Application of Multi-sensor Data Fusion Based on AIS to.doc

精品论文大全Application of Multi-sensor Data Fusion Based on AIS toLandslide Data Analysis1Xu Qiang, Yuan YongNational Laboratory of Geohazard Prevention and Geoenvironment Protection, ChengduUniversity of Technology, Chengdu P. R. China (610059)AbstractCurrent researches on landslide forecast are to a great extent dependent on the information from somekey monitoring points. How to effectively use the information from all the different monitoring points in different parts of the landslide is a basic subject of landslide forecast.After a brief survey of multi-sensor data fusion and AIS with emphasis on data fusion methods and immune algorithm, this paper applies the algorithm of multi-sensor data fusion based on AIS to dealing with the monitoring information, and very successful results are obtained in the case study.Keywords:Multi-sensor Data FusionAISLandslide1.IntroductionIn actual monitoring work, many monitoring instruments are usually installed in different parts of landslide and many groups of information reflecting the status of landslide are obtained. And we can use the most key information, displacement usually, of the most key point to forecast the tendency. This method, however, depends to a great extent on subjective judgments (the key point choosing and the key information choosing), thus causing a large waste of information obtained from monitoring, and even inevitably resulting in the challenge of the reliability of the forecast got from corresponding model.It is really not fortunate that we have no theory model or method of forecasting with multiple variables so far 1,2. Consequently, how to transfer the multidimensional data into one dimensional data is a valuable topic. Many researchers have tried to achieve something, but current researches are still not perfect 3. So there is indeed an obvious need for a more thorough study of this topic.In order to use the monitoring information as synthetically as possible, an algorithm of multi-sensor data fusion based on AIS is probed and Danba landslide is taken as a case study in this paper.2.Multi-sensor data fusionMulti-sensor data fusion is a functionally synthetical simulation of the human brain processing complex issues. In the fusion the information of a specific scene acquired by two or more sensors at the same time or separate times is combined to generate an interpretation of the scene not obtained from a single sensor 4,5. Alternatively, multi-sensor data fusion is done to reduce the uncertainty associated with the data from individual sensors 6.The essence of the fusion lies in using the information from different sensors, processing (including analyzing, integrating, allocating and applying) the data in accordance with the monitoring time series, optimizing the horary and special complementation and redundancy from among sensors in light of a certain optimal criterion, conquering the uncertainty and limitation consulted by the single sensor, and increasing the validity of the whole system and yielding an explanation and description of the monitoring environment 7. Mathematically, different monitoring values coming from different sensors make up a subspace, and each monitoring subspace projects to the fusion information in1Support by Specialized Research Fund for the Doctoral Program of Higher Education (SRFDP), China (Grant No.20040616005) and Application and Fundamental Research of Science and Technology of Sichuan Province, China (Grant No. 05JY029-087-1)-10-accordance with a given criterion. And the interpretation result is synthetical data based on the data from all the sensors used in the data fusion.Fusion can be performed at either the data, feature, or symbol level of representation. In data-based fusion, data from different sensors are combined to create new data with a better data-to-noise ratio than the original data. Feature-based fusion consists of extraction feature from different data on the basis of improving the performance of data processing tasks. Symbol-level or decision-level fusion consists of merging information at a higher level of abstraction. Usually, data fusion is made at feature or symbol level because of the heterogeneous multi-sensors. And the multi-sensor system has the virtue of more reliability, robustness, expansibility and discernibility 8,9.Algorithms for fusion of sensory information fall into the categories of weakly coupled fusion and strongly coupled fusion. In weakly coupled fusion, different sensor modules are assumed to be independent, and the fusion consists of combing the outputs from each of the sensor modules. In the strongly coupled fusion the decision of one sensory module is affected by the output of another sensory module 6.3.Artificial immune systemBIS, Biological Immune System, is the complex system in the animal body that enables animals to resist diseases. This system, functioning in maintaining the balance of internal mechanism, has the capacity to adaptively recognize, to eliminate foreign antigenic body, to learn, to memorize, and to control self-adaptability.Inspired by BIS, experts in AI have established AIS-Artificial Immune System 10. In 1970s, Jerne first put forward the idiotypic network theory 11，12, and Perelson developed the algorithm of AIS bysetting forth the concept of dynamic feedback immune memory 13. The commonly adopted algorithms comprise negative selection algorithm, clone selection algorithm, immune learning algorithm etc. Up to now AIS has been applied to robot controlling, computer security, machinelearning, and many other fields 1417. The application of AIS to data analysis was first put forwardby by Hunt and Coole who set up an algorithm to instruct machine learning in the classification of DNA 18. The results of their experiment showed that AIS is superior to ANN and other algorithms except that it is short of model for general purposes, and the parameters are weak in the presentation of real-time adaptability. Timmis improved Hunt and Cooles model and algorithm and put forward his conception of RLAIS which defined the fundamental compositions of a net as ARB and established computational formula of the degree of stimulation between ARBs. The stability and robustness were all improved to a great extent, the system has developed the ability to proceed with its learning 19.AIS, a simulation to the BIS, is made up of numerous molecules, cells and immune organs. All parts of the system are interrelated and interact on each other and implement immunity together. AIS can process tremendous information and has the attractive traits as follows 20:Ability of pattern recognition: AIS can recognize different Antigen (Ag) and produce relevant stimulation. In the pattern recognition, self and non-self can be distinguished.Ability of feature extraction: Antigen presentation cells (APCs) can extract the feature ofAntigen. And the process is equivalent to filtering.Ability of diversity keeping: Factors such as heredity immune system can generate enough lymphocytes to defend known or unknown Antigens.Ability of learning: The system can learn the structure of specific Ag by experiences and the learning ability is dependent on the supplementation mechanism (Clonal expansion ).Ability of memorizing: When the lymphocytes are activated, some cells change into specific andaddressable memory cells. Immunocytes with dynamic life-span need continuous stimulation coming from the remainder Ag so as to keep the balance of the system.Other abilities such as self-adapting, specificity and robustness are the other important features of AIS.3.1 Immune algorithmMost of T cells, B cells and Ab are integrated into a detector simulating the Ag processing in biological system, including Ag generating, self tolerance, clonal expansion and immune memory. The following is the main procedures of this algorithm 21:Antigen defining: Abstracting the problem needed to solve into the Ag, then Antigen recognition is the corresponding question solving.Initial Antibody generating: Defining the solution of the problem as the Ab group so that theaffinity between Ag and Ab is corresponding to assessment of the solution-the higher affinity the better solution. Similarly to GA, first generate the initial Ab to match the random solution of the problem.Affinity computing: Computing the affinity between Ag and Ab, usually a certain distance.Clone selection: Reproducing the Ab having high affinity with Ag, restraining the Ab with too higher consistency (equal to avoiding the local optimum solution) and eliminating the Ab with lower affinity, getting solution with high diversity (equal to looking for global optimum solution). Ab suffers mutation when it was cloning, so in the process of clone selection, Ab promotion and clone deletion are correspondent to optimum solution promotion and non-optimum solution deletion respectively.New Ab group evaluating: If the result does not meet the terminal condition then turn to the second procedure and restart, otherwise, the current solution is the best one for the problem. Following is theflow chart:Import AntigenGenerate Antibody randomlyAffinity ComputingPromotion and Suppression of AntibodyMemorize the Antibody withhigh AffinityAdd Some New Antibody Generated randomly to Antibody SetNMeet terminal conditionYEndFig1. Main Procedures of Immune AlgorithmEssentially, both immune algorithm and evolutionary computation are group search strategies and they all emphasize the information exchange between individuals.4.Multi-sensor data fusion based on AISIn terms of solution methodologies, neural net-based approaches seem to be the most predominant ones in data fusion application. This may be due more to the fact that neural nets have become generally available in the form of commercial packages than to any proven superiority over all the other tools available in the information fusion area 22. Admittedly the neural net-based approaches are not exhaustive, even the multi-sensor data fusion.The optimality rules of multi-sensor data fusion are diverse. Such as Least Square, Weighted Least Square Method, Least Mean Square error, Maximum Likelihood and Bayesian criterion, and the like. Its processing modes include Least Square, Weighted Least Square and other batch methods, at the same time, include Kalman, Alpha-Bata and other sequence processing methods 23-26.Because immune algorithm needs no apriori knowledge of monitoring data and can get the fusion result with least mean square error only in the light of monitoring data from the sensors, this paper adopted this method to realize the multi-sensor data fusion and find the optimal fusion result.Following is the flow chart:Sensor1X1Sensor2X2MMMulti-sensorData fusionBasedWXOn ImmuneAlgorithmSensor nXnFig2. The flow chart of multi-sensor data fusion based on AISL 2 , 2 , 2If the n monitoring data are X1, X2,Xn, and the relevant deviations are 12n ， eachtwo of the series are independent on each other. If the optimal weight of each sensor makes up ofthe vector W, whereW = (1 ,2 ,Ln ) ，irepresents the ith weight of the ith sensor, thefusion result of the system will be shown thatX = WX T =nni =1i X iwherei = 1i =1, the total deviation can also be given as 2 = E( X X )2 Because the elements of series of X1, X2, , Xn, are independent on each other, and each one of the series is the unbiased estimation of X, soE( X X i )( X X j ) = 0(i therefore,j, i, j = 1, 2L n)nn 2 = E X X 2 = En22i =1i X 2i =1i X i = En2 2i =1i ( X X i ) = i =1 2i iVirtually, thein the above equivalent is a quadratic function of many variables. If thenextreme values of 2 subject to the constrainti = 1 are founded, then the fusion result willi =1be got. And so the issue really comes down to finding extremum of the function subject to a constraint. The weight finding can be viewed as Ag and the optimal weight as Ab, then apply the immune algorithm to find the optimal weight satisfied the extremum with least mean square error.Finally the fusion result, the global optimum X under the condition of least mean square errorcan be obtained.5.Case study5.1 Danba landslideDanba landslide, a large-scale landslide located in Danba town, Sichuan province, southwest of China, occurred in 2005. The landslide, an oversize stack layer landslide, developed from fossil landslide, the landslide surface taking the shape of a half ellipse with lengthwise 290m, acreage of0.08km2, thickness 18-45.23m, average thickness 30m, cubage 220 thousand m3.The landslide deformation glided as a whole unit, and the gliding was powerful. The landslide deformation first pulled apart from the back edge and formed a split between the tail and the back edge around Feb. 15th, and shearing cracks on both sides took shape and extended to further parts of the slope. The monitoring data showed that the main body of the landslide moved along the central axis at a fast speed, and the displacement rate of the rest was similar. Once the whole landslide moved down, more than half of the town4,000 people and their properties would be destroyed.According to the monitoring data Sichuan Geotechnical institute provided, from Jan 24 to Apr 6,2005, the table can be shown as follows:Table1 The Monitoring Data of Danba LandslideMonitoringDatesensor1sensor 2sensor 5sensor 6sensor 7sensor 8sensor 9sensor 112005-1-22000000002005-1-23-1.8-2-6.5-2.3-3.3-5.6-5.4-7.52005-1-245.88.6-4.37.83.1-3.22.1-6.82005-1-2510.113.8-1.215.85.23.45.45.32005-1-269.918.21.421.87.45.18.76.62005-1-2715.828.33.930.59.77.810.211.42005-1-2822.336.52.242.18.68.99.213.72005-1-292946.15.344.610.810.612.310.22005-1-3028.348.22.354.113.98.79.77.32005-1-3135.457.35.56415.69.412.610.52005-2-137.866.711.171.314.511.918.720.62005-2-243.67720.581.926.524.226.831.72005-2-351.986.623.990.53326.63442.22005-2-456.295.530.910038.735.444.652.32005-2-563.3104.938112.651.241.953.868.62005-2-675.9122.853.1129.862.457.875.984.62005-2-780.1133.959.9139.970.968.787.21