Machine learning for personal care devices

Abstract

Systems, devices, and methods are described that apply machine learning to provide relevant feedback for personal care devices. System, device, and method embodiments can identify data from a device, such as a razor that includes multiple sensors on the razor, use a machine learning model trained using data from the multiple sensors to predict one or more characteristics, such as shaving characteristics, based on the data, and provide shaving feedback based on the one or more characteristics, such as shaving characteristics. System, device, and method embodiments can train the machine learning model, execute the trained machine learning model, ingest and / or process data for the machine learning model, provide feedback based on the output of the trained machine learning model, or perform a combination thereof.

Description

Technical Field

[0001] (Cross - reference to related applications) This application claims priority to U.S. Provisional Patent Application No. 63 / 043,451, filed on Jun. 24, 2020, and claims priority to International Patent Application PCT / US2021 / 038856, filed on Jun. 24, 2021, the contents of which are hereby incorporated by reference in their entirety.

[0002] The following generally relates to personal care, and more specifically, to machine learning for personal care devices.

Background Art

[0003] Many people use various instruments and devices for personal care and grooming. The effectiveness of any such instrument or device depends on factors such as how it is used, whether the instrument or device is in good condition, or whether repair or replacement (of any or all of its components) is necessary. For this reason, the useful life of such an instrument or device (or at least one or two or more of its components) is often limited, and there is a possibility that its effectiveness may decrease towards the end of its useful life.

[0004] As long as the user continues to use the product beyond its useful life or after the product generally requires component replacement or repair, the user may have an unsatisfactory experience and potentially suffer injury. Therefore, there is a need for a system and method for obtaining and providing user feedback regarding the use and technology of such personal care and grooming products and the status of such instruments and / or devices.

[0005] A non-exclusive example of a personal care and grooming device or instrument is a razor. Many people use razors for personal care and grooming. The effectiveness of a razor can depend on factors such as how it is used and whether the razor blade is sufficiently sharp. In other words, there is a limit to the effective life of a razor blade, and its effectiveness may decrease near the end of its effective life.

[0006] If a user's shaving technique is inadequate, or if the razor is nearing the end of its effective life, the user may have a poor user experience or even injure themselves. Therefore, there is a technical need for systems and methods for obtaining and providing user feedback regarding shaving technique and razor status. [Overview of the project]

[0007] Personal care devices may include those used for grooming or cleaning, such as razors, trimmers, dermaplaning tools, facial and body brushes and scrubbers, waterpiks, dental cleaning tools, and toothbrushes. Personal care devices may also include those that require replacement because they have an effective life, whether the replacement is for the entire personal care device or for its components.

[0008] Methods, apparatus, non-temporary computer-readable media, and systems for machine learning of personal care devices are described. Embodiments of the methods, apparatus, non-temporary computer-readable media, and systems can identify data from a personal care device that includes multiple sensors on the personal care device, predict one or more usage features based on the data using a machine learning model trained with data from the multiple sensors, and provide feedback based on one or more usage features.

[0009] The device may include a processor, memory that communicates electronically with the processor, and instructions stored in the memory. The instructions can enable the processor to identify data from a personal care device including multiple sensors, predict one or more usage features based on the data using a machine learning model trained with data from the multiple sensors, and provide feedback based on one or more usage features.

[0010] Non-temporary computer-readable media can store codes for personal care. In some examples, the codes include instructions that can be executed by a processor to identify data from a personal care device including multiple sensors, predict one or more usage features based on the data using a machine learning model trained with data from the multiple sensors, and provide feedback based on one or more usage features.

[0011] In some examples, the multiple sensors include inertial measuring units (IMUs), capacitance-to-digital converters, force sensors, rinse sensors, grip sensors, temperature sensors, accelerometers, gyroscopes, magnetometers, resistive contacts, piezoelectric contact microphones, Hall effect sensors, external cameras, strain gauges, load cells, reed switches, linear variable differential transformers, or any combination thereof.

[0012] In some examples, the machine learning model is trained using additional data from at least one additional sensor not included in the personal care device. In some examples, the at least one additional sensor includes a force sensor.

[0013] Some examples of methods, apparatus, non-transient computer-readable media, and systems may further include the step of receiving data via a wireless connection to a personal care device. In some examples, a machine learning model is trained using survey data collected from users, which includes usage quality information, device end-of-life information, or both.

[0014] In some examples, one or more features may include shaving features such as the number of shaving strokes, shaving quality information, razor end-of-life information, or any combination thereof. In some examples, feedback may include shaving feedback such as instructions to improve shaving quality. In some examples, shaving feedback may include razor replacement feedback.

[0015] Methods, apparatus, non-transient computer-readable media, and systems for machine learning for personal care devices are described. Embodiments of the methods, apparatus, non-transient computer-readable media, and systems can collect data from multiple sensors on a personal care device, identify one or more usage features based on the data, identify a subset of sensors that does not include at least one of the sensors, and train a machine learning model to predict one or more features using data from the subset of sensors.

[0016] The device may include a processor, memory that communicates electronically with the processor, and instructions stored in the memory. The instructions can enable the processor to operate to collect data from multiple sensors on a personal care device, identify one or more usage features based on the data, identify a subset of sensors that do not include at least one of the sensors, and train a machine learning model to predict one or more features using the data from the subset of sensors.

[0017] Non-temporary computer-readable media can store code for personal care. In some examples, the code includes instructions that can be executed by a processor to collect data from multiple sensors on a personal care device, identify one or more usage features based on the data, identify a subset of sensors that do not contain at least one of the sensors, and train a machine learning model to predict one or more features using the data from the subset of sensors.

[0018] In some examples, at least one of the sensors includes a force sensor. In some examples, one or more features may be shaving features including stroke count, stroke direction, or both.

[0019] Some examples of the methods, apparatus, non-temporary computer-readable media, and systems described above may further include using data from a subset of sensors to provide feedback to a trained machine learning model, at least partially.

[0020] Methods, apparatuses, and systems for personal care devices are described. Embodiments of methods, apparatuses, and systems may include a hub and a personal care device. The personal care device may provide data directly to the hub, and the hub may provide data to a server. The data may be raw data, and the server may reside in the cloud. The server may implement one of the machine learning methods described above to generate feedback based on the data. The server may also implement one of the machine learning methods to generate replacement orders for components of the personal care device. The feedback may be provided to a computer device linked to the hub and / or the personal care device.

[0021] In some examples, data can be provided wirelessly from a personal care device to a hub. The data can be generated by at least one of a plurality of sensors disposed within the personal care device. Feedback can be provided wirelessly to a hub and / or a computer device linked to the personal care device.

[0022] In some examples, the link between the hub and the server can be made secure. In an embodiment, raw data can be sent from the hub to a dedicated IP address of the server.

Brief Description of the Drawings

[0023] [Figure 1] FIG. is a diagram showing an example of a system for personal care and / or grooming according to an aspect of the present disclosure. [Figure 2] FIG. is a diagram showing an example of a razor according to an aspect of the present disclosure. [Figure 3] FIG. is a diagram showing an example of interactive razor hardware according to an aspect of the present disclosure. [Figure 4] FIG. is a diagram showing the architecture of an Internet of Things (IoT) razor system according to an exemplary embodiment of the concept of the present invention. [Figure 5] FIG. is a diagram showing a hub according to an exemplary embodiment of the concept of the present invention. [Figure 6] FIG. is a diagram showing an example of a feedback server for a personal care system according to an aspect of the present disclosure. [Figure 7] FIG. is a diagram showing an example of a feedback machine learning process for a personal care device according to an aspect of the present disclosure. [Figure 8] FIG. is a diagram showing an example of a sensor reduction server for a personal care system according to an aspect of the present disclosure. [Figure 9] FIG. is a diagram showing an example of a sensor reduction machine learning process for a personal care device according to an aspect of the present disclosure.

Modes for Carrying Out the Invention

[0024] The present disclosure generally relates to personal care, wellness, hygiene, and grooming, and more specifically to using machine learning methods to obtain information related to personal care devices and provide feedback.

[0025] Non-limiting embodiments of the present disclosure use machine learning techniques to identify razor data, predict shaving characteristics based on the razor data, and provide shaving feedback. In some cases, the razor data can include lifespan information for predicting the end of the lifespan of a cartridge. Thus, in an exemplary embodiment, the end of the lifespan of a cartridge can be predicted by using the razor data. The razor data can include any type of signal indicating that the end user should replace the cartridge. Data-driven cartridge replacement recommendations can result in an overall better shaving experience for the end user. Therefore, predicting the end of the lifespan of a cartridge can be achieved by obtaining data from the user and optionally from a specific end user. In some examples, sensor data from a razor or other personal care device can be collected and analyzed without using machine learning to predict feedback.

[0026] Embodiments of the present disclosure may utilize deep machine learning models. Such models may include interactive personal care or grooming appliances, tools, or devices such as interactive razors. The learning model may also be referred to as limited sensor post-processing software development. One embodiment may eliminate unnecessary sensors while retaining sufficient data to predict information useful to the end user, such as the end of life of replaceable components, such as the end of life of a razor cartridge, and additional feedback that may be of interest to the user may also be sourced through data generated by the device, tool, or appliance. Furthermore, simplifying device designs, such as razor designs, by removing at least some of the sensors that can be avoided by training machine learning algorithms with fewer sensors, can reduce the manufacturing cost of devices such as razors and enable competitive economic production.

[0027] An exemplary embodiment of the inventive concept described herein describes an Internet of Things (IoT) razor system capable of automatically providing a user with replacement blades and / or feedback based on user behavior. Briefly, the IoT razor system includes a razor equipped with multiple sensors that generate raw data when a user shaves their beard, and a short-range wireless communication device that outputs the raw data. The IoT razor system further includes a hub configured to receive the raw data from the razor and provide the raw data to a server via the cloud. The server then processes the raw data to predict, for example, the end of life of a razor cartridge and, based on the predicted shaving characteristics (e.g., predicted end of life), generate a replacement order that is automatically sent for a new razor cartridge.

[0028] Furthermore, notifications based on predicted shaving characteristics can be provided to the user. Notifications may include, but are not limited to, diagnostic feedback (e.g., instructions regarding alternative shaving procedures, current lifespan / usage level, etc.), recommendations for various consumer products (e.g., different types of cartridges based on predicted shaving characteristics, pre- or post-use skincare products such as shaving aid preparations based on predicted shaving characteristics, etc.), confirmations for replacement orders (e.g., order status, delivery status, approval or acceptance of the order, etc.), messages (e.g., visual, auditory, tactile, etc.), and the like. In one embodiment, notifications can be transmitted to the razor user via an IoT device located geographically in the same location as any or all components of the IoT razor system. For example, notifications may be with a smart mirror, smart faucet or smart shower head, user computer platform (e.g., smartphone, smartwatch, etc.), diodes in the razor and / or hub, and others.

[0029] Embodiments of this disclosure utilize one or more artificial neural networks (ANNs). In some cases, the ANN can be enhanced in its ability to handle a large number of input parameters or features. The ANN can learn from all input parameters and features, and thus the relationships can have higher dimensions. The ANN can be enhanced in its ability to implicitly detect complex nonlinear relationships between dependent and independent variables. The ANN can also be enhanced in its ability to detect all feasible interactions between input parameters and features.

[0030] Some embodiments of this disclosure may utilize variations of ANNs, including deep neural networks and shallow neural networks. In some cases, deep neural networks can implicitly learn from raw data features without relying on the generation of quadratic (e.g., derived) features. Deep neural networks can accommodate interactions that are difficult for the user to interpret but are still possible. Shallow neural networks, on the other hand, can include quadratic features predetermined by domain knowledge. Shallow neural networks can be more interpretable but, as a trade-off for being simpler models, offer less interaction.

[0031] As used herein, the terms “device,” “personal care device,” “appliance,” “instrument,” and “tool” may be used interchangeably. “Stroke count” may refer to the number of strokes in a single use, such as shaving strokes. “Direction” may refer to the changes and duration of the vertical and horizontal position of an appliance, tool, or implement, such as a razor, and the observed trend of the handle grip during use (such as during shaving). “Temperature” may refer to the temperature of an appliance, tool, or implement and its components during use, such as a razor. “Duration” may refer to the typical amount of time a consumer spends using an appliance, tool, or implement. “Shaving duration” may refer to the typical amount of time a consumer spends shaving. “Force” may refer to the amount of pressure applied to the face or body and the amount of drag across the surface while using an appliance, tool, or implement, such as during grooming or shaving. "Pivot measurement" can refer to the angle of the tool head (engaged to the face or body) relative to the handle (held by the user) during personal care or grooming, such as the razor head relative to the handle during shaving. "Rinse duration" can refer to the length of time spent under a stream of water from a sink, bath, or shower. "Grooming date and time" can refer to the date and time history of use of grooming tools, instruments, or equipment. "Shaving date and time" can refer to the shaving date and time history, i.e., a record of the days of the week and time of day when shaving usually occurred for the user. "Number of times" can refer to the number of times a tool, instrument, or device was used over a specific period, such as daily, weekly, or monthly. "Number of shaves" can refer to the number of shaves over a specific period, e.g., monthly or weekly. "Geographic location" can refer to the "geo-tagged" history of a consumer's personal care or grooming, such as shaving at home and while traveling.

[0032] As used herein, processing data may include date, duration, stroke count (obtained from force and / or IMU sensors), rinse count (obtained from head capacitance sensors), total rinse time, ambient temperature, minimum, maximum, and / or average force, stroke count in each direction (e.g., vertical upward, vertical downward, horizontal), and / or stroke force histograms. Processing data may further include stroke data (e.g., start index, end index, maximum force, gravity axis), analysis graphs, and / or filtered data.

[0033] As used herein, processed shaving data may include shaving date, shaving duration, stroke count (obtained from force and / or IMU sensors), rinse count (obtained from head capacitance sensors), total rinse time, ambient temperature, minimum, maximum and average force, stroke count and / or stroke force histograms in each direction (e.g., vertical upward, vertical downward, horizontal). Processed shaving data may further include stroke data (e.g., start index, end index, maximum force, gravity axis), analysis graphs and / or filtered data.

[0034] Figure 1 shows an example of a system for personal care and / or grooming according to an aspect of the present disclosure. The illustrated embodiment includes a user 100, a device 105, a communication network 110, storage 115, a server 120, and an artificial neural network (ANN) 125.

[0035] Users may use personal care devices and / or grooming devices (instruments, tools, or devices) such as device 105. In the illustrated example, device 105 is a razor configured for shaving or other grooming activities, but may include, for example, trimmers, dermaplaning devices, facial and body brushes and scrubbers, waterpiks, dental cleaning tools, toothbrushes, etc. Device 105 can communicate with server 120 via communication network 110. Communication network 110 may include communication networks such as mobile networks, wireless networks, and the Internet. Server 120 may include ANN 125. In some embodiments of this disclosure, ANN 125 can be trained using data from multiple sensors.

[0036] For example, device 105 could be a training device that provides data (d) from its (n) sensors (n is an integer) to train a machine learning algorithm to determine one or more shaving characteristics. However, device 105 could also be a production device that provides less data (e.g., d-1) from n sensors and / or fewer sensors (e.g., n-1) to enable a machine learning algorithm (e.g., trained) to specifically predict one or more shaving characteristics for production razors. Thus, for example, ANN125 could predict one or more shaving characteristics based on input data to indicate end-of-life information for device 105. In another example, ANN125 could predict feedback information related to brushing teeth, drying hair, or any other grooming process based on the type of device being used (e.g., a toothbrush), input data, and other factors.

[0037] Device 105 may include an example or embodiment of one or more corresponding elements described with reference to Figure 2.

[0038] The communication network 110 may include networks such as mobile networks, wireless networks, the Internet, or any combination thereof. According to some embodiments of this disclosure, information about the device 105 (e.g., sensor data) can be transmitted via the communication network 110 and collected by the ANN 125 of the server 120.

[0039] One embodiment of the present disclosure establishes a communication link between a hub and a ready-made personal care or grooming device, such as a razor, having an inertial measuring unit (IMU) (or a group of sensors sufficient to collect information necessary for determination) or a tool configured to generate raw usage data. The communication link may be a BT link. Other communication interfaces are also possible. The communication link may be encrypted (e.g., one-way, two-way, etc.).

[0040] The storage 115 may include an example or embodiment of the corresponding one or more elements described with reference to Figures 3, 7, and 9.

[0041] Server 120 may include an example or embodiment of the corresponding one or more elements described with reference to Figures 6 and 8. Server 120 may use machine learning techniques such as ANN125 to identify input data from a source (e.g., sensor data), predict one or more features such as shaving features based on the input data, and provide feedback and shaving feedback such as cartridge replacement and / or notification, end-of-life information for device 105, and others. One embodiment of the present disclosure is a feedback server 600, as shown by Figure 6. The feedback server 600 can train a machine learning model to provide feedback such as shaving feedback based on one or more features such as shaving features. In another embodiment, a sensor reduction server 800 can train a machine learning model to predict one or more shaving features using data from a subset of sensors.

[0042] ANN125 may include examples or embodiments of the corresponding one or more elements described with reference to the training components 630 and 835 in Figures 6 and 8. An artificial neural network (ANN) can be a hardware or software component comprising multiple connected nodes (also known as artificial neurons) that can roughly correspond to neurons in the human brain. Each connection or edge can transmit signals from one node to another (like physical synapses in the brain). When a node receives a signal, it can process the signal and transmit the processed signal to other connected nodes. In some cases, the signals between nodes may contain real numbers, and the output of each node may be calculated as a function of the sum of its inputs. Each node and edge may be associated with one or more node weights that determine how signals are processed and transmitted.

[0043] During the training process, these weights can be adjusted to improve the accuracy of the results (i.e., by minimizing a loss function that somehow corresponds to the difference between the current result and the target result). Edge weights can increase or decrease the intensity of the signal transmitted between nodes. In some cases, a node may have a threshold below which no signal is transmitted at all. Nodes can also be aggregated into layers. Different layers can perform different transformations on their inputs. The first layer may be known as the input layer, and the last layer as the output layer. In some cases, a signal may pass through a particular layer multiple times.

[0044] ANN125 can consider primary features that include all raw sensor data features. ANNs can also consider secondary (e.g., derived) features. Secondary features (e.g., derived) can include features that are more interpretable from the raw sensor data. ANN125 can incorporate deep learning models and shallow learning models. ANN125 can determine shaving length from IMU sensor data. For example, ANN125 can use acceleration and time data to calculate the usage length of personal care or grooming routines, such as shaving length.

[0045] In some cases, the feedback predicted by ANN125 may be based on shaving of a specific area of ​​the body (i.e., face, legs, or other body part). The length of shaving can be used as a predictor of the end of life of device 105. Machine learning models, methods, and systems can use input data and features that indicate the end of life, as well as sensors that generate the most valuable data.

[0046] Several embodiments may utilize deep learning architectures, including deep neural networks (DNNs) and recurrent neural networks (RNNs). In one embodiment of this disclosure, a force-predicting recurrent neural network (RNN) may be used. An RNN is a multilayer perceptron that learns how to process data over time. An RNN can be designed to learn how to manage memory for time-series data. Filtered sensor data can be processed as input data for an RNN. In some cases, an RNN can process fixed-length time-series sensor data. Fixed-length time-series sensor data may come from IMU data and capacitive sensor data, or from IMU data alone. An RNN can output perceived force for a fixed-length input. Note that IMU data can be the most important force indicator, as shown by the calculation of the mean absolute error (the mean difference between the predicted and target values). A first model using IMU data and capacitive sensor data and a second model using IMU data only can be improved over multiple training epochs (e.g., 50) (e.g., the error becomes smaller). Therefore, for example, in a sensor configuration, a force sensor may not be necessary.

[0047] According to one embodiment, the high-density neural network can be a multilayer perceptron that functions similarly to a logistic regression model. The DNN can be designed to learn meaningful combinations of input features. Quadratic features can be processed as input to the DNN.

[0048] One embodiment of this disclosure describes a network architecture for learning. The architecture can be a stacked long-short-term memory RNN. A long-short-term memory (LSTM) network is a type of recurrent neural network that can learn order dependencies in sequence prediction problems. An LSTM can be designed to learn how long to stay at each data point. Speed ​​and accuracy can be traded off by increasing the width (number of nodes) and depth (number of layers). The width and depth of the neural network are typically adjusted to the size and complexity of the dataset. For example, experiments are conducted on a dataset related to grooming appliances, tools, or equipment such as device 105. In this dataset, a diminishing return on accuracy may occur beyond 2 layers × 24 nodes. Some embodiments can use a 2 × 50 stacked LSTM-RNN.

[0049] Therefore, an LSTM-RNN can be approximated as a function of the input data, along with weights and biases such as Y=F(X,W,B). The input data can, for example, pass through an LSTM-RNN consisting of a neural network initialized with a specific width and depth. In one embodiment, the number of nodes and / or links connecting nodes between layers decreases as the network passes from input to output of the LSTM-RNN. For example, each node in the initial or input layer may have a link to each node in the immediately following layer, while a single node in the final or output layer may have only two input links from each of the two nodes in the preceding layer.

[0050] Furthermore, each point or channel of the input data can be weighted with the same or different weights and biased with the same or different biases. For example, if input data (X) passes through a model (F) and the model output (Y) is arbitrarily compared to expected output data (Ygt), the weights (W) and biases (B) at time 0 can be randomly initialized values ​​close to 0.0, selected by the Glorot Uniform initialization algorithm. Additionally or alternatively, IMU x acceleration data at time 1 can be weighted with weight (W1) and biased with bias (B), and IMU z gyro data at time 1 can be weighted with weight (W6) and biased with the same or different bias (B(n)), and so on, where each bias value is maintained independently and can be updated during training. Initial and updated parameters such as width, depth, links, weights and / or biases can be based on predicted usage features (for example, the weights and / or biases used may differ for training a force model compared to a distance model). Furthermore, data preprocessing (e.g., filtering) and processing (e.g., training) may occur on a single data point (or channel), but the same steps can be achieved on a window or set of data that can be consecutive or overlapping. For example, IMU x-accelerometer data at time 1 can be used, or five points of IMU x-accelerometer data from time 1-5 can be used. The reference data will correspond to the same time window as the input data.

[0051] During training, the model's output data is compared to reference data, and the percentage error and / or confidence level can be determined from the comparison between the model output data (Y) and the reference data (Ygt). If the percentage error or confidence level is unacceptable (e.g., based on a threshold), the weights and biases can be adjusted, and the input data can be passed through the updated LSTM-RNN again to determine whether an acceptable percentage error and / or confidence level has been reached. The process can be repeated until an acceptable error and / or confidence level is achieved with the trained machine learning model that produces the minimum or acceptable error. Thus, in one example, the variables Y, X, W, and B can be represented as tensors in computer memory or storage that can refer to an N-dimensional set of numerical data (e.g., normalized and scaled between [0,1] or [-1,1]). On the other hand, the error (e.g., loss) can refer to the averaged mean squared difference between the model output data (Y) and the reference data (Ygt). Thus, a learning algorithm (e.g., Adam optimizer) can take the calculated loss as input and determine the direction and magnitude to which each weight and bias of the training model is to be adjusted.

[0052] Before selecting a model architecture that yields the minimum or acceptable error, the trained model can also be evaluated to verify its reproducibility. For example, the entire set of training data can be split so that a relatively small portion of the training data is used for evaluation and not for training. The relatively small portion of the training data is passed through a candidate trained LSTM-RNN and compared to reference data to evaluate the error or confidence of a specific usage feature (e.g., the mean error of cumulative force across all shavings, the mean error from the total distance traveled across all shavings, etc.). The trained machine learning model can then be deployed to predict and / or enhance the functionality of usage features using production input data (e.g., virtually in real time while the device is in use). In this regard, usage features can be predicted from a subset of sensors (e.g., using only IMU data), and personal care devices such as device 105 can exclude, disable, and / or make unnecessary conventional sensors for identifying, deriving, or predicting usage features.

[0053] In one example of training a model for force prediction, the input training data passing through the LSTM-RNN consists only of time-series IMU data from a personal care device, such as device 105 (e.g., acceleration (x,y,z) and gyroscope (x,y,z)), while the reference data consists of sensor data, such as time-series force data, including force applied to the head from the personal care device's force sensor, but not limited to this case. If an acceptable percentage error and / or confidence level for force is achieved using only the IMU data, then the personal care device, such as device 105, can exclude, disable, and / or make unnecessary force sensors to provide usage features (e.g., cumulative force). The predicted usage features of force can then be used to feed production input data through the deployed, trained machine learning model to predict other usage features and / or enhance the user experience, including providing recommendations, control functions, and replenishment.

[0054] In another example of training a model for predicting distance traveled, the input training data passing through the LSTM-RNN consists only of time-series IMU data (e.g., acceleration (x,y,z) and gyroscope (x,y,z)) from a personal care device such as device 105, while the reference data consists of corresponding time-series data obtained using a 6-degrees-of-freedom (6DoF) tracking solution coupled to the personal care device (e.g., via a serial connection). The data from the tracking solution can be converted into measured values ​​of the actual distance traveled. In an example of 146 seconds of time-series data, the true distance was 4178.22 cm, the predicted distance was 4252.33 cm, and the error was 1.77%. Notably, the trained machine learning model can provide an average error of 2.79% for predicting the total distance traveled by the personal care device. Then, by passing production input data through the deployed trained machine learning model using the predicted use features of distance traveled, other use features can be predicted and / or recommended, enhancing the user experience, including providing control functions, replenishment, etc.

[0055] In yet another example of training a predictive model for tasks performed by a personal care device, the input training data passed through the LSTM-RNN consists only of time-series IMU data from a personal care device, such as device 105 (e.g., acceleration (x,y,z) and gyroscope (x,y,z)), and the reference data consists of a combination of sensor data, such as time-series force data, such as force applied to the head from the personal care device's force sensor, and corresponding time-series data obtained using a 6DoF tracking solution coupled to the personal care device (e.g., via a serial connection). However, the tasks performed by the personal care device can be determined using an arithmetic logic unit (ALU) or other processor configured to independently multiply the predicted force by the predicted distance traveled. Then, by passing the production input data through the deployed trained machine learning model using the predicted use features of the work, other use features can be predicted and / or recommended, and the user experience can be enhanced, including the provision of control functions, replenishment, etc.

[0056] Notably, data preparation can include clock synchronization between sensor data and reference data using corresponding signals in two data streams for each channel of data (e.g., IMU z-acceleration data and corresponding reference z-position data from a tracking device). For example, the signal of the IMU data can be matched with the signal of the reference data to eliminate the time offset and match the two datasets. As an example, the same time offset used to synchronize one channel can be applied to all data channels in the same dataset. In this way, the clock can be synchronized between datasets (e.g., for each of 100 shavings) by eliminating the offset between datasets. However, each sensor platform can utilize its own internal time clock synchronized to the operating system of the same external computer platform. This type of synchronization can render a relatively high but acceptable error rate of travel distance predictions from a model configured to use IMU data as input for travel distance predictions (e.g., during shaving, etc.) by passing production input data through a deployed and trained machine learning model. Additionally or alternatively, inactive periods longer than a predetermined duration (e.g., 5 seconds, etc.) in the reference data can be removed, and the remaining active periods can be split into subsets. Similar to clock synchronization, removal and splitting can be applied to all data channels from the same dataset.

[0057] Therefore, algorithms that predict any usage feature can generally be substantially similar to differences based on the specific usage feature being predicted. For example, a general algorithm that predicts any usage feature may comprise requirements definition, data collection, data preparation, model development (e.g., model selection, model training, model optimization), model evaluation, and / or model deployment. On the other hand, differences in general algorithms may include differences in data collection (e.g., different subsets of data used for training, such as 6DoF motion data not required to train a force measurement model), differences in data preparation (e.g., force can be calculated over a single time step (10 ms), while distance can be calculated over each overlapping window of 128 time steps (1280 ms)), differences in models (e.g., the values ​​of selected hyperparameters (model architecture) and learning parameters (weights and biases) can be unique to each model) and similar.

[0058] Here, the components of a system such as a razor system are described, followed by exemplary applications of the system. The razor system is shown for illustrative purposes only, and it is understood that one or more embodiments of the razor system can be implemented by any system (e.g., a toothbrush system), apparatus (e.g., a personal care device, a hub, a server, an IoT device, etc.), and / or method (e.g., feature prediction, etc.) according to the embodiments.

[0059] Figure 2 shows an example of device 105 according to an aspect of this disclosure. Device 105 may include an example or embodiment of the corresponding one or more elements described with reference to Figure 1.

[0060] Device 105 may include a cartridge 205, a cartridge retaining assembly 210, a rinse sensor lug 215, a connecting head 220, a finger guard 225, a razor body 230, a head hinge 235, a force sensor 240, a silicone pad 245, a logo insert 250, a battery 255, a microcontroller unit (MCU) 260, a charging and / or data transfer port such as Bluetooth® or WiFi 265, a USB or USB-C port 270, a power supply 275, and an end cap 280. The force sensor 240, the battery 255, and the MCU 260 may include examples or embodiments of the corresponding one or more elements described with reference to Figure 3.

[0061] In one embodiment, an interactive razor, such as device 105 in Figure 1, may have multiple sensors, some of which can be removed. For example, the force sensor 240 can be removed, and device 105 can still be used to accurately predict the number of strokes and classify the stroke direction (e.g., horizontal, vertical upward, vertical downward, etc.). In one embodiment, the trained model architecture can be selected for force prediction such that the input production data (e.g., while the user is shaving) passing through an LSTM-RNN to accurately predict the force consists only of time-series IMU data (e.g., acceleration (x,y,z) and gyroscope (x,y,z)). As shown in Figures 2 and 3, in one embodiment, the capacitive sensor 325 can be removed, and device 105 can be left removed to accurately classify the rinse. In another embodiment, the shaving length can be determined from IMU sensor 320 data. For example, the IMU sensor 320 can acquire and / or provide acceleration and time data to calculate the shaving length. In one example, the trained model architecture can be selected for predicting distance traveled such that the input production data (e.g., while the user is shaving) passing through the LSTM-RNN to accurately predict the distance traveled consists only of time-series IMU data (e.g., acceleration (x,y,z) and gyroscope (x,y,z)). In some embodiments, shaving length can be used as a predictor of the end of life. In yet another example, the trained model architecture can be selected for predicting work so that the input production data (e.g., while the user is shaving) passing through the LSTM-RNN to accurately predict the work consists only of time-series IMU data (e.g., acceleration (x,y,z) and gyroscope (x,y,z)). However, the work performed can be determined using an ALU or other processor configured to independently multiply the predicted force by the predicted distance traveled.

[0062] Accordingly, Figure 2 shows the components of a razor according to an exemplary embodiment of the concept of the present invention. The razor in Figure 2 is an exemplary configuration. In other words, the razor may include more or fewer components than those shown in Figure 2.

[0063] Figure 3 shows an example of interactive razor hardware 310 according to an aspect of the present disclosure. The illustrated example includes a programmer 300, interactive razor analysis software 305, and interactive razor hardware 310. The razor is shown for illustrative purposes only, and it is understood that one or more aspects of the razor hardware can be implemented by any system (e.g., a toothbrush system), apparatus (e.g., a personal care device, a hub, a server, an IoT device, etc.), and / or method (e.g., feature prediction, etc.) according to the embodiment.

[0064] The interactive razor hardware 310 may include a sensor input 315, a controller 350, a power manager 375, an external flash 392, and an RGB LED 394. The sensor input 315 may include an inertial measuring unit (IMU) 320, a capacitive sensor 325, a force sensor 330, a rinse sensor 335, a grip sensor 340, and an NTC thermistor 345. The controller 350 may include an MCU 355, an RTC 360, a BLE radio 365, and an internal flash 370. The MCU 355 may include an example or embodiment of one or more corresponding elements described with reference to Figure 2.

[0065] The power manager 375 may include a USB 380, a battery manager 382, ​​a battery 384, a source selector 386, a regulator 388, and a system power bus 390. The battery 384 may include one or more corresponding elements or embodiments thereof as described with reference to Figure 2.

[0066] As shown in Figures 2 and 3, according to some embodiments, the force sensor 240 can be removed, and the device 105 in Figure 1 can still be used, for example, to predict the number of strokes and classify the stroke direction (e.g., horizontal, vertically upward, vertically downward, etc.). The capacitance sensor 325 can be removed, and the device 105 in Figure 1 can still be used, for example, to classify the rinse. In another embodiment, the shaving length can be determined from data from the IMU sensor 320. For example, by acquiring and / or providing acceleration and time data, the IMU sensor 320 can calculate the shaving length. In some embodiments, the shaving length can be used as a predictor of the end of life.

[0067] Accordingly, Figure 3 illustrates the hardware of a razor according to an exemplary embodiment of the concept of the present invention. As shown in Figure 3, the razor hardware includes, among other things, a sensor input 315, a controller 350 which can be a 2.4GHz wireless microcontroller (e.g., CC2640R2) that supports BLE, and a power manager 375.

[0068] As illustrated, the sensor input 315 may include, but is not limited to, an inertial measuring unit (IMU) 320, a capacitance-to-digital converter, a force sensor 330, a rinse sensor 335, a grip sensor 340, and a temperature sensor (e.g., an NTC thermistor 345). The IMU 320 may include, for example, an accelerometer, a gyroscope, and / or a magnetometer. An IMU such as the IMU 320 may include at least one accelerometer per axis in three axes and / or at least one gyroscope per axis in three axes. For example, the IMU 320 may include six micro-electromechanical system (MEMS) modules, where one MEMS corresponds to one accelerometer or gyroscope for each axis (e.g., x, y, z axes, vertical axis, orthogonal axis, etc.).

[0069] Preferably, the IMU includes a magnetometer that lacks a magnetometer and is disabled, and / or a magnetometer whose data is ignored or otherwise filtered so that it does not reach an analog-to-digital converter, hardware registers, memory, storage, processor, trained or trained machine learning model, or a combination thereof. In one example, the filter may include a bandpass filter. Furthermore, the IMU may include a processor physically coupled to an accelerometer, gyroscope, and / or magnetometer. For example, IMU320 can be a module having a processor physically coupled to an accelerometer and gyroscope (e.g., via a data connection, via an analog-to-digital converter, etc.). Additional sensors for use with the razor according to exemplary embodiments of the concept of the present invention may include resistive contacts, piezo contact microphones, Hall effect sensors, external cameras, strain gauges, load cells, reed switches, linear variable differential transformers, etc.

[0070] The wireless microcontroller may include a main central processing unit (e.g., MCU355), a BLE radio 365, built-in flash memory 370, and other peripherals and modules (e.g., a real-time clock (RTC360)). For example, the CC2640R2 SimpleLink® BT microcontroller includes, among other things, a microcontroller (e.g., having a central processing unit, non-volatile memory, random access memory (RAM), and cache memory), an ultra-low power sensor controller (e.g., a sensor controller engine, analog-to-digital converter, RAM, etc.), and a radio frequency (RF) section (e.g., a transceiver, RF interface, etc.). In one example, the MCU355 can be physically coupled to the accelerometer and gyroscope of the IMU320 via a connection between the controller 350 and the IMU320. The wireless microcontroller can be connected to an external flash device 392 and a light-emitting element (e.g., a status RGB LED 394). The wireless microcontroller can also be connected to a programmer 300 and analysis software 305.

[0071] When shaving using the razor described above, raw sensor data can be collected. This raw data can be stored in the non-volatile memory on the razor. In the embodiment shown in Figure 3, the raw data can be downloaded to a desktop application via a serial connection. However, as described above, it is also possible to send the raw data directly to the hub. The raw data communicated between the razor and the hub can be encrypted. The raw data captured by the sensors may include, but are not limited to, time in seconds (e.g., UTC-0), IMU data (e.g., acceleration (x,y,z) and gyroscope (x,y,z)), force applied to the head, capacitance of the metal insert in the head, case temperature, gravity axis (e.g., rough orientation of the razor), and force sensing register (FSR) calibration values.

[0072] When the aforementioned razors are used for shaving during production, or more generally when personal care devices are used to perform their functions outside of the training process, the processor (e.g., one or more cores, a central processing unit, a graphics processing unit, or a combination thereof) can be configured to run a machine learning model trained to predict usage characteristics. For example, the processors in the IMU320 and / or MCU355 can be configured to run a trained LSTM-RNN that is selected to accurately predict force such that the input production data passing through the LSTM-RNN (e.g., while the user is shaving) consists only of time-series IMU data (e.g., acceleration (x,y,z) and gyroscope (x,y,z)). Additionally or alternatively, the processors of the IMU320 and / or MCU355 may be configured to run a trained LSTM-RNN selected to predict distance traveled, such that the input production data passing through the LSTM-RNN (e.g., while a user is shaving) consists only of time-series IMU data (e.g., acceleration(x,y,z) and gyroscope(x,y,z)) for accurately predicting distance traveled. Additionally or alternatively, the processors of the IMU320 and / or MCU355 may be configured to run a trained LSTM-RNN selected to predict work performed by a razor, such that the input production data passing through the LSTM-RNN (e.g., while a user is shaving) consists only of time-series IMU data (e.g., acceleration(x,y,z) and gyroscope(x,y,z)) for accurately predicting work. The processors of the IMU320 and / or MCU355 may, however, be configured to calculate work based on independently predicted force and distance traveled (e.g., work = force × distance).

[0073] Similarly, the processors of the IMU320 and / or MCU355 use predicted usage characteristics (e.g., predicted force, predicted distance traveled, etc.) to predict, but not limited to, stroke force (e.g., force generated in individual strokes, force generated in total strokes during shaving, etc.), stroke distance (e.g., stroke distance (e.g., distance traveled in individual strokes, distance traveled in total strokes during shaving, etc.), number of strokes (e.g., number of strokes during shaving, total number of strokes during operation, etc.), rinse events (e.g., individual rinses of the device, total number of strokes, etc., number of rinse events during shaving, total number of rise events during operation, etc.), end of life (e.g., prediction of when one or more components remain operational, prediction of when replacement is recommended based on thresholds, etc.), or combinations thereof, for personal care The IMU320 and / or MCU355 processors can be configured to perform transformations that predict other usage features associated with the device. These transformations can be achieved using functions (e.g., databases, matrices, pointers, mathematical functions, etc.) that associate the predicted usage features with other predicted usage features, such as stroke count or lifetime value. The IMU320 and / or MCU355 processors can, but are not limited, run additional machine learning models trained to predict other usage features based on one or more predicted usage features, such as when a model is run that is trained to predict stroke count or lifetime based on predictive power (from a trained model), predicted distance traveled (from a trained model), or a combination thereof. The IMU320 and / or MCU355 processors can, however, be configured to predict one or more other usage features (e.g., stroke count, lifetime, etc.) based on input production data (e.g., while a user shaves) that has passed through a trained LSTM-RNN consisting only of time-series IMU data (e.g., acceleration (x,y,z) and gyroscope (x,y,z)), and run additional machine learning models trained to accurately predict the other usage features.

[0074] The predicted usage characteristics can then be used to provide feedback such as control functions on the personal care device (e.g., fluid supply, pulses), usage recommendations (e.g., user guidance), control functions on IoT devices adjacent to the personal care device (e.g., faucets, lighting from fixtures), replenishment functions (e.g., automatic ordering), replacement information (e.g., cartridge usage level), or a combination thereof. For example, the predicted usage function may be used to communicate recommendations to the user regarding the use of the personal care device, such as changing the water flow rate or temperature from a faucet via an IoT faucet (e.g., via the personal care device, via an IoT device such as a smart mirror, via a mobile device), to provide stroke recommendations, to automatically replenish disposable components of the personal care device (e.g., order replacement cartridges), to notify the user of the predicted lifespan of disposable components, and to enable the user to replace the disposable components.

[0075] Accordingly, for example, the processors of the IMU320 and / or MCU355 can be configured to initiate control instructions for controlling elements of a personal care device. In one example, a control instruction may include an interrupt for modulating the operation of an element, an enable signal for enabling the operation of an element, a disable signal for disabling the operation of an element, or a combination thereof. For example, a razor device may include a razor blade and a cartridge optionally containing a shaving aid composition. Furthermore, the razor device may include a dispenser containing a personal care composition (e.g., shaving liquid) that can be selectively released to enhance the shaving experience. In this regard, the processors of the IMU320 and / or MCU355 can also be configured to control the dispenser by initiating control commands (e.g., generating commands, transferring commands, calling commands) to a controller coupled to the cartridge, and by initiating control commands to a controller coupled to the dispenser in order to selectively apply the personal care composition during shaving based on how one or more components of the cartridge come into contact with the user (e.g., force, distance traveled, strokes, number of strokes) based on expected usage characteristics (e.g., blade contact).

[0076] A processor (e.g., one or more cores, a central processing unit, a graphics processing unit, etc., or any combination thereof) coupled to one or more sensors (e.g., an accelerometer, a gyroscope, etc.) may optionally be coupled to an RF interface for transmitting raw data, predictions and / or feedback, memory or storage for storing raw data, predictions and / or feedback, input / output devices (e.g., a display, a speaker, a haptic device, a switch, etc.), or a sensor control device that controls prediction-based feedback via a combination thereof. In this regard, conventional communication connections include, but are not limited to, connections on a circuit board, serial connections between hardware modules, local area network connections, wide area network connections, or combinations thereof. Feedback signals include, but are not limited to, interrupts, enable signals, disable signals, machine code instructions (e.g., assembly code), input / output device instructions, or combinations thereof.

[0077] Furthermore, while processors coupled to accelerometers and / or gyroscopes (or IMUs) have been described as being physically coupled to personal care devices via communication connections, processors configured to perform one or more embodiments of the embodiments (e.g., machine learning models trained to predict usage characteristics related to personal care devices, data collection, model training, feedback, etc., or a combination thereof) may be located in one or more other devices such as hubs, servers, IoT devices, mobile computer platforms, or a combination thereof. Thus, for example, a processor can be communicatively coupled to a set of sensors on a personal care device (e.g., accelerometers, gyroscopes, etc.) via a connection between the hub and the personal care device. In another example, a processor can also be located on a server communicatively coupled to a hub in close proximity to the personal care device and / or the personal care device.

[0078] Figure 4 shows an architecture of an IoT razor system according to an exemplary embodiment of the concept of the present invention. The architecture shown in Figure 4 is provided as an example, but embodiments of the concept of the present invention are not limited thereto.

[0079] As shown in Figure 4, the IoT razor system includes a consumer component 410, a hub 420, and a cloud component 430. The consumer component 410 includes a smart razor and a mobile device. It is understood that the consumer component 410 may include additional devices with wireless communication technology, such as a smart mirror, smart faucet, or showerhead. The smart razor and mobile device (e.g., a smartphone) can communicate wirelessly with each other via Bluetooth® or Bluetooth® Low Energy (BLE). It will be understood that the smart razor and mobile device may employ a variety of other short-range wireless communication technologies, including but not limited to ZigBee®, Near Field Communication (NFC), Li-Fi, Ultra Wideband (UWB), and Radio Frequency Identification (RFID). The smart razor can correspond to one of the razors described above, with reference to Figures 1-3.

[0080] The smart razor wirelessly provides raw data to the hub 420, for example, via Bluetooth® or BLE. As mentioned above, the raw data can be provided via a number of other short-range wireless communication technologies. The raw data transmitted from the razor to the hub 420 is data collected by sensors within the razor. Sensors and raw data will be described in more detail later. The hub 420 stores the raw data and transmits the raw data to the cloud component 430. For example, the hub 420 can transmit the raw data to the cloud component 430 via a dedicated and secure IP protocol. More specifically, the raw data can be transmitted via a long-range connection (e.g., cellular connection, fiber optic connection, copper connection, etc.) to a dedicated IP address of a server in the cloud (e.g., a cloud computing network).

[0081] More specifically, in one embodiment, raw data is provided directly from the smart razor to the hub 420, which then provides the raw data directly to a server in the cloud. In this embodiment, a mobile device is not used as an intermediary for communicating raw data between the hub 420 and the server. Nor is a mobile device used as an intermediary for communicating raw data between the smart razor and the hub 420. As will be described later, in such embodiments, the mobile device is used only to receive information from the server. The mobile device does not receive raw razor data.

[0082] In Cloud 430, a Platform-as-a-Service (PaaS) provides an application gateway, such as the Cloud Gateway / IoT Hub in Figure 4, which receives raw data and preprocesses it before inputting it into a neural network model, such as the ANN125 in Figure 1. The model used can correspond to one or more of the machine learning models described herein. In stream preprocessing, force and capacitance sensor data can be passed through a Butterworth filter to flatten the frequency response. Furthermore, the input to the neural network can be normalized to speed up the neural network. Additionally, filtered force sensor values ​​can be scaled down to mitigate explosive gradients during training.

[0083] The model uses raw data (which can be pre-processed data) to predict shaving features. In one embodiment, a neural network model uses pre-processed data to predict the end of the effective lifespan of a razor cartridge. For example, taking filtered time-series sensor data as input, the neural network model generates an output indicating the probability that the razor will reach the end of its lifespan. An example of a neural network model will be described in detail later.

[0084] The output of the neural network model is stored in a database, for example, the data store in Figure 4, which may include a Hi-Child database, a relational database, an object-oriented database, etc. In one embodiment, the data store includes an Azure SQL database. The data output from the model can be used by the illustrated Application Services (App Services) to generate, deploy, and / or use an application to provide replacement orders for new razor cartridges. The model's output can also be used by the Application Services to generate, deploy, and / or use an application to provide other feedback to users. Although not illustrated, other service platforms or computer devices that access the data store and / or application data (App Data) can provide services (e.g., replacement services) and / or information (e.g., feedback) based on predicted features. Furthermore, business intelligence and / or application tools (e.g., Power Apps / BI) within a business analytics and / or subscription service platform (e.g., Microsoft 365) can utilize the data output from the model to build and / or use applications, and use business intelligence and / or data mining tools to gain insights from the data, make decisions, generate reports, etc.

[0085] The application service can, for example, facilitate communication with the consumer component 410 and provide informant to the user. Users of personal care or grooming devices, instruments, or tools such as razors can, for example, be notified via a software application on a mobile device that a new cartridge is in the way. For example, the display can provide a graphical user interface (GUI) in which the user can view notifications such as one or more recommendations, replacement order confirmations and / or acceptances, order status, end-of-use messages (e.g., predicted end date, usage level, etc.), or other usage information about the device or user. Any other consumer component and / or hub 420, however, can communicate with any service platform to receive feedback via a GUI, diodes, and / or haptic devices configured to provide information related to replenishment, feedback, etc.

[0086] Therefore, although this model has been described as being used to generate replacement orders for razor cartridges, it is understood that this model can be used for a variety of other applications. For example, the neural network model can utilize the data provided therein to identify usage habits (such as shaving habits) that can lead to customized recommendations. More specifically, the neural network can be used to generate product recommendations, adjustments to product usage, estimated remaining display counts for quality usage (such as shaving), etc. Thus, the services and / or feedback provided may relate to, for example, brushing teeth, drying hair, and / or any other personal grooming process, based on the type of device used (e.g., toothbrush), training, sensor configuration, input data, usage parameters / functions / metadata, etc., or a combination thereof. These additional features and other embodiments of Figure 4 will be described in more detail later. Furthermore, the illustrated PaaS is one non-limiting embodiment of cloud infrastructure. Therefore, one or more other infrastructures, such as Infrastructure as a Service (IaaS) or Software as a Service (SaaS), can be implemented additionally or alternatively. Similarly, one or more aspects of the Cloud 430 can be implemented additionally or alternatively in non-cloud computing networks, such as on-premises or corporate computer networks.

[0087] Figure 5 shows a hub 510 according to an exemplary embodiment of the concept of the present invention. As shown in Figure 5, the hub 510 includes a short-range wireless communication module 512, a Wi-Fi module 514, a microcontroller 516, a memory 518, and a power supply 520.

[0088] The short-range wireless communication module 512 may employ Bluetooth® or BLE. It is understood that various other short-range wireless communication technologies, including but not limited to ZigBee®, NFC, UWB, and RFID, may also be employed by the hub 510. The short-range wireless communication module 512 enables the hub 510 to communicate with a smart razor 502, a mobile device 504, or other nearby devices such as a smart mirror, a smart faucet or shower head, lighting, a thermostat, or other wireless communication devices 506. According to exemplary embodiments of the concept of the present invention, communication between a smart device (such as a razor 502) and the hub 510 can be encrypted. For example, a smart device (such as a razor 502) and the hub 510 can be paired using Secure Simple Pairing (SSP) to establish a BLE link, and once paired, encrypted raw data can be transmitted over the BLE link.

[0089] Data sent from a device (such as a razor 502) to the hub 510 can be stored in the hub's memory 518. The hub 510 can receive data from the device (such as a razor 502) in real time while the user is shaving. Data transfer may occur in response to requests from the hub 510. For example, the microcontroller 516 of the hub 510 can be configured to periodically request the device (such as a razor 502) to send the raw data stored in its onboard memory. Periodic data requests may be based on time of day (TOD), for example.

[0090] The Wi-Fi module 514 enables the hub 510 to communicate with, for example, a wireless home network 530. It will be understood that the hub 510 may include any other communication interface that enables communication with the network. For example, the hub 510 may include a module that enables power line communication (PLC).

[0091] Hub 510 can transmit raw data to a server in the cloud 534 via the wireless router 532. Data can be transmitted to the cloud 534 in real time as soon as it is received from a device (such as a razor 502). Hub 510 can also queue data and, after a predetermined period or after a predetermined amount of data has been collected, transfer it to the cloud 534 at a specific time. These transfer times can be set by the user via an application on a mobile device. Data transfer may also occur in response to data requests from the cloud server. These data requests can be transmitted after a predetermined period or after a predetermined amount of data has been collected, at a specific time.

[0092] Communication between Hub 510 and the cloud server is secure. For example, all raw data provided from Hub 510 to the server can be encrypted. Furthermore, all communication from the server to Hub 510 can be encrypted. More specifically, raw data can be transmitted through a secure channel between Hub 510 and the cloud server. The secure channel can employ any protocol such as Advanced Message Queuing Protocol (AMQP), Message Queuing Telemetry Transport (MQTT), or Hypertext Transfer Protocol Secure (HTTPS). SASL (Simple Authentication and Security Layer) can be used for authentication and data security between Hub 510 and the cloud server.

[0093] Hub 510 may include its own power source, such as a rechargeable or non-rechargeable battery. Hub 510 may also include a USB connector for power, and the hub may utilize inductive charging. The USB connector may also function as a data store. Hub 510 may also include a plug that connects to a power receptacle. As shown in Figure 5, Hub 510 is a standalone device similar to an Ethernet bridge. Hub 510 can be plugged into a power receptacle, for example, so that a user can take Hub 510 on a trip. In one embodiment, Hub 510 may have the form factor of a substantially cubic wall adapter with built-in power prongs. Hub 510 can be configured as a receptacle device for instruments such as shaving equipment. For example, Hub 510 may be a dock, stand, or cradle. In this case, Hub 510 can be used for charging a device (such as a razor 502), cleaning or sterilizing a device (such as a razor 502), and evaluating device failure and / or degradation.

[0094] Figure 6 shows an example of a feedback server 600 for a personal care system according to an aspect of this disclosure.

[0095] Server 600 can be an example of the corresponding one or more elements described with reference to Figures 1 and 8. Server 600 may include a processor component 605, a memory component 610, a data acquisition component 615, a feedback prediction component 620, a feedback component 625, and a training component 630.

[0096] The processor component 605 may be an example of, or an embodiment thereof, of one or more corresponding elements described with reference to Figure 8. The processor may include intelligent hardware devices (e.g., general-purpose processing components, digital signal processors (DSPs), central processing units (CPUs), graphics processing units (GPUs), microcontrollers, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), programmable logic devices, discrete gate or transistor logic components, discrete hardware components, or any combination thereof). In some cases, the processor may be configured to operate a memory array using a memory controller. In other cases, the memory controller may be integrated into the processor. The processor may be configured to perform various functions by executing computer-readable instructions stored in memory. In some examples, the processor may include application-specific components for modem processing, baseband processing, digital signal processing, or transmission processing. In some examples, the processor may consist of a system-on-a-chip.

[0097] The memory component 610 may be an example of, or an embodiment thereof, of one or more corresponding elements described with reference to Figure 8. Computer memory can store information for various programs and applications on a computer device. For example, storage may include data for running an operating system. Memory may include both volatile and non-volatile memory. Volatile memory may include random access memory (RAM), and non-volatile memory may include read-only memory (ROM), flash memory, electrically erasable programmable read-only memory (EEPROM), digital tape, hard disk drives (HDDs), and solid-state drives (SSDs). Memory may include any combination of readable and / or writable volatile memory and / or non-volatile memory, along with other possible storage devices.

[0098] The data acquisition component 615 can identify data from a razor containing a set of sensors (for example, an example of device 105 in Figure 1). The data acquisition component 615 may be an example of, or include embodiments thereof, of, one or more corresponding elements described with reference to Figure 8. Thus, in some examples, the set of sensors may include an IMU 320, a capacitance-to-digital converter, a force sensor 240, a rinse sensor 335, a grip sensor 340, a temperature sensor, an accelerometer, a gyroscope, a magnetometer, a resistive contact, a piezo contact microphone, a Hall effect sensor, an external camera, a strain gauge, a load cell, a reed switch, a linear variable differential converter, or any combination thereof. The data acquisition component 615 may also receive data via a wireless communication connection to the razor.

[0099] In one embodiment, the data collection component 615 collects data from a set of sensors on a device (e.g., a razor) such as device 105 in Figure 1. In some embodiments, the data collection component 615 may identify and collect other data through means such as user participation, surveys, and individual subjective opinions. In one example, a participant regularly shaves with the same cartridge using a razor (e.g., an example of device 105 in Figure 1) until they indicate that the cartridge no longer provides an effective shaving experience. After each shave, the participant may answer survey questions regarding their subjective opinions on their shaving experience. Thus, the data collected by the data collection component 615 may include sensor data from device 105 illustrated in Figure 1 (e.g., 60Hz, 120Hz, etc.) and / or hedonic survey data (sometimes also called SIMS data) regarding the shaving experience. Furthermore, the longevity shaving test may include conducting two separate trials (e.g., A and B). Trial A may be limited to 30 days, while Trial B may not be time-limited. Both trials include 26 unique participants, and the same participants may be present in each trial.

[0100] In one embodiment, the data collection component 615 collects data from a set of sensors on a personal care device other than a razor (e.g., device 105 in Figure 1). For example, a participant may brush their teeth or perform any other personal grooming activity until the device no longer provides an effective experience. After each session, the participant may answer survey questions regarding their subjective opinions on the experience.

[0101] In one example, ground truth data can be collected from a razor (or other shaving device) having multiple sensors (e.g., accelerometer, gyroscope, magnetometer, image capture device, Hall effect sensor, piezoelectric sensor, etc.) such as device 105 in Figure 1. When a user, such as user 100 in Figure 1, uses the shaving device, ground truth data can be captured by the sensors on the razor. Ground truth data can be compiled from sensor data. Thus, ground truth data can include information that is not noticeable or detectable to the human eye (e.g., slight movements, slight variations in various movements, multiple combinations of various movements, various movements captured simultaneously in real time, etc.). In another embodiment, ground truth data can include a dataset containing data from accelerometers, gyroscopes, Hall effect sensors, etc.

[0102] In embodiments of this disclosure, the data acquisition component 615 may include data cleanup, integrity checks (e.g., garbage in / garbage out), feature engineering, and extraction. The input data may be a plurality of subsets of filtered time-series sensor data. In another embodiment, the learning model may be developed by creating a neural network, where the input data are derived sensor data features (e.g., cumulative force, distance traveled, time under force), and the output data are the probability of the razor life ending.

[0103] In some embodiments, the processing data may include the shaving date, shaving time, number of strokes (from force and IMU sensors), number of rinses (from head volume sensor), total rinse time, ambient temperature, minimum, maximum, and average force, number of strokes in each direction (e.g., vertically upward, vertically downward, horizontal), and a stroke force histogram. Additional processing data may include stroke data (e.g., start index, end index, maximum force, gravity axis), analysis graphs, and filtered data.

[0104] In some embodiments, the processing data may include information about the individual's grooming activities other than shaving, such as the date, time, number of strokes (from force and IMU sensors), number of rinses (from head capacitance sensor), total rinse time, ambient temperature, minimum, maximum, and average force, number of strokes in each direction (e.g., vertically upward, vertically downward, horizontal), and force histograms for strokes other than shaving strokes (e.g., toothbrush strokes).

[0105] Referring to feature engineering, potential secondary (derived) features may include determining the start and end of stroke or rinse events, shaving characteristics, average force per stroke, cumulative force applied for razor life, cumulative duration of use for razor life, wall time between shaves, survey response delta (i.e., differences between respondents for normalization of survey data), shaving length or distance traveled, and the amount of work for razor life.

[0106] In one embodiment of this disclosure, determining the start and end of a stroke or rinse event allows for the use of data limited to this period for better metric purposes. In some cases, usage characteristics such as shaving characteristics may include the number of strokes, stroke direction (e.g., horizontal, vertical up, vertical down), and number of rinses. In embodiments, the respondent survey response delta may indicate the device condition, such as the condition of the razor, from an analysis of the end-user experience. For example, if the device usage quality (e.g., shaving) decreases by 1 point for two consecutive uses, the next use (e.g., shaving) will be the last use. Furthermore, if the shaving quality suddenly decreases by 3 points, the razor is dead. In one embodiment, the amount of work relative to the lifespan of the razor can be calculated as work = force × distance.

[0107] In some embodiments, the statistical correlation results may include quadratic features extracted from the described dataset. For example, the extracted quadratic features may include the mean value across all sensors per use, such as per shave; the deviation from typical sensor data for panelists; cumulative time, such as cumulative shaving time; and cumulative force.

[0108] The feedback prediction component 620 can predict one or more shaving features, at least partially, based on input data from a razor (e.g., an example of device 105 in Figure 1), using a machine learning model trained with data from at least a set of sensors from the razor (e.g., an example of device 105 in Figure 1). For example, the input data can be provided by a razor configured to provide sensor data from a subset of available sensors, a razor configured to provide a subset of sensor data from available sensor data, and / or only the IMU of the razor. The data used to train the machine learning model can include, for example, at least one additional sensor data that is not input data (e.g., data from more sensors than the subset of available sensors, a larger subset of sensor data, and / or from an IMU that provides more data than the trained IMU, etc.). The IMU of a commercially available razor can also provide all the raw sensor data for predicting shaving features. In some embodiments, the server 600 can predict shaving characteristics including the number of shaving strokes, shaving quality information, razor end-of-life information, or any combination thereof.

[0109] The feedback prediction component 620 can also predict one or more features related to grooming activities other than shaving, at least in part, based on input data from the device, using a machine learning model trained with data from at least multiple sensors. For example, the input data may be data from a device configured to provide sensor data from a subset of available sensors, a device configured to provide a subset of sensor data from available sensor data, and / or data from the IMU only. The data used to train the machine learning model may include, for example, at least one additional sensor data in addition to the input data. In some embodiments, the server 600 can predict features including stroke count, quality information, end-of-life information, or any combination thereof.

[0110] The feedback component 625 can provide usage feedback, such as shaving feedback, based on one or more features of the device or use, such as shaving features. In some examples, the feedback, such as shaving feedback, includes instructions for improving quality, such as shaving quality. In some examples, the feedback, such as shaving feedback, includes replacement feedback for at least one limited-life component of the device, such as razor (or razor cartridge) replacement feedback.

[0111] The feedback component 625 can provide feedback, such as shaving feedback, based on a trained machine learning model using data from a subset of sensors of a device (e.g., a razor). In one embodiment of the present disclosure, the feedback component 625 may include output data indicating a probability of the end of life of a device, such as the end of life of a device 105 used by user 100 in Figure 1 for periodic shaving.

[0112] The training component 630 may be an example of, or include embodiments thereof, of one or more corresponding elements described with reference to Figure 8. In some examples, the machine learning model is trained using survey data collected from the user, which includes shaving quality information, razor effective life information, or both. In an embodiment, the training component 630 may include a machine learning model such as ANN125, as illustrated in Figure 1. ANN125 can predict the effective life information of the device. In another embodiment, the learning model may be developed through the creation of ANN125, where the input data is derived sensor data features (e.g., cumulative force, distance traveled, time under force) and the output data is the likelihood of the razor reaching the end of its life.

[0113] In some examples, a machine learning model is trained using survey data collected from users, which includes quality information for grooming activities other than shaving, end-of-life information, or both. ANN125 can predict the end-of-life information of the device. In another embodiment, a learning model can be developed by creating ANN125, where the input data are derived sensor data features (e.g., cumulative force, distance traveled, time under force) and the output data is the likelihood of end-of-life.

[0114] In another embodiment, the learning model can be developed by creating ANN125, as shown in Figure 1, where the input data are derived sensor data features (e.g., cumulative force, distance traveled, time under force) and the output data are the likelihood of reaching the end of life (e.g., the end of razor life).

[0115] In one embodiment, ground truth data can be input to a machine learning (ML) algorithm to generate shaving implementation metadata. The ML algorithm can determine which parts of the ground truth data provide a specific confidence interval (CI) to accurately determine the metadata of the implementation usage parameters (e.g., shaving features). This can be an iterative process until a predetermined CI is reached. Thus, in each iteration, the ML algorithm can determine whether the CI for the implementation usage parameters is met. If the CI is met, the ML algorithm can proceed; otherwise, the ML algorithm can repeat the input step. In one embodiment, data from multiple sensors can be used to accurately determine the stroke. In one embodiment, data from multiple sensors is used to accurately determine the stroke. As an example, output data is compared with predicted data from the ML algorithm to verify whether the dataset (set of sensors) can accurately determine the stroke. If the dataset cannot accurately determine the stroke, the ML algorithm can repeat until a CI greater than n% (e.g., greater than 90%) is reached. The ML algorithm can use more data. The output data of the ML algorithm that satisfies the CI is used as the shaving implementation metadata. In some cases, machine learning models can be trained using additional data from at least one additional sensor not included in the razor.

[0116] In one embodiment, ground truth data can be input to an ML algorithm to generate tool metadata for devices other than razors. The ML algorithm can determine which parts of the ground truth data provide a confirmation threshold (CI) for accurately determining the metadata of tool usage parameters. The ML algorithm can then determine which parts of the data provide a CI for accurately determining the metadata of tool usage parameters. Therefore, in each iteration, the ML algorithm can determine whether the CI has been met for the implementation usage parameters. If the result satisfies the CI, the ML algorithm proceeds; otherwise, the ML algorithm can repeat the input step.

[0117] Figure 7 shows an example of a process for feedback machine learning for a personal care device according to aspects of this disclosure. In some examples, these operations can be performed by a system including a processor that executes a set of code to control the functional elements of the device. Additionally or alternatively, the process can be performed using application-specific hardware. In general, these operations can be performed according to the methods and processes described in accordance with aspects of this disclosure. For example, the operation can consist of various substeps or can be performed in combination with other operations described herein.

[0118] In operation 700, the server 600 can identify data from a device that includes a set of sensors on the device, such as device 105 in Figure 1 (e.g., a razor). In some cases, the operation in this step may be performed with reference to or by data acquisition components 615 or 815, as described with reference to Figures 6 and 8.

[0119] According to embodiments of this disclosure, the data acquisition component 615 may include data preparation, cleanup and integrity checks (e.g., garbage in / garbage out), feature engineering and extraction. A machine learning model, such as ANN125 in Figure 1, may include creating a deep neural network model for predicting force sensor output. Input data may include IMU sensor data and capacitive sensor data, and output data may include force sensor data. The raw data recorded may include time in seconds (UTC-0), IMU data (e.g., acceleration (X, Y, Z), gyroscope (X, Y, Z)), force applied to the head, capacitance of the metal insert in the head, case temperature, gravity axis (rough orientation of the device, such as a razor), and FSR calibration values. The raw data may include data from a 6-degree-of-freedom (6DoF) tracking device coupled to the personal care device, which can simultaneously acquire data from one or more sensors on the personal care device (e.g., tracking data from a user group) and data from the personal care device.

[0120] In some embodiments, the sensor set may include an IMU 320, a capacitance-to-digital converter, a force sensor 240, a rinse sensor 335, a grip sensor 340, a temperature sensor, an accelerometer, a gyroscope, a magnetometer, a resistive contact, a piezo contact microphone, a Hall effect sensor, an external camera, a strain gauge, a load cell, a reed switch, a linear variable differential converter, or any combination thereof.

[0121] In some embodiments, the processing data may also include the shaving date, shaving duration, number of strokes (from force and IMU sensors), number of rinses (from head volume sensor), total rinse time, ambient temperature, minimum, maximum, and average force, number of strokes in each direction (e.g., vertically upward, vertically downward, horizontal), and a force histogram of the strokes. Additional processing data may include stroke data (e.g., start index, end index, maximum force, gravity axis), analysis graphs, and filtering data.

[0122] In some embodiments, the processing data may also include information on grooming activities other than shaving, such as the session date, time, number of strokes (from force and IMU sensors), number of rinses (from head capacitance sensor), total rinse time, ambient temperature, minimum, maximum, and average force, number of strokes in each direction (e.g., vertically upward, vertically downward, horizontal), and force histograms for strokes. Additional processing data may include stroke data (e.g., start index, end index, maximum force, gravity axis), analysis graphs, and filtering data.

[0123] In embodiments, force and capacitance sensor data can be passed through signal processing filters, such as Butterworth filters, to flatten the frequency response. Input data to a neural network (e.g., ANN125) can be normalized around a mean value of zero. Doing so may allow the neural network to progress faster. For example, filtered force sensor values ​​can be scaled down by a factor of 100 to mitigate explosive gradients during training.

[0124] In operation 705, a server such as server 600 in Figure 6 can predict one or more features (e.g., shaving features) based on the data using a machine learning model trained with data from a set of sensors. In some cases, the operation in this step can be performed by referring to or by a feedback prediction component as described with reference to Figure 6. In some embodiments, one or more features may include stroke count, quality information, end-of-life information, or any combination thereof.

[0125] In operation 710, a server such as server 600 in Figure 1 may provide feedback based on one or more features. In some cases, the operation in this step may refer to, or be performed by, a feedback component 625 as described with reference to Figure 6.

[0126] In embodiments, a server, such as server 600 in Figure 1, can provide feedback based on one or more features. In some examples, the feedback may include instructions to improve the quality of the grooming session. The feedback may also include exchange feedback. The feedback may be based on a machine learning model trained using data from a subset of sensors. In one embodiment of this disclosure, the shaving feedback may include output data indicating the probability that device 105 is reaching the end of its lifespan.

[0127] In some cases, a server like server 600 in Figure 1 can provide feedback other than shaving. For example, a safety razor may have a power supply, a (unidirectional) short-range communication device (e.g., Bluetooth®), and an IMU (e.g., a multi-axis accelerometer or gyroscope). Raw data about the razor's movement (e.g., velocity, direction, acceleration) can be transmitted to a hub. The hub can transfer this raw data to a processor in the network described. The processor has an algorithm (e.g., applying machine learning or deep learning) to process the data and determine the type of motion. Actions may include a normal stroke (e.g., a shaving stroke), an accidental drop, the user placing the device (e.g., the razor) in storage between uses, or the user positioning the device (e.g., the razor) under a faucet or showerhead for rinsing. In the last example, a return signal is sent from the network to the hub, and the smart faucet or showerhead is activated while the device (e.g., the razor) is recognized as being in the "rinse" position. Real-time data and flow rate data from water taps and showerheads can be transmitted to an application on the user's smart device (e.g., phone, tablet, etc.) to educate the user about water consumption and suggest ways to reduce it. In some embodiments, additional capabilities of an Internet of Things (IoT) communication network can be used.

[0128] While independent blocks and / or specific sequences are shown for illustrative purposes, it will be understood that one or more blocks of the method in Figure 7 or components of the apparatus in Figure 6 may be combined, omitted, bypassed, rearranged, and / or flowed in any order, and may be placed on any device or a single device. For example, one or more components shown in Figure 6 and / or one or more blocks shown in Figure 7 may be additionally or alternatively present in or implemented in a personal care device, hub, server, IoT device, or a combination thereof.

[0129] Figure 8 shows an example of a sensor reduction server 800 for a personal care system according to an aspect of the present disclosure. The server 800 may be an example of, or include, one or more corresponding elements described with reference to Figure 1 and 6.

[0130] The server 800 may include a processor component 805, a memory component 810, a data acquisition component 815, a feature selection component 820, a sensor selection component 825, a feature prediction component 830, and a training component 835.

[0131] The server 800 may include an interactive device (e.g., a razor) deep learning model that predicts characteristics (e.g., shaving characteristics) while limiting the number of sensors on the device (e.g., device 105 in Figure 1). According to embodiments of the present disclosure, the server 800 can predict the end of life of a device or its components (e.g., a cartridge) even if some sensors are removed. For example, a razor (e.g., an example of device 105 in Figure 1) may have multiple sensors, some of which can be removed.

[0132] In one embodiment, even if a force sensor such as the force sensor 240 in Figure 2 is removed, the number of strokes can be predicted and the stroke direction (e.g., horizontal, vertical up, vertical down, etc.) can be classified. In another embodiment, a capacitive sensor such as the capacitive sensor 325 in Figure 3 can be removed, and the rinse can still be classified. In another embodiment, the duration or length of use, such as the duration or length of shaving, can be determined from IMU sensor data. For example, a user or a learning model can use acceleration and time data to calculate a length, such as the length of shaving. In some embodiments, a length, such as the length of shaving, can be used as a predictor of the end of life.

[0133] Processor component 805 may be an example of, or include embodiments of, one or more corresponding elements described with reference to Figure 6. Processor component 805 may be an example of, or include embodiments of, one or more corresponding elements described with reference to Figure 8. The processor may include intelligent hardware devices (e.g., general-purpose processing components, digital signal processors (DSPs), central processing units (CPUs), graphics processing units (GPUs), microcontrollers, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), programmable logic devices, discrete gate or transistor logic components, discrete hardware components, or any combination thereof). In some cases, the processor may be configured to operate a memory array using a memory controller. In other cases, the memory controller may be integrated into the processor. The processor may be configured to perform various functions by executing computer-readable instructions stored in memory. In some examples, the processor may include application-specific components for modem processing, baseband processing, digital signal processing, or transmission processing. In some examples, the processor may consist of a system-on-a-chip.

[0134] Memory component 810 may be an example of, or an embodiment of, one or more corresponding elements described with reference to Figure 6. Memory component 810 may be an example of, or an embodiment of, one or more corresponding elements described with reference to Figure 8. Computer memory can store information for various programs and applications on a computer device. For example, storage may contain data for running an operating system. Memory may include both volatile and non-volatile memory. Volatile memory may include random access memory (RAM), and non-volatile memory may include read-only memory (ROM), flash memory, electrically erasable programmable read-only memory (EEPROM), digital tape, hard disk drives (HDDs), and solid-state drives (SSDs). Memory may include any combination of readable and / or writable volatile memory and / or non-volatile memory, along with other possible storage devices.

[0135] The data collection component 815 may be an example of, or include, one or more corresponding elements as described with reference to Figure 6.

[0136] Some embodiments of this disclosure can enable one or more sensors configured based on implementation configuration data to generate raw usage data. This process can depend on the implementation configuration. For example, if the configuration involves manufacturing an implement such as a shaving implement with only the minimum number of sensors necessary to accurately determine shaving implement metadata for implement usage parameters, a user such as user 100 in Figure 1 would power on a device that powers the physically installed application sensors (which may be a minimum number). If the configuration involves selectively enabling or disabling the total number of sensors on a shaving device for the minimum number of sensors necessary to accurately determine the metadata of the shaving device, a user such as user 100 in Figure 1 would selectively power on or power off the minimum number of sensors. If the configuration involves selectively programming one or more sensors to capture only a subset of the data they can capture (for example, only a subset of gyro data such as only a subset of accelerometer data or only a subset of axis data), then a user such as user 100 in Figure 1 would selectively program it to capture only a subset of the data it can capture (i.e., capture only the data (which may be a subset) it can conventionally capture). However, in the above example, the device may additionally or alternatively be configured to automatically enable, disable, and / or program itself based on the instruction set executed by its processor and / or the instruction set stored in its memory.

[0137] The implementation metadata generated from the ML algorithm can be determined and used at the backend of the process, rather than at the frontend. For example, a commercially available razor with an IMU like the IMU320 in Figure 3, or a sensor suite sufficient to collect the information necessary for determination, can be used to collect as much raw data as possible. The raw data can be sent to a hub and then transmitted to a communication network such as the communication network 110 in Figure 1. The communication network can use the output (e.g., a signature of a subset of data that can accurately determine when a stroke occurs) to evaluate the raw sensor data, determine whether the shaving device is being used (e.g., strokes) and how it is being used, and make recommendations (e.g., refills, product changes, etc.) and / or automatically send products.

[0138] The feature selection component 820 can identify one or more features, such as shaving features, based on the data. In some examples, the one or more features, such as shaving features, include the number of strokes, the stroke direction, or both.

[0139] The sensor selection component 825 can identify a subset of sensors, where at least one sensor is not included in the subset. In some examples, at least one sensor includes a force sensor.

[0140] Embodiments of this disclosure may include shaving implementation metadata for configuring a shaving implementation. Configuration may occur in software, hardware, or a combination of both. In some cases, configuring an instrument, such as a shaving instrument, may include physically installing one or more application-specific sensors. For example, this could involve reconfiguring sensors using microcode, using configurable hardware such as a field-programmable gate array (FPGA), or selectively powering on or off components using voltage rails. In some examples, configuring an instrument, such as a shaving instrument, may include manufacturing a shaving instrument with only the minimum number of sensors necessary to accurately determine shaving instrument metadata for instrument usage parameters. This configuration may include selectively enabling or disabling the total number of sensors in an instrument, such as a shaving instrument, to the minimum number of sensors necessary to accurately determine instrument metadata, such as shaving instrument metadata. The configuration may also include selectively programming one or more sensors to capture only a subset of the data they can capture (e.g., only a subset of gyro data, such as only a subset of accelerometer data or only a subset of axial data). The advantages of the above configuration may include a simplified hardware stack, lower power consumption, and lower processing load. In one embodiment of this disclosure, the IMU can be physically installed in a shaving device. The IMU can be trained by a machine learning model (e.g., a deep neural network) and provide information that can be used to accurately determine strokes.

[0141] The feature prediction component 830 can use data from a subset of sensors from a device such as a razor (for example, the device 105 in Figure 1) to train a machine learning model to predict one or more features, such as shaving features. In some examples, the machine learning model can predict one or more features, such as shaving features including the number of strokes, stroke direction, or both.

[0142] The training component 835 may be an example of, or include embodiments thereof, of the corresponding one or more elements described with reference to Figure 6. In some embodiments, the machine learning model is trained using additional data from at least one additional sensor not included in the razor. In some examples, the at least one additional sensor includes a force sensor. In further examples, the last additional sensor is located on a 6-degree-of-freedom (6DoF) tracking device coupled to the personal care device, which can simultaneously capture data from one or more sensors on the personal care device (e.g., tracking data from a group of users) and data from the personal care device.

[0143] In one embodiment, ground truth data is input to an ML algorithm to determine a subset of data (and therefore a subset of sensors) that can accurately determine when a stroke occurs. In one example, the output data is compared with the predicted data from the ML algorithm to verify whether the subset of data (subset of sensors) can accurately determine a stroke. In such a case, the ML algorithm can be iterated until it reaches a CI of n% or more (e.g., 90% or more). The ML algorithm can then use more data. The output data of the ML algorithm that satisfies the CI is used as shaving implementation metadata.

[0144] The experiment is conducted on the sensor reduction server 800. In some examples, participants shave using a razor like device 105 in Figure 1, while an observer records the shaving characteristics. The generated data can include sensor data from a razor like device 105 in Figure 1 (e.g., 100Hz) and ground truth data of shaving characteristics (e.g., number of strokes, stroke direction, number of rinses). The 100Hz ground truth dataset may have a shave dataset count equal to 98. The shaving dataset can include 97 unique participants or other statistically significant populations. In another example, the 100Hz ground truth dataset may have a shave dataset count equal to 30. The shaving dataset can include 15 unique participants. Also, if an RNN (e.g., LSTM-RNN) is trained on 100Hz sensor data, the results from the force prediction neural network improve significantly as the sensor data collection frequency increases, providing accurate feature predictions. In one example, the predicted force curve follows the pattern of the true force curve, with a substantial peak alignment.

[0145] One embodiment describes a method for storing raw usage data on a server. In this case, the raw usage data may depend on the configuration of the implementation, or on one or more sensors configured based on the configuration data of the implementation. For example, if the configuration may involve manufacturing a shaving tool that has only the minimum number of sensors necessary to accurately determine shaving tool metadata for tool usage parameters, the data may be all raw data from an IMU provided to collect specific data (e.g., all accelerometer data only, all gyroscope data only, etc.). For example, a user may power on the device to supply power to physically installed application sensors (a minimum number may be required).

[0146] In some embodiments, if the configuration includes selectively enabling or disabling the total number of sensors in the shaving tool to the minimum number of sensors necessary to accurately determine the metadata of the shaving tool, the user will selectively power on or off the sensors from which data is collected. If the configuration includes selectively programming one or more sensors to collect only a subset of the data they can capture (e.g., only a subset of accelerometer data, only a subset of gyro data such as a subset of axis data), then data can be collected from the selectively programmed sensors. Alternatively, a commercially available razor may, by its capability, provide all the raw data.

[0147] Figure 9 shows an example of a process for sensor reduction machine learning for a personal care device according to aspects of this disclosure. In some examples, these operations can be performed by a system including a processor that executes a set of code to control the functional elements of the device. Additionally or alternatively, the operations can be performed using application-specific hardware. In general, these operations can be performed according to the methods and processes described in accordance with aspects of this disclosure. For example, the operations can consist of various substeps or can be performed in combination with other operations described herein.

[0148] In operation 900, the system collects data from a set of sensors on a device such as a razor (for example, the example device in Figure 1). In some cases, the operation of this step may be performed by referring to or by a data acquisition component as described with reference to Figures 6 and 8.

[0149] In operation 905, the system identifies one or more features, such as shaving features, based on the data. In some cases, the operation in this step may refer to or be performed by a feature selection component, such as those described with reference to Figure 8.

[0150] In operation 910, the system identifies a subset of sensors, and at least one of the sensors is not included in the subset. In some cases, the operation of this step may refer to, or be performed by, a sensor selection component as described with reference to Figure 8.

[0151] In some examples, the configuration of a shaving device may include manufacturing a shaving device that has only the minimum number of sensors necessary to accurately determine the shaving device metadata for the device usage parameters. The configuration may also include selectively enabling or disabling the total number of sensors in the shaving device down to the minimum number of sensors necessary to accurately determine the shaving device metadata. The configuration may also include selectively programming one or more sensors to capture only a subset of the data they can capture (e.g., only a subset of accelerometer data, a subset of gyroscope data, a subset of axis data, etc.). In one embodiment of this disclosure, the IMU may be physically installed in the shaving device. IMUs can be trained by machine learning models (e.g., deep neural networks) and provide information that can be used to accurately determine strokes.

[0152] In some examples, the configuration of a personal care device other than a razor may include manufacturing a device with the minimum number of sensors necessary to accurately determine the implement metadata for implement usage parameters. The configuration may also include selectively enabling or disabling the total number of sensors on the implementation to the minimum number of sensors necessary to accurately determine the relevant implementation metadata. The configuration may also include selectively programming one or more sensors to capture only a subset of the data they can capture (e.g., only a subset of accelerometer data, only a subset of gyroscope data, or a subset of axis data).

[0153] In one embodiment, ground truth data can be input to an ML algorithm to determine a subset of data (and therefore a subset of sensors) that can be accurately determined when a stroke occurs. In one example, the output data is compared to predicted data from the ML algorithm to verify whether the subset of data (subset of sensors) can accurately determine the stroke. In such a case, the ML algorithm can be iterated until it reaches a CI of n% or more (e.g., 90% or more). The ML algorithm can then use more data. The output data of an ML algorithm that satisfies the CI is impression data, such as shaving impression metadata.

[0154] In operation 915, the system trains a machine learning model to predict one or more features using data from a subset of sensors. In some cases, this step can be performed by referring to, or by, a feature prediction component as described with reference to Figure 8.

[0155] One embodiment of this disclosure describes inputting raw usage data into an algorithm to generate implementation usage data. There may be a step of comparing the raw usage data with reference data (e.g., shaving implementation metadata). For example, the raw data can be collected from a commercially available razor (having a sufficient number of sensors for specific usage parameters) or a configured shaving instrument. The raw usage data is input to a comparator with metadata output (signatures of a subset of ground truth data can be precisely determined when a stroke occurs) so that it can be determined whether or how (e.g., strokes) the shaving instrument is being used. The instrument usage data can be transferred to a counter (e.g., a stroke counter) to track usage.

[0156] Embodiments of this disclosure include extracting existing data from banks, compiling live data from users, live data from groups of users, time-series tracking data from users, tracking data from groups of users, or any combination thereof. By combining the data (i.e., data entered by users and data collected from users), end-of-life predictions can be developed.

[0157] One embodiment of this disclosure describes a method for using data. The same or different cloud servers described herein may transmit data to another network and may transmit data to memory (e.g., storage, cache, etc.). In one embodiment, the process of using data may include accessing user information or general reference information to determine whether it is appropriate to automatically ship replacement goods to the user and to provide recommendations. In one embodiment, an SQL server may be used to process information from a user. This may include receiving user preferences, payment information, calendar ratings, confirmations, etc., from user profiles from an enterprise server, or directly from the user via a personal computer device (e.g., via an application), or a combination thereof.

[0158] In one embodiment, raw usage data can be entered into a described system or network to accurately determine the % end of life of a device or its components, such as a razor cartridge. For example, actual usage data (e.g., showing stroke count) is used together with user data collected from user profiles, real-time statistics, or user input via an application (e.g., ordering patterns, travel patterns, cartridge preferences, actual strokes / cartridge, preference for strokes / cartridge, etc.). Additionally or alternatively, device usage data can be compared to general reference data (e.g., typical strokes per device (e.g., cartridge), specific age groups, specific geographical regions, etc.). Strokes can be used, for example, to determine the remaining life for a percentage life calculation (e.g., 90% of the remaining life), which is then used for recommendations (other products, other types of cartridges, companion products, etc.) or refills.

[0159] While independent blocks and / or specific sequences are shown for illustrative purposes, it is understood that any block of the method in Figure 9 or one or more components of the apparatus in Figure 8 can be combined, omitted, bypassed, rearranged, and / or flow in any order, or placed on any device or a single device. For example, one or more components shown in Figure 8 and / or one or more blocks shown in Figure 9 can additionally or alternatively reside or be implemented in a personal care device, hub, server, IoT device, or a combination thereof.

[0160] Additional Notes and Examples

[0161] Additional embodiments include systems, apparatus, non-temporary computer-readable media, and methods for training machine learning models, running machine learning models, acquiring and / or processing data for machine learning models, providing feedback based on the output of trained machine learning models, or a combination thereof.

[0162] Embodiment 1 may include a personal care device comprising a plurality of sensors and a processor coupled to the plurality of sensors, wherein the processor may include a system configured to run a machine learning model trained to predict usage characteristics associated with the personal care device based on data from the plurality of sensors.

[0163] Example 2 may include the system of Example 1, and the multiple sensors may include accelerometers and gyroscopes placed on the inertial measuring unit (IMU) of the personal care device.

[0164] Example 3 may include any system from Examples 1 to 2, wherein the IMU includes at least one accelerometer for each of the three axes and at least one gyroscope for each of the three axes.

[0165] Example 4 may include the system described in any one of Examples 1 to 3, wherein each of at least one accelerometer and at least one gyroscope includes its respective micro-electromechanical system (MEMS).

[0166] Example 5 may include the system described in any of Examples 1 to 4, wherein the IMU lacks a magnetometer.

[0167] Example 6 may include the system described in any of Examples 1 to 5, further comprising a filter to prevent data from the magnetometer from reaching one or more of the machine learning models being trained or the trained machine learning models.

[0168] Embodiment 7 may include the system described in any of Embodiments 1 to 6, wherein the processor is located on the IMU and is physically coupled to the accelerometer and gyroscope via a connector on the IMU.

[0169] Embodiment 8 may include any system from Embodiments 1 to 7, wherein the processor is located on the controller of the personal care device and is physically coupled to multiple sensors via a connection between the controller and the IMU. Embodiment 9 may include any system from Embodiments 1 to 8, wherein the processor is located on a hub adjacent to the personal care device and is communicably coupled to multiple sensors via a connection between the hub and the personal care device.

[0170] Embodiment 10 may include any system of Embodiments 1 to 9, wherein the processor is located on a server which is communicably coupled to one or more personal care devices or hubs adjacent to personal care devices, and the processor is communicably coupled to a plurality of sensors via a connection between the server and one or more personal care devices or hubs.

[0171] Example 11 may include the system described in any of Examples 1 to 10, wherein the personal care device provides data from multiple sensors to a hub via a short-range wireless connection unit, and the hub provides the data to a server located on a cloud computing network via a long-range connection unit.

[0172] Example 12 may include the system described in any of Examples 1 to 11, wherein the machine learning model is a trained stacked long-short-term memory recurrent neural network (LSTM-RNN).

[0173] Example 13 may include a system according to any one of Examples 1 to 12, wherein the usage characteristics include the force applied by the personal care device, the distance traveled by the personal care device, the work performed by the personal care device, or a combination thereof.

[0174] Example 14 is a first machine learning model in which the processor predicts a first usage feature related to a personal care device based only on accelerometer data and gyroscope data input to the machine learning model being trained, wherein the first usage feature includes force, A second machine learning model that predicts a second usage feature related to a personal care device based solely on accelerometer and gyroscope data input to the machine learning model being trained, wherein the second usage feature includes distance traveled. Alternatively, the system may include the system described in any one of Examples 1 to 13, which is configured to perform one or more of the following: a third machine learning model that predicts a third use feature related to a personal care device based only on accelerometer data and gyroscope data input to a machine learning model being trained, wherein the third use feature includes tasks performed by the personal care device; and a third machine learning model that includes tasks performed by the personal care device.

[0175] Example 15 describes a processor configured to predict, based on usage characteristics, one or more stroke counts, stroke force, stroke distance, rinse events, or end-of-life events associated with a personal care device. This may include the systems described in any of Examples 1 to 14.

[0176] Example 16 may include any system from Examples 1 to 15, wherein the processor is configured to perform one or more of the following: utilize a function that predicts one or more of the number of strokes, stroke force, stroke distance, rinse event, or end of life based on usage characteristics; or run an additional machine learning model that has been trained to determine one or more of the number of strokes, stroke force, stroke distance, rinse event, or end of life based on usage characteristics.

[0177] Example 17 may include the system described in any of Examples 1 to 16, wherein the personal care device comprises a razor device.

[0178] Example 18 may include a system according to any one of Examples 1 to 17, wherein the razor device is coupled in close proximity to and communicatively to an Internet of Things (IoT) device. Example 19 may include a system according to any one of Examples 1 to 18, further including a radio frequency interface, memory, storage, a sensor controller, an input / output device, or a combination thereof.

[0179] Example 20 may include the system described in any of Examples 1 to 19, wherein the processor is configured to perform one or more of the following: train a machine learning model to predict usage characteristics associated with a personal care device based on data from multiple sensors; collect data from multiple sensors; or provide feedback based on usage characteristics.

[0180] Example 21 may include at least one computer-readable storage medium containing a set of instructions that, when executed by a processor, cause the processor to implement any of the systems of Examples 1 to 20. For example, at least one computer-readable storage medium contains a set of instructions that, when executed, cause a processor to run a machine learning model trained to predict usage characteristics associated with a personal care device based on data from multiple sensors.

[0181] Example 22 may include a method for performing the operation of the system described in any of Examples 1 to 20. For example, this method may include capturing data through multiple sensors of a personal care device and running a machine learning model trained to predict usage characteristics associated with the personal care device based on the data.

[0182] Example 23 may include a method that includes means for performing the operation of the system described in any of Examples 1 to 20. For example, the method may include means for acquiring data through multiple sensors of a personal care device and means for running a machine learning model that has been trained to predict usage characteristics associated with the personal care device based on the data.

[0183] Embodiment 24 includes a feature prediction device comprising a feature prediction component configured to train a machine learning model to predict usage features associated with a personal care device using data from a subset of a set of sensors, wherein the feature prediction component is configured to initialize a machine learning model with an initial set of model parameters, input data from a subset of sensors into the machine learning model, run the machine learning model to generate output data corresponding to usage features, compare the output data with reference data to determine whether the machine learning model predicts usage features, iteratively update at least one model parameter, iteratively input data from a subset of sensors, and iteratively compare the output data with reference data until the machine learning model predicts usage features associated with the personal care device.

[0184] Example 25 may include the feature prediction device of Example 24, in which the initial model parameters include model width, model length, model link connections between model nodes, weights applied to the input data, biases applied to the input data, or a combination thereof.

[0185] Example 26 is a feature prediction device according to any one of Examples 24 to 25, wherein a subset of the sensor set includes an accelerometer and a gyroscope placed on the personal care device, and the feature prediction component is configured to initialize a machine learning model with an initial set of model parameters and to input data from the accelerometer and gyroscope into the machine learning model. The feature prediction device according to any one of Examples 24 to 25 may include running the machine learning model to generate output data corresponding to usage features, comparing the output data with reference data to determine whether the machine learning model predicts the usage features, iteratively updating at least one model parameter, iteratively inputting data from the accelerometer and gyroscope, and iteratively comparing the output data with reference data until the machine learning model predicts usage features related to the personal care device.

[0186] Example 27 may include a feature prediction device according to any one of Examples 24 to 26, wherein the accelerometer and gyroscope are located on the inertial measurement unit (IMU) of the personal care device.

[0187] Example 28 may include any of the feature prediction devices of Examples 24 to 27, wherein the feature prediction device is located on a server that is communicably coupled to one or more personal care devices or hubs adjacent to personal care devices, and the feature prediction components are communicably coupled to a subset of a set of sensors via a connection between the server and one or more personal care devices or hubs.

[0188] Example 29 may include a feature prediction device as described in any of Examples 24 to 28, wherein the personal care device provides data from multiple sensors to a hub via a short-range wireless connection unit, and the hub provides the data to a server located on a cloud computing network via a long-range connection unit.

[0189] Example 30 may include a feature prediction device according to any one of Examples 24 to 29, wherein the feature prediction component is located in one or more of the personal care device or a hub adjacent to the personal care device.

[0190] Example 31 may include a feature prediction device as described in any of Examples 24 to 30, wherein the machine learning model is a stacked long-short-term memory recurrent neural network (LSTM-RNN).

[0191] Example 32 may include any of the feature prediction devices from Examples 24 to 31, wherein the usage features include the force applied by the personal care device, the distance traveled by the personal care device, the work performed by the personal care device, or a combination thereof.

[0192] In Example 33, the feature prediction component is: Initialize and run a first machine learning model that predicts a first usage feature related to a personal care device based solely on accelerometer and gyroscope data input to the machine learning model being trained, wherein the first usage feature includes force, and the reference data includes data from a force sensor on the personal care device, data from a capacitive sensor on the personal care device, or a combination thereof. A second machine learning model is initialized and run to predict a second usage feature of a personal care device based solely on accelerometer and gyroscope data input to the machine learning model being trained, wherein the second usage feature includes distance traveled, and the reference data includes data from a 6-degree-of-freedom (6DoF) tracking device coupled to the personal care device. Alternatively, the feature prediction device may include the one or more of the following: initializing and running a third machine learning model to predict a third use feature related to a personal care device based solely on accelerometer and gyroscope data input to the machine learning model being trained, wherein the third use feature includes work performed by the personal care device, and the reference data includes data from a force sensor on the personal care device, data from a capacitive sensor on the personal care device, data from a 6DoF tracking device, or a combination thereof.

[0193] Example 34 may include any feature prediction device from Examples 24 to 33, wherein the feature prediction component is configured to predict one or more of the following related to the personal care device based on usage characteristics: stroke count, stroke force, stroke distance, rinse event, or end of life.

[0194] Example 35 may include a feature prediction device as described in any of Examples 24 to 34, wherein the feature prediction component is configured to perform one or more of the following: utilize a function that predicts one or more of the number of strokes, stroke force, stroke distance, rinse event, or end of life based on usage features; or run an additional machine learning model trained to determine one or more of the number of strokes, stroke force, stroke distance, rinse event, or end of life based on usage features.

[0195] Example 36 may include a feature prediction device according to any one of Examples 24 to 35, wherein the personal care device comprises a razor device.

[0196] Embodiment 37 may include any feature prediction device of Embodiments 24 to 36, further comprising a data acquisition device comprising a data acquisition component configured to collect data from a set of sensors, the set of sensors comprising an accelerometer of the personal care device, a gyroscope of the personal care device, additional sensors placed on the personal care device, or a combination thereof, or to prepare data before reaching a machine learning model, or to collect user profile data to be used to train a machine learning model; and a feedback device comprising a feedback component configured to provide feedback based on usage characteristics related to the personal care device.

[0197] Example 38 may include at least one computer-readable storage medium containing a set of instructions that, when executed by a processor, cause the processor to implement any of the feature prediction devices from Examples 24 to 37. For example, the at least one computer-readable storage medium, when executed, contains a set of instructions that cause the processor to implement a feature prediction component that trains a machine learning model to predict usage features associated with a personal care device using data from a subset of a set of sensors.

[0198] Example 39 may include a method for implementing the operation of any of the feature prediction devices from Examples 24 to 37. For example, the method may include training a machine learning model to predict usage features associated with a personal care device using data from a subset of a set of sensors.

[0199] Example 40 may include a method that includes means for performing the operation of a feature prediction device as described in any of Examples 24 to 37. For example, the method may include means for training a machine learning model to predict usage features associated with a personal care device using data from a subset of a set of sensors.

[0200] Embodiment 41 may include a data acquisition device that includes a data acquisition component configured to collect data from a set of sensors, the set of sensors including an accelerometer on a personal care device, a gyroscope on a personal care device, and additional sensors located on a personal care device, additional sensors located on a tracking device coupled to a personal care device, or a combination thereof.

[0201] Example 42 may include the data acquisition device of Example 41, wherein the data acquisition component is configured to provide access to data, and the data should be provided to one or more of the following: a comparator configured to receive output data from a machine learning model being trained to receive data from a subset of sensors, receive reference data from the set of sensors, and compare the output data with the reference data; and a machine learning model trained to predict usage characteristics related to a personal care device based on data from the subset of sensors.

[0202] Example 43 may include a data acquisition device from any of Examples 41 to 42, in which the sensor set includes an IMU, capacitance-to-digital converter, force sensor, rinse sensor, grip sensor, temperature sensor, accelerometer, gyroscope, magnetometer, resistive contact, piezoelectric contact microphone, Hall effect sensor, external camera, strain gauge, load cell, reed switch, linear variable differential converter, or a combination thereof.

[0203] Example 44 may include any of the data acquisition devices from Examples 41 to 43, wherein the data acquisition component is configured to prepare the data before it reaches the machine learning model.

[0204] Example 45 may include any of the data acquisition devices from Examples 41 to 44, wherein the data acquisition component is configured to perform flattening, normalization, rescaling, clock synchronization, removal and splitting, or a combination thereof.

[0205] Example 46 may include any of the data acquisition devices from Examples 41 to 45, wherein flattening includes passing the input data through a Butterworth filter to flatten the frequency response, normalization includes normalizing the input data around a mean value, and rescaling includes shrinking the input data to minimize the explosion gradient during training.

[0206] Example 47 may include any data acquisition device from Examples 41 to 46, wherein clock synchronization includes using corresponding signals in the input data and reference data to eliminate the offset between the input data and the reference data.

[0207] Example 48 may include any of the data acquisition devices from Examples 41 to 47, wherein the removal and splitting includes removing inactive periods longer than a predetermined period in the reference data and splitting the remaining active periods into subsets of activity periods.

[0208] Example 49 may include a data acquisition device as described in any of Examples 41 to 48, wherein the data acquisition component is configured to collect user profile data used to train a machine learning model.

[0209] Example 50 may include a data collection device from any of Examples 41 to 49, in which the user profile data includes survey data.

[0210] Example 51 may include any data acquisition device of Examples 41 to 50, wherein the data acquisition component is located in one or more of the following: a personal care device, a hub coupled in close proximity to a personal care device, or a server coupled to one or more of the personal care devices or hubs.

[0211] Embodiment 52 may include at least one computer-readable storage medium containing a set of instructions that, when executed by the processor, cause the processor to implement any of the data acquisition devices from Embodiments 41 to 51. For example, the at least one computer-readable storage medium, when executed, contains a set of instructions that cause the processor to implement a data acquisition component that collects data from a set of sensors, the set of sensors including an accelerometer on the personal care device, a gyroscope on the personal care device, and additional sensors located on the personal care device, additional sensors located on a tracking device coupled to the personal care device, or a combination thereof.

[0212] Example 53 may include a method for performing the operation of any of the data acquisition devices of Examples 41 to 51. For example, the method may include acquiring data from a set of sensors, the set of sensors including an accelerometer on a personal care device, a gyroscope on a personal care device, and additional sensors placed on a personal care device, additional sensors placed on a tracking device coupled to a personal care device, or a combination thereof.

[0213] Example 54 may include a method that includes means for performing the operation of any of the data acquisition devices of Examples 41 to 51. For example, the method may include means for acquiring data from a set of sensors, the set of sensors including an accelerometer on a personal care device, a gyroscope on a personal care device, and additional sensors placed on a personal care device, additional sensors placed on a tracking device coupled to a personal care device, or a combination thereof.

[0214] Example 55 includes a feedback device comprising a feedback component configured to provide feedback based on usage characteristics related to a personal care device, wherein the usage characteristics are predicted using a machine learning model trained with data from a subset of sensors.

[0215] Example 56 may include the feedback device of Example 55, where the feedback includes control functions on a personal care device, control functions on an Internet of Things (IoT) device adjacent to the personal care device, usage recommendations, replenishment functions, replacement information, or a combination thereof.

[0216] Example 57 may include a feedback device according to any one of Examples 55 to 56, wherein the feedback component is configured to initiate control commands for controlling elements of a personal care device.

[0217] Example 58 may include a feedback device from any of Examples 55 to 57, wherein the personal care device includes a razor device, the elements include one or more of the cartridge of the razor device or the dispenser of the razor device, and the feedback component includes either causing a control command to reach a controller module coupled to the cartridge of the razor device to change how one or more components of the cartridge touch the user, or causing a control command to reach a controller module coupled to the dispenser of the razor device to change how the personal care composition is dispensed from the razor device.

[0218] Example 59 may include a feedback device according to any one of Examples 55 to 58, wherein the feedback component is located in one or more of the following: a personal care device, a hub coupled in close proximity to a personal care device, or a server coupled to one or more of the personal care devices or hubs.

[0219] Embodiment 60 may include at least one computer-readable storage medium containing a set of instructions that, when executed by the processor, cause the processor to implement any of the data acquisition devices of Embodiments 55 to 59. For example, the at least one computer-readable storage medium may contain a set of instructions that, when executed, cause the processor to implement a feedback component that provides feedback based on usage characteristics related to a personal care device, the usage characteristics being predicted using a machine learning model trained with data from a subset of a set of sensors.

[0220] Example 61 may include a method for implementing the operation of any of the feedback devices of Examples 55 to 59. For example, the method may include providing feedback based on usage characteristics related to a personal care device, where usage characteristics are predicted using a machine learning model trained on data from a subset of a set of sensors.

[0221] Example 62 may include a method that includes means for performing the operation of a feedback device described in any of Examples 55 to 59. For example, the method may include means for providing feedback based on usage characteristics related to a personal care device, the usage characteristics being predicted using a machine learning model trained on data from a subset of a set of sensors.

[0222] The descriptions and drawings contained herein represent exemplary configurations and do not represent all implements within the scope of the claims. For example, operations and steps may be rearranged, combined, or otherwise modified. Structures and devices may also be represented in block diagram form to show the relationships between components and to avoid obscuring the described concepts. Similar components and features may have the same name but different reference numbers corresponding to different figures.

[0223] Several variations of this disclosure will be readily apparent to those skilled in the art, and the principles defined herein can be applied to other variations without departing from the scope of this disclosure. Accordingly, this disclosure is not limited to the examples and designs described herein, but should be given the broadest scope that is consistent with the principles and novel features disclosed herein.

[0224] The methods and / or components described herein can be implemented or executed by devices including general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic, discrete hardware components, or any combination thereof. A general-purpose processor may be a microprocessor, a conventional processor, a controller, a microcontroller, or a state machine. A processor can also be implemented as a combination of computer devices (e.g., a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors combined with a DSP core, or other configurations). Thus, the features described herein can be implemented in hardware or software and executed by a processor, firmware, or any combination thereof. When implemented in software executed by a processor, the features can be stored in the form of instructions or code on a computer-readable medium.

[0225] Computer-readable media include both non-temporary computer storage media and communication media, including any media that facilitates the transfer of code or data. Non-temporary storage media can be any available media that can be accessed by a computer. For example, non-temporary computer-readable media may include random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), compact discs (CDs) or other optical disc storage, magnetic disc storage, or any other non-temporary media that carries or stores data or code.

[0226] Furthermore, connection components can appropriately be referred to as computer-readable media. For example, when transmitting code or data from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), infrared, wireless, microwave signals, or other wireless technologies, coaxial cable, fiber optic cable, twisted pair, DSL, and wireless technologies are included in the definition of media. Combinations of media are also included in the scope of computer-readable media.

[0227] In this disclosure and the attached claims, the word “or” indicates an inclusive list, for example, a list of X, Y, or Z means X or Y or Z or XY or XZ or YZ or XYZ. Similarly, a list of one or more of X, Y, and Z, or a list of X, Y, and Z, or any combination thereof, means X or Y or Z or XY or XZ or YZ or XYZ. Also, the expression “based on” is not used to represent a closed set of conditions. For example, a step described as “based on condition A” may be based on both condition A and condition B. In other words, the expression “based on” is interpreted as “based at least partially.” Also, “a” or “an” indicates “at least one.” [Explanation of Symbols]

[0228] 100 users 105 devices 110 Communication Network 115 storage 120 servers 125 Artificial Neural Networks

Claims

1. It is a system, A personal care device comprising multiple sensors, wherein the multiple sensors include accelerometers and gyroscopes positioned on an inertial measuring unit (IMU) of the personal care device, and the personal care device is a personal care device for grooming or cleaning, configured to touch the user's body to groom or clean the body. A processor coupled to the plurality of sensors, wherein the processor is configured to run a machine learning model including an artificial neural network trained to predict usage characteristics related to the personal care device based on data from the plurality of sensors, Equipped with, The aforementioned usage characteristics include the force applied to the user's body by the personal care device, the distance the personal care device travels on the user's body, or a combination thereof. A system in which, when the machine learning model including the artificial neural network is executed by the processor, the force exerted on the user's body by the personal care device is predicted based solely on accelerometer and gyroscope data from the IMU input to a machine learning model trained to predict and output a measurement of the force, or the distance traveled by the personal care device on the user's body is predicted based solely on accelerometer and gyroscope data input from the IMU to a machine learning model trained to predict the distance traveled and output a measurement of the distance traveled, or a combination thereof.

2. The system according to claim 1, wherein the IMU does not have a magnetometer, or the system further includes a filter that prevents magnetometer data from a magnetometer from reaching one or more of the machine learning models being trained or the machine learning models being trained to predict the usage features.

3. The aforementioned processor, Displaced on the IMU and physically coupled to the accelerometer and the gyroscope via a connection on the IMU, Or, The controller of the personal care device is positioned on the controller and is physically coupled to the plurality of sensors via a connection between the controller and the IMU, or It is positioned on a hub adjacent to the personal care device and is connected to the plurality of sensors in a communicative manner via a connection between the hub and the personal care device, or The processor is located on a server that is communicably coupled to one or more of the personal care devices or the hubs adjacent to the personal care devices, and the processor is communicably coupled to the plurality of sensors via a connection between the server and one or more of the personal care devices or the hubs. The system according to claim 1.

4. The system according to claim 1, wherein the machine learning model including the artificial neural network is a stacked long short-term memory recurrent neural network (LSTM-RNN) trained using time-series IMU training data.

5. The system according to claim 1, wherein the processor is configured to calculate the work performed by the personal care device based on the force and the distance traveled, which are predicted independently of each other.

6. The aforementioned processor, A first machine learning model that predicts and outputs the measured value of the force based only on the accelerometer data and gyroscope data from the IMU input to the first machine learning model, or A second machine learning model that predicts and outputs the measured value of the travel distance based solely on the accelerometer data and gyroscope data from the IMU input to the second machine learning model, or A third machine learning model, which predicts and outputs measurements of the work performed by the personal care device based solely on accelerometer data and gyroscope data from the IMU input to the third machine learning model, The system according to claim 1, configured to perform one or more of the following.

7. The system according to claim 1, wherein the processor is configured to predict one or more of the following based on the usage characteristics: the number of strokes, stroke force, stroke distance, rinse events related to the personal care device, or end of life.

8. The aforementioned processor, A function is used to predict one or more of the following based on the usage characteristics: the number of strokes, the stroke force, the stroke distance, the rinse event, or the end of life. Alternatively, running an additional machine learning model trained to determine one or more of the following based on the usage characteristics: the number of strokes, the stroke force, the stroke distance, the rinse event, or the end of life. The system according to claim 7, configured to perform one or more of the following.

9. The system according to claim 1, wherein the personal care device includes a razor device, and the usage characteristics include the force applied to the user's body by the personal care device, and the distance the personal care device travels on the user's body.

10. The system according to claim 9, wherein the razor device is coupled in close proximity to and communicatively connected to an Internet of Things (IoT) device.

11. The system according to claim 1, further comprising a radio frequency interface, memory, storage, a sensor controller, an input / output device, or a combination thereof.

12. The system according to claim 4, wherein the machine learning model is trained to predict the usage characteristics associated with the personal care device based on the time-series IMU training data from the plurality of sensors and reference data from a six-degree-of-freedom tracking solution coupled to the personal care device.

13. A computer-readable storage medium comprising, when executed by a processor, a set of instructions causing the processor to execute a machine learning model including an artificial neural network trained to predict usage characteristics associated with a personal care device based on data from multiple sensors, The plurality of sensors include accelerometers and gyroscopes placed on the inertial measurement unit (IMU) of the personal care device, The personal care device is a personal care device for grooming or cleaning, configured to touch the user's body in order to groom or clean the body. The aforementioned usage characteristics include the force applied to the user's body by the personal care device, the distance the personal care device travels on the user's body, or a combination thereof. At least one computer-readable storage medium in which, when the machine learning model including the artificial neural network is executed by the processor, the force exerted on the user's body by the personal care device is predicted based solely on accelerometer and gyroscope data from the IMU input to a machine learning model trained to predict and output a measurement of the force, or the distance traveled by the personal care device on the user's body is predicted based solely on accelerometer and gyroscope data input from the IMU input to a machine learning model trained to predict and output a measurement of the distance traveled, or a combination thereof.

14. A method comprising the step of running a machine learning model which includes an artificial neural network trained to predict usage characteristics associated with a personal care device based on data from multiple sensors, The plurality of sensors include accelerometers and gyroscopes placed on the inertial measurement unit (IMU) of the personal care device, The personal care device is a personal care device for grooming or cleaning, configured to touch the user's body in order to groom or clean the body. The aforementioned usage characteristics include the force applied to the user's body by the personal care device, the distance the personal care device travels on the user's body, or a combination thereof. A method wherein, when the machine learning model including the artificial neural network is executed, the force exerted on the user's body by the personal care device is predicted based solely on accelerometer and gyroscope data from the IMU input to a machine learning model trained to predict and output a measurement of the force, or the distance traveled by the personal care device on the user's body is predicted based solely on accelerometer and gyroscope data input from the IMU to a machine learning model trained to predict and output a measurement of the distance traveled, or a combination thereof.

15. A machine learning model comprising an artificial neural network trained to predict usage characteristics associated with personal care devices based on data from multiple sensors, The plurality of sensors include accelerometers and gyroscopes placed on the inertial measurement unit (IMU) of the personal care device, The personal care device is a personal care device for grooming or cleaning, configured to touch the user's body in order to groom or clean the body. The aforementioned usage characteristics include the force applied to the user's body by the personal care device, the distance the personal care device travels on the user's body, or a combination thereof. When the machine learning model, including the artificial neural network, is executed, the force exerted on the user's body by the personal care device is predicted based solely on accelerometer and gyroscope data from the IMU input to a machine learning model trained to predict and output measurements of the force, or the distance traveled by the personal care device on the user's body is predicted based solely on accelerometer and gyroscope data input from the IMU to a machine learning model trained to predict and output measurements of the distance traveled, or a combination thereof.

Images