Patent Application
20210089947
(Not Applicable)
(Not Applicable)
The present invention relates to aquaculture which is the cultivation and farming of aquatic animals and plants in marine or freshwater environments in natural or constructed facilities on land, in coastal regions, or in open sea deep water areas.
The continued growth and productivity of the aquaculture industry demands that more automation and data-driven equipment be used in all forms and facilities of the farming of aquatic animal and plant species. The capacity and capability of conventional manual and computer techniques for collection, analysis, and management of data available in current and future aquaculture is being exceeded by the size and increasing growth rate of the raw data being collected. Currently there is no platform or consolidated system that integrates wide varieties of data, that applies machine learning technology to automate and increase the productivity of the integration and analysis of the data, and that generates profiles and action plans for operations management decisions action plans for improving throughput, return on investment, and output.
The global aquiculture industry, which is the farming of aquatic plants and animals in off-shore waters, on-shore ponds, and indoor facilities, has grown dramatically since 2000 and now represents more than 50% of the seafood produced globally. Along with this growth in production, there has been an accelerating growth in the amount of research and operational data being produced by research organizations worldwide, by corporations serving the industry, and by aquaculture farm operators who are deploying larger and more automated systems. Despite this growth in data, there has been limited success in optimizing the use and value of the data for the industry and for researchers even though several published research indicates that the productivity and sustainability of an aquaculture farm operation can be improved significantly (10% to 60%) with changes in feed, fertilizer, or operational procedures. As a result, there is a growing need for a practical but innovative method for converting this data into useful knowledge.
To meet this need and solve this problem, a decision support system has been created by the inventors that uses machine learning algorithms to process large and sometimes sparse research and operating data sets to provide aquaculture farm operators and their control equipment with the best decisions about the diets and process control parameters for the specific species being farmed and for the specific aquaculture farm facility.
The growth of the aquaculture industry has been due in great part to the necessity of satisfying a growing demand for seafood per capita globally as well as the leveling off and decline of the production of wild capture seafood. The global ocean fisheries have been fished to near exhaustion. As a result, for economic and ecological reasons, the aquatic farming industry which has been in a state of small but consistent growth for decades has accelerated. While there has been acceleration in the growth of aquaculture farms globally, the pace of innovation in this industry has not moved forward as fast. As a result, much of the growth in aquatic farm facilities and the aquatic feed industry that supports them has happened without the major breakthroughs that analogously accompanied the green revolution in the global agriculture industry.
Research by Mote Marine Laboratory and other organizations (Ref. 1 to 86) has shown that the productivity of an aquatic farm can vary significantly as a function of the constituency of the feed and fertilizer, the process variables of the aquatic farm facilities, the handling of the aquatic species between growth phases, the species of plant and animals, and other environmental factors. Research projects around the world continue to show that improvements over the conventional feeds and process controls are possible and compelling. However, there is no clear way for the implications of these findings to be effectively deployed to the thousands of aquatic farm operators around the world.
The purpose of this invention is to provide facility-specific and species-specific information about how to control key feed and fertilizers parameters and aquatic farm environmental operating parameters in order to achieve maximum productivity for a specific aquatic farm facility. Conventional approaches to productivity improvement for the aquatic farming industry are based on the dissemination of general best practices in aquatic farming operations from aquatic farm industry suppliers, academic institutions, and government agencies. The present invention involves the application of machine learning technology to fine tune and optimize farming practices for a specific aquatic species, aquatic farm facility, and geography by combining general industry best practices with data collected from each specific aquatic farm operation.
Optimized aquatic farming operations includes the feeding process and the growing environment process.
Best feeding recipes in this case means ensuring that the constituency of the feed or fertilizer being given the aquatic species at each daily stage of their growth process will use the most ecologically sound, economically balanced, and organically productive combination of feed ingredients. Most ecologically sound means that there will be a minimum of fish meal used as ingredients and that alternative sources of nutrients from plant, insect, and other animal sources will be used. Most economically balanced means that the lowest cost combination of ingredients will be used to reduce the cost for the aquatic farm operators. Most organically productive means that the aquatic species will grow larger and faster than they would with other feed or fertilizer ingredient combinations.
Best operational care for the aquatic species being farmed means in this case that the parameters important for growth are known and optimized. Such parameters include the temperature, alkalinity, salinity, and contamination of the water, the density of the aquatic species, the transport conditions when the aquatic species are moved from one station to the next, and the physiological health of the aquatic species.
Aquatic farms are becoming more automated and have many electronic devices that control the environmental and farming process machinery. The decision recommendations from the present invention can be used directly as inputs to these industrial controls or indirectly to the human operators in charge of setting and monitoring the automation equipment.
Conventional techniques used by aquatic farm operators is to use information provided to them by the suppliers of feed or fertilizers, by the makers of the aquatic farm equipment, by researchers who publish their findings of improvements, and by the records kept by the aquatic farm operators of their own operations successes or failures. There is no effective or convenient method or tool for combining all these sources of information or to learn from successes or failures of different combinations of parameters.
The present invention is an innovation and improvement over existing methods because it uses machine learning computational techniques and algorithms to process all data sets available to each aquatic farmer from all commercial, public domain, and privately collected sources. The training of the learning algorithms is a combination of supervised and unsupervised learning methods depending on the source and quality of the raw data sets. The present invention will provide the individual aquatic farmer with a decision tool that grows in its intelligence by optimizing the usefulness of all data sets available to the aquatic farmer from external sources and by learning from internal data sets such as the proprietary data that is unique to a specific aquatic farm facility.
The net benefit of the use of the invention by aquatic farm operators is sustainable productivity. The aquatic farm using the invention will be more productive by producing more aquatic animal or plant product (by weight) using feed or fertilizer ingredients and aquatic farm machinery that cost the same or less than conventional ingredients and machinery. The aquatic farm will be more ecologically sustainable because (1) it will be using feed and fertilizer that consist of alternative less on fish meal based nutrients and more on alternative sources of natural nutrients, (2) it will be using aquatic farm machinery more efficiently, and (3) it will be creating a diminishing amount of waste and harmful byproducts from its operation.
One or more specific embodiments of the present disclosure are described below. When introducing elements of various embodiments of the present disclosure, the articles “a,” “an,” “the,” and “said” are intended to mean that there are one or more of the elements. The terms “comprising,” “including,” and “having” are intended to be inclusive and mean that there may be additional elements other than the listed elements. Any examples of operating parameters and/or environmental conditions are not exclusive of other parameters and/or conditions of the disclosed embodiments.
The embodiments described herein relate to a computer implemented method for optimizing decisions for operating an aquaculture system with the multiple goals of optimizing operations which includes maximizing the growth rate of, maximizing the quality of, and minimizing the cost of the aquatic species crop for a specific aquaculture farm. The computer implemented method consists of a digital learning engine which learns how to optimize operations from data supplied by suppliers and researchers and collected from prior operations. After learning how to optimize operations from the data, the digital learning engine generates decision rules and management action plans for use by the humans and machines that control the aquatic farming facility.
An embodiment of basic operational steps in a typical aquaculture farm is shown in
The hatch or germinate step 221 in
The grow seedling step 231 in
The grow adults step 241 in
The final step 250 in
The block diagram in
Process steps data 620 includes steps that comprise the preparation of the feed or fertilizer recipe. These steps include the actions of mix 621, cook 622, package 624, store 625, and dispense 626. There may be other 623 steps as well.
The algorithms in the Regression 1220 digital library can be chosen from a variety of sources. Regression 1220 algorithms are designed to calculate coefficients for a polynomial that produces a best fit between the polynomial equation and many sets of data. This best fit polynomial then becomes the new or updated model for a Plan which is a set of Rules for how to grow a specific species in a specific facility. The calculations and simulations used to determine the best fit model is the training process for the new or updated Plan or set of Rules.
The algorithms in the Classification 1230 digital library can be chosen from a variety of sources. Classification 1230 algorithms are designed to split data into categories which have labels that have been discovered or predefined by human experts. There are a variety of classification algorithms which use different types of equations to determine best fit within a classification.
The mathematical approaches that can be used in Supervised 1210 algorithms for both Regression 1220 and Classification 1230 applications include Least Squares 1221, Bayesian 1222, Neural Nets 1223, Random Forests 1224, and Support Vectors 1225. Least Squares 1221 algorithms compute the coefficients for a polynomial that makes the distance between data points and the polynomial as small as possible. In Least Squares 1221 algorithms, there are no assumptions about what causes the differences between the data sets and the polynomial models. In Bayesian 1222 algorithms, assumptions are included that the causes of the differences between the data sets and the polynomial models are statistical in nature. The typical assumptions in Bayesian 1222 models include that the distribution is normal and that the mean and variance are known. In Neural Nets 1223 algorithms, regression or classification polynomial calculations are organized as a parallel processing problem by assigning and modifying the weights or coefficients of the polynomial terms they flow through one or more hidden layers of parallel states. In Random Forest 1224 algorithms, data sets are randomly selected, used to create several different decision trees often by different human experts, and then statistically merged or averaged together to produce a set of coefficients for matching polynomials or categories. In Support Vectors 1225 machines, the approach to classifying sets of data is to calculate a polynomial model surface that separates the categories of data best rather than calculating a polynomial surface that fits the data within a category best. The coefficients of the polynomial that describes the separating plane can be represented as a vector in matrix algebra.
The Unsupervised 1270 digital library of algorithms includes Clustering 1280 algorithms and Association 1290 algorithms. Unsupervised 1270 algorithms are called unsupervised because an assumption is made that there is no set of labels or categories predefined by human experts that can be used to supervise, guide, or set the starting point for the machine learning calculations. Unsupervised machine learning algorithms are sometimes called data mining algorithms because the algorithms are mining or searching for some type classification or labels from raw data.
Clustering 1280 machine learning algorithms include the use of mathematical techniques for grouping a set of data in such a way that data in the same group (called a cluster) are more similar (in some calculable sense) to each other than to data in other groups (clusters). Because the clustering approach is unsupervised, it usually requires several iterations of analysis until consistently clear categorizations and groupings can be identified from the data sets being analyzed.
The Clustering 1280 digital library includes the K-means 867 algorithm. The K-means 1281 calculates the average distance between the centroid of K clusters in a dataset. At the start of the analysis, a number is chosen for K. Every data point is allocated to each of the K clusters through reducing the in-cluster sum of squares difference from each of the centroids. This process is iterative and takes several steps to correct each centroid location and minimize the sum of squares of the distances from the data points in each cluster to the centroid. Then a lower value of K and a higher value of K can be chosen to see if either of those numbers of clusters produces a lower mean or tighter fit. The iterations end when a value of K is found which produces the lowest sum of squares difference.
Association 1290 machine learning algorithms include the use of correlation calculations to identify important relationships between categories or clusters of items in a data set. Relationships discovered by association machine learning algorithms can be used to generate new labels or categories for additional machine learning algorithm calculations.
Apriori 1291 is a digital library of algorithms that search for a series of frequent sets of relationship in datasets. For example, assume that a data set has five categories identified such as A, B, C, D, and E and that an association algorithm has identified a relationship between category A and B (e.g. if a data set has data in category A, 50% of the time it has data in a category B). An Apriori algorithm might find that if a data set has data in categories A and B, it has data in Category C 80% of the time.
Because it is not always possible to have data sets that can be analyzed with Supervised 1210 algorithms and because it is sometimes expensive and difficult to use only Unsupervised 1270 algorithms, an approach which speeds up the analysis process is to use a Semi-supervised 1240 approach to using machine learning algorithms. The Semi-supervised 1240 learning approach combines a small amount of data that can be used in a Supervised 1210 approach to a large amount of data that can be used in an Unsupervised 1270 approach. Markov 12151 algorithms, which are based on assumptions about the statistical randomness of the data being analyzed, can then be applied to complete the training calculations of the Semi-supervised approach.
The plan generator 830 includes software tools that makes plan changes 1410 that are either new or updated plans based on new or updated rules from rule generator 820 for facility-specific and species-specific recipe plans 1420, feed/fertilize plans 1430, and conditions control plans 1440. recipe control rules 1320, feed/fertilize control rules 1330, and conditions control rules 1340. Each time new data from the master knowledge base 450 is analyzed by learn 810, new or updated plans are added to master knowledge base 450.
This application is a nonprovisional application for a utility patent which claims priority from and the benefit of U.S. Provisional Application Ser. No. 62/926,081, entitled “Aquaculture Decision Optimization System Using A Learning Engine,” filed Oct. 25, 2019. Each of the foregoing applications is hereby incorporated by reference.
