Classification and Location of Cerebral Hemorrhage Points Based on SEM and SSA-GA-BP Neural Network

In this article, a method to fast classify (intradural hemorrhage, epidural hemorrhage, and cerebral parenchymal hemorrhage) and locate the bleeding points by using the singularity expansion method (SEM) and backpropagation (BP) neural network optimized by genetic algorithm (GA) and sparrow search algorithm (SSA) is proposed. In the simulation model, the bleeding spot with a radius of 3 mm is successfully identified by the approach. The test accuracy in the simulation for both the bleeding’s localization and classification are 98.0% and 97.4%, respectively. Head phantoms that have all been improved over the previous phantom established are used for experiments. A bleeding target with a volume of 3 mL can be identified in the microwave detection system. In the experiment, the accuracy of classification and localization of the bleeding type are 90% and 94.7%, respectively. The final results demonstrate the capability and effectiveness of the method. Faster determination of bleeding point type and orientation means that patients can be provided with different rescue measures accordingly.

stroke is 4-5 h, the earlier treatment means less harm to the patient [1].Therefore, to save more patients' lives, the time for early diagnosis and treatment of stroke is very precious.At present, the general medical instruments used to detect the bleeding point of stroke are computed tomography (CT) detectors and magnetic resonance imaging (MRI) detectors [2].However, these two instruments are complex to use and bulky, and they are not portable medical devices, so they are not convenient for outdoor rescue and other special scenes.
Microwave tomography (MWT) is a new imaging technique, which has been applied to the early diagnosis of breast cancer [3].At present, a novel stroke detection system using microwave imaging technology is being gradually deepened by various research groups.One of the advantages of microwave monitoring is equipment volume.Besides, the detection cost is low, and the detection speed is fast.So, the microwave detection system is expected to become a new generation of equipment to detect strokes for public [4], [5].MWT technology is a kind of non-destructive testing.In the early stage, different dielectric properties of various brain tissues responding to microwave signals can be used to reconstruct the real brain structure.Therefore, brain structures without abnormalities and those with hemorrhagic spots can also be distinguished due to their different electromagnetic properties [6].
There are now several research groups dedicated to early stroke detection and identification system research, in general, most of the research is looking for bleeding point targets in simple phantoms and imaging-detected targets [7], [8], [9].Some methods are time-consuming using inverse scattering.The inversion method among these methods can correctly evaluate the performance of the reconstruction model method.But the iterative time cost is fatal [2], [10], [11].Rodriguez-Duarte et al.'s [12], [13] team at the Polytechnic University of Turin has been working on microwave stroke classification detection.They proposed a method by using differential approximations and distorted Born approximations to image the target [12], [13].Mariano et al.'s [14] team at the Polytechnic University of Turin used a dataset generated by full-wave simulation to test the type and direction of stroke.The group of King's College London, Strand, London, applied the distorted Born iterative method, two-step iterative shrinkage thresholding (DBIM-TwIST) algorithm to differentiate the hemorrhagic and ischemic strokes [15], [16], [17].In the latest study, the team from Czech Technical University used support vector machines (SVMs) to establish a comprehensive training set, which can classify hemorrhagic stroke and ischemic stroke [18].The head phantom used in this reference only contains one material with dielectric properties equal to the average dielectric parameters of a human head.In this article, the raw signals were set as the input data of the learning-byexample (LBE) strategies.The study in [18] is a continuation of the previous research in [19].In [18], the raw signals were proposed by using the principal component analysis (PCA) algorithm to improve the accuracy of the classification.However, the algorithms proposed in [19] only trained on the data obtained from numerical simulations.The above research teams have made a great breakthrough in the research on the imaging part of the stroke.Stroke rescue detection needs the necessary conditions of urgency and speediness.Qualitative analysis of the bleeding type and bleeding location is in the first place.Analyzing the bleeding type first and then determining the location can narrow the imaging range, which is the necessary prerequisite for small-scale and highefficiency imaging in detection.Determining the type and the orientation of the bleeding point is also very important for early rescue and protection of patients.
To solve the above problems, this article proposed a method using singularity expansion method (SEM) and sparrow search algorithm-genetic algorithm-backpropagation (SSA-GA-BP) neural networks to quickly classify and locate the types of bleeding points.In this study, only the time-domain signal is needed to extract the feature singularity as the input data of the neural network.The attenuation factor and amplitude at the corresponding frequency are extracted from the time-domain signal.Classification efficiency can be significantly increased by using fewer features and fewer classification labels.This method combines the signal processing method and the neural network to identify the cerebral hemorrhage, which saves a lot of detection time.Table I provides a side-by-side comparison between the work presented in this article and the most recent related research.
The details of the research in this article are described as follows.In Section II, the details of establishing the 2-D head model by applying the finite-difference time domain (FDTD) method are shown.The details of the proposed method used in the simulation will be shown in Section III.The results of the simulation and discussions are shown in Section IV.Section V shows the experiment verification of the proposed method.The conclusion of this article is presented in Section VI.

II. HEAD MODELING IN SIMULATION
In this section, the details of the 2-D model of the brain in the simulation and the electromagnetic characteristics of brain tissues will be shown.The details of the antenna set around the head model will also be introduced in this section.

A. Electromagnetic Characteristics of the Brain Tissue
The structure of the human brain is complicated.Fig. 1 is an MRI of the real head, which shows the main structure of the human head.This MRI of the head is the key data for simulation modeling.It can be seen in Fig. 1 that the scalp is the outermost layer of the head, which is together with subcutaneous fat to protect the head.The skull plays a supporting and protecting role.Further inside are meningeal structures, including cerebrospinal fluid, and fat tissues between the skull and the brain.Deep inside the brain are the gray matter and white matter.In Fig. 1, the chosen main structures in the simulation modeling are labeled.
The electromagnetic characteristics of human brain tissue are the basis of analyzing the propagation of microwave signals in the brain.Human tissues have different electrical properties [20].Human biological tissues show different reaction effects in the external electromagnetic field due to their different electrical characteristics.Electrical characteristics include conductive characteristics and dielectric characteristics [21].
FDTD iterative analysis is applied in this study.The ideal frequency band is selected within 1-5 GHz according to relevant research results [22].The electromagnetic parameters of the brain tissue at 1 GHz are shown in Table II.

B. FDTD Simulation of Head Model
The head modeling in this article is based on the 2-D FDTD method.Before the iterative operation of FDTD, the analysis and processing of the MRI are carried out.The MRI is separated and labeled for each brain tissue so that the electromagnetic parameters can be assigned to each computing cell in the subsequent iterative operation [16], [23], [24].Fig. 2 shows the established 2-D head model and the location of the antenna array.Each color represents a tissue.The portion outside the head in the figure is set as an air layer.
To simulate the antenna position of the stroke detection equipment, the 16 antennas are evenly arranged around the scalp.During the detection, one antenna transmits a signal, the other antennas receive the signal, respectively.In this simulation, the emitted signal is set as the sixth-order Gaussian pulse signal [25].

III. SIGNAL DATA PROCESSING AND CLASSIFICATION
In this section, the approach of signal processing is specified.Different bleeding point types and orientations are set as different category labels for the collected raw signals.Then, the unique feature singularity information is extracted from the original signals.Finally, the GA-BP neural network is used in the simulation stage to classify the location and type of bleeding points.In the experimental detection phase, the SSA-GA-BP neural network is used to classify the location and type of bleeding targets.

A. Set Category Label
In this section, a bleeding spot with a radius of 3 mm is set in the brain model.Three bleeding types and four orientations  The central coordinate position of each bleeding point will be shown in Table III.In this article, each orientation of a particular kind of bleeding point is set as a label, so there are 12 different types of labels in total.After determining the tag, the FDTD method is applied to obtain the raw signals.
During the detection, one antenna transmits a signal, the other antennas receive the signal, respectively.A 16 × 15 signal matrix is obtained after a single detection.

B. Extraction of Feature Poles
After obtaining all the raw signals related to different types of bleeding points, the singular point expansion method is applied to the analysis of transient electromagnetic fields due to the natural resonance phenomenon of the electromagnetic system itself [26], [27].Signal pole extraction is a very important application in radar target recognition, nuclear magnetic resonance, speech recognition, and other fields.
The current model-based linear prediction pole extraction methods mainly include the Prony method, the general pencil of functions (GPOFs), and the matrix pencil method (MPM).In the direction of early breast tumor target recognition, the characteristic poles of the tumor are extracted through the Prony algorithm in the singular point expansion method, so that different growth stages of the tumor can be identified [28], [29], [30].Finally, this research chooses the Prony algorithm to extract the characteristic poles of the signals.
The Prony algorithm is actually a polynomial calculation method, and its most essential problem is to extract pole information from instantaneous data with equal time intervals.Here, the signal y(t) obtained by FDTD iteration is discretized to obtain y(n).Denote y(n) as a signal model of attenuation exponential sum The pole information is represented by In order to perform signal analysis, it is necessary to extract the characteristic components from the signal represented by (2), that is, it is necessary to extract the best M. Prony's theoretical definition constructs a sample function matrix where p is the real order of the Prony algorithm, generally take P e = N /2, and N is the time step of the signal.During FDTD iteration, the iteration time step is set as 6000 steps, so N = 6000 The following uses singular value decomposition (SVD) to determine the effective rank P of the matrix R. The dimension of the matrix S ( p) is ( p + 1) × ( p + 1) where σ 2 j j is singular values of matrix R v(i, j) is a factor on row i and column j of matrix v.
Finally, a and z need to be solved.The relationship between a and z is as follows: where a is represented by the following equation: S (− p) is the inverse matrix of S ( p) .At this time, assuming that z has been obtained, (1) can be simplified to the equation of parameter b in the following equation: The relationship between b and is • y.According to b, the information contained in the characteristic singularity can be obtained, including amplitude, attenuation factor, and frequency.The amplitude, attenuation factor, and frequency can be calculated by using the following equation: Then, the corresponding feature pole information, including amplitude, attenuation factor, and frequency of all the signals, can be obtained according to the abovementioned method.Table III shows the feature poles information.Different information corresponds to signals obtained from the simulation model with different bleeding points.The center coordinates are the positions of the bleeding points.They are the coordinates of the center of the bleeding point in the set 2-D model.The bottom-left corner of the 2-D model in Fig. 2 is defined as the coordinate origin.In Table III, only a set of information at 1 GHz corresponding to each signal is shown.There are 2200 sets of pole information obtained in the actual calculation.The data volume is so huge that it cannot be fully displayed here.

C. SSA-GA-BP Neural Network
This section shows the procedure of classifying the feature poles dataset obtained in Section III-B.There are many types of neural networks at present, but in this study, the SSA-GA-BP neural network is selected as the neural network model for classification.BP neural network is the most complete series of artificial neural network methods so far, and it is easier to apply to actual test scenarios [31], [32].BP neural network is mainly composed of an input layer, hidden layer, and output layer [33].
In this study, two networks are designed.The output layer of the first network has three outputs to help classify the type Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply. of bleeding target.The output layer of the second network has four outputs to help locate the bleeding target.
The overall performance of the BP network particularly depends on the initial weights and bias.So, how to obtain the initial weights and biases correctly is a key issue in improving the performance of the BP neural network [34].The most suitable weights and biases for the network need to be chosen.The number of hidden layers and the number of nodes in the BP neural network play a decisive role in the training speed and recognition accuracy of the entire network [35].SSA-GA optimization of the BP neural network can focus on solving these two problems.The SSA intelligence algorithm is a relatively novel swarm intelligence optimization algorithm.The core of the algorithm is inspired by sparrows' strategies of finding food and dealing with natural enemies.Compared with other popular optimization algorithms, the sparrow search algorithm (SSA) improves the search accuracy of the sample target and reduces the optimization time.
The genetic algorithm transforms the selection of initial weight and threshold into the problem of solving the optimal value of the function and assigns it to the BP neural network after selection.In terms of the selection of hidden layers, too many hidden layers and the number of nodes will have a negative impact on the network and negatively damage the processing capacity of the network.
The flowchart of the optimization algorithm is shown in Fig. 5.In the optimization process, the data are initialized first.Then, the SSA algorithm is used to find the appropriate number of hidden layer nodes.GA optimization algorithm optimizes the corresponding weight and bias.Then, the stopping condition is checked to decide whether to continue or to repeat the above loop.The stopping condition is whether the loss function (multiclass cross-entropy) in the network has converged within the set epochs.It will also check whether the current number of SSA iterations is the same as the set number of stops.
It is worth noting that the signal in the simulation stage is obtained under the ideal condition, so its characteristics are more obvious, and the optimal network can be obtained only by GA algorithm optimization in the simulation stage.The number of nodes, the weight, and the bias in the network are During the simulation, the final network structure optimized by the genetic algorithm is shown in Fig. 6.Fig. 6(a) is the first neural network for judging the type of bleeding point in the simulation phase.The number of nodes in the first hidden layer is 14.The number of nodes in the second hidden layer is 18.The number of nodes in the output layer is 3. Fig. 6(b) is the second neural network for judging the direction of the bleeding point in the simulation phase.This neural network is a single hidden layer neural network with six nodes, the number of nodes in the output layer is 4.
The ultimate research purpose is not only to determine the type of cerebral hemorrhage point but also to determine the location of the hemorrhage point, to facilitate the later imaging processing, or to provide corresponding protection measures for patients.The specific process of the overall detection method is shown in Fig. 7. First, the raw signals are collected Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.from FDTD models.Then, the raw signals are processed by SEM method to achieve the characteristic poles, which are set as the input data of the neural network.The first network classifies the type of cerebral hemorrhage, whereas the second network judges the orientation of cerebral hemorrhage.

IV. SIMULATION RESULTS AND DISCUSSION
This section contains the results and discussions of the proposed method in the simulation model.
The obtained 2200 sets of extreme point information were randomly divided into two groups, one group was used to test the first network for judging the type of bleeding point, and the other group was used to test the second network for judging the direction of bleeding point.Both networks have 1100 feature poles information.To improve the accuracy of judgment, the pole data are cleaned by the data cleaning method, which is also a method for data quality analysis.The data cleaning method is the process of reviewing and verifying data to remove duplicates, correct errors, and provide data consistency.After data cleaning, there are 1037 sets of data input into the first network and 1008 sets of data input into the second network.The data cleaning procedures of the two networks are different because the number of label categories is different.Although the number of cleaned data is different, it does not affect subsequent tests.In this study, 85% of the total data is selected as the training set and 15% as the test set.That is, the first network has 882 sets for training and 155 sets for testing.The second network has 857 sets for training and 151 sets for testing.The accuracy on the test set is 98.0%.Fig. 11 shows the relationship between loss function value (multiclass crossentropy) and epochs.
The above confusion matrixes show the feasibility of the method proposed in this article.To further verify the feasibility and practicability of the signal analysis method SEM and SSA-GA-BP network, the experimental verification of the brain phantom is carried out in Section V.

V. EXPERIMENTAL VERIFICATION
This section introduces the design, and test details of the experiment system.The experimental validation is a critical part.In this section, the details of establishing a phantom of the real human head are presented.The human head model consists of the skull, cerebrospinal fluid (CSF), white matter, gray matter, and blood in experiments.Antenna details around the model are also covered in this section.At the same time, the above algorithms are also verified in this experimental system.

A. Phantom Establishment of the Human Head
In this experimental system, a skull model is made based on real MRIs of real patients.The skull model is composed of two parts, one is a hollow skull and the other is a mandible.The main materials of this skull phantom are polyvinyl alcohol and calcium sulfate.The front of the skull phantom, the side of the skull phantom, the top of the skull phantom, and the interior of the skull phantom are shown in Fig. 12(a)-(d).This skull provides a container for filling the white matter and gray matter mimicking materials.simulated fluid with a volume of 3 mL.The composition of the red simulated blood material here is 1.5%-2% sodium alginate and 1% potassium chloride.Iron trioxide is used as a coloring agent to make the color similar to that of blood.Fig. 13(a)-(d) are the top-right dura mater hemorrhage, the bottom-right dura mater hemorrhage, the top-left dura mater hemorrhage, and the bottom-left dura mater hemorrhage.Fig. 13(e)-(h) are the top-right parenchymal hemorrhage, bottom-right parenchymal hemorrhage, top-left parenchymal hemorrhage, and bottom-left parenchymal hemorrhage.In the experiment, red simulated blood with similar electrical properties to the real blood solution was used to fill in the simulated material.The bleeding target is with a size of 5 mm in radius.The use of simulated liquid filling is to mimic the real bleeding case that occurred in the human head.
The head model used in this experiment is a simplified model to validate the capability of the proposed signalprocessing algorithm.Therefore, the skull, white matter, gray matter, and a small amount of CSF and blood are included.The electrical characteristics of the white matter, gray matter, CSF, and blood-mimicking materials inside the model are measured by the vector network analyzer in the 0-10-GHz frequency range at room temperature.Fig. 14 shows the comparison between the measured electrical characteristics of the developed materials and the electrical characteristics of actual brain tissues.In this study, the scale of all the materials of the simulated brain tissue is adjusted to mimic the real brain tissue.So, the difference in the measured dielectric parameters between the developed material and the actual brain tissue is smaller than before [36], as shown in Fig. 14.

B. Antenna Design
The antenna used in this system is a "12-line Archimedes antenna."The working frequency range of the antenna is 0.8-10 GHz.The antenna size is 60 × 60 mm.The antenna feeding method adopts balun feeding to realize the impedance matching between the coaxial line and the radiator.The antenna used in this experiment is designed by electromagnetic simulation software, as shown in Fig. 15

C. Experiment System Setup
The experiment system setup includes a vector network analyzer Agilent E5080B, a switch box, a laptop, an antenna array with eight antennas, and a head model.Fig. 19(a) shows the photograph of the experiment setup.The laptop is used to control the switch box.Four antennas are connected to port 1 of the vector network via the switching matrix, and the other four antennas are connected to port 2 of the vector network via the switching matrix.These eight antennas are evenly arranged around the head phantom.Fig. 19(b) shows the detailed diagram of the antenna arrangement and the location label of the bleeding target.
During the experiment, one signal was propagated in the head phantom through two antennas.A program in the laptop computer controls the RF switch box to switch the antenna.The obtained S 21 results are displayed on the vector network analyzer.Eight groups of signals are stored by the laptop.

D. Experimental Verification
In this experiment system, the raw data are S 21 which are detected in the range of 0.8-10 GHz.Fig. 20 shows four waveforms of S 21 .S 21 (o5 − 2) is the S 21 when antenna II emits a signal, antenna V receives the signal.S 21 (o5 − 3) is the S 21 received by antenna V, when antenna III emits a signal.The SSA is added in the experimental verification to optimize the number of nodes, and GA to optimize the weight and bias.As a result, the optimal network is given.In this way, SSA and GA are processed in parallel to form an overall optimization algorithm, which improves the accuracy when the number of training samples is small in the experiments.
Fig. 21(a) is the first neural network for judging the type of bleeding point.The number of nodes in the first hidden layer is 14, the number of nodes in the second hidden layer is 18, and the number of nodes in the output layer is 2. Fig. 21(b) is the second neural network for judging the direction of the bleeding point in the experimental verification phase.This neural network is a single hidden layer neural network with 16 nodes, and the number of nodes in the output layer is 4.      the bleeding point is 96.5%, and the final accuracy rate of the test set is 94.7%.Although there are some classification errors, the overall classification accuracy is good.The location of simulated bleeding targets in physical brain phantoms can be successfully detected and classified.The final results of the experiment prove that the detection algorithm proposed in this article is robust.

VI. CONCLUSION
Cerebral hemorrhage is a disease with a high fatality rate.The localization and classification of cerebral hemorrhage are judged directly by microwave signal, which saves precious life-saving time for patients.This study proposes a method for quickly determining the type and direction of the bleeding point.The simulation experiment and the designed microwave cerebral hemorrhage detection system are used to verify the feasibility of the Prony characteristic singularity signal processing algorithm combined with the SSA-GA-BP network.In the simulation, the test accuracy of classification and localization of the type of bleeding was 97.4% and 98.0%, respectively, which means the bleeding target with a radius of 3 mm can be located and classified successfully.
To validate the applicability of this method, experimental verification was conducted.The ability of the proposed method was demonstrated on a head phantom filled with developed materials of each brain tissue.In the experiment, the test accuracy of classification and localization of the bleeding type are 90% and 94.7%, respectively.The results show that the proposed method can successfully identify cerebral hemorrhage with a volume of 3 mL.After accurate computer timing, 20 s are taken for the processing procedure (including signal extraction, signal feature poles information extraction, and using the trained well network to get the final classification result).
The method proposed in this article still takes less time in the whole detecting procedure compared with the general medical instruments.All the results in this article show that the proposed method saved a lot of time for testing.In an emergency, it means more lives can be saved.
In this study, the original microwave signal data is not used to directly input the neural network for training.The feature pole information extracted by the signal processing algorithm in the early stage is the data input to the neural network.The reason why pole extraction is necessary can be divided into two categories.The first is the stability of the final results.The "end-to-end" mode is the original microwave signal directly input to the neural network.The neural network in this mode requires more parameters.This will cause more uncertainty.This type of judgment has more factors of interference, which will cause poor stability.While the input data is the processed signal rather than the original signal, it is more stable.The second is to use neural networks as an auxiliary tool for judgment.It means that signal processing is an important part of this study.The classification and localization do not depend entirely on the learning of neural networks.
In the future, different sizes of the hemorrhage point could be detected to help improve the applicability of this proposed method.More situations of the real human brain will be considered to improve the detection accuracy of the cerebral hemorrhage point.

Fig. 2 .
Fig. 2. Simulation model of the head and the location of the antennas.

Fig. 3 .
Fig. 3. Three different types of bleeding points with a radius of 3 mm are set in the 2-D model.(a) Epidural hemorrhage model with bleeding points in four directions.(b) Intradural hemorrhage model with bleeding points in four directions.(c) Parenchymal hemorrhage model with bleeding points in four directions.

Fig. 4 .
Fig. 4. Signals received by A 8 and emitted by A 1 in 12 bleeding models.

Fig. 6 .
Fig. 6.Final network structure optimized by the genetic algorithm.(a) First neural network for judging the type of bleeding point.(b) Second neural network for judging the direction of the bleeding point.

Fig. 7 .
Fig. 7. Specific process of the overall detection method.

Fig. 8
shows the confusion matrix diagram for the bleeding point type classification network.Fig. 8(a) is the training set's confusion matrix diagram, and Fig. 8(b) is the test set's confusion matrix diagram.The abscissa represents the predicted label of the sample, where 1 represents epidural hemorrhage, 2 represents intradural hemorrhage, and 3 represents parenchymal hemorrhage.The ordinate represents the

Fig. 10 .
Fig. 10.Confusion matrix diagram of the localization network.(a) Training set's confusion matrix diagram.(b) Test set's confusion matrix diagram.

Fig. 11 .
Fig. 11.Relationship between loss function value (multiclass cross-entropy) and epochs in the localization network.

Fig. 10
Fig. 10 shows the confusion matrix diagram for judging the localization network of bleeding points.Fig. 10(a) is the final training set's confusion matrix diagram of the judgment results, and Fig. 10(b) is the test set's confusion matrix diagram of the judgment results.The abscissa represents the predicted label of the sample, where 1 represents top left, 2 represents top right, 3 represents bottom left, and 4 represents bottom right.The ordinate represents the true label of the sample.The accuracy of the training set is 98.7%.The accuracy on the test set is 98.0%.Fig.11shows the relationship between loss function value (multiclass crossentropy) and epochs.The above confusion matrixes show the feasibility of the method proposed in this article.To further verify the feasibility and practicability of the signal analysis method SEM and SSA-GA-BP network, the experimental verification of the brain phantom is carried out in Section V.

Fig. 13
is a skull model filled with mimicked brain tissue materials.At the same time, there are eight head phantoms with eight hemorrhagic sites at different locations established to obtain raw signals.The bleeding target is assumed by red

Fig. 12 .
Fig. 12. Skull phantom used in this experiment.(a) Front of the skull phantom.(b) Side of the skull phantom.(c) Top of the skull phantom.(d) Interior of the skull phantom.

Fig. 14 .
Fig. 14.Measured relative permittivity of the developed materials and the actual relative permittivity of real brain tissues.
(a).Fig.15(b) is the prototype of the antenna used in the experimental system.Fig. 16 is the photograph of the measurement setup for the used antenna.The reflection coefficient (S 11 ) obtained from the simulation and the measured reflection coefficient (S 11 ) of
. The frequency range of the used antenna in which the reflection coefficient is lower than −10 dB is 0.8-10 GHz.

Fig. 18 .
Fig. 18.Antenna radiation patterns of E-plane at different frequencies.(a) E-plane radiation pattern at 1 GHz.(b) E-plane radiation pattern at 3 GHz.(c) E-plane radiation pattern at 5 GHz.(d) E-plane radiation pattern at 7 GHz.

Fig. 18 (
Fig. 18(a)-(d) is the simulated and measured far-field radiation patterns in the E-plane at 1, 3, 5, and 7 GHz.It can be observed in Fig. 18 that the simulated and measured far-field radiation patterns of the used antenna are similar over the whole frequency.The antenna has a wide half-power beamwidth (about 70 • ) at each measured frequency.

Fig. 19 .
Fig. 19.Measurement setup.(a) Photograph of the measurement setup.(b) Photograph of detailed diagram of the antenna arrangement and the location label of the bleeding target.

Fig. 20 .
Fig. 20.S 21 received by antenna V and emitted by antennas II, III, VII, and VIII, respectively, under the condition of simulating the bleeding point in position 1.

Fig. 22
shows the confusion matrix of the final training set and test set in the classification network.Fig. 22(a) is the final training set's confusion matrix of the classification results.Fig. 22(b) is the final test set's confusion matrix of the classification results.Fig. 23 shows the relationship between the multiclass cross-entropy and epochs during the training of the bleeding-type classification network.It can be seen in the figure that within the set epochs, the network converges.Fig. 24 shows the confusion matrix of the final training set and test set in the localization network.Fig. 24(a) is the training set's confusion matrix with the localization results.Fig. 24(b) is the test set's confusion matrix with the localization results.Fig. 25 shows the relationship between the

Fig. 21 .
Fig. 21.Final network structure optimized by the genetic algorithm and SSA algorithm.(a) First neural network for judging the type of bleeding point.(b) Second neural network for judging the direction of the bleeding point.

Fig. 23 .
Fig. 23.Relationship between the loss function value (multiclass cross-entropy) and epochs in the classification network.

Fig. 24 .
Fig. 24.Confusion matrix diagram of the localization neural network.(a) Training set's confusion matrix diagram.(b) Test set's confusion matrix diagram.

Fig. 25 .
Fig. 25.Relationship between the loss function value (multiclass cross-entropy) and epochs in the localization network.

TABLE III FEATURE
POLES INFORMATION OF DIFFERENT TYPES AND DIFFERENT LOCATIONS OF BLEEDING POINTS