Location sensing system for personal care or health device

Abstract

A location sensing system for a personal health device or personal care device is provided for sensing a location of the personal health device or personal care device during a use session. Motion information from the IMU up to a current point in time in a current use session is processed along with historical data relating to previous use sessions that includes location information over time for those previous use sessions. Historical data is based on post-session offline analysis of motion information of those previously used sessions. And obtaining the current position in real time based on the motion information and historical data.

Description

Technical Field

[0001] The present invention relates to position sensing systems and, in particular, to position sensing of handheld devices such as personal health devices or personal care devices. Background Art

[0002] Many handheld devices are equipped with inertial monitoring unit (IMU) sensors to monitor their movement and position. A typical example is a mobile phone. However, the use of position (i.e. positioning) sensing is also increasing in personal care devices or personal health devices or personal care devices such as electronic toothbrushes and shavers. For such devices, IMU data is used to locate where the user is applying the device.

[0003] For example, position sensing can be used to determine the tooth segment (i.e., adjacent group of teeth) that the user is currently brushing, which enables feedback to the user on the quality of their brushing session based on the time spent on different tooth segments. With this feedback, the user can improve their daily brushing habits.

[0004] IMU sensors only provide movement information, not absolute position information. In order to determine absolute position information, there needs to be a known starting position, and then the IMU movement information can be used to track movement from that position, thereby tracking changes in position.

[0005] For a toothbrush, the position information of interest is the segment of the teeth that the user is brushing as mentioned above, so a report on how well different areas are brushed can be provided. Detecting which segment the user is brushing based on the toothbrush IMU signal is a challenging problem because only the toothbrush motion is observed, but what is being observed is the relative motion between the toothbrush and the teeth of interest. When the user moves around in their bathroom, by measuring the movement of the head, this will generate additional signals that the algorithm (e.g., AI model) must learn to ignore.

[0006] Additionally, it is possible (in theory) to brush your teeth by holding the toothbrush still (or nearly still) and moving your head and mouth around the toothbrush. In this case, the toothbrush IMU would barely measure any acceleration signal, making it difficult or nearly impossible to detect these segments.

[0007] However, brushing in certain segments (and their transitions) is often done with specific motions, so the resulting IMU signals can be used to infer where the toothbrush is in the mouth. For example, it is known to detect motion patterns, and using these detected patterns, segment positions can be inferred throughout the brushing session.

[0008] In principle there are two ways to analyze IMU sensor signals:

[0009] Offline, where all IMU sensor data is recorded and, after the session is complete, the entire session is analyzed as a whole; and

[0010] Real-time (online), where IMU sensor data is used continuously by algorithms to provide predictions during a session.

[0011] Based on offline analysis, existing toothbrushes can determine how long the user spent brushing in each of the different (e.g., 12) tooth segments after a brushing session. In addition, the algorithm can also calculate pressure maps and brushing maps for each segment.

[0012] With the help of real-time analysis, new and smarter toothbrushes can be manufactured. Based on the real-time position, the toothbrush motor can be controlled or direct user feedback can be provided (for example with the help of lights or sounds). Although real-time evaluation enables valuable features, it is more difficult than offline analysis and therefore usually results in lower positioning accuracy.

[0013] For the offline algorithm, the prediction of the brushing segment at time T is based on the IMU data from the entire brushing session (t = [0..Tend]), thus using data before and after time T. In the real-time model, only past sensor data can be used, thus only sensor data from t = [0..T].

[0014] Furthermore, real-time models often need to be run inside a device, where cost and power constraints limit the available computer hardware. Offline models often run on a separate device, such as a mobile phone, or even in the cloud. As a result, the computational complexity of offline models can be much higher than real-time models, resulting in higher accuracy.

[0015] Therefore, there is a need to improve the accuracy of real-time positioning, and in particular, to reduce the accuracy gap between the offline model and the real-time model.

[0016] US2020 / 211414 discloses a toothbrush that monitors acceleration and generates an n-dimensional space of acceleration components. The acceleration components form clusters corresponding to different tooth segments. The cluster definitions are stored in a memory.

[0017] US2021 / 393026 discloses an oral care device having a motion sensor and an orientation sensor. The system uses a training oral care routine to learn how to interpret the motion sensor information and the orientation sensor information.

[0018] US2020 / 179089 discloses another motion tracking toothbrush. Machine learning is used to enable interpretation of motion sensor data. Calibration is used to provide a statistical model of the shape and size of a specific user's mouth.

[0019] EP 3 858 288 discloses an oral care system for interdental space detection, which combines position data with training data. Summary of the invention

[0020] The invention is defined by the claims.

[0021] According to an example according to aspects of the present invention, there is provided a position sensing system for a personal health device or a personal care device for sensing the position of the personal health device or the personal care device during a usage session, the position sensing system comprising:

[0022] an inertial monitoring unit to provide motion information related to the motion of a personal health device or a personal care device; and

[0023] A processor, configured to process the motion information to obtain position information, wherein the processor is configured to:

[0024] receiving movement information up to a current point in time in a current usage session;

[0025] receiving historical data related to previous usage sessions, the historical data including location information over time for the previous usage sessions, the historical data based on a post-session offline analysis of motion information for those previous usage sessions; and

[0026] Based on motion information and historical data, current location is determined in real time.

[0027] The system combines the flexibility of real-time location sensing with the accuracy of offline post-session analysis. This offline post-session analysis utilizes an algorithm that derives location from the session's prior and future movement information and thus reconstructs location data more accurately than real-time analysis. By combining this offline analysis with real-time motion sensing, more accurate real-time location sensing is possible.

[0028] The real-time analysis only has the current motion sensing signal and the previous motion sensing signal, and with only this analysis the location can be ambiguous, so there are various possible locations. Only by reasoning over the entire session (so this must be performed offline) is it possible to determine the correct location and therefore reduce the ambiguity. This ambiguity is reduced by injecting predictions from an offline model. This is often used for personal health devices or personal care devices, where the user performs a session in a consistent and repetitive manner, such as a fixed brushing or shaving routine.

[0029] The system can be operated in particular so that real-time performance improves gradually by using offline models, which are easier to train over time. Real-time models are usually simpler than offline models, which match the amount of computing power available in the device (real-time) compared to offline models (offline, in the application or in the cloud). The term offline is simply used to mean not real-time, and therefore one can freely choose where and when to perform processing, as latency in processing is not an issue.

[0030] For example, historical data includes an average of determined positions over time for a number of previous usage sessions. It therefore represents an average sequence of positions. It can be thought of as representing the probability of a particular position, or the position with the highest probability.

[0031] For example, the determined position comprises a position category, which is one of a set of position categories. For example, these categories are ranges of positions having a common meaning, such as all positions corresponding to a tooth or a set of teeth forming a specific tooth segment.

[0032] If the position sensing system is used for a shaver, each position class corresponds, for example, to a facial area.

[0033] The processor is configured, for example, to update the historical data after the current usage session is completed based on a post-session offline analysis of the motion information of the current usage session. Thus, the historical data can be updated using a more accurate analysis of the current session (using the offline model). This update can be performed after each usage session, or can be performed less frequently.

[0034] The processor is configured to update the historical information using a moving average update, for example. This updates the historical session information by combining the last current session information (after offline analysis) with the existing historical session information.

[0035] The inertial monitoring unit includes, for example, a three-axis accelerometer and a three-axis gyroscope, thereby collecting six-dimensional motion data.

[0036] A personal health system or personal care system may be provided that combines a personal health device or personal care device, and a position sensing system as defined above, and a computer program for performing an offline post-session analysis of motion information of previous use sessions and generating historical data related to those previous use sessions. The computer program runs on a remote device separate from the personal health device or personal care device, and the remote device and the personal health device or personal care device are capable of communicating with each other.

[0037] The present invention also provides a powered toothbrush, comprising:

[0038] a toothbrush handle having an interface for connecting to a toothbrush head;

[0039] A driving motor for driving the interface to drive the toothbrush head; and

[0040] The position system defined above is used to determine the position of the toothbrush head during a usage session.

[0041] The present invention also provides an electric shaver, which comprises:

[0042] shaver head;

[0043] a drive motor for driving the shaving head; and

[0044] The position system defined above is used to determine the position of the shaver head during a usage session.

[0045] The present invention also provides a method for determining the location of a personal health device or personal care device during a usage session, the method comprising:

[0046] receiving motion information related to motion of the personal health device or personal care device up to a current point in time in a current usage session;

[0047] receiving historical data related to previous usage sessions, the historical data including location information over time for the previous usage sessions, the historical data based on a post-session offline analysis of motion information for those previous usage sessions; and

[0048] Based on motion information and historical data, current location is determined in real time.

[0049] Historical data is based on offline analysis (rather than simply a storage of real-time analysis from previous sessions), so it can increase the accuracy of real-time analysis more than simply using historical real-time analysis. For example, real-time analysis is performed using the local processing power of the device, while offline analysis is performed remotely.

[0050] The method comprises, for example, deriving historical data by averaging the determined positions over time of a number of previous use sessions. Real-time sensor data from the current session is blended with the averages from previous sessions in this way. This avoids the need for the user to be highly consistent with their personal care routine. The historical data provides a set of expected positions over time.

[0051] The method includes, for example, updating the historical data based on a post-session offline analysis of the motion information of the current usage session after the current usage session is completed. For example, the update uses a moving average update.

[0052] The invention also provides a computer program comprising computer program code means adapted to implement the method defined above when said program is run on a computer.

[0053] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] For a better understanding of the invention, and in order to more clearly show how the invention may be put into practice, reference will now be made, by way of example only, to the accompanying drawings, in which:

[0055] Figure 1 A position system applied to a powered toothbrush is shown;

[0056] Figure 2 An offline evaluation algorithm is shown;

[0057] Figure 3 Shows how historical data is updated;

[0058] Figure 4 A real-time evaluation algorithm is shown;

[0059] Figure 5 A typical IMU signal during a complete brushing session is shown;

[0060] Figure 6 Shown based on Figure 5 Data of,identified brushing sessions over time;

[0061] Figure 7 The model used for offline evaluation is shown;

[0062] Figure 8 A model for real-time evaluation is shown;

[0063] Fig. 9 Four examples of different user sessions are shown (averaged over multiple sessions for each user);

[0064] Fig.10 shows the variation of segment probabilities over time, averaged over all users and all sessions in the database; and

[0065] Fig.11 A location determination method is shown. DETAILED DESCRIPTION

[0066] The present invention will be described with reference to the accompanying drawings.

[0067] It should be understood that the detailed description and specific examples, while indicating exemplary embodiments of the apparatus, system, and method, are for illustrative purposes only and are not intended to limit the scope of the invention. These and other features, aspects, and advantages of the apparatus, system, and method of the invention will be better understood through the following description, the appended claims, and the accompanying drawings. It should be understood that the drawings are merely schematic and are not drawn to scale. It should also be understood that the same reference numerals are used throughout the drawings to indicate the same or similar components.

[0068] The present invention provides a position sensing system for a personal health device or personal care device for sensing the position of the personal health device or personal care device during a usage session. Motion information from an IMU up to a current point in time in a current usage session is processed together with historical data related to previous usage sessions, the historical data including position information over time for those previous usage sessions. The historical data is based on post-session offline analysis of the motion information for those previous usage sessions. Based on the motion information and the historical data, the current position is derived in real time. This approach can avoid the need for intermediate motion pattern detection.

[0069] As explained above, offline evaluation of position is easier because all data of the session are available to interpret the movements, but the results of the analysis are only available after the session has been completed. For the example of providing analysis of brushing performance, the best that can be done is to tell the user that he / she should brush his / her teeth differently based on this information. The advantage of real-time analysis is that it provides immediate feedback to the user.

[0070] For real-time analysis, only session samples up to the current time t can be used, and information about future samples in the session is not available for interpretation. This is a well-known technical problem for real-time evaluation.

[0071] For the example of tracking the position of the toothbrush head, at the start of a brushing session, the Inertial Monitoring Unit (IMU) data does not provide any information about the position in the mouth. It only contains information about the orientation of the toothbrush head (which is absolutely known due to the known gravity vector). If the toothbrush head is oriented upwards, then it can be inferred that the user started brushing the chewing surfaces of the upper teeth, but it does not contain information about the horizontal position in the upper jaw. Clues about the horizontal position can only be drawn from future samples of the session. For example, if future samples indicate a significant horizontal movement of the brush head to the left, then it can be concluded that the user started somewhere on the right side of the jaw.

[0072] This example illustrates how the assessment of brush head position at time t benefits from the availability of session data recorded after time t. In the case of offline assessment, this "future" data is available, but not for real-time assessment.

[0073] For example, two algorithms for localization were trained using the same IMU dataset. The offline algorithm achieved a test accuracy of 84.0%, compared to 75.7% for the real-time evaluation.

[0074] A well-known compromise between real-time and offline analysis is to perform the analysis within a short delay Δt. In this case, slightly more data can be analyzed (until t+Δt), but the feedback to the user will also be delayed by the same amount of time. This delay is in addition to the computational time delay of the algorithm.

[0075] The present invention significantly improves the accuracy of real-time assessment without introducing a delay Δt. Only data available at time t is used.

[0076] Figure 1 The position system of the present invention is shown applied to a powered toothbrush.

[0077] The powered toothbrush 10 comprises a toothbrush handle 12 having an interface for connecting to a toothbrush head 14. The handle comprises an interface drive motor 16 for driving the toothbrush head.

[0078] A position sensing system is used which determines the position of the toothbrush head during a usage session. The position sensing system comprises an inertial monitoring unit 20 for providing motion information relating to the motion of the toothbrush.

[0079] The inertial monitoring unit includes, for example, a three-axis accelerometer and a three-axis gyroscope.

[0080] A processor 22 is provided as part of the toothbrush for processing the motion information to derive position information.

[0081] The processor receives motion information from the IMU up to the current point in time in the current use session. However, it also receives historical data relating to previous use sessions, the historical data including position information over time for the previous use sessions, the historical data being based on post-session offline analysis of the motion information for those previous use sessions. This data is stored in the database 24. However, it is preferably generated remotely, rather than using the processor 22.

[0082] Such processing may be performed, for example, by a processor of the mobile phone 30 or in the cloud 40. For example, the toothbrush sends data to the mobile phone 30 which acts as a user interface for the toothbrush, and it is used to convey information to the user via an application on the mobile phone.

[0083] Based on the motion information and historical data, the current position is determined in real time. Therefore, the system uses two inputs. One is the regular IMU sensor data of the current session up to the current time t, and the other is the historical data related to the information that has been extracted from the historical sessions.

[0084] Therefore, the solution employs two independent analysis algorithms:

[0085] The offline part creates historical data hd[0..t] for each user (which is a function of time);

[0086] A real-time part which derives a real-time evaluation from the IMU input IMU[0..t] and the historical data hd[0..t].

[0087] [0..t] specifies the time interval of data from the start of the session (t=0) to the current time t.

[0088] The purpose of the offline part is to capture information about the device movements that have been observed in previous sessions. The basis of the offline part is a more accurate offline evaluation algorithm. This algorithm is applied to complete the sessions that have been collected in the past. This analysis results in prediction data about the expected position over time during the session. For example, the prediction data of all previous sessions of the user are accumulated over time and averaged to generate historical data hd.

[0089] The historical data hd is a two-dimensional data array of size n_time_samples (rows) × n_classes (columns). Therefore, for each time sample, a value, specifically a probability value, is assigned to each location class.

[0090] For the application of the localization of the toothbrush head, the mouth is subdivided into 12 segments, for example, so that n_class = 12. The determined position is therefore not recorded as an absolute position, but is assigned to a position class, which is one of a set of position classes.

[0091] The value n_time_samples corresponds to the length of the longest recorded session, i.e. the maximum number of time samples, or the default maximum number of time samples.

[0092] The historical data is initialized with data from sessions of many users (e.g., users in the total training set). After the algorithm is deployed in the device, the historical data can be updated with newly recorded data from the users of the device, but again using an accurate offline evaluation algorithm.

[0093] By updating historical data, information becomes more personalized and more relevant to specific users, which leads to performance improvements in real-time evaluation.

[0094] Figure 2 An offline evaluation algorithm 50 is shown. The input to the algorithm is the IMU sensor data for a complete session. The offline evaluation algorithm is used to calculate the probability of each position category for the session. Therefore, the output is the predicted probability of each position category in the complete input session.

[0095] Figure 3 It is shown how these probabilities are added to the historical data. The update function is represented by module 52.

[0096] Figure 4 A real-time evaluation algorithm 54 is shown. The updated historical data is used as input for the real-time evaluation. The input to the algorithm is the IMU sensor data and historical data for the session at the current time t. The output is the predicted position at time t.

[0097] With this approach, there is no need to ask the user to provide his / her starting position and brushing sequence as is currently the case for guided brushing using some existing systems.

[0098] The method will now be described in more detail.

[0099] The algorithm for offline evaluation of a full session starts with the IMU data of the full session, which has the shape [0…t 结束 ,0..nb 通道 ] in the form of a two-dimensional array, which usually contains nb 通道 = 6 channels, corresponding to t 结束 The x / y / z components of the accelerometer and gyroscope signals for the time step. Therefore, the data from the IMU sensor is time series data, and the accelerometer and gyroscope measure their components in the x-direction, y-direction, and z-direction, resulting in 6 channels sampled at about 30Hz. Assuming a typical brushing time of about 2 minutes, this results in a matrix of about 3600×6 values.

[0100] Figure 5 A typical IMU signal during a complete brushing session is shown. The upper graph shows the x, y, and z components from the accelerometer. The lower graph shows the x, y, and z components from the gyroscope.

[0101] The offline model predicts the probability that the user brushes their teeth in segment i at each time step, thus:

[0102] P 离线 =M 离线 (x)

[0103] Among them, p 离线 is a shape [t 结束 ,nb 节段 ] two-dimensional array. x represents the date of complete tooth brushing.

[0104] Since the offline model can be complex and may require a lot of memory (>1MB), it will usually run on a mobile phone or in the cloud. Therefore, it can be updated without updating the real-time model. The performance of the real-time model will get better for the offline model that performs better.

[0105] The historical data hd is a data with the shape [t 结束 ,nb 节段 ] and contains the average probabilities from previous sessions. The historical data will start with an initial version that has not been customized for a specific user. As an initial version of the historical data, the average segment probability can be used, which is predicted by the offline model on a dataset of large brushing sessions (i.e., training data).

[0106] Since all sessions are of different lengths, only a predetermined time period may be used, such as a 2 minute time period. This may be the first 2 minutes, since most people brush their teeth for this recommended time. This then defines the maximum number of time steps. Alternatively, any number of time steps may be permitted, but when brushing lasts longer than the predetermined time period, the historical data for the last time step may be repeated.

[0107] Starting from the initial historical data, updates and personalization can be performed by computing an exponential moving average update using the probabilities from the offline model:

[0108] hd 新 =(1-α)·hd 旧 +α·p 离线

[0109] Each new prediction of the offline model does not necessarily lead directly to an update of the historical data. Multiple offline predictions can also be combined in a single update of the historical model.

[0110] The exponential multiplier α is, for example, in the range of 0.5 to 0.05.

[0111] The real-time model takes the current IMU sample x[t] and the historical data hd[t] at the same time as input and predicts the current segment probability p[t], so:

[0112] p 离线 [t] = M 实时 (x[t],hd[t])

[0113] Therefore, in this example, only a single point in time in the historical data is needed at any particular time to perform real-time processing. Therefore, the computational overhead of merging historical data is minimal. A single data set is used that indicates the probability of being in different segments at that particular time.

[0114] The historical data enables the real-time algorithm to choose when in doubt (between tooth segments). Thus, the historical data guides (i.e. biases) the decisions made by the simpler real-time model (in the absence of historical data).

[0115] When implementing the model using long short-term memory (LSTM) 实时 ), it usually needs to update the internal state / hidden state at each time step. LSTM uses this hidden state to provide context and consistency over multiple time steps. Therefore, the model gets one more input (containing the previous hidden state) and one more output (the updated hidden state), so the model actually has 3 inputs and 2 outputs:

[0116] P 在线 [t], hide [t] = M 在线 (x[t],hd[t],hd[t-1])

[0117] At t=0, the hidden state is usually initialized with zeros.

[0118] The hidden state is a representation of the current state. As such, it includes (in an abstract representation) information about the current position and the velocity and acceleration of the device, and possibly inferred head / body movement information.

[0119] Figure 6 Shown based on Figure 5 Data of brushing sessions identified over time. Different positions on the y-axis represent 12 separate session identifiers (ie, numbered 0 to 11).

[0120] The 12 sections are:

[0121] 0: Upper right outer side

[0122] 1: Front upper lateral

[0123] 2: Upper left outer side

[0124] 3: Lower left outer side

[0125] 4: Anterior lower lateral

[0126] 5: Lower right outer side

[0127] 6: Upper right inner side

[0128] 7: Front upper medial

[0129] 8: Upper left inner side

[0130] 9: Lower left inner side

[0131] 10: Anterior lower inner side

[0132] 11: Lower right inner side.

[0133] The y-axis counts from 11 down to 0 in the positive y-direction.

[0134] The -1 minimum value on the y-axis is a value used to indicate that the segment cannot be reliably determined.

[0135] Accurate position sensing can be used to measure the performance of the position system in determining where the user is brushing their teeth and for algorithm training.

[0136] There are two ways to measure (rather than estimate) which segment a person is brushing their teeth. A motion capture system can be used, or a film can be recorded and each video frame (at 30Hz) manually annotated. This measured position information can be used as ground truth data for training the algorithm.

[0137] Typical human action recognition methods are solved as classification problems. To train the algorithm, a sliding time window is extracted from the IMU signal, for example, and classification is performed based on the majority of ground truth labels in the window. This defines the input for algorithm training. Based on the IMU signal, the model predicts the most likely category (e.g., segment identifier) ​​at time t. Different sliding window labeling methods are possible. One is to use the most frequent label on the sequence, and the other is to use the label at the last timestamp.

[0138] The problem with this approach is that the windows may not be consistent with the transitions between the ground truth labels. Furthermore, applying such a classifier would require applying the network multiple times, each time with the time series shifted by 1, resulting in many identical computations.

[0139] A similar problem is encountered in computer vision algorithms when segmenting images (also known as semantic segmentation), for which fully convolutional networks have been developed. Labels are provided for each time step, so the labels for all time steps can be predicted simultaneously. For this purpose, an AI model or AI network can be used that will receive [nb_time_steps × 6] values ​​as input and will predict [nb_time_steps × 12] class probability values ​​as output.

[0140] The 6 inputs correspond to the number of channels in the IMU signal (x, y and z components of the accelerometer and gyroscope signals). The 12 outputs correspond to the number of predicted tooth segments and thus Figure 6 The number of category and segment identifiers in .

[0141] Such a network will be very complex because the correct segment cannot be inferred from data from a short time window around that moment. For example, in order to infer whether the user starts from the front upper teeth or the front lower teeth, data flipping from top to bottom or bottom to top needs to be included so that this information can be used to correctly predict the start position. In practice, this means that the model should be able to properly model long-distance dependencies / long-term dependencies.

[0142] In one example, the offline model is based on a so-called U-Net that has been modified to work on time series data.

[0143] U-Net has been the go-to jack-of-all-trades network for image segmentation using medical datasets.

[0144] Figure 7 The U-Net architecture is shown. In U-Net, the input data is fed through the first encoder (E0), resulting in a first feature representation. Next, this feature representation is downscaled (using max pooling) and fed through the second encoder (E1), resulting in a second feature representation. This process repeats itself several times, resulting in several feature representations at different scales. Next, the lowest resolution feature representation is upscaled (using nearest neighbor upsampling) and combined (concatenated) with the second lowest resolution feature representation, and pulled through the decoder.

[0145] By repeatedly upscaling, concatenating with the encoder feature representation of the same resolution and pulling the result through the decoder, we can eventually obtain the decoded feature representation of the original input resolution.

[0146] Two output heads are applied to this final representation (Output 1 and Output 2), one for the 12 segment classification and one for the position regression along the tooth arc.

[0147] The power of U-Net comes from its multi-scale analysis of representations, so the data is analyzed not only at the original input resolution, but also at a smaller resolution (2x). This enables effective modeling of long-distance dependencies, which become more local at lower resolutions. For example, actions taken 17 seconds before or after a moment are 17×30=510 samples apart at the input resolution, but at resolution level 9, they are only 1 time step apart. This makes it possible to model these long-distance dependencies with relatively small convolution kernels.

[0148] Additionally, skip connections (horizontal connections from encoder_x to decoder_x) ensure strong gradient flow when applying backpropagation (used in optimizing weights), thus making U-Net less problematic to optimize than other networks of similar depth.

[0149] For IMU time series data, the convolutional layers are modified to be one-dimensional instead of two-dimensional.

[0150] Many variants of U-Net are known in the literature, most of which maintain the same Figure 7 The same schematic structure is shown, but the specific layers used in the encoder / decoder blocks are changed.

[0151] As an alternative to U-Net, a bidirectional LSTM can be used. This type of LSTM can only be applied to offline sessions and can reason both forward and backward in time. When training a real-time model with this type of bidirectional LSTM instead of a normal LSTM, similar results are obtained as with U-net.

[0152] The model evaluated in real time is, for example, the conv-LSTM model. Figure 8 As shown, this model consists of three parts, an encoder 70, a long short-term memory (LSTM) layer 72 and a classifier 74.

[0153] The encoder is made of one or more convolutional layers that convert the raw IMU signal into a higher dimensional representation. The IMU signal (with 6 channels) is concatenated with the historical data (hd, with 12 segments / classes), resulting in an input signal for the encoder with 18 channels. The output of the encoder is used as the input to the LSTM.

[0154] In the LSTM layer 72, time-dependent data is processed in time steps. For each time step, the input data is combined with the previous hidden state to obtain a new updated hidden state and output.

[0155] Finally, in the classifier 74, the output of the LSTM is used as input to a pair of convolutional layers (kernel size = 1) and rectified linear units (ReLU) (thus effectively a multilayer perceptron) to predict segment probabilities.

[0156] As explained above, historical data consists of average predictions from previous sessions. The average predictions may be for all users stored in the database.

[0157] Fig. 9 Four examples of different user sessions are shown (averaged over multiple sessions for each user). In each case, the y-axis has 12 segments and the x-axis is a time sample (2 minutes total). The grey shading indicates the probability level.

[0158] Fig.10 The variation of segment probabilities over time, averaged over all users and all sessions in the database, is shown.

[0159] from Fig.10It can be concluded that most people in the database start at segment 2 (hence the high probability near the upper left corner of the graph). Therefore, segment 2 has the highest probability at t=0. In addition, people will follow or switch segments in response to the timer buzzer, resulting in light and dark patches. Most people brush their teeth for 2 minutes (because it is recommended), resulting in a strong signal until 120 seconds × 30 Hz = 3600 samples.

[0160] Since everyone brushes their teeth (and shaves) in different orders and for different durations, when averaging only the sessions of the same user, we get the same Fig.10 The average probability plot is very different from the population mean (such as Fig. 9 ). Some users are very consistent, in that each session aligns perfectly with the previous session, resulting in very bright blocks. Others have a very random brushing style, resulting in a more evenly distributed probability map.

[0161] Fig.11 A method for determining the location of a personal health or personal care device during a usage session is shown. The method comprises, in step 80, receiving motion information relating to the movement of the personal health or personal care device in a current usage session up to a current point in time. In step 82, historical data relating to previous usage sessions is received. It comprises location information over time for previous usage sessions, the historical data being based on post-session offline analysis of the motion information for those previous usage sessions. In step 84, a current location is determined in real time based on the motion information and the historical data.

[0162] The present invention has been tested to demonstrate improved location sensing accuracy. The offline model alone achieved a validation accuracy of 84.0%. Without historical data input, the normal real-time model achieved a validation accuracy of 75.7%, and the method of the present invention achieved a validation accuracy of 81.5%. Each model was trained on the same data, and validation was performed on similar data from different users. As a result, the validation accuracy gap of about 8% between the offline model and the real-time model was reduced to just below 3%.

[0163] The above example utilizes IMU data to track how users use a device. If people use a device in a similar way across different sessions, other information can be used, such as information about applied pressure and button presses.

[0164] The main focus of the invention is on personal health devices, such as toothbrushes and shavers. However, it can also be used for other user devices. As long as the user uses the device in a user-specific, more or less consistent way, the usage history data can enable improved real-time predictions.

[0165] Variations to the disclosed embodiments may be understood and effected by those skilled in the art in practicing the claimed invention, through a study of the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality.

[0166] The functions implemented by a processor may be implemented by a single processor or by multiple separate processing units, which may be considered together to constitute a "processor". In some cases, such processing units may be remote from each other and communicate with each other by wired or wireless means (optional)

[0167] The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

[0168] The computer program may be stored / distributed on a suitable medium, such as an optical storage medium or solid-state medium provided with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems. (optional)

[0169] If the term "suitable for" is used in the claims or the description, it should be noted that the term "suitable for" is intended to be equivalent to the term "configured to". If the term "arranged" is used in the claims or the description, it should be noted that the term "arranged" is intended to be equivalent to the term "system" and vice versa.

[0170] Any reference signs in the claims should not be construed as limiting the scope.

Claims

1. A position sensing system for a personal health device or a personal care device (10), for sensing the position of the personal health device or the personal care device during a usage session, the position sensing system comprising: an inertial monitoring unit (20) for providing motion information related to the motion of the personal health device or the personal care device; as well as A processor (22) for processing the motion information to obtain position information, wherein the processor is configured to: Receive the motion information of the current usage session up to the current point in time, Characterized in that the processor is further configured to: receiving historical data related to a previous usage session, the historical data including location information over time for the previous usage session, the historical data based on a post-session offline analysis of the motion information for the previous usage session; and Based on the movement information and the historical data, a current position is determined in real time. 2 . The system of claim 1 , wherein the historical data comprises an average of the determined positions over time for a plurality of previous usage sessions.

3. The system of any one of claims 1 to 2, wherein the determined location comprises a location category, the location category being one of a group of location categories.

4. The system of claim 3, wherein the position sensing system is for a toothbrush (10), and wherein each position category corresponds to a tooth segment.

5. The system of any one of claims 1 to 3, wherein the position sensing system is for a shaver, and wherein each position category corresponds to a facial region.

6. The system according to any one of claims 1 to 5, wherein the processor (22) is configured to update the historical data based on a post-session offline analysis of the motion information of the current usage session after the current usage session is completed.

7. The system of claim 6, wherein the processor (22) is configured to update the historical information using a moving average update.

8. The system according to any one of claims 1 to 7, wherein the inertial monitoring unit (20) comprises a three-axis accelerometer and a three-axis gyroscope.

9. A powered toothbrush comprising: A toothbrush handle (12) having an interface for connecting to a toothbrush head (14); A driving motor (12) for driving the interface to drive the toothbrush head; as well as A position sensing system according to any one of claims 1 to 8, for determining the position of the toothbrush head during a usage session.

10. An electric shaver, comprising: shaver head; A driving motor for driving the shaver head; as well as A position sensing system according to any one of claims 1 to 8, for determining the position of the shaver head during a usage session.

11. A method for determining the location of a personal health device or personal care device during a usage session, comprising: (80) receiving motion information related to motion of the personal health device or the personal care device up to a current point in time for a current usage session; Characterized in that the method further comprises: (82) receiving historical data associated with a previous usage session, the historical data including location information over time for the previous usage session, the historical data based on a post-session offline analysis of the motion information for the previous usage session; and (84) Determine the current location in real time based on the movement information and the historical data.

12. A method according to claim 11, comprising deriving the historical data by averaging the positions determined over time for a plurality of previous usage sessions.

13. The method according to claim 11 or 12, comprising updating the historical data based on a post-session offline analysis of the motion information of the current usage session after the current usage session is completed.

14. The method of claim 13, comprising updating the historical information using a moving average update.

15. A computer program comprising computer program code means adapted to implement the method according to any one of claims 10 to 14 when said program is run on a computer.

Images