The position fingerprint method mostly uses the received signal strength and the corresponding position information to train a machine learning or deep learning model, establishes a position fingerprint library, compares the strength signal received in real time with the position fingerprint library in actual application, and calculates the position corresponding to the signal strength. With the development of deep learning in recent years, intensity signals are increasingly classified by using a deep learning model including a convolutional neural network. At present, a position fingerprint method for positioning by utilizing a received intensity signal has K neighbor (machine learning) and a convolutional neural network (deep learning), but the method has the problems of low identification accuracy and inaccurate positioning.
The invention content is as follows:
the invention provides a wireless positioning fusion method and system based on deep learning, which can effectively improve the wireless positioning accuracy rate under the condition of not changing the original wireless positioning equipment and data acquisition mode, aiming at the problems of low identification accuracy rate and poor positioning effect when wireless positioning is carried out by utilizing Received Signal Strength (RSSI) at present.
The invention is achieved by the following measures:
a wireless positioning fusion method based on deep learning comprises a model building and training stage and a model application stage, and is characterized in that the model building and training stage executes the following steps:
step 1: acquiring original data, and determining a preprocessing mode according to the data, wherein the original data is received signal intensity data of a sensor, the preprocessing mode comprises a preprocessing mode A and a preprocessing mode B, the data processed by the preprocessing mode A is used for training and using a convolutional neural network model, and the data processed by the preprocessing mode B is used for training and using a K neighbor model;
step 2: training and generating a convolutional neural network model and a K neighbor model by respectively utilizing the data processed by the preprocessing mode A and the preprocessing mode B and the corresponding position data;
and step 3: determining an optimal threshold value through a data experiment;
the model application phase performs the following steps:
and 4, step 4: after the intensity data of the signals received by the sensor are processed according to the preprocessing mode A determined in the step 1, classifying the intensity data by using the convolutional neural network model obtained in the step 2 to obtain a position judgment result of the convolutional neural network model and a probability value of the result;
and 5: comparing the probability value with the optimal threshold value obtained in the step 3, and if the probability value is not lower than the optimal threshold value, receiving a position judgment result of the convolutional neural network model; and if the probability value is lower than the optimal threshold value, processing the intensity data of the signals received by the sensor according to the preprocessing mode B determined in the step 1, classifying the data by using the K neighbor model obtained in the step 2, and receiving the position judgment result of the K neighbor model.
The convolutional neural network model generated in the step 2 of the invention is a one-dimensional convolutional neural network model, and comprises a convolutional layer, a pooling layer, a Flatten layer, a merging layer, a full-link layer and a Dropout layer for avoiding over-fitting, wherein the Dropout layer is positioned behind the convolutional layer or between the full-link layers; a multilayer parallel network structure is adopted in the convolutional neural network model, data enter two parallel pooling layers after passing through a convolutional layer, and the two pooling layers are an average pooling layer and a maximum pooling layer respectively.
In step 1 of the invention, aiming at a K neighbor model, the step of determining a preprocessing mode B comprises the following steps:
step 1-1: randomly dividing all original data into a training set and a verification set;
step 1-2: filling the data missing values of the training set and the verification set with different b values each time, wherein the b value is smaller than the minimum value a of the original data;
step 1-3: training the training set data filled with the missing values in the step 1-2 to generate a K neighbor model, and testing the model accuracy by using the verification set filled with the missing values in the step 1-2 to obtain the model accuracy when each value b is filled with the missing values;
step 1-4: repeating the steps 1-1 to 1-3 for a plurality of times, wherein the number of times is not less than 10 times to ensure the effect, and obtaining a plurality of groups of accuracy rates under the condition that each b value is filled, and if the b values are repeated for 10 times, each b value corresponds to 10 accuracy rates;
step 1-5: averaging a plurality of accuracy rates under each b value, and selecting the b value corresponding to the highest accuracy rate average value as the filling of the missing value;
the preprocessing mode B can be determined through the steps 1-1 to 1-5, and the content of the preprocessing mode B is as follows: and filling the missing value with the b value corresponding to the highest accuracy mean value determined in the step 1-5. The model trained in the above steps 1-3 is only used for accuracy verification and determining the preprocessing mode B, and the trained model is not used in the actual application stage.
In step 1 of the invention, the determination of the preprocessing mode A aiming at the convolutional neural network model comprises the following steps:
step 1-1': judging whether the original intensity value is less than or equal to 0, if the minimum intensity value a in the original intensity value is less than or equal to 0, adding (-a +1) to all original data for translation, enabling the translated data minimum intensity value a' to be >0, and then randomly dividing the translated data into a training set and a verification set; if the original intensity data are positive values, the data are directly randomly divided into a training set and a verification set;
step 1-2': filling missing values with different c values, wherein c is less than or equal to 0;
step 1-3': generating a convolutional neural network model by using the training set data filled with the missing values in the step 1-2 ', and testing the model accuracy by using the verification set filled with the missing values in the step 1-2' to obtain the model accuracy when each c value is filled with the missing values;
step 1-4': repeating the steps 1-1 'to 1-3' for a plurality of times, wherein the number of times is not less than 10 times to ensure the effect, and obtaining a plurality of groups of accuracy rates under the condition that each c value is filled, and if the steps are repeated for 10 times, each c value corresponds to 10 accuracy rates;
step 1-5': averaging a plurality of accuracy rates under each c value, and selecting the c value corresponding to the highest accuracy rate average value as the filling of the missing value;
determining a pretreatment mode A through steps 1-1 'to 1-5', wherein the pretreatment mode A comprises the following contents: if the minimum intensity a of the original intensity data is less than or equal to 0, adding (-a +1) to all the original data for translation, and filling the missing value with a value c corresponding to the highest accuracy mean value determined in the step 1-5'; and if the original data are positive values, filling the missing values by using the c values corresponding to the highest accuracy mean values determined in the steps 1-5'.
The model trained in the step 1-3' is only used for accuracy verification and determining the preprocessing mode A, and the trained model is not used in the actual application stage.
The method for determining the optimal threshold value through a data experiment in the step 3 comprises the following steps:
step 3-1: randomly dividing all data into a training set and a verification set;
step 3-2: after being processed by the preprocessing mode A, the training set data is used for training a convolutional neural network model, the architecture and parameters of the convolutional neural network model are consistent with those of the convolutional neural network model in the step 2, after being processed by the preprocessing mode B, the training set data is used for training a K neighbor model, and the parameters of the K neighbor model are consistent with those of the K neighbor model in the step 2; step 3-3: sequentially trying different threshold values, calculating the accuracy of the K neighbor model and the accuracy of the fusion model under different threshold values by using a verification set, and solving the difference value of the accuracy of the fusion model and the accuracy of the single K neighbor model, namely the accuracy improvement value of the fusion model compared with the single K neighbor model;
wherein the fusion model performs the following: the verification set data is preprocessed in a mode A, a convolutional neural network model is used for judging, and if the probability value output by the convolutional neural network model is not lower than the attempted threshold value, the judgment result of the convolutional neural network model is received; if the output probability value is lower than the attempted threshold value, the verification set data is judged by using a K neighbor model in a preprocessing mode B, and the judgment result of the K neighbor model is accepted;
step 3-4: and (3) repeating the steps 3-1 to 3-3, namely randomly dividing all the data again, repeating the steps for training and testing to obtain multiple groups of accuracy rate improvement values of the fusion model under each attempted threshold value compared with the single K neighbor model. If repeating for 10 times, 10 fusion models are applied to the accuracy rate improvement values of the K neighbor model for each attempted threshold value, the average value of the accuracy rate improvement values corresponding to each threshold value is calculated, and the threshold value with the maximum average value of the improvement values is the optimal threshold value;
the K neighbor model, the convolutional neural network model and the fusion model trained in the steps 3-1 to 3-3 are only used for accuracy verification and determining an optimal threshold, and the trained model cannot be used in an actual application stage.
The invention also provides a system using the wireless positioning fusion method based on deep learning, which is characterized by comprising a data preprocessing determination unit, a classification model building and training unit, an optimal threshold determination unit and a result classification unit,
the data preprocessing determining unit is provided with a preprocessing mode A determining module and a preprocessing mode B determining module, the preprocessing mode A is obtained through the data preprocessing mode A determining module, and original data are used for training and using a convolutional neural network model after being processed by the preprocessing mode A; obtaining a preprocessing mode B through a data preprocessing mode B determining module, and using the original data processed by the preprocessing mode B for training and using a K neighbor model;
two models are arranged in the classification model building and training unit, and comprise a convolution neural network model generated by training the original data together with the corresponding position after being processed by a preprocessing mode A and a K neighbor model generated by training the original data together with the corresponding position after being processed by a preprocessing mode B;
the result classification unit processes the original data according to a pre-processing mode A determined in the data pre-processing determination unit, then uses a classification model to build and train a convolutional neural network model determined in the training unit for classification, the convolutional neural network model outputs a position judgment result and a probability, compares the probability value with an optimal threshold determined in the optimal threshold determination unit, if the probability is greater than or equal to the optimal threshold, the result classification unit outputs the position judgment result of the convolutional neural network as a final judgment result, if the probability value is less than the optimal threshold, the result classification unit processes the original data according to a pre-processing mode B determined in the data pre-processing determination unit, and then uses the classification model to build and train a K neighbor model determined in the training unit to obtain the final judgment result.
Compared with the existing algorithm, the invention has the following advantages: the method can improve the accuracy of wireless positioning and carry out more accurate identification; for the convolutional neural network algorithm, the method improves the structure and the preprocessing mode of the convolutional neural network, and for the K neighbor algorithm, the method improves the preprocessing mode; meanwhile, by introducing a probability judgment mechanism, the two algorithms are fused, and a more suitable single algorithm can be selected according to the situation by a newly generated fusion algorithm; in addition, the algorithm of the invention does not need to be additionally added on equipment, only changes the algorithm, can avoid the cost in the aspects of capital and time caused by equipment purchase and erection, and has higher practical value.
Description of the drawings:
FIG. 1 is a flow chart of the model building and training phase of the present invention.
FIG. 2 is a flow chart of the model application phase of the present invention.
Fig. 3 is a flowchart of the determination of the pretreatment mode in the present invention (wherein 3.1 is a determination flow of the pretreatment mode a, and 3.2 is a determination flow of the pretreatment mode B).
Fig. 4 is a block diagram of a convolutional neural network of the present invention.
Fig. 5 is a flow chart of optimal threshold determination in the present invention.
FIG. 6.1 is a block diagram of the system of the present invention, and FIG. 6.2 is a flow chart of the execution of the result classification unit of the present invention.
The specific implementation mode is as follows:
the invention is further described with reference to the following figures and specific embodiments.
The invention provides a wireless positioning fusion method based on deep learning, which comprises a model building and training stage and a model application stage.
The model building and training stage comprises the following steps:
step 1: acquiring original data, and determining a preprocessing mode according to the data, wherein the original data is received signal intensity data of a sensor, the preprocessing mode comprises a preprocessing mode A and a preprocessing mode B, the data processed by the preprocessing mode A is used for training and using a convolutional neural network model, and the data processed by the preprocessing mode B is used for training and using a K neighbor model;
step 2: training and generating a convolutional neural network model and a K neighbor model by respectively utilizing the data processed by the preprocessing mode A and the preprocessing mode B and the corresponding position data;
and step 3: determining an optimal threshold value through a data experiment;
the model application phase performs the following steps:
and 4, step 4: after the intensity data of the signals received by the sensor are processed according to the preprocessing mode A determined in the step 1, classifying the intensity data by using the convolutional neural network model obtained in the step 2 to obtain a position judgment result of the convolutional neural network model and a probability value of the result;
and 5: comparing the probability value with the optimal threshold value obtained in the step 3, and if the probability value is not lower than the optimal threshold value, receiving a position judgment result of the convolutional neural network model; and if the probability value is lower than the optimal threshold value, processing the intensity data of the signals received by the sensor according to the preprocessing mode B determined in the step 1, classifying the data by using the K neighbor model obtained in the step 2, and receiving the position judgment result of the K neighbor model.
In the step 1 of the invention, the data preprocessing is to fill (and transform) the missing value of the original data; this is because the data is received by the sensors, and there are some positions where all the signals of the transmitters cannot be received, and if there are N transmitters, the received signal strength of a position should have N values, but actually the position cannot receive the signals of some transmitters, so that there will be data missing. At this time, the minimum signal strength value which can be received is defined as a, a value slightly smaller than a needs to be selected to fill all missing values caused by signals which cannot be received, and the size of the value selected to fill the missing values affects the effects of the K neighbor model and the convolutional neural network model.
As shown in fig. 3.2, for the K-nearest neighbor algorithm, the invention randomly divides all original data into a training set and a verification set, fills missing values in the training set and the verification set with different values (b values) each time, generates a K-nearest neighbor model from the training set filled with each b value, tests the accuracy by using the verification set, obtains the corresponding accuracy when each b value is filled, and randomly divides all data again after recording. Repeating the steps to obtain the average value of the accuracy when each b value fills the missing value under multiple experiments, wherein the b value corresponding to the highest average value of the accuracy is used as the filling value of the missing value; once the B value corresponding to the highest accuracy mean value is selected, a preprocessing mode B is determined, and a method for filling a missing value by using the B value corresponding to the highest accuracy mean value is a data preprocessing method aiming at a K neighbor model and a preprocessing mode B;
as shown in fig. 3.1, for the convolutional neural network algorithm, all original data are randomly divided into a training set and a verification set, if the minimum value a of the intensity of the original signal is less than or equal to 0, all original signals are translated by adding a numerical value (-a +1) to the original signals, all the translated signals are positive numbers, and if the minimum value a is greater than 0, translation is not needed; then, similar to the method for determining the preprocessing mode B, different values (c values) are adopted to fill missing values in a training set and a verification set each time, the filled training set is used for training a convolutional neural network model, the verification set is used for testing the accuracy, the accuracy under the filling condition of different c values is compared, in order to enable the result to be more stable, the training set and the verification set are repeatedly divided, the previous steps are repeatedly carried out for multiple times, the average value of the accuracy under the filling condition of multiple different c values is obtained, and the c value corresponding to the highest average value of the accuracy is the value used for filling the missing value in actual use; determining a preprocessing mode A along with a c value corresponding to the highest accuracy mean value, wherein if the minimum value a of an original signal is less than or equal to 0, adding (-a +1) to all data, and filling a missing value by using the c value corresponding to the highest accuracy mean value; and if the minimum value a of the original signal is greater than 0, filling the missing value by directly using the c value corresponding to the highest accuracy mean value.
All models trained in fig. 3.1 and fig. 3.2 are only for determining the preprocessing mode a and the preprocessing mode B, and the trained models are not used in the application phase, and the model used in the application phase is formed by fusing the K nearest neighbor trained by using all data (not a divided training set) in step 2 and the convolutional neural network model.
The convolutional neural network model generated in the step 2 of the invention is a one-dimensional convolutional neural network model, and comprises a convolutional layer, a pooling layer, a Flatten layer, a merging layer, a full-link layer and a Dropout layer for avoiding over-fitting, wherein the Dropout layer is positioned behind the convolutional layer or between the full-link layers; a multi-layer parallel network structure is adopted in the convolutional neural network model, data enter two parallel pooling layers after passing through a convolutional layer, and the two pooling layers are an average pooling layer and a maximum pooling layer respectively.
As shown in fig. 5, in step 3 of the present invention, all data are randomly divided into a training set and a verification set, the training set is used to train a convolutional neural network model and a K-nearest neighbor model (the architecture and parameters of the trained convolutional neural network model and the K-nearest neighbor model should be consistent with those in step 2), and the verification set is used to obtain the optimal threshold of the fusion model; the fusion model is constructed by a convolution neural network model and a K neighbor model together: the fusion model is that the data to be tested is firstly preprocessed in a mode A and judged by using the convolutional neural network model, if the probability value output by the convolutional neural network model is not lower than the tried threshold value, the judgment result of the convolutional neural network model is received, and if the output probability value is lower than the tried threshold value, the data to be tested is judged by using the preprocessing mode B and the K neighbor model, and the judgment result of the K neighbor model is received.
The specific steps for determining the optimal threshold in step 3 of the invention are as follows:
after data are randomly divided, a convolution neural network model trained by a training set and a K neighbor model are utilized to form a fusion model, different thresholds are tried in sequence, a threshold value under each try is recorded, and the accuracy rate of the fusion model is compared with the accuracy rate of a single K neighbor model when a verification set is used for testing; then repeating the steps, randomly dividing all data again, training and testing again, obtaining the average value of the fusion model accuracy rate comparison single K neighbor model accuracy rate promotion values under each threshold value of multiple experiments, and selecting the threshold value with the maximum promotion value average value as the optimal threshold value when the fusion model is actually used; note that this step is only to determine the optimal threshold, and none of the trained models is used in the application phase.
As shown in figure 6, the invention also provides a system using the wireless positioning fusion method based on deep learning, which is characterized by being provided with a data preprocessing determination unit, a classification model building and training unit, an optimal threshold determination unit and a result classification unit,
the data preprocessing determining unit is provided with a preprocessing mode A determining module and a preprocessing mode B determining module, the preprocessing mode A is obtained through the data preprocessing mode A determining module, and original data are used for training and using a convolutional neural network model after being processed by the preprocessing mode A; obtaining a preprocessing mode B through a data preprocessing mode B determining module, and using the original data processed by the preprocessing mode B for training and using a K neighbor model;
two models are arranged in the classification model building and training unit, and comprise a convolution neural network model generated by training the original data together with the corresponding position after being processed by a preprocessing mode A and a K neighbor model generated by training the original data together with the corresponding position after being processed by a preprocessing mode B;
the result classification unit processes the original data according to a pre-processing mode A determined in the data pre-processing determination unit, then uses a classification model to build and train a convolutional neural network model determined in the training unit for classification, the convolutional neural network model outputs a position judgment result and a probability, compares the probability value with an optimal threshold determined in the optimal threshold determination unit, if the probability is greater than or equal to the optimal threshold, the result classification unit outputs the position judgment result of the convolutional neural network as a final judgment result, if the probability value is less than the optimal threshold, the result classification unit processes the original data according to a pre-processing mode B determined in the data pre-processing determination unit, and then uses the classification model to build and train a K neighbor model determined in the training unit to obtain the final judgment result.
Example 1:
a wireless positioning fusion method based on deep learning comprises the following steps:
a model training stage:
s1: collecting wireless intensity signals and position information data of each transmitter received at a sampling point;
s2: determining a preprocessing mode and preprocessing original data, wherein the preprocessing mode to be determined comprises a preprocessing mode A and a preprocessing mode B, the data processed by the preprocessing mode A is used for training and using a convolutional neural network model, and the data processed by the preprocessing mode B is used for training and using a K neighbor model;
s3: processing the received signal intensity data by using the preprocessing mode A determined in S2, and training the processed data and corresponding position information to generate a one-dimensional convolutional neural network model;
s4: processing the received signal strength data by using the preprocessing method B determined in S2, and training the processed data together with corresponding position information to generate a K neighbor model;
s5: determining an optimal threshold under the condition that a K neighbor model, a convolutional neural network model architecture and parameters are not changed;
and (3) a practical application stage:
s6: processing the acquired data by using the preprocessing mode A determined in S2, inputting the processed data into the convolutional neural network model obtained in S3, and obtaining a judgment result of the convolutional neural network model and a probability value of the result;
s7: comparing the probability value with the optimal threshold value determined in the S5, and if the probability value is greater than or equal to the optimal threshold value, taking the judgment result of the convolutional neural network model in the S6 as a final positioning result; and if the probability value is smaller than the optimal threshold value, processing the acquired data by using the preprocessing mode B determined in the S2, inputting the processed data into the K neighbor model obtained in the S4, obtaining a judgment result of the K neighbor model, and taking the result as a final positioning result.
In step S1, the data used in this example are signal strength data from 520 transmitters, in order to locate 735 rooms in which the target may be located, i.e. 735 classifications are made by 520 data; the received signal strength range is-104 to 0, and more missing values exist.
The method for determining the preprocessing mode B in the S2 includes the steps of randomly dividing all data into a training set and a verification set, sequentially filling missing values (in this example, 125-105) by using integers smaller than-104 in a preset range, training a K neighbor model (model parameters are fixed) by using the training set, and obtaining model accuracy by using the verification set; after the test is finished, the training set and the verification set are divided again, and the experiment is repeated; after repeating for 20 times, the numerical experiment result shows that the K nearest neighbor model has the highest average value of accuracy when the missing value is filled to-105; the pre-processing mode B thus determined is to fill the original signal missing value to-105.
The method for determining the preprocessing mode A comprises the steps of randomly dividing all data into a training set and a verification set, adding 105 to all signal intensities, changing the data range into 1-105 at the moment, sequentially filling missing values (20-0 in the example) by using integers which are less than or equal to 0 in a preset range, training a convolutional neural network model by using the training set, and obtaining the accuracy of the model by using the verification set; after the test is finished, the training set and the verification set are divided again, and the experiment is repeated; after repeating for 20 times, the numerical experiment result shows that the mean value of the accuracy rate of the convolution neural network model is highest when the missing value is filled to be-5; the pre-processing mode a thus determined is to add 105 to the total original data and fill the missing value to-5.
When trying to fill the missing value, a large-range and large-step search mode can be adopted, and then a small-range and small-step search mode can be adopted; for example, when the preprocessing mode B is determined in the embodiment, the missing values can be sequentially filled to be-105, -110, -115, -120, multiple experiments show that the corresponding accuracy mean value is gradually reduced, therefore, -105, -106 and-110 are sequentially selected in a small range from-105 to-110, filling is performed according to a small step length, the filling value corresponding to the highest accuracy mean value is searched by adopting the same method as the large range and the large step length, and the missing value is finally determined to be filled to be-105 with the best effect, so that the preprocessing mode B is obtained, and the preprocessing mode A is the same;
in the embodiment, the signal intensity range is-104-0, the optimal preprocessing method (namely, the preprocessing mode B) of the K neighbor model is determined to be-105 by data tests, and the optimal preprocessing method (namely, the preprocessing mode A) of the convolutional neural network model is to fill-5 after the original data is + 105;
in step S3, after the data is processed in the preprocessing mode a (the original data intensity is +105 and then the missing value is filled with-5) determined in step S2, the convolutional neural network is trained by using all the processed data and the corresponding positions (room numbers) thereof;
the structure of the convolutional neural network in this example is a one-dimensional convolutional layer (filter number 32, convolutional kernel size 6, activation function Relu), a Dropout layer (Dropout ratio 0.4), two one-dimensional pooling layers in parallel: a maximum pooling layer (pooling size 5) and an average pooling layer (pooling size 5), a Flatten layer to which each of the two pooling layers is connected, a fusion layer that merges the two Flatten layers, a fully connected layer (neuron number 200, activation function Relu), a Dropout layer (Dropout ratio 0.5), a fully connected layer (neuron number 735, activation function softmax). The number of neurons in the last layer is the number of classes to be performed, which in this example is the number of 735 rooms to be located;
in step S4, the data is processed using the preprocessing method B (filling the missing value of the raw intensity data with-105) determined in step S2, and the K neighbor model is trained using all the processed data and their corresponding positions (room numbers);
in step S5, randomly dividing all data into a training set and a verification set (the data ratio is preset), after the training of the convolutional neural network model and the K-nearest neighbor model by the training set data is completed, sequentially trying different thresholds within a certain range (0% to 100% in this example), and recording an increase value of the accuracy of the verification set data compared with the accuracy of a single K-nearest neighbor model by using a fusion model composed of the convolutional neural network model and the K-nearest neighbor model under each tried threshold; when 88% is determined as the threshold after repeating for 10 times, the fusion algorithm has the most significant effect of improving the accuracy (the average value of the accuracy improvement values is the highest) compared with the K neighbor model, so that the optimal threshold is 88% in the example,
the optimal threshold range determined in the data experiment can be determined by searching in a large range and a large step length first and then searching in a small range and a small step length; for example, the threshold value can be set to 0%, 10%, 20% to 100% in sequence, and after 10 times of experiments, it is found that the accuracy of the fusion model is improved most at 80% and 90% compared with the accuracy of the single K neighbor model, so that 81% and 82% to 90% are sequentially selected as the threshold value in the range of 80% -90%, the accuracy improvement value of the fusion model compared with the single K neighbor model is checked by adopting the same method, and 88% is determined as the optimal threshold value after 10 times of repeated experiments;
in step S6, in actual use, the raw signal intensity data is processed in a preprocessing mode a (signal intensity +105 and missing value-5 are filled), and then sent to the convolutional neural network model obtained in step S3 for judgment, and the position (class corresponding to the maximum value of the softmax full link) judged by the convolutional neural network model and the probability value (maximum value of the softmax full link) of the judgment are obtained.
In step S7, comparing the probability value determined by the convolutional neural network obtained in S6 with the optimal threshold value determined in S5, and if the probability value is greater than or equal to 0.88, taking the position determined by the convolutional neural network model as a final positioning result; if the probability value is smaller than the optimal threshold value of 0.88, the original signal intensity data is processed in a preprocessing mode B (missing values are filled with-105) and then sent to the K neighbor model obtained in the step S4 for judgment, and the judged position of the K neighbor model is obtained and is used as a final positioning result.
Compared with the prior algorithm, the method has the following accuracy:
compared with the existing algorithm, the invention has the following advantages: the method can improve the accuracy of wireless positioning and carry out more accurate identification; for a convolutional neural network algorithm, the method improves the structure and the preprocessing mode of the convolutional neural network, and for a K neighbor algorithm, the method improves the preprocessing mode; meanwhile, by introducing a probability judgment mechanism, the two algorithms are fused, and a more suitable single algorithm can be selected according to the situation by a newly generated fusion algorithm; in addition, the algorithm of the invention does not need to be additionally added on equipment, only changes the algorithm, can avoid the cost in the aspects of capital and time caused by equipment purchase and erection, and has higher practical value.
Claims (3)
1. A wireless positioning fusion method based on deep learning comprises a model building and training stage and a model application stage, and is characterized in that the model building and training stage executes the following steps:
step 1: acquiring original data, and determining a preprocessing mode according to the data, wherein the original data is received signal intensity data of a sensor, the preprocessing mode comprises a preprocessing mode A and a preprocessing mode B, the data processed by the preprocessing mode A is used for training and using a convolutional neural network model, and the data processed by the preprocessing mode B is used for training and using a K neighbor model;
step 2: training and generating a convolutional neural network model and a K neighbor model by respectively utilizing the data processed by the preprocessing mode A and the preprocessing mode B and the corresponding position data;
and step 3: determining an optimal threshold value through a data experiment;
the model application phase performs the following steps:
and 4, step 4: after the intensity data of the signals received by the sensor are processed according to the preprocessing mode A determined in the step 1, classifying the intensity data by using the convolutional neural network model obtained in the step 2 to obtain a position judgment result of the convolutional neural network model and a probability value of the result;
and 5: comparing the probability value with the optimal threshold value obtained in the step 3, and if the probability value is not lower than the optimal threshold value, receiving a position judgment result of the convolutional neural network model; if the probability value is lower than the optimal threshold value, processing the intensity data of the signals received by the sensor according to the preprocessing mode B determined in the step 1, classifying the data by using the K neighbor model obtained in the step 2, and receiving the position judgment result of the K neighbor model;
in step 1, the method for determining the preprocessing mode A aiming at the convolutional neural network model comprises the following steps:
step 1-1': judging whether the original intensity value is less than or equal to 0, if the minimum intensity value a in the original intensity value is less than or equal to 0, adding (-a +1) to all original data for translation, enabling the translated data minimum intensity value a' to be >0, and then randomly dividing the translated data into a training set and a verification set; if the original intensity data are positive values, the data are directly randomly divided into a training set and a verification set;
step 1-2': filling missing values with different c values, wherein c is less than or equal to 0;
step 1-3': generating a convolutional neural network model by using the training set data filled with the missing values in the step 1-2 ', and testing the model accuracy by using the verification set filled with the missing values in the step 1-2' to obtain the model accuracy when each c value is filled with the missing values;
step 1-4': repeating the steps 1-1 'to 1-3' for a plurality of times, wherein the times are not less than 10 times to ensure the effect, and obtaining a plurality of groups of accuracy rates under the condition that each c value is filled;
step 1-5': averaging a plurality of accuracy rates under each c value, and selecting the c value corresponding to the highest accuracy rate average value as the filling of the missing value;
the model trained in the step 1-3' is only used for carrying out accuracy verification and determining a preprocessing mode A, and the trained model cannot be used in an actual application stage;
in step 1, the step of determining the preprocessing mode B aiming at the K neighbor model comprises the following steps:
step 1-1: randomly dividing all original data into a training set and a verification set;
step 1-2: filling the data missing values of the training set and the verification set with different b values each time, wherein the b value is smaller than the minimum value a of the original data;
step 1-3: training the training set data filled with the missing values in the step 1-2 to generate a K neighbor model, and testing the model accuracy by using the verification set filled with the missing values in the step 1-2 to obtain the model accuracy when each value b is filled with the missing values;
step 1-4: repeating the steps 1-1 to 1-3 for a plurality of times, wherein the number of times is not less than 10 times to ensure the effect, and obtaining a plurality of groups of accuracy rates under the condition that each b value is filled;
step 1-5: averaging a plurality of accuracy rates under each b value, and selecting the b value corresponding to the highest accuracy rate average value as the filling of the missing value;
the model trained in the step 1-3 is only used for carrying out accuracy verification and determining a preprocessing mode B, and the trained model cannot be used in an actual application stage;
the step 3 of determining the optimal threshold value through a data experiment comprises the following steps:
step 3-1: randomly dividing all data into a training set and a verification set;
step 3-2: after being processed by the preprocessing mode A, the training set data is used for training a convolutional neural network model, the architecture and parameters of the convolutional neural network model are consistent with those of the convolutional neural network model in the step 2, after being processed by the preprocessing mode B, the training set data is used for training a K neighbor model, and the parameters of the K neighbor model are consistent with those of the K neighbor model in the step 2;
step 3-3: sequentially trying different threshold values, calculating the accuracy of the K neighbor model and the accuracy of the fusion model under different threshold values by using a verification set, and solving the difference value of the accuracy of the fusion model and the accuracy of the single K neighbor model, namely the accuracy improvement value of the fusion model compared with the single K neighbor model;
wherein the fusion model performs the following: the verification set data is preprocessed in a mode A, a convolutional neural network model is used for judging, and if the probability value output by the convolutional neural network model is not lower than the attempted threshold value, the judgment result of the convolutional neural network model is received; if the output probability value is lower than the attempted threshold value, the verification set data is judged by using a K neighbor model in a preprocessing mode B, and the judgment result of the K neighbor model is accepted;
step 3-4: repeating the step 3-1 to the step 3-3, namely randomly dividing all data again, repeating the steps for training and testing to obtain multiple groups of accuracy improvement values of the fusion model compared with the single K neighbor model under each attempted threshold, and solving the average value of the accuracy improvement values corresponding to each threshold, wherein the threshold with the maximum improvement value average value is the optimal threshold;
the K neighbor model, the convolutional neural network model and the fusion model trained in the steps 3-1 to 3-3 are only used for accuracy verification and determining an optimal threshold, and the trained model cannot be used in an actual application stage.
2. The deep learning-based wireless positioning fusion method according to claim 1, wherein the convolutional neural network model generated in step 2 is a one-dimensional convolutional neural network model, and includes a convolutional layer, a pooling layer, a Flatten layer, a merging layer, a fully-connected layer, and a Dropout layer avoiding over-fitting, and the Dropout layer is located behind the convolutional layer or between the fully-connected layers; a multilayer parallel network structure is adopted in the convolutional neural network model, data enter two parallel pooling layers after passing through a convolutional layer, and the two pooling layers are an average pooling layer and a maximum pooling layer respectively.
3. A system using the deep learning-based wireless positioning fusion method according to any one of claims 1 or 2, which is characterized by being provided with a data preprocessing determination unit, a classification model building and training unit, an optimal threshold determination unit and a result classification unit,
the data preprocessing determining unit is provided with a preprocessing mode A determining module and a preprocessing mode B determining module, the preprocessing mode A is obtained through the data preprocessing mode A determining module, and original data are used for training and using a convolutional neural network model after being processed by the preprocessing mode A; obtaining a preprocessing mode B through a data preprocessing mode B determining module, and using the original data processed by the preprocessing mode B for training and using a K neighbor model;
two models are arranged in the classification model building and training unit, and comprise a convolution neural network model generated by training the original data together with the corresponding position after being processed by a preprocessing mode A and a K neighbor model generated by training the original data together with the corresponding position after being processed by a preprocessing mode B;
the result classification unit processes the original data according to a pre-processing mode A determined in the data pre-processing determination unit, then uses a classification model to build and train a convolutional neural network model determined in the training unit for classification, the convolutional neural network model outputs a position judgment result and a probability, compares the probability value with an optimal threshold determined in the optimal threshold determination unit, if the probability is greater than or equal to the optimal threshold, the result classification unit outputs the position judgment result of the convolutional neural network as a final judgment result, if the probability value is less than the optimal threshold, the result classification unit processes the original data according to a pre-processing mode B determined in the data pre-processing determination unit, and then uses the classification model to build and train a K neighbor model determined in the training unit to obtain the final judgment result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010229645.8A CN113453153B (en) | 2020-03-27 | 2020-03-27 | Wireless positioning fusion method and system based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010229645.8A CN113453153B (en) | 2020-03-27 | 2020-03-27 | Wireless positioning fusion method and system based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113453153A CN113453153A (en) | 2021-09-28 |
| CN113453153B true CN113453153B (en) | 2022-05-17 |
Family
ID=77807719
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010229645.8A Active CN113453153B (en) | 2020-03-27 | 2020-03-27 | Wireless positioning fusion method and system based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113453153B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106599941A (en) * | 2016-12-12 | 2017-04-26 | 西安电子科技大学 | Method for identifying handwritten numbers based on convolutional neural network and support vector machine |
| CN108038471A (en) * | 2017-12-27 | 2018-05-15 | 哈尔滨工程大学 | A kind of underwater sound communication signal type Identification method based on depth learning technology |
| CN109151995A (en) * | 2018-09-04 | 2019-01-04 | 电子科技大学 | A kind of deep learning recurrence fusion and positioning method based on signal strength |
| CN110197127A (en) * | 2019-05-06 | 2019-09-03 | 安徽继远软件有限公司 | Wireless signal detection and electromagnetic interference categorizing system and method based on deep learning |
| CN110351244A (en) * | 2019-06-11 | 2019-10-18 | 山东大学 | A kind of network inbreak detection method and system based on multireel product neural network fusion |
| CN110502991A (en) * | 2019-07-18 | 2019-11-26 | 武汉理工大学 | Internal combustion engine health monitor method and system based on random convolutional neural networks structure |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180024239A1 (en) * | 2017-09-25 | 2018-01-25 | GM Global Technology Operations LLC | Systems and methods for radar localization in autonomous vehicles |
| TWI653605B (en) * | 2017-12-25 | 2019-03-11 | 由田新技股份有限公司 | Automatic optical detection method, device, computer program, computer readable recording medium and deep learning system using deep learning |
| CN110517482B (en) * | 2019-07-29 | 2021-06-29 | 杭州电子科技大学 | Short-term traffic flow prediction method based on 3D convolutional neural network |
-
2020
- 2020-03-27 CN CN202010229645.8A patent/CN113453153B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106599941A (en) * | 2016-12-12 | 2017-04-26 | 西安电子科技大学 | Method for identifying handwritten numbers based on convolutional neural network and support vector machine |
| CN108038471A (en) * | 2017-12-27 | 2018-05-15 | 哈尔滨工程大学 | A kind of underwater sound communication signal type Identification method based on depth learning technology |
| CN109151995A (en) * | 2018-09-04 | 2019-01-04 | 电子科技大学 | A kind of deep learning recurrence fusion and positioning method based on signal strength |
| CN110197127A (en) * | 2019-05-06 | 2019-09-03 | 安徽继远软件有限公司 | Wireless signal detection and electromagnetic interference categorizing system and method based on deep learning |
| CN110351244A (en) * | 2019-06-11 | 2019-10-18 | 山东大学 | A kind of network inbreak detection method and system based on multireel product neural network fusion |
| CN110502991A (en) * | 2019-07-18 | 2019-11-26 | 武汉理工大学 | Internal combustion engine health monitor method and system based on random convolutional neural networks structure |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113453153A (en) | 2021-09-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111476294B (en) | Zero sample image identification method and system based on generation countermeasure network | |
| CN111461038B (en) | Pedestrian re-identification method based on layered multi-mode attention mechanism | |
| CN111709311A (en) | Pedestrian re-identification method based on multi-scale convolution feature fusion | |
| CN114564982B (en) | Automatic identification method for radar signal modulation type | |
| CN107038713A (en) | A kind of moving target method for catching for merging optical flow method and neutral net | |
| CN110147745A (en) | A kind of key frame of video detection method and device | |
| CN109039503A (en) | A kind of frequency spectrum sensing method, device, equipment and computer readable storage medium | |
| CN113095370A (en) | Image recognition method and device, electronic equipment and storage medium | |
| CN110263731B (en) | Single step human face detection system | |
| CN111366820A (en) | Pattern recognition method, device, equipment and storage medium for partial discharge signal | |
| CN109492596A (en) | A kind of pedestrian detection method and system based on K-means cluster and region recommendation network | |
| CN111353377A (en) | Elevator passenger number detection method based on deep learning | |
| CN114861761A (en) | Loop detection method based on twin network characteristics and geometric verification | |
| CN111623797B (en) | Step number measuring method based on deep learning | |
| CN113453153B (en) | Wireless positioning fusion method and system based on deep learning | |
| CN117689995A (en) | Unknown spacecraft level detection method based on monocular image | |
| CN116311357A (en) | Double-sided identification method for unbalanced bovine body data based on MBN-transducer model | |
| KR102407834B1 (en) | Method and apparatus for property-based classification of long-pulse radar signals | |
| Si et al. | Radar signal recognition and localization based on multiscale lightweight attention model | |
| Wang et al. | FCM algorithm and index CS for the signal sorting of radiant points | |
| CN111274894A (en) | Improved YOLOv 3-based method for detecting on-duty state of personnel | |
| CN112949385A (en) | Water surface target detection and identification method based on optical vision | |
| CN113362372B (en) | Single target tracking method and computer readable medium | |
| CN116863529B (en) | Intelligent lamp control method based on facial expression recognition | |
| CN115205595A (en) | Train visual positioning test case generation method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |