doi:10.1016/j.compchemeng.2008.10.006 L J a b a A R R 1 A A K H C O P 1 i m a p i n b T i r t w h o m o V e p t t i 0 d Computers and Chemical Engineering 33 (2009) 371–378 Contents lists available at ScienceDirect Computers and Chemical Engineering j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / c o m p c h e m e n g earning HAZOP expert system by case-based reasoning and ontology insong Zhao a,∗, Lin Cui b, Lihua Zhao b, Tong Qiu a, Bingzhen Chen a Department of Chemical Engineering, Tsinghua University, Beijing, 100084, China College of Information Science and Technology, Beijing University of Chemical Technology, Beijing, 100029, China r t i c l e i n f o rticle history: eceived 1 November 2007 eceived in revised form 3 September 2008 a b s t r a c t To improve the learning capability of HAZOP expert systems, a new learning HAZOP expert system called PetroHAZOP has been developed based on the integration of case-based reasoning (CBR) and ontology that can help automate “non-routine” HAZOP analysis. PetroHAZOP consists of four modules including case base module, CBR engine module, knowledge maintenance module and user graphical interface ccepted 4 October 2008 vailable online 22 October 2008 eywords: AZOP ase-based reasoning module. Within the case base, HAZOP analysis knowledge is represented as cases which are organized with a hierarchical structure. Similarity-based case retrieval algorithm is also depicted to find the closest- matching cases. In order to enhance the case retrieval, a new set of ontologies for CBR-based HAZOP analysis is created by integration of existing ontologies reported in literature. Finally the application of PetroHAZOP is demonstrated by two case studies of industrial processes. t “ a w l t a H m e u s p H t t i s r ntology rocess safety . Introduction Safety is an important issue in process design and operation n the chemical process industry (CPI). It is even more critical for odern chemical manufacturing processes which are often oper- ted under extreme conditions to achieve maximum economic rofit, or have to undergo changes of customer demands. The mportance of safety analysis in process operation is well recog- ized after occurrence of several tragic accidents that could have een avoided if adequate process safety analysis had been done. o ensure safe operation, process hazard analysis (PHA) is very mportant to proactively identify the potential safety problems and ecommend feasible mitigation actions. Among the available PHA echniques, hazard and operability (HAZOP) analysis is the most idely used one in the CPI. HAZOP analysis done by human teams, owever, has the following shortcomings: time consuming, labori- us, expensive and inconsistent. To solve these problems, various odel and/or rule-based HAZOP expert systems have been devel- ped during the last decade, which was respectively reviewed by enkatasubramanian, Zhao, and Viswanathan (2000) and McCoy t al. (1999). These systems, however, can only address “routine” or rocess-generic HAZOP analysis. “Routine” HAZOP analysis means hat its reasoning logic can be applied to different processes while he “non-routine” HAZOP analysis means that its reasoning logic s process specific or plant specific. Generally analysis of devia- ∗ Corresponding author. Tel.: +86 10 62783109. E-mail address: jinsongzhao@tsinghua.edu.cn (J. Zhao). i w c v a p a f 098-1354/$ – see front matter © 2008 Elsevier Ltd. All rights reserved. oi:10.1016/j.compchemeng.2008.10.006 © 2008 Elsevier Ltd. All rights reserved. ions generated by using guidewords “other than”, “as well as” and part of” are “non-routine”. As a result, these kinds of deviations re hardly addressed in literature about HAZOP expert systems. In the CPI, “routine” HAZOP analysis roughly occupies 60–80% hile “non-routine” HAZOP analysis occupies 20–40%. Due to the ack of self-learning capability in existing HAZOP experts systems, he knowledge of “non-routine analysis” can be hardly formulized nd reused for similar chemical processes, and the “non-routine” AZOP analysis still needs to be addressed by human experts. To evaluate the output quality of the signed directed graph (SDG) odel based HAZOP expert system HAZID developed by McCoy t al. (1999), five industrial plant systems which had not been sed during the model development stage were selected as a test et (McCoy, Wakeman, Larkin, Chung, & Rushton, 2000). The out- ut of HAZID was compared against the results of conventional AZOP study which was done by human teams. Table 1 shows he test results which are quite interesting. According to Table 1, he percentage of scenarios identified by conventional HAZOP also dentified by HAZID ranged from 33% to 60%, and the percentage of cenarios identified by HAZID which were judged to be corrected anged from 33% to 83%. Moreover, the percentage of scenarios dentified by HAZID which were judged to be correct and of interest as much lower, ranging from 9.5% to 29%. In other words, severe ompleteness and correctness issues existed in HAZID. The unfa- orable performance of HAZID was attributed to a few factors such s the quality of the unit model, lack of fluid property data, the lant complexity and human judgment variance. In fact, there is nother more important factor that was unrecognized and that actor is knowledge representation limitation. The “non-routine” http://www.sciencedirect.com/science/journal/00981354 http://www.elsevier.com/locate/compchemeng mailto:jinsongzhao@tsinghua.edu.cn dx.doi.org/10.1016/j.compchemeng.2008.10.006 bahamut 下划线 bahamut 下划线 372 J. Zhao et al. / Computers and Chemical Engineering 33 (2009) 371–378 Table 1 Summary analysis of trial results of HAZID (McCoy et al., 2000). Plant systems Absorber system �-trichloroethane system Propane rectification system Benzene storage Separation system Scenarios identified by conventional HAZOP also identified by HAZID (%) 36 33 60 50 53 Scenarios identified by HAZID which were judged to be correct (%) 49 33 69 83 53 Scenarios identified by HAZID which were judged to be correct and 9.5 29 24 27 N/A P H T k c e d 2 a 1 t s & i m t o e b m a I e t w c 5 2 t H t n t “ a H P H s r t f b a p F t w i a P r c k p t time, effort and money involved in performing HAZOP, there exists considerable incentive to develop intelligent systems for automat- ing the process hazards analysis of chemical process plants. An intelligent system can reduce the time, effort and expense involved of interest (%) rotections identified by HAZID which were judged to be correct (%) 9.5 29 AZOP analysis could not be represented by the existing models. herefore, it is clear that there is much room for improvement in nowledge representation for HAZOP expert systems. It is worth noting that consistence and completeness are criti- al in HAZOP analysis because neglect of any potential hazard may ven result in disasters. Investigation results of past industrial acci- ents, e.g. the tragic BP Texas city plant accident occurred in March 005, have proved that poor quality of PHA is a major root cause of ccidents occurred in the CPI. Recently case-based reasoning (CBR) technology (Kolodner, 993; Aha, 1998) has been integrated into HAZOP automation echnology by researchers at Purdue University to enhance the elf-learning capability of HAZOP expert systems (Zhao, Bhushan, Venkatasubramanian, 2005). However, the case-based reason- ng they proposed aimed to facilitate modification of the existing odels and creation of new models based on the knowledge in he existing models. The “non-routine” HAZOP analysis still relied n the human team. To solve this problem, a new learning HAZOP xpert system called PetroHAZOP has been developed in this paper ased on the integration of CBR and ontology that can help auto- ate both “routine” and “non-routine” HAZOP analysis. This article is organized as follows. In Section 2, HAZOP analysis nd related work on the automation of HAZOP are briefly described. n Section 3, the integrated methodology for HAZOP analysis is xplained. Section 4 contains two industrial application examples hat illustrate how the proposed HAZOP expert system can help ith improvement of “routine” and “non-routine” analysis. Finally, ontributions of this work are summarized and discussed in Section . . HAZOP HAZOP was firstly introduced by ICI (Imperial Chemical Indus- ries, UK) for identifying hazards in chemical plants in 1960s. AZOP study is accomplished by a HAZOP team through a collec- ive brainstorming effort that stimulates creativity and brings about ew ideas of the potential hazards including their cause–effect rela- ionships. Generally the chemical process is divided into sessions called analysis nodes” before study. Then meaningful deviations in every nalysis nodes are generated by combining process parameters and AZOP guidewords including MORE OF, LESS OF, NONE, REVERSE, ART OF, AS WELL AS and OTHER THAN. For each deviation, the AZOP team has to identify all of its credible causes and all of pos- ible adverse consequences. Once the causes and consequences are ecorded, the team has to list the existing safeguards for the iden- ified hazards and give necessary recommendations accordingly or hazard mitigation if the required risk level cannot be achieved N/A 77 N/A y the safeguards. The process is repeated deviation by deviation nd node by node until the analysis of the whole process is com- leted. The conventional HAZOP study procedure is presented in ig. 1. To complete the HAZOP analysis of a typical chemical process, it akes about 1–8 weeks for a HAZOP team with 4–8 members. It is idely accepted that HAZOP analysis is an extremely time consum- ng and effort consuming process. An estimated including direct nd indirect costs $5 billion is spent annually by the CPI to perform HAs and related activities. The estimated cost of process hazards eviews is about 1% of sales or about 10% of profits for a big chemical ompany. Moreover, the quality of HAZOP analysis depends on the nowledge and experience of the HAZOP team. Therefore, incom- leteness and inconsistence usually are the drawbacks with regards o HAZOP done by human teams. Given the enormous amounts of Fig. 1. HAZOP study procedure. bahamut 下划线 mical i s o a H “ l a m a 3 o 3 t s “ a r t a s I r p R r a t b s m u I v b r F a f M b v 3 c d d d a r t a f i s f a g J. Zhao et al. / Computers and Che n a HAZOP, make the analysis more thorough, detailed, and con- istent, minimize human errors, and free the team to concentrate n the more complex aspects of the analysis which are unique nd difficult to automate (Venkatasubramanian et al., 2000). The AZOP analysis that is difficult to automate generally refers to non-routine” analysis discussed in the above section. In what fol- ows, case-based reasoning and ontology are integrated for the utomation of HAZOP analysis that was considered difficult to auto- ate before by using the traditional model-based or rule-based pproaches. . HAZOP expert system by the integration of CBR and ntology .1. Case-based reasoning (CBR) Experts often find it easier to relate stories about past cases han to formulate rules. Similarly it is true in the HAZOP analy- is domain that rules or models are hard to construct to automate non-routine” analysis. To overcome this problem, an important rtificial intelligence technique – CBR is adopted to augment the easoning machines embedded in the existing HAZOP expert sys- ems. CBR is both a pattern for computer-aided problem solvers and model of human cognition. The central idea is that the problem olver reuses the solution from past cases to solve a new problem. n this way, valuable experiences that are difficult to formulate into ules or models could be utilized for solving new problems. In a CBR system, the problem solving process includes four hases (Aamodt and Plaza, 1994), namely 4R’s: Retrieve, Reuse, evise and Retain as shown in Fig. 2. Basically, knowledge and expe- ience are stored in the form of cases in a case base. The content of case is made up of three parts: the problem/situation description, he solution, and the outcome. The outcome is not needed but could e added to suggest solutions that work and use cases with failed olutions to warn of potential failures. When a new problem is sub- itted, CBR system retrieves the similar cases from the case base i a e o Fig. 2. Problem solving Engineering 33 (2009) 371–378 373 sing certain similarity algorithm based on the predefined indexes. ndexes should be abstract enough to retrieve a relevant case in a ariety of future situations, and also should be concrete enough to e easily recognizable in future situations. Then the solutions of etrieved cases are adapted to solve the new problem if necessarily. inally, the new problem description and its solutions are retained s a new case in the case base for future use. Most of the CBR applications do not go through all the above our phases (López-Arévalo, Bańares-Alcántara, Aldea, Rodríguez- artínez, & Jiménez, 2007). In this paper, the adaptation is done y the users because it is highly domain dependent and requires erification of the solution performance. .2. Ontology Human experts are indispensable in HAZOP analysis of any hemical processes even though various expert systems can be esigned to facilitate the process. Different experts especially from ifferent organizations have different jargons with regards to the escriptions of the analysis objects and results including causes nd consequences of hazards. That is to say, there is no standard to epresent the HAZOP analysis domain information. This increases he difficulty of CBR for different users. To settle the terminological nd conceptual incompatibility problem, a new set of ontologies or CBR-based HAZOP analysis (CHA) is created in this paper by ntegration of existing ontologies reported in literatures. “Ontology” is a term of philosophy originally, which refers to the ubject of existence. Artificial intelligence (AI) borrows the old term rom philosophy and gives it new wonderful meanings. In AI, there re a number of definitions of ontology. However, the definition iven by Gruber is accepted by majority of researchers: an ontology s an explicit specification of a conceptualization (Gruber, 1993). HAZOP analysis of chemical process needs knowledge from reas such as chemistry, chemical process engineering, safety ngineering, electrical engineering and so on. A large-scale ontol- gy OntoCape was constructed by the research group of Professor phases with CBR. bahamut 下划线 bahamut 下划线 bahamut 下划线 bahamut 下划线 bahamut 下划线 bahamut 下划线 bahamut 下划线 bahamut 下划线 374 J. Zhao et al. / Computers and Chemical Engineering 33 (2009) 371–378 ow of M M t r ( s D o 2 1 ( t H u r c c s o p o f d s r d i c c d P t 3 a t a t s o l o a u e 3 i s o g Fig. 3. Workfl arquardt for chemical process engineering (Morbach, Yang, & arquardt, 2007). To facilitate the information sharing among he HAZOP analysis expert systems developed by the labo- atory of Professor Venkatasubramanian at Purdue University Venkatasubramanian et al., 2000), process simulation packages uch as Aspentech’s BatchPlus and documentation tool such as yadym’s PHAPro, operational related ontologies and safety related ntologies were created (Zhao, Bhushan, & Venkatasubramanian, 003). Batres et al. created an upper level ontology based on ISO 5926 that had already been used for knowledge queries in HAZOP Batres et al., 2007). Based on the major concepts and ideas from he above ontologies, six ontologies are created for case-based AZOP analysis. They respectively are process ontology, process nit ontology, unit operation ontology, equipment ontology, mate- ial ontology and HAZOP ontology. Within each of the ontologies, oncepts are organized in a hierarchy where concept nodes are onnected by is-a links. Synonyms are given for concepts whose ynonyms are available in the CPI. Process ontology is built based n the classification of chemical plant types such as hydrocracking lant, FCCU plant, ethylene plant, ammonia synthesis plant and so n. Process unit ontology and unit operation ontology are basically rom OntoCape. In equipment ontology, equipment is specified by esign properties such as design temperature, pressure and some tructural specifications. Since MSDS information of each mate- ial is indispensable in HAZOP analysis, material ontology not only epicts the chemical species information, but also contains MSDS nformation for each material. In HAZOP ontology, HAZOP related onceptions such as nodes, parameters, guidewords, deviations, auses, consequences, safeguards, recommendations and risk are escribed. After the ontologies were created, the ontology editor d g s d t PetroHAZOP. rotégé (Stanford Medical Informatics, 2006) was used for verifica- ion. .3. Integrated reasoning framework for HAZOP analysis As stated above, the existing HAZOP expert systems could only ddress “routine” and generic process analysis. Due to the lack of he ability of machine learning, they could not “remember” the nalysis, especially “non-routine” analysis that has been done so hat the HAZOP team has to do the analysis once again even if imilar HAZOP analysis scenarios have been discussed before. To vercome this problem, a HAZOP expert system PetroHAZOP with earning capability is built by integrating CBR and the above six ntologies. The HAZOP analysis process of the case-based expert system pproach is illustrated in Fig. 3. PetroHAZOP consists of four mod- les, as shown in Fig. 4. In what follows, the four modules will be xplained. .3.1. Case base module Construction of the case base to a large extent determines the ntelligence level of a CBR system. Each case instance generally con- ists of two parts: the problem and the solution. Inside the case base f PetroHAZOP, the problem part contains the HAZOP analysis back- round information of a particular deviation while the solution part escribes its abnormal causes, adverse consequences, risk, safe- uards, recommendations and some other auxiliary information uch as the name of HAZOP team members and HAZOP analysis ate. Stored in a relational database, each case holds a unique iden- ification number. It is not uncommon that there are hundreds or J. Zhao et al. / Computers and Chemical e i c s s a m s o t c fi s t t H f c t c s d a o p i m a c s i t a 3 p p T t t a t T a a s a d d c ( ( Fig. 4. Configuration of PetroHAZOP. ven thousands of deviations that need to be combed through dur- ng a HAZOP analysis of a typical chemical process. Therefore, the ase base will grow to a very large size with time. To facilitate the imilarity-based case retrieval that is described in the following ection, a hierarchical case structure is introduced in this paper s an indexing method to partition a huge number of cases into ultiple hierarchical subordinate case bases (SCB). HAZOP analy- is cases can be categorized into the first level SCBs by the types f the chemical processes specified in the process ontology while hose cases within a same type of chemical process can be further lassified into the second level SCBs by the equipment types speci- ed in the equipment ontology. Cases inside a second level SCB can till be divided into the third level SCBs according to the deviation ypes specified in the HAZOP ontology. This hierarchical case struc- ure can be regarded as a knowledge-based indexing method where AZOP domain-specific knowledge is applied, important features or quick and accurate retrieval of past cases (Barletta, 1991). Each case in the case base is defined by indexes of four major ategories: equipment with its design parameters, materials con- ained in the equipment, operating conditions, and stream context onditions. The equipment design parameters such as design pres- ure and design temperature describe the equipment where the eviation being analyzed occurs. The equipment type must be vailable in the equipment ontology. Each case contains a list f materials present in the equipment. Hazardousness related hysical–chemical characters of materials such as flash point, boil- ( Fig. 5. Similarity algorithms for diff Engineering 33 (2009) 371–378 375 ng point and toxicity are distilled from MSDS to represent the aterial characters. Operating conditions include parameters such s operating temperature, pressure and level. The stream context onditions reflect the equipment types of both up stream and down tream of the equipment. According to the standard IEC61882, the case solution which s the HAZOP analysis results should include information such as he deviation’s root causes, adverse consequences, the safeguards vailable in the P&IDs, the recommendations for hazard mitigation. .3.2. CBR engine module The CBR engine module (CEM) is the core of the indispensable hase of CBR systems, i.e. retrieval. When a new deviation analysis roblem is presented to the system, the CBR engine is activated. he engine starts from selecting the corresponding SCB that fits he problem through the hierarchical indexing mechanism. Within he chosen SCB, all past cases are compared with the new problem, nd scored based on the similarity-based case retrieval algorithm hat is described in the following to find the closest-matching cases. o define the similarity between the past case and new problem, measure is needed first to assess the closeness between the ttributes belonging to them. Basically there are five types of attributes for each case: object uch as equipment, string such as the material name, numeric such s operating temperature of equipment, interval-numeric such as esign parameters and set object such as materials. Fig. 5 shows ifferent similarity algorithms that are employed in this paper to alculate the similarities of different types of attributes. 1) The similarity of string attributes is simply calculated by string matching algorithm. If string attributes are same, then their similarity is 1, otherwise 0. 2) In a HAZOP case, all numeric attributes are transformed to non- negative values. For example, the temperature unit is Kelvin temperature while the absolute pressure is used to represent pressure attributes. The mathematics similarity formula for non-negative numeric attributes xi and yi is: sim(xi, yi) = 1 − ∣∣xi − yi ∣∣ max(xi, yi) (1) 3) Similarity between sets Usually, there is more than one material involved in a piece of process equipment. Comparison of HAZOP cases usually requires comparison of the material sets present in the equip- ments where the cases originate. We proposed a new approach erent types of case attributes. bahamut 下划线 bahamut 下划线 bahamut 下划线 bahamut 下划线 3 mical ( ( m o r l P s n T t m s w B s i e O Q e i e t 3 i k q t t H c c d b p “ a e j t f 3 t • • • • • • • • • h I e s g u S t fi done, the caption of button is changed to “Show similar cases”. PetroHAZOP also has a user menu for project management, public 76 J. Zhao et al. / Computers and Che (Set-Similarity method) to compute the material set similarity between different cases. The approach is expressed as follow. Assume there are two material sets A and B, which belong to two different cases, respectively. Set A contains m materials: MA1, MA2, . . ., MAm, and set B contains n materials: MB1, MB2, . . ., MBn. Then the similarity of A and B could be computed by Eq. (2). sim(A, B) = N∑ i=1 Si max(m, n) (2) where N = Min(m, n) Si is the maximum similarity between the ith material in one set and each material in the other material set, 1 ≤ i ≤ N. if N = m then, Si = n Max j=1 {sim(MAi, MBj )} (3) if N = n then Si = m Max j=1 {sim(MAj , MBi)} (4) In Eqs. (3) and (4), sim(MAi,MBj) represents the similarity between material MAi and material MBj, which can be computed by Eq. (5) sim(MAi, MBj ) = K∑ k=1 (Wksim(attAik, attBjk)) (5) where K represents the number of index attributes of a material, Wk represents the weight of the kth index attribute, 1 ≤ k ≤ K, attAik, attBik are respectively the kth numeric index attributes of materials MAi and MBj, and sim(attAik, attBik) can be calculated by Eq. (1). 4) The similarity algorithm of interval-numeric feature is extended-Euclidian algorithm. Suppose there are two interval- numeric attributes A = [a1,a2], B = [b1,b2], then their similarity can be calculated as follows: sim(A, B) = (a1b1 + a2b2) max((a1)2 + (a2)2, (b1)2 + (b2)2) (6) 5) Object similarity HAZOP cases have object attributes such as equipment and aterials. The object similarity calculation takes advantage of the ntologies described in Section 3.2. Ontology based similarity algo- ithms have been reported in literature. In this paper, the path ength measure is used to calculate the object similarity (Pedersen, akhomov, Patwardhan, & Chute, 2007). It essentially computes the imilarity between two object nodes by counting the numbers of odes on the shortest path between them in the ontology hierarchy. he shortest path includes both the object nodes. Mathematically, he similarity of two object nodes A and B using the path-length easure (path) is defined as: im(A, B) = 1 (7) p here p is the number of nodes on the shortest path between A and within an ontology hierarchy. For example, in the equipment ontology, if equipment A is a ubclass of equipment P (subclass is equivalent to a is-a relationship d v m C n Engineering 33 (2009) 371–378 n ontology), and equipment B is a subclass of equipment Q while quipment P and equipment Q are two subclasses of equipment . The shortest path from equipment A to equipment B is A-P-O- -B. There are five nodes on the path. Therefore, the similarity of quipment A and equipment B is 1/5 (see Fig. 5). Finally the case similarity is the sum of each case attribute sim- larity multiplied by its weight which is determined by domain xperts. The weights are adjustable through the knowledge main- enance module which is described below. .3.3. Knowledge maintenance module The effectiveness of a CBR system depends largely on the qual- tative and quantitative richness of its stored cases which are the nowledge repository of past experiences. That is to say, the more uality cases stored in the case base, the more effectively the sys- em reasons. Initially, there are few cases in the case base. Through he knowledge maintenance module (KMM), cases from the past AZOP analysis records can be manually input to the subordinate ase bases to facilitate CBR. Case retaining is another feature of KMM. Revised cases can be onverted to new cases and retained in the corresponding subor- inate case bases. Before a new case is stored, term translation ased on ontology is to be done if necessary. For example, “tem- erature too high” and “high temperature” will be translated into more temperature” in the case base. Since weight factors used in the similarity based case retrieval re predefined according to the knowledge of authors and a few xpert consultants, it will not be surprised that there is a need to ustify them according to the feedbacks of users in industrial prac- ices. A user interface is designed in KMM to modify the weight actors through a certain level of authorization. .3.4. Graphical user interface (GUI) module The GUI module contains graphical user interfaces to perform he following functions: Creating HAZOP analysis project Specifying equipment Specifying materials Selecting parameters and HAZOP guidewords Editing HAZOP analysis results Retrieving and reusing similar cases Retaining cases Reporting HAZOP analysis results Administrating users Fig. 6 is a snapshot of a main GUI of PetroHAZOP. On its left- and side is a treeview like the Microsoft’s Windows Explorer tree. t displays a hierarchical collection of labeled items such as nodes, quipments, process variables and deviations. On its right-hand ide is the area where causes, consequences, risk, safeguards, sug- estions and comments for a certain deviation can be edited by sers or automatically loaded from selected similar cases. Case earching button on the top-right corner of the snapshot implies hat the system is searching similar cases based on the speci- cations of the deviation being analyzed. Once the searching is ata entry and user administration. On the bottom of the tree- iew, there are two more nodes, i.e. materials and report. In the aterials node, the process materials can be specified by users. ustomizable HAZOP reports can be generated by using the Report ode. J. Zhao et al. / Computers and Chemical Engineering 33 (2009) 371–378 377 hical 4 t u t i w p p “ 4 i t u a a p p r r h r s e f o f m c o o a r t c t 4 Fig. 6. Snapshot of a main grap . Application examples PetroHAZOP is programmed with Java allowing concurrent mul- iple users to manipulate the system through intranet. This multiple ser mode greatly improves the work efficiency of the HAZOP eam. Recently, PetroHAZOP has been successfully installed and mplemented at one of the largest oil companies in China. There ere more than 900 cases in the case base at the time when this aper was written. The following examples demonstrate how the roposed HAZOP expert system PetroHAZOP can help with both routine” and “non-routine” analysis. .1. Example 1: “non-routine” analysis Acrylonitrile production process is a highly hazardous process n the petrochemical industry. Most of the 16 materials involved in he process including raw materials, intermediate products, prod- cts and by-products are flammable, toxic or/and volatile. There re about 14 major equipments such as reactor, chilling tower, bsorption column, recycle column, distillation columns. The whole rocess was divided into six nodes. Totally 87 deviations of 59 key arameters had been analyzed by a HAZOP team. We transfer the esults through the knowledge maintenance module into cases, esulting in 87 cases in the case base. H t s Fig. 7. Snapshot of case ‘Ignition failure user interface of PetroHAZOP. The vinyl chloride production process (VCPP) is another highly azardous chemical process which consists of three sections: chlo- ine/hydrogen processing section (CHPS), hydrochloride synthesis ection (HSS) and vinyl chloride synthesis section (VCSS). One xample of the ‘non-routine’ analysis of this process was ‘ignition ailure of the HCl synthesis furnace’ since it is hard to be modeled r generalized with rules. With PetroHAZOP, a similar case ‘ignition ailure of acrylonitrile startup furnace’ from the case base was auto- atically retrieved with the similarity degree of 0.602. The found ase is analogous in the equipment type in the equipment ontol- gy and the deviation type in the HAZOP ontology. If the user clicks n the button “Show similar cases”, he/she can find out the causes nd consequences of similar case (Fig. 7). This similar case can be eused if the button “Reuse” on Fig. 7 is pressed. Here reuse means hat the causes and consequences are automatically loaded to the auses and consequences of the new case. The user then can edit he causes and consequences if necessary. .2. Example 2: “routine” analysis Another example is the completeness checking even for routine AZOP analysis by means of CBR. A HAZOP team was assigned o HAZOP a polymer production process. Node 1 containing a tirred tank was analyzed first and it took the team about 2 days to of acrylonitrile startup furnace’. 3 mical c o w t r t s c T l S c s 5 p T a r t T t l c C c e E o i o t l a t t p a t A T a R A A B B C G K L M M M P S V Zhao, C., Bhushan, M., & Venkatasubramanian, V. (2003). Roles of ontology in 78 J. Zhao et al. / Computers and Che omplete its analysis. In the third day, the team started the analysis f Node 2 which also contained a stirred tank. When the team as about to close the analysis of the ‘Low Pressure’ deviation in he stirred tank of Node 2, the leader asked the recorder who was esponsible for manipulating PetroHAZOP to retrieve similar cases o check if anything was missed. A similar case ‘Low pressure in the tirred tank of Node 1’ was discovered in the case base and one of its auses, ‘axial sealing leak’, was not considered for the current case. he consequence of the cause was “oxygen entering the stirred tank eading to contamination and deactivity of the catalyst in the tank”. ince the cause and consequence could also happen in the current ase, they decided to reuse the found case, and minor modification uch as change of the catalyst name was made for the new case. . Conclusions HAZOP analysis requires high accuracy, consistency and com- leteness because any ignorance would lead to catastrophic losses. herefore, the HAZOP team must ensure that it would not lose ny resources that are available to help them meet the above equirements. As a solution, this article offers an integrated solu- ion for the complex problems in the path of automating HAZOP. he proposed HAZOP expert system PetroHAZOP not only facili- ate “routine” analysis but also “non-routine” analysis due to its earning capability by which the HAZOP analysis quality can be ontinuously improved during practice. Due to the adoption of BR mechanism, PetroHAZOP can map past experiences to the new ases. Therefore, it is more adaptive in HAZOP analysis and greatly ases the knowledge management and knowledge dissemination. ven though HAZOP has been widely practiced in the CPI of devel- ped countries, it just starts to get recognized by the practitioners n China. Lack of HAZOP experts hinders the wide implementation f HAZOP in the chemical plants. Hopefully, PetroHAZOP can facili- ate the industrial exercises of HAZOP in China and contribute to the oss prevention in the largest developing country where chemical ccidents are posing a serious threat to its fast development. Automatic adaptation for HAZOP analysis is an outstanding task hat could be addressed by introducing other artificial intelligence echnologies to CBR. Layered digraph model (LDG) has been pro- osed by authors to perform both “non-routine” and routine HAZOP nalysis (Cui, Zhao, Qiu, & Chen, 2008). Future work will be oriented o integrate CRB with LDG based reasoning. Z Engineering 33 (2009) 371–378 cknowledgements This work is supported by Program for New Century Excellent alents in University. Anonymous reviewers of this paper are highly ppreciated for their helpful comments and suggestions. eferences amodt, A., & Plaza, E. (1994). Case-based reasoning: Foundational issues. Methodological variations, and system approaches. Artificial Intelligence Com- munications, 7(1), 39–59. ha, D. W. (1998). The omnipresence of case-based reasoning in science and appli- cation. Knowledge-Based Systems, 11(5–6), 261–273. arletta, B. (1991). An introduction to case-based reasoning. AI Expert, 6(8), 42–49. atres, R., West, M., Leal, D., Price, D., Masaki, K., Shimada, Y., et al. (2007). An upper ontology based on ISO 15926. Computers & Chemical Engineering, 31(5–6), 519–534. ui, L., Zhao, J., Qiu, T., & Chen, B. (2008). Layered digraph model for HAZOP analysis of chemical processes. Process Safety Progress, 27(4), 293–305. ruber, T. R. (1993). A translation approach to portable ontology specifications. Knowledge Acquisition, 5(2), 199–221. olodner, J. (1993). Case-based reasoning [M]. San mateo, CA: Morgan Kaufmann Publishers, Inc. ópez-Arévalo, I., Bańares-Alcántara, R., Aldea, A., Rodríguez-Martínez, A., & Jiménez, L. (2007). Generation of process alternatives using abstract models and case- based reasoning. Computers & Chemical Engineering, 31(8), 902–918. cCoy, S. A., Wakeman, S. J., Larkin, F. D., Jefferson, M. L., Chung, P. W. H., Rushton, A. G., et al. (1999). HAZID, a computer aid for hazard identification 1. The STOPHAZ package and the HAZID code: An overview, the issues and the structure. Process Safety and Environmental Protection, 77(6), 317–327. cCoy, S. A., Wakeman, S. J., Larkin, M. L., Chung, P. W. H., & Rushton, A. G. (2000). HAZID, a computer aid for hazard identification: 4. Learning set, main study system, output quality and validation trials. Process Safety and Environmental Protection, 78(2), 91–119. orbach, J., Yang, A., & Marquardt, W. (2007). OntoCape—A large-scale ontology for chemical process engineering. Engineering Applications of Artificial Intelligence, 20(2), 147–161. edersen, T., Pakhomov, S. V. S., Patwardhan, S., & Chute, C. (2007). Measure of semantic similarity and relatedness in biomedical domain. Journal of Biomedical Informatics, 40(3), 288–299. tanford Medical Informatics. (2006). The protégé ontology editor and knowledge acquisition system. Available at http://protege.stanford.edu. enkatasubramanian, V., Zhao, J., & Viswanathan, S. (2000). Intelligent systems for HAZOP analysis of complex process plants. Computers & Chemical Engineering, 24(9–10), 2291–2302. automated process safety analysis. Computer Aided Chemical Engineering, 14, 341–346. hao, C., Bhushan, M., & Venkatasubramanian, V. (2005). PHASUITE: An automated HAZOP analysis tool for chemical processes Part I: Knowledge Engineering Framework. Process Safety and Environmental Protection, 83(B6), 509–532. http://protege.stanford.edu/ Learning HAZOP expert system by case-based reasoning and ontology Introduction HAZOP HAZOP expert system by the integration of CBR and ontology Case-based reasoning (CBR) Ontology Integrated reasoning framework for HAZOP analysis Case base module CBR engine module Knowledge maintenance module Graphical user interface (GUI) module Application examples Example 1: "non-routine" analysis Example 2: "routine" analysis Conclusions Acknowledgements References