Fire Alarm Systems - Essay Example

 Fire Alarm Systems Chapter 1 Introduction Of all the dangers at sea, fire is the most terrible. Unlike a building in a city where help can arrive swiftly from many immediate quarters, a ship's response is self-contained. Thus, early alert is vital in shipboard fire situations. Traditional ship fire alarm systems, however, fail to provide reliable, immediate and effective alert because they identify fire conditions only when the output of a single sensor reaches a preset value. In addition, conventional systems are prone to false alarms. Repeated false alarms have been known to cause seafarers to switch off the fire alarm system, thereby endangering the ship. Therefore, the main objectives of this study are to reduce the probability of false alarms and to shorten the time between fire onset and alarm response in shipboard fires, with applicability to other fire situations as well. Wang et al. (2000), presented a ship fire alarm system based on a fuzzy neural network (FNN), a multi-sensor detection algorithm and two fire parameters (temperature and smoke density). Utilization of more than one sensor in a fire alarm system provides additional information concerning fire condition and, if properly employed, results in fewer erroneous alarms. Our study follows Wang and uses dual sensors: a temperature-sensing K-type thermocouple and an analog photoelectric smoke sensor. A fire-detection system using a self-learning FNN was introduced by Chen and Zhao (2000, 242). In Chen's system, the fuzzy knowledge base and parameters were optimized by FNN on-line learning of real-time data, giving the system adaptive and self-learning capability. Additional examples of fuzzy rule-based classification systems can be found in (Setnes, 2000, 509), and clearly demonstrate that fuzzy theory can handle different situations, avoiding or dramatically reducing misinterpreted signals. The relevant research has been proposed by Khananian et al. (2001, 1861) and Neubauer (1997, 9). Grey theory was first proposed by Deng in 2002 (p. 289). Since then, the grey model (GM) has been successfully and widely used (Kuo, 2001, 55). GM is often applied to prediction in time-varying non-linear systems. Kuo and Wu (2001, 59) used GM to predict deformation of thin ship panels. Yuan et al. (2000, 13) proposed a method based on grey theory to predict gas-in-oil concentrations in an oil-filled transformer. This present paper improves upon the above studies by endowing a dual-sensor fuzzy detection system with GM predictive ability, thereby shortening alarm response time and increasing alarm accuracy. Fig. 1 presents a block diagram of the proposed system, showing a temperature sensor, a smoke sensor, individual GM circuits for each sensor and a fuzzy classification system. The adaptive fuzzy classification system is established by the use of various fire/non-fire data configurations (readily obtained from the Taiwan National Fire Administration Ministry) and from this builds a set of fuzzy rules. Next, two grey-prediction models are set up, one for each sensor. Each GM system interprets the ongoing changes in the dynamic behaviour of its respective sensor. Raw data trends, which indicate possible fire situations, i.e., rapid increase or continued rise in smoke or temperature slope, result in higher GM predictive output values than the equivalent raw sensor output values. Thus, incorporation of GM between sensor and computer anticipates future sensor values, allowing the system to make fire alarm response before actual alarm conditions, without increasing susceptibility to false alarms. Fig. 1. Structure of the grey-fuzzy fire-detection system. Chapter 2 Adaptive Fuzzy Rule-Based Classification System The adaptive fuzzy rule-based classification algorithm in this study is modified from Nozaki, (1996, 238) and Ishibuchi, (1999, 1040). The modified algorithm is used to classify situations as fire or non-fire and consists of three procedures: (1) an automatic procedure for generation of fuzzy rules; (2) a classification procedure; (3) a fuzzy rule self-learning procedure. First, the fuzzy system is established by inputting data for various fire conditions (i.e. hot smokeless fire, smoky cool fire, etc.) and generating a fuzzy rule base. After that, the same data but in different order are tested in the adaptive fuzzy classification system to determine classification accuracy. During the classification procedure, the system employs a self-learning mechanism which adjusts the fuzzy rules according to a reward–punishment principle. The three procedures are introduced in detail in the following. 2.1. Automatic Fuzzy Rule Generation The basic structure of the proposed adaptive fuzzy system consists of two inputs and one output. The two inputs derive from analog-to-digital conversion of data from temperature and smoke sensors. These input data are linearly correlated with the environmental conditions being monitored. The single output is an alarm-trigger signal having two digital values, fire or non-fire. Thus, the classification situation is a two-class (fire/non-fire) problem in the two-dimensional pattern space [0,1]×[0,1]. Hence, the system judges fire or non-fire situations depending on temperature and smoke density. We assume that m patterns xp=(xp1,xp2), p=1,2,...,m, are given as training patterns from the two classes. Namely, each pattern xp belongs to either class. In each pattern xp=(xp1,xp2), xp1 and xp2 represent attribute values of different pattern spaces. The present fuzzy classification system employs fuzzy if–then rules for two-class pattern classification. The fuzzy rules are illustrated as follows: Rule Rij: If xp1 is Ai and xp2 is Aj, then xp belongsto Cij with GoC=GoCij, i=1,2,…,K, j=1,2,…,K, where K is the number of fuzzy subsets in each pattern space. The fuzzy rules show that the two pattern spaces have the same number of fuzzy subsets. Ai and Aj are the fuzzy subsets of the antecedent parts of the rule Rij in the unit interval [0,1]. In the consequent part, Cij represents one attribute of the two classes and GoCij is the grade of certainty of the fuzzy rule Rij, which represents the possibility of fire/non-fire. In the fuzzy rule antecedents, symmetric triangle-shaped membership functions for Ai and Aj are used. The pattern space is evenly partitioned into K fuzzy subsets A1,A2,…,AK, as shown in Fig. 2: (1) (2) Where, Fig. 2. Two-class classification and membership function of fuzzy subsets. Derived from Eq. (1) and (2), the membership functions of Ai and Aj are μi(xp1) and μj(xp2), respectively. The fuzzy rule consequents Cij and GoCij of fuzzy rule Rij can be determined by the following calculations: Step 1: Calculate the compatibility of each training pattern xp=(xp1,xp2) with fuzzy rules Rij via the following product operation: μij(xp)=μi(xp1)μj(xp2), p=1,2,…,m. (3) Step 2: Calculate the sum of the compatibility of the m training patterns with the fuzzy rules Rij for each class: (4) where βClass,h(Rij) is the sum of the compatibility of the training patterns in class h with the fuzzy rules Rij. Step 3: Find Cij that has the maximum value in βClass,h(Rij): (5) Cij is thus determined by class in Eq. (5). Step 4: Determine the grade of certainty GoCij: (6) Step 5: Repeat K×K calculations from steps 1 to 4 to determine the consequent Cij and GoCij of each fuzzy rule. Thus, the fuzzy rule base is generated automatically based on m training patterns. Such a procedure can save trial-and-error time when adjusting each fuzzy rule. 2.2. Classification Procedure In this procedure, a rule set is established to form a fuzzy rule-based classification system. When a rule set S is given, a test pattern xp=(xp1,xp2) is classified by a single winner rule Rij in S, which can be obtained by the following calculations: Step 1: Calculate γClass,T for T=fire, non-fire as (7) Step 2: Use the winner rule to find class X in which input pattern xp belongs: γClass,X=max{γClass,FIRE,γClass,NON-FIRE}. (8) Thus, the correct classification by fuzzy inference is obtained if the fuzzy rule has the maximum value μi(xp1)μj(xp2)GoCij. 2.3. Fuzzy Rule Self-Learning Procedure To provide a fuzzy classification system with on-line and self-learning ability, the grade of certainty (GoC) of each fuzzy rule is adjusted during the classification process based on the reward–punishment learning principle. When the selected fuzzy rule is able to correctly classify the pattern xp, the grade of certainty is increased: GoCijnew=GoCijold+δ1 GoCijold. (9) In Eq. (9) δ1 is a positive learning factor for increasing GoC and δ1 GoCijold is the reward for the correct classification. On the contrary, if the selected fuzzy rule fails to correctly classify the pattern xp, GoC is reduced: GoCijnew=GoCijold+δ2 GoCijold. (10) As shown in Eq. (10), δ2 is a negative learning factor for decreasing GoC and δ2 GoCijold is the punishment for misclassification. In our approach, a fuzzy rule base becomes more adaptive and flexible for different types of classification after implementing the above procedures. 2.4 Grey Temperature/Smoke Density Predictors The dynamic state of temperature and smoke density are important parameters in a fire-detection system. During monitoring, changes in temperature and smoke density arise from convection and diffusion, conveying crucial messages to the fire-detection system. Conventional fire alarms cannot anticipate fire conditions because they monitor a parametric value at a momentary point in time, with no reference to the history of the value. The momentary value is compared to a preset trigger value. If the trigger value is exceeded, the alarm response is triggered. The momentary logic of conventional systems is unable to detect temperature and smoke dynamic trends. It has been shown that monitoring the dynamic state of temperature and smoke density can improve alarm response speed and accuracy. Raw data regarding temperature and smoke density from sensors are sent to an adaptive fuzzy rule-based classification system which has been trained by experience with the general type of data and uses this experience to improve decision appropriateness when deciding whether or not to trigger an alarm. However, conventional fuzzy systems, like conventional fire alarm logic, use only momentary data, with no immediate sense of data trends. This lack of immediate sense of trend precludes a predictive sense which can anticipate the attainment of alarm-trigger conditions from, for example, an adequately rapid rise of temperature in an adequately brief time period. Humans have an innate sense of prediction from immediately prior data trends, and have attempted to model this sense mathematically. Many statistical methods have been devised to make predictions such as regression analysis and square difference analysis combined with prediction. However, conventional statistical methods typically require a large volume of sample data. This requires a long time to collect and analyse, making this method unsuitable for real-time classification systems. Further, some statistical methods require regular samples, e.g. exponential, linear, or logarithmic distributions. If the samples are distributed at random, the prediction values may become unstable and make the system predict inaccurately. To avoid the limitations of traditional statistical methods, this paper explores grey theory for prediction of temperature and smoke density. A grey prediction extends from past information to the future, i.e. its analysis of past output trends is used to predict future output. In this paper, two grey-prediction models, the GM(1,1) unified-dimensional new message model and the GM(1,1) new message model, are employed and compared, as shown in Fig. 3. The “(1,1)” in the term GM(1,1) indicates a GM with one variable in a first-order grey differential equation. Each sensor's data is treated separately by its own individual but identical GM(1,1) equation. Through accumulated generating operation (AGO) and inverse accumulated generating operation (IAGO), the GM have the capability of predicting upcoming temperature and smoke density. The construction of the predictors is introduced in the following. Fig. 3. Construction diagram of the grey temperature predictor. First, two rows Eq. ((11) and (12)) of temperature (T(0)) and smoke density (S(0)) data are collected from the sensors mounted in the shipboard engine room and used as inputs in each prediction model: T(0)=(T(0)(1),T(0)(2),T(0)(3),…,T(0)(n)), (11) S(0)=(S(0)(1),S(0)(2),S(0)(3),…,S(0)(n)), (12) where the superscript (0) denotes the original detected data. During the predictive process in the GM(1,1) model, the number of data points in the temperature and smoke density rows should be determined first. For example, if the number of data points is 6, there will be six elements in the temperature and smoke density rows. That is, n is 6 in Eq. (11) and (12). Through 1-AGO, the new accumulated temperature row (13) and smoke density row (14) are obtained: (13) (14) where T(1)(1)=T(0)(1), S(1)(1)=S(0)(1), and superscript (1) denotes the accumulated row. Hence, the first-order ordinary differential equation of the GM(1,1) needs to be established as (15) (16) According to Eq. (15) and (16), we assume the parameter matrix is â. The relationships between parameter matrix â, accumulated matrix B and constant matrix yN are shown as follows: â=[a,u]T, (17) â=(BTB)−1BTyN, (18) (19) (20) yNT=[T(0)(2),T(0)(3),…,T(0)(n)], (21) yNS=[S(0)(2),S(0)(3),…,S(0)(n)]. (22) The tendency of this time sequence can be approximated by an exponential function whose dynamic behaviour is similar to that of a first-order differential equation. An exponential Eq. (23) of accumulated temperature is inferred from Eq. (15), (17), (18), (19) and (21), and an exponential Eq. (24) of accumulated smoke density is inferred from Eq. (16), (17), (18), (20) and (22): (23) (24) Substituting k=0,…,n−1 in Eq. (23) and (24), the new accumulated temperature row (25) and smoke density row (26) are obtained: (25) (26) The temperature row and the smoke density row are changed into the non-accumulated temperature row and smoke density row after IAGO calculation, signified as αT(1) and αS(1). Superscripts (0) and (1) denote the number of IAGO calculations. Non-accumulated temperature is deduced from accumulated temperature as given in Eq. (29). Likewise, the relationship between accumulated and non-accumulated smoke density is given as Eq. (30): (27) (28) (29) (30) Finally, non-accumulated temperature row (31) and smoke density row (32) are obtained via Eq. (29) and (30): (31) (32) In this paper, predictive values for temperature and smoke density are generated by two different kinds of grey-prediction models, the GM(1,1) new message model and the GM(1,1) unified-dimensional new message model. The major difference in these two models is that GM(1,1) unified-dimensional new message model eliminates the first element, or , while the other model does not. Thus, the GM(1,1) new message model accumulates errors that the GM(1,1) unified-dimensional new message model removes. The prediction temperature and prediction smoke density are added to the temperature row and smoke density row, respectively. A new temperature row and smoke density row are then formed as the following equations: (33) (34) As for the GM (1,1) unified-dimensional new message model, the first and in Eq. (33) and (34) are eliminated, yielding the following equations: (35) (36) A residual checking method is used to compare the predicted data with the actual data after the predicted data are derived from GM(1,1) models. The error rows of the two models are referred to as eT(k) and eS(k), shown in Eq. (37) and (38). The model with the smaller residual ratios will be chosen to combine with the fuzzy rule-based classification to function as an optimal fire-detection system: (37) (38) Chapter 3 Experimental Results Experimental evaluation of the proposed system involved: (1) hardware implementation (dual sensors: K-type thermocouple/temperature and analog photoelectric/smoke); (2) recording of sensor data from two types of controlled shipboard engine-room fires (open flame and smoldering); and (3) software implementation. Final system evaluation was based on inputting the recorded shipboard fire data to a PC containing software embodying several of the best fuzzy-grey configuration alternatives and also the threshold values for the sensors when operated in their commercial “dumb” mode. The results of this allow comparison of the response behaviour of the various fuzzy-grey options, and also comparison with the response of the dual-sensor (photo/heat) pair as conventionally operated. Software implementation involved fuzzy rule generation and then selection of the single best fuzzy rule base. For fuzzy rule generation, three different kinds of fuzzy rule bases were established, the antecedent and consequent parts being determined via 1000 fire/non-fire training samples. Next, to select the best fuzzy rule base, the same set of 1000 training samples but in different order was input to each rule base to establish optimal values for positive learning factor δ1 and negative learning factor δ2, giving the GoC of each fuzzy rule in each of the three bases. The combined results for each fuzzy rule base were then compared to determine the optimal fuzzy rule base. 3.1. Establishing three fuzzy rule bases In establishing a fuzzy rule base, its pattern space is first partitioned. The pattern space can be partitioned by different numbers of fuzzy subsets. In our approach, K=3, 5 and 7 are used as the fuzzy partition numbers for the three rule bases. Thus, the first rule base has a partition number of 3, the second a partition number 5 and the third a partition number of 7. Approximately 25 temperature/smoke density curves for a mixed collection of open-flame and smoldering fires were obtained from the Taiwan National Fire Administration Ministry. Smoke density for these samples and in the remainder of this paper is reported as percent obscuration per meter using the following equation [12]: (39) where I is the intensity of the transmitted light under test conditions, I0 is the intensity of the transmitted light under normal ambient conditions, and d is the distance between the light source and the receiving instrument. A total of 1000 momentary temperature/smoke density points were selected from these curves. The fire/non-fire state of these points was determined by reference to temperature and smoke density alarm-trigger values as used in a common commercial alarm (temp=70°C, SMOKE=15%). The fire/non-fire state of each two-value point was added to the point, making a three-value training sample, each training sample thus containing temp data, smoke density and fire/non-fire state. The training samples are fed to each of the three fuzzy rule bases to determine the fire(1) or non-fire(0) attribute Cij and the GoC in the consequent part of each fuzzy rule. Table 1, Table 2 and Table 3 show the training results for partition number K=3, 5 and 7, where SS (smaller small), S (small), SM (smaller medium), M (medium), ML (medium large), L (large), and LL (larger large) represent the different fuzzy subsets. As can be seen, if the tested values of temperature and smoke density are smaller, then the attribute Cij is non-fire(0) and GoC is 1; in other words, there is no chance of fire under this data combination. When the sample values of temperature and smoke density become larger, the attribute Cij varies from non-fire(0) to fire(1) and GoC is larger. In Table 3, when the temperature is L (large) and smoke density is L (large), the attribute Cij is fire(1) and GoC is 96%. After training, the three established rule tables properly correlate the variations of temperature and smoke density sensor data with the probability of a real fire. Table 1. 3×3 fuzzy rule base Temp: temperature; smoke: smoke density. Table 2. 5×5 fuzzy rule base Temp: temperature; smoke: smoke density. Table 3. 7×7 fuzzy rule base Temp: temperature; smoke: smoke density. A fuzzy classification system is built around each of the three fuzzy rule bases described above, yielding three fuzzy classification systems. The one built around the rule base with K=3 (K is the partition number) is thus equipped with 3×3 rules to classify these 1000 test samples. The classification system built around the K=5 rule base has 5×5 rules, while the K=7 rule base has 7×7 rules. 3.2. Determining Positive/Negative Learning Factors and the Best Rule Base After establishing the three fuzzy rule bases, we need to determine positive learning factor δ1 and negative learning factor δ2 Eq. ((9) and (10)). The absolute value of δ can range from 0 to 1. A pair of likely values (based on our experience) is selected and applied to the three classification systems. The 1000 training samples were randomized as to order and, in this form, given the designation “test samples”. The test samples were then input to the three fuzzy classification systems with a given δ-pair. This process yields the data of Table 4. Results of Table 4 allow easy comparison of the correct-detection percentages obtained by the δ-pair in the three fire classifications. The highest correct-detection percentage tells us the best δ-pair and fuzzy classification system. Table 4. Correct classification ratios for different learning factors and fuzzy subsets Table 4 shows δ1 and δ2 for a range 0.001–0.15. When the fuzzy classification system equipped with the 3×3 rule base is used to classify the 1000 test samples, the correct-detection percentage varies from 78.5% to 96.9%. When 5×5 rule base is applied, the correct-detection percentage varies from 81.3% to 99.1%. In the 7×7 rule base case, the correct-detection percentage varies from 88.1% to 99.7%. Thus, classifying the 1000 fire/non-fire test samples with the 7×7 (49 fuzzy rules) base with δ1=0.01 and δ2=−0.13 yields the highest correct-detection percentage, i.e. 99.7%. 3.3. Testing the Algorithm in Two Fire Situations Following Wang et al. (2000), a dual sensor system is employed, one sensor for smoke and the other for temperature. One was the analog photoelectric smoke detector. The other was the K-type thermocouple. These two sensors were mounted on the ceiling of the engine room of a 25 m fibreglass coastal fishing trawler. The engine room was 14 m2, 3 m tall. Sensors were mounted 15 cm below the flat ceiling. Both sensors were analog output. Raw output was converted to digital format via an FX2N-8AD programmable logic controller (Mitsubishi, Japan) and then input to an IBM-compatible PC via the RS232 input. Sampling data were 1 Hz. The optimal fuzzy classification system and the two GM are implemented on the PC. The same computer also contained the sensor trigger values for the commercial alarm mentioned above (temp=70°C, SMOKE=15%), allowing comparison of the results of the proposed and commercial systems to the same set of input data. The above hardware was used to record 600 s of data each for two different types of fire, an “open fire” and a “smoldering fire”. Burning was performed at floor level in a 1 m×1 m×10 cm stainless steel tray. The same amount of wood of the same type was used for both fires. Commercial wax-based fire starters of the same type and number were used for both fires, and continued to burn throughout the test. The smoldering condition was generated first spraying the wood with water from the city water supply and continuously spraying more water on the wood as it dried and started to burn normally. For shipboard safety considerations, the ship's engine was not working during the experiments. The basic temperature and smoke data as recorded in our experiments can be seen in Fig. 4 for the open fire and in Fig. 11 for the smoldering fire. Interestingly, although one might intuitively assume the smoldering fire to have more smoke than the open fire, our recorded data show that the detected smoke level approximately equalled or even exceeded that of the smolder. This is despite the subjective fact that to the human eye, the smoldering fire appeared to produce much greater quantities of smoke. It would be of interest to investigate the optical and aerosol nature of this difference. Nevertheless, actual shipboard fires are highly unpredictable due to cause and materials. Regardless of the nature of the difference, our experimental fires were measurably different enough to create a valid challenge to our detection system, especially given the early state of the detector design herein. Fig. 4. Measured open fire with: (a) data POINTS=3; (b) data POINTS=6; and (c) data POINTS=9. Next, to determine the better GM and also the best predictive period to use in the GM, the implemented GM at three reasonable predictive values experimentally tested the recorded fire data from our two fires. At this point, it would be instructive for those unfamiliar with grey modelling to look briefly at the summarized results of Table 5 and Table 6 (Table 5 summarizes open fire results, Table 7 summarizes smolder fire). As seen, the residual ratios (value of error divided by correct value, Eq. (37) and (38)) are compared for both GM and for the three prediction periods, and the best combination is easily identified. Table 5. Residual checking for first experiment Table 6. Earlier alarm triggering time (GM(1,1) unified-dimensional new message model) Table 7. Residual checking for the second experiment To arrive at Table 5, the as-sampled open fire temp erature and smoke data of Fig. 4 were processed by both GM at predictive periods of 20, 40 and 60 s and for 3, 6 and 9 data points. Raw data values and grey-prediction values are plotted in Fig. 5, Fig. 7 and Fig. 9 for the GM(1,1) unified-dimensional new message model and in Fig. 6, Fig. 8 and Fig. 10 for the GM(1,1) new message model. Smolder-fire results of Table 7 are arrived at similarly from the as-sampled data of Fig. 11. Fig. 5. Prediction diagrams of GM(1,1) unified-dimensional new message model with predictive PERIOD=20 s and data POINTS=3. Fig. 6. Prediction diagrams of GM(1,1) new message model with predictive PERIOD=20 s and data POINTS=3. Fig. 7. Prediction diagrams of GM(1,1) unified-dimensional new message model with predictive PERIOD=40 s and data POINTS=3. Fig. 8. Prediction diagrams of GM(1,1) new message model with predictive PERIOD=40 s and data POINTS=3. Fig. 9. Prediction diagrams of GM(1,1) unified-dimensional new message model with predictive PERIOD=60 s and data POINTS=3. Fig. 10. Prediction diagrams of GM(1,1) new message model with predictive PERIOD=60 s and data POINTS=3. Fig. 11. Measured signals of a smoldering fire with: (a) data POINTS=3; (b) data POINTS=6; and (c) data POINTS=9. Fig. 5, Fig. 7, Fig. 9, Fig. 6, Fig. 8 and Fig. 10 can be used to compare the predictive characteristics of the two GM for all tested conditions. In Fig. 5(a), it is clear that predicted values for smoke and temperature match well the detected values. The level of match deteriorates with increasing data points but remains reasonable. On the other hand, Fig. 6 shows clear mismatch between predicted and detected values, a condition that holds true regardless of the number of data points. Thus, the GM(1,1) unified-dimensional new message model with three data points seems to be the optimal grey modelling methodology, an evaluation confirmed more technically in lowest residual ratio (error rate) of 7.6% for smoke density and 3.1% for temperature given in Table 5. From the above, we selected what seems to be the superior fuzzy classification system and the superior GM. However, the choice of fuzzy classification was based only on training data. We would like to confirm its use under real data conditions. Also, we would like to confirm our choice of data point number and predictive period for the unified-dimensional new message GM. Thus, we need to combine the two systems and observe the results with real data. The same two real experimental fire data sets that were collected earlier were subjected to computer analysis using: (1) alarm-trigger values for a conventional system (see below); (2) the combined fuzzy/grey fire system with predictive periods and data point combinations from Table 5 for which residual values were less than 20%. The conventional alarm PC simulation applied alarm-trigger values for smoke of 15% or temperature of 70°C, as used in the commercial dual-sensor (photo/heat) alarm. The results for this are shown in Table 6, which summarizes comparative fire-detection performance in terms of the number of seconds earlier a fire condition is identified. All grey-fuzzy systems indicated fire conditions earlier than the conventional dual-sensor (smoke/temperature) fire alarm. For the predictive period of 60 s, three data points give the best early detection time, 62 s. However, as seen in Fig. 9, there are significant overshoots of predicted temperature and smoke density, which could result in false alarms. Overshoots are also found for the 20 and 40 s prediction times (Fig. 5 and Fig. 7, respectively), with the smallest overshoot found for Fig. 5(a). Overshoot is best avoided for increased accuracy. Fortunately, the levels found in Fig. 9 (60 s predictive time, three data points) are workable at our current levels of accuracy. Thus, it is shown that the proposed fire-detection system accurately triggers a fire alarm significantly earlier than a conventional dual-sensor alarm. Further, it is shown that the optimal configuration is a GM(1,1) unified-dimensional new message model using a predictive period of 60 s and three data points. Data are summarized in Fig. 11, Fig. 12, Fig. 13, Fig. 14, Fig. 15, Fig. 16 and Fig. 17 and Table 7 and Table 8. No dramatic differences are found. The final results again indicate an optimal configuration of a GM(1,1) unified-dimensional new message model using a predictive period of 60 s and three data points. The early time for the smoky fire is 68 and 62 s for the hot fire. Fig. 12. Prediction diagrams of GM(1,1) unified-dimensional new message model with predictive PERIOD=20 s and data POINTS=3. Fig. 13. Prediction diagrams of GM(1,1) new message model with predictive PERIOD=20 s and data POINTS=3. Fig. 14. Prediction diagrams of GM(1,1) unified-dimensional new message model with predictive PERIOD=40 s and data POINTS=3. Fig. 15. Prediction diagrams of GM(1,1) new message model with predictive PERIOD=40 s and data POINTS=3. Fig. 16. Prediction diagrams of GM(1,1) unified-dimensional new message model with predictive PERIOD=60 s and data POINTS=3. Fig. 17. Prediction diagrams of GM(1,1) new message model with predictive PERIOD=60 s and data POINTS=3. Table 8. Earlier alarm triggering time (GM(1,1) unified-dimensional new message model) Notably, overshoot for the open flaming fire occurred both in the temperature prediction and the smoke density prediction, while for the smoky fire overshoot occurred primarily in the smoke density prediction. The level of overshoot seems correlated with the amount of initial slope of the variable in question. Thus, a high and prolonged slope for a variable results in higher predicted values and higher overshoot for the variable. Overshoot, however, is not a problem for our system, since our fire alarm is presently concerned only with earliest trigger time and it almost takes place after a fire alarm is triggered. The plot or prediction of fire data after triggering is of no relevance. For more sophisticated fire alarm, analysis and monitoring systems, however, overshoot would presumably need to be taken into consideration. 3.4. Conclusion The function of a shipboard fire-detection system is to give reliably the earliest possible warning when unwanted fire conditions are present on board. This paper has presented a novel system whose simulated and experimental results demonstrate that real-time shipboard fire detection based on a combined fuzzy-classification/grey-prediction algorithm can give correct and rapid alarms for both smolder and open fires. Earlier work has shown multi-sensor fuzzy-neural fire alarms to be more accurate than conventional alarms. The present work has added a grey-prediction algorithm to a fuzzy system, making the proposed system demonstrably faster than other “classical” (heat/smoke) fire alarm systems, providing accurate alarms in the range of 60 s earlier. The current work demonstrates the feasibility of grey/fuzzy systems as applied to shipboard fire alarms. Future research may focus on refinement and implementation in real situations. One obvious factor that needs to be taken into account for shipboard applications is ambient conditions, since ships and ship compartments may experience anything from arctic or refrigerated to tropical or engine-room conditions. Possible extensions and refinements of temperature, environmental and sensor-type are plentiful, and it is expected that within the near future, the marketplace will see many “smart” fire-detection systems incorporating far more than the grey/fuzzy model presented herein. References Chen S, Yi J, Zhao Y. Self-learning fuzzy neural network and its application to fire auto-detecting in fire protection system. Proceeding of the Third World congress on Intelligent Control and Automation, Hefei, China, June 28–July 2, 2000, pp. 244-250. H. Ishibuchi and T. Nakashima, Improving the performance of fuzzy classifier systems for pattern classification problems with continuous attributes. IEEE Trans Ind Electron 46 6 (1999), pp. 1038–1057. H.C. Kuo and L.J. Wu, Prediction of deformation to thin ship panels for different heat sources. J Ship Prod 17 2 (2001), pp. 52–61. J. Deng, Control problems of grey system. Systems Control Lett 5 (2002), pp. 288–294. K. Nozaki, H. Ishibuchi and H. Tanaka, Adaptive fuzzy rule-based classification systems. IEEE Trans Fuzzy Systems 4 3 (1996), pp. 238–250. M. Setnes and H. Roubos, GA-FUZZY modeling and classification: complexity and performance. IEEE Trans Fuzzy system 8 5 (2000), pp. 509–522. Neubauer A. Genetic algorithms in automatic fire detection technology. Second International Conference on Genetic Algorithms in Engineering Systems: Innovations and Applications, Galesia, September 2–4, 1997, pp. 9. UL 268, Standard for smoke detectors for fire protective signaling systems, 2nd ed. Northbrook, IL: Underwriters Laboratories, 2003, pp. 3-5. Wang X, Xiao J, Bao M. A ship fire alarm system based on fuzzy neural network. Proceeding of the Third World Congress on Intelligent Control and Automation, Hefei, China, June 28–July 2, 2000. Yuan B, Guo J, Tang G, Wang L. Using grey theory to predict the gas-in-oil concentrations in oil-filled transformer. Proceedings of the 6th international conference on properties and applications of dielectric materials, Xi’an, China, June 21–26, 2000, pp. 10-15. Z. Li, A. Khananian, R.H. Fraser and J. Cihlar, Automatic detection of fire smoke using artificial neural networks and threshold approaches applied to AVHRR imagery. IEEE Trans Geosci Remote Sensing 39 9 (2001), pp. 1859–1870. Read More