Method for providing a human-machine interface for cooking appliances based on machine learning

Abstract

Computer-implemented method for detecting clusters and / or behavior patterns in cooking appliances, comprising at least the steps: • Collecting and / or receiving first data sets from at least one cooking appliance; • Creating at least one prediction model based on an automatic analysis of the first data sets, in particular based on machine learning and / or artificial intelligence.

Description

The invention relates to a method for detecting behavior patterns in the selection and selection behavior in cooking appliances and for providing a human-machine interface for this purpose.In particular in large kitchens, cooking devices are used which, in addition to the manual execution of cooking processes in which the user manually changes cooking parameters at the respective point in time, make it possible to carry out automatic cooking processes based on cooking programs.In this case, it can be provided that the cooking program is created manually before the cooking process is carried out. It is also known that a user is enabled to select from automatic cooking programs before the cooking process is carried out.For carrying out such automatic cooking programs, the cooking devices are equipped with so-called intelligences, i.e. a multiplicity of programs or routines which can access a multiplicity of data. The data can be made available, for example, in the form of stored tables, read in from third-party devices and / or obtained from measurements by means of sensors. These data are processed by the cooking appliance in such a way that functional units, such as heat sources, steam generators, ventilation and / or venting units, water supply and / or discharge devices and / or other cooking chamber climate control devices, are controlled and / or regulated in order to achieve a desired cooking result.FIG. 1 ashows an exemplary cooking appliance in the form of an intelligent combi damper 1, in which hot air and / or steam can be applied to items to be cooked and can thus be cooked. At least one microwave source can also be provided for heating or cooking by means of microwaves. The combi damper 1 has a cooking chamber 2, which can be closed by means of a handle 4 via a cooking chamber door 3. Furthermore, an operating screen 5 with at least one display device in the form of a display 6 and at least one input device and / or interface, for example in the form of at least one central dial 7, a voice control and / or a touch panel, is provided. Furthermore, the cooking appliance can have interfaces, not shown, such as in particular digital, data interfaces, such as LAN, WLAN, network interfaces, mobile radio (4G or 5G) or the like. The cooking appliance 1 can also comprise at least one USB connection 8; at least one air filter 9; a hand shower 10 for cleaning the cooking chamber 2, and / or a drawer 11 and / or a chemical connection for cleaning chemistry. At least one heating device; at least one device for supplying moisture, comprising as a rule at least one steam generator and / or at least one steam nozzle for injecting water into the cooking chamber 2 with one or more steam stages, can be provided not visible from the outside, namely behind the operating panel 5. At least one blower wheel can be fluidically operatively connected to the cooking chamber with one or more fan stages, and at least one baffle; and a closable flap for connecting the cooking chamber 2 to the external atmosphere can be contained. Furthermore, sensors are provided, such as for measuring the temperature and moisture in the cooking chamber and the core temperature in a product to be cooked.FIG. 1b shows a home screen displayed on the display 6. Preferably, the display 6 comprises a touchscreen, so that a control of the cooking appliance can be carried out with a plurality of selection buttons in an upper row. For example, the selection keys comprise at least one key 51 for selecting a manual cooking, keys 52 to 54 for selecting automatic programs, at least one key 55 for selecting a programming mode, in particular for creating new sequences, and / or at least one home key 56.When manual cooking is selected, input parameters for cooking with hot air and / or steam can be set, and intelligent operating modes 57, favorite programs 58 and cooking methods 59 for various items for cooking can be selected for automatic cooking. In a lower row, a date / time display 60, a setting display 61, a help key 62 and a return / cancel key 63 can also be seen on the display 6 of FIG. 1 b.In order to assist the user or customer in setting the cooking appliance, for example by means of programs and / or parameters, it is possible to resort to methods which make it possible to assign users, customers and / or the special cooking appliance to a behavior pattern in order to be able to make more accurate suggestions to the user and / or customer via the human-machine interface (MMI). In this case, the data which the cooking appliance, such as the combi damper 1, collects are evaluated. An MMI can in particular comprise a user interface which displays parameters.In particular, methods are known from the prior art which are based on the assumption that behavior patterns are sufficiently known by customers and can therefore be derived from the data of the cooking appliance, for example by means of specific parameters and / or a parameter combination, which are detected by queries during the use of the appliance.Parameters can be different settings and / or the use history of a cooking appliance and / or a fleet or a group of cooking appliances. Parameters of a data set of a cooking appliance can in turn also be input parameters, cooking parameters (that is to say, for example. The information processing unit may comprise sensor measurements) of one or more settings, one or more sensor data, frequencies, cooking paths, cooking programs, one or more working steps of a cooking path and / or cooking program, recipe suggestions and / or other data of the cooking appliance. Furthermore, parameters can also comprise properties of the data set itself, such as preferably size, number of entries and / or usage statistics.These parameters can be stored, for example, in a cooking appliance by means of one or more log files. For clusters and / or behavior patterns, the parameters or settings of the customer or user of a cooking appliance in relation to the cooking programs and / or cooking paths are of interest in particular.In this case, a cooking path is the sequence of different stages in a cooking process. An example of this would be the recipe listed in Table 1 for the cooking path "Pizza", which is summarized below as follows: Table 1 Table 11Preheatinghot air20602Bakinghot air191803Bakinghot air18620An intelligent cooking path performs dynamic adjustments during the run time of the regular cooking path. This change in the values can be based on the load (for example. Weight measurement) and other influences such as the boiling point at the location of the device and / or other sensor measurements (for example. The temperature and / or humidity) of the cooking appliance may be based on the temperature and / or humidity of the cooking appliance. The intelligent cooking paths can be adjusted with the aid of input parameters. There may be different types and varying number of input parameters for each intelligent cooking path. In the example of Table 1, the input parameters that can be adjusted are the browning stage and the time input.A cooking program may include one or more cooking paths having respective input parameters. Recipes may in turn comprise one or more cooking programs.Although the methods in the prior art, as known, for example, from EP 2 101 229 B2, have proven their worth in principle, they also have several disadvantages. The mapping results from derivatives of manually generated and rigid patterns. Thus, neither new behavior patterns can be derived nor new relevant parameters and / or parameter combinations can be recognized. In addition, a derivation of the behavior pattern can only be achieved after obtaining sufficient data. Another problem is that they must be laboriously manually worked.The object of the present invention is therefore to provide an improved method which overcomes the disadvantages of the prior art.In a first aspect, this object is achieved by a computer-implemented method for detecting clusters and / or behavior patterns in cooking appliances, wherein the method comprises at least the steps:• Collecting and / or receiving first data sets from at least one cooking appliance;• Creation of at least one prediction model based on an automatic, in particular machine-learning and / or artificial intelligence-based, analysis of the first data sets.Furthermore, it can be provided in the method that the creation of the prediction model comprises at least the following steps:• At least one first, in particular automatic, analysis of the first data sets by at least one first analysis algorithm, preferably with the receipt of clusters and / or behavior patterns;• labeling the data records with at least one label in order to obtain second, labeled data records,wherein in particular the at least one label corresponds to the cluster and / or behavior pattern to which the respective data record was assigned in the first analysis;• At least one second analysis of clusters and / or behavior patterns by means of at least one second analysis algorithm in order to obtain and / or create the prediction model.Alternatively or additionally, it can be provided that the collection of first data sets for further evaluation by means of a first analysis comprises only data sets which exceed one or more threshold value(s) and / or the first data sets comprise log files, in particular application logs, AWE logs.The one or more threshold values may be defined based on at least one or any combination of the following properties of a first data set:• a log file size of the cooking appliance and / or of the customer profile;• the number of entries in a log file of the cooking appliance and / or of the customer profile;• the useful life of the cooking appliance;• Number of Used Cooking Programs and / or Cooking Paths;• Number of individually created and / or individualized cooking programs and / or cooking paths; and / or• Number of Cleanings by Means of a Hygiene Programme.Alternatively or additionally, the method can comprise: preparing the first data sets, in particular before the first analysis, wherein the prepared first data sets are preferably used for the first analysis.In this case, the processing of the first data sets, in particular before the first analysis, can comprise:• Dimensionality reduction of the first data sets, preferably by filtering the first data sets, wherein redundancies are preferably reduced and / or eliminated and / or unnecessary data is removed from the first data sets and / or not read out from the data sets;wherein preferably not required data comprise: data, in particular personal data, of users and / or customers of one or more cooking appliances, self-generated cooking programs and / or cooking paths, data of hygiene programs and / or assignment of a cooking appliance to a customer;• Extracting and / or reading one or more parameters from the first data sets for the first analysis to obtain edited data sets,wherein input parameters, in particular manual and / or intelligent cooking paths and / or cooking programs and / or individual data stored on the device are preferably extracted and / or read out from the first data sets,wherein the extracted and / or read data comprise in particular the following input parameters: time input, time indication, use of the core temperature sensor, browning stage, core temperature input, cooking chamber temperature input, frying temperature, cooking level, cooking chamber temperature stage microwave power and / or moisture stage, wherein furthermore preferably one or more parameters, in particular input parameters, in particular of the manual and / or intelligent cooking paths and / or cooking programs, are filtered with regard to the most frequently used settings of a parameter, in particular of an input parameter;• presorting and / or adapting the first data sets, preferably with respect to device generation and / or device types; and / or• Filtering and / or sorting out empty incomplete and / or damaged first data sets.Alternatively or additionally, it is proposed that the method comprises: separate and / or joint normalization and / or normalization of the first data sets and / or of the first prepared data sets, in particular before the first analysis.The normalization and / or normalization may include:• Z-score normalization and / or point normalization;• Mapping at least one score between 0 and 1 of one or more or each parameter over one or more first data sets and / or over the entire first data sets;• dimensional normalization and / or normalization within and / or over a plurality and / or over the entire first data sets;wherein a dimension comprises in particular one or more cooking paths, intelligent cooking paths, manual cooking paths, cooking programs, recipes, parameters and / or input parameters;• Standardization and / or normalization of the first data sets which are used for the at least one first analysis, based on a value between [0:1] for each column and / or for each row of a data matrix, which is preferably created from one and / or more and / or all first data sets and / or is created in particular for the at least one first analysis, wherein the normalization of a column and / or row of the data matrix is preferably carried out in such a way that the sum of the column and / or row yields "1" in each case; and / or• Discarding outliers, in particular with too great a deviation from the median and / or mean value, wherein a quantile, in particular the 90% or 95% quantile, is preferably used.Furthermore, in one embodiment, it can alternatively or additionally be provided that the at least one first analysis algorithm is an algorithm which is based on unsupervised learning, in particular is a clustering algorithm, wherein in particular the first analysis algorithm is based at least on one or a combination of the following methods:• center-based clustering methods, preferably K-means;• Density-based clustering methods, preferably DB-SCAN;• hierarchical clustering methods, preferably agglomerative clustering;• Principal Component Analysis, PCA; and / or• Independent Component Analysis, ICA.Alternatively or additionally, in one embodiment, the method may comprise the following steps:• Determination of a behavior pattern number and / or cluster number, in particular for the first analysis, and / or the prediction model, preferably based on the second analysis,wherein preferably the behavior pattern and / or cluster number is based on at least:verifying the clusters and / or behavior patterns obtained from the at least one first analysis, preferably by means of a statistical analysis, in particular a Silhouette analysis, and / or test data, preferably by means of one or more Silhouette coefficients,◯ the declared variance, wherein in particular the selected behavior pattern and / or cluster number explains at least 80% to 99% of the variance;◯ in connection with dimensionality of the first data sets used for clustering or obtaining the behavior patterns and / or used edited first data sets; and / or Experience values, in particular experience values based on additional parameters, in particular customer categories and / or customer and / or usage statisticsbestimmt wird,wherein preferably the number of 3 to 20, more preferably 5 to 15, particularly preferably 3, 5, 7, 8, 9, 10, 11 or 12 is selected as the behavior pattern and / or cluster number.It is also proposed in one embodiment that the method, preferably after labeling (400) the data records, can also alternatively or additionally comprise: multiplying the second data records, preferably while maintaining the at least one label and / or all labels of the second data records;wherein in particular an identical distribution of data records over the clusters and / or behavior patterns, the frequency of the labels and / or the frequency of one or more parameters of the data records is achieved;wherein in particular a weighted multiplication of the data sets is carried out based on at least one of the following properties dimensionality of the data set and / or occurrence and / or quantity of the adaptation of one or more parameters, in particular input parameters, of cooking paths, cooking programs and / or settings; and / orwherein in particular the multiplication is effected on the basis of a fixed multiplier or minimum multiplier.Alternatively or additionally, the method can, in particular after labeling (400) the data sets, further comprise: manipulation (500) of at least a part of the second data sets, preferably while maintaining the at least one label and / or all labels of the second data sets, in particular comprising: (i) simulation of unused cooking applications, cooking programs and / or cooking paths on a device, in particular by adapting one or more second data sets, preferably by removing and / or non-reading, in particular during and / or after the multiplication of the second data sets according to the previous embodiment, of individual or more cooking applications, cooking programs and / or cooking paths and / or the associated information and / or one or more data points and / or parameters, in particular input parameters, settings and / or other data relating to a cooking path or cooking path types, from the one or more second data sets and / or (ii) setting one or more parameters, in particular input parameters and / or settings, which reflect a cooking path and / or a cooking program on factory setting; and / or (iii) setting one or more parameters, in particular input parameters and / or settings, which reflect a cooking path and / or a cooking program on first data sets assigned by a or corresponding mean values and / or median from the totality of the first data sets and / or a cluster in relation to this cooking program and / or this cooking path.Alternatively or additionally, the method can further comprise: defining cooking devices which are assigned to a customer and / or customer account as individual customers in each case.Alternatively or additionally, in one embodiment, the second analysis and / or the creation of the prediction model can be based on a supervised-learning algorithm.The supervised learning algorithm can be based on at least one or more of the following methods: neural networks, convolutional neural network, decision tree, principal component regression, PCR, independent component regression, ICR, partial least square regression, PLSR, supported vector machine, SVM, and / or decision tree.Alternatively or additionally, the second analysis and / or the creation of the prediction model can comprise: passing through, in particular iteratively, different combinations of first and second parts of the second data sets in order to create the prediction model.Alternatively or additionally, the second analysis and / or the creation of the prediction model can comprise a leave-one-out cross-validation or another type of cross-validation.Furthermore, the method can alternatively or additionally comprise the step of: reducing the prediction model in its dimension, wherein in particular the model is reduced to one or more parameters which have a high weighting according to the prediction model, in particular have a positive and / or negative weighting, wherein preferably the parameters with the 3rd to 10th highest weighting, in particular a magnitude highest weighting, are used and / or parameters with the 3rd to 10th highest positive weighting and / or with the 3rd to 10th highest negative weighting are selected.In addition, alternatively or additionally, the prediction model can comprise one or more prediction vectors and / or matrices, wherein in particular by means of the one or more prediction vectors and / or matrices a cluster and / or behavior pattern assignment is made possible when applying, in particular multiplying, cross product and / or convolution of the prediction model to an, in particular new, first data set.It is also alternatively or additionally proposed that the prediction model comprises a look-up table.The look-up table can comprise: values for suggestions corresponding for each cluster and / or behavior pattern of the prediction model, wherein the values comprise in particular one or more parameters, wherein these are preferably determined by the statistically most frequently occurring parameter, and / or a preferably weighted mean value and / or median of the parameters, and / or a range thereof, a median and / or mean value of the range from the first and / or second data sets which are assigned to the respective behavior patterns and / or clusters according to the labels.Alternatively or additionally, the one or more parameters can be determined by a distance metric, in particular a geometric distance metric; and / or wherein the one or more parameters are not normalized.In one embodiment, the method may alternatively or additionally comprise periodically updating, preferably annually, monthly or weekly, the acquired first data sets and / or repeating the method according to any of the previously described embodiments to obtain a current prediction model.Furthermore, the method can alternatively or additionally comprise: providing the prediction model to cooking devices, preferably via an Internet-based, in particular automatic, update and / or via a manual software update; preferably together with the look-up table.Alternatively or additionally, the method can at least further comprise: receiving at least one first data set of a cooking appliance, in particular for the purpose of predicting and providing an action instruction based on the prediction model and / or the look-up table via a cloud solution.Furthermore, in a second aspect, a further computer-implemented method for intelligently modifying one or more parameters of a cooking appliance is proposed, wherein the further method comprises the following steps:• Providing and / or acquiring (1000) a data set for prediction by a prediction model• Application (2000) of the prediction model, wherein the prediction model is preferably based on a prediction model according to the method according to one of Claims 1 to 24, on a first data record of the cooking appliance,• obtaining a classification of the cooking appliance with respect to a cluster and / or behavior pattern based on the prediction model,• Modify (4000), based on the obtained classification of the cooking appliance, one or more parameters and / or a human machine interface, MMI, based on the one or more parameters.Here, modifying and / or proposing a changed setting may comprise:• changing one or more parameters, in particular input parameters, settings and / or other data of at least one cooking program, cooking path and / or recipe; and / or• Suggestion of at least one or more parameters, in particular input parameters, settings and / or other data of at least one cooking program, cooking path, and / or recipe,• wherein the suggestions and / or changes are preferably displayed to the user by the MMI, in particular by a user interface, UI, and / or are preferably limited to default settings, in particular thus parameters, input parameters, settings and / or other data which are located in the factory setting.Alternatively or additionally, the further method can further comprise: editing and / or pre-optimizing the data set, in particular before applying the prediction model:• Dimensionality reduction of the data set, preferably by filtering, of the data set, wherein redundancies are preferably reduced and / or eliminated and / or unnecessary data are removed from the data set and / or are not read out from the data setwherein preferably not required data comprise: data, in particular personal data, cooking programs and / or cooking paths created by users and / or customers of one or more cooking appliances themselves, data of hygiene programs and / or assignment of a cooking appliance to a customer;• Extraction and / or Readout of one or more parameters from the data record for the classification, in particular in order to obtain edited data records,wherein input parameters, in particular manual and / or intelligent cooking paths and / or cooking programs and / or individual data stored on the device are preferably extracted and / or read out from the first data sets,wherein the extracted and / or read data comprise in particular the following input parameters: time input, use of the core temperature sensor, browning stage, core temperature input, cooking chamber temperature input, frying temperature, cooking level, cooking chamber temperature stage, microwave power and / or moisture stage, wherein furthermore preferably one or more parameters, in particular input parameters, in particular of the manual and / or intelligent cooking paths and / or cooking programs, are filtered with respect to the most frequently used settings of an input parameter, in particular of the input parameter; and / or• Presorting and / or adapting the data set, preferably with respect to device generation and / or device types.Alternatively or additionally, the further method can further comprise: separate and / or joint normalization and / or normalization of the data set and / or the edited data set before the application of the prediction model.The normalization and / or normalization may include:• Z-score normalization and / or point normalization;• Mapping at least one score between 0 and 1 of one or more or each parameter over one or more first data sets and / or over entire first data sets; and / or• dimensional normalization and / or normalization within and / or over a plurality and / or over the entire first data sets,• wherein a dimension comprises in particular one or more cooking paths, intelligent cooking paths, manual cooking paths, cooking programs, recipes, parameters and / or input parameters;• Normalising and / or normalising the data set which are used for the at least one first classification based on a value between [0:1] for each column and / or for each row of a data matrix which is preferably produced from the data set and is in particular produced for the at least one classification, wherein the normalisation of a column and / or row of the data matrix is preferably carried out in such a way that in each case the sum of the column and / or row yields 1;• Discarding outliers, in particular with too great a deviation from the median and / or mean value, wherein a quantile, in particular the 90% or 95% quantile, is preferably used.Alternatively or additionally, the further method can further comprise:• querying at least one customer information group, wherein the customer information group preferably corresponds to at least one cluster and / or behavior pattern of the prediction model;• Determination of values for the modification of one or more parameters and / or of the UI, in particular of a display, based on the obtained classification of the cooking appliance, preferably based on the classification and / or of the customer information group, in particular by the prediction model, preferably by means of a look-up table;• receiving the prediction model;• Providing first data sets for the creation of the prediction model.In a third aspect, a cooking appliance, preferably comprising at least one processor, at least one memory and at least one human-machine interface, is proposed, wherein the cooking appliance, in particular the processor, is configured to carry out the further method described above.The cooking appliance can have at least one interface which is preferably configured to update the prediction model, preferably at regular intervals, particularly preferably weekly, monthly, or annually, and / or to search for updates of the prediction model.The foregoing aspects have several advantages. First, the use of an automatic, computer-assisted creation of a prediction model provides the surprising finding that a wide variety of clusters or behavior patterns can be detected with such a model, which clusters or behavior patterns cannot be covered by rigid behavior patterns. This not only recognizes the recognition of new patterns on the basis of the available data, but also enables a better assignment. This further leads to better proposals when using the prediction model, which leads to increased user satisfaction and also increases the desired cooking or cooking results.In particular, it is advantageous that, for example, by a first analysis of first data sets, clusters or behavior patterns of actual use trends of the operating users and / or customers can be derived from cooking devices which, insofar, cannot be derived by logical expectation of a possible user behavior. As a result, behavior patterns or clusters are not defined by predefined parameters and / or parameter combinations or by a predefined labeling grid, but rather on the basis of the actual statistically differentiable usage behavior. This has the advantage that by a corresponding reclassification, i.e. for instance a new application of the first analysis to time-updated individual data sets, corresponding trend changes and associated new clusters and / or behavior patterns can be recognized directly and automatically.Furthermore, the present method offers the advantage that, when the individual data sets are provided, they can be prefiltered by means of a threshold value check. The threshold value check is used to separate meaningful individual data sets from those which are not meaningful. In this case, a check is made as to whether a first data record exceeds one or more defined threshold values. Such threshold values can be defined, for example, on the basis of various parameters in the data sets.The advantage of expanding, i.e. multiplying and / or in, preferably subsequent, manipulating the labeled data set is that this allows corresponding usage scenarios to be included in a behavior pattern, which may occur in particular at the beginning or at the beginning, i.e. in an early usage phase of the respective cooking appliance. By providing such extended data sets, it is possible that, for example, by a second analysis, which serves to generate the prediction model, an improved assignment of devices with a short use duration and / or of devices which have just or recently been put into operation and / or of devices in which only a small number of cooking programs or cooking paths has been used is facilitated or made possible.The above embodiments may be used singly or in any combination to provide the apparatus and method according to the present invention.Other aspects and advantages of the invention will become apparent from the following description in which preferred embodiments are described with the aid of the accompanying drawings: FIG. 1 ashows an exemplary cooking appliance which can provide first data sets for the method according to the invention and on which the prediction model can be used as soon as its processor is adapted for this purpose; FIG. 1 bshows a display on a display of the cooking appliance of FIG. 1 a; FIG. 2 shows an exemplary sequence of the method according to the invention for automatic pattern recognition and for training the prediction model; and FIG. 3 shows a schematic representation of the method according to the invention when applying the created prediction model.In a first step 100, first data sets are provided for pattern recognition. The acquisition of data is possible reliably via the networking of the cooking appliances with a data sink (data sink).In this case, for example, a cloud solution can be used, such as the system of the applicant, which is referred to as connected cookieing. In this case, a user and / or customer can use, in particular, the following functions which offer important information both for the adaptation based on and for the recognition of clusters and / or behavior patterns:• Recipe management: The customer and / or user can store and make usable on all his devices a fixed set of cooking paths, as well as individualized sequences;• Device management: the customer and / or user can remotely start cooking paths and cleanings and monitor the status of the device; and / or• Hygiene Management: The customer can generate or download from the cloud the Haszard Analysis Critical Control Point (HACCP) data requested by regulatory authorities.However, a different networking of devices and / or customers is also possible. Thus, data can be acquired, for example, centrally and / or de-centrally and / or secured by means of a blockchain and / or a distributed ledger.The transmitted data sets can comprise, for example, log files of the device, which provide information about the use of the different cooking programs and / or cooking paths and the input parameters used therein, for example a cooking path and / or cooking program, in particular time input, core temperature, browning stage, cooking chamber temperature input, frying temperature, cooking degree, cooking chamber temperature stage and / or moisture stage.Thus, the data can be generated in the form of various log files. In order to be able to determine clusters and / or behavior patterns, the selected cooking paths and input parameters are relevant information in particular. These can be read out from application logs (AWE logs) or are present in them. The AWE logs document all information about sensors, actuators, user and / or customer settings and the data points created during operation of the device. For example, when selecting a cooking path, such an AWE log can be recorded. An example reduced section of a device's AWE log is shown in Table 2: Table 220,8021,25...20,8121,27...20,8121,33............Alternatively or additionally, individual cooking programs and / or cooking paths created by the user and / or the customer can also be transmitted and can be at least part of one or more first data sets. In this case, parameters, for example both individual working steps and input parameters, settings and / or other data, can be transmitted in the data record and / or the log file, in particular for one or more or each of these individual cooking programs and / or cooking paths, and can in particular be at least part of one or more first data records.Depending on the embodiment, the user or customer of a networked device can decide which data, in particular whether the device and / or the devices which are linked, for example, to a customer account, participate in the provision of data. In order to obtain first data sets for the first pattern recognition or first analysis that are as accurate and comprehensive as possible, in one embodiment data sets can be frozen at specific times and transmitted to the data sink (frozen data sets). This enables different stages of use to be obtained precisely at the beginning of the use of a new cooking appliance. This in turn facilitates or enables a recognition of behavior patterns in the early use phase of a device in the pattern recognition or in the first analysis.The transmission of such frozen data sets can take place, for example, at regular time intervals, in particular automatically, such as daily, weekly, monthly, semi-annual, annually and / or at any other desired intervals. However, a combination of different time intervals is also possible. For example, frozen data sets, for example in the case of a cooking appliance which is newly put into operation, can first be transmitted weekly and after a specific time interval, for example after 6 or 12 months, a monthly frozen data set.In the subsequent step 200, the first data records of the customers and / or of the devices are edited.Thus, the data can be reduced in dimensionality. For this purpose, information can be compressed. Redundancies in the first data sets, such as a log file, are removed. In addition, unnecessary data can be removed from the data set and / or cannot be read out from the data sets in order to reduce its dimension. Such remote data and / or data not read from log files or data sets can be, for example, the device generation and / or the device number ID. The dimension reduction can also be effected by adapting the first data sets, so that somewhat the most frequently used cooking programs and / or cooking paths can be determined over all first data sets and only the associated data are extracted. A maximum limit value can be defined in this case. This can comprise, for example, the cooking programs and / or cooking paths most frequently used.On the other hand, data can also be extracted from the first data sets in a targeted manner, for example from the log files. Thus, parameters, preferably input parameters, settings and / or other data, in particular intelligent cooking paths, and / or individual data stored on the device, for example that of the log file, can be read out. These extracted data may include, for example, input parameters: time input, use of the core temperature sensor, browning stage, core temperature input, cooking chamber temperature input, flash temperature, degree of cooking, cooking chamber temperature stage, and / or humidity stage.Alternatively or additionally, the first data sets can be normalized and / or normalized individually and / or jointly. A possible normalization and / or normalization can be, for example, a z-score normalization and / or a point normalization. Where z-score returns a number for each element of a data matrix such that the columns of the matrix are centered to have an average of 0 and scaled to have a standard deviation of 1. Alternatively or additionally, the normalization can also comprise that outliers are discarded, in particular by discarding data points with too great a deviation from the median and / or mean value. A quantile, in particular the 90% or 95% quantile, can be used here.Furthermore, such normalization and / or normalization may be carried out, for example, in such a way that a score between 0 and 1 is assigned to one or more or each parameter over an entire first data set and / or over the entire first data sets used for clustering. The normalization and / or normalization can additionally or alternatively also be effected as follows: First, all data of a first data set and / or all data sets, for example a data matrix, can be normalized in the dimensions. These dimensions can comprise, for example, the one or more cooking paths and / or intelligent cooking paths and / or cooking programs and / or recipes.For example, a correspondingly standardized data record for sections of various intelligent cooking paths can be seen in Table 3, which is based on the data record of Table 2: Table 3 Table 30,90,0500,020,030,40,20,10,10,20,10,350.050,50As can be seen from the example in Table 3, a column-by-column normalization to the range [0; 1] can be effected. Additionally or alternatively, a further adaptation can also be carried out by calculating a normalization to the range [0; 1] also via a line (i.e. in the example of table 8 device-specific), so that a row sums up to the value 1 in each case.A particular simplification can be carried out if the method according to the invention is to be used to generate a device-specific behavior pattern for each device in the fleet of a customer and based thereon a device-specific prediction model. Thus, the individual data records of an individual device can then be defined as a customer in each case and the assignment of the device to a fleet can be canceled or omitted. This is particularly useful when certain cooking devices in the fleet are used more frequently or exclusively for a specific cooking type, while other cooking devices in the fleet are used exclusively or predominantly for specific cooking types different therefrom.In subsequent step 300, the data of the first data sets are analyzed by means of at least one first analysis algorithm, in particular with respect to clusters and / or behavior patterns. The behavior patterns (clusters) obtained therefrom can then be verified statistically, for example by way of a Silhouette analysis and / or test data. A sufficient behavior pattern and / or cluster number can then be determined for the at least one first analysis. In this case, this behavior pattern and / or cluster number can be determined, for example, via a Silhouette coefficient.This first analysis algorithm may be an algorithm that uses unsupervised learning, in particular a clustering algorithm. Thus, the first analysis can be based on at least one or a combination of the following methods: center-based and / or partitioning clustering methods, in particular K-means, density-based clustering methods, in particular DB-SCAN, hierarchical clustering methods, in particular agglomerative clustering, principal component analysis (PCA),Independent component analysis (ICA), grid-based clustering methods and / or combined clustering methods.Alternatively and / or additionally, the sufficient behavior pattern and / or cluster number can be determined on the basis of the explained variance and / or in connection with the dimensionality of the data used for clustering or obtaining the behavior patterns. Thus, the number of clusters and / or behavior patterns should be, in particular, a multiple as the dimensionality of the data sets in order to prevent over-fitting.Alternatively or additionally, experience values can also be used for determining the number of clusters or behavior patterns. These empirical values can be derived, for example, from additional information and / or parameters, such as customer categories and / or other customer and / or usage statistics. Thus, for example, a number of 3 to 20, preferably 5 to 15, particularly preferably 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 or 12 behavior pattern and / or cluster numbers can be selected.Furthermore, a further first analysis can also be carried out, which is based, for example, on frozen data sets, in order to record behavior patterns which result in the course of, in particular at the beginning of, the use of a device. However, the frozen data sets can also be used in part and / or in their entirety or not used at all in the at least one first analysis.In step 400, the first data sets are labeled, wherein the first data sets may be previously unlabeled, in particular. This results in second, labeled data sets. Each label corresponds to the cluster to which the respective data record was assigned. Data records on a plurality of labels can also be obtained, even if a single label is preferred for the respective data record. Thus, data records can obtain, in particular, the labels, for example, from frozen data records of the same customer and / or device that are earlier or later. The use of a plurality of analyses is also possible, for example based on different models and / or directed at different aspects and / or different parts of the first data sets. The further labels and / or analyses can also be used on the target device to make various, in particular alternative, suggestions for the user on the basis thereof. For this purpose, Baysian decision tree can be used.Subsequently, the second data records can additionally be duplicated in a further step 500 while maintaining the one or more labels. Such multiplication can be used, for example, to adapt the data records of the training data for the subsequent second analysis for the purpose of predictive model generation, in order to obtain an identical distribution therein for the different behavior patterns or clusters of the at least one first analysis. However, a weighted multiplication of the data sets can also take place. This can be advantageous if different behavioral aspects are to be more learned. For example, parts of data sets relating to specific, for example frequently used or frequently changed cooking programs and / or intelligent cooking paths can be duplicated in a targeted manner in order to make the analysis more robust in this respect. Furthermore, the multiplication can also be carried out on the basis of a fixed multiplier or minimum multiplier.Furthermore, alternatively or additionally, data sets can also be manipulated in this step 500. This can include, among other things: the simulation of unused cooking applications, cooking programs and / or cooking paths on a device by adapting one or more second data sets. The second data sets are changed in such a way that, as a result, for individual or multiple cooking applications, cooking programs and / or cooking paths and / or the associated information or data points, such as the parameters, preferably input parameters, setting and / or other data, relating to a cooking path or cooking path types, are removed from the data set and / or are not read out from the labeled data set to be duplicated. Such manipulated data sets can depict different use cases that occur just at the beginning of the use of a new, newly installed and / or reset cooking appliance. In this way, in addition to the possible frozen data sets, a possibility is created of achieving a better assignment of a device or a customer to a cluster.Furthermore, it is also possible for one or more parameters which reflect a cooking path and / or a cooking program to be reset to the factory setting. This represents a particularly efficient possibility for simulating unused cooking programs.The second, labeled data sets processed (in particular manipulated and / or duplicated) in this way are then supplied to at least one second analysis algorithm in a further step 600. This algorithm uses the labeled individual data sets to create a prediction model. This can be made possible by means of so-called supervised learning, in which, in particular, the assignment to one of the clusters and / or behavior patterns is learned by the second analysis on the basis of the labels of the individual data sets.Possible algorithms that can be considered are neural networks, convolutional neural network, decision tree, principal component regressions (principal component regression), partial least square regression (PLSR), supported vector machine (SVM) and / or decision tree.At least a part of the data sets can be used as a training set, while an at least second part is used to predict the cluster, in order to check in particular the correct function of the prediction model. In this case, the algorithm can pass through a plurality of different combinations of first and second parts, in particular iteratively, in order to generate the prediction model. For example, a leave-one-out cross-validation or another type of cross-validation is possible.The prediction model can comprise, in particular after completion of the second analysis, one or more prediction vectors and / or matrices, by means of which a cluster and / or behavior pattern assignment is possible when applying, for example multiplying, cross product and / or convolution of the model onto a new first data set.It is also possible for the dimension of the model-i.e. the entries and / or data points from, for example, a matrix of a data set and / or a data set which are required-to be further reduced in an intermediate step (not shown), depending on whether corresponding parameters which are represented by data points in the data set receive high weighting (for example a strong score and / or weight in a prediction vector) according to the prediction model or receive low weighting (for example score or weight in a prediction vector). It is decisive here that the weighting factor is different from "0", in particular. Thus, for example, the parameters with the 3rd to 10th highest weighting, i.e. in particular a magnitude-based highest weighting (for example, can be used. Score or weight), preferably comprised in a prediction vector of the prediction model. Alternatively or additionally, parameters with the 3rd to 10th highest positive weighting and / or with the 3rd to 10th negative weighting can also be selected.After completion of the training phase, the prediction model can then comprise, for example, one or more prediction vectors and / or one or more matrices, which can respectively provide factors for the prediction for individual data points from the first data sets. These can then be applied to a new data set to classify it according to the learned model.In order to be able to make proposals for each cluster or behavior pattern of the prediction model, corresponding values must be provided for the clusters or behavior patterns of the prediction model. These values can comprise parameters, in particular in addition to input parameters for individual and / or intelligent cooking programs and / or cooking paths, also comprise entire cooking programs and / or cooking paths and / or other settings of the cooking appliance, which are proposed to the customer or user of a appliance.These values can thus be determined by determining them within the data, data points and / or data records of a cluster and / or behavior pattern by means of a standard statistics tool. For this purpose, for example, the mean value, the median, and / or the setting that is used most frequently can be used. The latter is particularly suitable for discrete values, for example the input parameters (for example the browning stages) in cooking paths and / or cooking programs.However, values or parameter ranges can also be used. Thus, for "quasi-continuous" values, such as the temperature, a discretization can be performed in the data in advance (e.g. 2, 3 and / or 5° C. intervals) and the most frequent interval within a cluster and / or behavior pattern and / or the mean value or median resulting from the limit values of the most frequent interval can be defined as a value.Alternatively or additionally, the values can also be generated using a distance metric. In this case, the first and / or second data sets which belong to a cluster and / or behavior pattern can be approximated geometrically or by other methods. From this approximation, the values for the cluster and / or behavior pattern are then determined. This approach can allow for more accurate mapping of a cluster.The values from the first and / or second data sets may not be normalized and / or normalized. This saves a subsequent conversion with respect to the data set applied.A look-up table for characterizing the individual clusters of the prediction model can then be generated from the values. The training phase is thus completed.After completion of the training phase, the created prediction model is then transmitted to the target system in a step 700 of the method according to the invention. This can be, in particular: the devices that have provided data sets and / or devices that are updated and / or populated with the latest driver software in which the prediction model is integrated. This updating can be carried out either manually or automatically, in particular either via a data carrier or a data interface, such as a network connection. Alternatively, the prediction model can also be provided individually on devices. The look-up table can be transmitted together with the prediction model or individually to a target system (for example a cooking appliance, which is in particular put into operation anew). The transmission takes place in the standard update of the device software and generally requires a suitable AI framework, a little tensorflow, MATLAB or PyTor. The update (of the model) can either be an entire device software update or relate only to one or more individual functionalities.Alternatively, this, as well as the prediction model, can be provided via a cloud solution. In such a case, individual data sets and / or the relevant parts are provided for prediction by a cooking appliance according to the prediction vector. After receipt on a server, the prediction model is then applied there and suggested changes are transmitted to the cooking appliance according to the cluster and / or behavior pattern assignment, in particular according to a look-up table. Alternatively, the look-up table may be present on the device, so that only the result of the classification by the prediction model needs to be provided from the cloud.Fig. 2 shows the application of the overprinted prediction model by means of the second method according to the invention.First, in a first step 1000, data must be collected for the application of the prediction model. In this case, a pre-optimized data record can be generated, for example, from the log file or files of the system. The pre-optimized data record preferably corresponds to the dimension of the training data records. However, all data points or parameters relevant to the prediction model should be present in the data set. In addition, the data sets are intended to pass through the same steps of normalization in order to allow efficient prediction.Subsequently, in step 2000, classification determination (in the sense of a prediction) can then be carried out by using the prediction model.In a further step (not shown), an updated prediction model can also be received before or after and during the collection of the data, for example via an update function. For this purpose, the relevant cooking appliance is connected to a network via an interface, preferably a cloud solution of the device provider, in order to correspondingly comprise the latest version of the prediction model. This can additionally alternatively or additionally also comprise an update of the suggestions, settings and / or parameters determined for a behavior pattern, in particular in the form of a look-up table. The transmission can be carried out by means of the method described in Abs. 61 are described. Training "on the edge", i.e. a further individual refinement of the model on the relevant cooking appliance, can furthermore also take place, wherein, in particular as far as possible, the preprocessing of the data takes place on the relevant appliance, such that preferably as small as possible data packets can be transferred. Alternatively, however, this can also be completely excluded in order to conserve computing capacities and / or resources.After classification has been made, a customer information group associated with the cluster can be queried in a further step 3000. In this case, this customer information group is determined on the basis of the classification by the prediction model.Thus, a customer information group can be assigned for each behavior pattern and / or cluster. This includes values, i.e. for example. Parameters, in particular input parameters, recipe suggestions, cooking paths, cooking programs, adjusted cooking parameters, settings and / or other data. These values for a customer information group can be provided, for example, via a look-up table.In this case, the look-up table can contain data representative in particular of one or more dimensions, data points, parameters, cooking paths, cooking programs and / or recipes. This representative data can be determined, for example, based on the occurrence (the most frequently used value of a cooking program, for example) in the training data of a specific cluster and / or behavior pattern. Thus, in particular, the most frequently occurring value can be used. Alternatively or additionally, this look-up table can be generated, for example, by a geometric representation of and / or approximation to the respective determined behavior pattern and / or cluster, among other things a density and / or distance metric. However, it is also possible here to additionally or alternatively employ extensible Al methods, for example a decision tree which generates the look-up table on the basis of the associated cluster.For each cluster and / or behavior pattern derived from the first analysis, corresponding values, i.e., in particular parameters, supplementary, new, modified cooking programs, cooking paths, parameters for cooking paths and / or parameters for intelligent cooking paths, can be provided in the look-up table. However, the values can also further comprise: corresponding proposal options, in particular for cooking programs, cooking paths, intelligent cooking paths and / or recipes, which are provided to the user and / or customer via an MMI, in particular a user interface (UI), preferably a tab, of the UI.For this purpose, the values for a customer information group can be determined from the training datasets, which were used, for example, in the training phase of the method according to FIG. 1, which are assigned to a cluster (for example by their label), by means of a statistics tool. In this case, for example, the mean value and / or the median can be used. Alternatively or additionally, the most widely used setting can be used. In this case, in the case of discrete values, additionally or alternatively an interval assignment (for example temperature intervals, that is to say for example, can also be carried out. 3° C. intervals, associated with their frequency in a cluster and / or behavior pattern), and thus an interval can also be proposed to the user. Alternatively or additionally, the values can also be determined by means of a, preferably geometric and / or modified, distance metric, for example by means of n-nearest neighbor. To this end, all and / or one or more parts of the data points in a cluster or behavior pattern can be used. Thus, a cluster and / or behavior pattern can be provided via a small number of meaningful representatives. This can then provide a dynamic and / or individual look-up table for a cluster or behavior pattern by means of application, for example convolution, of the previous data and / or settings of the relevant cooking appliance to the distance metricAdditionally or alternatively, the data present in the respective device, in particular individual settings already carried out, can also be taken into account and / or incorporated in the generation of the look-up table. Thus, for example, settings or parameters already individualized by the user cannot be changed or frozen in the look-up table, while other parameters are adapted and / or determined based on the cluster assignment by means of the values of the look-up table. The latter has the advantage that the customer group information based on the cluster can additionally take into account individual customer desires in the form of the look-up table.In a further step 4000, changes, for example in the default settings, are proposed to the customer and / or carried out automatically according to the customer information group assigned thereto via the behavior pattern / cluster. For example, input parameters of a cooking path, in particular of an intelligent cooking path or individual cooking path, can be used. For this purpose, the MMI can be adapted. For this purpose, the UI can propose new cooking paths, cooking programs based on the look-up table, for example. For this purpose, an additional tab "other customers frequently also used" can be provided in the UI.The features of the invention disclosed in the above description, the drawings and the claims can be essential both individually and in any combination for the realization of the invention in its various embodiments.List of reference characters1 Combi damper 2 Cooking chamber 3 Cooking chamber door 4 Handle 5 Operating panel 6 Display 7 Setting wheel 8 USB connection 9 Air filter 10 Hand shower 11 Drawer 51- 56 Buttons 57 Intelligent operating modes 58 Favorite programs 59 Cooking method 60 Date / time display 61 Setting display 62 Help button 63 Return / break button 100- 700 Steps in the method for creating the prediction model 1000- 4000 Steps in the method when applying the prediction model.References included in the specificationThis list of documents cited by the applicant has been produced in an automated manner and is only included for the better information of the reader. The list is not part of the German patent application or utility model application. The DPMA does not take any adhesion for any faults or omissions.Patent Literature citedEP 2 101 229 B2

[0015]

Claims

Computer-implemented method for detecting clusters and / or behavior patterns in cooking appliances, comprising at least the steps: • collecting and / or receiving (100) first data sets from at least one cooking appliance; • creating at least one prediction model (600) on the basis of an automatic, in particular machine-learning and / or artificial intelligence-based, analysis of the first data sets.The computer-implemented method according to claim 1, (i) wherein creating the prediction model comprises at least the following steps: • At least one first, in particular automatic, analysis (300) of the first data sets by at least one first analysis algorithm, preferably obtaining clusters and / or behavior patterns; • Labeling (400) the data sets with at least one label to obtain second labeled data sets, wherein in particular the at least one label corresponds to the cluster and / or behavior pattern to which the respective data set was assigned in the first analysis; • At least one second analysis (600) of clusters and / or behavior patterns by means of at least one second analysis algorithm to obtain and / or create the prediction model; (ii) wherein the collection (100) of first data sets for further evaluation by means of a first analysis comprises only data sets that exceed one or more threshold value(s) and / or (iii) wherein the first data sets comprise log files, in particular application logs, AWE logs.The computer-implemented method according to claim 2(ii), wherein the one or more threshold values are defined based on at least one or any combination of the following properties of a first data set: • a log file size of the cooking appliance and / or of the customer profile; • the number of entries in a log file of the cooking appliance and / or of the customer profile; • the useful life of the cooking appliance; • number of used cooking programs and / or cooking paths; • number of individually created and / or individualized cooking programs and / or cooking paths; and / or • number of cleanings by means of a hygiene program.The computer-implemented method according to any one of claims 1 to 3, wherein the method further comprises: editing (200) the first data sets, in particular before the first analysis, wherein preferably the edited first data sets are used for the first analysis.The computer-implemented method according to claim 4, wherein the processing (200) of the first data sets, in particular before the first analysis, comprises: • dimensionality reduction of the first data sets, preferably by filtering the first data sets, wherein preferably redundancies are reduced and / or eliminated and / or unnecessary data are removed from the first data sets and / or are not read out from the data sets; wherein preferably unnecessary data comprise: data, in particular personal data, of users and / or customers of one or more cooking devices, self-generated cooking programs and / or cooking paths, data of hygiene programs and / or assignment of a cooking device to a customer; • Extraction and / or reading of one or more parameters from the first data sets for the first analysis in order to obtain prepared data sets, wherein input parameters, in particular manual and / or intelligent cooking paths and / or cooking programs and / or individual data stored on the device are preferably extracted and / or read out from the first data sets, wherein the extracted and / or read data comprise in particular the following input parameters: time input, time indication, use of the core temperature sensor, browning stage, core temperature input, cooking chamber temperature input, frying temperature, cooking degree, cooking chamber temperature stage microwave power and / or moisture stage, wherein furthermore preferably one or more parameters, in particular input parameters, in particular of the manual and / or intelligent cooking paths and / or cooking programs, are filtered with regard to the most frequently used settings of a parameter, in particular of an input parameter; • presorting and / or adapting the first data sets, preferably with respect to device generation and / or device types; and / or • filtering and / or discarding empty incomplete and / or damaged first data sets.The computer-implemented method according to any one of the preceding claims, wherein the method further comprises: separate and / or joint normalization and / or normalization of the first data sets and / or the first edited data sets, in particular before the first analysis.The computer-implemented method of claim 6, wherein the normalization and / or normalization comprises: • Z-score normalization and / or point normalization; • mapping at least one score between 0 and 1 of one or more or each parameter over each one or more first data sets and / or over the entire first data sets; • dimensional normalization and / or normalization within and / or over a plurality and / or over the entire first data sets; wherein one dimension comprises in particular one or more cooking paths, smart cooking paths, manual cooking paths, cooking programs, recipes, parameters and / or input parameters; • Standardization and / or normalization of the first data sets which are used for the at least one first analysis, based on a value between [0:1] for each column and / or for each row of a data matrix, which is preferably created from one and / or more and / or all first data sets and / or is in particular created for the at least one first analysis, wherein the normalization of a column and / or row of the data matrix is preferably carried out in such a way that the sum of the column and / or row in each case yields "1"; and / or • discarding outliers, in particular with too great a deviation from the median and / or mean value, wherein recourse is preferably made to a quantile, in particular the 90% or 95% quantile.The computer-implemented method according to any one of the preceding claims, wherein the at least one first analysis algorithm is an algorithm based on unsupervised learning, in particular a clustering algorithm, wherein in particular the first analysis algorithm is based on at least one or a combination of the following methods: • center-based clustering methods, preferably K-means; • density-based clustering methods, preferably DB-SCAN; • hierarchical clustering methods, preferably agglomerative clustering; • principal component analysis, PCA; and / or • independent component analysis, ICA.The computer-implemented method according to any one of the preceding claims, wherein the method further comprises: determining a behavior pattern and / or cluster number, in particular for the first analysis, and / or the prediction model, preferably based on the second analysis, wherein preferably the behavior pattern and / or cluster number is based on at least: • verifying the clusters and / or behavior patterns obtained from the at least one first analysis, preferably by means of a statistical analysis, in particular a Silhouette analysis, and / or test data, preferably by one or more Silhouette coefficients, • the declared variance, wherein in particular the selected behavior pattern and / or cluster number accounts for at least 80% to 99% of the variance; • In connection with dimensionality, the first data sets used for clustering or obtaining the behavior patterns and / or used prepared first data sets; and / or • Empirical values, in particular empirical values, are determined based on additional parameters, in particular customer categories and / or customer and / or usage statistics, wherein preferably the number of 3 to 20, more preferably 5 to 15, particularly preferably 3, 5, 7, 8, 9, 10, 11 or 12 is selected as the behavior pattern and / or cluster number.The computer-implemented method according to any one of the preceding claims, wherein the method, preferably after labeling (400) the data sets, further comprises: multiplying the second data sets, preferably while maintaining the at least one label and / or all labels of the second data sets; wherein in particular an identical distribution of data sets over the clusters and / or behavior patterns, the frequency of the labels and / or the frequency of one or more parameters of the data sets is achieved; wherein in particular a weighted multiplication of the data sets is effected based on at least one of the following properties dimensionality of the data set and / or occurrence and / or quantity of the adaptation of one or more parameters, in particular input parameters, of cooking paths, cooking programs and / or settings; and / or wherein in particular the multiplication is effected on the basis of a fixed multiplier or minimum multiplier.The computer-implemented method according to any one of the preceding claims, wherein the method, in particular after labeling (400) the data sets, further comprises: manipulation (500) of at least a part of the second data sets, preferably while maintaining the at least one label and / or all labels of the second data sets, in particular comprising: (i) simulation of unused cooking applications, cooking programs and / or cooking paths on a device, in particular by adapting one or more second data sets, preferably by removing and / or non-reading, in particular during and / or after multiplying the second data sets according to claim 10, of individual or more cooking applications, cooking programs and / or cooking paths and / or the associated information and / or one or more data points and / or parameters, in particular input parameters, Settings and / or other data on a cooking path or cooking path types, from the one or more second data sets and / or (ii) setting one or more parameters, in particular input parameters and / or settings, which reflect a cooking path and / or a cooking program to factory setting; and / or (iii) setting one or more parameters, in particular input parameters and / or settings, which reflect a cooking path and / or a cooking program to first data sets assigned by a or corresponding mean values and / or median of the totality of the first data sets and / or a cluster with respect to this cooking program and / or this cooking path.The computer-implemented method of any preceding claim, wherein the method further comprises: defining cooking devices associated with a customer and / or customer account as individual customers, respectively.The computer-implemented method according to any one of the preceding claims, wherein the second analysis and / or the creation of the prediction model is based on a supervised learning algorithm.The computer-implemented method of claim 13, wherein the supervised learning algorithm is based on at least one or more of the following methods: neural networks, convolutional neural network, decision tree, principal component regression, PCR, independent component regression, ICR, partial least square regression, PLSR, supported vector machine, SVM, and / or decision treeThe computer-implemented method according to any one of the preceding claims, wherein the second analysis and / or the creation of the prediction model comprises: passing through, in particular iteratively, different combinations of first and second parts of the second data sets in order to create the prediction model.The computer-implemented method of any preceding claim, wherein the second analysis and / or the creation of the prediction model comprises a leave-one-out cross-validation or another type of cross-validation.The computer-implemented method according to one of the preceding claims, wherein the method comprises the step of: reducing the prediction model in its dimension, wherein in particular the model is reduced to one or more parameters which according to the prediction model have a high weighting, in particular a positive and / or negative weighting, wherein preferably the parameters with the 3rd to 10th highest weighting, in particular a magnitude-based highest weighting, are used and / or parameters with the 3rd to 10th highest positive weighting and / or with the 3rd to 10th highest negative weighting are selected.The computer-implemented method according to one of the preceding claims, wherein the prediction model comprises in particular one or more prediction vectors and / or matrices, wherein by means of the one or more prediction vectors and / or matrices a cluster and / or behavior pattern assignment is enabled when applying, in particular multiplying, cross product and / or convolution of the prediction model to an, in particular new, first data set.The computer-implemented method of any preceding claim, wherein the prediction model comprises a look-up table.The computer-implemented method according to claim 19, wherein the look-up table comprises: values for suggestions corresponding for each cluster and / or behavior pattern of the prediction model, wherein the values in particular comprise one or more parameters, wherein these are preferably determined by the statistically most frequently occurring parameter, and / or a, preferably weighted, mean value and / or median of the parameters, and / or a range thereof, a median and / or mean value of the range from the first and / or second data sets which are associated with the respective behavior patterns and / or clusters according to the labels.The computer-implemented method according to any one of the preceding claims 19 or 20, wherein the one or more parameters are determined by a distance metric, in particular a geometric distance metric; and / or wherein the one or more parameters are not normalized.The computer-implemented method according to any one of the preceding claims, further comprising periodically updating, preferably annually, monthly or weekly, the acquired first data sets and / or repeating the method according to any one of claims 1 to 21 to obtain a current prediction model.The computer-implemented method according to any one of the preceding claims, wherein the method further comprises: providing (700) the prediction model to cooking devices, preferably via an Internet-based, in particular automatic, update and / or via a manual software update; preferably together with the look-up table.The computer-implemented method according to any one of the preceding claims 1 to 23, wherein the method at least further comprises: receiving at least one first data set of a cooking appliance, in particular for the purpose of predicting and providing an action instruction based on the prediction model and / or the look-up table via a cloud solution.Computer-implemented method for intelligently modifying one or more parameters of a cooking appliance, the method comprising the following steps: • providing and / or acquiring (1000) a data set for a prediction by a prediction model • applying (2000) the prediction model, wherein the prediction model is preferably based on a prediction model according to the method according to one of claims 1 to 24, to a first data set of the cooking appliance, • obtaining a classification of the cooking appliance with respect to a cluster and / or behavior pattern based on the prediction model, • modifying (4000), based on the obtained classification of the cooking appliance, one or more parameters and / or a human machine interface, MMI, based on the one or more parameters.The computer-implemented method according to claim 25, wherein the modifying and / or proposing of a changed setting comprises: • changing one or more parameters, in particular input parameters, settings and / or other data of at least one cooking program, cooking path and / or recipe; and / or • proposing at least one or more parameters, in particular input parameters, settings and / or other data of at least one cooking program, cooking path and / or recipe, wherein preferably the proposals and / or changes are displayed to the user by the MMI, in particular by a user interface, UI, and / or are preferably limited to default settings, in particular thus parameters, input parameters, settings and / or other data which are in factory setting.The computer-implemented method according to claim 25 or 26, wherein the method further comprises: preparing and / or pre-optimizing the data set, in particular before applying the prediction model: • dimensionality reduction of the data set, preferably by filtering, of the data set, wherein redundancies are preferably reduced and / or eliminated and / or unnecessary data are removed from the data set and / or are not read out from the data set, wherein preferably unnecessary data comprise: data, in particular person-related data, cooking programs created by users and / or customers of one or more cooking devices themselves and / or cooking paths, data of hygiene programs and / or assignment of a cooking device to a customer; • Extraction and / or reading of one or more parameters from the data record for classification, in particular in order to obtain prepared data records, wherein input parameters, in particular manual and / or intelligent cooking paths and / or cooking programs and / or individual data stored on the device are preferably extracted and / or read out from the first data records, wherein the extracted and / or read data comprise in particular the following input parameters: time input, use of the core temperature sensor, browning stage, core temperature input, cooking chamber temperature input, frying temperature, cooking degree, cooking chamber temperature stage, microwave power and / or moisture stage, wherein furthermore preferably one or more parameters, in particular input parameters, in particular the manual and / or intelligent cooking paths and / or cooking programs, are filtered with regard to the most frequently used settings of an input parameter, in particular of the input parameter; and / or • presorting and / or adapting the data set, preferably with respect to device generation and / or device types.The computer-implemented method of any of the preceding claims 25 to 27, wherein the method further comprises: separately and / or jointly normalizing and / or normalizing the dataset and / or the edited dataset prior to applying the prediction model.The computer-implemented method of claim 28, wherein the normalization and / or normalization comprises: • Z-score normalization and / or point normalization; • mapping at least one score between 0 and 1 of one or more or each parameter over one or more first data sets and / or over all first data sets, respectively; and / or • dimensional normalization and / or normalization within and / or over a plurality and / or over all first data sets, wherein one dimension comprises, in particular, one or more cooking paths, smart cooking paths, manual cooking paths, cooking programs, recipes, parameters and / or input parameters; • Standardization and / or normalization of the data set which are used for the at least one first classification, based on a value between [0:1] per column and / or per row of a data matrix which is preferably produced from the data set and is in particular produced for the at least one classification, wherein the normalization of a column and / or row of the data matrix is preferably carried out in such a way that in each case the sum of the column and / or row yields 1; • Discarding outliers, in particular with too great a deviation from the median and / or mean value, wherein recourse is preferably made to a quantile, in particular the 90% or 95% quantile.The computer-implemented method according to any one of the preceding claims 25 to 29, wherein the method further comprises: • querying (3000) at least one customer information group, wherein the customer information group preferably corresponds to at least one cluster and / or behavior pattern of the prediction model; • determining values for modifying one or more parameters and / or the UI, in particular a display, based on the obtained classification of the cooking appliance, preferably based on the classification and / or the customer information group, in particular by the prediction model, preferably by means of a look-up table; • receiving the prediction model; • providing first data sets for creating the prediction model.Cooking appliance, preferably comprising at least one processor, at least one memory and at least one human-machine interface, wherein the cooking appliance, in particular the processor, is configured to carry out a method according to one of claims 25 to 30.Cooking appliance according to one of Claim 31, wherein the cooking appliance has at least one interface which is preferably designed to update the prediction model, preferably at regular intervals, particularly preferably weekly, monthly or annually, and / or to search for updates of the prediction model.

Images