NO170372B - STATISTICAL DISCRIMINATOR FOR FIRE FIGHTING - Google Patents

Description

The present invention relates to a statistical discriminator circuit for fire detection and a fire detection system which includes a pair of such statistical discriminator circuits. Furthermore, the invention relates to a method for statistically discriminating between stimuli from fire and non-fire sources when processing detected radiation in the time range.

Sensing the presence of a fire using photoelectric transducers is a relatively simple task. However, this becomes more difficult when one has to reliably discriminate between stimuli from a natural fire and other heat or light stimuli from a non-fire source. Radiation from the sun, ultraviolet light, sweeps, incandescent sources and the like often present particular problems with regard to false alarms generated in fire detection systems.

It has been found that improved discrimination can be developed by limiting the spectral response of the photodetectors used in the system. A plurality of signal channels having different spectral response bands have been used in a number of prior art systems using various methods to solve the problem of developing appropriate sensitivity for fire sensing while reliably discriminating against non-fire stimuli. However, the solutions in question have generally not realized the degree of efficiency that is necessary for a successful and reliable fire detection system that is not excessively exposed to the generation of false alarms.

TJS Patent 3,931,521 (Cinzori) relates to a dual channel fire and explosion detection system that uses a long wavelength radiant energy responsive detection channel and a short wavelength radiant energy responsive channel and introduces a condition of coincident signal detection to eliminate the possibility of false triggering. US Patent 3,825,754 (Cinzori et al) adds to the previously mentioned patent description the feature of discriminating between large explosive fires on the one hand and high energy flashes/explosions that do not cause any fire on the other hand. However, this specialized system is not easily convertible to more general fire detection system applications, such as the present invention.

US Patent 4,296,324 (Kern and Cinzori) relates to a dual spectrum infrared fire sensing system in which a long wavelength channel responds to radiant energy in a spectral band greater than about 4 micron and a short wavelength channel that responds to radiant energy in a spectral band less than approx. 3.5 micron, with at least one of the channels responsive to an atmospheric absorption wavelength associated with at least one combustion product of the fire or explosion to be detected.

TJS patent 3,665,440 (McMenamin) relates to a fire detector using ultraviolet and infrared light detectors and a logic system whereby an ultraviolet detection signal is used to suppress the output signal from the infrared detector. In addition, filters are provided in series with both detectors to respond to fire flicker frequencies of approx. 10 Hz. As a result, an alarm signal is developed only if flickering infrared radiation is present. A threshold circuit is also included to exclude low-level infrared signals, such as from a match or cigarette lighter, and a delay circuit is included to prevent spurious signals of short duration from triggering the alarm. However, such a system can be confused by other flickering sources, as simple and common as sunlight reflected off a shimmering water surface or a rotating fan chopping up sunlight or light from an incandescent lamp.

US patents 3,739,365 and 3,940,753 (Muller) deal with dual channel detection systems that use photoelectric sensors which respectively respond to different spectral ranges of incident radiation, the signals from these being filtered for the detection of flapping within a frequency range of about. 5 to 25 Hz. A differential amplifier generates an alarm signal in one of these systems when the signals in the respective channel deviate by more than a predetermined amount from a selected value or range of values. In the second system, the output signals from the differential amplifier are fed to a phase comparator with threshold circuit and time delay. An alarm signal is provided only if the input signals are in phase, of amplitude exceeding the threshold level, and of sufficient duration to exceed the preset delay. However, such a system may be ineffective for discrimination against non-fires, such as jet engine exhaust (which has a flapper content), in the presence of flashing or cloud-modulated sunlight.

US Patent 3,609,364 (Paine) uses several channels which are particularly suitable for the detection of hydrogen fires on board a rocket for high altitude with particular attention directed to discrimination against solar radiation and rocket engine smoke (plume) radiation.

US Patent 4,249,168 (Muggii) uses dual channels that respond to wavelengths in the range of 4.1 to 4.8 microns and 1.5 to 3 microns, respectively. Signals in both channels are subjected to a bandpass filter with a transmission range between 4 and 15 Hz for flame flapping frequency response. Both channels are connected to an AND port so that coincidence of detection in both channels is required for a fire alarm signal to be developed.

US Patent 4,220,857 (Bright) relates to an optical flame and explosion detection system having first and second channels which respectively respond to different products of combustion. Each channel has a narrow band filter to limit the spectral response. Level detectors in each channel signal detected radiation that exceeds selected threshold levels. A ratio detector provides an output when the ratio of signals in the two channels exceeds a certain threshold. When all three thresholds are exceeded by detected radiation, a fire signal is generated.

Other fire alarm and fire detection systems are discussed in US Patent 3,995,221 (MacDonald), US Patent 4,206,454 (Schapira et al), US Patent 3,122,638 (Steel et al), US Patents 2,722,677 and 2,762. 033 (Krueger), US Patent 4,101,767 (Lennington), US Patent 4,280,058 (Tar), and US Patents 4,160,163 and 4,160,164 (Nakauchi).

Despite the abundance of systems in the prior art with respect to fire detection, the fact remains that no system has been shown to be completely effective in discriminating against false alarms. In those systems where sensitivity is improved, there appears to be a concomitant reduction in other performance parameters, such as false alarm immunity. The present system is directed to techniques for analyzing radiation detection data to improve the reliability of fire detection.

Under certain circumstances, man-made phenomena or random natural phenomena can duplicate the characteristics of a fire in the frequency range. For example, the radiation from a light bulb (or other non-flammable source that emits both light and heat) can appear to a detector as fire in the frequency range if the light is chopped at a constantly varying rate. Sunlight reflecting ripples on a body of water can develop the same effect. The prior art fire detection systems currently known employ a frequency range analysis method for fire detection. The present invention involves processing amplitude information from each separate detection channel statistically in the time domain to eliminate the possibility of enhancement and error from radiation detection in the frequency domain. The invention uses special statistical methods to achieve this result.

The basic technique involves modulating a fire as a random process and using selected statistical mechanisms to test for random process characteristics. As a parameter for use in representing the "randomness" of a fire, the amplitude-distribution of the peak or change-in-slope point of the time-domain signal is chosen. Other parameters could also be used, such as zero-crossing interval, second-derivative-equal-zero point, etc. In order to develop data for the application of time domain statistical methods, one needs to keep a running tabulation of the peaks of the detected radiation signals. This is done by sampling the signal at the change-in-slope points. When the first derivative of the signal waveform changes sign, a sample is taken. In a particular embodiment of the invention, these sample signals over the last five seconds are stored in microprocessor storage locations. Approximately 40 to 50 data points, if developed for less than 5 seconds, are sufficient for the analysis. During the storage in memory, data points from more than five seconds previously are discarded. Periodically (approximately once per second) a calculation is made using the data points stored in storage.

Once a collection of data points is stored in storage, various statistical mechanisms can be used to determine whether or not the distribution of data points fits known random processes. A parameter that has been shown to be very definitive of the randomness of fire relative to the non-randomness of periodic radiation sources is the Kurtosis parameter. Kurtosis is a measure of how the collection of data is concentrated around its mean value. Large values of Kurtosis represent distributions with data points that are widely spread from the mean.

To determine the mean, the variance (or the standard deviation which is \/ji2) and the Kurtosis, if x^ represents the different data points, i=l, ...N, then one gets:

Kurtosis is defined as the ratio between the fourth central moment and the square of the second central moment:

where the fourth component is the average of all deviations raised to the fourth power, and the second central component is the average of all deviations raised to that power. As will be shown later here, Kurtosis is quite different for fires and non-fires. However, the quadrature and quadrature apparatus take quite a lot of computation time in a microprocessor implementation and a simplified version would be desirable for use with small microprocessors.

Just as multiple definitions exist to express the most likely value a statistically varying parameter may have (mean, median, mode, etc), more than one definition exists to express the degree to which the data points are spread about this "average" value . Each data point has a deviation, or difference, between its own value and that of the sample mean, taken here to be the arithmetic mean. A popular parameter for expressing the total deviation is the standard deviation (a) which is the root-mean-square value of a series of deviations. For a series of N samples, x^ t.o.m. Xjj, the mean (x) is given by the definition: and the standard deviation by:

This is a useful definition because that the squares of the deviations entail positive components so that the deviations of the opposite polarity will not equalize. Moreover, the quadratic function can be easily treated with algebra.

Another definition replaces the quadratic expression with the one for absolute value, thereby retaining a positive contribution from each deviation. This is known from the mean deviation:

It is less popular than the standard deviation because absolute-value function, defined to always return a positive result:

is sometimes difficult to handle in algebraic manipulations. However, it has strong appeal for microprocessor applications due to that polarity reversal in binary notation (complement and add 1 (LSB) (least significant bit)) is far easier to realize a squaring and square root functions.

With the definition of a measure for the deviation of data about the mean value, it is desirable to define a corresponding characteristic to express the extent to which the individual deviations are spread about the mean deviation. Two contrasting signals illustrate the need for this: A broadband Gaussian noise source and a square wave that has a zero-to-peak value equal to the mean deviation or standard deviation of the noise source. These have identical mean deviations, but nevertheless show radically different time characteristics and probability distribution functions (PDF) due to that the square wave has all its data points gathered at the same deviation.

For the special case of a square wave, all deviations are equal and the Kurtosis takes a value of 1. As the deviations become increasingly spread out, those greater than a will contribute more to P4 than those less than a subtracted from ^4. This is because the nonlinearity from the fourth power lies in JJ4. JJ2<2> in the denominator can be thought of as a normalization factor that causes K to be unitless and independent of the actual value of ^2 or a.

Another means of evaluating the dispersion of data around its standard deviation (or mean deviation, depending on what has been chosen) is to find the mean "deviation of the deviation", i.e. the average amount by which each individual deviation deviates from the mean (or standard) deviation. Again, the absolute difference will be used to preserve a positive contribution from each sample. With each individual deviation given by: |xj_ - x| as before, the mean difference between individual deviations and the mean deviation (hereafter defined by the term "spread" for lack of a better one) is expressed as:

This can be normalized by dividing by D and will be termed "modulation" as the parameter is now fairly analogous to that of amplitude modulation of a carrier wave. An unmodulated carrier wave (even with varying frequency) has a spread, and therefore modulation, equal to zero. The maximum steady state spread is equal to the mean deviation and therefore modulation can be varied from zero to one, or $100.

The preceding definition of modulation is intended to allow the evaluation of a signal for the same quality that Kurtosis provides, but without the need for multiplication (squaring and fourth powers) or extracting square roots. If the mean deviation is used for D, an integer power of 2 is used for N, and a constant fixed degree of modulation is used for a decision criterion, no true deviations need be performed. The apparent division of N becomes a series of right shifts (performed before summation to avoid overflow). The threshold circuit becomes a comparison between dispersion and a fixed fraction of D, again obtained by right shift (and possibly addition to get the desired fraction). A split will only be performed if an analog measure of modulation is desired for survey purposes. Thus, realization of this "simplified Kurtosis" makes it possible to use small inexpensive microprocessors to perform the real-time requirements of a fire detector's statistical discriminator.

In order to make the data collection practical according to the present invention, a device and input of data from the detected radiation signals includes a hysteresis circuit. The effect of this hysteresis circuit is to "clean" the data to separate the primary information from small disturbances or noise that may be present. The hysteresis circuit generates an output signal that trails the input signal with a fixed offset until a slope reversal occurs and a dead zone has been crossed. at which point the output begins to follow the input with a trailing shift of the opposite polarity. This ensures that small signal fluctuations of less than one to three percent of full scale do not give rise to a resampling of the trailing peak detector. The slope reversal indication in the output is stored in a peak detector. The real-time signal deviations are obtained by comparing the maximum and minimum sampling output signals to the sample mean. Comparing these results with the mean deviation followed by smoothing, again with a first-order delay, gives a dispersion value that will lie between zero and a value equal to the mean deviation.By dividing by one analog part, the modulation ratio S/D becomes available and can be compared with a fixed reference threshold. The final binary output is then a logical TRUE when the modulation is adequate to be that of a flap signal, indicating fire detection.

Another parameter that can be used to judge whether the set of data points in memory is arbitrarily distributed is the output from a simple up-down counter. If this counter is programmed to count down at, for example, a 3 Hz rate and up at the rate at which data is received from the waveform peaks, then low-frequency waveforms will not exceed a predetermined count threshold, regardless of whether they are arbitrary or not. As the waveform from a fire is known to have higher frequency components, this up-down counter parameter represents a small but additional criterion for distinguishing fires from non-fires.

Another parameter that can be used to judge randomness involves what is known as the Chi-squared "goodness of fit" test. In statistics, if one can say with a 95$ confidence level that a given result could not have occurred by chance, the result is said to be statistically "significant". Similarly, a $99 security level is "very significant".

By applying the Chi-square test to the collection of data points in memory, with a 95% confidence level, one can say that the given data points are normally distributed to a "significant" degree if the Chi-square test proves positive. The chi-square test is an assessment of how close to an arbitrary distribution the data points represent. The Chi-square test thus works well with the Kurtosis parameter to further rule out non-fire waveforms. For example, a waveform with a few large, narrow peaks, but most of its information concentrated near zero, could have a large kurtosis due to fourth power effect of the big peaks. However, the Chi-square test would recognize that the data points are not randomly distributed.

On the other hand, a periodic signal could have its amplitude modulated in a pseudo-arbitrary manner to the point where a collection of data points might be able to fit a Chi-square test. This may be the case in particular if the Chi-square test did not have many data points to work with and if the data points have gathered something about the mean value. However, the kurtosis parameter will detect that the "arbitrariness" is clustered around the mean, even with ten or fewer data points, and then fills the gap in the Chi-square test where few data points are available.

The statistical discriminator circuit is characterized, according to the invention, by a low-pass filter for connection to a radiation detector that responds to radiation in a preselected wavelength range,

peak detector means connected to the output of said filter for detecting the peaks against the remaining signal components, means for processing the peak signals to develop respective estimated mean values and mean deviation values of the peak signals,

means coupled to the processing means for combining said peak signals with said estimated mean values and mean deviation values to develop a signal spread level, and

means coupled to receive said signal spread level and a corresponding mean deviation value for dividing the signal spread

the level with the mean deviation value to determine the radiation modulation.

The fire detection system is characterized, according to the invention, in that each is connected to the output of a corresponding detector channel which comprises a radiation detector and associated amplifier, the radiation detector in a first of said channels being chosen to respond to radiation which has a long wavelength in the range of 7- 25 microns and the radiation detector in the second of said channels is selected to respond to radiation having a short wavelength in a preselected range.

The initially mentioned method is characterized, according to the invention, by the steps: receiving signals from a radiation detector which has a response to radiation within a preselected wavelength range,

filtering said received signals to remove components above a selected frequency,

to detect the peaks of the remaining signal components,

combining the peak signals to develop estimated mean values and mean deviation values of the peak signals,

combining individual peak signals with the estimated mean and estimated mean deviation values to develop a signal spread level, and

dividing the signal dispersion level by the estimated mean deviation value to give an output value for radiation signal modulation.

An alternative to this method is characterized, according to the invention, by the steps: deriving a series of sequential data signals by sampling detecting radiation waveforms according to a preselected parameter,

processing said signals according to at least one selected statistical analysis mechanism to test the randomness property of said detected radiation,

to compare the result of said treatment with a preselected threshold level, and

providing an output indicating the sensation of a fire when the result of said processing exceeds said threshold level.

Further embodiments of the discriminator circuit, the fire detection system and the method, according to the invention appear from the patent claims.

A better understanding of the present invention can be obtained from a consideration of the following detailed description, taken in connection with the attached drawings, where: Fig. 1 is a time domain plotting of waveforms from a flickering fire in channels for both long and short wavelength. Fig. 2 is a time domain plot of comparable waveforms from a warm, dull light bulb that is arbitrarily chopped. Fig. 3 is a diagram of waveforms of detected radiation

from a flickering fire in the frequency range.

Fig. 4 is another frequency range plot of detected radiation from a warm, dull light bulb chopped at a fixed frequency. Fig. 5 is a plot corresponding to that in fig. 4, but with

the radiation chopped up indiscriminately.

Fig. 6 is a flowchart illustrating a typical program that uses a special arrangement according to the present invention. Fig. 7 is a functional block diagram representing a particular arrangement according to the present invention. Fig. 7A is a block diagram showing a particular arrangement

which can be realized as a supplement to fig. 7.

Fig. 8 is a block diagram illustrating use of the present invention in a dual spectrum frequency responding fire detector of the cross-correlator type. Figures 9-16 are plots illustrating various waveforms included to illustrate the application of the present invention. Fig. 17 is a flowchart illustrating a combined counter

and Kurtosis test for fire detection.

Fig. 18 is a flowchart representing a Chi-square test

for fire detection.

Figures 1 and 2 are time domain plots of detected radiation and are provided to show the differences in detected radiation between a flickering fire and an artificial source. Fig. 1 shows a time domain plot of detected radiation from a flickering fire. The waveforms in fig. 1 represents detection in two channels. The upper waveform illustrates the signal from a short-waveform detector that has a response in the range of 0.8-1.1 microns. The lower waveform shows the output from a long waveform detector that has a response in the 7-25 micron range. Correlation on a time basis between the upper and lower waveforms is obvious. The amplitude of a given waveform is quasi-arbitrary. Fig. 2 shows the time domain plot of a detected arbitrarily chopped radiation from a warm, dull light bulb. The time scale is extended, relative to fig. 1, and the two waveforms are interchanged, i.e. the lower waveform 1 fig. 2 represents the output from a short wavelength detector in the range of 0.8-1.1 microns while the upper waveform represents the output from a long wavelength detector, in the range of 7-25 microns. Fig. 3 represents the plots of detected radiation from a flickering fire in the frequency range from 0 to 25 Hz. The upper waveform represents the radiation with the shorter wavelength, while the lower waveform represents the radiation with the longer wavelength. The time span for collecting this data is ten seconds and it will be noted that the peaks and valleys change from time to time. However, the general tendency is attenuation at the higher frequencies. Fig. 4 shows the waveforms of detected radiation from a warm, dull light bulb which is chopped at 2.6 Hz. The waveform with the longer wavelength is the upper waveform in the right part of the figure. There are clear peaks at 2.6 Hz, 7.8 Hz, and 13 Hz, corresponding to odd harmonics of the chopping frequency. Fig. 5 shows plots of detected radiation from a warm, dull light bulb, as in fig. 4, except that the chopping of the radiation is arbitrary rather than at a fixed frequency. The waveform of longer is the upper waveform in the left half of the figure. No clear peaks are present and the frequency range plotting is very similar to that in fig. 3. Figs 2 to 5 show that operation in the frequency range over the ten second sample integral does not provide sufficient information to allow one to distinguish between the light from a fire and chopped light from a light bulb. Time range processing is needed.

As a chopped waveform has relatively equal positive and negative peaks, peak detection was used to develop the data to be processed. When mechanizing the treatment, an Intel 2920 signal processor was chosen. Because of. the limiting mathematical ability of the aforementioned 2920, true Kurtosis calculation of j^/j<jg2> was not possible at 100 samples per second. Thus, the approximation of true kurtosis (referred to as "modulation") was used for the first embodiment. This approach proved quite successful for separating the random fire signal in Fig. 1 and 3 from the chopped light bulb measurement in fig. 2, 4 and 5.

A flowchart is shown in fig. 6 and represents a typical program that can be used to perform the modulation test described above, where the spread S is determined from the equation:

which is then normalized by dividing by D to develop modulation. The particular program represented in fig. 6 has been realized on an Intel 2920 signal processor using a 100 sample/second input rate, a five second smoothing time constant, and a modulation threshold equal to 38% for deciding whether the input signal corresponds to chopped or random radiation.

The incoming data samples, taken every 0.01 seconds, are passed through a 3-pole 4 Hz low-pass filter which is realized by recursive digital filter technology. The filter is quite similar to a Gaussian configuration, but has somewhat higher attenuation of the conjugate pole pair to ensure a lack of overshoots from rapid input changes. In addition, the slope polarity is taken from the difference between output samples separated by four sample intervals to further reduce the disturbance from noise transients above the desired signal passband.

The slope polarity is used to determine when a filtered data sample can be retained as a new positive peak (xp) or negative peak (xn). To be retained, it must occur after a signal change of at least 1% of full scale after the preceding peak. This dead zone reduces the likelihood that small fluctuations will reduce the usefulness of the peak data. Positive and negative peaks are independently smoothed by a 2.5 second time constant, single pole filter as an approximation of true averages, Xp and xn.

From these two values, the sample mean value x is estimated as 1/2 (xP + x<n>) and the mean deviation is estimated as D = 1/2 (xp-xn). With these, each peak sample, xP or x<n>, gives an individual deviation x-[ - x which can be used to calculate the spread and modulation as previously described. The smoothing time constant applied to S and M is 5 seconds. It must be longer than that used to derive x and D so that under transient conditions S cannot exceed D, giving rise to M negative or greater than one. In the threshold test, if M > 3/8 D, modulation is considered sufficient to indicate fire flapping signal.

It should be noted that in this embodiment the lack of second and fourth powers of the input signal will avoid the dynamic range problems associated with a true realization of the Kurtosis function. For example, an input signal range of 30:1 is typical for a usable range of 3 feet to 100 feet (91.44 cm to 30.48 m) with 30 dB of automatic gain control (AGC) compensation. Taken to the fourth power, this requires a dynamic range of 810,000:1 or 118 dB plus an additional 10 to 20 dB for waveform resolution within the weakest possible signal. Obviously, this requires a microprocessor with considerably more arithmetic capability than the aforementioned 2920 for a fire detector application. The modulation approach requires only the dynamic range of the signal plus the added 10 to 20 dB for waveform resolution, totaling 40 to 50 dB.

The functional block diagram in fig. 7 represents another possible realization of a modulation detector for the kurtosis approach. This is shown to include an input stage which has a low-pass filter 20 with a cut-off frequency equal to 4 Hz. This is followed by a hysteresis circuit 22 from which the signal is split into positive and negative parts for supply to respective peak detectors 24, 25. Each of the detectors is connected to a corresponding low-pass filter 26 or 27 which has a time constant equal to 2.5 seconds. These low-pass filters 26, 27 perform a summation operation on Xp and xn in analog form instead of in digital form, such as summation of x^ for the purpose of calculating an average, as follows:

These are in turn, in their respective channels, connected to attenuators 28 or 29 and operational amplifiers 30, 31. The output from the amplifier 30 is fed to another pair of operational amplifiers 32, 33 which are connected to receive respectively, on the remaining inputs , signals from the top detectors' 24, 25 outputs. The attenuator stages 34, 35 are connected respectively to the outputs of the amplifiers 32, 33 and are connected to provide inputs to a summing amplifier 36 which is also connected to the output of the amplifier 31. The output from the amplifier 36 is connected to a low-pass filter 38 which has a five second time constant which in turn is connected to an analog divider 40 which receives a second input from the amplifier 31 output. A comparator 42 is connected to the output of the divider 40 and also has a reference level input.

In a preferred arrangement according to the invention, detectors 24 and 25 are peak detectors which respond to a change in slope of the input waveform. As alternatives, the blocks 24, 25 can for example represent zero-crossing detectors, for determining zero-crossing time intervals, or other derivative-equal-zero detectors. Such detectors 24, 25 develop data in the form of selected sample signals which are then processed from an analysis of the input waveform according to the invention. In the particular discussion of the embodiment of FIG. 7 and 7Å, the circuits will be described in terms of top detectors 24, 25. However, it will be understood that these detectors 24, 25 may just as easily be other of the aforementioned types.

In the circuit in fig. 7, the input signal is filtered to below 4 Hz to remove high frequency noise and then applied to the hysteresis circuit 22. These stages, which can be fabricated with an assortment of integrators, diodes and offsets, as known in the art, generate an output that follows the input by a fixed displacement until a slope reversal occurs and a dead zone has been crossed. At that point the output begins to follow the input with a trailing shift of the opposite polarity. This ensures that small signal swings of less than one to three percent of full scale do not give rise to resampling of the following peak detector. Whenever a slope reversal occurs after a turn greater than 1%, referenced to the previous slope reversal, the new peak value (positive or negative) is stored in a peak detector. The resulting staircase-like waveforms are independently smoothed with a first-order delay filter having a time constant equal to 2.5 seconds. The following circuits 28, 29, the summing amplifier 30 and the difference amplifier 31 combine half of the sum of Xp and xn to get the average and also half of the difference to get the center-to-peak swing, or standard deviation. The stair step values from the maximum and minimum samples (Xp and xn) are compared with the sample mean value to obtain the real-time deviations. Comparing these with the mean deviation and smoothing, again with a first-order delay, gives a value of spread S that will lie between zero and a value equal to the mean deviation. By dividing by an analog divider 40, the modulation ratio, S/D, becomes available and can be compared to a fixed reference threshold in the comparator 42. The binary output is then a logical TRUE when the modulation is adequate to be that of a flapping signal.

The equations for S and D given earlier were realized as shown in fig. 6 and 7 to be adapted to the strengths of the 2920 signal processor. Thus, low-pass filters were used instead of calculated averages such as:

to avoid storing N data points. For other microprocessors that have larger memories, a direct calculation based on the equations can be directly applied.

Fig. 7A is a block diagram representing a particular circuit according to a feature of the present invention which may be included as a supplement to the circuits of Figs. 7.

Fig. 7A shows an up/down counter 72 which is driven in the UP direction by signals derived from the sampled waveform and in the DOWN direction by a clock. The circuit in fig. 7A can be connected in the circuit in fig. 7 in the manner indicated.

Signals to drive the counter 72 in the upward direction are taken from the positive and negative top detectors 24, 25 in fig. 7 before waveform smoothing is applied. These signals are fed into an OR gate 74 and then to the UP input of the counter 72. The DOWN input to the counter comes from a clock signal operating at approximately 3 Hz (for the circuit in Fig. 7 where the signals are cut off above 4 Hz by the low-pass filter 20 ). The count established in the counter 72 is fed to a threshold stage 76 which has a preselected reference level input for signal comparison. The output from the threshold stage 76 is supplied to an AND gate 78 which is connected to receive as a second input the output from the comparator stage 42 in fig. 7. Only when both inputs to AND gate 78 are TRUE will the logic output from AND gate 78 be TRUE, thereby indicating a fire.

With the counter 72 counting down at the clock rate equal to 3 Hz and counting up at the rate at which data is received from the waveform peaks at the peak detectors 24, 25, low frequency waveforms will not exceed the predetermined count threshold at step 76-, regardless of whether they are random or not. However, when a waveform from a fire is detected, higher frequency components of such a waveform cause the count to exceed the preset reference level at threshold stage 76, thereby applying a TRUE signal to AND gate 78.

Fig. 8 is a block diagram showing the implementation of statistical discriminators according to the present invention in a dual spectrum frequency responsive fire detector, as described in US Patent Application No. 592,611 (Mark T. Kern), entitled "Dual Spectrum Frequency Respond-ing Fire Sensor", transferred to the assignee for this application. The content of US patent application no. 592,611 (Norwegian patent application no. 854568) is incorporated herein by this reference. The circuit in fig. 8 corresponds to fig. 5 in application no. 592,611, where statistical discriminators according to the present invention replace the periodic signal detectors in said fig. 5 and with the addition of a cross-correlation detector, as discussed in FIG. 5 of US Patent Application No. 735,039 entitled "Fire Sensor Cross-Correlator Circuit and Method"

(also transferred to the assignee for this application). The content of that application is also included here with this reference.

In fig. 8 shows system 50 having n dual narrowband channels 1, 2, ...n, each set being on a different narrowband filter spectral passband F^, F2, ...Fn. Each of the narrow-band channels includes dual signal channels extending respectively from amplifier 55, coupled to the short-wavelength detector 53, and amplifier 56, coupled to the long-wavelength detector 54, to a ratio detector 57. As indicated, the short-wavelength detector 53 responds to wavelengths in the range of 0.8 to 1.1 microns, and the long-wavelength detector 54 responds to wavelengths in the range of 7-25 microns. Alternatively, the short wavelength detector 53 can be set to respond to wavelengths in the range of 1.3 to 1.5 microns.

Each of the signal channels includes a narrow-band filter, a full-wave rectifier and a low-pass filter connected in series between the amplifiers 55 or 56, as the case may be, and the input of the ratio detector stage 57. The outputs of the ratio detectors 57 in the n narrow-band channels 1, 2, .. .n is applied to a voting logic stage 59 which generates an output signal which is either TRUE or FALSE according to the majority of ratio detector outputs from the n narrowband channels. This output is coupled as an input to an AND gate 60, whose other inputs are the output from a cross-correlation detector 62 and the outputs from a pair of statistical discriminators 64, 65, supplied through inverter stages 66, 67. The output from the AND stage 61 is supplied to a delay stage 70 which provides the output from the sensor system 50.

The statistical discriminators 64, 65 in fig. 8 corresponds to the circuit shown in fig. 7. These replace the periodic signal auto-correlation detectors mentioned in our previous application, and provide improved recognition of artificially chopped sources, thereby developing better security against false alarms. In the circuit in fig. 8, an artificially chopped signal is recognized as such by the statistical discriminators 64, 65 which thereby block the AND gate 60 to prevent the circuit from developing a TRUE signal as a false alarm at the output. The statistical discriminators according to the present invention can be used instead of periodic signal detectors in other fire detection equipment to achieve a more restrictive response to artificially chopped radiation sources.

According to statistical theory, a true random process will have a Kurtosis equal to 3.0. To see how certain fire signals and certain non-fire signals compare to an arbitrary process, some analysis was performed by calculating the Kurtosis in sections of recorded data.

Fig. 9-16 show different waveforms illustrating this Kurtosis calculation performed according to the present invention, based on selected real-time signals. In these figures, the waveform in fig. 9 a pure sine wave, provided for comparison. The waveforms in fig. 10 and 11 correspond to chopped radiation from a warm, dull light bulb that is chopped. The chopping for the waveform in fig.

10 varies in frequency. The waveform in fig. 12 corresponds to solar radiation on a clear day. The waveform in fig. 13, 14 and 15 correspond to radiation from fires at varying distances of 30.5 metres, 15.25 meters and 6.10 metres, respectively. Finally, the waveform in fig. 16 derived from sunlight on a partly cloudy day.

In these cases, the calculations are based on the true Kurtosis equation:

and not on the approximation of spread S derived by dividing by D, as described above. Each calculation of the waveforms in Fig. 9-16 represent 20 data points (10 positive, 10 negative). The data, in millivolts after amplification, appear in the following Table 1, where certain signals are amplified more than others to achieve adequate resolution.

Each of the signals in table 1 and as represented in the waveforms in fig. 9-16 rides on a DC voltage level of approx. 1 volt. This makes no difference, as the data points for the mean (x) are subtracted out to obtain the variance and the kurtosis. As is clear from table 1, the chopped waveforms in fig. 9-11, although they vary in frequency as in fig. 10, a Kurtosis that is very close to a pure sine wave (Fig. 9). On the other hand, the fires, even at a distance of 30.5 meters, have a radically different Kurtosis (K = 2.5 to 3.2) and a value very close to that of a true random process.

Sunlight signals, which are shown in fig. 12 and 16, appear as random signals instead of chopped signals. The smaller sunlight signal in fig. 12 has a Kurtosis that falls in the area between a fire and a chopped signal. On the other hand, the larger sunlight signal in fig. 16 (a 15 point calculation instead of a 20 point calculation) a Kurtosis smaller than that of a fire. This is due to its arbitrary relatively chopped up nature. In a fire sensor system application, the high kurtosis of cloud-modulated sunlight allows a fire to be detected by other mechanisms, such as those that are the subject of the two pending applications mentioned above, even in the presence of direct sunlight.

The flowchart in fig. 17 illustrates how the Kurtosis test is mechanized together with the up/down counter test (see Fig. 7A). A decision heading for a 1/3 second elapsed time represents a 3 Hz counter 72 counting down, while peak signals generated from slope polarity changes energize the counter to count up. A threshold with a count of 4 is used as the decision point regarding whether data from slope changes is received quickly enough to represent a fire.

Likewise, a decision point with a Kurtosis equal to 2.4 is used to indicate whether the data points are correctly distributed to indicate a fire. The 2.4 reference level is derived empirically from the variation in Kurtosis for a fire which is in the range of 2.5 to 3.2 from Table 1, with that for non-fires which is in the range of 1.0 to 1.9.

Fig. 18 is a flowchart representing the performance of a Chi-square test on sampled data from received radiation to detect the presence of a fire. Pre-programmed into fig. 18 is K, the number of bins to use when calculating Chi-square. Also pre-programmed into fig. 18 is the expected number of samples per container expressed as a percentage of N, the total number of samples in stock. Thus, if e<1> is known, the container edges are calculated in the form of x and a and all data points in storage are sorted into the K containers. bjj is then the number of samples sorted in the k'th bin. Chi-square is then calculated and compared to the decision value c, which is also pre-programmed in fig. 18 through knowledge of K.

As an example, consider the case from Table 1 for the column headed fig. 15 where N=20 samples have been taken and K=6 intervals are to be used for testing the hypothesis that they come from a normal probability distribution with a 95% confidence level. The fire interval boundaries, , can be chosen (arbitrarily) to be equally spaced by: x - a, x - a/2, x, x + a/2, andx+2. From a table of the normal curve of error, the number of samples that can be expected to fall in these intervals are: e^ to e^ = 3.2 respectively; 3.0; 3.8; 3.8; 3.0 and 3.2.

From table 1 for fig. 15, the test samples will be sorted into these same intervals with the following counts b^ to b^ = 3, 2, 7, 3, 2 and 3 respectively. Chi-square can be calculated as follows:

From a Chi-square table using 3 degrees of freedom at the 95$ probability level, the decision value is c = 7.81. The example from table 1 is smaller than this; therefore, the 20 data points in the example are considered to be normally distributed to a 9556 confidence level. For a value of Chi-square close to c, such as in the column for fig. 13, a decision test can be applied based on the number of data samples in stock. For a number of data samples less than 20, the Chi-square test becomes less reliable. Thus, for fewer than 20 samples in stock, the Chi-square value can be ignored if it conflicts with the Kurtosis/Counter test results. For more than 20 data points in stock, the Chi-square test output can be combined with that from the Kurtosis/Counter test for additional reliability.

In summary, the present invention applies statistical analysis to detected radiation signals as a further means of discriminating between fire sources and artificial radiation sources. By applying this statistical analysis to the radiation in the time domain, the invention provides an added dimension to the capability of the frequency range sensing systems that have been developed previously, whereby combinations with such systems are enabled to operate with increased sensitivity by providing the added protection against false alarms. Statistical discriminators according to the present invention provide signal sampling and processing of data in a microprocessor, using selected statistical analysis parameters which are taken care of by the microprocessor. In a method according to the present invention, the true Kurtosis equation is followed. In another method, according to the present invention, Kurtosis is approximated by a simplified solution that eliminates the need for multiplication, squaring, fourth powers or extracting square roots, which are operations that make processing in the microprocessor slower. In another method, an up/down counter is used to prevent low frequency signals, which cannot be fires, from confusing the signal processing. In a further method, the Chi-square test is used as a further test of the incoming waveform.

Although specific solutions of a fire detector's statistical discriminator according to the invention have been described above in order to illustrate the way in which the invention can be used with advantage, it will be understood that the invention is not limited to this. Accordingly, any and all modifications, variations or equivalent solutions that may appear to those skilled in the art, such as other tests based on arbitrary treatment, shall be considered to lie within the scope of the invention as defined in the appended claims.

Claims (47)

1. Statistical discriminator circuit for fire detection, characterized by: a low-pass filter for coupling to a radiation detector which responds to radiation in a preselected wavelength range, peak detector means coupled to the output of said filter for detecting the peaks against the remaining signal components, means for processing the peak signals to develop respective estimated mean values and mean deviation values of the peak signals, means coupled to the processing means for combining said peak signals with said estimated mean values and mean deviation values to develop a signal spread level, and means coupled to receive said signal spread level and a corresponding mean deviation value for dividing the signal spread level by the mean deviation value to determine the radiation modulation. 2. Circuit as stated in claim 1, characterized in that the peak detector means comprise a pair of peak detectors of opposite polarity connected to the output of said filter to separate the signal peaks according to polarity and to supply peak signals of opposite polarity to a pair of parallel signal channels, also comprising means coupled to the two signal channels for combining said peak signals of positive and negative polarity with said estimated mean values to develop signal levels corresponding to the deviation of individual peak signals from the estimated mean value, and means for combining the individual peak signal deviations with said estimated mean deviation value. 3. Circuit as stated in claim 2, characterized by also comprising means connected between the low-pass filter and the peak detectors to establish a dead band to block the response of the peak detectors to small signal deviations. 4. Circuit as stated in claim 3, characterized in that said means for establishing a dead band comprise a hysteresis stage coupled to react to output signals from the low-pass filter, said hysteresis stage having a predetermined sensitivity level. 5. Circuit as stated in claim 2, characterized in that each of the two signal channels includes a low-pass filter stage connected to the output of its corresponding peak detector. 6. Circuit as stated in claim 2, characterized in that each of the parallel channels is connected to provide signal inputs to a first pair of amplifiers for development of the estimated mean value and the estimated mean deviation value as respective outputs from said amplifier. 7. Circuit as set forth in claim 6, characterized by further including a second pair of amplifiers, coupled to receive as respective inputs the estimated mean value and a corresponding one of the positive and negative peak signals from the peak detectors and to provide individual deviation signals corresponding to the deviations of individual peak signals from the estimated mean value. 8. Circuit as set forth in claim 7, characterized by further including a summation stage for combining said individual deviation signals with the estimated mean deviation value and a low-pass filter coupled to the output of the summation stage for equalization of output signals therefrom to develop the signal spread value. 9. Circuit as stated in claim 1, characterized by also including means connected to the output at the signal propagation level, the dividing means for comparing the modulation with a fixed reference threshold and developing an output signal indicating fire detection for modulation that exceeds said reference threshold. 10. Circuitry as set forth in claim 1, further comprising an up/down counter, means for coupling peak signals from the peak detector means to an input of the counter to cause it to count in a first direction, a clock signal coupled to the second input on the counter to cause it to count in a second direction, and a threshold stage coupled to the output of the counter to compare said output with a preselected reference level and develop a logical TRUE signal when the count state in said counter exceeds said preselected reference level, thereby indicating detection of a fire. 11. Circuit as set forth in claim 10, characterized by further comprising a comparator stage coupled to receive a signal indicating the radiation modulation for the comparison with a preselected reference level and to develop a logically true output indicating detection of a fire when the radiation modulation exceeds the reference level of the comparator stage. 12. Circuit as set forth in claim 11, characterized by further including an AND gate coupled to receive the outputs from the threshold stage and the comparator stage and provide a logical TRUE output denoting detection of a fire by the simultaneous occurrence of logically the same outputs from said threshold stage and said comparator stage. 13. Fire detection system including a pair of statistical discriminator circuits as stated in claim 1, characterized in that each is connected to the output of a corresponding detector channel comprising a radiation detector and associated amplifier, the radiation detector in a first of said channels being selected to respond to radiation having a long wavelength in the range of 7-25 microns and the radiation detector in the second of said channels is selected to respond to radiation having a short wavelength in a preselected range. 14. System as stated in claim 13, characterized in that said preselected range is between 0.8 and 1.1 microns. 15. System as stated in claim 13, characterized in that said preselected areas are between 1.3 and 1.5 microns. 16. System as stated in claim 13, characterized by also comprising a cross-correlation detector which is connected in parallel with the two statistical discriminators the circuits to provide a combined output indicating the detection of radiation from a fire. 17. System as stated in claim 16, characterized in that the cross-correlation detector is connected to receive signals from both detector channels via separate inputs and to provide a fire detection output in parallel with output signals from the statistical discriminator circuits. 18. Method for statistically discriminating between stimuli from fire and non-fire sources when processing detected radiation in the time domain, characterized by the steps of: receiving signals from a radiation detector having a response to radiation within a preselected wavelength range, filtering said received signals to remove components above a selected frequency, to detect the peaks of the remaining signal components, to combine the peak signals to develop estimated mean values and standard deviation values of the peak signals, to combine individual peak signals with the estimated mean and estimated mean deviation values to develop a signal spread level, and to divide the signal spread level by the estimated mean deviation value to provide an output value for radiation signal modulation. 19. Method as set forth in claim 18, characterized in that the detecting step comprises separating the peak signals according to their polarity, and that it further comprises the steps of filtering the positive peak signals and the negative peak signals separately to develop respective estimated mean values for the positive and negative peak signals, combining an estimated mean value with individual peak signals of opposite polarity to develop respective individual deviation signals for the positive and negative peak signals, and combining said individual deviation signals with the estimated mean deviation value to develop the signal spread level. 20. Method as stated in claim 18, characterized by also comprising the step of comparing the modulation value with a preselected threshold reference level to develop an output that indicates the detection of a fire when the modulation value exceeds said reference level. 21. Method as stated in claim 18, characterized in that said selected frequency is 4 Hz. 22. A method as set forth in claim 21, further comprising the step of establishing a dead band for signals of opposite polarity to block the detection of signal peaks for signal changes less than a predetermined level. 23. A method as set forth in claim 20 further comprising combining the output of the modulation comparison with the output of a cross-correlator stage coupled to receive signals corresponding to detected radiation in a preselected wavelength range to provide a TRUE fire sensing signal only upon the simultaneous occurrence of outputs from the cross-correlator and the statistical discriminator stages. 24. Method as stated in claim 18, characterized in that the radiation detector is chosen to have a radiation response in the range of 7-25 microns. 25. Method as stated in claim 18, characterized in that the radiation detector is chosen to have a radiation response in the range of 0.8 - 1.1 microns. 26. Method as stated in claim 18, characterized in that the radiation detector is chosen to have a radiation response in the range of 1.3 - 1.5 microns. 27. Method as set forth in claim 20, characterized by further including the step of supplying peak signals to an input of a counter to drive the counter in the first direction, supplying clock signals with a repetition rate slightly less than said selected frequency to drive the counter in the opposite direction, and comparing the count state of the counter with a predetermined reference level to develop a logic output corresponding to the sensing of a fire when the count state exceeds said reference level. 28. Method as set forth in claim 27, characterized by further comprising the steps of combining the logical output from the count comparison with a logical output from the modulation value comparison to develop a logical TRUE signal indicating fire sensing in the event that both said combined signals indicate sensing a fire. 29. Method as set forth in claim 28, characterized by further comprising applying a Chi-square test to a plurality of peak signals by developing values of Chi-square for said signals, comparing the values of Chi-square with a selected reference level, and providing an output indicating the sense of a fire for Chi-square values less than said reference level. 30. Method for statistically discriminating between stimuli from fire and non-fire sources by processing detected radiation in the time domain, characterized by the steps: deriving a series of sequential data signals by sampling detected radiation waveforms according to a preselected parameter, processing said signals according to to at least one selected statistical analysis mechanism to test the randomness property of said detected radiation, to compare the result of said treatment with a preselected threshold level, and to provide an output indicating the sensation of a fire when the result of said treatment exceeds said threshold level. 31. Method as stated in claim 30, characterized in that the processing step includes deriving an average value for a selected number of said data signals, using said average value to calculate the variance of said selected number of data signals, and using said average value and said variance to calculate The kurtosis of said selected number of data signals, and that the comparison step comprises comparing the calculated kurtosis with the preselected threshold level as the basis for indicating the feeling of a fire. 32. Method as stated in claim 31, characterized by also including the step of requiring that the calculated Kurtosis exceeds said preselected threshold level for a predetermined interval before providing said output indicating the feeling of a fire. 33. Method as stated in claim 32, characterized by also including the step of supplying said signals, together with clock pulses, to an up/down counter for calculating the Kurtosis. 34. Method as stated in claim 30, characterized by said derivation step comprises detecting changes in slope polarity of a detected radiation waveform and sampling said waveforms upon detection of a slope polarity change to develop said data signals. 35. Method as set forth in claim 34, characterized by further comprising the step of using said slope polarity change signals' to increment a counter and using clock signals to decrement the counter prior to said signal processing step. 36. Method as stated in claim 32, characterized by also including the step of storing said data signals derived within a predetermined time interval in a storage. 37. Method as stated in claim 36, characterized in that said storage step comprises updating the data stored in the storage to retain the stored signals on a first-in, first-out basis. 38. Method as set forth in claim 37, characterized in that said processing step includes dealing with the signals that are stored in storage within a predetermined time interval prior to the time of processing. 39. Method as stated in claim 38, characterized in that the calculation of said average value, the variance and the kurtosis is performed approximately once per second. 40. Method as stated in claim 31, characterized in that the sampling of a detected radiation waveform is carried out at zero crossings of said waveform. 41. Method as stated in claim 31, characterized in that the sampling of a detected radiation waveform is carried out at points where the waveform changes slope polarity in order to detect positive and negative peaks in the waveform. 42. Method as stated in claim 31, characterized in that the sampling of a detected radiation waveform is carried out by detecting the points where the second derivative of the waveform is equal to zero. 43. Method as stated in claim 31, characterized in that the amplitude distribution of the waveform peaks is selected as the parameter for determining the sampling of the radiation waveform. 44. Method as stated in claim 30, characterized in that the step of processing said signals includes the calculation of the Kurtosis of a selected series of data signals to determine the degree of arbitrariness of a detected radiation waveform as a criterion for providing the output indication of fire sensing. 45. Method as set forth in claim 44, characterized by applying a Chi-square test to a plurality of peak signals when developing values of Chi-square for said signals, comparing the value of Chi-square with a selected reference level, and providing an output signal which indicates the feeling of a fire for Chi-square values less than said reference value. 46. Method as stated in claim 30, characterized in that the step of processing said signals comprises calculating the spread of data signals and dividing by the mean deviation to determine the modulation of the detected radiation waveform as a criterion for providing the output indication for fire detection. 47. Method as set forth in claim 46, characterized by further comprising applying a Chi-square test to a plurality of peak signals by developing values of Chi-square for said signals, comparing the value of Chi-square with a selected reference level, and providing an output signal indicating the sense of a fire for Chi-square values less than said reference level.