Capgemini-海洋数据和人工智能用于物种保护（英）-2022.10-14正式版.doc

《Capgemini-海洋数据和人工智能用于物种保护（英）-2022.10-14正式版.doc》由会员分享，可在线阅读，更多相关《Capgemini-海洋数据和人工智能用于物种保护（英）-2022.10-14正式版.doc（14页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、Ocean Dataand AI forspeciesconservation1Abstract.The problem and at the same time ourmotivation is the loss of species diversity in theocean, which is often also referred to as ”invisibledying”. With our approach, we pursue the goal ofmaking this dying visible and thus preventable.To make events in

2、the ocean visible, we needto identify patterns depending on the oceandepth and then recognize deviations from thesepatterns. With the Norwegian institute Lofoten-Vesteralen, we are analyzing ocean data, to helpdetect anomalies. This should enable a betterunderstanding of the ocean ecosystem. Thissho

3、uld help to identify the consequences ofhuman intervention in nature, such as dwindlingfish stocks.OCEAN DATA AND AI FOR SPECIES CONSERVATION I OCTOBER 20222IntroductionAccurate observation of ecosystems enable detailed oceanographic research, allowing anomalies to be identified in enormous amounts

4、of data with the help of artificial intelligence (AI).The Lofoten-Vesteralen (LoVe) Ocean Observatory is located west of Hovden Vesteralen in the northern part of Norway. It is located in an ecological, geological, oceanographic and economic ”hotspot”. A network of submarine cables and seven sensor

5、nodes covers a cross-section from the mainland to the deep sea. It includes a land-based station and seven sensor platforms, covering a gradient from sea level to a depth of 200m. The system continuously provides valuable online data on the marine environment in northern Norway, and has been active

6、since 2013.The system is both, a national research infrastructure, basic and applied research, as well as a test infrastructure, where industry partners can test new underwater sensors and technologies. The Lofoten-Vesteralen Ocean Observatory has collected over 100 terabytes of sensor data (tempera

7、tures, currents, echograms) over the years.The teamThomas Rammis a Software Engineer at Capgemini. He created the initial infrastructure and GitHub integration.Majed Alaitwniis a software developer. The focus of his bachelor thesis was interactive visualization for anomaly detection in ocean measure

8、ment data. He created the visualization.Geir Pedersenis a researcher at the LoVeOcean Ocean Observatory and supports the project on the Norwegian side.Tom Hattonis a Data Scientist, his masters thesis explored the use of unsupervised AI models for anomaly detection in high-dimensional ocean measurem

9、ent data. He continues to develop the AI model.Sophie Baderis a molecular biologist specializing in oceanic ecosystems, as well as a software developer. She assisted with the infrastructure.parMustapha Mustaphais a software developer at Capgemini. He has done planning work and helped design the orig

10、inal AI model.Daniel Friedmannis a software developer at Capgemini. He is an expert in containerization and Docker and provides content support for the project.Nils Olav Handegaardis a researcher at the LoVe Ocean Observatory. His research focuses on the application of new methodologies and data pro

11、cessing techniques to the fields of marine ecology and fisheries oceanography.Eldar Sultanow is Enterprise Architect Director at Capgemini. His main focus is on modern software architectures, digitalization and enterprise architecture management. He developed the code with Thomas Ramm in the initial

12、 phase and is now supervisor of the project and oversees research work.OCEAN DATA AND AI FOR SPECIES CONSERVATION I OCTOBER 20223OceAIn was created as a team name to participate in Capgeminis Global Data Science challenge (GDSC) 2021. The goal of OceAIn was to develop an AI model that gains new insi

13、ght into seasonal correlating patterns in ecosystems using the time series data which was collected by ocean sensors. This should help to build better models and understand the climate of our planet.The AI model processes data of the cross section from the mainland to the deep sea. They are collecte

14、d by four different sensors that measure (1) directional pulsating sounds in specific areas with a scientific echo sounder, (2) a so-called hydrophone, i.e. an underwater microphone that records sounds in the environment, (3) an Acoustic Doppler (ADCP) that detects thestep and cleaned and transforme

15、d in the next. These steps will take place in Docker containers, with each type of data (hydrophone, biomass detection, etc.) having its own container in every step. The collaboration of the containers is defined inApache Airflow, which works on the ”Configuration As Code” principle. Airflow allows

16、the definition of infrastructure using DAGs (Directed Acyclic Graph). It is also worth mentioning that Docker containers are actually managed by Kubernetes, which in turn is defined by Airflow. Finally, the transformed data is persisted in the form of CSV files. They are processed by the AI to detec

17、t anomalies, which are displayed with the data in the form of image files through an interactive web interface.Implementation of the AI model OceAIn includes an AI model that detects anomalies in ocean data. The sensor data for the AI model are highly variable in the type of information, as well as

18、in their duration. In addition, some types of anomalies are only detectable if the data points are considered interconnected from the beginning.The initial idea was to focus on individual models that could handle different types of data and later combine the separate models. Due to the variety of da

19、ta, this approach showed some disadvantages, such as increased complexity in aggregating the individual models.speed and direction of ocean currents using the Doppler effect, and finally, (4) point sensors that provide real-time physical, biological and chemical observations.The Identification of re

20、peating seasonal patterns and anomalies allows scientists to better monitor the marine environment. This involves widespread exploration of the anomalies and their influencing factors and drawing conclusions from the bigger picture, such as differences in fish populations, varying current patterns,

21、or the influence of climate change.While large volumes of raw data are difficult to process manually and the results are highly error-prone in the process, AI models allow filtering of this data for relevant events. In addition, AI enables continuous analyzation of incoming data, resulting in a stre

22、am of data to the researchers.Architektur des Systems Even though the AI model is the core of OceAIn, there are other components that make up the platform. Our next goal is to make them work together according to the cloud-native architecture concept, which will create a future-proof and flexible da

23、ta pipeline. The raw data provided by the institute will be collected in oneTable 1. Results of the Baseline ModelsMODELF1 SCOREPRECISIONRECALLMACROAVG F1always true1.000.300.470.23always false0.000.000.000.41uniform random0.540.400.460.58stratified0.170.210.190.44OCEAN DATA AND AI FOR SPECIES CONSE

24、RVATION I OCTOBER 20224The idea that ultimately prevailed was to use a deep-learning neural network that analyzes all the data in their entirety for anomalies. At first, unsupervised models were supposed to be used, but this approach turned out not to be feasible.The use of unsupervised AI models ha

25、s the advantage that the training of those models does not rely on the presence of labeled training data (anomaly vs. normal).However, those models are extremely vulnerable to data noise and corruption 1. The underlying ocean measurement data are also subject to these characteristics, which are furt

26、her exacerbated by their whole-scale nature and the high data dimensionality that accompanies them.The contribution of labeled data by the researchers has enabled the use of supervised AI models. Basically, supervised AI models outperform their unsupervised counterparts in anomaly detection, as they

27、 are particularly capable of detecting application-specific anomalies 2.To check the performance of the unsupervised models later, four baseline models were created first, the results of which can be seen in 1. These are a common tool in the evaluationone-dimensional convolutions in each of the hidd

28、en layers, where the amount of filters is determined by a hyperparameter. The second architecture uses so-called LSTM layers and the third is a fully connected autoencoder.Table 2. Results of all reconstruction modelsRECONSTRUCTIONCLEAN DATAFULL DATAdenselstmconvlstmconvdense W32lstm W32 HL100dense

29、W16lstm W32 HL100conv W16HL400 4040 20 4HL100 40 20 440 20 4HL1000 100 10dense W32lstm W32 HL100dense W32lstm W32 HL100conv W16HL400 40 v240 20 4 v2HL100 40 20 440 20 4 v2HL1000 400 1004dense W8 HL400dense W32lstm W32 HL40conv W1640 20 4HL2000 1000lstm W32 HL400HL100 40 20 4100 40 20 4100 10conv W32

30、dense W32lstm W32 HL64HL100 40 20 4HL400 4032 16conv W64dense W32lstm W8 HL100HL100 40 20 4HL400 40 v340 20 4of machine learning models and are naive solutions for a classification problem.By comparing the results of a real model with those of the baseline models, conclusions about the correctness o

31、f the real models can be made. Of the baseline models used here, one always classifies false, one always classifies true, one splits the datasets in half between true and false, and the last stratifies the data based on their labels.During the model development, two types of models were created. The

32、se are reconstructionbased models and predictive models. For the former, three different microarchitectures were created. The first usespredictive models also fall into three categories. These are also LSTM and fully connected hidden layers, as well as an architecture that uses convolution and max-p

33、ooling. The naming of the models follows a fixed pattern. First, it is specified which architectural scheme a model corresponds with. Then a ”W” and a number is given, which represents the window size. ”W32” thus describes a neural network with a window size of 32. This is followed by further number

34、s, which indicate the size of the hidden layers. Optionally, there is also a version number at the end, which indicates models that had delivered promising results in the first instance, which is why they are run several times. Since the models described above do not in themselves detect anomalies,

35、there is anothercomponent that is responsible for exactly that. The results of these models were sobering. Hardly any of the models could exceed an F1 score of 0.5, which means that they were no better than the baseline models with random division of the values. In fact, there was only one model, ”l

36、stm W32 HL128128 128” which could (minimally) exceed this limit. All tested models can be seen in the tables 2 and 3, while 4 shows the average results of the different types of models.There are several reasons for the comparatively poor results of unsupervised models. The data itself is rather unsu

37、itable for an unsupervised model. Gaps, noise, and data corruption greatly degrade the results of unsupervised models.OCEAN DATA AND AI FOR SPECIES CONSERVATION I OCTOBER 20225Table 3. Results of all prediction modelsRECONSTRUCTIONCLEAN DATAFULL DATAdenselstmconvlstmconvdense W32lstm W32 HL128conv W

38、32lstm W32 HL128conv W32 HL32HL256*5 v2128 128HL128 128 128128 12832 32conv W32HL128 128 128 v2conv W32 HL3232 32Table 4. Average results of all model typesMODELQUANTILEF1 SCOREPRECISIONRECALLMACRO AVG F1reconstruction dense clean data0.730.430.340.730.41reconstruction lstm clean dat0.850.410.410.67

39、0.4reconstruction conv full data0.850.460.380.680.5reconstruction dense full data0.850.280.320.390.46reconstruction lstm full data0.850.40.330.620.46forecast conv clean data0.850.470.320.90.35forecast dense clean data0.730.480.320.950.33forecast lstm clean data0.730.480.320.980.3forecast lstm full d

40、ata0.730.470.310.940.31forecast conv full data0.850.470.320.950.33OCEAN DATA AND AI FOR SPECIES CONSERVATION I OCTOBER 20226In addition, the datasets are highly dimensional, with a large number of sensors collecting information simultaneously. The more complex a dataset is, the harder it is to train

41、 AI models to produce correct results. Presumably, our models would produce far better results if the data was less complex. Another issue is the limited computational capacity we have. The compression ratio of the models is high because adding more layers would have required significantly more comp

42、utation time, which was not feasible. Lastly, it should be mentioned that extensive hyperparameter optimization has not taken place yet. This could mean that the models themselves are actually better than assumed.For all these reasons, it can be concluded that unsupervised models are unsuitable for

43、anomaly detection in ocean data as it is generated by LoVe. Thus, OceAIn shall use a model that learns in a supervised manner. This model will later be continuously retrained based on data generated by the researchers during operation.A snippet of code for the model is shown in listing 1. Here, the

44、datasets that deviate significantly from the expected normal are detected. This is done by first iterating over all the datasets, while all datapoints that are detected as deviations are stored in the array anomalousdataindices.Listing 1. Code snippet of the anomaly detection module12345678910111213

45、1415# Detect all the samples which are anomalies. anomalies = test_mae_loss thresholdprint(Number of anomaly samples: , np.sum(anomalies)print(Indices of anomaly samples: , np.where(anomalies)plt.plot(x_test0)plt.plot(x_test_pred0, alpha=0.7)plt.show()anomalous_data_indices = for data_idx in range(T

46、IME_STEPS - 1, len(X_test) - TIME_STEPS + 1):if np.all(anomaliesdata_idx - TIME_STEPS + 1 : data_idx):anomalous_data_indices.append(data_idx)anomalous_data_indicesOCEAN DATA AND AI FOR SPECIES CONSERVATION I OCTOBER 20227Visualization of the vast amounts of dataThe visualization of the results generated by the AI models and the associated raw data is done in the form of a web applic