Roasting a food item

Abstract

A method (200) for roasting a food item (125) comprises steps of detecting (210) a degree of roasting to which the food item (125) is to be roasted; cooking, in particular roasting (220), the food item (125) by means of a domestic appliance; detecting (225) a temperature profile with respect to the food item (125) during cooking, in particular during roasting; determining (230), on the basis of the temperature profile, a point in time at which the food item (125) is expected to be roasted to the degree of roasting; and providing (235) an indication of the determined point in time.

Description

[0001] Roasting a foodstuff

[0002] The present invention relates to the roasting of a foodstuff. In particular, the invention relates to the roasting of a foodstuff to a predetermined degree of roasting.

[0003] When preparing a dish, it may be necessary to roast one of the ingredients or an intermediate product. Roasting involves exposing a plant or animal food to high heat to remove moisture, alter its flavor, darken its color, increase its shelf life, or change its texture. The degree to which the food should be roasted can be specified as the degree of roasting. For example, a recipe might indicate that a roux should be lightly or heavily roasted.

[0004] To achieve a predetermined level of roasting, a specific timeframe can be specified, after which the desired level of roasting should be reached. This timeframe can be determined based on the quantity of food. Additionally, a temperature can be specified that the food should have reached when the desired level of roasting is achieved. If the level of roasting is too low, the dish will not cook properly; if it is too high, it may spoil.

[0005] An individual food preparation process can differ from a reference process, which is used to determine when the predetermined degree of browning is reached, in numerous parameters. These parameters can relate to the quality, moisture content, starting temperature, condition, or quantity of the food used. The browning process often cannot be reliably controlled based solely on duration, resulting in the food ending up with a browning level higher or lower than desired. An example is browning ground meat in the Bosch Cooklt. The correct browning time depends not only on the water and fat content of the ground meat but also on its quantity. After preparation, the ground meat can be in three states: "not browned," "well browned," and "burnt." Traditionally, cooking times are specified for each step of the Cooklt's cooking process.If the manufacturer of the Cooklt or a provider of recipes for the Cooklt specifies a browning time X, variations in water and fat content, as well as the amount of ground meat, will lead to different results for each user. For some users, the ground meat will be "not browned" (Type I error), for others it will be "well browned" (desired result), and for still others it will be "burnt" (Type II error). If a user wants to brown ground meat to cook chili con carne, a Type I error will result in an edible dish that might taste a little bland in terms of umami roasted flavors. The user is unlikely to complain to the Cooklt manufacturer or the recipe provider about this. However, a Type II error will very likely result in a complaint. Therefore, it is advisable to choose a browning time X that is slightly too short.However, this means that in most cases the minced meat will not be browned enough.

[0006] One of the problems underlying the present invention is to provide an improved technique for roasting a foodstuff. The invention solves this problem by means of the subject matter of the independent claims. Dependent claims describe preferred embodiments.

[0007] According to a first aspect of the present invention, a first method for roasting a foodstuff comprises the steps of determining the desired degree of roasting; cooking, in particular roasting, the foodstuff using a household appliance; determining the temperature profile of the foodstuff during cooking, in particular during roasting; determining, based on the temperature profile, a time at which the foodstuff is expected to be roasted to the desired degree of roasting; and providing an indication of the determined time. The household appliance may, in particular, be an automatic cooking appliance, e.g., the Cooklt.

[0008] Under the influence of heat, food can successively reach different degrees of roasting. For example, food can be lightly roasted, then seared, dark roasted, very dark roasted, and finally burnt. Burnt food is usually no longer suitable for consumption. The degree of roasting can be indicated numerically on a predetermined scale, for example, from zero (not roasted) to six (burnt). Examples of food include meat, fish, poultry, shellfish, vegetables, nuts, seeds, coffee or cocoa beans, grains, malt, chickpeas, bread, a roux, or popcorn.

[0009] By taking the temperature profile into account, the point at which a predetermined degree of roasting is reached can be determined more accurately. Accordingly, it is easy to determine how long it will take to reach the desired degree of roasting. Using the proposed method, a cooking process that requires roasting the food to the predetermined degree of roasting can be implemented more efficiently. In one embodiment, the user can be notified of the precise time, allowing them to manually end the roasting process or begin a subsequent cooking step.

[0010] A basic idea is to divide the cooking process into different phases and use sensor measurements to determine which phase the cooking process is currently in or from which phase it is transitioning to the next. For example, a sensor can be placed on the bottom of the appliance's cooking vessel (e.g., the bottom of the Cooklt pot) to measure the temperature of the base (hereinafter referred to as the base temperature). In addition, a sensor can be provided to measure the temperature on the vessel wall, side, or rim (hereinafter referred to as the wall temperature). Another sensor (e.g., an infrared sensor) can be used to determine the temperature of the ambient air or the air inside the cooking vessel (hereinafter referred to as the ambient temperature).Typically, the floor temperature, the corresponding wall temperature, and the corresponding ambient temperature will not be independent of each other, but rather exhibit a certain correlation. Preferably, the sensor measurements (e.g., the floor temperature, the corresponding wall temperature, and the corresponding ambient temperature) are aggregated into latent features. Latent features can be thought of as higher-order features. In the case of object recognition, for example, pixels of an image are fed into a neural network. An edge would then be, for example, a latent feature. The pixel level has therefore already been left behind, but no object has yet been recognized.

[0011] In some embodiments, the temperature profile with respect to the food includes a bottom temperature, a wall temperature and an ambient temperature of a cooking vessel in which the food is cooked, in particular roasted, wherein the cooking vessel is preferably a pot.

[0012] In some embodiments, the cooking process of food is divided into at least two phases based on the temperature profile of the food, with roasting occurring in at least one of these phases. For example, when browning ground meat, there is a water phase, a frying phase, and a burning phase. In the water phase, the meat slowly begins to release water. The bottom temperature stabilizes at approximately 100 °C due to evaporation. When the water has largely evaporated, the ground meat enters the frying phase, in which the bottom temperature rises above 100 °C. As soon as the bottom temperature rises above approximately 150 °C, the ground meat is in the burning phase, at which point the user should be prompted to end the cooking process, for example, by adding liquid or other ingredients. Roasting occurs in both the frying and burning phases.

[0013] Unfortunately, it is often not entirely straightforward to definitively determine the cooking phase from sensor measurements. A fundamental approach is to assign at least parts of different instances of sensor measurement data, particularly temperature profiles, to a specific cooking phase. This process is also frequently referred to as labeling in machine learning.

[0014] When all training data is labeled by humans, it is often referred to as supervised learning. However, supervised learning algorithms require a large amount of labeled training data to produce well-performing classifiers. In many use cases, only a large amount of unlabeled data is available. The process of data labeling is often difficult, tedious, expensive, or time-consuming, as it requires the use of human experts or specialized equipment.

[0015] Semi-supervised learning (SSL) attempts to overcome the bottleneck of labeled data by allowing the learning model (i.e., the machine learning technique) to integrate some or all of the available unlabeled data into the learning process. The goal is to improve the model's learning performance through the additional use of machine-labeled measurements while simultaneously reducing manual effort. In contrast to supervised learning, SSL involves experts labeling only a small portion of the data. The manually labeled and unlabeled data can then be used by the semi-supervised learning algorithm (SSL algorithm). This algorithm selects a specific subset of the unlabeled data where a label prediction can be made with the highest degree of certainty and labels it automatically.

[0016] For example, the following self-training algorithm, which is a type of SSL algorithm, can be used within the context of the invention described here: We have a set of labeled data called L and a set of unlabeled data U. The classifier is trained on L and then automatically generates labels for U. The generated labels are sorted according to their predictive confidence. For example, the data for which the classifier was very confident that the assigned label was correct can appear at the top of a list, while the data for which a different label could just as easily have been assigned appear at the bottom. Subsequently, a set S containing the most confident predictions is created. In the aforementioned list, the n top entries would thus be transferred to the set S.Set S is then deleted from set U and assigned to set L, after which the classifier is retrained with set L. This process allows for a richer and more robust set of training data, enabling continuous improvement of the classifier.

[0017] In some implementations, a SMOTE algorithm (SMOTE = Synthetic Minority Oversampling Technique) can be used. The rationale behind this is that many real-world problems addressed by artificial intelligence suffer from unbalanced data. A dataset is considered unbalanced when the classification categories (such as "undercooked," "well-cooked," and "burnt") are not evenly represented. Often, the datasets consist predominantly of "normal" observations and only a small percentage of anomalous or interesting observations. Furthermore, the cost of incorrectly predicting an anomalous example is often much higher than the cost of making the opposite error.For example, if you want to teach a car to drive autonomously on the highway using cameras, it's not surprising that recording data during a highway drive generates a large amount of data showing the car following another car at a certain distance, while a child running into the road is a relatively rare occurrence. Similarly, in the example of browning ground meat mentioned above, it's to be expected that field measurements will much more frequently indicate "not browned" than "burnt."

[0018] To address the problem of unbalanced data, different costs can be assigned to observations of different classes. Assigning higher costs to observations of the smaller class forces the model to consider these observations sufficiently during training, leading to more balanced predictions. Another option is to resample the original dataset. This can be achieved, for example, by oversampling the minority class and / or undersampling the majority class. Oversampling is a technique that deliberately overrepresents a certain class of observations, thus compensating for unbalanced data. Conversely, undersampling involves deliberately underrepresenting a portion of the data.A classic method for oversampling the minority class is to perform repeated sampling with replacement of the respective observations. The resulting dataset thus contains copied observations of the smaller class and therefore a larger number of observations of the original minority class.

[0019] The SMOTE algorithm performs oversampling to balance the size of the individual classes. However, instead of simply replicating the minority class instances, a key idea of ​​SMOTE is to introduce synthetic examples. This synthetic data is generated by interpolating between multiple minority class instances located in a defined neighborhood.

[0020] Based on these considerations, further embodiments of the method are provided.

[0021] In some implementations, a trained classifier determines the current stage of the cooking process. This classifier can reside on the appliance itself or at a central location, such as a server on the internet or in a cloud infrastructure. Initially, a classifier trained solely through supervised learning can be used. This means that all training data was initially labeled by human experts. Then, when a large number of appliances in the field, located with a wide variety of users, send data to the central location, the classifier can be continuously improved through semi-supervised learning. This can potentially enable a continuous learning and improvement process, also known as the AloT cycle, where AloT stands for "Artificial Intelligence" and "Internet of Things."The appliance manufacturer can collect information regarding the use of its appliances and learn from it in order to provide the user with an improved user experience of the appliance.

[0022] In some embodiments, the method includes the further steps of transferring the temperature profile, in particular from the household appliance or other household appliances of the same or similar design, to a central location and storing the temperature profile at the central location, in particular in a data lake. A data lake can be characterized by the fact that the data to be stored does not have to have a predefined structure at the time of storage, as is the case, for example, with a classic database. A data lake can therefore store structured, unstructured, or even semi-structured data. This can enable flexible analyses using machine learning. It is conceivable that unstructured data is only structured later, for example, during reading. This approach can be described as schema-on-read.

[0023] In some embodiments, a semi-supervised learning algorithm, in particular a self-training algorithm, preferably the self-training algorithm described above, is used to train the classifier.

[0024] In some embodiments, a large number of temperature profiles stored centrally are automatically labeled and evaluated based on the predictive accuracy of the assigned label, in particular by sorting. Temperature profiles whose predictive accuracy exceeds a predetermined threshold can be assigned to a training dataset, which is used to train a classifier. An SMOTE algorithm can be used to assign the automatically labeled temperature profiles to the training dataset. The SMOTE algorithm can oversample the temperature profiles that are particularly critical, such as those that resulted in burnt ground meat. This oversampling can be achieved, for example, by inserting synthetic temperature profiles into the training dataset, which, for instance,obtained by interpolation between two measured temperature profiles, which in particular each resulted in burnt minced meat.

[0025] Particularly if the classifier has been improved, i.e., further trained, at least once at the central location, it can be advantageous to transfer this improved classifier to one or more household appliances. Accordingly, in some implementations, the classifier is trained with the training data set at the central location, and the trained classifier can then be transferred from the central location to the household appliance.

[0026] In some embodiments, the method can be used, for example, for the automated preparation of a dish by determining the end of a cooking step that includes roasting based on the method and automatically controlling a subsequent cooking step once roasting is complete. The dish can be prepared using an integrated cooking appliance that has a heat source and / or a drive motor to heat, stir, or otherwise mechanically process the food. Such integrated cooking appliances are typically networked, for example, using transmission standards such as WiFi (IEEE 802.11) and / or Bluetooth. A particularly well-known integrated cooking appliance is, for example, the Bosch Cooklt.

[0027] The instruction can be given to an operator, allowing them to end the roasting step in time or begin a subsequent step. The instruction can also be evaluated automatically. In particular, the heat input can be controlled to ensure that the roasting process continues for precisely the duration required to achieve the predetermined degree of roasting, based on the instruction.

[0028] The timing is preferably determined (as already explained) using a machine learning technique (such as the classifier mentioned above) that is trained on a large number of temperature profiles and their corresponding roasting levels. In some embodiments, training data is obtained during cooking processes, each comprising a cooking step that includes roasting the food. Preferably, the cooking steps are identical or comparable, and more preferably, the cooking processes follow the same recipe.

[0029] In an environment where the same recipe is prepared by different users using comparable appliances, training data can be easily collected and processed for use in training machine learning systems. For example, the cooking process can be carried out multiple times independently using integrated kitchen appliances employed by different users to prepare a dish. During preparation, predefined data can be measured and collected from the kitchen appliances. This data can then be further processed and used to train the machine learning system. For networked, integrated kitchen appliances already in the customer's possession, it is particularly advantageous to collect training data using big data methods and store it in the aforementioned data repository, especially a data lake, for further use.

[0030] Since different foods behave differently when roasted, training data for machine learning techniques is preferably assigned to a specific food item. The timing can also be determined in relation to a quantity of the food. An example of two foods that behave differently when roasted are ground beef and goulash. Both may be made from pork, but even the different size of the pieces can influence the roasting behavior.

[0031] For different foodstuffs, the machine learning technique, particularly the classifier, can be specifically trained. The quantity of the foodstuff can be an influencing factor that affects the determination of the time. The quantity can be a parameter with respect to which the machine learning technique, especially the classifier, is trained or which is specified as a boundary condition. Furthermore, the determination using the machine learning technique can be directed at a predetermined quantity, and a specific time can be adjusted if a different quantity is present. For example, a certain duration until the desired degree of roasting is achieved can be multiplied by a factor that depends on the deviation of the actual quantity of the foodstuff from an assumed quantity.In one embodiment, the machine learning technique includes a recurrent network, such as LSTM (Long Short Term Memory) or GRU (Gated Recurrent Unit). Other examples of such recurrent neural networks include an Elman network, a Jordan network, a Hopfield network, a recurrent neural network, and a fully interconnected neural network. A transformer can also be used to enable parallelization via multi-head attention.

[0032] A recurrent neural network has a kind of memory because it uses information from previous inputs to influence a current input or output. The output of a recurrent neural network therefore depends on the previous elements within the sequence. Feedback to a recurrent neural network can be direct, indirect, lateral, or total. Recurrent networks share parameters across all layers of the network. While feedforward networks have different weights across each node, recurrent neural networks share the same weight parameter within each layer. However, these weights are still adjusted through the processes of backpropagation and gradient descent to facilitate reinforcement learning.

[0033] Therefore, a recurrent neural network can be used to improve the analysis of time-dependent trends in measured values. In this way, it is possible to predict a likely future trend of a parameter based on its past behavior. This property of the recurrent network can be advantageously used to determine the time of a given event.

[0034] In some implementations, machine learning techniques include a transformer and / or a foundation model, such as a Generative Pre-trained Transformer (GPT) and / or a Large Language Model. While a recurrent neural network operates sequentially, a transformer can parallelize the processes, potentially resulting in speed advantages.

[0035] In some embodiments, a time at which the food reaches the desired degree of roasting is estimated based on the foodstuff and / or a quantity thereof; and a further (or second) indication of the estimated time may be provided. The estimated time may be determined based on information from a cooking recipe according to which a dish is prepared, the recipe including cooking steps, one of which involves roasting a foodstuff. Such an estimated time can provide a good guideline, which can be refined by the time determined based on the temperature profile. If the indication is provided to a user, the first and second times may be indicated. In particular, the first time may be given as a guideline and the second as a deviation from the guideline.The user can choose whether to rely on the guideline value or also take the deviation into account.

[0036] If the deviation of the second time from the first time is not relevant, for example because it is below a predetermined time span of, for example, approximately 5 seconds, then providing a reference to the specific time can be omitted, and the estimated time can be used for preparing the food.

[0037] According to a further aspect of the present invention, a method for training a machine learning technique (hereinafter also referred to as the second method) comprises steps of acquiring a first data set comprising temperature profiles relating to first foodstuffs while the first foodstuffs are roasted to associated roast levels; and of training the machine learning technique on the basis of the data of the first data set to determine and / or predict the roast level based on the temperature profile.

[0038] A machine learning technique trained using the second method can be applied in a first method described herein. The technique is preferably trained on a large number of temperature profiles relating to initial foodstuffs with associated roast levels. The roast levels can be assessed, for example, by human employees, preferably experts such as a chef. In other words, the roast level of a foodstuff undergoing a roasting process can be evaluated by a person. The data in the first dataset can be systematically determined. For example, in one data set, the same quantity of the same foodstuff can be roasted to different roast levels under the same heating conditions. In another data set, the same quantities of the same foodstuff can be roasted to the same roast level under different temperatures.Other data series can be created in a similar manner.

[0039] Creating training data with the involvement of a human expert is complex and expensive, so further training data can be generated automatically. In a further development of the present invention, a second data set can be acquired, comprising temperature profiles relating to a second foodstuff while the second foodstuff is roasted. This data can be acquired by integrated kitchen appliances and may not include the degree of roasting to which the second foodstuff was roasted.

[0040] The roasting levels of the data from the second dataset can then be determined based on data from the first dataset. A known technique, such as Extra Trees or XGBoosting, can be used for this purpose. Optionally, the machine learning technique can be trained only on the second dataset. The second dataset can be collected automatically, particularly if it encompasses cooking processes recorded by identical kitchen appliances and, even more preferably, when preparing a dish according to the same recipe.

[0041] The preferred technique is machine learning, in particular the classifier mentioned above, which is trained to predict, based on a temperature profile of a foodstuff during roasting, a point in time at which the foodstuff will have reached a predetermined degree of roasting. XGBoost or time series analysis, for example, can be used for this purpose.

[0042] Machine learning techniques, particularly classifiers, are increasingly being trained or used to predict timing at predetermined intervals. For example, a roasting process that takes approximately...

[0043] The process will take 8 minutes, be tracked for approximately 5 minutes, before the exact time is predicted based on a recorded temperature profile. Further predictions can then be made at intervals of approximately 1 or 2 minutes, for example.

[0044] According to yet another aspect of the present invention, a first device, in particular a household appliance, for roasting food to a predetermined degree of roasting comprises a heat source for heating, in particular for cooking, the food; a temperature sensor for detecting a temperature profile with respect to the food during cooking, in particular during roasting; an output device for providing an indication of a specific time; and a processing unit. The processing unit is configured to determine, based on the temperature profile, a time at which the food is expected to be roasted to the predetermined degree of roasting.

[0045] The device is preferably configured to carry out a method described herein, either partially or completely. All features or advantages of the first device can be transferred to all embodiments of the described methods, and vice versa. The embodiments can, of course, also be combined with one another.

[0046] The heat source of the aforementioned device preferably acts on a cooking vessel for holding the food, e.g., a pot. Several temperature sensors can be provided at different locations within the cooking vessel, and the processing unit can determine the timing of temperature profiles at these different locations. The first device can be part of an integrated food processor. The food processor is preferably designed to assist and partially automate the preparation of food by a user.

[0047] A first temperature sensor can be located, for example, at the bottom of the cooking vessel, and a second temperature sensor can be located on the wall, side, or edge of the cooking vessel. Preferably, a third temperature sensor is provided that measures the temperature of the air inside the cooking vessel or a temperature outside the cooking vessel.

[0048] According to a further aspect of the present invention, a further device, in particular a central unit, for training a machine learning technique, especially the classifier mentioned, is proposed, which is hereinafter also referred to as the second device. The second device comprises a data storage unit for receiving data sets, in particular a data lake; wherein a data set comprises a temperature profile relating to a foodstuff while the foodstuff is roasted to a corresponding degree of roasting; and a processing unit. The processing unit is configured to train the machine learning technique, in particular the classifier, on the basis of the data sets in order to determine, on the basis of a temperature profile relating to a foodstuff during its roasting, an estimated time at which a predetermined degree of roasting is reached.All features or advantages of the second device can be transferred to all embodiments of the described methods (as well as to the first device), and vice versa. The embodiments can, of course, also be combined with one another.

[0049] In some embodiments, the processing device is configured to automatically label the corresponding temperature profiles for a large number of data records stored in the data memory; to evaluate the data records based on the predictive accuracy of the assigned label, in particular to sort them, and to assign the data records of temperature profiles whose predictive accuracy exceeds a predetermined threshold to a training data set with which the classifier is trained.

[0050] In some embodiments, the second device is equipped with an interface for transferring the trained classifier from the device to a household appliance for roasting a food.

[0051] The second device is preferably configured to partially or completely execute a method described herein, in particular a training method. All features or advantages of the second device can (as already stated) be transferred to the method, or vice versa.

[0052] A processing device described herein is preferably implemented electronically and may, for example, comprise a programmable microcomputer or microcontroller. A method described herein may be in the form of a computer program product with program code means. The computer program product may also be stored on a computer-readable data carrier. All methods can therefore be computer-implemented. Non-limiting embodiments of the invention are now described in more detail with reference to the accompanying figures, in which:

[0053] Figure 1 shows a system with a food processor;

[0054] Figure 2 shows a flowchart of a first procedure; and Figure 3 shows a flowchart of a second procedure.

[0055] Figure 1 shows a system 100 with a food processor 105 and an external central unit 110. The food processor 105 is designed to assist a user in preparing a dish according to a predetermined recipe. The food processor 105 includes a heat source 115 that acts on a cooking vessel 120, into which food 125 can be placed. An optional drive motor 130 can move the food 125 within the cooking vessel 120. An optional weight sensor 135 can determine the quantity of food 125 in the cooking vessel 120. An optional interaction device 140 is designed to provide a message to a user as an output device and / or to capture user input as an input device. In the illustrated embodiment, the interaction device 140 is implemented as an optical display, which is preferably touch-sensitive.The interaction device 140 can be used to display a cooking step of a cooking recipe, which is to be implemented using the kitchen machine 105.

[0056] A processing device 145 comprises or implements a machine learning technique 150. The processing device 145 can, in particular, include the aforementioned classifier. At least one first temperature sensor 155 and preferably a second temperature sensor 160 are provided on the cooking vessel 120. The temperature sensors 155 and 160 are located at different points on the cooking vessel 120. In the illustrated embodiment, the first temperature sensor 155 is located, for example, in the area of ​​the bottom of the cooking vessel 120, and the second temperature sensor 160 is located somewhat higher, on a side wall of the cooking vessel 120. Optionally, one or more additional temperature sensors can be provided on the cooking vessel 120. An optional third temperature sensor 165 is configured to determine an ambient temperature.In addition, an interface 170 can be provided, which allows the kitchen machine 105 to communicate with the external central point 110.

[0057] The external location 110 preferably comprises a corresponding interface 175 for communication with the food processor 105, a processing unit 180, and an optional data storage device 185. The external location 110 is preferably configured to communicate with a large number of food processors 105.

[0058] It is proposed to use machine learning technique 150, in particular the classifier, to improve the timing of the roasting process of a foodstuff 125 with a food processor 105, ensuring that the foodstuff 125 reaches a predetermined degree of roasting. For this purpose, machine learning technique 150 is provided with at least one temperature profile from the food processor 105, and from this, technique 150 determines the current degree of roasting of the foodstuff 125 and / or a time at which the foodstuff 125 is expected to reach a predetermined degree of roasting.

[0059] The technology 150 may have been previously trained for this purpose by the external central authority 110. Training data for the technology 150 can be determined by a variety of kitchen machines 105 while comparable cooking steps are carried out, including a corresponding roasting process. Figure 2 shows a flowchart of a first method 200 for roasting a foodstuff 125. The method 200 can be carried out, in particular, using a kitchen machine 105.

[0060] In step 205, a machine learning technique 150, in particular the aforementioned classifier, can be provided which is trained to determine the current degree of roasting of a foodstuff 125 based on at least one temperature profile during a roasting process. Preferably, the technique 150 is configured to determine a time at which the foodstuff 125 has reached a predetermined degree of roasting. The provision of the technique 150 is described in more detail below with reference to Figure 3.

[0061] In step 210, a cooking recipe can be entered, which includes cooking steps to prepare a predetermined dish using the food processor 105. One of the cooking steps involves roasting a food item 125, and this cooking step includes an indication of how much the food should be roasted, in the form of a degree of roasting.

[0062] In step 215, a roasting time can be determined, which specifies how long the food 125 should be roasted under predetermined conditions to achieve a predetermined degree of roasting. This roasting time can be included as a specification in the cooking recipe.

[0063] In step 220, the food 125 can be roasted by placing it in the cooking vessel 120 while the heat source 115 heats the cooking vessel 120. Typically, the cooking vessel 120 is brought to a predetermined temperature before the food 125 is added.

[0064] During roasting in step 220, the temperature of the food (125) or the cooking vessel (120) is determined as a time-dependent profile in step 225. Temperature values ​​can be recorded at a predetermined sampling rate. This sampling rate can be, for example, approximately twice per second to approximately ten times per second.

[0065] Based on the temperature profile, a predicted end to the roasting process can be determined in step 230, such that the food 125 has reached a predetermined degree of roasting. The predicted end can be determined within a predetermined prediction horizon, for example, up to approximately 3 minutes. Within this horizon, the determination can be made in real time. In addition to the temperature profile, one or more further parameters can be included in the determination of the predicted end. Such parameters can include, for example, the type of food 125 or its quantity. Furthermore, it can be taken into account whether and how the food 125 is stirred in the cooking vessel 120.

[0066] In step 235, a user can be informed when the current roasting step will end. This step can be executed multiple times, updating the information provided to the user. First, the roasting time from step 215 can be displayed. After a portion of the roasting time has elapsed, the anticipated end of the roasting process can be determined for the first time in step 230. If the anticipated end time differs significantly from the actual roasting time, a message can be displayed recommending that the user adjust the predetermined roasting time accordingly. If the anticipated end time from step 230 differs only insignificantly from the end time determined in step 215, no further message can be displayed to the user.

[0067] Determining the expected end of the roasting process in step 230 can be repeated periodically. With each determination, the notification to the user can be updated, activated, or deactivated. Roasting is usually ended by the user, for example, by adding more food 125 to the cooking vessel 120 or removing the cooking vessel 120 from the food processor 105.

[0068] Figure 3 shows a flowchart of a second procedure 300 for providing a machine learning technique 150 for the first procedure 200 or the kitchen machine 105. For illustrative purposes, it is assumed that the second procedure 300 is executed by the external central authority 110.

[0069] In step 305, initial data 310 are provided, which may be stored in the data memory 185. The initial data 310 each include a temperature profile 325. A degree of roasting 330 is also specified. The initial data 310 describe a roasting process carried out on a food item 125 using a food processor 105. Optionally, parameters 335 that influence the roasting process are also included in the initial data 310. Such parameters may relate, for example, to the type of food item 125, its quantity, or an ambient temperature.

[0070] The roast level 330 of the first data 310 could have been determined by a human expert. For this purpose, the expert could have observed the roasting process and, in particular, recorded visual, acoustic, and / or olfactory data on the food 125. A large number of first data 310 can be compiled into a first data set 340. Good results were achieved with approximately 100 first data 310 in the first data set 340.

[0071] The second set of data 320 is structured similarly to the first set of data 310, but initially does not include a roast level 330. The second set of data 320 may have been sampled from kitchen machines 105 while corresponding roasting processes were carried out with corresponding foodstuffs 125. For the evaluation of the second set of data 320, no information on the achieved roast level is available, and usually no optical, acoustic, or olfactory data about the roasting process are available either. It is proposed to determine the roast levels 330 in the second set of data 320 in a step 350 based on the first set of data 310. For this purpose, a technique such as XGBoost or a time series analysis can be used. The aforementioned self-training algorithm can also be employed.

[0072] A large number of second data sets 320 can be compiled into a second data set 345. Data sets 310 and 320 from data sets 340 and 345 can be compared for consistency, and the procedure in step 350 can be adjusted so that the first and second data sets 310 and 320 are consistent with each other. Good results were achieved with approximately 6000 second data sets 320.

[0073] In step 355, the machine learning technique 150 can be trained based on data 310, 320. The training is carried out in such a way that the technique 150 can determine a current roast level and / or an expected time at which roasting to a predetermined roast level is completed, based on a temperature profile.

[0074] The trained technique 150 can then be connected to a kitchen machine 105. There, the technique 150 can be used to accompany a roasting process, as described in more detail with reference to the first method 200.

[0075] Reference sign

[0076] 100 System

[0077] 105 Kitchen machine

[0078] 110 external position

[0079] 115 Heat source

[0080] 120 cooking vessels

[0081] 125 Foods

[0082] 130 drive motor

[0083] 135 Weight sensor

[0084] 140 Interaction device, output device

[0085] 145 Processing unit

[0086] 150 Machine Learning Techniques

[0087] 155 first temperature sensor

[0088] 160 second temperature sensor

[0089] 165 third temperature sensor

[0090] 170 interface

[0091] 175 interface

[0092] 180 processing unit

[0093] 185 data storage devices

[0094] 200 first procedure

[0095] 205 Kl provide

[0096] Record 210 cooking recipes

[0097] 215 Determine roasting time

[0098] Roast 220 foods

[0099] 225 Determine temperature profile

[0100] 230 Determine the expected end of roasting

[0101] 235 Give a note to the user

[0102] 300 second procedure 305 provide first data

[0103] 310 first data

[0104] 315 provide second data

[0105] 320 second data 325 temperature profile

[0106] 330 roast level

[0107] 335 parameters

[0108] 340 first data set

[0109] 345 second data set 350 determine roast level

[0110] 355 Training

Claims

PATENT CLAIMS 1. Method (200) for roasting a foodstuff (125), wherein the method (200) comprises the following steps: - Determining (210) a degree of roasting to which the food (125) is to be roasted; - Cooking, especially roasting (220), of the food (125) using a household appliance; - Recording (225) a temperature profile with reference to the food (125) during cooking, especially during roasting; - Determine (230), based on the temperature profile, a time at which the food (125) is expected to be roasted to the desired degree of roasting; and - Providing (235) a reference to the specified time.

2. Method (200) according to claim 1, wherein the temperature profile with reference to the food comprises a bottom temperature, a wall temperature and an ambient temperature of a cooking vessel in which the food is cooked, in particular roasted, wherein the cooking vessel is preferably a pot.

3. Method (200) according to one of the preceding claims, wherein a process of cooking the foodstuff based on the temperature profile with reference to the foodstuff is divided into at least two phases, wherein roasting takes place in at least one of the phases.

4. Method (200) according to one of the preceding claims, comprising the further steps - Transferring the temperature profile to a central location and - Storing the temperature history at a central location, especially in a data lake.

5. Method (200) according to claim 4, wherein a plurality of temperature profiles stored at the central location are machine-labeled and evaluated, in particular sorted, based on the prediction reliability of the assigned label, wherein the temperature profiles whose prediction reliability exceeds a predetermined threshold are assigned to a training data set with which a classifier is trained.

6. Method (200) according to claim 5, wherein a SMOTE algorithm is used for assigning the machine-labeled temperature profiles to the training data set.

7. Method (200) according to claim 5 or 6, wherein the classifier is trained with the training data set at the central location and the trained classifier is transferred from the central location to the home device.

8. Method (200) according to any one of claims 3 to 8, wherein a trained classifier determines which phase the cooking process is currently in.

9. Method (330) according to claim 8, wherein the classifier is trained, among other things, to predict, on the basis of a temperature profile relating to the food (125) while the food (125) is being roasted, a time at which the food (125) is expected to have reached a predetermined degree of roasting (330).

10. Device (105), in particular a household appliance, for roasting a foodstuff (125) to a predetermined degree of roasting, the device comprising the following elements: - a heat source (115) for heating, in particular for cooking, the food (125); - a temperature sensor (155, 160) for recording a temperature profile with reference to the food (125) during cooking, especially during roasting; - a processing device (145) designed to determine, based on the temperature profile, a time at which the foodstuff (125) is expected to be roasted to the predetermined degree of roasting; and - an output device (140) for providing a reference to the specified time.

11. Device (105) according to claim 10, wherein the heat source (115) acts on a cooking vessel (120) for receiving the food (125), in particular on a pot; wherein several temperature sensors (155, 160) are provided at different locations of the cooking vessel (120); and wherein the processing device (145) determines the time with respect to temperature profiles at the different locations.

12. Device (105) according to claim 11, wherein a first temperature sensor is arranged at the bottom of the cooking vessel (120) and a second temperature sensor is arranged on the wall, on a side or on the edge of the cooking vessel (120), wherein preferably a third temperature sensor is provided which measures the temperature of air in the cooking vessel or a temperature outside the cooking vessel.

13. Device (110), in particular a central component, for training a machine learning technique (150), in particular a classifier; wherein the device comprises the following elements: - a data storage (185) for storing data records (310, 320), in particular a data lake; - wherein a data set (310, 320) comprises a temperature profile with reference to a foodstuff (125) while the foodstuff (125) is roasted to an assigned degree of roasting; and - a processing unit (180) which is set up to train the machine learning technique (150) on the basis of the data sets in order to determine, on the basis of a temperature profile relating to a foodstuff (125) during its roasting, an expected time at which a predetermined degree of roasting of the foodstuff will be reached.

14. Device (110) according to claim 13, wherein the processing device (180) is configured to - to automatically label the corresponding temperature profiles for a large number of data records (310, 320) stored in the data storage (185); - to evaluate the data sets based on the predictive reliability of the assigned label, in particular to sort them, and - to assign the temperature trend data sets, whose predictive accuracy exceeds a predetermined threshold, to a training data set, which is used to train a classifier.

15. Device (110) according to claim 13 or 14, comprising an interface (175) for transferring the trained classifier from the device to a household appliance for roasting a food (125).

Images