A system and method for tracking objects
The IoT-based system addresses the limitations of existing logistics monitoring by inferring zones of interest using environmental data to predict and manage risks, ensuring the safety and security of goods in transit.
Patent Information
- Application Number
- PCT/EP2026/062623
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-08
- Filing Date
- 2026-04-30
- Publication Date
- 2026-11-12
AI Technical Summary
Current methods for monitoring and controlling the flow of goods in logistics chains are inadequate in mitigating risks such as damage, delay, and theft, as they rely on limited data from online mapping tools like Google Maps, which do not account for real-time environmental conditions and human intervention.
A system utilizing IoT devices to gather environmental data, including RF and physical parameters, which infers zones of interest using clustering algorithms and machine learning, enabling proactive risk management by predicting and alerting users to potential risks through a central server.
The system minimizes product risks by continuously updating inferred zones based on real-time data, allowing for proactive risk mitigation and optimizing the safety and security of goods in transit.
Smart Images

Figure EP2026062623_12112026_PF_FP_ABST
Abstract
Description
[0001] 29 April 2026
[0002] A System and Method For Tracking Objects
[0003] The present invention relates to a system and method for tracking objects. The invention also relates to an Internet of Things (loT) device for use within the system and method.
[0004] In embodiment, the invention relates to a tracking device, a server system and a method of operating a server system for control of a plurality of loT (internet of things) devices, such as tracking devices. In embodiments, the invention relates to a distribution system for optimising distribution of goods in a global supply chain.
[0005] Artificial Intelligence (Al) has now become a key driver for business transformation. Big consulting firms and all listed companies with significant market capital valuations are developing strategies to benefit from recent Al advancements. The long-term future of Al will rely on computational power or resource as a key metric of performance. According to Sam Altman, in the future, if more computational resource can be used for a task, it will be used. The cost of the computing will become the key limiting factor due to economics. This is because, aside from other growth factors in Al use, given the choice, it follows that the maximum intelligence will be used for any task.
[0006] In the supply chain industry, which is the world's biggest industry and a key beneficiary of Al value improvements, the majority of data is collected by sensors in realtime at the edge, i.e. away or remote from the central data processing resource which will typically be in a cloud environment. Goods that are moved around the world in logistics chains are subject to many different types of risk. These include risk of damage to the product due to environmental conditions that they may encounter (for example exposure to temperatures outside a safe zone), risk of damage due to delay, i.e. for products that have a limited shelf life such as fresh food or pharmaceuticals, but also risk from human intervention, i.e. unauthorised accessing of the goods or even theft.
[0007] Current ways to monitor and control the flow of goods attempt to mitigate these risks but often fall short. Examples include the use of online mapping tools such as29 April 2026
[0008] Google maps and the like. The data they provide can be used to determine risk of delay to a product if, say, congestion in shown on a road being used to transport some goods. However, the data such systems can provide is limited and accordingly residual risk to goods remains high, or higher than is optimal for the goods being transported.
[0009] According to a first aspect of the present invention, there is provided a system for tracking objects, the system comprising plural loT devices each associated with a corresponding object; a server in communication with each of the loT devices; each of the loT devices being arranged to gather environmental data associated with the object with which it is associated and transmit the gathered environmental data to the server; the system being arranged, in dependence on the gathered data, to infer zones of interest in relation to the loT devices.
[0010] A system is provided in which multiple loT devices, each associated with some real world twin, are able to communicate data regarding their position and / or environment to a central sever. Based on the data received or collected by the central server zones can be inferred into which the loT devices are arranged. Risks associated with the zones can be calculated which can take into account the products associated with an individual loT.
[0011] A specific and non-limiting example might be food stuffs being transported which are known not to be suitable for exposure to a certain temperature for a specified amount of time. The model could identify or infer a zone and determine that any foodstuffs that enter it are at a high risk of exposure to the temperature and thus damage. Action can then be taken to avoid the risk by diverting the loT and real world twin. This could be done automatically or manually. An alarm could be raised and communicated to someone or something with control of the loT and associated goods.
[0012] In an example, the environmental data is selected from data relating to the RF environment and data relating to the physical environment of the loT device.
[0013] In an example, the data relating to the RF environment includes one or more of Cellular, Wi-Fi, Lora, UWB or BLE signal strength experienced by the loT device.29 April 2026
[0014] In an example, the data relating to the physical environment includes one or more of temperature, humidity, pressure, movement, and acceleration.
[0015] In an example, the system is arranged to determine zones using a clustering algorithm based on data gathered from the plural loT devices.
[0016] In an example, zones are inferred on the basis of gathered data from multiple loT devices and the application of rules-based clustering.
[0017] In an example, rules-based clustering is executed based on gathered environmental data and / or geographical location.
[0018] In an example, upon introduction of an loT data is collected by the loT device and communicated to the server.
[0019] In an example, based on the received data the server establishes the geographic location or the RF zone in which an loT device is located.
[0020] In an example, the data gathered from the loT is added to stored data from other loTs.
[0021] In an example, the formation of a new inferred zone comprises the steps of: receiving at the server the data from multiple loT devices; defining clustering criteria; apply clustering criteria to identify loT devices belonging to respective clusters; and, establish boundaries of clusters based on identification of loT devices belonging to the clusters.
[0022] In an example, the data received from the loT devices is one or more of geographic data and data relating to a detected RF environment.
[0023] In an example, the clustering criteria are modified if the number of identified loT devices is not within a desired range for an identified cluster.29 April 2026
[0024] In an example, a risk to products associated with each of the loT devices is determined in dependence on the inferred zone in which the loT device in question is located.
[0025] In an example, the system comprises developing a model of multiple inferred zones and determining likely trajectory of a particular loT in dependence on previous movements.
[0026] In an example, the system comprises building the model of inferred zones expected movements of and / or loTs using machine learning. The volume of data that will be gathered and processed is large, related to the large number of loT devices likely to be in use or circulation at any point in time. This enables the use of machine learning to infer the zones. Preferably a neural network can be used such as an RNN to develop the model of, i.e. infer, the zones. Transformer can be used as well or instead. The skilled person will understand that other suitable forms of machine learning can be used both to identify and infer the zones from the data gathered at the server or to develop a model of movement of the loTs between zones. This is explained in greater detail below.
[0027] In an example, the model includes likely risks to the loT within one or more of the inferred zones or on its passage through and between inferred zones.
[0028] In an embodiment, the gathered data associated with the object with which it is associated is encoded at the loT device.
[0029] In an embodiment, the encoding is selected to reduce data transmission to the server.
[0030] In an embodiment, the reduction of data transmission comprises reduction in data volumes, bandwidth or rate.
[0031] In an embodiment, the gathered data is encoded to an indication of one or more of Human Presence Event, Change in Custody and Modality in relation to the loT device in question.29 April 2026
[0032] A system is provided in which complex information gathered by an loT device for transmission to the server is encoded in such a way that a machine learning tool at the server is able to make the best use out of it and advantageously such that it is not necessary to transport as much data to the central server from the loT devices.
[0033] Encoding and data processing is done at the loT devices which is considered or referred to as embedding. The encoding or embedding is in effect the substitution of a smaller and simpler data entity for multiple pieces of data or data points relating to specified parameters. Thus, the volume or rate of data needed to enable transfer to the central server can be substantially reduced. Since data transmission is also power intensive, the substitution enables a reduction in power requirements for the loT devices.
[0034] According to a second aspect of the present invention, there is provided an loT device for use in tracking objects in a system for tracking objects including plural loT devices each associated with a corresponding object, a server in communication with each of the loT devices, in which the loT devices are arranged to gather data associated with the object with which it is associated and transmit the data to the server, in which the server is arranged, in dependence on the received data, to infer zones of interest in relation to the loT devices, the IOT device comprising: an RF detector to detect an RF environment of the loT device; a geographic position detector to detect an RF environment of the loT device; a transmitter to transmit to the server the gathered data.
[0035] In an example, the loT Device comprises a processor and a rechargeable power source to provide power to the components of the loT tag.
[0036] In an embodiment, the gathered data associated with the object with which it is associated is encoded at the loT device.
[0037] In an embodiment, the encoding is selected to reduce data transmission to the server.
[0038] In an embodiment, the reduction of data transmission comprises reduction in data volumes, bandwidth or rate.29 April 2026
[0039] In an embodiment, the gathered data is encoded to an indication of one or more of Human Presence Event, Change in Custody and Modality in relation to the loT device in question.
[0040] An loT device is provided that is arranged to gather information and to encode the gathered data (such as environmental data) for transmission to the server. The machine learning tool at the server is able to make the best use out of it and advantageously such that it is not necessary to transport as much data to the central server from the loT devices. Encoding and data processing is done at the loT devices which is referred to as embedding. This, enables the volume or rate of data needed to enable transfer to the central server to be substantially reduced.
[0041] According to a third aspect of the present invention, there is provided method for tracking objects, the method comprising: providing plural loT devices each associated with a corresponding object to be tracked; providing a server in communication with each of the loT devices; each of the loT devices gathering data such as environmental data associated with the object with which it is associated and transmitting the gathered data to the server; at the server, in dependence on the gathered data, inferring zones of interest in relation to the loT devices.
[0042] In an example, the environmental data is selected from data relating to the RF environment and data relating to the physical environment of the loT device.
[0043] In an example, the data relating to the RF environment includes one or more of Cellular, Wi-Fi, Lora, UWB or BLE signal strength experienced by the loT device.
[0044] In an example, the data relating to the physical environment includes one or more of temperature, humidity, pressure, movement, and acceleration.
[0045] In an example, the method comprises determining zones using a clustering algorithm based on data gathered from the plural loT devices.29 April 2026
[0046] In an example, the method comprises inferring zones on the basis of gathered data from multiple loT devices and the application of rules-based clustering.
[0047] In an example, the method comprises executing rules-based clustering based on gathered environmental data and / or geographical location.
[0048] In an example, upon introduction of an loT data is collected by the loT device and communicated to the server.
[0049] In an example, based on the received data the server establishes the geographic location or the RF zone in which an loT device is located.
[0050] In an example, the data gathered from the loT is added to stored data from other loTs.
[0051] In an example, the formation of a new inferred zone comprises the steps of: receiving at the server the data from multiple loT devices; defining clustering criteria; apply clustering criteria to identify loT devices belonging to respective clusters; and, establishing boundaries of clusters based on identification of loT devices belonging to the clusters.
[0052] In an example, the data received from the loT devices is one or more of geographic data and data relating to a detected RF environment.
[0053] In an example, the clustering criteria are modified if the number of identified loT devices is not within a desired range for an identified cluster.
[0054] In an example, a risk to products associated with each of the loT devices is determined in dependence on the inferred zone in which the loT device in question is located.
[0055] In an example, the method comprises developing a model of multiple inferred zones and determining likely trajectory of a particular loT in dependence on previous movements.29 April 2026
[0056] In an example, the method comprises building the model of expected movements of loTs using machine learning.
[0057] In an example, the model includes likely risks to the loT within one or more of the inferred zones or on its passage through and between inferred zones.
[0058] In an embodiment, the gathered data associated with the object with which it is associated is encoded at the loT device.
[0059] In an embodiment, the encoding is selected to reduce data transmission to the server.
[0060] In an embodiment, the reduction of data transmission comprises reduction in data volumes, bandwidth or rate.
[0061] In an embodiment, the gathered data is encoded to an indication of one or more of Human Presence Event, Change in Custody and Modality in relation to the loT device in question.
[0062] As above, the encoding and data processing is done at the loT devices which enables the volume or rate of data needed to enable transfer to the central server to be substantially reduced.
[0063] The use of such a system and loT device enables a continual updating of a model of inferred zones which in practice will enable the minimisation or elimination of risk to products that are being moved around the world or from one location to another.
[0064] A system is also provided in which a model is maintained and updated based on data received from one or more edge devices, such as the tracking or loT devices referred to above.29 April 2026
[0065] Typically, Al such as Machine learning is used on the database and factors such as expected movement trajectories of the individual IOT device can be determined.
[0066] Embodiments of the present invention will now be described in detail with reference to the accompanying drawings, in which:
[0067] Figure 1 A is a schematic view of a logistics system including an HGV lorry and a warehouse connected to a cloud server;
[0068] Figure 1 B is a schematic view of a logistics system similar to that of Figure 1 A, but in this example utilising the concept of embedding, as described below;
[0069] Figure 10 is a schematic view of an loT device arranged to utilise the concept of embedding described with reference to Figure 1B;
[0070] Figure 2 is a schematic view showing the data flow in the present system;
[0071] Figure 3 is a schematic view of process for creating zones of significance in the present method;
[0072] Figure 4 is a schematic view of process for finding intersection of zones through RF environment and signal strength optimisation;
[0073] Figure 5 is a schematic view of a plurality of identified zones in a logistics distribution system;
[0074] Figures 6 and 7 are schematic flow diagrams showing the processes of model preparation and model usage according to the present invention.
[0075] As explained above, the present system is a system for use in the logistical distribution of product as typically might happen around the world on lorries, boats airplanes or any other form of transport. The system relies on the use of IOT (internet of things) devices which are used to collect data from goods being transported and29 April 2026
[0076] communicate that data to a remote server for processing. The remote server functions to infer zones of interest for the products or devices as they are moved around the world.
[0077] Furthermore, the identification and presence of inferred zones based on data from loT devices that are in the system, enable predictions for other loT devices earlier in their journeys regarding a sequence of zones that are to be entered. With knowledge of inferred zones yet to be entered for a particular loT device, analysis of risks for products associated with the loT device in question can be made. Alerts are preferably generated to notify users of risks that goods-in-transport might face as they pass through or into a particular inferred zone.
[0078] The method relies on two related processes, the first of which is the generation of a model and the inferring of the zones based on the gathered data from the IOT devices. The second, is use of the model and the pattern and existence of the inferred zones to improve the safety, reliability and security of goods that subsequently travel through the inferred zones. For a specific IOT device moving through the inferred zones, it is possible to generate alerts and thereby optimise the safety of products that are being moved around the world.
[0079] The system thus relates to IOT devices and the real-world objects and constructs that they represent as digital twins. Typically, the real-world constructs would be shipments of goods, packages or assets and their supporting cloud services in logistics, supply chain and asset management.
[0080] The method provides an up-to-date operating model of real-world environmental behaviour and allows users and autonomous services to analysis IOT devices as they move through the real world. Interventions can be made based on predictions and levels of statistically determined confidence in relation to a particular IOT device and the goods with which it is associated.
[0081] During transit of goods, shipments may use established lanes and ensure that goods are only transported along those established lanes. On a specific lane, a shipment or container of goods may enter inferred zones of interest that involve or generate material change to the shipment.29 April 2026
[0082] Examples of material changes to the shipment include a change of custody, a dwell of a shipment, and an area responsible for environmental excursions such as temperature, humidity, pressure and the like. Importantly, in practice, the zones of interest are not fixed. In time, due to changes in many factors, the zone of interest might change. The present system provides flexibility in that zones can be varied and changed as further data is gathered from the plurality of IOT devices that are moving around the world.
[0083] In practice, hundreds of thousands, or millions, of such IOT devices are in active use of any point in time. Thus, the volume of data that it is possible to gather is significant. This data can be modelled and, based on it, zones of interest can be inferred. Indeed, the zones may or may not be known or specifically documented, but can be significant as they present themselves in the gathered data at a central server.
[0084] Referring then to figure 1A, the system 2 in this nonlimiting example is exemplified by a heavy goods vehicle (HGV) 4 shown carrying a number of totes 6 each containing goods being transported.
[0085] An IOT device 8 is provided associated with the HGV 4. The loT devices 8 are preferably tags, to be described in detail below, capable of gathering, storing and transmitting data about their environment and conditions to which they are exposed. In practice, individual IOT devices might be associated with each of the totes 6 but in the schematic example shown, a single IOT device 8 is associated with the HGV 4.
[0086] Typically, an loT device is provided in the form of a device which includes a PCB having various desired components fixed thereon. For example, the loT device might include some processing capacity, e.g. a microprocessor together with memory and a power supply such as a rechargeable battery.
[0087] The loT device 8 is arranged to scan the RF environment wherever it is positioned at the point in time when the scan is taking place. Sensor data is collected which relate to the values of the detected parameters. For example, it could include the identification and strength of a Wi-Fi signal or the strength of a local Bluetooth signal.29 April 2026
[0088] There is also shown, schematically, a warehouse 10 including a forklift truck 12 and, again, a number of totes 14 containing various goods. An IOT device 16 is provided, associated with the warehouse 10. Again, individual IOT devices can also be provided in association with one or more totes 14 within the warehouse 10. However, in this example and for simplicity, only a single IOT device 16 is shown associated with each of the warehouse 10 and the HGV 4.
[0089] A server 18, remote from both the HGV 4 and the warehouse 10 is provided. The server 18 is arranged to receive data communicated to it from the loT devices 8 and 16 at the HGV 4 the warehouse 10, respectively. As explained above, in an actual system there will typically be 100s of thousands, or even millions, of such IOT devices distributed at any point in time all over the world.
[0090] The IOT devices 8 and 16 are preferably arranged to include a plurality of sensors for detecting various data associated with their environment. This could, for example, include information relating to an RF environment, temperature, handling data and additional sensor data. The RF environment can be thought of as an RF fingerprint indicating all of the available RF signals associated with a region or “zone” in which the IOT device is located.
[0091] In both the case of the IOT device 16 associated with the warehouse 10 and the IOT device 8 associated with the HGV 4, it will be appreciated that the RF environment will be a changing entity. The HGV RF environment specifically will change as the HGV moves. It will encounter different Wi-Fi signals from buildings that is passes and also different cellular transmissions from mobile base stations and the like. The IOT device 8 associated with the HGV 4 (or IOT devices associated with individual totes within the HGV 4) will detect the changing RF environment as the HGV moves and can be arranged at desired frequencies to sample this and communicate it to the server 18. The sever 18 gathers data like this and uses it to build a model of inferred zones.
[0092] The IOT device 16 associated with the warehouse 10, although comparatively static (as compared to the IOT device 8 associated with the HGV 4), will also encounter29 April 2026
[0093] a varying RF environment since the signals that will be detected within the warehouse will likely change with time.
[0094] Accordingly, the data gathered both by the IOT device 16 associated with the warehouse and the IOT device 8 associated with the GV 4 is communicated to a server 18, which is preferably remote from both the HGV 4 and the warehouse 10. The server 18 is preferably a server located in the cloud. The data can be communicated from the IOT devices to the server 18 in any suitable form or using any appropriate protocol. Typically, it will be communicated wirelessly to the cloud using known communication protocols.
[0095] Considering the IOT device 8 associated with the HGV 4, it will be appreciated that the IOT device will move wherever the HGV 4 moves.
[0096] These zones may represent actual geographical locations or represent radio frequency (RF) zones which typically consist of Cellular, Wi-Fi, Lora, UWB or BLE zones that represent a proxy for a geographical location generally outdoors but potentially also indoors.
[0097] Responsible parties for shipping goods, do not always know or understand where these zones are. They either do not document all of the zones, are incapable of doing so or do not have enough statistically significant data themselves to form the zones. Where a responsible party is able to both know and predict entry of a shipment into a zone or series of zones, the party or parties can actively manage a risk to a shipment.
[0098] Each loT device that is sent along or associated with a shipment of goods will be arranged and configured to collect data about the environment around it using one or more of a combination of sensors. Our copending application GB2411301 , the entire contents of which are hereby incorporated by reference, describes a method and system by which individual loT devices and plural loT devices acting as a system are able to obtain and transmit environmental data to a central server whilst minimising loT device power usage and maximising loT device battery life.29 April 2026
[0099] In the present system, and as will be described in greater detail below, the collected data at the server (or remotely from the server) is used to perform clustering either to infer or create a new zone, or to attribute or assign a particular loT device from which data has been gathered to an existing zone. The zones can be created manually or inferred through machine learning. Accordingly, it will be understood that zones are determined or defined in dependence on data gathered from loT devices and clustered according to some defined algorithm or process. In practice clustering algorithms are used to infer the zones.
[0100] The present system improves upon existing ideas of zones and geozones created both automatically or manually because of its ability to both find the optimum statistically significant size (in terms of devices within a zone) and the ability to add both raw data and derived data into the zone for the purpose of grading the zone for risk to goods in-transit and / or assets. The system preferably combines both geozones and RF zones. In other words, the system uses both information relating to an loT device’s geographical position on the earth’s surface and information relating to its detected RF data to generate a zone or allocate a particular loT device to an existing zone.
[0101] The system also improves over traditional route prediction used in logistics. This is because the calculations and predictions of the present system are based upon loT device movement between zones and inferring of the zones in the first place based on gathered data from multiple other loT devices. These defined or inferred zones, based on data collected from loT devices within the zones, are able to be graded for risk. The method and system, as will be described in greater detail below, enables predicting passage into or from a zone and scoring a zone for risk.
[0102] It is known that existing logistics systems utilise the prediction of journeys for loT devices and for risk analysis of the journey predicted. However, the present method and system differs in both the two key areas. First, an optimum geo zone size is calculated adaptively during operation of the system i.e. as loT devices move and gather data and communicate the data back to the central server 18. The zones are not static and in fact are able to be continually updated as more data is gathered from other loT devices. Second, the system and method uses RF zones for the same application. Accordingly,29 April 2026
[0103] the method and system, in preferred embodiments, uses the combination of Geo and RF zones for the same application.
[0104] The present system and method rely on use of plural devices that are arranged and configured to track and provide date regarding vehicles or units with which they are associated. For example, a device could be associated with a vehicle such as an HGV lorry or with a pallet of goods within a container on a goods tanker conveying goods from, say Europe to Asia.
[0105] As described above, the system and method can be implemented using tracking devices such as those described in our copending GB patent application GB2411301. The device disclosed therein comprises a processor and a memory, and the device is arranged to receive data from a central server, the received data representing a matrix subset corresponding to the expected location of the device. The device is further arranged to detect physical parameters from a local environment. The physical parameters might typically an RF signature or fingerprint of the environment in which the device is situated. In dependence on a comparison of the detected local communications parameters and the received subset determining an operation.
[0106] The devices are deployed globally, monitoring logistics movements for businesses and organizations. The devices, which may be referred to as edge devices as they are remote from the central server 18. They are typically battery-powered with either rechargeable or primary (non-rechargeable) cells. Techniques described in GB2411301 are employed to manage battery use, allowing for longevity and small size while collecting and transmitting sufficient sensor data back to the central server 18 typically arranged in the cloud, for respective use cases.
[0107] Figure 1 B is a schematic view of a logistics system similar to that of Figure 1 A, but in this example utilising the concept of embedding. Similar components are numbered in corresponding manner to Figure 1 A. the system is arranged to operate in similar manner to the system shown in Figure 1 A however the data communicated from the loT devices to the central server 18, e.g. a server in the cloud, is different.29 April 2026
[0108] Complex information for transmission to the server 18 is encoded in such a way that the machine learning tool at the server 18 is able to make the best use out of it and such that it Is not necessary to transport as much data to the central server 18 from the loT devices. This some encoding and data processing is done at the loT devices to enable this encoding or embedding. The encoding or embedding is in effect the substitution for multiple pieces of data or data points relating to specified parameters which enables a substitution to be made for the entirety of the data in question.
[0109] Thus, it will be understood that an loT device (or multiple such devices) that are attached to assets or associated with goods in a shipment are arranged to collect data about the environment around the device using a combination of sensors. The data collected is used to attribute a device either to an existing zone (either created manually or inferred through machine learning) or to a new zone.
[0110] In some situations, as explained, the data collected by loT device sensors can be too large to transmit over a cellular connection effectively (especially in areas of poor signal). Examples of sensor data that can be too large include:
[0111] 1. Audio data
[0112] 2. Video data
[0113] 3. Accelerometer data
[0114] To effectively transfer the data and use it within models, in a preferred embodiment, the data is sampled and pre-processed as embeddings on the loT device (i.e. edge processing) before being transmitted to the remote (typically cloud-based) server.
[0115] Embeddings for the above services will rely on models local to the loT device to categorize data before embedding. Examples include:
[0116] Audio and accelerometer data combined to determine mode of transport (vibration and ambient sound).
[0117] Accelerometer data analysed locally to determine handling and / or change of custody29 April 2026
[0118] Thus, embeddings are a way to represent complex and large data which can be used within machine learning models. For distributed data collection and battery maximisation, it is preferable that data is not continuously streamed to the server. In some cases if there are bandwidth limitations it might not even be possible. This specific non-limiting embodiment uses machine learning models deployed at the server that can operate on the embeddings without requiring transport of the raw data from the loT devices.
[0119] For example, a model on the loT device is able to determine (understand) if the shipment is now on a plane instead of a truck. This is determined by, for example, working out on the device through a model that utilises audio and accelerometer data. The combination of this data enables the loT device to communicate to the server 18 that, say, “custody has changed” or “on a plane”. Instead of sending large volumes of data such as the details of the audio or movement detected by the loT device, the captured data is encoded into a shorter substitution or embedding. In the example without use of such embedding this determination is done at the server or at least, remotely from the loT device. In the present example, the loT device itself is able to make this determination and provide the reduced amount of data to the server 18.
[0120] This type of embedding also simplifies the job of creating further machine learning models as it is possible to base the models off the substituted shorthand or the embedding “on a plane” instead of on the large number, typically thousands or more, of individual audio samples and vibration samples. This further reduces the power requirement for transmission of the data from the loT devices to the server 18.
[0121] Figure 1C is a schematic view of an loT device arranged to utilise the concept of embedding described with reference to Figure 1B. A number of raw data sources 17 are arranged in use to be received by the loT device 15 in question. In the example of Figure 1A this data 17 would be transmitted to the server 18 to enable identification or inference of zones. In the present example though, the various collections of data is encoded as embedding “Human Presence Event”, “Change in Custody” and “Modality”. Each of these embeddings is determined by the loT device 15 based on the raw data 17. Once the embedding has been identified this embedding 19 is preferably transmitted29 April 2026
[0122] back to the server in place of the raw data 17 itself, which would be of much larger volume by comparison. Of course in some examples if desired both the embedding and the raw data are transmitted but this is done only if there is some specific need for bothy data types at the server 18.
[0123] In an example the loT device itself (or themselves) includes a model for computing an embedding vector as an encoding of the associated data. Embedding vectors are typically numerical representations of data (like text, images, or audio) that capture its semantic meaning and relationships. They can be used by machine learning models to understand and compare data points, allowing for tasks like similarity searches, clustering, and recommendations. In the present example, the embedding vector represents the context of gathered data. It is then the vector that is transmitted to the server (instead of the raw data) and used to retrieve similar embeddings associated with prior tracked objects or zones.
[0124] The embedding vector can thus be used to perform one or more tasks at the server, i.e. classifying a shipment risk, predicting journey outcomes, or recommending data acquisition policy. In preferred examples, the model for computing an embedding vector is partially trained or updated at the loT device using local data, or via updates received from the server. The model is thus distributed (it may be considered a distributed model). Preferably, the loT device is arranged to participate in a federated learning framework, such that updates to the embedding model are computed locally at the loT device using gathered data, and only model updates or gradients are transmitted to the server, thereby improving the embedding model across multiple devices without requiring transmission of raw data (between loT devices and the sever).
[0125] Figure 2 is a schematic view showing the data flow in the present system. The process is described with reference to an exemplary physical system such as that shown in and described with reference to Figure 1A.
[0126] In practice, initially, each of the IOT devices is arranged to collect 20 data from its local environment. At step 22, the IOT device is arranged to communicate the collected data to the central server 18. In the example of Figures 1B and 1C instead of raw data29 April 2026
[0127] being transmitted to the server, the embedding is used which can enable a smaller volume of data to be transmitted, as explained above.
[0128] The sever 18 may be a cloud-based server or it may be local to a particular environment such as a warehouse or commercial facility. The collection of data, or transmission of data from the loT devices, can be done on a periodic cyclic basis at some predefined frequency or, alternatively it can simply be done every time there is a change in one of the other gathered parameters. For example, if an loT device detects a change in its RF environment or detects that it has moved geographically, it can be arranged to “switch on”, e.g. from a sleep mode, and take measurements for transmission to central server 18.
[0129] The location of the IOT device is identified 24 either with the use of internet locating services or via GPS data supplied by the IOT device. Both the location of the RF environment and the RF environment are considered zone candidates. At step 26 existing zones are searched to determine if, based upon the location or RF gathered data, the data should be placed within the zone for analysis. The collected data is added to the existing zone 28 if appropriate based on the determined parameters.
[0130] Alternatively if the data does not fit into an existing zone then it is stored for further analysis 30.
[0131] Figure 3 is a schematic view of process for creating (“inferring”) zones of significance in the present method.
[0132] Figure 3 is a process that is utilised in the event that the gathered data (or embedding) does not fit into an existing zone as identified at the central server 18.
[0133] Initially at step 32 a background service is started. The latitude and longitude of the IOT device which has provided the data is noted. At step 36, clustering criteria are defined as a tool to help define the extent of a particular cluster to be created. Initially an arbitrary number can be used in this step, but as the system develops optimal clustering numbers can be identified.
[0134] Next, at step 38, a suitable clustering Application is used to calculate the actual clusters and identify which of the lOTs is within each cluster. Applications such as29 April 2026
[0135] HDBSCAN can be used to calculate or determine the clusters. If the number of IOT devices identified is too large, then appropriate methods or tools can be used to cluster at this massive scale. This typically occurs when there are too many records for a typical clustering algorithm to operate. It functions by partitioning the group and thereby enabling clustering in this way. For example a statistical sample of all the available loT devices may be used, or data from every other one of the loT devices may be discarded and not used by the clustering algorithm.
[0136] At step 40, a review is made of the identified loTs within a specified cluster and a feedback loop can be used to optimise the output 42. If the clustering identified is significant then the feedback mechanism will simply output an indication of the significance and the boundary of the cluster 44 is thus established. If the feedback mechanism determines that the cluster is not considered significant, the process returns to step 36 as it continues in a search for a new cluster. The feedback loop 42 repeats steps 36 to 40 until one is identified.
[0137] Finally, with the boundary of the cluster established at step 44, the inferred geozone is specified at step 46.
[0138] The process defined above for creating new zones of significance is arranged to operate in dependence on the latitude and longitude of each individual IOT device in consideration. As well as, or instead of using this process, it is possible to perform a similar sequence of steps based on the RF environment. This takes into account reported RF environment data from each of the IOT devices.
[0139] This is now described with reference to Figure 4. At step 48, with a background service started, the RF environments of the individual loT devices is received at the central server 18. An analogous to the latitude and longitude based zoning methodology used in the process of figure 3, in the present example, a rules based clustering of the RF environments 50 is performed. Again, partitioning 52 may be applied if the number of IOT devices within a shared RF environment is too big.
[0140] At step 54, the clustered environments are filtered based on the significance of the associated business rules. Finally, at step 56, the inferred RF zones are specified.29 April 2026
[0141] Accordingly, referring to both figures 3 and 4, it can be seen that two parallel methodologies have been used for generating clusters of IOT devices. These are based both on the physical positioning of the devices (e.g. latitude and longitude) and the RF environments of the loT devices that are detected by the devices themselves.
[0142] Once the RF and physical location zones have been specified, the previous iteration of the zones that was already stored within the system can be updated based upon the collected data. If there was no previous version, then new zones are simply created.
[0143] The gathered geo-location data and RF environment data, including both the raw data and information derived from the raw data is compiled into a machine learning model. Neural networks such as RNNs or Transformer can be used as the machine learning model. It will be appreciated that the uploading of data to the central server is a continuous process such that zones are updated based upon collected data. This enables at any specific point in time the model to be used to provide an accurate risk assessment for goods within the inferred zone.
[0144] The type of data relating to the environment that might be derived from a sensor within an loT device can typically be the maximum temperature to which the loT has been exposed, the median temperature to which it has been exposed, a number of temperature excursions from some defined range, and the average length of time at a specific temperature or within a temperature range.
[0145] In addition, separate from the temperature, other parameters can be gathered. Such parameters include an indication of the rough handling of goods, e.g. based on accelerometer data, the number of times the goods dwelled in passage, the average length of time of the dwell and any changes of custody location. The system will be aware that any particular IOT device must adhere to specific rules relating to risk. This is of course related to the products with which the loT device is associated. The developed machine learning model can therefore return a grade with a confidence in respect of a particular IOT device or an inferred zone.29 April 2026
[0146] Figure 4A shows a schematic scorecard for a particular loT device to enter a zone (labelled just for convenience ’’Zone A”). In this example, according to the system rules it is known that the loT device must not dwell for more 30 minutes in a zone.
[0147] However, the zone model has determined that an loT device is likely (with 90% confidence) to dwell for 2 hours in the zone. In this situation some action might be advisable, such as generating an alert to a system controller or the owner of the goods associated with the loT in question.
[0148] The parameter associated with zone A show that no rough handling of goods can be allowed an average dwell time of two hours is specified. A maximum temperature of thirty degrees centigrade is allowed. Accordingly, in this example, the IOT device must not dwell for more than thirty minutes in a zone. However, the zone model has determined that an IOT device within this zone is likely to dwell for two hours and this is stated with ninety percent confidence. The risk score is therefore given a grade F.
[0149] Evasive action might be possible, such as rerouting of the goods.
[0150] The system operates to learn, via machine learning, likely sequences of zones that an loT will enter. This can be done based on the gathered data from a large number of other loT devices.
[0151] The zones that are entered by a device are processed as sets of sequences and then appropriate machine learning techniques such as a recurrent neural network (RNN) or deep learning architecture such as Transformer are preferably used to create a neural network that can predict the next zone entry.
[0152] For example:
[0153] Input: [A, B, C] -> Target: [D]
[0154] Input: [A, B] -> Target: [C]
[0155] Input: [F, B, A, B] -> Target: [C]
[0156] In which A,B,C,D (letters) are zones entered, both inferred, manually provided and both GeoZones and RF Zones.29 April 2026
[0157] The result is ultimately a machine learning trained model that is able to predict multiple next Zone entries as a sequence:
[0158] For example,
[0159] Input: [A, B, C, D, E]
[0160] Output: [{D: 0.8, E: 0.1, F: 0.1}, {F: 0.8, G: 0.1, H: 0.1}, {G: 0.8, I: 0.1, J: 0.1}] Prediction sequence: [D, F, G]
[0161] Based on the input the sequence of next three zones produced as an output is D, F and G. This can be seen from the exemplary data which shows as the first “next” zone, the probabilities for each of the tree candidates D, E and F are, respectively, 0.8, 0.1 and 0.1. Thus, the model is able to predict that the next zone to be entered, with a probability of 80% is zone D.
[0162] When a zone prediction is performed, and a grading is also possible given an IOT device, risk rules can be used. The appropriate reports and alerts can be raised to users and other machine-based agents to take appropriate action. Both the confidence of the next zone predictions, the risk grading of the predicted zone and the confidence of the risk rating can be taken into account.
[0163] Figure 5A is a schematic view of a plurality of identified or inferred zones in a logistics distribution system. Initially an IOT device has entered zone A 60 and then moved to B 62. Both zones are previously calculated by collation of data and have been determined or inferred to be zones.
[0164] Models are trained and exist for each of the zones. Given the risk criteria (risk rules) of the IOT device, a grade is calculated for each zone. The calculated risk grade for zone A 60 is B whereas the calculated risk grade for zone B 62 is D. Given the machine learnt model and the passage of other IOT devices which has been used to train the model, IOT device is predicted to enter zone C 64 with an eighty percent probability. The calculated risk grade of zone C is F.29 April 2026
[0165] Figure 5B shows the risk scoring for an IOT device entering zone C. As can be seen, the risk score is F and the IOT device is considered ninety percent likely to enter zone C next. There is a ninety percent probability that the IOT device will be exposed to temperature outside of its range for twenty minutes. Thus, the system is able to determine that the IOT device is at some significant risk. In the example given above of the transport of food it can be understood that this information, in real time, is extremely useful.
[0166] The IOT device is not an abstract device moving within a logistics chain, but rather is a digital twin of some real world goods or products that are moving associated with the IOT device. Accordingly, knowledge of the risk or probability of an IOT device moving from one zone to another and the risks that it is likely to encounter along the way, is a twin for knowledge of the risks to the associated goods. It is therefore extremely useful in making technical determinations regarding the management and control of the movement of the products or goods themselves. Specifically, if it is determined that there is a ninety percent chance of the IOT device entering zone C which itself has a ninety percent risk of exposure outside its temperature range for twenty minutes, then alternative action can be taken and the IOT device and associated products can become instead, redirected to an alternative destination. Thus, the process and system provides a way in which actively and in real time, the movement of goods can be controlled in such a way as to minimise risk to the goods during transit.
[0167] The example described above was with relation to temperature of the products. Specifically, the example was given of the risk of excursion beyond the accepted temperature range and for some specified time period. This is merely exemplary and alternative risks could be, for example, the chance that the goods are placed within an environment where physical tampering is highly likely. This could relate, to a particular warehouse or transit station.
[0168] Figures 6 and 7 are schematic flow diagrams showing the processes of model preparation and model usage according to the present system and method.
[0169] Referring to figure 6, data is collected by an IOT device 68 and processed 70 into zones, as described above. Grading scores are generated based on the zone machine29 April 2026
[0170] learning model, again as described above with reference to Figures 4A, 5a and 5B. In parallel at step 74, the zone sequence is recorded. In other words the sequence of zones through which the loT device actually moved is recorded preferably based on data it transmits back to the central server 18.
[0171] At step 76, machine learning is trained so as to predict the sequence of zones through which IOT devices are expected to move. At steps 78 and 80, the zone grading model and the sequence model are both deployed for use. At step 82, the models are ready for usage in the real world. It will be understood that the models are not static. In real time loT devices are moving and so continually providing new and updated data to the sever 18 for processing. The models that the server or system therefore generate are continually being updated.
[0172] Referring to figure 7, in practice, an IOT device is paired with its “digital twin” 84. This can be a particular container of goods for transit. At 86, the risk rules of the IOT device are established which might typically relate to the type of product being transported. For example, if the product being moved is a consumable foodstuff, the rules might specify “no exposure to a temperature above X degrees centigrade for a period of more than Y minutes”. If it is a sophisticated piece of computer hardware, the rules might relate to limits on the movements or acceleration to which it can be subject.
[0173] At step 88, the IOT device reports to the server 18 which enables the most recent zone established 90 to be allocated. The model identified at step 80 (figure 6) can then be used to predict 92 the next zone for that particular IOT device.
[0174] The predicted zones grading models are matched against the risk rules for the IOT device 94. At step 96, a determination is made as to whether or not there is any risk rule exception. If there is not, a store grade is predicted and the journey is continued zone 98. If a risk is identified 100 a risk notification is produced which again feeds to the store grade prediction 98 before the IOT device is able to continue its journey 102.
[0175] Thus, by this process, in real time as an IOT device moves in transit, risk notifications are produced which can be used to modify the route of an IOT device to minimise the risk to the products associated with the device. The risk notification can be generated29 April 2026
[0176] automatically and transmitted to a controller (human or computer) to determine alternative or evasive action.
[0177] Embodiments of the present invention have been described with particular reference to the examples illustrated. However, it will be appreciated that variations and modifications may be made to the examples described within the scope of the present invention.
Claims
29 April 2026Claims1. A system for tracking objects, the system comprisingplural loT devices each for association with a corresponding object;a server for communication with each of the loT devices;each of the loT devices being arranged to gather data associated with the object with which it is in use associated and transmit the gathered data to the server;the system being arranged, in dependence on the gathered data, to infer zones of interest in relation to the loT devices.
2. A system according to claim 1 , in which the environmental data is selected from data relating to the RF environment and data relating to the physical environment of the loT device.
3. A system according to claim 2, in which the data relating to the RF environment includes one or more of Cellular, Wi-Fi, Lora, UWB or BLE signal strength experienced by the loT device.
4. A system according to claim 2 or 3, in which the data relating to the physical environment includes one or more of temperature, humidity, pressure, movement, and acceleration.
5. A system according to any of claims 1 to 4, arranged to determine zones using a clustering algorithm based on data gathered from the plural loT devices.
6. A system according to any of claims 1 to 5, in which zones are inferred on the basis of gathered data from multiple loT devices and the application of rules-based clustering.
7. A system according to claim 6, in which rules-based clustering is executed based on gathered environmental data and / or geographical location.
8. A system according to any of claims 1 to 7, in which upon introduction of an loT data is collected by the loT device and communicated to the server.29 April 20269. A system according to claim 8, in which based on the received data the server establishes the geographic location or the RF zone in which an loT device is located.
10. A system according to claim 9, in which the data gathered from the loT is added to stored data from other loTs.
11. A system according to any of claims 1 to 10, in which the formation of a new inferred zone comprises the steps of:receiving at the server the data from multiple loT devices;defining clustering criteria;apply clustering criteria to identify loT devices belonging to respective clusters; and,establish boundaries of clusters based on identification of loT devices belonging to the clusters.
12. A system according to claim 11, in which the data received from the loT devices is one or more of geographic data and data relating to a detected RF environment.
13. A system according to claim 11 or 12, in which the clustering criteria are modified if the number of identified loT devices is not within a desired range for an identified cluster.
14. A system according to claims 1 to 13, in which a risk to products associated with each of the loT devices is determined in dependence on the inferred zone in which the loT device in question is located.
15. A system according to any of claims 1 to 14, comprising developing a model of multiple inferred zones and determining likely trajectory of a particular loT in dependence on previous movements.
16. A system according to claim 15, comprising building the model of expected movements of loTs using machine learning.29 April 202617. A system according to claim 15 or 16, in which the model includes likely risks to the loT within one or more of the inferred zones or on its passage through and between inferred zones.
18. A system according to any of claims 1 to 17, in which the gathered data associated with the object with which it is associated is encoded at the loT device.
19. A system according to claim 18, in which the encoding is selected to reduce data transmission to the server.
20. A system according to claim 19, in which the reduction of data transmission comprises reduction in data volumes, bandwidth or rate.
21. A system according to any of claims 18 to 20, in which the gathered data is encoded to an indication of one or more of Human Presence Event, Change in Custody and Modality in relation to the loT device in question.
22. A system according to any of claims 18 to 21, in which the loT device includes a model for computing an embedding vector as an encoding of the associated data, the embedded vector representing the context of gathered data, and wherein the embedding vector is transmitted to the server and used to retrieve similar embeddings associated with prior tracked objects or zones.
23. A system according to claim 22, wherein the embedding vector is used to perform one or more of: classifying a shipment risk, predicting journey outcomes, or recommending data acquisition policy.
24. A system according to 22 or 23, wherein the model for computing an embedding vector is partially trained or updated at the loT device using local data, or via updates received from the server.
25. A system according to any of claims 22 to 24, wherein the loT device is arranged to participate in a federated learning framework, such that updates to the embedding29 April 2026model are computed locally at the loT device using gathered data, and only model updates or gradients are transmitted to the server, thereby improving the embedding model across multiple devices without requiring transmission of raw data.
26. An loT device for use in tracking objects in a system for tracking objects including plural loT devices each for association with a corresponding object, a server in communication with each of the loT devices, in which the loT devices are arranged in use to gather data associated with the object with which it is associated and transmit the data to the server, in which the server is arranged, in dependence on the received data, to infer zones of interest in relation to the loT devices, the IOT device comprising:an RF detector to detect an RF environment of the loT device;a geographic position detector to detect an RF environment of the loT device a transmitter to transmit to the server the gathered data.
27. An loT Device according to claim 26, comprising a processor and a rechargeable power source to provide power to the components of the loT tag.
28. An loT device according to claim 26 or 27, in which the gathered data associated with the object with which it is associated, is encoded at the loT device.
29. An loT device according to claim 28, in which the encoding is selected to reduce data transmission to the server.
30. An loT device according to claim 29, in which the reduction of data transmission comprises reduction in data volumes, bandwidth or rate.
31. An loT device according to any of claims 28 to 30, in which the gathered data is encoded to an indication of one or more of Human Presence Event, Change in Custody and Modality in relation to the loT device in question.
32. An loT device according to any of claims 26 to 31, in which the loT device includes a model for computing an embedding vector as the encoding of the associated data, the embedded vector representing the context of gathered data, and wherein the29 April 2026embedding vector is transmitted to the server and used to retrieve similar embeddings associated with prior tracked objects or zones.
33. An loT device according to claim 32, wherein the embedding vector is used to perform one or more of: classifying a shipment risk, predicting journey outcomes, or recommending data acquisition policy.
34. An loT device according to 32 or 33, wherein the model for computing an embedding vector is partially trained or updated at the loT device using local data, or via updates received from the server.
35. An loT device according to any of claims 32 to 34, wherein the loT device is arranged to participate in a federated learning framework, such that updates to the embedding model are computed locally at the loT device using gathered data, and only model updates or gradients are transmitted to the server, thereby improving the embedding model across multiple devices without requiring transmission of raw data.
36. A method for tracking objects, the method comprising:providing plural loT devices each associated with a corresponding object to be tracked;providing a server in communication with each of the loT devices;each of the loT devices gathering environmental data associated with the object with which it is associated and transmitting the gathered environmental data to the server;at the server, in dependence on the gathered data, inferring zones of interest in relation to the loT devices.
37. A method according to claim 36, in which the environmental data is selected from data relating to the RF environment and data relating to the physical environment of the loT device.29 April 202638. A method according to claim 37, in which the data relating to the RF environment includes one or more of Cellular, Wi-Fi, Lora, UWB or BLE signal strength experienced by the loT device.
39. A method according to claim 37 or 38, in which the data relating to the physical environment includes one or more of temperature, humidity, pressure, movement, and acceleration.
40. A method according to any of claims 36 to 39, comprising determining zones using a clustering algorithm based on data gathered from the plural loT devices.
41. A method according to any of claims 36 to 40, comprising inferring zones on the basis of gathered data from multiple loT devices and the application of rules-based clustering.
42. A method according to claim 41, comprising executing rules-based clustering based on gathered environmental data and / or geographical location.
43. A method according to any of claims 36 to 42, in which upon introduction of an loT data is collected by the loT device and communicated to the server.
44. A method according to claim 43, in which based on the received data the server establishes the geographic location or the RF zone in which an loT device is located.
45. A method according to claim 44, in which the data gathered from the loT is added to stored data from other loTs.
46. A method according to any of claims 36 to 45, in which the formation of a new inferred zone comprises the steps of:receiving at the server the data from multiple loT devices;defining clustering criteria;apply clustering criteria to identify loT devices belonging to respective clusters; and,29 April 2026establishing boundaries of clusters based on identification of loT devices belonging to the clusters.
47. A method according to claim 46, in which the data received from the loT devices is one or more of geographic data and data relating to a detected RF environment.
48. A method according to claim 46 or 47, in which the clustering criteria are modified if the number of identified loT devices is not within a desired range for an identified cluster.
49. A method according to any of claims 36 to 48, in which a risk to products associated with each of the loT devices is determined in dependence on the inferred zone in which the loT device in question is located.
50. A method according to any of claims 36 to 49, comprising developing a model of multiple inferred zones and determining likely trajectory of a particular loT in dependence on previous movements.
51. A method according to claim 50, comprising building the model of expected movements of loTs using machine learning.
52. A method according to claim 50 or 51, in which the model includes likely risks to the loT within one or more of the inferred zones or on its passage through and between inferred zones.
53. A method according to any of claims 36 to 52, in which the gathered data associated with the object with which it is associated, is encoded at the loT device.
54. A method according to claim 53, in which the encoding is selected to reduce data transmission to the server.
55. A method according to claim 54, in which the reduction of data transmission comprises reduction in data volumes, bandwidth or rate.29 April 202656. A method according to any of claims 53 to 55, in which the gathered data is encoded to an indication of one or more of Human Presence Event, Change in Custody and Modality in relation to the loT device in question.
57. A method according to any of claims 53 to 56, in which the loT device includes a model for computing an embedding vector as the encoding of the associated data, the embedded vector representing the context of gathered data, and wherein the embedding vector is transmitted to the server and used to retrieve similar embeddings associated with prior tracked objects or zones.
58. A method according to claim 57, wherein the embedding vector is used to perform one or more of: classifying a shipment risk, predicting journey outcomes, or recommending data acquisition policy.
59. A method according to claims 57 or 58, wherein the model for computing an embedding vector is partially trained or updated at the loT device using local data, or via updates received from the server.
60. A method system according to any of claims 57 to 59, wherein the loT device is arranged to participate in a federated learning framework, such that updates to the embedding model are computed locally at the loT device using gathered data, and only model updates or gradients are transmitted to the server, thereby improving the embedding model across multiple devices without requiring transmission of raw data.
