Alarm system

Abstract

A detector (10) for an alarm system (100) is operable to monitor a monitored area for disturbances. The detector includes a first neural network that has been trained to distinguish between false alarms that should not cause an alarm to be triggered by the alarm system (100) in a normal operating mode and disturbances for which an alarm should be triggered in a normal operating mode. The detector (10) is configured to receive input from a user instructing the detector (10) to enter a training mode. The input includes a classification setting that instructs the detector (10) to mark subsequent disturbances recorded while the detector (10) is in the training mode as disturbances indicating a false alarm or as disturbances for which an alarm should be triggered in a normal operating mode.

Description

Technical Field

[0001] The present invention relates to a detector for an alarm system comprising a neural network, a training system for an alarm system and a method of training an alarm system. The neural network of the detector is trained to distinguish between false alarms that should not cause an alarm to be triggered by the alarm system in a normal operating mode and disturbances that should trigger an alarm in a normal operating mode. The present invention may be particularly useful for, but not limited to, intruder alarm systems for detecting intruders in a monitored area. Background Art

[0002] It is common for certain areas to be equipped with alarm systems. These areas include private homes, offices, shops or other commercial locations. When these areas are unattended, alarm systems are usually equipped to detect the presence of unwanted disturbances (such as intruders).

[0003] It is desirable to reduce the false alarm rate in alarm systems. False alarms may cause law enforcement to delay in being able to respond to real emergencies by dealing with false alarms. Some systems register with local law enforcement, and when the alarm is triggered, law enforcement will arrive. If false alarms are too frequent, these places may be removed from the registry, and law enforcement will no longer arrive when the alarm is triggered. Similarly, if the alarm is activated too often, neighbors may ignore it. If the system is a monitored system, and too many false alarms are triggered, there may also be costs. Users may also become reluctant to equip alarms due to the excessive number of false alarms. This then makes the area vulnerable to intruders.

[0004] It will be appreciated that it is also important to reduce the occurrence of failure to trigger an alarm when a real intruder is present.

[0005] Many traditional security systems use motion detectors that detect passive infrared (PIR) signals generated by body heat (infrared energy). "Passive" means that the system's sensors detect signals generated by other objects, but do not emit those signals themselves. Passive infrared sensors are widely used in home security systems. A typical system's sensors can detect heat and movement of heat sources in the area it monitors, dividing the area into protective "grids." If a moving object appears in too many grid zones and / or the infrared energy levels change rapidly across the grid zones, the sensor trips, triggering an alarm.

[0006] Other security systems utilize active signals, such as ultrasound or radio frequency fields. "Active" means that the sensor transmits a signal and detects a corresponding reflected signal. For example, a sensor can transmit radio pulses and measure the reflection from a moving object to detect its movement. Sensors can also emit infrared signals from LEDs following a similar principle.

[0007] Typical alarm systems, such as home alarm systems, are designed to detect human intruders in the monitored area. However, other disturbances may also trigger the alarm. For example, pets, robotic vacuum cleaners, moving fans, or similar devices may trigger the alarm. Movement of objects in the monitored area caused by airflow from the heating system, air conditioning system, or open windows / doors may also trigger the alarm. In addition, mechanical vibrations, electromagnetic disturbances, thermal airflow, and white light radiation can trigger the alarm.

[0008] Typical security systems are extensively tested before being sold to customers. Testing often involves simulated scenarios in a test area. For example, these simulations might be an intruder breaking into a house or store. The simulations are run in a controlled environment to allow the security system to detect the intruder and trigger an appropriate response.

[0009] Testing continues until the security system achieves the required level of accuracy. A certain level is required before the alarm can be certified for use. Testing of security systems can be a lengthy process, and simulated scenarios are often not a perfect representation of real-world situations. There may be many other environmental factors present when installing a security system that are not accounted for in simulated scenarios.

[0010] Therefore, there is a need for a detector and associated alarm system to achieve better accuracy at the installation site. It is also desirable to achieve this improved accuracy without the expense of increased testing time, and therefore it would be advantageous to provide a detector and associated alarm system that can be quickly brought to market. Summary of the invention

[0011] According to a first aspect, the present invention provides a detector for an alarm system, wherein the detector is operable to monitor a monitored area for disturbances, wherein the detector comprises a first neural network defined by a topology and a first set of parameters, wherein the first neural network has been trained to distinguish between false alarms that should not result in an alarm being triggered by the alarm system in a normal operating mode and disturbances for which an alarm should be triggered in the normal operating mode, wherein the detector is configured to receive an input from a user instructing the detector to enter a training mode, and wherein the input comprises a classification setting instructing the detector to mark subsequent disturbances recorded while the detector is in the training mode as disturbances indicating a false alarm or as disturbances for which an alarm should be triggered in the normal operating mode.

[0012] Thus, the user may set the detector to a training mode of operation. When doing so, the user may set a classification setting that is applicable to that particular session. The classification setting may either instruct the detector to mark subsequent disturbances recorded while the detector is in training mode as disturbances indicating false alarms ("false alarm" training mode), or as disturbances for which an alarm should be triggered in normal operating mode ("true alarm" training mode).

[0013] For example, a user may ensure that no one is inside a monitored area, making the monitored area inaccessible to people, and then may set the detector to a training mode. In this case, when setting the detector to a training mode, the user may set a classification setting to specify that any disturbance that occurs while the detector is in training mode should be flagged as indicating a false alarm (i.e., a disturbance that the alarm should not be activated when in a normal operating mode). In operation, this establishes a baseline of environmental factors that can potentially cause false alarms. These environmental factors may be specific to the particular monitored area in which the detector is installed.

[0014] Alternatively, the user can place the detector into training mode and can set the classification settings to specify that any disturbance that occurs while the alarm is in training mode should be flagged as a disturbance that would trigger the alarm in normal operating mode ("true alarm" training mode). The user can then simulate various activities that an intruder might perform (e.g., opening windows or doors and moving around the monitored area).

[0015] The detector may include a communication interface configured to send and receive information. The communication interface may be any wired or wireless interface. Examples of wireless interfaces include WiFi, Bluetooth, ZigBee, etc.

[0016] The detector may be configured to detect at least one of infrared radiation, sound waves, ultrasonic waves, and radio frequency waves. For example, the detector may be a PIR sensor. The detector may utilize active signals such as ultrasonic, infrared, or radar, and may emit a pulse signal and measure the reflection of a moving object in order to detect its movement.

[0017] Thus, the detector may comprise at least one transducer operable to receive at least one of infrared radiation, sound waves, ultrasound waves and radio frequency waves, and operable to convert the received quantity into an electrical signal.

[0018] The detector may include at least one transducer operable to generate at least one of infrared radiation, sound waves, ultrasound waves, and radio frequency waves.

[0019] The detector can convert the detected quantity into an electrical signal. These electrical signals can be processed to extract characteristic parameters from the electrical signals. For example, the characteristic parameters can include at least one of frequency, phase shift, time delay, maximum value, minimum value and average value, amplitude, phase spectrum component, intensity and duration.

[0020] The detector may include a processor operable to process electrical signals received by the detector.

[0021] In addition to the training mode, the detector may also be operable in a normal operating mode. The normal mode may correspond to the detector being armed, i.e., operable to detect disturbances and generate an alarm if appropriate (i.e., if the disturbance is not dismissed as a false alarm). Such a normal mode may be activated, for example, when the monitored space is secure and no one is in the monitored area. For example, in the case where the detector is set up in a home, the detector may be set to operate in the normal mode (i.e., the detector may be armed) when the resident leaves the home during the day or when the resident goes to bed at night.

[0022] The detector may also have an inactive mode (i.e., off mode). This mode is typically set when there are authorized personnel in the monitored area. For example, in the case where the detector is set up in a home, the detector may be set to operate in an inactive mode when the resident is present and active in the home during daytime hours.

[0023] In a normal operating mode, the detector can be configured to input characteristic parameters of the electrical signal into a neural network and calculate the neural network response to determine whether the detected disturbance is a false alarm (and can therefore be ignored without triggering an alarm), or whether the disturbance is potentially likely to be caused by an intruder such that an alarm should be triggered.

[0024] The detector can be configured to send characteristic parameters of the electrical signal and associated labels (i.e., marked as a false alarm, or marked as a disturbance that should trigger an alarm in normal operating mode, depending on the classification setting selected when setting the training mode) to the storage device when in training mode.

[0025] Alternatively, in the training mode, the detector is configured to input the characteristic parameters of the electrical signal into the neural network and calculate the neural network response to determine the "expected response" - that is, to determine whether the detected disturbance will be considered a false alarm when the alarm is operating in the normal mode, or whether the disturbance will be considered to be potentially caused by an intruder, so that the alarm should be triggered in the normal mode. The detector can then compare the "expected response" with the classification setting, and if the two are different, the characteristic parameters of the electrical signal and the associated label can be sent to the storage device only. For example, if the "expected response" is that the disturbance is potentially caused by an intruder, but the classification setting is "false alarm", the detector can send the characteristic parameters of the electrical signal and the associated label to the storage device. However, if the "expected response" is that the disturbance is a false alarm, and the classification setting is "false alarm", the detector may not send the characteristic parameters of the electrical signal and the associated label to the storage device.

[0026] The present invention extends to an alarm system comprising: a detector of the first aspect, optionally having any of the optional features described above; and a storage device configured to store characteristic parameters of the electrical signal and an associated label (i.e., labeled as a false alarm, or labeled as a disturbance for which an alarm should be triggered in normal operating mode, depending on the classification setting selected when setting the training mode).

[0027] The alarm system may include multiple detectors. Each detector may have its own corresponding neural network. Each detector may be trained individually (i.e., each detector may be placed in training mode independently of the other detectors). Alternatively, the system may be configured so that all detectors in the system are in training mode simultaneously.

[0028] One possibility is to include the storage device as part of the detector. However, this increases the size and cost of the detector. An alternative is to provide a storage device remote from the detector. The storage device and the detector may communicate with each other, for example, via a local network (wired or wireless). In this case, the storage device may be near the detector and the monitored area. Alternatively, the storage device and the detector may communicate with each other, for example, via the Internet. In this case, the storage device may be located at a great distance from the detector.

[0029] The storage device may be any suitable means for storage of information including, for example, a solid state drive (SSD), a hard disk drive (HDD), or a cloud service.

[0030] The storage device may include a communication interface configured to send and receive information. The communication interface may be any wired or wireless interface. Examples of wireless interfaces include WiFi, Bluetooth, ZigBee, etc.

[0031] The alarm system may include a user interface that allows a user to place the detector in a training mode. The user interface may allow the user to set a classification setting that indicates that subsequent disturbances recorded while the detector is in the training mode should be marked as disturbances indicating a false alarm, or as disturbances for which an alarm should be triggered in a normal operating mode (a "true alarm"). The user interface may be configured to be run by a user device having a communication interface configured to send and receive information.

[0032] The user device may be, for example, a mobile phone, a tablet computer, a laptop or desktop computer, a smart watch or any similar device. The user interface may include an application or software program run by the user device.

[0033] The alarm system may include a training system, wherein the training system includes a communication interface configured to send and receive information, wherein the training system includes a second neural network, the second neural network having the same topology as the first neural network and defined by a second set of parameters, wherein the second neural network has been trained to distinguish between false alarms that should not cause an alarm to be triggered by the alarm system in a normal operating mode and disturbances for which an alarm should be triggered in the normal operating mode, wherein the training system stores an existing training set of disturbances, the set of disturbances in the existing training set being labeled as disturbances indicating false alarms or as disturbances for which an alarm should be triggered in the normal operating mode of the alarm system ("true alarms"), wherein the training system is configured to: receive a set of newly labeled disturbances recorded by the detector; add the set of newly labeled disturbances to the existing training set of disturbances to form an updated training set; retrain the second neural network using the updated training set to generate an updated set of neural network parameters; and send the updated set of neural network parameters to the detector.

[0034] The training system can be provided remote from the detector (and monitored area). In this case, the training system can communicate with other components of the alarm system via the Internet.

[0035] The training system may include a processor and a storage medium, such as an SSD, HDD, or cloud service.

[0036] The "topology" of a neural network refers to the way the neurons of the neural network are connected. The first and second neural networks use multiple layers and various activation functions, including ramp or sigmoid functions. For example, the neural network can be a two-layer feedforward neural network with a unipolar activation function.

[0037] The alarm system may be an intruder alarm system trained to detect an intruder (eg, a human intruder) in a monitored area.

[0038] It is possible that the security system may be in "false alarm" training mode when a real disturbance caused by an intruder occurs. In this case, such an intrusion will be marked as a false alarm. If this data is subsequently used to retrain the neural network, the performance of the alarm system may be degraded. To avoid this, the user interface can have a review function, allowing the user to confirm that the recorded disturbances did not include any real disturbances. Only when the data has been verified in this way will the information be sent to the remote training system.

[0039] According to another aspect of the present invention, there is provided a method for training an alarm system, comprising:

[0040] installing a detector to monitor a monitored area for disturbances, wherein the detector comprises a first neural network defined by a topology and a first set of parameters, wherein the neural network has been trained to distinguish between false alarms that should not cause an alarm to be triggered by the alarm system in a normal operating mode and disturbances that should cause an alarm to be triggered in the normal operating mode;

[0041] instructing the detector to operate in a training mode;

[0042] instructing the detector to flag a disturbance according to a classification setting, wherein the classification setting indicates whether the disturbance should be flagged as a false alarm or as a disturbance for which an alarm should be triggered in a normal operating mode (a "true alarm");

[0043] Recording perturbations and labeling them according to the classification setting to generate a set of labeled perturbations;

[0044] sending the set of labeled perturbations to a training system, the training system comprising a second neural network having the same topology as the first neural network;

[0045] adding the set of labeled perturbations to an existing training set of perturbations to form an updated training set;

[0046] retraining the second neural network using the updated training set to generate an updated set of neural network parameters;

[0047] sending the updated set of neural network parameters to the detector; and

[0048] The first neural network parameter set is replaced in the first neural network with an updated neural network parameter set.

[0049] The training system may be provided remote from the detector (and monitored area). In this case, the training system may communicate with other components of the alarm system via the Internet.

[0050] Prior to sending the set of labeled perturbations to the training system, the set of labeled perturbations may be stored by a storage device, as disclosed above.

[0051] The alert system may be trained iteratively such that the second neural network is retrained multiple times, each time generating a new updated set of neural network parameters.

[0052] The alarm system can be retrained after each training session. For example, a user can run a training session after a false alarm or after a failure to detect an intruder.

[0053] Alternatively, the alarm system may be retrained when the set of flagged disturbances includes a number of disturbances exceeding a predetermined threshold.

[0054] The alarm system described above may be configured to perform the above method (optionally including any of the described optional features). The above method may be performed by the alarm system described above (optionally including any of the described optional features).

[0055] According to another aspect of the invention, a user interface for an alarm system is provided, wherein the user interface allows a user to place the alarm system into a training mode, and wherein the user interface allows the user to set a classification setting indicating that subsequent disturbances recorded while the detector is in said training mode should be marked as disturbances indicating a false alarm, or as disturbances for which the alarm should be triggered in a normal operating mode ("true alarms").

[0056] The user device may be, for example, a mobile phone, a tablet computer, a laptop or desktop computer, a smart watch or any similar device. The user interface may include an application or software program run by the user device.

[0057] According to another aspect of the invention, there is provided a training system for an alarm system, wherein the training system comprises a trained neural network having a parameter set, wherein the trained neural network has been trained to distinguish between false alarms that should not cause an alarm to be triggered by the alarm system in a normal operating mode and disturbances for which the alarm should be triggered in the normal operating mode, wherein the training system stores an existing training set of disturbances, the disturbances in the existing training set being marked as disturbances indicating false alarms or marked as disturbances for which the alarm should be triggered in the normal operating mode of the alarm system, wherein the training system is configured to: receive a set of newly marked disturbances recorded by a detector of the alarm system; add the set of newly marked disturbances to the existing training set of disturbances to form an updated training set; retrain the trained neural network using the updated training set to generate an updated set of neural network parameters; and send the updated set of neural network parameters to the detector.

[0058] Due to the learning capabilities of the detectors and / or associated alarm systems, embodiments of the present invention may allow for detectors and / or associated alarm systems that may achieve improved accuracy at the installation location. That is, the detectors and alarm systems may be effectively customized for the specific environment in which they are installed. The detectors and alarm systems may have the ability to learn a baseline of environmental factors that may potentially cause false alarms at a specific installation location. Moreover, the detectors and alarm systems may have the ability to allow a user to simulate various activities that may be performed by an intruder at a specific installation location. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Certain embodiments of the present invention will now be described, by way of example only, with reference to the accompanying drawings, in which:

[0060] Figure 1A security system is shown;

[0061] Figure 2 shows the steps performed by the detector; and

[0062] Figure 3 The steps performed by the training system are shown. DETAILED DESCRIPTION

[0063] Figure 1 A safety system 100 is shown. The system comprises a plurality of neural network based detectors 10, a storage device 20, a control device 30 and a remote training system 40.

[0064] In this case, the control device 30 is a mobile phone, but could for example be a tablet, laptop, desktop PC, smart watch or any similar device.The control device 30 comprises a user interface allowing the user to place the detector 10 into a training mode.

[0065] Figure 1 The security system 100 is intended for use in any type of area to be monitored. The system can be used in a user's home, alternatively it can be used in a public area or commercial location, such as an office or store or any other area where it would be desirable to detect unauthorized intruders. Thus, the user of the system 100 can be a resident of a residence or an owner of a commercial location.

[0066] The plurality of neural network-based detectors 10 may be any type of detector capable of detecting the presence of an intruder in the monitored area. For example, the detector may include infrared, radar, ultrasonic, acoustic or optical sensors. The plurality of neural network-based detectors 10 may also include a combination of different sensors, depending on the type of unwanted disturbance to be detected in the monitored area.

[0067] The plurality of neural network based detectors 10 are positioned throughout the monitored area. The number of neural network based detectors 10 required will depend on the size and configuration of the monitored area.

[0068] Typically, the monitored area may include multiple sub-areas, such as individual rooms within a home or office. Depending on the size of a particular sub-area, multiple neural network-based detectors 10 may be required to provide sufficient coverage to monitor the sub-area. Each of the individual neural network-based detectors 10 may include the same type of sensor. Alternatively, each sub-area may have multiple neural network-based detectors 10, each of which is configured to measure a different attribute. A specific sensor type may be selected with consideration of expected environmental factors in the monitored area.

[0069] In operation, the neural network-based detector 10 monitors disturbances in a monitored area. The neural network-based detector 10 includes a neural network that is trained to distinguish between false alarms (for which an alarm should not be triggered) and true disturbances (e.g., disturbances caused by an intruder). Prior to installation in a monitored area, the neural network present in the neural network-based detector 10 is trained under simulated scenarios. However, such scenarios are limited in number and do not take into account the specific conditions of the specific area in which the detector 10 is installed. Therefore, false alarms may occur.

[0070] After installation, and / or after a false alarm occurs, the user may decide to retrain the neural network present in the neural network based detector 10 in the system 100 .

[0071] The user places the neural network based detector 10 into a training mode using the control device 30. The user then sets the classification setting to either "false alarm" or "true alarm", depending on whether the user wishes to train the neural network based detector 10 to learn the background environmental conditions in the monitored area (i.e., learn "false alarms") or whether the user wishes to simulate a "true alarm".

[0072] The purpose of the "false alarm" training mode is to record data on ambient conditions within the monitored area and determine any routine disturbances that may occur when the security system is activated. This allows the system 100 to learn expected or acceptable disturbances that should not trigger an alarm.

[0073] If the user sets the classification setting to "false alarm", the user should ensure that the monitored area is safe and that no one is present in the monitored area. The user can then run the system 100 in a training mode, allowing the neural network-based detector 10 to accumulate data on any disturbances that occur. Such disturbances may be caused by pets (e.g., cats or dogs), moving household devices (e.g., a robotic vacuum cleaner or a rotating fan), or any object that moves due to air flow, thermal air flow, white light, electromagnetic or mechanical disturbances, etc.

[0074] The electrical signal measured by the neural network based detector 10 is processed to determine characteristic parameters of the electrical signal. The characteristic parameters of the accumulated disturbances are marked as "false alarms" and sent to the storage device 20.

[0075] From the storage device 20 , characteristic parameters of the accumulated disturbance and “false alarm” labels are sent to the remote training system 40 .

[0076] The remote training system 40 includes a training neural network having the same topology as the neural network of the neural network-based detector 10. The remote training system 40 stores an existing training set of perturbations, the perturbations in the existing training set are labeled as "false alarms" or "true alarms", and the trained neural network has been trained to distinguish between these. The remote training system 40 receives a set of newly labeled perturbations recorded by the neural network-based detector 10 of the alarm system and adds these perturbations to the existing training set. The remote training system 40 then retrains the trained neural network using the updated training set to generate an updated set of neural network parameters. The updated set of neural network parameters is then sent to the neural network-based detector 10.

[0077] The neural network based detector 10 then benefits from the retrained neural network which has been retrained using the data accumulated during the "false alarm" training mode.

[0078] The user may perform this process after installing the alarm system 100. The user may repeat this process until they are satisfied with the accuracy of the alarm system 100. The user may also repeat the process if the general conditions in the monitored area change. For example, the user may run the training mode if a new pet is present in the monitored area, or if a new mobile device is used in the area. In addition, the user may repeat the training process after a false alarm or after the system fails to detect a real intruder.

[0079] The purpose of the True Alarm training mode is to simulate and record the activities that an intruder might perform, such as entering through a door or window, and moving through the monitored area. The specific location of doors, windows, etc. in the monitored area may affect how sensitive the alarm is to an intruder, and therefore the True Alarm training mode can be used to improve accuracy in detecting intruders.

[0080] If the user sets the classification setting to "True Alarm", then the user should simulate intruder activity, allowing the neural network based detector 10 to accumulate data on the disturbances that occurred.

[0081] The electrical signal measured by the neural network based detector 10 is processed to determine characteristic parameters of the electrical signal. The characteristic parameters of the accumulated disturbance are marked as "true alarms" and sent to the storage device 20.

[0082] From the storage device 20 , characteristic parameters of the accumulated disturbance and a “true alarm” label are sent to the remote training system 40 .

[0083] The remote training system 40 receives the set of newly labeled perturbations and adds them to the existing training set. The remote training system 40 then retrains the trained neural network using the updated training set to generate an updated set of neural network parameters. The updated set of neural network parameters is then sent to the neural network-based detector 10.

[0084] The neural network based detector 10 then benefits from the retrained neural network which has been retrained using the data accumulated during the "true alarm" training mode.

[0085] In this example, the neural network based detector 10, the storage device 20 and the control device 30 communicate via a local WiFi network. The remote training system 40 communicates with the network based detector 10 and the storage device 20 via the Internet.

[0086] In some cases, the training pattern will not cover every possible disturbance type, and therefore false alarms may still occur.

[0087] Figure 2 A method for training the security system of the present invention is shown as performed by the neural network based detector 10 and the storage device 30. Using the example of a house, once the security system 100 has been installed in the house (which may involve positioning a neural network based detector in each room of the house), the user activates the training mode via the user interface on the control device 30 at block 1000.

[0088] The user sets the training mode classification at box 2000. In this example, the user selects the classification as "false alarm." In the example of a house, the user would request training mode and false alarms when leaving the house during the day.

[0089] The neural network based detector 10 will then record disturbances throughout the day (block 3000). The disturbances may be caused by normal ambient conditions within a particular house.

[0090] This collected data will then be stored in a storage device as "false alarm" data (block 4000).

[0091] This newly stored data will then be sent to the remote training system 40 (at block 5000).

[0092] When the user returns home, the training mode is deactivated using the user interface on the control device 30 .

[0093] The additional steps of the method of training the alarm system 100 are Figure 3 As shown in—these steps are performed by the remote training system 40.

[0094] The remote training system 40 receives the set of labeled perturbations at block 6000 and adds the set of labeled perturbations to the existing training set (block 7000). Then, at block 8000, the neural network is retrained using the updated training set to generate an updated set of neural network parameters. The updated set of neural network parameters is then sent to the detector (block 9000).

[0095] It is possible that the security system may be in "false alarm" training mode when a real disturbance caused by an intruder occurs. In this case, such an intrusion will be marked as a false alarm. If this data is subsequently used to retrain the neural network, the performance of the alarm system may be degraded. To avoid this, the user interface can have a review function, allowing the user to confirm that the recorded disturbances did not include any real disturbances. Only when the data has been verified in this way will the information be sent to the remote training system.

Claims

1. An alarm system comprising a detector, wherein the detector is operable to monitor a disturbance in a monitored area, wherein the detector comprises a first neural network defined by a topology and a first set of parameters, wherein the first neural network has been trained to distinguish between false alarms that should not cause an alarm to be triggered by the alarm system in a normal operating mode and disturbances that should cause an alarm to be triggered in the normal operating mode, wherein the detector is configured to receive input from a user instructing the detector to enter a training mode, and wherein the input includes a classification setting instructing the detector to mark subsequent disturbances recorded while the detector is in the training mode as disturbances indicating a false alarm or as disturbances for which the alarm should be triggered in the normal operating mode, The alarm system further comprises: a storage device configured to store characteristic parameters of the marked disturbed electrical signal recorded by the detector, wherein the storage device comprises a communication interface configured to send and receive information, A remote training system, wherein the remote training system comprises a communication interface configured to send and receive information, wherein the remote training system comprises a second neural network, the second neural network having the same topology as the first neural network and defined by a second set of parameters, wherein the second neural network has been trained to distinguish between false alarms that should not cause an alarm to be triggered by the alarm system in a normal operating mode and disturbances that should trigger an alarm in the normal operating mode, wherein the remote training system stores an existing training set of disturbances, the set of disturbances in the existing training set being marked as disturbances indicating false alarms or as disturbances that should trigger the alarm in the normal operating mode of the alarm system, wherein the remote training system is configured to: receiving a set of newly labeled perturbations recorded by the detector; adding the set of newly labeled perturbations to the existing training set of perturbations to form an updated training set; retraining the second neural network using the updated training set to generate an updated set of neural network parameters; and The updated set of neural network parameters is sent to the detector.

2. The alarm system of claim 1, wherein the detector includes a communication interface configured to send and receive information.

3. An alarm system as claimed in claim 1 or 2, wherein the detector is configured to detect at least one of infrared radiation, sound waves and radio frequency waves, and is configured to convert the detected quantity into the electrical signal.

4. An alarm system as claimed in claim 3, wherein the sound waves are ultrasonic waves.

5. The alarm system of claim 3, wherein the detector is configured to extract the characteristic parameter from the electrical signal.

6. An alarm system as claimed in claim 5, wherein in a normal operating mode, the detector is configured to input the characteristic parameters of the electrical signal into the neural network and calculate the neural network response to determine whether the detected disturbance is a false alarm or whether the alarm should be triggered.

7. An alarm system as claimed in claim 5 or 6, wherein when in the training mode, the detector is configured to send the characteristic parameters of the electrical signal to a storage device.

8. The alarm system of claim 1 , comprising a user interface that allows the user to place the detector in a training mode, and wherein the user interface allows the user to set a classification setting that indicates whether subsequent disturbances recorded while the detector is in the training mode should be marked as disturbances indicating a false alarm or as disturbances for which the alarm should be triggered in the normal operating mode, Wherein the user interface is configured to be executed by a user device having a communication interface configured to send and receive information.

9. The alarm system of any one of claims 1, 2 and 8, wherein the alarm system is an intruder alarm system trained to detect an intruder in the monitored area.

10. A method for training an alarm system, comprising: installing a detector to monitor a monitored area for disturbances, wherein the detector comprises a first neural network defined by a topology and a first set of parameters, wherein the neural network has been trained to distinguish between false alarms that should not cause an alarm to be triggered by the alarm system in a normal operating mode and disturbances that should cause an alarm to be triggered in the normal operating mode; instructing the detector to operate in a training mode; instructing the detector to flag a disturbance according to a classification setting, wherein the classification setting indicates whether the disturbance should be flagged as a false alarm or as a disturbance for which an alarm should be triggered in a normal operating mode; recording perturbations and labeling them according to the classification setting to generate a set of labeled perturbations; sending the set of labeled perturbations to a remote training system, the remote training system comprising a second neural network having the same topology as the first neural network; adding the set of labeled perturbations to an existing training set of perturbations to form an updated training set; retraining the second neural network using the updated training set to generate an updated set of neural network parameters; sending the updated set of neural network parameters to the detector; as well as The first neural network parameter set is replaced in the first neural network with the updated neural network parameter set.

11. The method of claim 10, wherein the alert system can be iteratively trained such that the second neural network can be retrained multiple times, each time generating a new updated set of neural network parameters.

12. A method as claimed in claim 10 or 11, wherein the alarm system can be retrained after each training session, or wherein the alarm system can be retrained when the set of marked disturbances includes a number of disturbances exceeding a predetermined threshold.

Images