Detecting a break-in attempt using an autoencoder

Abstract

It is presented a method for enabling detecting a break-in attempt. The method comprises: obtaining (40) training acceleration data from at least one accelerometer (10) to detect vibrations of an openable barrier (15), when the openable barrier is subjected to harmless use; deriving (42) a plurality of candidate events, wherein each candidate event comprises event acceleration data being extracted from the training acceleration data, wherein the start of each candidate event is determined based on a section of the training acceleration data having an amplitude that is greater than a first threshold amplitude; determining (44) when a sufficient amount of candidate events has been obtained; selecting (46) a set of training events, yielding a set of selected training events comprising at least some of the candidate events; and training (48) a machine learning, ML, model based on the set of selected training events, wherein the ML model is an autoencoder.

Description

DETECTING A BREAK-IN ATTEMPT USING AN AUTOENCODERTECHNICAL FIELD

[0001] The present disclosure relates to the field of detecting a break-in attempt based on accelerometer data, and in particular to detecting a break-in attempt based on accelerometer data using an autoencoder.BACKGROUND

[0002] It is a continuous problem with break-ins in homes and commercial properties. Various sensors exist in the prior art to detect such break-ins. Some sensors detect when a window or door is opened or when glass is broken, and other sensors detect movement.

[0003] One type of such sensor is based on accelerometers. These are used for detecting vibrations that occur when a break-in attempt occurs. In this way, an alarm can be raised prior to major structural damage occurring. However, it is very difficult to find the balance between an acceptable harmless activities and actual break-ins. False alarms are very stressful and result in undermined trust of the alarm system. On the other hand, a missed detection of a break-in is even worse, since the whole point of such a sensor is to detect break-ins.

[0004] Additionally, many break-ins occur when residents are home. It would be of great benefit if there would be a possibility to be able to detect break-ins even if one or more windows or doors are open. Indoor motion sensors are quite restricted in their use when people are home.SUMMARY

[0005] One object of embodiments presented herein is to improve the ability to distinguish a break-in attempt from harmless events.

[0006] According to a first aspect, it is provided a method for enabling detecting a break-in attempt, the method being performed by a break-in detector. The method comprises: in a training phase, performing the following: obtaining training acceleration data fromat least one accelerometer configured to detect vibrations of an openable barrier, when the openable barrier is subjected to harmless use; deriving a plurality of candidate events, wherein each candidate event comprises event acceleration data being extracted from the training acceleration data, wherein the start of each candidate event is determined based on a section of the training acceleration data having an amplitude that is greater than a first threshold amplitude; determining when a sufficient amount of candidate events has been obtained; selecting a set of training events, yielding a set of selected training events comprising at least some of the candidate events; and training a machine learning, ML, model based on the set of selected training events, wherein the ML model is an autoencoder

[0007] The method may further comprise: in an operational phase, performing the following: obtaining operational acceleration data from at least one accelerometer configured to detect vibrations of an openable barrier; determining that an operational event has started by the operational acceleration data exceeding a threshold amplitude, the operational event comprising operational acceleration data starting from the start of the operational event; calculating a reconstruction error by comparing operational acceleration data of the operational event that is input to the ML model with output data from the ML model; and detecting a break-in attempt based on the reconstruction error being greater than an anomaly threshold

[0008] The accelerometer that is used as source for the training acceleration data may be the same as the accelerometer that is used as a source for the operation acceleration data.

[0009] The selecting a set of training events may comprise: evaluating, for each one of the candidate events, how different the candidate event is from any events that are already part of the set of selected training events; and selecting, for each one of the candidate events, the event to form part of the set of selected training events, based on the event being sufficiently different from any events that are already part of the set of training events.

[0010] The evaluating how different the event is may comprise comparing at least one parameter of the event with any events that are already part of the set of selectedtraining event, wherein the at least one parameter is selected from the group consisting of duration, minimum amplitude, maximum amplitude, variance and standard deviation.[oon] The determining when a sufficient amount of candidate events has been obtained may be based on determining that a variation indicator of the event acceleration data of the candidate events is greater than a variety threshold.

[0012] In the deriving a plurality of candidate events, the end of each candidate event may be determined based on the training acceleration data having an amplitude that is below a second threshold amplitude.

[0013] The method may further comprise, prior to the training phase: selecting a base instance of an autoencoder based on characteristics of the openable barrier; wherein the base instance is a starting point for the autoencoder for which training is performed.

[0014] According to a second aspect, it is provided a break-in detector for enabling detecting a break-in attempt. The break-in detector comprises: processing circuitry; and memory circuitry storing instructions that, when executed by the processing circuitry, cause the break-in detector to: in a training phase, execute instructions to: obtain training acceleration data from at least one accelerometer configured to detect vibrations of an openable barrier, when the openable barrier is subjected to harmless use; derive a plurality of candidate events, wherein each candidate event comprises event acceleration data being extracted from the training acceleration data, wherein the start of each candidate event is determined based on a section of the training acceleration data having an amplitude that is greater than a first threshold amplitude; determine when a sufficient amount of candidate events has been obtained; select a set of training events, yielding a set of selected training events comprising at least some of the candidate events; and train a machine learning, ML, model based on the set of selected training events, wherein the ML model is an autoencoder;

[0015] The instructions may further comprise instructions that, when executed by the processing circuitry, cause the break-in detector to: in an operational phase, executethe instructions to: obtain operational acceleration data from at least one accelerometer configured to detect vibrations of an openable barrier; determine that an operational event has started by the operational acceleration data exceeding a threshold amplitude, the operational event comprising operational acceleration data starting from the start of the operational event; calculate a reconstruction error by comparing operational acceleration data of the operational event that is input to the ML model with output data from the ML model; and detect a break-in attempt based on the reconstruction error being greater than an anomaly threshold

[0016] The accelerometer that is used as source for the training acceleration data may be the same as the accelerometer that is used as a source for the operation acceleration data.

[0017] The instructions to select a set of training events may comprise instructions that, when executed by the processing circuitry, cause the break-in detector to: evaluate, for each one of the candidate events, how different the candidate event is from any events that are already part of the set of selected training events; and select, for each one of the candidate events, the event to form part of the set of selected training events, based on the event being sufficiently different from any events that are already part of the set of training events.

[0018] The instructions to evaluate how different the event is may comprise instructions that, when executed by the processing circuitry, cause the break-in detector to compare at least one parameter of the event with any events that are already part of the set of selected training event, wherein the at least one parameter is selected from the group consisting of duration, minimum amplitude, maximum amplitude, variance and standard deviation.

[0019] The instructions to determine when a sufficient amount of candidate events has been obtained may be based on determining that a variation indicator of the event acceleration data of the candidate events is greater than a variety threshold.

[0020] The instructions to derive a plurality of candidate events may comprise instructions that, when executed by the processing circuitry, cause the break-in detectorto determine the end of each candidate event based on the training acceleration data having an amplitude that is below a second threshold amplitude.

[0021] The break-in detector may further comprise instructions that, when executed by the processing circuitry, cause the break-in detector to, prior to the training phase: select a base instance of an autoencoder based on characteristics of the openable barrier; wherein the base instance is a starting point for the autoencoder for which training is performed.

[0022] According to a third aspect, it is provided a computer program for enabling detecting a break-in attempt. The computer program comprises computer program code which, when executed on a break-in detector causes the break-in detector to: in a training phase, execute instructions to: obtain training acceleration data from at least one accelerometer configured to detect vibrations of an openable barrier, when the openable barrier is subjected to harmless use; derive a plurality of candidate events, wherein each candidate event comprises event acceleration data being extracted from the training acceleration data, wherein the start of each candidate event is determined based on a section of the training acceleration data having an amplitude that is greater than a first threshold amplitude; determine when a sufficient amount of candidate events has been obtained; select a set of training events, yielding a set of selected training events comprising at least some of the candidate events; and train a machine learning, ML, model based on the set of selected training events, wherein the ML model is an autoencoder.

[0023] According to a fourth aspect, it is provided a computer program product comprising a computer program according to the third aspect and a computer readable means comprising non-transitory memory in which the computer program is stored.

[0024] Generally, all terms used in the claims are to be interpreted according to their ordinary meaning in the technical field, unless explicitly defined otherwise herein. All references to "a / an / the element, apparatus, component, means, step, etc." are to be interpreted openly as referring to at least one instance of the element, apparatus, component, means, step, etc., unless explicitly stated otherwise. The steps of anymethod disclosed herein do not have to be performed in the exact order disclosed, unless explicitly stated.BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Aspects and embodiments are now described, by way of example, with reference to the accompanying drawings, in which:

[0026] Fig 1 is a schematic diagram showing an environment in which embodiments presented herein can be applied;

[0027] Figs 2A-C are schematic diagrams illustrating embodiments of where the break-in detector 1 can be implemented;

[0028] Figs 3A-B are flow charts illustrating embodiments of methods for enabling detecting a break-in attempt;

[0029] Fig 4 is a schematic graph illustrating how candidate events are determined in conjunction with acceleration data, e.g. in embodiments covered by the flow chart of Figs 3A-B;

[0030] Fig 5 is a schematic diagram illustrating components of the break-in detector of Figs 2A-B; and

[0031] Fig 6 shows one example of a computer program product 90 comprising computer readable means.DETAILED DESCRIPTION

[0032] The aspects of the present disclosure will now be described more fully hereinafter with reference to the accompanying drawings, in which certain embodiments of the invention are shown. These aspects may, however, be embodied in many different forms and should not be construed as limiting; rather, these embodiments are provided by way of example so that this disclosure will be thorough and complete, and to fully convey the scope of all aspects of invention to those skilled in the art. Like numbers refer to like elements throughout the description.

[0033] According to embodiments presented herein, it is enabled to input acceleration data from an accelerometer to a machine learning (ML) model to detect when a break-in is in progress. The ML model is an autoencoder. The training of the ML model is achieved by capturing acceleration data when the openable barrier is not subjected to any break-in attempts, deriving events from the acceleration data, and training the ML model based on these events. The training can be unsupervised, since the events that occur during the training are not break-in attempts, whereby no labelling is needed. In operation, a break-in attempt can be detected as an anomaly by the trained autoencoder, since the ML model is trained on only harmless situations. The unsupervised training enables simple yet customised provisioning, while providing reliable detection of break-in attempts. This enables an alarm system to be constantly active, even when occupants are in the building, greatly reducing, or even eliminating, the risk of harmless events triggering the alarm.

[0034] Fig 1 is a schematic diagram showing an environment in which embodiments presented herein can be applied. Access to a physical space 6 is restricted by a physical openable barrier 5. Optionally, the openable barrier is also selectively controllable to be in a locked state or an unlocked state. The openable barrier 5 can be a door, window, window door, gate, hatch, cabinet door, drawer, etc. The openable barrier 5 is provided in a surrounding structure 8 (being a wall, fence, ceiling, floor, etc.) and is provided between the restricted physical space 6 and an accessible physical space 7. The surrounding structure 8 forms part of a building. It is to be noted that the accessible physical space 7 can be a restricted physical space in itself, but in relation to this openable barrier 5, the accessible physical space 7 is accessible. A handle 3 is provided on the barrier to allow a person to open and close the barrier.

[0035] In order to unlock the barrier 5, a lock 4 is optionally provided. The lock 4 can be a traditional mechanical lock or an electronic lock. It is to be noted that the lock 4 can be provided in the openable barrier 5 (as shown) or in the surrounding structure 8 (not shown).

[0036] An accelerometer 10 is provided to detect vibrations in the openable barrier 5 and / or in the surrounding structure 8. The accelerometer 10 can detect vibrations in three geometric dimensions (X, Y and Z), providing acceleration data containing threecomponents corresponding to the three geometric dimensions. The structure in which vibrations are detected can be the openable barrier 5 and / or surrounding structure 8. In this way, as explained in more detail below, in operation, it can be determined if the acceleration data from the accelerometer 10 indicates a harmless event or a break-in attempt. The accelerometer 10 can form its own sensor device, or the accelerometer can form part of the lock 4 or a sensor device integrated in (or provided on) the openable barrier 5 or frame around the openable barrier 5. Alternatively, the accelerometer is provided in or by a striking plate.

[0037] A crossing component of the acceleration data is defined along a direction through the openable barrier. In this example, the crossing component is along the z dimension. A vertical lateral component of the acceleration data is defined as a vertical component in a lateral direction of the openable barrier. In this example, the vertical lateral component is along the y dimension. A horizontal lateral component of the acceleration data is defined as a horizontal component in a lateral direction of the openable barrier (when closed). In this example, the horizontal lateral component is along the x dimension.

[0038] Figs 2A-C are schematic diagrams illustrating embodiments of where the break-in detector 1 can be implemented.

[0039] In Fig 2A, the break-in detector 1 is shown as implemented in the lock 4. The lock 4 is thus the host device for the break-in detector 1. In this embodiment, the lock 4 is an electronic lock. The break-in detector 1 is also connected to the accelerometer 10. The accelerometer 10 can form part of the lock 4, or the accelerometer 10 could be external to the lock 4.

[0040] In Fig 2B, the break-in detector 1 is shown as implemented in the accelerometer 10. The accelerometer 10 is thus the host device for the break-in detector1.

[0041] In Fig 2C, the break-in detector 1 is shown as implemented as a stand-alone device. In this embodiment, the break-in detector 1 does not have a host device. The accelerometer can be external to the break-in detector. The connection between thebreak-in detector i and the accelerometer can be wireless, e.g. based on BLE, or any other of the possible wireless protocols mentioned below with reference to Fig 5 that might be supported by the break-in detector 1.

[0042] It is to be noted that part or all of the break-in detector 1 does not need to be located near the accelerometer. For instance, the break-in detector 1 can be implemented in a remote server, which can be implemented in what is known as the cloud, with connectivity with the accelerometer, e.g. as part of an Internet of Things framework. In one embodiment, the break-in detector 1 forms part of a control panel. The control panel can also be used for other purposes, such as alarm, access control, smart home control, etc.

[0043] Optionally, the break-in detector 1 is provided as a multi-component system comprising one part implemented in a server and one part implemented in a local device. In this case, the training can occur in the server and the operational detection can occur in the local device (e.g. as part of the sensor or as a stand-alone device) in the immediate vicinity of the openable barrier, e.g. when training is not possible or desired to be performed by the local device.

[0044] Figs 3A-B are flow charts illustrating embodiments of methods for enabling detecting a break-in attempt. The method is performed by a break-in detector 1, such as the break-in detector 1 of Figs 2A-C.

[0045] Turning first to Fig 3A, there is first an optional select base instance step 39, in which the break-in detector 1 selects a base instance of an autoencoder based on characteristics of the openable barrier. The characteristics can include the type of openable barrier (e.g. window, wood door, metal door, window door, etc.) and / or installation environment (e.g. concrete building, wood building, etc.). The base instance is then used as a starting point for the autoencoder for which training is subsequently performed.

[0046] After the select base instance step 39, when performed, there is both a training phase 36 and an operational phase 38. In the training phase 36, a number of steps are performed.

[0047] In an obtain training acceleration data step 40, the break-in detector 1 obtains training acceleration data from at least one accelerometer 10. The accelerometer(s) 10 is / are configured to detect vibrations of an openable barrier 15. The training acceleration data is captured during a time when the openable barrier is subjected to harmless (normal) use. In other words, the openable barrier is not subjected to any break-in attempt in the time period that the training acceleration data covers. There is no requirement for labelling any of the training acceleration data, or other manual preparation of the training acceleration data. The training acceleration data can be in the form of data for three geometric dimensions (x, y, z) or a subset of these dimensions, e.g. one or two geometric dimensions.

[0048] In a derive candidate events step 42, the break-in detector 1 derives a plurality of candidate events. Each candidate event comprises event acceleration data being extracted from the training acceleration data, wherein the start of each candidate event is determined based on a section of the training acceleration data having an amplitude that is greater than a first threshold amplitude. The event acceleration data is extracted as a subset of the training acceleration data, covering a short time period. The time period is intended to cover one specific event that the openable barrier 5 is subjected to. It is envisioned that the entire training acceleration data can cover days, weeks or even months of acceleration data, while the candidate event only covers a time period of less than a second to a small number of seconds.

[0049] Optionally, the end of each candidate event is determined based on the training acceleration data having an amplitude that is below a second threshold amplitude. Alternatively, the end of each candidate event is determined to be a preconfigured amount of time from when the candidate event begins. Optionally, when the training occurs in the server, the candidate events can be transferred to the server.

[0050] In a conditional sufficient step 44, the break-in detector 1 determines whether a sufficient amount of candidate events has been obtained. It is beneficial if candidate events for all common harmless situations that may result in accelerometer data are covered. Such situations may comprise opening of the openable barrier, closing of the openable barrier, locking of the openable barrier, unlocking of the openable barrier, people walking nearby the openable barrier on the inside, people walkingnearby the openable barrier on the inside, ball from kids playing bouncing on the openable barrier or the surrounding structure, etc.

[0051] In order to achieve sufficient coverage of different types of situations, the determination can be based on determining that a variation indicator of the event acceleration data of the candidate events is greater than a variety threshold. This can be achieved by, for each candidate event, determining one or more parameters for the candidate event. Such parameters can include one or more of duration, minimum amplitude, maximum amplitude, variance and standard deviation. These one or more parameters are then stored in a database. A variation between candidate events can then be determined by finding variations between the one or more parameters of the candidate events. In this way, if there is a lot of variation in the event acceleration data, this will result in variations in the parameters of the candidate events, which can be determined based on the calculated variation.

[0052] When a sufficient amount of candidate events has been obtained, the method proceeds to a select training events step 46. Otherwise, the method returns to the obtain training acceleration data step 40.

[0053] In a select training events step 46, the break-in detector 1 selects a set of training events. This yields a set of selected training events. The set of selected training events comprises at least some of the candidate events. Only the training events in the set of selected training events are subsequently used for the training. In one embodiment, all candidate events are selected.

[0054] In a train model step 48, the break-in detector 1 trains a machine learning (ML) model based on the set of selected training events. The ML model is an autoencoder. As known in the art per se, the autoencoder is a neural network that compress input data into a lower-dimensional representation and then provides data based on attempting to reconstruct the input data. In anomaly detection, an autoencoder identifies outliers by comparing the reconstruction error, indicating the difference between the output data and the input data. This anomaly detection is based on unusual data that is input into the autoencoder will have a higher error due to poor reconstruction.

[0055] The ML model takes the x, y, and z components of acceleration data as input. Each component is first processed through one-dimensional convolutional layers with NF filters, e.g. 6 to 8 filters. Each filter has NFCOE coefficients (e.g. 7 to 9 coefficients), which act as filters to expand the number of features, potentially identifying relevant frequency components in the signal.

[0056] The features extracted from each component are then concatenated into a single set of features. This single set of features is passed through a dropout layer to reduce overfitting. The dropout layer is a regularization technique used in neural networks, including autoencoders. It works by randomly “dropping out” (i.e., setting to zero) a fraction of the neurons’ outputs in a layer during each training iteration. This prevents the network from becoming too dependent on any specific neurons, encouraging the model to learn more robust and generalizable features. This is followed by batch normalization to stabilize and accelerate training. Subsequently, a dense layer compresses the feature set into a lower-dimensional encoded representation with ND dimensions (e.g. 8 to 10 dimensions).

[0057] Finally, the encoded data is passed through transposed convolutional layers, which reconstruct the original input signal.

[0058] Once the training phase 36 is done, the detection a break-in attempt is enabled and the ML model is ready to be used operationally. Thus, in the subsequent operational phase 38, a number of steps are performed to apply the ML model to detect break-in attempts.

[0059] In an optional obtain operational acceleration data step 50, the break-in detector 1 obtains operational acceleration data from at least one accelerometer 10 configured to detect vibrations of an openable barrier (15). Optionally, the accelerometer 10 that is used as source for the training acceleration data is the same as the accelerometer 10 that is used as a source for the operation acceleration data.

[0060] In an optional determine start of operational event step 51, the break-in detector 1 determines that an operational event has started by the operational acceleration data exceeding a threshold amplitude (which can be the same as the firstthreshold mentioned above that is used to define the start of a candidate event), see also Fig 4, described below. Similar to the training events, the operational event comprises operational acceleration data starting from the start of the operational event. It is to be noted that the end of the operational event does not need to be determined at this time. At some point, the end of the operational event can be determined, e.g. as described below with reference to Fig 4, to stop sending data to the ML model. It is to be noted that it is also possible to stream the data through the model continuously, in which case the start of operational event is not required either.

[0061] In an optional calculate reconstruction error step 52, the break-in detector 1 calculates a reconstruction error by comparing operational acceleration data of the operational event that is input to the ML model with output data from the ML model. The calculation of the reconstruction error can e.g. be based on a mean squared error (MSE) calculation.

[0062] When an operational event occurs that is similar to the training events upon which the ML model was trained (harmless events), the ML model (autoencoder) will produce a small error, since such an operational event is what the ML model was trained on. In contrast, when the operational event, such as a break-in, differs from the training events upon which the ML model was trained, the ML model (autoencoder) produces a larger error. This is a property of the autoencoder type of ML models that is exploited by embodiments presented herein.

[0063] In an optional conditional break-in step 54, the break-in detector 1 determines whether the operational acceleration data indicates a break-in attempt. This determination is based on the reconstruction error. Specifically, when the reconstruction error is greater than an anomaly threshold, this is interpreted as a break- in (which is an anomaly to the events upon which the ML model was trained). For instance, the reconstruction error can be monitored per data chunk or sliding window of a certain length, and the decision about whether or not a break-in occurs can optionally be taken based on some additional logic. For instance, a break-in can be triggered if more than n reconstruction occurs over period of t time (e.g. measured in clock time duration, number of data chunks or number of samples of acceleration data).

[0064] When a break-in is determined, the method proceeds to an optional trigger alarm step 56. Otherwise, the method returns to the optional obtain operational acceleration data step 50.

[0065] Each iteration of steps 50, 52 and 54 can be performed based on data thus far of an operational event; the event does not need to end prior to feeding the captured operational acceleration data of the event to the ML model. This achieves a fast response, since the entire event does not need to occur when an alarm is triggered.

[0066] In an optional trigger alarm step 56, the break-in detector 1 triggers an alarm, indicating that a break-in is in progress. This can involve signalling an alarm signal to an alarm system.

[0067] Turning now to Fig 3B, this illustrates optional sub-steps of the select training events step 46. The sub-steps are performed for each one of the candidate events that have been derived.

[0068] In an optional evaluate difference step 46a, the break-in detector 1 evaluates how different each candidate event is from any events that are already part of the set of selected training events. Optionally, this step comprises comparing at least one parameter of each event with any events that are already part of the set of selected training events. The at least one parameter is selected from the group consisting of duration, minimum amplitude, maximum amplitude, variance and standard deviation. Optionally, these one or more parameters can be reused from if they were obtained for the conditional sufficient step 44.

[0069] In an optional conditionally select event step 46b, the break-in detector 1 determines whether the event evaluated in the evaluate difference step 46a is sufficiently different from any (i.e. zero or more) events that are already part of the set of selected training events. If this is the case, the break-in detector 1 selects that event to form part of the set of selected training events. If the event is not sufficiently different, the event is not selected, i.e. does not form part of the set of selected training events.

[0070] By only selecting candidate events that are different from the already selected events, the problem of overfitting is significantly reduced or even eliminated. Forautoencoders, overfitting is when the ML model is tailored too much to a particular input pattern. Consider, for instance, the situation that 99 per cent of the candidate events relate to opening, closing, unlocking and locking of the openable barrier. The remaining one percent relates to a ball bouncing on the openable barrier. If all candidate events are used for training, there is a large risk that, during operation, a (harmless) ball bouncing on the openable barrier could result in a reconstruction error that is so large that it is considered to be an anomaly, resulting in an erroneous break-in detection.

[0071] Hence, during training, by only selecting candidate events that are sufficiently different from any of the previously selected events, the problem of overfitting is greatly reduced.

[0072] Fig 4 is a schematic graph illustrating how candidate events 2oa-f are determined in conjunction with acceleration data 18, e.g. in embodiments covered by the flow chart of Figs 3A-B. The acceleration data 18 consists of three components x, y, z indicating acceleration for a respective corresponding geometric dimension.

[0073] In this example, there are six (candidate training or operational) events 2oa-f that are determined. The break-in detector 1 determines the start of each event 2oa-f based on when the acceleration data has an amplitude (in at least one of the components x, y, z) that is greater than a first threshold amplitude.

[0074] The end of each event can be determined based on the acceleration data having an amplitude that is below a second threshold amplitude. This evaluation can be based on all components x, y, z having their respective amplitudes below the second threshold amplitude. Alternatively, the end of each event can be determined to be a preconfigured amount of time from when the candidate event begins.

[0075] The acceleration data 18 also comprises smaller variations 19 that are not sufficient to trigger a candidate event to be determined.

[0076] Fig 5 is a schematic diagram illustrating components of the break-in detector 1 of Figs 2A-B. It is to be noted that when the break-in detector 1 is implemented in a host device, one or more of the mentioned components can be shared with the host device. Processing circuitry 60 is provided using any combination of one or more of asuitable central processing unit (CPU), graphics processing unit (GPU), multiprocessor, neural processing unit (NPU), microcontroller, digital signal processor (DSP), etc., capable of executing software instructions 67 stored in memory circuitry 64, which can thus be a computer program product. The processing circuitry 60 could alternatively be implemented using an application specific integrated circuit (ASIC), field programmable gate array (FPGA), etc. The processing circuitry 60 can be configured to execute the method described with reference to Figs 3A-B above.

[0077] The memory circuitry 64 can be any combination of random-access memory (RAM) and / or read-only memory (ROM). The memory circuitry 64 also comprises non- transitory persistent storage, which, for example, can be any single one or combination of magnetic memory, optical memory, solid-state memory or even remotely mounted memory.

[0078] A data memory 66 is also provided for reading and / or storing data during execution of software instructions in the processing circuitry 60. The data memory 66 can be any combination of RAM and / or ROM.

[0079] The break-in detector 1 further comprises an I / O interface 62 for communicating with external entities, e.g. via a wireless interface such as Bluetooth or Bluetooth Low Energy (BLE), ZigBee, any of the IEEE 802. nx standards (also known as Wi-Fi), etc., and / or using a wire-based interface such as for Ethernet or optic fibres.When the break-in detector 1 is provided in a server, the I / O interface could be configured to mainly support wire-based communication.

[0080] Other components of the break-in detector 1 are omitted in order not to obscure the concepts presented herein.

[0081] Fig 6 shows one example of a computer program product 90 comprising computer readable means. On this computer readable means, a computer program 91 can be stored in a non-transitory memory. The computer program can cause processing circuitry to execute a method according to embodiments described herein. In this example, the computer program product 90 is in the form of a removable solid-state memory, e.g. a Universal Serial Bus (USB) drive. As explained above, the computerprogram product could also be embodied in a memory of a device, such as the computer program product 64 of Fig 5. While the computer program 91 is here schematically shown as a section of the removable solid-state memory, the computer program can be stored in any way which is suitable for the computer program product, such as another type of removable solid-state memory, or an optical disc, such as a CD (compact disc), a DVD (digital versatile disc) or a Blu-Ray disc.

[0082] The aspects of the present disclosure have mainly been described above with reference to a few embodiments. However, as is readily appreciated by a person skilled in the art, other embodiments than the ones disclosed above are equally possible within the scope of the invention, as defined by the appended patent claims. Thus, while various aspects and embodiments have been disclosed herein, other aspects and embodiments will be apparent to those skilled in the art. The various aspects and embodiments disclosed herein are for purposes of illustration and are not intended to be limiting, with the true scope being indicated by the following claims.

Claims

CLAIMS1. A method for enabling detecting a break-in attempt, the method being performed by a break-in detector (i), the method comprising: in a training phase (36), performing the following: obtaining (40) training acceleration data from at least one accelerometer (10) configured to detect vibrations of an openable barrier (15), when the openable barrier is subjected to harmless use; deriving (42) a plurality of candidate events, wherein each candidate event comprises event acceleration data being extracted from the training acceleration data, wherein the start of each candidate event is determined based on a section of the training acceleration data having an amplitude that is greater than a first threshold amplitude; determining (44) when a sufficient amount of candidate events has been obtained; selecting (46) a set of training events, yielding a set of selected training events comprising at least some of the candidate events; and training (48) a machine learning, ML, model based on the set of selected training events, wherein the ML model is an autoencoder2. The method according to claim 1, further comprising: in an operational phase (38), performing the following: obtaining (50) operational acceleration data from at least one accelerometer (10) configured to detect vibrations of an openable barrier (15); determining (51) that an operational event has started by the operational acceleration data exceeding a threshold amplitude, the operational event comprising operational acceleration data starting from the start of the operational event; calculating (52) a reconstruction error by comparing operational acceleration data of the operational event that is input to the ML model with output data from the ML model; and detecting (54) a break-in attempt based on the reconstruction error being greater than an anomaly threshold3. The method according to claim 2, wherein the accelerometer (10) that is used as source for the training acceleration data is the same as the accelerometer (10) that is used as a source for the operation acceleration data.

4. The method according to any one of the preceding claims, wherein the selecting (46) a set of training events comprises: evaluating (46a), for each one of the candidate events, how different the candidate event is from any events that are already part of the set of selected training events; and selecting (46b), for each one of the candidate events, the event to form part of the set of selected training events, based on the event being sufficiently different from any events that are already part of the set of training events.

5. The method according to claim 4, wherein the evaluating (46a) how different the event is comprises comparing at least one parameter of the event with any events that are already part of the set of selected training event, wherein the at least one parameter is selected from the group consisting of duration, minimum amplitude, maximum amplitude, variance and standard deviation.

6. The method according to any one of the preceding claims, wherein the determining (44) when a sufficient amount of candidate events has been obtained is based on determining that a variation indicator of the event acceleration data of the candidate events is greater than a variety threshold.

7. The method according to any one of the preceding claims, wherein in the deriving (42) a plurality of candidate events, the end of each candidate event is determined based on the training acceleration data having an amplitude that is below a second threshold amplitude.

8. The method according to any one of the preceding claims, further comprising, prior to the training phase (36): selecting (39) a base instance of an autoencoder based on characteristics of the openable barrier; wherein the base instance is a starting point for the autoencoder for which training is performed.

9. A break-in detector (i) for enabling detecting a break-in attempt, the break-in detector (1) comprising: processing circuitry (60); and memory circuitry (64) storing instructions (67) that, when executed by the processing circuitry, cause the break-in detector (1) to: in a training phase (36), execute instructions to: obtain training acceleration data from at least one accelerometer (10) configured to detect vibrations of an openable barrier (15), when the openable barrier is subjected to harmless use; derive a plurality of candidate events, wherein each candidate event comprises event acceleration data being extracted from the training acceleration data, wherein the start of each candidate event is determined based on a section of the training acceleration data having an amplitude that is greater than a first threshold amplitude; determine when a sufficient amount of candidate events has been obtained; select a set of training events, yielding a set of selected training events comprising at least some of the candidate events; and train a machine learning, ML, model based on the set of selected training events, wherein the ML model is an autoencoder;10 The break-in detector (1) according to claim 9, wherein the instructions further comprises instructions (67) that, when executed by the processing circuitry, cause the break-in detector (1) to: in an operational phase (38), execute the instructions to: obtain operational acceleration data from at least one accelerometer (10) configured to detect vibrations of an openable barrier (15); determine that an operational event has started by the operational acceleration data exceeding a threshold amplitude, the operational event comprising operational acceleration data starting from the start of the operational event; calculate a reconstruction error by comparing operational acceleration data of the operational event that is input to the ML model with output data from the ML model; and detect a break-in attempt based on the reconstruction error being greater than an anomaly threshold11. The break-in detector (1) according to any one of claims 10, wherein the accelerometer (10) that is used as source for the training acceleration data is the same as the accelerometer (10) that is used as a source for the operation acceleration data.

12. The break-in detector (1) according to claim 9 or 10, wherein the instructions to select a set of training events comprise instructions (67) that, when executed by the processing circuitry, cause the break-in detector (1) to: evaluate, for each one of the candidate events, how different the candidate event is from any events that are already part of the set of selected training events; and select, for each one of the candidate events, the event to form part of the set of selected training events, based on the event being sufficiently different from any events that are already part of the set of training events.

13. The break-in detector (1) according to claim 12, wherein the instructions to evaluate how different the event is comprise instructions (67) that, when executed by the processing circuitry, cause the break-in detector (1) to compare at least one parameter of the event with any events that are already part of the set of selected training event, wherein the at least one parameter is selected from the group consisting of duration, minimum amplitude, maximum amplitude, variance and standard deviation.

14. The break-in detector (1) according to any one of claims 9 to 13, wherein the instructions to determine when a sufficient amount of candidate events has been obtained is based on determining that a variation indicator of the event acceleration data of the candidate events is greater than a variety threshold.

15. The break-in detector (1) according to any one of claims 9 to 14, wherein the instructions to derive a plurality of candidate events comprise instructions (67) that, when executed by the processing circuitry, cause the break-in detector (1) to determine the end of each candidate event based on the training acceleration data having an amplitude that is below a second threshold amplitude.

16. The break-in detector (1) according to any one of claims 9 to 15, further comprising instructions (67) that, when executed by the processing circuitry, cause the break-indetector (1) to, prior to the training phase (36): select a base instance of an autoencoder based on characteristics of the openable barrier; wherein the base instance is a starting point for the autoencoder for which training is performed.

17. A computer program (67, 91) for enabling detecting a break-in attempt, the computer program comprising computer program code which, when executed on a break-in detector (1) causes the break-in detector (1) to: in a training phase (36), execute instructions to: obtain training acceleration data from at least one accelerometer (10) configured to detect vibrations of an openable barrier (15), when the openable barrier is subjected to harmless use; derive a plurality of candidate events, wherein each candidate event comprises event acceleration data being extracted from the training acceleration data, wherein the start of each candidate event is determined based on a section of the training acceleration data having an amplitude that is greater than a first threshold amplitude; determine when a sufficient amount of candidate events has been obtained; select a set of training events, yielding a set of selected training events comprising at least some of the candidate events; and train a machine learning, ML, model based on the set of selected training events, wherein the ML model is an autoencoder.

18. A computer program product (64, 90) comprising a computer program according to claim 17 and a computer readable means comprising non-transitory memory in which the computer program is stored.

Images