TECHNICAL FIELD

[0001] The present disclosure belongs to the technical field of smoke detectors, and more particularly to smoke detector for identification of smoke detectors obstructed of light in an ambient.

BACKGROUND

[0002] A key aspect of fire protection is to identify a developing fire emergency in a timely manner, and to alert the building's occupants and fire emergency organizations. This is the role of fire detection and alarm systems. These systems also self-monitor, identifying where within the building(s) alarms originate from and detecting when errors occur in wiring and connections that may hinder the system from working correctly.

[0003] In essence, a fire alarm system has four key functions: detect, alert, monitor and control. These sophisticated systems use a network of devices, appliances, and control panels to carry out these four functions.

[0004] The way a fire alarm system detects a fire is through its initiating devices, which detect smoke or a fire. These devices include smoke detectors of various kinds, heat detectors of various kinds, sprinkler water flow sensors, and pull stations.

[0005] Automatic fire detection systems are used to detect fires and to trigger an alarm in case of a detected fire. In order to be able to ensure the operational capability of automatic fire detection devices at all times, there is a need for manual or automatic testing of the fire detection device.

[0006] Document DE102012201703 providing an automatic fire detector for detecting fires, having a housing, wherein the housing comprising a measuring chamber for detecting smoke particles, having a sensor system for detecting a measured variable for evaluating the operational capability of the automatic fire detector, having an evaluation device for evaluating the operational capability of the automatic fire detector on the basis of the measured variable, the sensor system comprising at least one flow sensor for detecting a flow as the measured variable for evaluating the operational capability of the automatic fire detector.

[0007] Document EP2624229B1 providing an invention relates to an automatic fire detector for detecting fires, having a housing which comprises a measuring chamber for detecting smoke particles, having a sensor system for detecting a measured variable for evaluating the operational capability of the automatic fire detector and having an evaluation device for evaluating the operational capability of the automatic fire detector on the basis of the measured variable.

[0008] Document US9959748B2 describes a system and method for the monitoring and trending the rate at which fire detection devices get dirty. This information is used to determine which devices are clogged or getting clogged and to establish that the chambers are open to air flow because they are accumulating dirt over time. Air flow through the detection chamber is proven using this analysis. Further self-testing is also employed for the fire detection devices by including modules that simulate the smoke interference with the light. This can be accomplished in two ways. In one example, light from the chamber light source can be reflected toward the scattered light photodetector to simulate alarm conditions. In another example, an additional chamber light source can be added to the detection chamber that can generate light to simulate alarm conditions.

[0009] Document EP2270762B1 discloses an invention related to a smoke detector with a housing which has smoke passage openings and comprises a smoke detector and an alarm signalling device. The invention further relates to a method for checking the contamination of smoke passage openings of a smoke detector with a housing.

[0010] These facts are disclosed in order to illustrate the technical problem addressed by the present disclosure. GENERAL DESCRIPTION

[0011] The present disclosure discloses a smoke detector arranged to identify when obstructed in an ambient, comprising: a dark chamber with one or more inlet openings for receiving smoke of the ambient; a light-sensor arranged within the dark chamber; and an electronic data processor arranged to: capture an electrical signal output of the light-sensor for sampling the light o over a period of time; process the captured light sample signal with a pretrained machine-learning model to determine whether the light sample signal matches one or more predetermined light patterns of an obstructed smoke detector; determine a warning if the captured light sample is determined to be at least in part comprising a predetermined light pattern of an obstructed smoke detector.

[0012] The fire alarm panel is connected to the system's initiating devices through either 2- or 4-wire circuits. This circuitry allows the control panel to monitor the state of its initiating devices, usually by zones, identifying whether the devices are in normal or alarm mode.

[0013] When a fire starts, the smoke or heat will activate one of the initiating devices, or someone will activate the manual pull station, alerting the fire alarm system to the fire and putting it in alarm mode. The control panel shows these readings on its display panel.

[0014] Properly installed and maintained smoke alarms are one of the best and least expensive means of providing an early warning of a potentially deadly fire and could reduce by almost half the risk of dying from a fire.

[0015] It is also disclosed a smoke detector arranged to identify when obstructed in an ambient, comprising: a dark chamber with one or more inlet openings for receiving the smoke of the ambient; a light-sensor arranged within the dark chamber; and an electronic data processor arranged to: capture an electrical signal output of the lightsensor for sampling the light of the over a period of time; process the captured light sample signal with a pretrained machine-learning model to determine whether the light sample signal matches one or more predetermined light patterns of an obstructed smoke detector; determine a warning if the captured light sample signal is determined to be at least in part comprising a predetermined light pattern of an obstructed smoke detector.

[0016] In an embodiment, said light patterns are periodical daily patterns over said period of time.

[0017] In an embodiment, the electronic data processor comprises a performance core configured to carry out light processing tasks and a low-power core configured to carry out smoke detection tasks.

[0018] In an embodiment, the electronic data processor is further arranged to input the captured light sample signal to the pretrained machine-learning model by aggregating the captured electrical signal output for a subperiod of time comprised within said period of time.

[0019] In an embodiment, the sub-period of time is an hour.

[0020] In an embodiment, the period of time is twenty-four hours.

[0021] In an embodiment, the electronic data processor is further arranged to feed the captured light sample signal to the pretrained machine-learning model by inputting the aggregated captured electrical signal output to an embedding neural network for increasing dimensionality.

[0022] In an embodiment, the light-sensor is a photoresistor, photodiode and/or phototransistor and/or wherein the light-sensor is a light sensor of a light-scattering detection arrangement.

[0023] In an embodiment, the pretrained machine-learning model is an artificial neural network.

[0024] In an embodiment, the pretrained machine-learning model is an artificial recurrent neural network, a convolutional neural network, and/or a long short-term memory.

[0025] It is also disclosed a computer-based method for providing a warning for identification of a smoke detector obstructed in an ambient, said smoke detector comprising a dark chamber with one or more inlet openings for receiving the light of the ambient, a light-sensor arranged within the dark chamber and an electronic data processor, said method comprising the steps of: capturing by the electronic data processor an electrical signal output of the light-sensor for sampling the light over a period of time; processing by the electronic data processor the captured light sample signal with a pretrained machine-learning model to determine whether the light sample signal matches one or more predetermined light patterns of an obstructed smoke detector; determining by the electronic data processor a warning if the captured light sample signal is determined to be at least in part comprising a predetermined light pattern of an obstructed smoke detector. In an embodiment, the light patterns are periodical daily patterns over said period of time.

[0026] In an embodiment, said method further comprising inputting the captured light sample signal to the pretrained machine-learning model by aggregating the captured electrical signal output for a subperiod of time comprised within said period of time, in particular aggregating by averaging the captured electrical signal output.

[0027] In an embodiment, the method comprises the sub-period of time of an hour.

[0028] In an embodiment, the method comprises the period of time of the twenty-four hours.

[0029] In an embodiment, the method comprises feeding the captured light sample signal to the pretrained machine-learning model by inputting the aggregated captured electrical signal output to an embedding neural network for increasing dimensionality.

[0030] In an embodiment, the method comprises the pretrained machine-learning model is an artificial neural network, in particular the pretrained machine-learning model being an artificial recurrent neural network, a convolutional neural network, and/or a long short-term memory.

BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The following figures provide preferred embodiments for illustrating the disclosure and should not be seen as limiting the scope of invention.

[0032] Figure 1: Schematic representation of an embodiment for considering pitched ceiling limitations to place a detector (in National Fire Alarm and Signalling Code). [0033] Figure 2: Schematic representation of the process behind the identification of obstructed detectors by using Al models.

[0034] Figure 3: Schematic representation of an embodiment approach of Al models' accuracy for identification of obstructed detectors.

[0035] Figure 4: Schematic representation of an embodiment of a model pipeline.

[0036] Figure 5: Schematic representation of an embodiment approach of values forthe 24 features of the optl for an obstructed and an unobstructed detector.

[0037] Figure 6: Schematic representation of an embodiment of LSTM model architecture.

[0038] Figure 7: Schematic representation of an embodiment of CNN model architecture.

DETAILED DESCRIPTION

[0039] The present disclosure relates to a smoke detector for identification of smoke detectors obstructed of light in an ambient spot.

[0040] A smoke alarm should be installed and maintained according to the manufacturer's instructions. When installing a smoke alarm, many factors influence where you will place the alarm, including how many are to be installed. Consider placing alarms along the escape path to assist in egress in limited-visibility conditions. In general, alarms should be placed in the centre of a ceiling or, if placed on a wall, should be installed not more than 12 inches away from the ceiling. If ceiling is pitched, the alarm should be installed within 36 inches of the peak but not within the apex of the peak (4 inches down from the peak) as shown in Figure 1.

[0041] In general, smoke alarms must not be placed within:

300mm of a corner of a ceiling and a wall

300mm of a light fitting

400mm of an air-conditioning vent

400mm of the blades of a ceiling fan [0042] To have a fully operational fire alarm system, with all the conditions to work as expected, it is important that these and all the installation's rules are verified. Nowadays, only an in-situ inspection could verify if those requirements are fulfilled. It is thus a need for an automatic process capable to alert for undesired situations that could put at risk the buildings and people's safety.

[0043] Using Artificial Intelligence (Al) it is possible to learn from data and provide an accurate model capable to detect automatically if a detector is obstructed. This solution has the advantage that we can easily and quickly detect a problem which could be the root cause of some system faults or false alarms. This type of solution also avoids the expenses and disturbances caused by the in-situ inspections, which sometimes are very difficult to schedule with the customer and technical teams.

[0044] Optical smoke detectors use the scattered-light method. A LED transmits light to the measuring chamber, where it is absorbed by the labyrinth structure. In the event of a fire, smoke enters the measuring chamber, and the smoke particles scatter the light from the LED. The amount of light hitting the photo diode is converted into a proportional electrical signal.

[0045] High and simultaneous variations on optical sensor's data are expected in a fire occurrence, otherwise optical values tend to follow a linear pattern, being variations around 10 units considered a normal behaviour.

[0046] An obstructed detector will not be able to precisely check for smoke particles, having thus a risk to not work properly and be responsible for some system fault. With the right data and technology, it is possible to create accurate and robust models capable to automatic identify the obstructed detectors. The main idea behind this solution (Figure 2) consists of using a test system installed in a controlled environment and cover some of the detectors. It is expected that obstructed and unobstructed detectors report different optical values over time. Just looking to the data sometimes is not easy to identify these differences, but with the right technology and software, it is possible to prepare the data for the application of Machine (ML) and Deep Learning (DL) models. These models are capable to learn from data and identify the "hidden" patterns behind the data. [0047] The first step to create the desired solution was to install a fire alarm test system, in a controlled environment. The smoke detectors were installed following all the installation rules' requirements, but to have valid data to train Al models and create a robust solution, half of the installed sensors were obstructed for 2 months. After that period some of those covers were removed and used in other sensors for the same timeperiod.

[0048] Data collected from these experiments were split in training and test datasets for Al models application. Al models were trained in a training set to recognize certain types of patterns. They use various types of algorithms to reason over and learn from data, with the overarching goal of providing a reliable classification algorithm capable to distinguish between obstructed and unobstructed detectors.

[0049] Random Forest (RF), Support Vector Machines (SVM), Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) are some of the Al models used to develop our idea.

[0050] RF are an ensemble learning method (use multiple learning algorithms to obtain better predictive performance) for classification, regression and tasks that operates by constructing multiple decision trees at training time. For classification tasks, the output of the random forest is the class selected by most trees. For example, each tree in the classifications takes input from samples, features are then randomly selected, and are used in growing the tree at each node. In addition, every tree in the forest should not be pruned until the end of the task when the prediction is reached with a decision.

[0051] As major advantages this model reduces overfitting problem in decision trees and the variance which improves the accuracy, is usually robust to outliers and can handle them automatically. It is also a very stable algorithm - if a new data point is presented, the overall algorithm is not affected, it can impact one tree, but it is very hard to influence all the trees.

[0052] In machine learning, SVM are supervised learning models with learning algorithms that analyse data for classification and regression analysis. In brief, a supportvector machine creates a hyperplane or set of hyperplanes in a high- or infinitedimensional space, which can be used for classification, regression, or tasks such as outlier's detection. A separation is achieved by the hyperplane that has the largest distance to the nearest training data point of a class. The SVM provides possible kernels, thus, we can choose a function which is not necessarily linear and can have different forms in terms of different data.

[0053] SVM can be used for data that is not regularly distributed and have unknown distribution, generally avoid overfitting and performs well when there is a clear separation between classes. SVM can handle high dimensional data.

[0054] A CNN is a class of artificial neural network (ANN). CNN are regularized versions of multilayer perceptron and multilayer perceptron, typically are fully connected networks, where each neuron in one layer is connected to all neurons in the next layer. In addition, is typically composed of three types of layers (or building blocks): convolution, pooling, and fully connected layers. The convolution and pooling layers perform feature extraction, whereas the fully connected layer, maps the extracted features into final output, as for example classification. The architecture consists of repetitions of a stack of several convolution layers and a pooling layer. Training a CNN network is a way of finding kernels in convolution layers and weights in fully connected layers that minimize differences between output predictions and the ground truth labels on a training set. Also, a model performance through kernels and weights is considered using a loss function through forward propagation on a training dataset. The learnable parameters, kernels and weights, are updated according to the loss value through an optimization algorithm called backpropagation and gradient descent, where the goal is to minimize the loss.

[0055] CNN have accomplished great achievements across a variety of domains in particular, the CNN model allows the use of a global average pooling. The advantages of applying global average pooling are: (1) the capacity to reduce the number of learnable parameters and (2) enables the CNN to accept inputs of variable size.

[0056] LSTM is an artificial recurrent neural network (RNN) architecture used in deep learning. On the contrary of the standard feedforward neural networks, LSTM has feedback connections. In neural networks we have a stack of layers containing nodes, where the input data (features) feed the nodes of the input layers and the information is combined as weights and passed to the next layer, until arrives the output layer. At the end, the expected output (target) is compared with the model's output and the weights are updated. During the training process of a network, the goal is to minimize loss (e.g. in terms of error). The gradient is calculated, which is, loss with concerning a particular set of weights, the weights are adjusted, and the process is repeated until the loss is minimum for an optimal set of weights. This process is called backtracking.

[0057] LSTM networks are well-suited to perform predictions based on time series data, since there can be lags of unknown length between events in a time series. LSTM also handles noise and continuous values.

[0058] All the tested models show great performance and be able to automatically identify patterns, reaching accuracies above 88% (figure 3). These results show the promising solution to identify the obstructed smoke detectors using Al models. This identification followed by a trouble event and a fire panel message will help to provide timely information to the maintenance teams to work on predictive solutions, allowing to resolve the problems before they arise. The connection between fire panel, remote maintenance and apps makes possible to provide an efficient maintenance routine and apply for fast solutions.

[0059] This invention is supported under the Safe Cities Project Innovation (POCI-Ol- 0247-FEDER-041435) which ends in 2022.

[0060] The following pertains to Data collection. To achieve the obstructed state, it was put a cover on the detectors (covered detector represents an obstructed one). The data forthe obstructed detectors model was recorded in the Bosch test system. It is collected data from 15 devices, of which 7 are covered and 8 are uncovered. The state (covered or uncovered) of each detector changes in time. This means that the cover in the devices is put on or taken off. The date in which these modifications are done is recorded to later label the recorded data.

[0061] The metric recorded is the optical (optl). The sensors register the data within intervals of 15 to 90 minutes.

[0062] The following pertains to data pre-processing. Like mentioned above, the data is received in intervals of about 15 to 90 minutes. In order to make the frequency between timestamps equal it was done an aggregation of the values to hourly intervals using the mean to calculate the value of the metric.

[0063] Then it was performed a conversion of each hour of the day to separate features, resulting in 24 features for each day of data. Different data arrangements were tested but it was concluded that this achieved the best results. In Figure 5 it is possible to see one observation of the optl metric from an obstructed and an unobstructed detector. In the x-axis it is represented the hours of the day (features) and in the y-axis the value that corresponds to each feature.

[0064] Finally, the data was labelled as obstructed or unobstructed. In order to train and test the models it was performed an 80/20 and an 70/30 train-test data split. The models were later tested with both distributions to assess which produced the best results.

[0065] The following pertains to LSTM model. Long short-term memory (LSTM) is an architecture of a recurrent neural network (RNN) that can preserve information from the beginning of the sequence and carry it forward. Unlike a normal RNN, the LSTM contains recurrent units that allows to process sequence data.

[0066] LSTM work well when performing classification tasks on time series data, since there can be a variable duration between events in a time series.

[0067] The LSTM model developed is composed of an embedding layer, a LSTM layer and a dense layer to give the output of the model. The embedding layer allows to convert the features into a vector representation of all the features. The model is then trained based on this vector.

[0068] The parameters of each of the layers are computed using a hyperparameter optimization framework. The parameters that are optimized are the activation function and number of neurons of the LSTM layer and the batch size and number of epochs of the model.

[0069] The optimal parameters depend on the data that is being train to the model. To give an example, for the 80/20 train-test split the optimal parameters are:

Activation function: ReLU

Number of neurons: 120

Batch size: 50 Number of epochs: 10

[0070] In Figure 6 it is possible to check the model architecture and parameters. The parameters that are described as Optuna are the ones that are optimized with the optimization framework.

[0071] The following pertains to CNN model. Convolutional neural network (CNN) is a class of artificial neural network (ANN). The CNN has three main types of layers: convolutional layer, pooling layer and fully-connected layer. A model can have several convolutional and pooling layers but only one fully-connected layer (last layer). Each convolutional layer processes increasingly more complex information (starts by processing easier features). The pooling layers allow to reduce the dimensional complexity of the model while keeping the most significant information. The fully- connected layer maps the extracted features into a final output.

[0072] Like in the LSTM model, the CNN starts by the embedding step. The data is then fed to three pairs of convolutional and pooling layers. Finally, a dense layer outputs the prediction of the model.

[0073] The max pooling layer summarizes the presence of features in the input sequence. Pooling layers provide an approach to down sampling feature maps. The output after a max-pooling layer would be the feature map containing the most prominent features of the previous feature map.

[0074] The term "comprising" whenever used in this document is intended to indicate the presence of stated features, integers, steps, components, but not to preclude the presence or addition of one or more other features, integers, steps, components or groups thereof.

[0075] The disclosure should not be seen in any way restricted to the embodiments described and a person with ordinary skill in the art will foresee many possibilities to modifications thereof. The above-described embodiments are combinable. The following claims further set out particular embodiments of the disclosure.