WO2022178197A1 - Neural network-based anomaly detection system and method - Google Patents

Abstract

An exemplary method and system are disclosed to detect a fire anomaly (e.g., presence of fire, smoke, heat, carbon monoxide, air quality monitoring) using a trained neural network model. In some embodiments, the exemplary method and system may employ variational auto-encoders configured in an unsupervised learning setting. In addition to fire anomaly, the exemplary method and system may be beneficial for other indoor environmental anomaly detection, e.g., building security and intruder detection.

Description

Neural Network-based Anomaly Detection System and Method

Statement of Government Interest

[0001] This invention was made with government support under GR00003481 awarded by the National Aeronautics and Space Administration (NASA). The government has certain rights in the invention.

Related Application

[0002] This PCT International Application claims priority to, and the benefit of, U.S.

Provisional Patent Application No. 63/150,648, filed February 18, 2021, entitled “Deep LSTM Variational Autoencoder for Fire Detection,” which is incorporated by reference herein in its entirety.

Background

[0003] Conventional fire detection system uses fire signatures such as those from smoke, heat, carbon monoxide, or radiation sensors, to identify early signs of fire and trigger an alarm. There can be significant costs associated with missed fire detection with respect to injuries, fatalities, and property damage when signs of fire are not detected early for the unfolding fire to be neutralized or contained at an early stage. There are also costs associated with false alarms. For buildings configured with fire sprinkler suppression systems, false alarms can cause unnecessary damage from their own fire suppression system. Fire alarm horns and sirens are intended to invoke a reaction of people, and false alarms can be a nuisance to the occupants of the building. In the United States, local fire departments respond to over 2million nuisance alarms yearly. Nuisance alarms are a leading cause of disconnected detectors.

[0004] Current fire detection system typically employs a fixed trigger mechanism that uses a single or a pre-defined finite set of measured readings to trigger the alarm. The current rate of false alarms is still too high. Using conventional fixed trigger mechanisms, a tradeoff may exist between sensitivity and specificity. Presently, lowering the nuisance alarm rate compromises the early detection of fires and can come at the cost of missed detection.

[0005] There is a benefit to improving or maintaining the accuracy and specificity of fire alarm systems while also reducing nuisance alarms.

Summary

[0006] An exemplary method and system are disclosed to detect a anomaly (e.g., the presence of fire, smoke, heat, carbon monoxide, poor air quality) using a trained neural network model. The exemplary method and system provide a detection architecture (e.g., fire-related detection) that leverages soft computing techniques and deep learning tool integration to detect an anomaly. The exemplary method and system may employ variational auto-encoders configured in an unsupervised learning setting. In addition to fire anomaly, the exemplary method and system may be beneficial for other indoor environmental anomaly detection, e.g., building security and intruder detection.

[0007] The exemplary method and system are configured to determine a score (e.g., a fire score) generated as a comparison, e.g., as a difference, between a reconstructed signal model and the actual sensor signal. As the difference between the synthetic and the actual signals continues and/or increases, the system accumulates evidence that an anomaly has occurred or is occurring, and the fire score is increased. When this increase becomes sufficiently high to be clearly distinguishable from noise, an analogy alarm (e.g., fire alarm) is triggered. The signature of anomalies (e.g., fires) or nuisances is reflected in the time- varying differences between the two signals (e.g., “fire score” hereafter). The exemplary system and method provide additional degrees of freedom to improve the accuracy and specificity of alarm systems while also reducing nuisance alarm and is in stark contrast with the existing memoryless detection principle.

[0008] In some embodiments, the exemplary detection method and system employ a deep Long-Short Term Memory (LSTM) neural networks and variational autoencoder (VAE) that is configured to meet or outperform the increasingly stringent requirements of existing fire detection methods. A study was conducted, via computation analysis, to evaluate the performance of the exemplary fire detection method and system as compared to alternative dynamic and fixed-temperature heat detectors. The study observed that the LSTM-VAE based fire detection robustly outperforms the other detection methods with statistically significant shorter alarm time lags, no missed detection, and no false alarms.

[0009] In some embodiments, the trained neural network is trained at the manufacturer site (or remote site) to be deployed at the local edge device. In other embodiments, a reduced or simplified model can be deployed at the local edge device based on that trained neural network.

[0010] In some embodiments, the system is configured to retrain the neural network at the edge device. In other embodiments, the neural network is retrained at a local or global controller, e.g., over a cloud infrastructure or home automation system, over a communication channel, and then updated at the edge device.

[0011] In some embodiments, the exemplary method may be used for various advanced anomaly detection tools for a broad range of applications (with sequential multisensor data), including air quality monitoring and other environmental monitoring. [0012] In an aspect, a method is disclosed to detect a fire anomaly (e.g., presence of fire, smoke, heat, carbon monoxide, air quality monitoring), the method comprising receiving, by a processor, sensor data from a sensor device (e.g., smoke alarm, fire alarm, carbon dioxide alarm, heat alarm, air quality monitoring alarm); generating, by the processor, via a trained neural network model (locally or globally trained) or a local model derived therefrom (e.g., a global model), reconstructed sensor data using the sensor data; and determining, by the processor, a value associated with a presence or non-presence of a fire anomaly by comparing the reconstructed sensor data to the sensor data, wherein the determined value is used to generate an alarm associated with the fire anomaly.

[0013] In some embodiments, the method includes comparing, by the processor, the determined value associated with a presence or non-presence of a fire anomaly to a predefined threshold (e.g., first pre-defined threshold), wherein an alarm (e.g., audible alarm or light alarm) or notification (e.g., communication message, e.g., to a nearby sensor or local/remote central controller) is generated based on the comparison.

[0014] In some embodiments, the method further includes outputting, by the processor, a signal associated with the determined value associated with a presence or nonpresence of a fire anomaly; and comparing, via an electronic circuit, the outputted signal to a pre-defined signal threshold, wherein an alarm (e.g., audible alarm or light alarm) or notification (e.g., communication message, e.g., to a nearby sensor or local/remote central controller) is generated based on the comparison.

[0015] In some embodiments, the neural network model comprises a variational autoencoder with a deep LSTM network.

[0016] In some embodiments, the neural network model comprises at least one of a recurrent neural network, a convolutional neural network, a machine learning model, and a combination thereof.

[0017] In some embodiments, the determined value is denoised following the comparison between the reconstructed sensor data and the sensor data to produce a smoothed score associated with the fire anomaly.

[0018] In some embodiments, the determined value is denoised via a Kalman filter.

[0019] In some embodiments, the trained neural network model was trained in a nonfire condition using sensor data from (i) the sensor device or (ii) a sensor device having a similar or same sensor type. [0020] In some embodiments, the trained neural network model is retrained based on an input command generated from a mechanical input (e.g., button or capacitive sensor) located on the sensor device.

[0021] In some embodiments, the trained neural network model is retrained at a remote computing system (e.g., cloud infrastructure or local controller (e.g., in-home controller)), the sensor device comprising a network interface that is operatively connectable over a network (e.g., physical or wireless network in the local area network and through a wide-area network) to the remote computing system.

[0022] In some embodiments, the fire anomaly includes at least one of the presence of smoke, the presence of elevated heat, the presence of fire, the presence of carbon monoxide, and a combination thereof.

[0023] In some embodiments, the pre-defined threshold is varied during the operation of the sensor device.

[0024] In some embodiments, the method further includes comparing, by the processor, the determined value associated with a presence or non-presence of a fire anomaly to a second pre-defined threshold, wherein the alarm or notification is generated based on the second comparison, and wherein the second pre-defined threshold has an associated threshold time, and wherein the associated threshold time of the second pre-defined threshold is different from an associated threshold time of the pre-defined threshold (e.g., first pre-defined threshold).

[0025] In another aspect, a system (e.g., sensor device) is disclosed comprising at least one sensor configured to detect one or more fire-associated measurements (e.g., fire, smoke, heat, carbon dioxide, air quality monitoring); a processor (e.g., microprocessor, microcontroller, low-cost ASIC, programmable logic device, etc.); and memory having instructions stored thereon, wherein execution of the instructions by the processor causes the processor to: receive sensor data from a sensor device (e.g., smoke alarm, fire alarm, carbon dioxide alarm, heat alarm, air quality monitoring); generate, via a trained neural network model or a local model derived therefrom, reconstructed sensor data using the sensor data; and determine a value associated with a presence or non-presence of a fire anomaly by comparing the reconstructed sensor data to the sensor data, wherein the determined value is used to generate an alarm associated with the fire anomaly.

[0026] In some embodiments, the system further includes a speaker operatively connected to the processor, wherein execution of the instructions by the processor further causes the processor to (i) compare the determined value associated with a presence or non- presence of a fire anomaly to a pre-defined threshold and (ii) direct output of a signal to energize the speaker based on the comparison.

[0027] In some embodiments, the system further includes a speaker operatively connected to the processor, wherein execution of the instructions by the processor further causes the processor to output a signal associated with the determined value associated with a presence or non-presence of a fire anomaly; and an electronic circuit configured to receive the signal and compare the signal to a pre-defined signal threshold and direct an output to energize the speaker based on the comparison.

[0028] In some embodiments, the neural network model comprises a variational autoencoder with a deep LSTM network, a recurrent neural network, a convolutional neural network, a machine learning model, or a combination thereof.

[0029] In some embodiments, execution of the instructions by the processor further causes the processor (e.g., by executing a Kalman filter) to denoise or smoothen the determined value following the comparison between the reconstructed sensor data and the sensor data.

[0030] In some embodiments, the trained neural network model was trained in a nonfire condition (e.g., using sensor data from (i) the sensor device or (ii) a sensor device having a similar or same sensor type).

[0031] In some embodiments, the system further includes a housing; and a mechanical input located on the housing, the mechanical input being coupled to associated circuitry that is operatively coupled to the processor, wherein execution of the instructions by the processor further causes the processor to initiate, based on the mechanical input, retaining of the trained neural network model (e.g., locally, or remotely).

[0032] In some embodiments, the sensor includes a photoelectric sensor or detector

(e.g., photoelectric heat or photoelectric smoke detector, air quality monitoring); a sampling tube smoke sensor or detector; a duct smoke sensor or detector; a carbon monoxide sensor or detector; an ionization sensor or detector; a temperature sensor; a resistance temperature detector (RTD) sensor; a thermistor sensor; an air quality monitoring sensor or detector, or a combination thereof.

[0033] In some embodiments, execution of the instructions by the processor further causes the processor to compare the determined value associated with a presence or nonpresence of a fire anomaly to a second pre-defined threshold, wherein the alarm or notification is generated based on the second comparison, and wherein the second predefined threshold has an associated threshold time, and wherein the associated threshold time of the second pre-defined threshold is different from an associated threshold time of the predefined threshold.

[0034] In another aspect, a method is disclosed to configure a sensor device to detect a fire anomaly, the method comprising training, a neural network model, using sensor data acquired from a plurality of sensor devices each having a similar or same sensor type as the sensor device (e.g., wherein the sensor data are acquired during non-fire condition); and storing the neural network model or a local model derived therefrom to a memory of the sensor device, wherein the trained neural network model or the local model derived therefrom is subsequently used to generate reconstructed sensor data during operation of the sensor device, an wherein the reconstructed sensor data is compared to the sensor data to generate an alarm associated with the fire anomaly.

[0035] In another aspect, a system is disclosed comprising at least one sensor configured to detect one or more fire-associated measurements (e.g., fire, smoke, heat, carbon dioxide, air quality monitoring); a processor (e.g., microprocessor, microcontroller, low-cost ASIC, programmable logic device, etc.); and memory having instructions stored thereon, wherein execution of the instructions by the processor causes the processor to perform any one of the above-discussed methods.

[0036] In another aspect, a non-transitory computer-readable medium is disclosed having instructions stored thereon, wherein execution of the instructions by a processor of a sensor device causes the processor to perform any one of the above-discussed methods.

Brief Description of the Drawings

[0037] The skilled person in the art will understand that the drawings described below are for illustration purposes only.

[0038] Fig. 1 shows an example building safety monitoring system 100 for anomaly detection in accordance with an illustrative embodiment.

[0039] Figs. 2A - 2C each shows an example system to configure the trained neural network of Fig. 1 to perform anomaly detection in accordance with an illustrative embodiment.

[0040] Fig. 3 shows an example anomaly detection algorithm configured with a trained neural network to perform anomaly detection in accordance with an illustrative embodiment.

[0041] Figs. 4A and 4B each shows an example method to train a neural network for anomaly detection in accordance with an illustrative embodiment. [0042] Figs. 4C and 4D each shows an example method to operate a building safety monitoring system configured with a trained neural network for anomaly detection in accordance with an illustrative embodiment.

[0043] Fig. 5A shows an overview of a study to evaluate the building safety monitoring system and methods in accordance with an illustrative embodiment.

[0044] Fig. 5B shows performance results as a receiver operating characteristic

(ROC) curve of the various detection methods in an aspect of the study in accordance with an illustrative embodiment.

[0045] Figs. 6A and 6B show performance results of the building safety monitoring system and methods based on simulation-based data.

[0046] Figs. 7A and 7B show performance results of the building safety monitoring system and methods based on real-world data comprising fire events and non-fire false alarm events.

Detailed Specification

[0047] Some references, which may include various patents, patent applications, and publications, are cited in a reference list and discussed in the disclosure provided herein. [0048] Some references, which may include various patents, patent applications, and publications, are cited in a reference list and discussed in the disclosure provided herein. The citation and/or discussion of such references is provided merely to clarify the description of the disclosed technology and is not an admission that any such reference is “prior art” to any aspects of the disclosed technology described herein. In terms of notation, “[n]” corresponds to the nth reference in the list. For example, [4] refers to the first reference in the list, namely, K. Muhammad, J. Ahmad, I. Mehmood, S. Rho, and S. W. Baik, “Convolutional neural networks-based fire detection in surveillance videos,” IEEE Access, vol. 6, pp. 18174- 18183, 2018. All references cited and discussed in this specification are incorporated herein by reference in their entireties and to the same extent as if each reference was individually incorporated by reference.

[0049] Example System

[0050] Fig. 1 shows an example building safety monitoring system 100 (shown as

100a) for anomaly detection in accordance with an illustrative embodiment. The system 100a includes a sensor 102 that is operatively coupled to a processor 104 that executes trained neural network 106. The sensor 102 may be based on a photoelectric sensor or detector (e.g., photoelectric heat or photoelectric smoke detector), a sampling tube smoke sensor or detector, a duct smoke sensor or detector, a carbon monoxide sensor or detector, an ionization sensor or detector, a temperature sensor, a resistance temperature detector (RTD) sensor, a thermistor sensor, an air quality monitoring sensor or detector, or a combination thereof. The sensor 102 is configured to provide sensor signals 108 to be used for the detection of the presence of fire, smoke, heat, poor air quality, or carbon monoxide via a comparison with a sensor model data set or signal 110 generated by the trained neural network 106.

[0051] In generating a sensor model data or signal 110, the trained neural network

106 is configured to leverage contextual and dynamic information in the streaming sensor data to improve the detection of a fire anomaly with high accuracy and specificity, no false alarms, and quick detection time lags. The exemplary system and method contrast with memory-less fixed threshold detection systems typically employed in conventional fire detection systems.

[0052] In some embodiments, the trained neural network 106 is configured to generate a reconstructed signal, as the sensor model data set or signal 110, of the acquired signal data. The system 100a is configured to detect discrepancies, e.g., as the difference, between the measured signal and the reconstructed signal and employ those discrepancies as evidence of a fire anomaly. The system 100a can measure the discrepancy to establish the sufficiency of evidence of a fire anomaly as a score via smoothened operator to be compared to a score threshold that then triggers the output of an alarm or notification of a fire anomaly. [0053] In the example shown in Fig. 1, system 100a includes the sensor 102, the processor 104, a memory 112, a driver circuit 114, a speaker/hom 116, and a battery or power adapter 118 (shown as “Battery /Power Adapter” 118). The driver circuit 114 includes circuitries to interface the sensor 102 with the processor 104 and/or memory 112 and to interface the processor 104 with an output alarm (shown as a speaker/hom 116) comprising audible and visual outputs. In some embodiments, the output alarms include a network interface to generate a signal or a message to a local central controller or a global monitoring system. The output alarm may include audible outputs as noted as well as LEDs and/or strobe outputs. The battery or power adapter 118 includes a power converter to convert external electrical energy (e.g., AC input or DC input) from an external source to power the device 100a and to draw from an internal battery, e.g., when the external power is absent. In some embodiments, the battery or power adapter 118 is configured to recharge the batteries configured as a rechargeable energy storage module.

[0054] As used herein, processing unit and processor refers to a physical hardware device that executes encoded instructions for performing functions on inputs and creating outputs, including, for example, but not limited to, microprocessors (MCUs), microcontrollers, graphical processing units (GPUs), and application- specific circuits (ASICs). Thus, while instructions may be discussed as executed by a processor, the instructions may be executed simultaneously, serially, or otherwise executed by one or multiple processors.

[0055] The memory 112 includes instructions to execute a fire anomaly detection algorithm 120 (shown as 120a) comprising an input interface module 122 (shown as “Sensor Interface” 122), the trained neural network 106 (shown as “Neural Network Model” 106), a fire anomaly detector 124, and an output interface module 126 (shown as “Alarm Interface” 126). The sensor interface 122 is configured to interface to the sensor 102 to acquire the sensor signals 108 and then provide the sensor signals 108 to the neural network model 106 to generate the sensor model data set or signal 110.

[0056] The trained neural network 106 can be a variational autoencoder with a deep

LSTM network, a recurrent neural network, a convolutional neural network, a machine learning model, or a combination thereof. In the example shown in Fig. 1, the trained neural network 106 is configured as a “LSTM VAE” 106a. The LSTM VAE 106a includes an encoder to encode the input signal (e.g., 108, 108a) to determine/leam the most salient features or properties of the input signal. The LSTM VAE 106a includes a decoder rebuild the input signal based on the previously learned lower-dimensional representation of the signal. The trained neural network 106 serves as an intelligence layer of the processing that was trained in a non-fire condition using either local sensor data, e.g., acquired from the sensor 102 or a training data set of sensor data acquired from a set of sensor devices having a similar or same sensor type. In determining the features of the non-fire condition, the trained neural network 106 can reduce the complexity of the training operation and neural network infrastructure in trying to capture the lower-dimensional representation of the signal. In addition, the trained neural network 106 can be generated using a substantially reduced set of training data as compared to the usage of a data set having a fire event, which can have more complex set of scenarios.

[0057] The fire anomaly detector 124 (shown as 124a) is configured to determine the discrepancies between the measured sensor data 108 (shown as 108a) and the sensor model data 110 (shown as 110a) and establish the sufficiency of evidence of a fire anomaly. In operating with a trained neural network that is trained to detect for normal conditions within a building or structure, the fire anomaly detector 124a is configured to determine when the non-normal condition starts to appear within the sensor readings while taking account of historical non-fire events. [0058] In the example shown in Fig. 1, the fire anomaly detector 124a includes a comparator 128, a smoothed filter 130, and a detector module 132. The comparator 128, in some embodiments, is configured as a difference operator to provide a difference output 133 (shown as “Fire Score” 133) between current values of sensor signal value 108a and values of the sensor model data 110a corresponding to the same time period as a fire score. The smoothing filter 130 may be implemented as a Kalman filter to generate a smoothed score 134. The detector 132 may be implemented as a threshold operator to compare the smoothed score 134 to one or more thresholds 136 (not shown) to trigger an alert or notification.

[0059] Example Fire- Anomaly Neural Network Training System

[0060] Fig. 2A shows an example system 200 (shown as 200a) to configure the trained neural network (e.g., 106) of Fig. 1 to perform fire anomaly detection in accordance with an illustrative embodiment. Fig. 4A shows an example method 400 to train a neural network (e.g., of Fig. 2A) for anomaly detection (e.g., fire, smoke, heat, carbon monoxide, poor air quality, etc.) in accordance with an illustrative embodiment.

[0061] In the example shown in Fig. 2A, the system 200a includes a training module

202 that is operatively connected to a data store 204 that includes training data sets of measured sensor data from sensor devices 100a (shown as 100b)) or the like. The sensor devices 100b may include devices having sensors the same or similar to those of devices 100a. The sensor data 206 may include sensor data acquired in a non-fire condition. In certain embodiments, the data store 204 includes simulated data 208, e.g., from a simulation module 210, that is used for the training operation (e.g., by module 202). In yet other embodiments, the data store 204 includes publicly available data sets 212, e.g., those published in R. W. Bukowski, R. D. Peacock, J. D. Averill, T. G. Cleary, N. P. Bryner, and P. A. Reneke, “Performance of Home Smoke Alarms, Analysis of the Response of Several Available Technologies in Residential Fire Settings,” 2003, which is incorporated by reference herein in its entirety.

[0062] The training module 202 may generate the neural network detector 205 (e.g., executing the fire anomaly detection 120 (shown as 120b)) comprising a trained neural network (e.g., 106, 106a) using gradient descent methodologies. In some embodiments, once the neural network (e.g., 106, 106a) has been trained, the system 200a includes a validation module 210 configured to validate the neural network.

[0063] In the example shown in Fig. 4 A, method 400 includes training (402), via a neural network model, using sensor data acquire from a plurality of sensor devices, each having a similar or same sensor type as the sensor device, e.g., wherein the sensor data are acquired during a non-fire condition. Method 400 includes storing (404) the neural network model or a local model derived therefrom to a memory of the sensor device, wherein the trained neural network model or the local model derived therefrom is subsequently used to generate reconstructed sensor data during the operation of the sensor device, and wherein the reconstructed sensor data is compared to the sensor data to generate an alarm associated with the fire anomaly.

[0064] The training procedure may employ, for example, an unsupervised training using the Adam optimizer, e.g., as described in D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv: 1412.6980, 2014, which is incorporated by reference herein in its entirety.

[0065] LSTM-VAE fire detection method

[0066] Fig. 3 shows an example anomaly detection algorithm 300 (previously shown in Fig. 1 as 120a and in Fig. 2 A as 120b) configured with a trained neural network 106 (shown as FSTM VAE neural network 106a’) in accordance with an illustrative embodiment. In the example shown in Fig. 3, the anomaly detection algorithm 300 (e.g., for detection of fire, smoke, heat anomaly as described herein) includes (i) the trained FSTM-VAE neural network 106a’ comprising a variational autoencoder configured with deep FSTM networks 302 (shown as “Encoder FSTM layer 1” 302a, “Encoder FSTM layer 2” 302b, “Decoder FSTM layer 1” 302c, “Decoder FSTM layer 2” 302d) in an encoder 304 and decoder 306 and (ii) a denoising Kalman filter 308 implemented as a smoothing filter 130. The FSTM-VAE neural network 106a’ employs the encoder 304 to encode the input signal 108 (shown as 108b). During the training, the encoder 304 is configured to learn the most salient features or properties of the input signal 108b and, in the example of Fig. 3, includes the two FSTM layers 302a, 302b connected by a latent state 310 (shown as “Encoder latent state” 310a). [0067] The decoder 306 rebuilds the input signal 108b using a previously-learned lower-dimensional representation of the signal determined during the training operation to build a sensor model data 110 (shown as “reconstructed signal, y(t)” 110b, also referred to herein as a rebuild signal). In the example shown in Fig. 3, the decoder 306 includes two FSTM layers 302c, 302d connected by the latent state 310 (shown as “Decoder latent state” 310b). The encoder and decoder FSTM networks (e.g., 302a-302d) are trained using normal, non-fire operating conditions (or non-smoke, non-elevated heat, etc.) to have the FSTM-VAE learn the dynamics of the sensed quantity in non-anomalous conditions and subsequently to reconstruct the signal based on these learned non-fire condition dynamics (or other dynamics). Once configured, the trained FSTM-VAE 106a’ can be used to supervise the sensor signal 108b and provide its lower-dimensional representation as the rebuild signal

During run-time, the sensor signals 108b can be employed by the LSTM-VAE encoder 304 to determine the encoded mean ( ) 312a and standard deviation

312b to provide the hidden state 314 (shown as “VAE hidden state” layer 314). These hidden states (e.g., in layer 314) can include random variables sampled from a Gaussian distribution with the encoded mean (312a) and standard deviation (312b). [0068] Following signal reconstruction, in the example shown in Fig. 3, the anomaly detection algorithm 300 includes a difference operator 316 to compare the rebuild signal

( ) 110b and the sensor signal ^(^) 108b to determine a score 133 (shown as “Fire Score

133a) based on their difference per Equation 1 to provide estimation or likelihood value/score of an anomaly.

(Eq. 1) [0069] In Equation 1,

is the root mean square error (RMSE) of the training sequence and reflects or indicates the noise level in the training performance of the network. This training RMSE can be calculated using the training sensor signal and the LSTM-VAE reconstructed signal. [0070] The difference operator (or comparison operator) facilitates the use of a trained neural network configured using training data set acquired during normal, non-fire operating conditions. In doing so, the training data is substantially reduced in complexity as fire events can have a more complex set of scenarios while non-fire events have fewer. In being able to identify contextually or with historical data, the fire detection system can employ that model as normal baseline signal to determine when an anomaly that deviates from that normal baseline signal has occurred. [0071] Stated differently, under non-anomalous conditions, the LSTM-VAE or other neural network described herein, having learned these nominal environmental dynamics from the times series data, can accurately reconstruct the sensor signal (subject to the noise reflected in the RMSE), and the corresponding score (e.g., fire score, smoke score, etc.) is small. This condition changes when an anomaly occurs: the environmental dynamics as captured by the sensor can shift, and the signal reconstruction signal ) 110b would

increasingly diverge at every time step from the input signal 108b. The signal

reconstruction 110b would increasingly deviate from the sensor signal because the

former signal 108b is based on the learned environmental dynamics under the nominal conditions that trained LSTM-VAE, whereas the latter measured signal y(t) now measures different, anomalous (fire) dynamics. This topology provides a basis for anomaly detection with temporal depth that is in contrast to memoryless threshold-based detection. The growing difference between the measured signal and reconstructed signal

110b accumulates to provide evidence, via a measured score or metric (133a), that an anomaly has occurred. When the increase becomes sufficiently large and clearly distinguishable from the noise, as defined by a threshold, the fire alarm can then be triggered. [0072] Variational Autoencoder. As discussed above, the trained LSTM-VAE neural network 106a’ (also referred to as VAE 106a’) of Fig. 3 includes deep LSTM networks 302 (shown as “Encoder LSTM layer 1” 302a, “Encoder LSTM layer 2” 302b, “Decoder LSTM layer 1” 302c, “Decoder LSTM layer 2” 302d) in an encoder 304 and decoder 306. It is contemplated that other autoencoder architectures or neural network architectures or other detection techniques, including those described herein, may be employed to generate the sensor model data.

[0073] In the example shown in Fig. 3, the input layer within LSTM layer 302a of the sensor signal of the trained LSTM-VAE neural network 106a’ can be encoded as random variables in the hidden state by the encoder 304. The coded posterior distribution can be a diagonal Gaussian distribution, e.g., per Equation 2, and the encoder can map the input xit) into mean m(ΐ) and standard deviation ai t) values of the hidden distribution, e.g., per Equation 3. The hidden state z(t) can be sampled from the previously coded hidden distribution. The decoder 306 can reconstruct the signal x'it) based on the sampled hidden state, e.g., per Equation 4. The VAE loss function can then be calculated per Equation 5.

[0074] In Equation 5, the first part of the VAE loss function,

, is a sum square error (SSE) of the reconstruction x

) compared with the input layer

The second part, can be characterized as the Kullback-Leibler (KL)

divergence between and the standard normal distribution. The VAE may be

trained to provide an accurate reconstruction of the original signal and ensure the hidden state distribution is more likely to be Gaussian. [0075] The trained LSTM-VAE neural network 106a’ may be configured via an unsupervised learning operation to infer potentially complex distributions of the input layer. The VAE is an advanced ML anomaly detection method that is highly robust to signal noise. As noted previously, the use of LSTM can make the detection method more sensitive; however, it can also increase the false alarm rate [20]. The instant method and system may leverage this VAE to reduce the likelihood of false alarms. Further description of VAE may be found in [36].

[0076] Deep LSTM Network Encoder and Decoder. Fig. 3 shows the two-layer deep

LSTM networks (302a, 302b and 302c, 302d, respectively) implemented in each of the encoder 304 and the decoder 306 in the variational autoencoder 106a’. The parameters may be trained with the Adam optimizer [34]. Depth in autoencoder networks has been experimentally shown to provide significantly better performance than shallow autoencoders [32]. Additional or fewer LSTM layers may be implemented.

[0077] As discussed above, the input 318 of the encoder 304 includes the input signal

(y) 108b, and the outputs include the mean value m(ΐ) 312a and the standard deviation ait)

312b of the hidden Gaussian distribution. The input 320 of the two-LSTM-layer deep decoder 306 includes the sampled VAE hidden state 314, and the output is the reconstructed signal (y) 110b. For both the encoder and decoder (304, 306), the first LSTM layer (e.g., 302a, 302c) may map the input to a latent state, and this latent state serves as the input to the second LSTM layer for the output [26]. In the example shown in Fig. 3, the LSTM includes a hidden state h(t) 326, a cell state C(t) 324, and four gates: forget date 328, input gate 330, cell gate 332, and output gate 334.

[0078] The forget gate fit) (328) is configured to determine what the LSTM should forget from the previous cell state, C(t -1), as shown in Equation 6. The input gate i(t) (330) determines the updated value to the cell state C(t), as shown in Equation 7. The cell gate Cit) (332) uses current state input and the previous hidden state to generate a candidate for the new cell state, as shown in Equation 8. The output gate o(t) (334) determines the output and updates to the hidden state, h(t), as shown in Equation 9.

[0079] The updated hidden state h_t (326) and cell state G (324) are used to model the state of the current system and propagate to the next cell for the prediction of system future performance as per Equations 10 and 11.

[0080] LSTM is a recurrent neural network (RNN) that can operate on sequential data and capture dependencies over multiple time scales, including long past data. LSTM can learn with high accuracy the environmental dynamics of the sensed quantity under nominal conditions, and consequently, it is sensitive to shifts in these dynamics for detecting anomalies and triggering the fire alarm. Further description of LSTM may be found in [20, 22, 31, 32, 33]

[0081] Kalman Filter. To denoise the generated score 133 (e.g., fire score S(t) 133a), the anomaly detection algorithm 300 includes the Kalman filter 308 (e.g., as an example of a smoothing filter 130 or the like) and to produce a smoothed score 134 (shown as “Smoothed Fire Score 134a) to wit the detection decision (e.g., fire detection decision, smoke

detection decision, etc.) can be made. The smoothed fire score 134a may be used in a subsequent detection decision, e.g., comprising one or more score thresholds. The Kalman filter or other smoothen filters can mitigate the sensor noise distribution having small probability tail events with large values, which can cause large variations in the fire score and, in turn, increase false alarms.

[0082] Kalman filters may be employed for system state estimation and prediction with disturbance and measurement noise, e.g., per Equation 12, where is the state variable, u is the control input, w is the state disturbance, y is the measurement, and v is the measurement noise. The terms A, B, C are the system matrices for the state transition, control, and measurement.

(Eq. 12)

[0083] The noise can be assumed to belong to a centered Gaussian distribution with fixed covariance, R, and Q per Equation 13.

[0084] The parameters of the denoising Kalman filter include

and they may be automatically tuned by the expectation-maximization method, e.g., as described in [31]. The parameters of the model may be tuned to minimize the error of the posterior state estimation in which denotes the estimated state, and its covariance

given the measurements. The posterior state estimation may include five equations, per Equations 14-18, which can be characterized as the prior prediction and the correction update. The prediction process may propagate the current estimation to the next time step, e.g., per Equations 14 and Eq. 15, in which

and P¹ denote the prior measurement predictions of the state variable and error covariance, respectively.

[0085] Following each next measurement, the correction process may update the prior prediction, per Equations 16 and 17 in which K_t is the Kalman gain per Equation 18.

[0086] Kalman filtering employs linear quadratic estimation (LQE) that can employ a series of measurements observed over time and corrupted with statistical noise to estimate the future trend of the temporal signal. The exemplary method may use the Kalman filter to denoise the fire score in the exemplary fire detection method with the objective to lower a potential false alarm rate without compromising the improved sensitivity obtained by the LSTM-VAE. Further description may be found in [37, 38, 39].

[0087] Alarm Condition Based on Smoothed Fire Score. In the example shown in

Fig. 3, once the smoothed fire score 134a is generated, e.g., per the Kalman filter 308, the detector 132 (shown as 132a) may evaluate the smooth fire score 134a based on one or more thresholds to generate an alarm output or a pre-alarm output.

[0088] In some embodiments, two or more different thresholds may be established, e.g., based on the fire score and a corresponding confidence interval. The threshold(s) can be continuously varied. In one example, e.g., per Table 1, three thresholds may be established at 90%, 95%, and 99% confidence intervals, respectively.

[0089] Fire- Anomaly Neural Network Training Example #2

[0090] Fig. 2B shows another example system 200 (shown as 200b) to configure the trained neural network (e.g., 106, 106a) of Fig. 1 to perform anomaly detection (e.g., fire, smoke, heat, carbon monoxide, poor air quality) in accordance with an illustrative embodiment. Fig. 4B shows an example method 410 to train a neural network (e.g., of Fig. 2B) for fire anomaly detection in accordance with an illustrative embodiment. In the example shown in Fig. 2B, the system 200b includes the training module 202 and the data store 204, e.g., a described in Fig. 2A. Following the generation of the training module 202, system 202b includes a model reduction module 204 configured to generate a local model derived from the neural network (e.g., 106).

[0091] In some embodiments, the model reduction module 204 is configured to implement (i) the neural network 106 (shown as 106c), e.g., a recurrent neural network, a convolutional neural network, a machine learning model, and (ii) Kalman filter (e.g., 134) and/or other smoothing filters, using a fixed-point operator and data type. The reduction may allow the neural network algorithm (e.g., via module 106) and Kalman filter operations (e.g., via module 130) to be implemented or executed on a fixed-point microcontroller or the like (e.g., an 8-bit or 16-bit microcontroller). Example description of complexity reduction is described in Cotton, Nicholas J., Bogdan M. Wilamowski, and Gunhan Dundar. “A neural network implementation on an inexpensive eight bit microcontroller,” 2008 International Conference on Intelligent Engineering Systems. IEEE, 2008, which is incorporated by reference herein.

[0092] In the example shown in Fig. 4B, the method 410 includes training (402), via a neural network model, using sensor data acquired from a plurality of sensor devices, each having a similar or same sensor type as the sensor device, e.g., wherein the sensor data are acquired during a non-fire condition, e.g., as described in relation to Fig. 4A. Method 400 then includes validating 412 the trained neural network.

[0093] Method 410 then includes modifying (414) the neural network model (e.g.,

106) to operate on a fixed-point microcontroller. In some embodiments, the floating-point operators and floating-point data can be modified to fixed-point operators and fixed-point data. In some embodiments, look-up tables may be employed to allow for numerical estimation of certain floating-point operators.

[0094] Fire- Anomaly Neural Network Training Example #3

[0095] Fig. 2C shows yet another example system 200 (shown as 200c) to configure the generate the trained neural network (e.g., 106, 106a) of Fig. 1 to perform anomaly detection (e.g., fire anomaly, smoke anomaly, etc.) in accordance with an illustrative embodiment. Figs. 4C and 4D each shows an example method 420 and 430, respectively, to operate a building safety monitoring system 100c configured with a trained neural network for anomaly detection (e.g., fire anomaly, smoke anomaly, etc.) in accordance with an illustrative embodiment. In the example shown in Fig. 2C, the building safety monitoring system 100 (shown as 100c) includes a network interface 214 configured to operate with an external or global controller 216.

[0096] An external controller (e.g., 216) may be a local controller (e.g., a home or building automation controller) that operatively couples to the building safety monitoring system 100c through a local area network. The global controller (e.g., 216) may be a cloud infrastructure or server that operatively couples to the building safety monitoring system 100c through a wide-area network. In some embodiments, the network interface 214 is configured to communicate over a wired communication channel such as Ethernet or over a wireless communication channel such as ZigBee, WiFi, WiPan, and the like. In some embodiments, the network interface 214 is configured to communicate over a broadband cellular network such as 5G. The external controller and global controller 216 includes additional computing resource and/or training modules (e.g., 202) to retrain the trained neural network 106 using local data or measurements acquired at the building safety monitoring system 100c.

[0097] Cloud computing may comprise providing computing services via a network connection using dynamically scalable computing resources. Cloud computing may be supported, at least in part, by virtualization software. A cloud computing environment may be established by an enterprise and/or dynamically adjusted on an as-needed basis from a third-party provider. Some cloud computing environments may comprise cloud computing resources owned and operated by the enterprise as well as cloud computing resources hired and/or leased from a third-party provider.

[0098] It should be appreciated that the logical operations described above for the anomaly detector (e.g., fire anomaly detector, smoke anomaly detector, etc.) can be implemented (1) as a sequence of computer- implemented acts or program modules running on a computing system and/or (2) as interconnected machine logic circuits or circuit modules within the computing system. The implementation is a matter of choice dependent on the performance and other requirements of the computing system. Accordingly, the logical operations described herein are referred to variously as state operations, acts, or modules. These operations, acts, and/or modules can be implemented in software, in firmware, in special purpose digital logic, in hardware, and any combination thereof. It should also be appreciated that more or fewer operations can be performed than shown in the figures and described herein. These operations can also be performed in a different order than those described herein.

[0099] One or more programs may implement or utilize the processes described in connection with the presently disclosed subject matter, e.g., through the use of an application programming interface (API), reusable controls, or the like. Such programs may be implemented in a high-level procedural or object-oriented programming language to communicate with a computer system. However, the program(s) can be implemented in assembly or machine language, if desired. In any case, the language may be a compiled or interpreted language, and it may be combined with hardware implementations.

[0100] In the example shown in Fig. 2C, the anomaly detection algorithm 120 (shown as 120c) includes the neural network model (e.g., 106) and anomaly detector (e.g., 124), e.g., as described in relation to Fig. 2A. The fire anomaly detection algorithm 120c further includes a retraining/update module 218. In some embodiments, the module 218 is configured to update the neural network model 106 by providing an updated neural network from the external controller and global controller 216. The updated neural network may have been generated using the methods, e.g., described in relation to Fig. 2B.

[0101] In other embodiments, the module 218 is configured to direct the acquisition and storage of sensor measurements from sensor 102. The module 218 can direct the transmission of the acquired sensor measurements to the external or global controller 216 to retrain the trained neural network using local data or measurements acquired at the building safety monitoring system 100c. The module 218 can then receive the retrained neural network model from the external or global controller 216 and then update the neural network model 106 with the retained neural network.

[0102] The updated or retained neural network may have the same neural network architecture configured with adjusted weights, and revised weights may be transmitted for the update. In other embodiments, the updated or retained neural network may have an updated neural network architecture. [0103] In yet other embodiments, the module 218 is configured to direct the retraining of the trained neural network (e.g., 106). The module 218 may update the current train neural network with updated sensor measurements acquired locally by the system 100c. In the example shown in Fig. 2C, the system 100c includes a neural network training module 220 configured to perform the training. The neural network training module 220 may be triggered via an external command or signal or may be triggered via a user-input button located on the system 100c. In the example shown in Fig. 2C, the system 100c includes a user input button 222 that operatively connects to the processor 104 through the driver 114 to trigger a retraining command, e.g., when the system 100c is in a normal non-fire non-smoke condition. The system 100c may store a few minutes of data or 5-15 minutes of data, in one embodiment, to retrain the neural network 106.

[0104] As discussed above, Figs. 4C and 4D each shows an example method 420 and

430, respectively, to operate a building safety monitoring system 100c configured with a trained neural network for anomaly detection in accordance with an illustrative embodiment. Specifically, in Fig. 4C, the method 420 includes receiving (422), by a processor, sensor data from a sensor device (e.g., smoke alarm, fire alarm, carbon dioxide alarm, heat alarm, air quality monitoring alarm). The method 420 then includes generating (424), by the processor, via a trained neural network model (locally or globally trained), reconstructed sensor data using the sensor data. The method 420 then includes determining (426), by the processor, value associated with a presence or non-presence of a fire anomaly by comparing the reconstructed sensor data to the sensor data. The method 420 then includes outputting (428) an alarm based on a comparison of the determined value. Additional embodiments and examples of method 420 are described in relation to Fig. 2A.

[0105] Fig. 4D shows a method 430 to update the trained neural network of Fig. 2C in accordance with an illustrative embodiment. The method 430 includes storing (432), by the processor, e.g., within the local memory 112, local sensor data from a sensor device (e.g., smoke alarm, fire alarm, carbon dioxide alarm, heat alarm, air quality monitoring alarm) in a non-fire condition. The method 430 then includes retraining, via the processor, the trained neural network model (e.g., executing in the operation of Fig. 4C) using the stored local sensor data.

[0106] Experimental Results and Additional Examples

[0107] A study was conducted to evaluate the operation of the exemplary fire detection method and system. In the study, a trained LSTM-VAE neural network was employed and evaluated via simulation-based experiments as well as real-world fire and non fire datasets.

[0108] To evaluate the effectiveness of the exemplary method, the study conducted computational experiments with high-fidelity large eddy simulation (LES) data. The study also used real-world fire and non-fire datasets [29]. The study compared the exemplary fire detection with alternative methods, including standard LSTM detection [27], CUSUM fire detection [13, 14], exponentially weighted moving average (EWMA) anomaly detection [30], and fixed-temperature heat detection [7, 8, 9].

[0109] Specifically, the study performed a trained LSTM-VAE neural network in a set of simulation-based computational experiments with different fire and non-fire scenarios using Fire Dynamic Simulation (FDS) computational fluid dynamic (CFD) software [32] developed by National Institute of Standards and Technology (NIST). The study also evaluated a trained LSTM-VAE and other methods using real-world fire and non-fire datasets provided by the National Institute of NIST [19]. In the second part of the study, 69 real datasets were used, including 27 fire (flaming, smoldering, and cooking oil fires) and 42 nuisance non-fire experiments to evaluate alarm time lag, missed detection rates, false alarms rates, and FI scores for all the detection methods. An example real fire dataset includes data from a flaming chair experiment (SDC02) that NIST had conducted.

[0110] The study compared the exemplary LSTM-VAE ( Iv ) (at different confidence levels Ivi, lv2, and lv₃ ), along with alternative methods, including a standard LSTM (“Is") anomaly detection method [27]; a CUSUM (“ cu ") fire detection method [3, 4], an exponentially weighted moving average (EWMA) (“ew”) anomaly detection method [30], and two currently fixed-temperature detectors with thresholds set at 47°C and 58°C, respectively

The results indicate that neural network models, such as LSTM-VAE, can robustly outperform the other detection methods on all performance metrics with statistically significant shorter alarm time lags, no missed detection, and no false alarms. The study also observed that detection techniques such as CUSUM detection performance can degrade more so with sensor noise and can be prone to false alarm. The study also observed that the fixed heat detectors, while having no false alarm or low false alarm rate, have poor detection rates at least within the conditions of the conducted experiments.

[0111] Because of its high-performance results, the study further suggest that the exemplary LSTM-VAE and other neural networks as described herein can accommodate a host of various sensors, including smoke detectors, CO₂, and CO sensors as well as the integration of trained LSTM-VAE neural network with other sensor processing operation such as sensor data fusion. [0112] Fig. 5A shows an overview of a study 500 to evaluate the building safety monitoring system and methods (e.g., of Figs. 1-4) in accordance with an illustrative embodiment. Specifically, Fig. 5A shows a study 500 comprising a Fire Dynamic Simulation (FDS) computational fluid dynamic (CFD) software (shown using “NIST FDS Data” 502) and real-world fire and non-fire NIST datasets (shown as using “NIST Fire/Nuisance Data” 504). [0113] FDS Computational Experiments. The FDS computational experiments evaluated 4 test scenarios: a basic fire scenario (506); an obstacle fire scenario (508); a light perturbation but no fire scenario (510); and a medium perturbation but no fire scenario (512). In scenario “1” (506), the simulation includes a propane fire at the center part of the floor (514) of an adiabatic room (4m x 4m x 8m) with a temperature sensor (516) located at the center part of the ceiling. The initial and normal operating condition temperature inside the room is established at 20℃ and the propane fire is simulated with a reaction defined by 1(C₃H₈) + 4.81308 (O₂+3.7619 N₂) → 1(18.10631 N₂ + 2.81813 CO₂ + 3.9899 H₂O) + 0.20208 C_0.9H_0.1 [42], which releases heat, as fire signatures, and the

fire ignition occurs at ^ = 0c. Scenarios “2,” “3,” and “4” simulate obstructions in the room with the inclusion of (ii) a 2i × 2i × 2i adiabatic obstacle located at the center of the room (508), (iii) a human being person (modeled as a 1m x 1m x 1.8m volumetric heat source having a temperature at 37.5℃) in the windowed room (3m x 3m 20℃ constant temperature source) with no fire to evaluate a false alarm condition (510), and (iv) the same conditions as scenario “3” but with additional people (i.e., 4 people model) present in the room (512). To model real-world conditions, sensor signal noise (∈) is added as Gaussian noise

e~N(0,1), where m is the variance for three noise levels (0.4°C, 1.0°C, and 2.0°C). The study ran 100 simulations for each of the scenarios. [0114] The results of the simulation comprising the temperature profile at the sensor are shown as 506a, 508a, 510a, and 512a, respectively. The simulations were conducted over a 100-second period with a time step of 0.01s. As shown in the fire-event data (e.g., 506a, 508a, 512a), the temperature is continuously rising. Plots 506a and 508a are shown on the same scale to each other on the y-axis. And plots 510a and 512a are shown on the same scale to each other on the y-axis. The temperature disturbance is weak but still captured by the sensor.

[0115] Figs. 6A and 6B show performance results of the building safety monitoring system and methods based on simulation-based data. Specifically, Figs. 6A and 6B show the mean and standard deviation of the alarm lag time and accuracy for each of the assessed fire detection methods for scenarios “I” (fire) and “II” (fire in obstructed room), respectively. In Figs. 6A and 6B, it can be observed that the exemplary LSTM-VAE (702b) does trigger an alarm in the evaluated conditions along with the other evaluated detectors and does so earlier. [0116] The mean value of At refers to the overall alarm time lag and its sensitivity for a given evaluated method. Fire alarm lag time (mean) was determined and was calculated as refers to a time when the alarmed was triggered, and

is the

time the fire was ignited. The standard deviation of At reflects the stability or consistency of the fire detection method. The exemplary LSTM-VAE ( Iv ) along with alternative methods appears to capture a fire event, however, the exemplary LSTM-VAE appears to be able to do so more quickly.

[0117] Real-World and Non-Fire NIST Study. Table 2 shows the results from the conducted NIST data study, including average fire alarm lag time, missed detection rate, false alarm rate, and the FI score.

[0118] For the real-world fire and non-fire NIST dataset study, 69 real-world NIST datasets were employed in fire and non-fire experiments. The experiments were conducted for different types of fire: flaming (13 experiments), smoldering (11 experiments), and cooking oil fires (3 experiments).

Table 2

[0119] It can be observed from Table 2, that the exemplary LSTM-VAE

and performed the most robust (no missed detection or false alarms). The

exponentially weighted moving average (“ew”) algorithm was the next most robust (no missed detection but had several false alarms). Other detection methods had missed detections. The data suggest that the exemplary LSTM-VAE is quicker and has higher sensitivity and specificity than other evaluated methods, which may otherwise still be commercially viable.

[0120] ROC Curve. Fig. 5B shows performance results as a receiver operating characteristic (ROC) curve of the various detection methods in the real-world fire and non fire NIST dataset study. The plot was generated from a study of 69 NIST real-world datasets. It can be observed that the exemplary LSTM-VAE detection operation (518) has the highest sensitivity (y-axis) and specificity (x-axis). The results are shown in comparison to the LSTM (520), CUSUM (522), and EWMA (524).

[0121] NIST-DATA Fire and Non-Fire Experiment. The NIST data study also evaluated specific triggering time on specific NIST data sets, including of the temperature sensor signal (514) collected during a NIST flaming chair fire experiment (SDC02) [29]. [0122] Figs. 7A and 7B show performance results of the building safety monitoring system and methods based on real-world data comprising fire events and non-fire false alarm events. Specifically, Fig. 7A shows a time-profile result (e.g., detection score) for a fire event for the various detection operations, including the exemplary LSTM-VAE (702a), LSTM (704a), CUSUM (706a), EWMA (708a), and fixed thresholds (710a) are shown. In each of the plots, the alarm triggering point (712) and associated time are also shown. It can be observed that the exemplary LSTM-VAE (702b) does trigger an alarm in the evaluated conditions along with the other evaluated detectors and does so earlier.

[0123] For the evaluation, the data (e.g., 514 of Fig. 5A) included 515 temperature observations sampled at 0.5 Hz with the fire being ignited at t = 0s. The data (514) included a nominal/non-fire phase (temperature ~ 20°C), a fire ramp-up phase (temperature rises to 51.2°C at t = 206s), and a fire exhaustion phase (the temperature decreased) can be observed. The first 300 samples in the nominal phase were used to train both the LSTM-VAE and LSTM anomaly detection and to calculate/set the EWMA threshold [30]. The trained LSTM- VAE and other detection methods were then used to supervise the fire event after ignition (t = 0s) and to evaluate their performance.

[0124] Fig. 7B shows a time-profile result (e.g., detection score) for a non-fire event for the various detection operations, including the exemplary LSTM-VAE (702b), LSTM (704b), CUSUM (706b), EWMA (708b), and fixed thresholds (710b) are shown. In each of the plots, the alarm triggering point (712) and associated time are also shown. It can be observed that the exemplary LSTM-VAE (702b) does not trigger a false alarm in the evaluated conditions while false alarms are triggered with the LSTM (704b) and CUSUM (706b) associated detectors.

[0125] For the evaluation, the data (e.g., 516 of Fig. 5A) included MHN42 nuisance- alarm dataset provided by NIST. The nuisance dataset included 4500 data points sampled at a frequency of 5 Hz. The first 300 samples were used to train both the FSTM-VAE and FSTM anomaly detection and to calculate/set the EWMA threshold [30]. The trained ME models are used to supervise the temperature signal and to evaluate the possibility of false alarms.

[0126] Indeed, the results using both the simulation-based computations and the real- world fire and non-fire experiments are complementary, and they indicate that the LSTM-VAE robustly outperforms the other detection methods on all performance metrics with statistically significant shorter alarm time lags, no missed detection, and no false alarms.

[0127] Examples of various fire detection systems are disclosed in U.S. Patent No.

5450066, U.S. Patent No. 5966077, U.S. Patent No. US6166647, Chinese Patent No 105975991, and EP Patent No. EP3695392A1, which is incorporated by reference herein in their entirety. The exemplary system may be incorporated into the controls and operations of these systems in some embodiments.

[0128] Discussion

[0129] Background on fire detection. Significant advances have occurred in recent years in building automation and information systems. The operation of buildings has become more complex, and several of its functions require more effective and reliable monitoring of the environment within buildings (their internal state) increasingly. Fire detection, as a critical component of a building safety monitoring system, uses fire signatures such as smoke, heat, C02, or radiation to identify early signs of fire and trigger alarms. There can be significant costs associated with missed fire detection in terms of loss of life and property damage when signs of fire are not detected early enough to neutralize or contain the unfolding accident at an early stage. There can also be significant costs associated with false alarms if, for example, sprinklers are triggered unnecessarily, and water damage to the building occurs. Advanced fire detections use statistical models and optimization methods to improve detection accuracy and enhance the understanding of fire event development [1-5]. [0130] Effective anomaly detection in general and fire detection, in particular, remains a critical research area with substantial practical relevance [6]. Its focus is to improve, among other things, the sensitivity of the detection scheme and reduce its false alarm rate and missed detection. The sensitivity of fire detection stands for the ability to detect early signs of fire. The reduction of missed detection is related to sensitivity. For example, a highly sensitive detection system is capable of detecting early, small signatures of fire, having a low missed detection rate low. However, its drawback is that these small, detected signatures might be ambiguous and non-fire related, and as a result, the false alarm rate can be high. In short, a tradeoff is generally understood to mediate between these performance metrics of a fire detection system, its sensitivity on the one hand, and its false alarm rate (the complement of specificity) on the other handl.

[0131] Fire detection methods can be classified into two broad categories based on the alarm triggering mechanism. The first category is memoryless threshold-based detection [7-11]. In this category, the fire alarm decision is made based on the comparison between the present sensor signal and a pre-defined threshold value above which the alarm is triggered. Only the present sensor output is accounted for in this decision. For example, fixed temperature heat detectors [7, 8, 9] belong to this category. They use materials with different melting points to set different temperature thresholds to achieve different sensitivity. This first category is subsumed under the broader heading of point anomaly detection [12]. The second category is history-based fire detection. In this category, the fire alarm is triggered based on past and present sensor output, the information contained in or extracted from the time-series data, not just the sensor’s present output as in the previous, memoryless threshold category. One popular history-based method is the cumulative sum control chart (CUSUM) for fire detection [13, 14]. The CUSUM detector calculates a partial sum of the abnormal sensor signal and triggers the fire alarm when the sum exceeds a given threshold. In this case, there is a memory of past sensor outputs in the alarm triggering decision. The challenge for this category of fire detection methods is to probe the dynamics of the sensor’s output and extract meaningful features that are reliably predictive of fire occurrence. This second category is subsumed under the broader heading of contextual anomaly detection [12].

[0132] Conceptually, this second category of fire detection methods can be viewed as seeking to “accumulate evidence” over time before making a decision, whereas point anomaly detection methods operate with a single observation, e.g., “exhibit A,” as the primary and only evidence in support of the decision to trigger the fire alarm. The present disclosure can be characterized as belonging to the second category of methods of contextual anomaly detection: the objective is to leverage state-of-the-art machine learning tools to improve both the sensitivity and reliability of fire detection without compromising the false alarm rate. [0133] Machine learning for anomaly detection. There is a broader context within which instant work on fire detection is situated. It is related to advances in machine learning (ML) in general and unsupervised learning, in particular, for reliability and safety applications. Recent applications of Machine learning models in reliability engineering include methodology development, system diagnostic, remaining useful life estimation and prognostic health management [16-21].

[0134] Unsupervised learning includes examining datasets with only input variables or features, and no labels or response variable. Its general objective is to explore the feature space and find patterns in the dataset.

[0135] Two major sub-categories or tasks of unsupervised learning are based on the nature of the patterns sought: clustering and anomaly detection. Clustering consists in dividing the observations into clusters that share some similarities in the feature space. Anomaly detection consists in identifying unexpected observations in a dataset. The term anomaly in this ML contest is used in a broader sense than how it is understood in reliability and safety contexts.

[0136] Within this ML context, anomaly detection refers to “the problem of finding patterns in data that do not conform to expected normal behavior” [12]. Anomaly detection algorithms have found applications in many domains because they produce critical information that can be acted upon and prompt meaningful intervention. For example, anomaly detection is used in cyber- security and intrusion detection [22], in banking and insurance fraud detection, in a host of medical applications [23], and increasingly in reliability and safety applications, which is where the instant work on fire detection fits in. Anomaly detection is particularly well-suited for and used in early fault detection of equipment and structures. It is related to sensor data, and for industrial machinery and equipment, the data typically comes in a streaming fashion. Early detection of anomalies is essential in some contexts to prevent further damage and preempt catastrophic failures. The literature includes applications of anomaly detection in support of prognostic and health management (PHM) for different systems, for example, aircraft flight data recorders [24], industrial gas turbines [25], spacecraft operation and health monitoring [26, 27], and induction motors with a focus on ball-bearing faults [28]. Applications of anomaly detection algorithms for structural damage detection also abound, a discussion of which can be found in [12].

[0137] Herein, fire (as well as smoke, heat, carbon monoxide, radiation, poor air quality) can be considered as an anomaly in the monitored environment, and the system leverages advanced ML anomaly detection algorithms, Long-Short Term Memory (LSTM), and variational autoencoder (VAE) to improve the sensitivity and reliability of fire detection. An exemplary architecture is disclosed along with the supporting analytics for the next- generation fire detection system.

[0138] Although example embodiments of the present disclosure are explained in some instances in detail herein, it is to be understood that other embodiments are contemplated. Accordingly, it is not intended that the present disclosure be limited in its scope to the details of construction and arrangement of components set forth in the following description or illustrated in the drawings. The present disclosure is capable of other embodiments and of being practiced or carried out in various ways.

[0139] It must also be noted that, as used in the specification and the appended claims, the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise. Ranges may be expressed herein as from “about” or “ 5 approximately” one particular value and/or to “about” or “approximately” another particular value. When such a range is expressed, other exemplary embodiments include from the one particular value and/or to the other particular value.

[0140] By “comprising” or “containing” or “including” is meant that at least the name compound, element, particle, or method step is present in the composition or article or method, but does not exclude the presence of other compounds, materials, particles, method steps, even if the other such compounds, material, particles, method steps have the same function as what is named.

[0141] In describing example embodiments, terminology will be resorted to for the sake of clarity. It is intended that each term contemplates its broadest meaning as understood by those skilled in the art and includes all technical equivalents that operate in a similar manner to accomplish a similar purpose. It is also to be understood that the mention of one or more steps of a method does not preclude the presence of additional method steps or intervening method steps between those steps expressly identified. Steps of a method may be performed in a different order than those described herein without departing from the scope of the present disclosure. Similarly, it is also to be understood that the mention of one or more components in a device or system does not preclude the presence of additional components or intervening components between those components expressly identified. [0142] The term “about,” as used herein, means approximately, in the region of, roughly, or around. When the term “about” is used in conjunction with a numerical range, it modifies that range by extending the boundaries above and below the numerical values set forth. In general, the term “about” is used herein to modify a numerical value above and below the stated value by a variance of 10%. In one aspect, the term “about” means plus or minus 10% of the numerical value of the number with which it is being used. Therefore, about 50% means in the range of 45%-55%. Numerical ranges recited herein by endpoints include all numbers and fractions subsumed within that range (e.g., 1 to 5 includes 1, 1.5, 2, 2.75, 3, 3.90, 4, 4.24, and 5).

[0143] Similarly, numerical ranges recited herein by endpoints include subranges subsumed within that range (e.g., 1 to 5 includes 1-1.5, 1.5-2, 2-2.75, 2.75-3, 3-3.90, 3.90-4, 4-4.24, 4.24-5, 2-5, 3-5, 1-4, and 2-4). It is also to be understood that all numbers and fractions thereof are presumed to be modified by the term “about.”

[0144] The following patents, applications, and publications as listed below and throughout this document are hereby incorporated by reference in their entirety herein.

[1] X. Huang and L. Du, "Fire Detection and Recognition Optimization Based on Virtual Reality Video Image," IEEE Access, vol. 8, pp. 77951-77961, 2020.

[2] G. Xu, Q. Zhang, D. Liu, G. Lin, J. Wang, and Y. Zhang, "Adversarial adaptation from synthesis to reality in fast detector for smoke detection," IEEE Access, vol. 7, pp. 29471- 29483, 2019.

[3] Y. Xie, J. Zhu, Y. Cao, Y. Zhang, D. Feng, Y. Zhang, and M. Chen, "Efficient Video Fire Detection Exploiting Motion-Flicker-Based Dynamic Features and Deep Static Features," IEEE Access, vol. 8, pp. 81904-81917, 2020.

[4] K. Muhammad, J. Ahmad, I. Mehmood, S. Rho, and S. W. Baik, "Convolutional neural networks based fire detection in surveillance videos," IEEE Access, vol. 6, pp. 18174- 18183, 2018.

[5] G. Roque and V. S. Padilla, "LPWAN Based IoT Surveillance System for Outdoor Fire Detection," IEEE Access, vol. 8, pp. 114900-114909, 2020.

[6] Z. Liu and A. K. Kim, "Review of recent developments in fire detection technologies," Journal of Fire Protection Engineering, vol. 13, no. 2, pp. 129-151, 2003.

[7] Engineers, N. (n.d.). Heat detector design. Retrieved June 23, 2020, from https://www.ny- engineers.com/system-components/heatdetector- design

[8] Automatic Fire Detection: Opal Heat detector, fixed 78C. (n.d.). Retrieved June 23, 2020, from https://www.notifierfiresystems.co.uk/product.asp?id=574

[9] S. Nam, "Predicting response times of fixed-temperature, rate-of-rise, and rate- compensated heat detectors by utilizing thermal response time index," Fire safety journal, vol. 41, no. 8, pp. 616-627, 2006. [10] J. Geiman and D. T. Gottuk, "Alarm thresholds for smoke detector modeling," Fire Safety Science, vol. 7, pp. 197-208, 2003.

[11] M. Jackson and I. Robins, "Gas sensing for fire detection: Measurements of CO, C02, H2, 02, and smoke density in European standard fire tests," Fire safety journal, vol. 22, no. 2, pp. 181-205, 1994.

[12] V. Chandola, A. Banerjee, and V. Kumar, "Anomaly detection: A survey," ACM computing surveys (CSUR), vol. 41, no. 3, pp. 1-58, 2009.

[13] E. Zervas, A. Mpimpoudis, C. Anagnostopoulos, O. Sekkas, and S. Hadjiefthymiades, "Multisensor data fusion for fire detection," Information Fusion, vol. 12, no. 3, pp. 150- 159, 2011.

[14] O. Sekkas, S. Hadjiefthymiades, and E. Zervas, "A multi-level data fusion approach for early fire detection," in 2010 International Conference on Intelligent Networking and Collaborative Systems, 2010: IEEE, pp. 479-483.

[15] A. Tharwat, "Classification assessment methods," Applied Computing and Informatics,

2020.

[16] Z. Xu and J. H. Saleh, "Machine Learning for Reliability Engineering and Safety Applications: Review of Current Status and Future Opportunities," arXiv preprint arXiv:2008.08221 , 2020.

[17] Y. Shi, Z. Lu, R. He, Y. Zhou, and S. Chen, "A novel learning function based on Kriging for reliability analysis," Reliability Engineering & System Safety, vol. 198, p. 106857, 2020.

[18] J. Zhou, L.-Y. Zheng, Y. Wang, and C. Gogu, "A Multistage Deep Transfer Learning Method for Machinery Fault Diagnostics Across Diverse Working Conditions and Devices," IEEE Access, vol. 8, pp. 80879-80898, 2020.

[19] B. Chinomona, C. Chung, L.-K. Chang, W.-C. Su, and M.-C. Tsai, "Long Short-Term Memory Approach to Estimate Battery Remaining Useful Life Using Partial Data," IEEE Access, vol. 8, pp. 165419-165431, 2020.

[20] L. Ren, L. Zhao, S. Hong, S. Zhao, H. Wang, and L. Zhang, "Remaining useful life prediction for lithium-ion battery: A deep learning approach," IEEE Access, vol. 6, pp. 50587-50598, 2018.

[21] Z. Gao, C. Ma, J. Zhang, and W. Xu, "Enhanced Online Sequential Parallel Extreme Learning Machine and its Application in Remaining Useful Life Prediction of Integrated Modular Avionics," IEEE Access, vol. 7, pp. 183479-183488, 2019. [22] C.-F. Tsai, Y.-F. Hsu, C.-Y. Lin, and W.-Y. Lin, "Intrusion detection by machine learning: A review," expert systems with applications, vol. 36, no. 10, pp. 11994-12000, 2009.

[23] A. L. Beam and F S. Kohane, "Big data and machine learning in health care," Jama, vol. 319, no. 13, pp. 1317-1318, 2018.

[24] K. K. Reddy, S. Sarkar, V. Venugopalan, and M. Giering, "Anomaly detection and fault disambiguation in large flight data: A multi-modal deep auto-encoder approach," in Annual Conference of the Prognostics and Health Management Society, 2016, vol. 2016.

[25] W. Yan and L. Yu, "On accurate and reliable anomaly detection for gas turbine combustors: A deep learning approach," arXiv preprint arXiv: 1908.09238, 2019.

[26] S. Fuertes, G. Picart, J.-Y. Tourneret, L. Chaari, A. Ferrari, and C. Richard, "Improving spacecraft health monitoring with automatic anomaly detection techniques," in 14th International Conference on Space Operations, 2016, p. 2430.

[27] K. Hundman, V. Constantinou, C. Laporte, I. Colwell, and T. Soderstrom, "Detecting spacecraft anomalies using lstms and nonparametric dynamic thresholding," in Proceedings of the 24^th ACM SIGKDD international conference on knowledge discovery & data mining, 2018, pp. 387-395.

[28] T. Ince, S. Kiranyaz, L. Eren, M. Askar, and M. Gabbouj, "Real-time motor fault detection by 1-D convolutional neural networks," IEEE Transactions on Industrial Electronics, vol. 63, no. 11, pp. 7067-7075, 2016.

[29] R. W. Bukowski, R. D. Peacock, J. D. Averill, T. G. Cleary, N. P. Bryner, and P. A. Reneke, "Performance of Home Smoke Alarms, Analysis of the Response of Several Available Technologies in Residential Fire Settings," 2003.

[30] Z.-G. Zhou and P. Tang, "Improving time series anomaly detection based on exponentially weighted moving average (EWMA) of season-trend model residuals," in 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 2016: IEEE, pp. 3414-3417.

[31] F. A. Gers, N. N. Schraudolph, and J. Schmidhuber, "Learning precise timing with LSTM recurrent networks," Journal of machine learning research, vol. 3, no. Aug, pp. 115-143, 2002.

[32] I. Goodfellow, Y. Bengio, A. Courville, and Y. Bengio, Deep learning (no. 2). MIT press Cambridge, 2016.

[33] R. Pascanu, T. Mikolov, and Y. Bengio, "Understanding the exploding gradient problem," CoRR, abs/1211.5063, vol. 2, p. 417, 2012. [34] D. P. Kingma and J. Ba, "Adam: A method for stochastic optimization," arXiv preprint arXiv: 1412.6980, 2014.

[35] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, and L. Antiga, "Pytorch: An imperative style, high-performance deep learning library," in Advances in neural information processing systems, 2019, pp. 8026- 8037.

[36] C. Doersch, "Tutorial on variational autoencoders," arXiv preprint arXiv: 1606.05908, 2016.

[37] F. Aghili and C.-Y. Su, "Robust relative navigation by integration of ICP and adaptive Kalman filter using laser scanner and IMU," IEEE/ASME Transactions on Mechatronics, vol. 21, no. 4, pp. 2015-2026, 2016.

[38] T. Moore and D. Stouch, "A generalized extended kalman filter implementation for the robot operating system," in Intelligent autonomous systems 13: Springer, 2016, pp. 335- 348.

[39] N. Thacker and A. Lacey, "Tutorial: The kalman filter," Imaging Science and Biomedical Engineering Division, Medical School, University of Manchester, p. 61,

1998.

[40] T. K. Moon, "The expectation-maximization algorithm," IEEE Signal processing magazine, vol. 13, no. 6, pp. 47-60, 1996.

[41] K. McGrattan, S. Hostikka, R. McDermott, J. Floyd, C. Weinschenk, and K. Overholt, "Fire dynamics simulator user’s guide," NIST special publication, vol. 1019, no. 6, 2013.

[42] M. J. Hurley, D. T. Gottuk, J. R. Hall Jr, K. Harada, E. D. Kuligowski, M. Puchovsky, J. M. Watts Jr, and C. J. WIECZOREK, SFPE handbook of fire protection engineering. Springer, 2015.

[43] E. L. Lehmann and J. P. Romano, Testing statistical hypotheses. Springer Science & Business Media, 2006.

[44] Y. Sasaki, "The truth of the f-measure. 2007," ed, 2007.

[45] F. J. Massey Jr, "The Kolmogorov-Smimov test for goodness of fit," Journal of the American statistical Association, vol. 46, no. 253, pp. 68-78, 1951.

[46] J. Chambers, W. Cleveland, B. Kleiner, and P. Tukey, "Assessing distributional assumptions about data," Graphical Methods for Data Analysis, 1983.

[47] O. A. Grigg, V. Farewell, and D. Spiegelhalter, "Use of risk-adjusted CUSUM and RSPRTcharts for monitoring in medical contexts," Statistical methods in medical research, vol. 12, no. 2, pp. 147-170, 2003. [48] F. Gustafsson and F. Gustafsson, Adaptive filtering and change detection. Citeseer,

2000.

Claims

What is claimed is:

1. A method to detect a fire anomaly, the method comprising: receiving, by a processor, sensor data from a sensor device; generating, by the processor, via a trained neural network model or a local model derived therefrom, reconstructed sensor data using the sensor data; and determining, by the processor, a value associated with a presence or non-presence of a fire anomaly by comparing the reconstructed sensor data to the sensor data; wherein the determined value is used to generate an alarm associated with the fire anomaly.

2. The method of claim 1 further comprising: comparing, by the processor, the determined value associated with a presence or non presence of a fire anomaly to a pre-defined threshold, wherein an alarm or notification is generated based on the comparison.

3. The method of claim 1 further comprising: outputting, by the processor, a signal associated with the determined value associated with a presence or non-presence of a fire anomaly; and comparing, via an electronic circuit, the outputted signal to a pre-defined signal threshold, wherein an alarm or notification is generated based on the comparison.

4. The method of any one of claims 1-3, wherein the neural network model comprises a variational autoencoder with a deep LSTM network.

5. The method of any one of claims 1-3, wherein the neural network model comprises at least one of a recurrent neural network, a convolutional neural network, a machine learning model, and a combination thereof.

6. The method of any one of claims 1-5, wherein the determined value is denoised following the comparison between the reconstructed sensor data and the sensor data to produce a smoothed score associated with the fire anomaly.

7. The method of claim 6, wherein the determined value is denoised via a Kalman filter.

8. The method of any one of claims 1-7, wherein the trained neural network model was trained in a non-fire condition using sensor data from (i) the sensor device or (ii) a sensor device having a similar or same sensor type.

9. The method of any one of claims 1-8, wherein the trained neural network model is retrained based on an input command generated from a mechanical input located on the sensor device.

10. The method of any one of claims 1-9, wherein the trained neural network model is retrained at a remote computing system, the sensor device comprising a network interface that is operatively connectable over a network to the remote computing system.

11. The method of any one of claims 1-10, wherein the fire anomaly includes at least one of presence of smoke, presence of elevated heat, presence of fire, presence of carbon monoxide, and a combination thereof.

12. The method of any one of claims 2-11, wherein the pre-defined threshold is varied during operation of the sensor device.

13. The method of any one of claims 2-12 further comprising: comparing, by the processor, the determined value associated with a presence or non presence of a fire anomaly to a second pre-defined threshold, wherein the alarm or notification is generated based on the second comparison, wherein the second pre-defined threshold has an associated threshold time, and wherein the associated threshold time of the second pre-defined threshold is different from an associated threshold time of the pre-defined threshold.

14. A system comprising: at least one sensor configured to detect one or more fire-associated measurement; a processor operatively coupled to the at least one sensor; and memory having instructions stored thereon, wherein execution of the instructions by the processor causes the processor to: receive sensor data from a sensor device; generate, via a trained neural network model or a local model derived therefrom, reconstructed sensor data using the sensor data; and determine a value associated with a presence or non-presence of a fire anomaly by comparing the reconstructed sensor data to the sensor data; wherein the determined value is used to generate an alarm associated with the fire anomaly.

15. The system of claim 14, further comprising: a speaker operatively connected to the processor, wherein execution of the instructions by the processor further causes the processor to (i) compare the determined value associated with a presence or non-presence of a fire anomaly to a pre-defined threshold and (ii) direct an output of a signal to energize the speaker based on the comparison.

16. The system of claim 14 further comprising: a speaker operatively connected to the processor, wherein execution of the instructions by the processor further causes the processor to output a signal associated with the determined value associated with a presence or non-presence of a fire anomaly; and an electronic circuit configured to receive the signal and compare the signal to a pre defined signal threshold and direct an output to energize the speaker based on the comparison.

17. The system of any one of claims 14-16, wherein the neural network model comprises a variational autoencoder with a deep LSTM network, a recurrent neural network, a convolutional neural network, a machine learning model, or a combination thereof.

18. The system of any one of claims 14-17, wherein execution of the instructions by the processor further causes the processor to denoise or smoothen the determined value following the comparison between the reconstructed sensor data and the sensor data.

19. The system of any one of claims 14-18, wherein the trained neural network model was trained in a non-fire condition.

20. The system of any one of claims 14-19, further comprises: a housing; and a mechanical input located on the housing, the mechanical input being coupled to associated circuitry that is operatively coupled to the processor, wherein execution of the instructions by the processor further causes the processor to initiate, based on the mechanical input, retaining of the trained neural network model.

21. The system of any one of claims 1-10, wherein the at least one sensor includes: a photoelectric sensor or detector; a sampling tube smoke sensor or detector; a duct smoke sensor or detector; a carbon monoxide sensor or detector; an ionization sensor or detector; a temperature sensor; a resistance temperature detector (RTD) sensor; a thermistor sensor; an air quality sensor or poor quality detector; or a combination thereof.

22. The system of any one of claims 15-21 wherein execution of the instructions by the processor further causes the processor to compare the determined value associated with a presence or non-presence of a fire anomaly to a second pre-defined threshold, wherein the alarm or notification is generated based on the second comparison, wherein the second pre defined threshold has an associated threshold time, and wherein the associated threshold time of the second pre-defined threshold is different from an associated threshold time of the pre defined threshold.

23. A method to configure a sensor device to detect a fire anomaly, the method comprising: training, a neural network model, using sensor data acquired from a plurality of sensor devices each having a similar or same sensor type as the sensor device; storing the neural network model or a local model derived therefrom to a memory of the sensor device, wherein the trained neural network model or the local model derived therefrom is subsequently used to generate reconstructed sensor data during operation of the sensor device, and wherein the reconstructed sensor data is compared to the sensor data to generate an alarm associated with the fire anomaly.

24. The method of claim 23, further comprising any of the methods of claims 1-13.

25. A system comprising: at least one sensor configured to detect one or more fire-associated measurement; a processor operatively coupled to the at last one sensor; and memory having instructions stored thereon, wherein execution of the instructions by the processor causes the processor to perform any one of the methods of claims 1-13 or 23- 24.

26. A non-transitory computer-readable medium having instructions stored thereon, wherein execution of the instructions by a processor of a sensor device causes the processor to perform any one of the methods of claims 1-13 or 23-24.