Application of Knowledge-Based System for Grading Meat

1. Field of the Invention

Assessing meat yield and quality in livestock carcasses is important for exporting meat and for the domestic consumer. It is important at all stages of the meat marketing chain from farm to retail sale. Since the meat grading system was first put into use in the United States in 1927, meat has been graded by human graders. The method used to grade meat is very subjective and, because of this subjective nature, it is very difficult (if not impossible) to achieve consistency and equity. The development of instruments to assist the human grader in evaluating grade factors has been attempted without much success.

This invention relates to a knowledge-based system for grading carcass meat.

2. Description of the Prior Art

In carcass beef grading, the carcasses are graded according to their degree or relative development of the desirable physical characteristics. The grade of the carcass is to reflect the differences in the portion of the desirable to less desirable parts of the carcass or cut, to reflect the ratio of meat to bone, and also to evaluate the characteristics of the meat which are associated with its ultimate palatibility. The USDA beef grading standards are designed to measure the quality (palatability) and the yield (cuttability) of the product as presented in carcass form. The quality grades are Prime, Choice, Select (formerly Good), Standard, Commercial, Utility, Cutter, and Canner, in order of decreasing palatability. The yield grades are Grades 1 through 5 , with Grade 1 having the highest yield of retail cut and Grade 5 the lowest yield of retail cut.

In present practice, meat graders reason about the final yield and quality grade of the carcasses using judgmental rules and empirical associations along with visual observation of the conditions of the carcasses. There are usually no algorithmic solutions except a formula for computing the preliminary yield grade of the carcass as given in the Meat Evaluation Handbook (National Live Stock and Meat Board, 1983). Since the grading of meat is accomplished by qualitative reasoning techniques rather than by mathematical or data processing procedures, this task represents an ideal application of a knowledge-based expert system.

Current developments in artificial intelligence programming, particularly in the expert systems field, make it possible to devise computer programs to assist humans in rendering decisions to complex judgemental type problems. As the name suggests, an expert system solves complex reasoning tasks that normally require an expert. A computer program in an artificial intelligence programming language deals with symbolic, nonalgorithmic methods of problem solving. These problem solving techniques relate symbols or objects through judgmental rules, or heuristics, as well as through theoretical laws and definitions. Expert systems programming uses artificial intelligence programming methods to solve problems conventional programming techniques cannot tackle.

A rule-based expert system, which is sometimes called an IF-THEN or situation-action system, consists of a knowledge-base and an inference engine. The knowledge-base contains information, consisting of facts and rules, about the specific area of interest and activity that the expert system deals with. Rules describe the logical relationships among elements of information within this area. The inference engine is the program that processes the rules and information and makes the inferences to arrive at specific conclusions.

Using advanced computer technology, we have now developed an applications system to assist meat graders in deciding the yield and quality grades of carcasses. The applications system comprises: (1) means for inputting information relating to characteristics of a carcass to be graded; (2) a program comprising a knowledge base of facts and rules relating to grading of the carcasses; (3) an inference engine for processing the input information against the facts and rules of the knowledge-base to arrive at yield and quality grades for the carcass; and (4) a means for outputting the carcass grades.

In accordance with this discovery, it is an object of the invention to provide an applications system for assisting meat graders in deciding the yield and quality grades of meat carcasses.

Another object of the invention is to provide a system for objectively and reliably determining the yield and quality of meat on a carcass based upon input information relating to carcass characteristics.

It is also an object of the invention to relieve human carcass graders of mental fatigue associated with assigning grades to carcasses over a prolonged period of time.

A further object of the invention is to employ a computerized grading system which is amenable to rapid information input and processing.

Another object of the invention is to provide a system for training inexperienced human meat graders, and for verification or testing of experienced human meat graders.

Other objects and advantages of the invention will be readily apparent from the ensuing description.

FIG. 1 is a block diagram of a preferred embodiment of the applications system of the invention.

The expression "applications system" is used herein in reference to the total carcass grading system of the invention. The expression "knowledge-based subsystem" refers to a component of the applications system. This subsystem is a rule-based program comprising a knowledge base of the facts and rules needed to generate both yield and quality grades based upon the input information relating to carcass characteristics and the inference engine for processing the input information in relation to the knowledge base. The term "fact-list" is descriptive of that portion of the knowledge base allocated to storing the facts about carcass meat and the grading of carcass meat. The expression "production rule" is used to describe any program statement or sequence of program statements using an "IF, THEN" syllogism.

In a preferred embodiment of the invention illustrated in FIG. 1, the application system 1 comprises a knowledge-based subsystem 10 of the type described above together with a voice recognition subsystem 3. The voice recognition subsystem 3 is employed for the purpose of inputting information into the knowledge-based system 10. The voice recognition system may be any device as known in the art which is capable of digitizing the human voice and emulating the keyboard input. In most commercially-available systems, the voice input is accomplished by means of a microphone or similar device for converting the voice into an analog electrical signal. The signal is thereafter subjected to accoustic processing, to include signal conditioning and analog-to-digital conversion. The resultant digital signal is passed through decoder processing whereby input sounds or sound combinations become associated with an output string. Most of the currently available voice recognition systems must be "trained" to recognize the user's voice by construction of a library or vocabulary file of each word or phrase to be used in the task. Voice patterns of these training words are preferably stored in the system's memory and saved on disk for repeated use. The words and phrases in the vocabulary file are selected to match those defining the parameters used in the production rules preprogrammed into the knowledge-based subsystem described below. For the grading of beef, for example, descriptive words are taken from the Official United States Standards for Grades of Carcass Beef (USDA-FSQS, 1980).

It is advantageous to divide the vocabulary into logical groups or menus of related characteristics. One reason is that the speed of the matching process is increased if only part of the total vocabulary is active and there are fewer words for comparison with the input patterns. Secondly, similar sounding words that are unrelated can be put into different groups, reducing the risk of an incorrect path. Words and phrases that are frequently used may be kept active at all times; otherwise, they are kept in a submenu and are activated only when necessary.

Upon comparison of the digitized voice input with the vocabulary file, recognized output strings are converted to simulated keystrokes by a keyboard emulator or the like. These simulated keystrokes are analogous to input made by keyboard entry and are processed in the knowledge-based system. The input information includes carcass characteristics and measurements as observed by the grader. These data are used by the knowledge-based system to determine both the quality and the yield grades.

The knowledge-based system (subsystem) 10 depicted in FIG. 1 comprises as its essential components a working memory 11 (short-term memory); the knowledge base 12 (long-term memory) comprising a set of validation rules 13, fact-list 14, and a set of rules for grading meat 15; and an inference engine and control strategy 16. The validation rules are production rules for checking the accuracy of input data against facts in the fact-list. The rules for grading meat are production rules used to generate the quality grade 20, the yield grade 21, an optional review and explanation of the grade determination 22, and permanent output records of the grade determinations to printer and disk, 23. The role of each of these components will become apparent in the ensuing Example. The reasoning process employed by the system is of course a function of the particular program. Some programs employ a forward-chaining process, others backward-chaining, and others a combination of the two. After grading is complete, the results are output to a screen, printer, hard disk, or other suitable device.

The Example, below, describes an implementation of the invention for the grading of beef carcasses using a knowledge-based system developed in the `C` Language Integrated Production System (CLIPS) environment.

A system for grading carcass beef was developed by integrating a voice input subsystem with a knowledge-based subsystem. Both subsystems were installed in, and operated by, a "COMPAQ DESKPRO 386" computer with two megabytes of memory and a 40-megabyte hard disk.

The VoiceScribe-1000 Speech Recognition System manufactured by Cherry Corporation of Waukegan, IL, and developed by Dragon System, Inc. of Newton, MA, was used. This is a discrete utterance, speaker dependent, recognition system. Table I lists the vocabulary of the voice input subsystem. This vocabulary matches the parameters used in the production rules of the knowledge-based subsystem. They are descriptive words and phrases taken from the Standards (supra) and in the Meat Evaluation Handbook (supra).

Since the voice recognition subsystem is speaker dependent, a grader must "train" the system to his or her voice for each word or phrase in the vocabulary. The voice patterns of these training words are saved on computer hard disk for routine usage.

After all of the patterns are formed, recognition can be activated. Once in recognition mode, a pattern is formed for the incoming utterance and is compared to all of the reference patterns in the active vocabulary. The word or phrase that matches the incoming pattern with the highest correlation score is considered the "winning" word or phrase. If the score exceeds the reject threshold, the speech driver outputs the results to the computer as if they were typed in through the keyboard.

The prompts in the program which solicit a response by voice input from the grader are as follows:

Enter, the following for quality grade

carcass-id (ID code)

marbling-is? (abundant, moderate, modest, small, traces, etc.)

maturity-is? (A B C D E)

firmness-is? (firm, slightly-firm, mod-firm, soft, etc.)

color-is? (light-red, mod-light-red, cherry-red, etc.)

texture-is? (very-fine, fine, mod-fine, slightly-coarse, etc.)

Enter, the following for calculating the yield grade

carcass-weight-is (numerical value)

back-fat-is (numerical value)

rib-eye-area-is (numerical value)

kph-fat-is (numerical value)

Optionally, enter, the following to input preliminary yield grade, then optionally adjust up or down

yield-grade-is (numerical value)

or back-fat-is (numerical value)

adjust-by (numerical value)

Enter, the command . . . finished

A complete printout of the operating instructions for the meat grading production system is set forth in Appendix A, below.

CLIPS, which is created by the Artificial Intelligence Section of the Mission Planning and Analysis Division at NASA/Johnson Space Center, is a development and delivery expert system tool. It is currently in widespread public and private use and is available from Cosmic, 382 East Broad Street, Athens, GA 30602. CLIPS provides a complete environment for the construction of rule-based expert systems. Its inference engine uses forward-chaining problem-solving strategy, based on the Rete algorithm. It is written in `C` language for portability, high speed, and high performance. Knowledge-bases can be compiled into CLIPS to produce turnkey executable programs.

CLIPS is a data-driven program where facts are the data that stimulate execution. Thus, the production rules can read any input, at any time. Also when the program reads input, the input is asserted into the fact-list, some of the production rules are processed, and all the matched facts are put into a conflict set, immediately; thereby permitting rapid processing. Most of the facts needed to arrive at a grade are coded into the working memory; and additional knowledge is asserted into the memory only when it is needed.

CLIPS also provides a utility called Cross Reference, Style, and Verification (CRVS). This utility comprises basic software development aids such as cross-referencing of relations and style checking and the ability to aid in the process of verifying the knowledge-base, which greatly improves the efficiency of the program.

Although the production system computations depend on the data inputs, some controls are needed for efficient operation. This is accomplished by organizing all the rules into different frames such as quality-frame, maturity-frame, and yield-frame. The frame is evoked only when it is needed for obtaining the desired goal or subgoal, by asserting the frame. The frame is retracted when the goal or subgoal is obtained. Some facts are asserted into the fact-list by the production rule when they are needed. These operations are orchestrated by the inference engine control.

For the present application of grading beef carcasses, the knowledge base in the CLIPS program is equipped with the vocabulary of Table I together with two sets of facts in the fact-list. Table II, below, is a first set of facts installed in the program, specifying values for classes and grades of carcass beef as well as values for the carcass characteristic parameters (degree of marbling, degree of firmness, and maturity) used in arriving at the quality grade. Table III, below, is a second set of facts relating to carcass beef grade parameters and their possible marbling and maturity values. The facts in these Tables are used for both verifying the voice input and in changing the grade up or down one or more steps. They are asserted when a reset is performed. For example, the fact that high-choice beef can be moderate marbling and A or B maturity is asserted into the fact-list and the fact that values for marbling can be abundant, moderate abundant, slightly abundant, moderate, modest, slight, traces, and practically devoid is also asserted into the fact-list.

The CLIPS program employs 61 production rules including the validation rules and the grading rules as set forth in the program printout attached as Appendix B. One rule is written so that the input of the carcass identification number is asserted to be the first fact. This identification number subsequently becomes the file name for storing carcass data and resulting grades.

The voice-input or keyboard input data are checked for their validity by the validation rules. Most parameter inputs are also checked. The invalid inputs are removed, while the valid inputs are converted to multifield variables. The production rule performing this task is fired first because of its high salience value. When the last input, `finished,` is given, the quality frame for evaluating quality of the carcass is evoked.

The quality grade of beef carcass is conventionally determined by the degree of marbling and firmness as observed on the cut surface of the rib-eye, in relation to the maturity of the carcass. In the present system, the meat grader inputs the degree of marbling on the exposed rib-eye. If the firmness of the lean tissue is not entered, the knowledge-based system assumes that it is comparably developed with the degree of marbling, and the quality will be determined on maturity and marbling alone. This is done by matching the marbling and maturity with knowledge listed in the quality-parameter fact-list (Table III) and arriving at a preliminary quality grade.

Maturity can be arrived at in three different ways. Firstly, it can be directly input with a letter degree such as A, B, C, D, or E. Secondly, if maturity is not known, then skeletal and/or lean characteristics, such as conditions of the chine bones, vertebrae, rib bones, and color and texture of the meat are input. This approach is taken when the grader cannot decide on a letter stage of maturity for the carcass, usually because the stage of maturity is on the borderline between two stages. At this point, the maturity frame is evoked and the facts of possible skeletal characteristics are asserted and checked against the inputs of the skeletal characteristics by the meat grader. The maturity is then determined by the conditions of the bones, with emphasis being placed on the conditions of the chine bone, followed by the thoracic vertebrae, lumbar vertebrae, and sacral vertebrae. However, if the conditions of lean maturity are also input, then meat color and meat texture are also used to determine the maturity of the carcass. The latter parameters are used mostly to decide different percents of the same stage of maturity. Thirdly, if no letter maturity or skeletal characteristics are entered, a default value of `A` is assigned for the stage of maturity.

For maturity groups besides `A,` an increase in degree of marbling is required to compensate for the progressive increase in maturity in each grade. To accommodate this feature, the grader should include a numeric value for the degree of marbling and stage of maturity inputs to the knowledge-based system. For example, an input of "B 80" indicates that the degree of maturity is 80% of the way between an "A 100" or "B 0" and a "B 100" or "C 0" carcass. If no percent is in the input, the system assumes the percent to be 100. If a skeletal feature with a specific condition is input, then a fact list, pertaining to that feature is asserted into working memory. This fact list contains the feature, its possible physical conditions, and the stage of maturity that relate to that condition.

Once the stage of maturity has been determined and the degree of marbling has been input, the quality frame is reasserted and the quality grade is determined by rules which correspond to FIG. 1, page 11 of the Standards, supra, herein incorporated by reference. When the check-firmness frame is evoked, rules are used to check if firmness complies with maturity and marbling conditions. If not, a rule is fired to decrease the preliminary quality grade. This is done by asserting the facts for the necessary conditions for each grade and comparing with the input firmness parameter.

Rules are provided for downgrading the quality grade if the carcass class is either from cow or is dark cutting beef; and the quality grade is eliminated if the carcass class is bull, because bulls are not eligible for a quality grade.

When the quality grade is determined, the yield frame is evoked by asserting the fact of yield-frame. The yield grade of the carcass is determined based on the values of the carcass weight; the backfat thickness on the exposed rib-eye; the percent kidney, pelvic, and heart (KPH) fat; and the area of the rib-eye. A direct method of computation is employed if all four of these facts are input by the meat grader. The results are obtained according to the formula described in the Meat Evaluation Handbook (supra). If the rib-eye area is not known, a rib-eye area is computed by its body weight. If the KPH fat is not given, a default value of 3.5% is assigned. For the indirect method of computing the yield grade, the meat grader may elect to give either a preliminary yield grade or else a backfat thickness from which the program computes the preliminary yield grade. In either case, the preliminary yield grade can be further adjusted up or down by the meat grader with numerical inputs or subjective estimates of the rib-eye area and KPH fat.

The system described above was used to grade two carcasses. The input data and the grade analyses are shown in Table IV. This system completed processing the final yield and quality grades in real time, without taking any measurable time right after the last input (`finished`) was given by the meat grader. Thus, the limiting factor for the speed of processing was the amount of time required to vocally input the values and characteristics of a carcass.

With this system, a trained carcass characteristics inspector could perform the task of a certified meat grader. The system would relieve the person responsible for grading from the intensive mental work in assigning the grades to the carcasses. Since the computer is doing the reasoning, the process would not deteriorate after long hours of continuous work. It is expected that the time required to grade a carcass will be greatly reduced as the grader acquires experience and becomes more familiar with the system.

Since the rules are based on the specifications of the official meat grading rules, the resulting grades should always be equally or more consistent than those assigned by human meat graders. This will result in more equitable grades. However, the consistency of the final yield or quality grades is dependent on the consistency of the meat characteristics observed by the meat graders. The rules are easily changed or upgraded to implement new standards. Also, the meat characteristics and yield and quality grades are permanently recorded and processes of the decision making can be easily reviewed. This system not only can be used on the industrial livestock kill floor, but is also very useful as a training tool for the future meat graders.