System for generating a human-perceived explanatory output for an anomaly predicted by an anomaly detection module on high-frequency sensor data or derived quantities of an industrial manufacturing process, and method and computer program for monitoring an artificial intelligence-based anomaly detection during an end-of-line acoustic test of a gearbox.

Abstract

System (HF x-Al) for generating a human-perceivable (Expert) explanatory output (EA) for an anomaly (IFA3) predicted by an anomaly detection module (IF-Anom) on high-frequency sensor data (Data2) or derived quantities of an industrial manufacturing process (IF), wherein the anomaly detection module (IF-Anom) predicts the anomaly when the anomaly detection module (IF-Anom) classifies the sensor data (Data2) or areas of the sensor data describing a state (Z) of a machine (PM), component and / or product (G) of the manufacturing process (IF) as deviating from normally expected observable data, wherein the system (HF x-Al) is configured • to calculate an explanation for the predicted anomaly (IFA3) based on sensor data (IFA1) or areas of sensor data processed by the anomaly detection module (IF-Anom), a machine learning model (IFA2) that the anomaly detection module (IF-Anom) executes on the sensor data (Data2) or areas of sensor data, and the anomaly (IFA3) predicted by the machine learning model (IFA2); • to determine an explanation mask (EM) for the explanation calculation, which modifies the sensor data (IFA1) or areas of the sensor data on which the anomaly (IFA3) was predicted in such a way that the anomaly detection module (IF-Anom) classifies the sensor data masked with the explanation mask (EM) as normal; • To optimize the values ​​(W) of the explanation mask (EM) by iteratively processing the machine learning model (IFA2) on the respective masked sensor data; • to send the optimized explanation mask (EM) as the explanation output (EA) perceptible to the human (Expert), whereby the human (Expert) recognizes, based on the optimized explanation mask (EM), which sensor data (IFA1), areas of the sensor data or derived quantities are responsible for the anomaly (IFA3) predicted by the anomaly detection module (IF-Anom).

Description

[0001] The invention relates to a system for generating a human-perceived explanation for an anomaly predicted by an anomaly detection module based on high-frequency sensor data or derived parameters of an industrial manufacturing process. The invention also relates to a method and a computer program for monitoring artificial intelligence-based anomaly detection during end-of-line acoustic testing of a gearbox.

[0002] The following definitions, descriptions and explanations retain their respective meaning for the entire disclosed subject matter of the invention and apply to the entire disclosed subject matter of the invention.

[0003] DE 10 2019 135 608 A1 discloses a method, a device and a system for detecting anomalous operating conditions of a device, in particular a motor vehicle.

[0004] Industrial manufacturing processes are variable and complex, and therefore susceptible to errors and failures. These processes and / or individual steps include, for example, primary forming, forming, cutting, joining, coating, and / or material property modification processes according to DIN 8580. For instance, end-of-line testing of a gearbox, such as a vehicle gearbox or a wind turbine gearbox, encompassing functional and / or acoustic testing, is an industrial manufacturing process. Simultaneously, the integration of sensors into manufacturing processes is increasing as part of the so-called Fourth Industrial Revolution, ensuring comprehensive data-driven process mapping.

[0005] Ensuring the flawless execution of an industrial process step is associated with considerable effort and high costs, as both error detection and correction are often manual processes and are also reactive. The German application with file number 10 2021 210 107.0 discloses computer-implemented methods, modules, and a system for anomaly detection in industrial manufacturing processes. These methods enable the monitoring of industrial manufacturing processes and / or individual steps within an industrial manufacturing process to improve the detection of deviations or anomalies. The disclosed anomaly detection system is a comprehensive system whose core components are analysis models, or anomaly detection models. Based on the data, these anomaly detection models evaluate a condition and calculate an anomaly value. The anomaly value represents the degree of the anomaly; see also the German application with file number 10 2021 210 106.2.This enables the data-driven detection of anomalies in industrial manufacturing processes, particularly through ordered, numerical data, such as sonograms. The anomaly detection algorithms, anomaly detection modules, and the anomaly detection system disclosed therein are explicitly part of this disclosure.

[0006] For example, anomaly detection analysis models are based on machine learning algorithms. Logistic regression is just one example of such an analysis model. Other models include isolation trees (also called isolation forests), autoencoders, generative adversarial networks (also called generative adversarial networks), convolutional networks, and support vector machines. Isolation forests are an anomaly detection algorithm that identifies anomalies through isolation. Anomalies are isolated using binary trees. Isolation forests work well in situations where the training set contains few or no anomalies, making them a particularly advantageous analysis model.

[0007] The German application with file number 10 2021 210 106.2 discloses a computer-implemented method and a system for anomaly detection, as well as a method for anomaly detection in the final acoustic testing of a gearbox, by means of which a degree of state anomaly can be represented based on state data. The anomaly detection models and methods for calculating the anomaly value disclosed therein are explicitly included in the content of this disclosure.

[0008] A common type of data in industrial manufacturing processes is high-frequency data, meaning data acquired by sensors such as temperature sensors, vibration sensors, force sensors, pressure sensors, acoustic sensors, measuring sensors, current sensors, and / or voltage sensors at a high sampling rate. The sampling rate depends on the specific application. For example, in the acoustic field, frequencies above 20 kHz are considered high-frequency, in electrical engineering above 9 kHz, and in medical applications even from 1 kHz. What these high-frequency data have in common is that their processing differs from other data. One reason for this is the number of data points, another is the significance of frequencies in the data being analyzed. With high-frequency data, the frequency spectrum is often analyzed, for example, using Fourier transforms.

[0009] Based on this data, one of the most fundamental tasks, preferably performed by artificial intelligence algorithms, is the detection of anomalies within a manufacturing process, a production step, or during the inspection of components and / or products. To ensure this detection has the necessary robustness and inspires user confidence, the algorithmically detected anomalies must be effectively monitored and evaluated by a human, at least in part. This human oversight serves, for example, to prevent or minimize safety risks. Due to the complexity of the algorithmic decision-making process, this evaluation is difficult for a human to perform without assistance, especially when using artificial intelligence algorithms. This is all the more true the more complex the data, such as high-frequency data.

[0010] Artificial intelligence encompasses machine learning. Machine learning is a technology that teaches computers and other data processing devices to perform tasks by learning from data, rather than being programmed for those tasks. Examples of machine learning algorithms include artificial neural networks, such as convolutional networks, support vector machines, and random forest models. The general task of artificial intelligence algorithms is to unlock the potentially hidden information within data. A common task is classifying datasets into categories. This can involve using pre-existing categories; for example, classifying animal images by species.In other cases, the categories are only roughly known, for example in anomaly detection relevant to this invention, where only one category is known (normal) and the task is to recognize deviating data examples (non-normal).

[0011] The object of the invention was to find a way to generate an explainability of the outputs of an algorithm from the field of artificial intelligence, i.e., an additional output that makes it possible to understand why the algorithm made a certain classification or categorization, for a human being, in particular how an explainability output understandable to a human being can be generated that represents the decision-making of the algorithm on processed high-frequency data, especially in the classification of data sets obtained with sensors from industrial manufacturing processes into normal and non-normal data.

[0012] The subject matter of the independent / dependent claims solves this problem. Advantageous embodiments of the invention will become apparent from the definitions, the dependent claims, the drawings, and the description of preferred embodiments.

[0013] According to one aspect, the invention provides a system for generating a human-perceived explanatory output for an anomaly predicted by an anomaly detection module on high-frequency sensor data or derived quantities of an industrial manufacturing process. The anomaly detection module predicts the anomaly when it classifies the sensor data, or portions of the sensor data describing a state of a machine, component, and / or product of the manufacturing process, as deviating from normally expected observable data. The system is configured • to calculate an explanation for the predicted anomaly based on sensor data or areas of sensor data processed by the anomaly detection module, a machine learning model that the anomaly detection module executes on the sensor data or areas of sensor data, and the anomaly predicted by the machine learning model; • to determine an explanation mask for the explanation calculation that modifies the sensor data or areas of the sensor data on which the anomaly was predicted in such a way that the anomaly detection module classifies the sensor data masked with the explanation mask as normal; • To optimize the values ​​of the explanation mask by iteratively processing the machine learning model on the respective masked sensor data; • to send the optimized explanation mask as the explanation output perceptible to humans, whereby the human recognizes, based on the optimized explanation mask, which sensor data, areas of the sensor data or derived quantities are responsible for the anomaly predicted by the anomaly detection module.

[0014] The system can be abbreviated as HF-xAl. HF refers to the high-frequency sensor data. x denotes both the data itself and the fact that the artificial intelligence can be explained by it; in this latter context, "x" is short for explainable. Al stands for artificial intelligence and indicates that anomalies are predicted using artificial intelligence.

[0015] The anomaly detection module is, for example, the one disclosed in the German application with file number 10 2021 210 107.0 and shown here in Fig. 2 IF-Anom anomaly detection module described as an exemplary embodiment.

[0016] Sensor data is acquired using sensors. The term "sensor" encompasses both physical sensors and sensor models. Sensor models simulate physical sensors. These sensors can include, for example, temperature sensors, vibration sensors, force and pressure sensors, optical sensors (including camera, infrared, lidar, and radar sensors), current sensors, voltage sensors, and / or sensors for measuring the size of a component. The sensors or sensor models are technical components. Which sensors are used in a specific case may depend on the respective process step. According to one aspect of the invention, the functions of a machine, as well as the properties of a component, are monitored by means of sensors, including the aforementioned sensors. According to another aspect, the invention includes the following sensor configurations: • Component-related sensor measurement: geometry, appearance; • Machine-related measurement: Vibrations and temperatures of one or more components; • Sensor in a component test: noise, vibration, current, pressure.

[0017] According to one aspect, the sensors transmit sensor data to each other and / or to individual components / modules of the solution according to the invention and / or to a cloud infrastructure using Internet of Things technology. This enables automated or autonomous anomaly detection. The cloud infrastructure comprises a cloud-based data storage system. The cloud infrastructure can be, for example, a public, private, or hybrid cloud infrastructure.

[0018] An example of a quantity derived from sensor data is a sonogram. For instance, a time signal recorded by a sensor is converted into a frequency domain using a Fast Fourier Transform. In the sonogram, each entry represents the amplitude or energy of a frequency within a time interval; that is, it results in a matrix of numerical values ​​where each column corresponds to a time interval and each row to a frequency. The sonogram is typically a two-dimensional matrix with a time axis and a frequency axis. The entries of a sonogram are thus ordered.

[0019] In the example of a sonogram, the explanation mask is a matrix, usually a two-dimensional one. The explanation mask serves to modify the original sensor data or the original sonogram in such a way that the machine learning model of the anomaly detection module recognizes the data set in question as normal. This approach is based on the scenario of a targeted adversarial attack. The individual values ​​of the explanation mask are modified so that the machine learning model of the anomaly detection module no longer recognizes the original data set as an anomaly. The explanation mask then provides the explanation by representing the changes necessary to turn the original data from non-normal to normal.

[0020] The optimized explanation interface presents areas of high-frequency sensor data or a sonogram that are relevant for explaining anomaly detection. The system designs the presentation of the explanation output to enable, improve, or facilitate human perception of these decision-relevant areas in a specific way.

[0021] The system according to the invention can be a distributed system, for example, a cloudified system. For instance, the sensors are arranged locally at production steps of the industrial manufacturing process, and the anomaly detection module and the calculation of the explanation output are performed in the cloud. Anomaly detection is thus executed as functional software in the cloud infrastructure. In another example, the sensors and the anomaly detection module are arranged locally within the industrial manufacturing process, and the calculation of the explanation output takes place in the cloud. In this context, "based on the sensor data, the machine learning model, and the predicted anomaly" means that an evaluation module executed in the cloud receives the sensor data, the machine learning model, and the predicted anomaly from the anomaly detection module via appropriate transmit and receive units.

[0022] The modules comprise hardware and / or software modules. The hardware modules include electronic units, integrated circuits, embedded systems, microcontrollers, multiprocessor systems-on-chip, central processing units and / or hardware accelerators, for example, graphics processors, data storage units and connectivity elements, for example, WLAN modules, RFID modules, Bluetooth modules, NFC modules.

[0023] The definition of what constitutes an anomaly can, for example, be based on the definition of what is meant by a state in an industrial manufacturing process. A state describes the nature / properties of a process, a component, and / or a (production) machine at a specific point in time or within a specific time interval. This state is detected by measurements using suitable sensors, as described above. Which sensors are suitable depends on the definition of the state as well as the system to be described (component, machine, etc.). Conversely, the availability of suitable sensors also influences the definition of the state. A state that is not accessible through appropriate sensors is not a meaningful definition of a state. A state can have different characteristics in an industrial manufacturing process; the invention includes, among others, the following definitions of state: • The state of a machine can be defined by the totality of sensor data within a time interval, for example the last 10 seconds, as well as by component-specific parameters. • The condition of a component can be defined by the results of a test, as well as by other component parameters, encompassing different component variants.

[0024] Depending on the specific definition of a state, a distribution of states is determined. An anomaly is a state, specified by corresponding sensor data, that is both rare and, based on the recorded data, deviates from almost all other states, for example, a detected state characterized by high-frequency recorded vibrations. An anomaly is rare and different. Nevertheless, a precise distinction between normal and non-normal can often be difficult and, in many cases, even impossible with the available data, so that ultimately a state is only ever non-normal with a certain probability. To account for this, a human reviewer / evaluator is proposed.

[0025] The explanatory output refers to the explainability of artificial intelligence algorithms. For a specific classification of a data set into a category, such as "non-normal," the explanatory output is generated as an additional output that allows a person to understand precisely why this classification was made. The system enables the explainability of artificial intelligence algorithms when applied to high-frequency data. High-frequency data presents a particular challenge for humans because it deviates from everyday experience, as is the case with, for example, frequently encountered images and tabular data. The system helps people understand the decision-making process of an artificial intelligence algorithm based on the processing of high-frequency data.Traceability fulfills various tasks that are important in an industrial context, for example: • Justification: The explanation output provides a rationale for the classification made or the predicted anomaly. This can build trust in the use of automation based on artificial intelligence and, if necessary, help meet regulatory requirements. • Verification: Datasets used to train artificial intelligence algorithms are rarely perfect; therefore, it must be ensured that the aspects such an algorithm considers for decision-making are relevant to the problem and thus robust. This ensures that the algorithm also delivers meaningful results on examples outside the data used for initial training and testing. In particular, this allows for the development of approaches if an algorithm produces an incorrect result. • Improvement: The better a system is understood, the better further improvements can be developed. The explainability of algorithms in the field of artificial intelligence therefore plays a crucial role in their development and further refinement. • Detection: A crucial question for anomaly detection in industrial manufacturing is what causes the anomaly and, consequently, how it can be eliminated. This is known as root cause analysis. The explainability of sensor data classified as non-normal can provide crucial clues in this process.

[0026] The aforementioned aspects become even more relevant when the system is designed to operate autonomously in certain areas, which is usually the goal in industrial applications.

[0027] Following one aspect, the explanatory output is sent to the user via a human-machine interface of the system. The system thus supports the user in performing the anomaly detection check through this human-machine interface.

[0028] The system allows the anomaly detection module to be monitored by individuals. This human oversight serves to prevent or minimize security risks that arise when errors occur in the anomaly detection module.

[0029] Following a further aspect, anomalies are annotated to provide a higher level of detail. An annotation is structured into several (detail) levels, for example, three levels. The first level comprises a binary classification by the inspector, such as a false positive event. In the second level, the inspector can, for example, specify the affected component in more detail. In the third level, the inspector can, for example, provide more detailed information about their inspection / result. Following a further aspect, the annotations are reported back, for example, online in the form of a table / file.

[0030] Another aspect of the invention relates to the use of the system according to the invention during the model development of an anomaly detection module.

[0031] Based on an explanatory output from the system, a machine learning model, executed by the anomaly detection module, is trained to ensure that no false, random, or irrelevant aspects of the training data are learned. This allows for the control of datasets used to train artificial intelligence algorithms. It ensures that the aspects used for decision-making are meaningful and robust.

[0032] The system can also be used in the following deployment scenarios: • Use in prototypical operation of anomaly detection: Gaining the trust of human employees by demonstrating that the anomaly predictions are based on relevant aspects of the data; • Use during operation in random samples: Verify the results of anomaly detection and thus ensure high quality of the method and confidence in the method.

[0033] Additionally, the explanation outputs can be stored together with the anomaly detection results for later, more in-depth analyses to meet potential regulatory requirements, such as transparency requirements. Such requirements are stipulated, for example, in the so-called EU Act on the regulation of artificial intelligence.

[0034] A method can also be provided for monitoring anomaly detection based on artificial intelligence in high-frequency sensor data or derived parameters of an industrial manufacturing process. The method comprises the following steps: • Receiving sensor data processed by an anomaly detection module, on which a machine learning model executed by the anomaly detection module has predicted an anomaly, receiving the machine learning model and receiving the predicted anomaly; • Generating an explanation output for the predicted anomaly; • Sending the explanation output to a human reviewer; • Human evaluation of the predicted anomaly into a correctly or incorrectly predicted anomaly based on the explanation output; • Sending the assessment to a system, the anomaly detection module and / or to a relevant production step of the industrial manufacturing process via a human-machine interface.

[0035] According to one aspect, the explanation output is generated using a system according to the invention.

[0036] According to another aspect, the invention provides a method for monitoring an anomaly detection based on artificial intelligence during end-of-line acoustic testing of a gearbox. The method comprises the following steps: • Driving the gearbox with a speed profile, measuring the resulting structure-borne sound using acoustic sensors and obtaining structure-borne sound time series; • Transforming the structure-borne sound time series into a frequency space and obtaining sonograms; • Classifying sonograms into normal and non-normal classes based on predefined limits for specific frequency ranges; • Processing the sonograms through an artificial neural network trained to predict anomalies in the sonograms; • If the artificial neural network predicts an anomaly on a sonogram classified as normal, generate an explanation output for the predicted anomaly; • Based on the explanation output, display to a human evaluator the areas of the sonogram that are responsible for the predicted anomaly.

[0037] According to one aspect, the explanation output is generated using a system according to the invention.

[0038] The methods according to the invention can be computer-implemented.

[0039] A computer program can also be provided for monitoring an AI-based anomaly detection in high-frequency sensor data or derived variables of an industrial manufacturing process. The computer program includes instructions that cause a computer to execute the steps of the procedure for monitoring AI-based anomaly detection in high-frequency sensor data or derived variables of an industrial manufacturing process when the computer program is running.

[0040] According to another aspect, the invention provides a computer program for monitoring an AI-based anomaly detection during end-of-line acoustic testing of a gearbox. The computer program includes commands that cause a computer to execute the steps of the method for monitoring AI-based anomaly detection during end-of-line acoustic testing of a gearbox when the computer program is running on the computer.

[0041] The instructions of computer programs comprise machine instructions, source code, or object code written in assembly language, an object-oriented programming language such as C++, or in a procedural programming language such as C. In one respect, computer programs are hardware-independent application programs that are delivered, for example, via data carriers or data carrier signals using Software Over The Air (SOA) technology.

[0042] In another aspect, the system is configured to present the optimized explanation mask—a combination of initial sensor data or derived quantities and the optimized explanation mask, or a side-by-side display of the initial sensor data or derived quantities and the optimized explanation mask—in a way that is visually perceptible to humans. A combination, for example, involves adding the optimized explanation mask to the initial / original sensor data or multiplying them accordingly. An explanation mask with a lower dimension than the initial sensor data can also be combined. This prepares the explanation output for visual perception. The side-by-side display of the initial sensor data or derived quantities and the optimized explanation output enables the evaluator to better identify and assess the relevant points in the initial sensor data for the decision.

[0043] In one respect, the system is configured to generate the explanation output as an audible signal. In this case, the human evaluator is able to hear and assess differences in noise.

[0044] In addition to the visual or audio output of the explanation, the system is also configured to generate a report. For example, a report accompanying a sonogram-based explanation describes in text form at which frequencies and times the amplitude / excitation was too high and by what percentage.

[0045] According to another aspect, the system is configured to optimize the values ​​of the explanation mask under the boundary conditions that • the change in initial sensor data or derived quantities is minimal; • the change in the initial sensor data or derived quantities is focused on localized areas of the sensor data; • favors changes in contiguous areas of sensor data or derived quantities over point-by-point sensor data; and / or • the changed values ​​are within the range of expected observable values ​​or physically possible values.

[0046] The boundary conditions or restrictions generate added value with regard to explainability. For example, favoring related areas improves human interpretability. These boundary conditions can be incorporated into the optimization process, for instance, in the form of regularization terms. The precise design and weighting of the various restrictions can be configured by the system user according to one aspect. This allows the user to select different configurations for calculating the explanation output, thereby generating an explanation tailored to the specific use case through an iterative process. Alternatively, the user can choose from a range of standard configurations that often yield good results. Once configured, the selected configuration is then applied by default to all subsequently detected anomalies.It is also possible to save multiple configurations, so that the applied configuration is chosen depending on, for example, metadata of the example to be explained.

[0047] Another aspect of the system is its ability to perturb initial sensor data, derived quantities, and / or the explanation mask with random values. For example, the sonogram modified by the explanation mask is perturbed with random values ​​in each optimization run. This is achieved, for instance, by adding a matrix of the same dimension as an original sonogram, consisting of random values, to the sonogram and / or explanation mask. Such perturbation results in improved explanation masks because it prevents the explanation mask from settling on only seemingly optimal values / local minima. It has been recognized that this improves the explanation output. Perturbation with random values ​​can influence the explanation result. The system starts with parameters that yield good results in many cases. Furthermore, users of the system iteratively refine the parameters to suit specific applications.

[0048] Another aspect is that the explanation mask has the same dimension as the initial sensor data or derived quantities, and the system is configured to iteratively add the explanation mask to the initial sensor data or derived quantities and / or iteratively multiply the explanation mask by the initial sensor data or derived quantities. For example, if a sonogram is derived from the initial sensor data in the form of a two-dimensional matrix, then the explanation mask is congruent with the sonogram. Addition and multiplication can lead to different explanation results. This allows the system user to test different configurations.

[0049] Another aspect is that the initial explanation mask has a lower dimension than the initial sensor data or derived quantities, and the system is configured to determine an explanation mask with the same dimension as the initial sensor data or derived quantities using upsampling. For example, a sonogram consists of a very large number of individual values. This results in numerous possibilities for influencing an explanation mask intended to reverse the classification result of the machine learning model of the anomaly detection module. However, to generate an explanation mask that provides added value in terms of explainability, the resulting structure should be as simple and regular as possible. To this end, degrees of freedom are reduced in the explanation mask by starting with an explanation mask of a lower dimension than the sonogram to be explained.Only through deterministic upsampling is an explanatory mask generated that has the same dimension as the original sonogram.

[0050] According to another aspect, the system includes • a data input module that receives the sensor data or ranges of sensor data processed by the anomaly detection module, the machine learning model that the anomaly detection module executes on the sensor data or ranges of sensor data, and the anomaly predicted by the machine learning model; • an evaluation module that receives the data and machine learning model obtained from the data input module and generates the explanation output based on this, wherein the evaluation module includes a first transmission module that sends the explanation output to the human; • a storage module comprising a receiving module that receives a human-made evaluation of the predicted anomaly as a correctly or incorrectly predicted anomaly based on the explanation output, wherein the storage module receives and stores the data and machine learning model obtained from the data input module, the explanation output generated by the evaluation module, and the evaluation received from the receiving module; • a second transmitting module that sends the human-made assessment of the predicted anomaly into a correctly or incorrectly predicted anomaly to the anomaly detection module to improve the machine learning model.

[0051] The data input module can be a data interface between the anomaly detection module and the system. In a cloudified implementation, where the system runs in a cloud, the data interface is, for example, a wireless interface. Because the system receives the machine learning model from the anomaly detection module via the data input module, it can iteratively optimize the explanation mask locally, for example, in the cloud, or using the system's evaluation module. During optimization, no data exchange with the anomaly detection module is necessary, thus saving computing time.

[0052] According to another aspect, the data input module is configured to check the validity of the data received from the anomaly detection module. For example, if the data type for the sensor data is specified as floating-point numbers, the data input module checks whether the received sensor data is in floating-point form. Metadata may be in a different data format. Based on this check, according to another aspect, the data is corrected at this point as far as possible, or a message is sent to an operator describing the problem, and further processing is paused until the required correction is made.

[0053] The evaluation module corresponds to an explanation generator and can include hardware accelerators configured to optimize the explanation form. For example, the evaluation module can include one or more graphics processors. Depending on the aspect, the evaluation module comprises several submodules, for example, four submodules: a first submodule for explanation calculation, a second submodule for similarity calculation, a third submodule for explanation conversion, and a fourth submodule for processing the explanation output. The first submodule optimizes the explanation form.

[0054] The second submodule compares currently assessed anomalies with those already assessed via data exchange with the storage module. The third submodule can convert the explanatory output into a corresponding data space to further support the evaluator in their assessment. The fourth submodule prepares the explanatory output for visual or auditory perception.

[0055] In one respect, the storage module is a cloud storage device.

[0056] According to another aspect, the system is configured to compare, when explaining a currently predicted anomaly, sensor data or ranges of sensor data for this current anomaly with sensor data of predicted anomalies stored in a memory module of the system; to compare an explanation output for this current anomaly with explanation outputs of predicted anomalies stored in the memory module; and / or to compare a combination of the sensor data and the explanation output, each for this current anomaly, with corresponding combinations of predicted anomalies stored in the memory module. The system is configured to determine the comparisons by executing an algorithm in which distance measures are implemented. The comparison can thus be performed on three levels.This ensures that structural similarities are found both between basic measurements and between defect patterns, and finally also between actually similar products / components that have resulted in different measurements due to errors during production.

[0057] For example, the comparison or algorithm is performed by the second submodule for a similarity calculation. The algorithm can perform the comparison based on a simple distance measure, such as Euclidean distance. Following an aspect, the algorithm performs a cosine distance calculation on AutoEncoder embeddings.

[0058] The cosine distance, also called cosine similarity, is a measure of the similarity between two vectors. It is calculated by determining the cosine of the angle between the two vectors. Cosine similarity is advantageous for comparing data, particularly regarding semantic relationships. In one respect, the algorithm is the Latent Semantic Analysis algorithm disclosed in US 4,839 / 853, which can compare a search query with a file or document using cosine similarity.

[0059] To place a suspected anomaly in the context of previously reviewed potential anomalies, the system displays the three most similar previously reviewed anomalies by default, based on each aspect. This ensures that reviewers can work efficiently and that a certain degree of consistency is achieved in the assessments. If the previously assessed anomalies are displayed, the reviewer can also subsequently revise their assessment if new information has come to light. For example, findings about components where an anomaly was detected by the system and which were subsequently examined more closely in the laboratory to locate the fault can be recorded.

[0060] In another respect, the system is configured to display initial sensor data in a different data space and to generate the explanation output in this data space. For example, the display in a different data space is achieved using the third submodule for explanation conversion. This provides an important source of information for evaluating the explanation and thus the potential anomaly. If the original signal on which the anomaly detection was based is, for example, a vibration / time signal, such as an acoustic signal, then the calculated explanation will initially also be of this type. The system provides a reviewer with the option of also examining this signal in the frequency domain, for example, using a Fourier transform. This allows a reviewer to visually observe and evaluate what is essentially an acoustic signal.Conversely, it is also possible to convert a signal that was originally present in the frequency domain into an acoustic signal, and thus listen to the abnormal noise.

[0061] If underdetermination occurs when converting an original signal to a different data space, the system must implement strategies to address this. Firstly, the system may, based on experience, assume default values ​​that work well for a given problem. These could be a single value for an open parameter or a selection of values, allowing the evaluator to then choose the most suitable result. Secondly, this creates a potential link to similarity calculations. If parameter underdetermination occurs during explanation conversion, the free parameters can be populated with values ​​that have yielded good results with similar explanations and original data.

[0062] Another aspect is that the system is configured to initiate a higher-frequency sampling of sensor data, ranges of sensor data, or derived variables responsible for the anomaly predicted by the anomaly detection module. This further improves data acquisition. For example, the system could issue a recommendation, perhaps in the form of a metadata file, that the most relevant areas should be sampled at a higher frequency.

[0063] The invention is illustrated in the following exemplary embodiments. They show: Fig. 1 a schematic representation of an exemplary embodiment of an industrial manufacturing process, Fig. 2 a first embodiment of an anomaly detection module, Fig. 3 an embodiment of a system according to the invention, Fig. 4a a detailed representation of the system from Fig. 3, Fig. 4b an embodiment of a method according to the invention for monitoring an anomaly detection based on artificial intelligence in high-frequency sensor data or derived quantities of an industrial manufacturing process, Fig. 5a an example of initial sensor data in the form of a sonogram, Fig. 5b Example values ​​of the sonogram from Fig. 5a, Fig. 5c an example of an initial explanation mask, Fig. 6a the sonogram from Fig. 5a, Fig. 6b an example of an optimized explanation mask, Fig. 6c an example of a visually perceptible explanatory output, Fig. 7a an example of a sonogram, Fig. 7b an explanation output for the sonogram generated according to the invention Fig. 7a and Fig. 8 an embodiment of a method according to the invention for monitoring an anomaly detection based on artificial intelligence during an end-of-line acoustic testing of a gearbox.

[0064] In the figures, identical reference symbols denote identical or functionally similar reference parts. For clarity, only the relevant reference parts are highlighted in each figure.

[0065] Industrial manufacturing is generally structured into production lines, which are further subdivided into individual, sequential production steps (PS). A production line has a specific production capacity, for example, 100 products per day. Production can be scaled up beyond this capacity by using multiple identical production lines, each manufacturing the same product. In addition to the main line, there are secondary lines where individual components of the main product are manufactured. A single production step (PS) can have different characteristics; for example, it might involve screwing a component together, milling a part, or performing a quality inspection. Fig. Figure 1 shows an industrial manufacturing process (IF) comprising one main line and two secondary lines. The main line includes production steps PS1, PS2, and PS3. One of the secondary lines also includes three production steps (PS1, PS2, and PS3) and relates, for example, to production step PS2 of the main line. Another secondary line also includes three production steps (PS1, PS2, and PS3) and relates, for example, to production step PS3 of the main line. An end-of-line (EOL) inspection is performed at the end of the main line.

[0066] The anomaly detection module IF-Anom monitors, for example, the production steps PS, PS1, PS2, PS3 and / or the respective associated production machines PM with regard to occurring anomalies, see Fig. 2.

[0067] The anomaly detection module IF-Anom integrates the following steps or modules: Data-based definition of a state Z of the process to be monitored, data storage and processing of all necessary data, training of an anomaly detection system based on historical data Data1, integration of a human reviewer Expert to check detected / predicted anomalies and to generate further annotations for targeted expansion of the training data set Data1, detection / prediction of anomalies, processing of the evaluation of the human reviewer Expert and, based on this, continuous adaptation and improvement of the anomaly detection system.

[0068] The basic functionalities, such as saving and expanding the data set, setting up configurations of trained models (Config-Model), and requesting annotations to expand the data basis, are handled by a first module, IF-Anom Core.

[0069] The task of detecting / predicting potential anomalies is performed by a second module, the IF-Anom Anomaly Detector. Data and / or test results corresponding to the defined state Z are continuously sent to IF-Anom from the production step PS or production machine PM being monitored. IF-Anom's response then classifies state Z as "normal / not normal," optionally with a confidence estimate, and is sent to a human expert for review via a third module, the IF-Anom Annotation Manager.

[0070] The extent to which IF-Anom's suggestions are actually reviewed by a human expert depends both on the use case and on how far IF-Anom's training has progressed, i.e., how high the expected accuracy of IF-Anom's assessments is.

[0071] As an additional optional fourth module, IF-Anom Data, the anomaly detection module IF-Anom offers the possibility of integrating further data sources into a production step (PS). This could, for example, be an additional vibration sensor coupled to a specific excitation. The data generated in this way can be transmitted to the anomaly detection module IF-Anom during operation and, once a sufficient amount of data has been generated, serves as an additional data source for model training.

[0072] Model training can be carried out, for example, using the following method, which is implemented using a computer: • Obtaining data Data1, Data2 of states Z of a process, a component and / or a production machine PM in at least one process step PS, PS1, PS2, PS3 to be monitored of an industrial manufacturing process IF at a specified time or in a specified time interval, comprising sensorially measured data from similar process steps and / or from downstream processes and / or data obtained from simulating the process; • Determining a state definition based on the data Data1, Data2; • Saving the data and the state definition; • Training at least one machine learning model ISA2 for anomaly detection, wherein the data Data1, Data2 are fed into the machine learning model ISA2 and the machine learning model ISA2 determines a distribution of the states Z based on the state definition and, based on the distribution, classifies the states Z that are rare based on the data Data1, Data2 and / or deviate from other states Z as anomalies; • Evaluating the trained machine learning model, with an expert review instance annotating true positive results and / or false positive results among the classified anomalies; • Extending the data Data1, Data2 with the annotations and saving the extended data; • Saving the trained machine learning model ISA2 and its evaluation.

[0073] The in Fig. Scenario 1, as depicted, could, for example, involve the case of a production machine PM breaking down. In the known state of the art, repairs are only carried out in such a case after the production machine PM has failed, thus incurring costs due to production downtime. If the production machine PM does not fail immediately, defective components can still be produced undetected, processed further, and then lead to problems in downstream production steps PS, PS1, PS2, PS3, or in the products themselves.

[0074] In Fig. 2. The integration of sensors S into the industrial manufacturing process IF and / or inspections, along with the corresponding availability of data, enables the intelligent anomaly detection module IF-Anom and / or its respective modules IF-Anom Core, IF-Anom Anomaly Detector, and IF-Anom Annotation Manager to use artificial intelligence Al to capture the state Z of a production step PS, PS1, PS2, PS3, or an inspection in a data-driven manner, and subsequently detect and report abnormal behavior. This allows for, firstly, an early response to emerging changes, ideally before a process comes to a standstill or a defective component continues through the process chain, and secondly, better identification of the underlying defect.

[0075] In Fig. 2. The anomaly detection module IF-Anom monitors, for example, a production step PS or a production machine PM. This could be, for instance, a lathe machining a component or a person screwing a component together. The application process of the anomaly detection module IF-Anom is, for example, as follows: Once the process step to be monitored has been defined, the relevant data set, Data 2, is reviewed. Based on this, a data-driven definition of the state Z to be monitored is established. Then, the initial training data, Data 1, for the anomaly detection module IF-Anom is selected.

[0076] The primary training data source, Data1, typically consists of historical data from the process in question and upstream processes, insofar as these are relevant for assessing state Z. This could include, for example, geometric information from the components involved. Additionally, if sufficient data is unavailable for the process being monitored, data from comparable processes can be used as a secondary training data source for Data1. In some cases, synthetic, simulated data can be used as a third training data source. This involves artificially generating / simulating data, for example, via a physical model of the process in question.

[0077] Regarding the second and third data sources of the training data Data1, transfer learning can be used, for example. The processing of the data Data1 and Data2, as well as the model training, is handled by the first module, IF-Anom Core.

[0078] During operation of the anomaly detection module IF-Anom, data (Data2) from the monitored process is continuously sent to the second module, IF-Anom Anomaly Detector. The second module, IF-Anom Anomaly Detector, receives the current models from the first module, IF-Anom Core, as well as a use-case-specific configuration (Config-Model) that specifies, for example, the necessary preprocessing of the incoming data (Data2). Using these models, the second module, IF-Anom Anomaly Detector, then performs an assessment of the current state (Z).

[0079] This assessment is passed on to the third module, the IF-Anom Annotation Manager. The third module, IF-Anom Annotation Manager, is responsible for deciding whether and when a state Z is submitted to a human expert for evaluation, and if so, to which expert. Generally speaking, detected anomalies are almost always passed on, and examples identified as normal are also passed on under certain circumstances. The assessments by the human expert(s) are then sent back to the third module, IF-Anom Annotation Manager, which decides in what form the annotations are sent back to the first module, IF-Anom Core, to expand the database accordingly.Ideally, the third module, IF-Anom Annotation Manager, has access to meta-information regarding the assessment, such as an anonymized identification of the assessing expert (particularly necessary for data protection reasons) and a timestamp indicating when the assessment was written. Details of the third module, IF-Anom Annotation Manager, are described below.

[0080] To build an expanded database, the data sent to the second module, IF-Anom Anomaly Detector, is forwarded by it to the first module, IF-Anom Core, for processing and data storage. The first module, IF-Anom Core, handles the regular retraining or re-training of the ISA2 machine learning models. For this purpose, the first module, IF-Anom Core, can also send annotation requests to the third module, IF-Anom Annotation Manager.

[0081] If a predicted anomaly is confirmed as such by the reviewer expert, or if the detected anomaly is released without review, two reactions are triggered. First, the monitored production step PS is examined more closely based on the identified abnormal condition Z to determine whether and when maintenance / repair is necessary. Second, the components processed during and after the anomaly occurred can be proactively removed and / or inspected separately.

[0082] An exemplary application of the fourth module IF-Anom data for additional data generation in this scenario is a specially designed test program that generates machine states that do not occur under standard production conditions, but contain a high degree of information about the health of the components of the production machine PM being monitored.

[0083] For example, high-frequency time signals are often transformed into the frequency domain using Fast Fourier Transform or Wavelet Transform. With time signals in general, it may be necessary to synchronize them before processing to ensure comparability. Image data represents a fundamentally different data source with its own specific processing steps. These steps are highly dependent on the specific application, but the basic principles of the IF-Anom anomaly detection module described here remain unchanged.

[0084] First, the time signal is converted into a sonogram (Sona) using a Fast Fourier Transform. Within this sonogram (Sona), regions can be assigned to specific transmission components and defined by limit values ​​that must not be exceeded. Based on these regions, known as features, and their associated limit values, the state Z of a transmission (G) is determined to be OK if no limit value is exceeded, and not OK if at least one limit value is exceeded.

[0085] In a sonogram (Sona), each entry corresponds to the amplitude or energy of a frequency within a time interval. This results in a matrix of numerical values ​​(W) where each column represents a time interval and each row a frequency. The entries of a sonogram (Sona) are thus ordered.

[0086] Defining the characteristics and their associated limits is a complex process, resulting both from the design of the gearbox and from experience gained during operation. Consequently, there are always some gearboxes with unusual characteristics that go undetected. For this reason, the established gearbox acoustic testing is enhanced by anomaly detection using the IF-Anom anomaly detection module.

[0087] A model that uses an autoencoder to represent the states Z typically observed in the training data. An autoencoder is a model with a two-part neural network architecture. One part of the model learns to generate a low-dimensional representation of a state, for example, a sonogram Sona, and the second part of the model learns to reconstruct the original state from this representation. The error made in this process is larger for rarely occurring states than for the majority of frequently occurring states Z. This allows the model to be used for anomaly detection.

[0088] Fig. Figure 3 shows the interaction of the anomaly detection module IF-Anom with the HF x-Al system according to the invention. During operation of HF x-Al, data is exchanged between the anomaly detection module IF-Anom and HF x-Al. The data relevant for anomaly detection, namely the high-frequency sampled sensor data Data2 from a process step PS or from a product inspection, for example, vibration measurements of a machine or product component, as well as metadata relevant for anomaly detection, such as material number and machine parameters, are transmitted from the respective process step PS to the anomaly detection module IF-Anom via a correspondingly implemented interface. Based on this data, IF-Anom generates a classification into classes. At a minimum, the classes normal / non-normal are considered, but a higher level of detail, for example, classification into different types of anomalies, is also possible.

[0089] The HF x-Al system according to the invention fulfills the task of making the prediction process comprehensible. To this end, the anomaly detection module IF-Anom sends the data it processes, IFA1, the Al model used, ISA2, and the result of the classification, for example, a predicted anomaly, ISA3, to HF x-Al. HF x-Al then calculates an explanation output, EA, which makes the result of the anomaly detection comprehensible to a human expert. The explanation output EA is often visual in nature, but is not limited to this. In addition to the result of the anomaly detection, the explanation output is then sent to a human reviewer, Expert, who validates the result of the anomaly detection in order to initiate appropriate steps.

[0090] If a detected anomaly is confirmed, production can be stopped, potentially defective parts can be rejected, and the underlying error can be rectified based on a root cause analysis shown in process step A3. If the detected anomaly is classified as a false prediction, the results of the evaluation can be incorporated into an improvement of the machine learning model ISA3 in process step A1. In process step A2, the evaluation contributes to an improvement in data acquisition. For example, the evaluation can enable data to be sampled at a higher frequency. In process step A4, the trust of human employees in the HF x-Al system and / or the anomaly detection module IF-Anom is built, for example, in a prototypical operation of the anomaly detection system, by demonstrating that the anomaly predictions are based on relevant aspects of the data.

[0091] Fig. Figure 4a shows the HF x-Al system. The data ISA1, ISA2, ISA3 sent from the anomaly detection module IF-Anom to the HF x-Al system first enter a data input module, Data. There, the validity of the data ISA1, ISA2, ISA3 is checked, for example, with regard to data type. Based on this check, the data is corrected at this point where possible, or a response is triggered to an operator describing the problem, and further processing is paused until the required correction is made.

[0092] The data ISA1, ISA2, and ISA3 are then sent to a storage module Mem and to the explanation generator Gen, the central instance for generating the explanation output EA. The explanation generator Gen consists of, for example, four submodules Sub1-Sub4, which are briefly described below and in more detail in the following sections. One of the modules, Sub1, is used to calculate the explanation; another, Sub4, is used to generate the explanation output EA, which is processable by a human expert; and two further modules, Sub2 and Sub3, are used to extend the calculated explanation. The calculated explanation output EA is sent to the storage module Mem and added there to the input data ISA1, ISA2, and ISA3.

[0093] The processed explanation output (EA) is sent to a human evaluator, Expert. Expert possesses the necessary expertise to ensure a high-quality classification of the potential anomaly. Expert uses the explanation output (EA) to verify whether the classification by the anomaly detection module (IF-Anom) was correct. To ensure a high-quality explanation, Expert can manually refine the explanation calculation iteratively by adjusting various parameters (see process step B2). The final classification of the potential anomaly is entered into a human-machine interface (HMI) and sent to the memory module (Mem) in process step B3, resulting in a complete data set regarding the anomaly and its evaluation.The parameters used for the final explanation are stored by the explanation generator Gen and used as a standard for future explanations. Based on the expert's assessment, measures concerning the production process are implemented in the event of a confirmed anomaly.

[0094] In process step B4, the assessment of the evaluator Expert, in particular its classification, is fed back to the AI ​​models to improve them. In process step B5, the classification of the evaluator Expert into corresponding processes of the industrial manufacturing process IF is fed back to improve the handling of anomalies. In process step B1, data is exchanged between the explanation generator Gen and the storage module Mem, for example, for a similarity calculation based on a comparison of a currently predicted anomaly IFA3 with historically predicted anomalies IFA3 stored in the storage module Mem and their respective associated data.

[0095] In the Fig. In the procedure step C1 shown in 4b, the anomaly detection module IF-Anom is used. Fig. 4a. Processed sensor data IFA1, on which the machine learning model IFA2, executed by the anomaly detection module IF-Anom, predicted an anomaly, is received. The machine learning model IFA2 and the predicted anomaly IFA3 are then obtained. In process step C2, the explanation output EA for the predicted anomaly IFA3 is generated. In process step C3, the explanation output EA is sent to the human evaluator Expert. In process step C4, the predicted anomaly IFA3 is humanly evaluated as either correctly or incorrectly predicted, based on the explanation output EA. In process step C5, the evaluation is sent to the system HF x-Al, the anomaly detection module IF-Anom, and / or to a relevant production step PS of the industrial manufacturing process IF via a human-machine interface (HMI).

[0096] Fig. 5a- Fig. Figure 5c shows an example of calculating the explanation output EA. The data IFA1 processed by the anomaly detection module IF-Anom is located in Fig. 5a in the form of a sonogram Sona. Fig. Figure 5b shows in a simplified representation that the sonogram Sona is a matrix of time / frequency values ​​W. Fig. 5c shows an explanatory mask EM, which has the same dimension as the sonogram Sona. Fig. 5a, Fig. 5b and which is initialized with zeros, for example.

[0097] Fig. 6b shows an optimized explanation mask EM, which explains the in Fig. The non-normal sonogram Sona shown in 6a is formed. The sonogram Sona of Fig. 6a was classified as non-normal by the anomaly detection module IF-Anom. Based on the initialization, see for example... Fig. 5c, the evaluation module Gen of the HF x-Al system iteratively modifies the explanation mask EM such that the explanation mask EM transforms an initially non-normal sonogram Sona into a normal sonogram Sona. In the example shown, the focus is primarily on the upper region of the sonogram Sona. Fig. 6a Values ​​W will be reduced. Fig. Figure 6c shows the explanation output EA, namely the visual representation of the optimized explanation mask of the Fig. 6b.

[0098] Fig. Figure 7a shows an example of a sonogram Sona classified as non-normal by the anomaly detection module If-Anom during the end-of-line (EOL) acoustic testing of a gearbox G. Fig. Figure 7b shows the explanation output EA calculated by the HF x-Al system. Dark areas correspond to negative values ​​in the explanation mask EM. Light areas correspond to positive values ​​in the explanation mask EM.

[0099] The flawless condition of a finished gearbox G is verified as part of the end-of-line (EOL) test. This test includes, for example, a functional and / or acoustic test. During the acoustic test, a speed ramp is applied in various gears, and the resulting structure-borne noise is measured at the gearbox housing. This corresponds to process step D1 in Fig. 8.

[0100] The established assessment of the gearbox condition is then carried out by converting the acceleration / time signal into the frequency domain using a Fast Fourier Transform, see procedure step D2 in Fig. 8, and subsequent application of limit values ​​to specific time / frequency ranges. The result of this test is then a classification as OK or NOT. This corresponds to procedure step D3 in Fig. 8. Since the established transmission noise test evaluation cannot detect all defective transmissions, the following is done in the Fig. In procedure step D4, shown in Figure 8, an AI-based anomaly detection is performed on the gearbox tests classified as OK. For anomaly detection, algorithms based on Convolutional Neural Networks (CNNs) are used, for example, the results of which are reviewed by IFA3 Expert evaluators.

[0101] The CNN operates on the sonograms (Sona) of the gearbox noise tests and results in a classification of the sonograms (Sona) into the classes normal / non-normal. If a gearbox (G) tested as OK is classified as non-normal by the CNN, this classification can be reviewed by a human evaluator (Expert). To ensure this review is targeted, HF x-Al is used to show the reviewer (Expert) areas of the sonogram (Sona) that were relevant for the classification as anomaly. This corresponds to the [missing information]. Fig. 8 shown process step D5.

[0102] In this use case, the explanation mask EM is added to the original sonogram Sona and adjusted under the following boundary conditions so that the sum of the original sonogram Sona and the explanation mask EM is no longer classified as non-normal by the CNN used. The optimization is performed, for example, using the gradient descent method. The boundary conditions used are: • The sum of the absolute values ​​of all values ​​W in the explanation mask EM should be as small as possible. This ensures that the explanation mask handles changes sparingly and focuses specifically on the truly relevant areas of anomaly classification. • The changes between adjacent values ​​W in the explanation mask EM should be as small as possible. This ensures that the explanation mask EM favors related areas and thus produces a representation that is easier for a human expert to interpret. • The changes induced by the explanation mask EM are limited. This means that the sum of the original sonogram Sona and the explanation mask EM, at the level of each individual value W, must not exceed or fall below certain values. This prevents the solution algorithm from generating explanation masks EM that lead to unrealistic sonograms Sona. The goal here is for the explanation mask EM to remove / detect the non-normal structures in the sonogram Sona without generating overly unrealistic sonograms Sona. • Further terms of this type in the optimization may be useful and must be designed depending on the exact application scenario.

[0103] In the Fig. In the procedure step D6 shown in step 8, the areas of the sonogram Sona that are responsible for the predicted anomaly IFA3 are then displayed to a human evaluator Expert based on the explanation output EA. Reference sign IF industrial manufacturing process PS1-PS3 Production Steps PS Production step PM production machine G gearbox EOL End Of Line Test S Sensor HF x-Al System EA Explanatory Edition EM Explanation Mask HMI Human-Machine Interface IF-Anom anomaly detection module IF-Anom Core first module IF-Anom anomaly detector second module IF-Anom Annotation Manager third module IF-Anom data fourth module Expert Supervisor / Human Evaluator IFA1 sensor data processed by the anomaly detection module IFA2 Machine Learning Model IFA3 predicted anomaly Data1 first data Data2 high-frequency sensor data Data3 third data Sona sonogram Data Data Entry Module Gene Evaluation Module / Explanation Generator Sub1 first submodule for explanation calculation Sub2 second submodule for similarity calculation Sub3 third submodule for explanation conversion Sub4 fourth submodule for processing the explanation output Memory module Z State Config Model Configuration of Trained Models Artificial Intelligence CNN Convolutional Neural Network A1-A4 Procedure steps B1-B4 Procedure steps C1-C5 process steps D1-D6 process steps

Claims

[1] System (HF x-Al) for generating a human (Expert) perceptible explanation output (EA) for an anomaly (IFA3) predicted by an anomaly detection module (IF-Anom) on high-frequency sensor data (Data2) or derived quantities of an industrial manufacturing process (IF), wherein the anomaly detection module (IF-Anom) predicts the anomaly when the anomaly detection module (IF-Anom) classifies the sensor data (Data2) or areas of the sensor data describing a state (Z) of a machine (PM), component and / or product (G) of the manufacturing process (IF) as deviating from normally expected observable data, wherein the system (HF x-Al) is configured, • to calculate an explanation for the predicted anomaly (IFA3) based on sensor data (IFA1) or areas of sensor data processed by the anomaly detection module (IF-Anom), a machine learning model (IFA2) that the anomaly detection module (IF-Anom) executes on the sensor data (Data2) or areas of sensor data, and the anomaly (IFA3) predicted by the machine learning model (IFA2); • to determine an explanation mask (EM) for the explanation calculation, which modifies the sensor data (IFA1) or areas of the sensor data on which the anomaly (IFA3) was predicted in such a way that the anomaly detection module (IF-Anom) classifies the sensor data masked with the explanation mask (EM) as normal; • To optimize the values ​​(W) of the explanation mask (EM) by iteratively processing the machine learning model (IFA2) on the respective masked sensor data; • to send the optimized explanation mask (EM) as the explanation output (EA) perceptible to the human (Expert), whereby the human (Expert) recognizes, based on the optimized explanation mask (EM), which sensor data (IFA1), areas of the sensor data or derived quantities are responsible for the anomaly (IFA3) predicted by the anomaly detection module (IF-Anom). [2] System (HF x-Al) according to claim 1, configured to display the optimized explanation mask (EM), a combination of initial sensor data (IFA1) or derived quantities and the optimized explanation mask (EM), or a side-by-side display of the initial sensor data (IFA1) or derived quantities and the optimized explanation mask (EM) in a way that is visually perceptible to humans (Experts). [3] System (HF x-Al) according to one of the preceding claims, configured to optimize the values ​​(W) of the explanation mask (EM) under the boundary conditions such that • the change in initial sensor data (IFA1) or derived quantities is minimal; • the change in the initial sensor data (IFA1) or derived quantities is focused on localized areas of the sensor data; • favors changes in contiguous areas of sensor data or derived quantities over point-by-point sensor data; and / or • the changed values ​​(W) are within the range of expected observable values ​​or physically possible values. [4] System (HF x-Al) according to claim 3, configured to disrupt initial sensor data (IFA1) or derived quantities and / or the explanation mask (EM) with random values. [5] System (HF x-Al) according to one of the preceding claims, wherein the explanation mask (EM) has the same dimension as initial sensor data (IFA1)) or derived quantities and the system (HF x-Al) is configured to iteratively add the explanation mask (EM) to initial sensor data (IFA1) or derived quantities and / or to iteratively multiply the explanation mask (EM) with initial sensor data (IFA1) or derived quantities. [6] System (HF x-Al) according to any one of claims 1 to 4, wherein the explanation mask (EM) initially has a lower dimension than initial sensor data (IFA1) or quantities derived therefrom and the system (HF x-Al) is configured to determine an explanation mask (EM) by means of upsampling which has the same dimension as initial sensor data (IFA1) or quantities derived therefrom. [7] System (HF x-Al) according to any one of the preceding claims, comprising • a data input module (Data) that receives the sensor data (IFA1) or areas of the sensor data processed by the anomaly detection module (IF-Anom), the machine learning model (IFA2) that executes the anomaly detection module (IF-Anom) on the sensor data (Data2) or areas of the sensor data, and the anomaly (IFA3) predicted by the machine learning model (IFA2); • an evaluation module (Gen) that receives the data (IFA1, IFA3) and the machine learning model (IFA2) obtained from the data input module (Data) and generates the explanation output (EA) based on this, wherein the evaluation module (Gen) includes a first transmission module that sends the explanation output (EA) to the human (Expert); • a memory module (Mem) comprising a receiver module that receives an evaluation of the predicted anomaly (IFA3) as a correctly or incorrectly predicted anomaly (IFA3) performed by the human (Expert) based on the explanation output (EA), wherein the memory module (Data) receives and stores the data (IFA1, IFA3) and the machine learning model (IFA2) obtained from the data input module (Data), the explanation output (EA) generated by the evaluation module (Gen), and the evaluation received from the receiver module; • a second transmitting module that sends the human (Expert) assessment of the predicted anomaly (IFA3) into a correctly or incorrectly predicted anomaly (IFA3) to the anomaly detection module (IF-Anom) to improve the machine learning model (IFA2). [8] System (HF x-Al) according to any of the preceding claims, configured to compare sensor data or ranges of sensor data for a currently predicted anomaly (IFA3) with sensor data of predicted anomalies (IFA3) stored in a memory module (Mem) of the system (HF x-Al) when explaining a currently predicted anomaly (IFA3), to compare an explanation output (EA) for this current anomaly (IFA3) with explanation outputs (EA) of predicted anomalies (IFA3) stored in the memory module (Mem) and / or to compare a combination of the sensor data and the explanation output (EA) for this current anomaly (IFA3) with corresponding combinations of predicted anomalies (IFA3) stored in the memory module (Mem), wherein the system (HF x-Al) is configured to determine the comparisons by executing an algorithm in which distance measures are implemented. [9] System (HF x-Al) according to one of the preceding claims, configured to display initial sensor data (IFA1) in another data space and to generate the explanation output (EA) in that data space. [10] System (HF x-Al) according to any of the preceding claims, configured to initiate a sampling at an increased frequency for the sensor data, ranges of the sensor data or derived quantities responsible for the anomaly (IFA3) predicted by the anomaly detection module. [11] Use of a system (HF x-Al) according to one of the preceding claims during a model development of an anomaly detection module (IF-Anom), wherein, based on an explanation output (EA) of the system (HF x-Al) in a training process of a machine learning model (IFA2) executed by the anomaly detection module (IF-Anom), it is ensured that no false, random or irrelevant aspects of training data are learned. [12] Method for monitoring an artificial intelligence (AI)-based anomaly detection during an end-of-line (EOL) acoustic test of a gearbox (G), the method comprising the steps: • Driving the gearbox (G) with a speed profile, measuring the resulting structure-borne sound using acoustic sensors (S) and obtaining structure-borne sound time series (D1); • Transforming the structure-borne sound time series into a frequency space and obtaining sonograms (Sona) (D2); • Classifying sonograms (Sona) into normal and non-normal classes (D3) based on predefined limits for specific frequency ranges; • Processing the sonograms (Sona) by an artificial neural network (AI, ISA2) trained to predict anomalies on the sonograms (Sona) (D4); • if the artificial neural network (AI, ISA2) predicts an anomaly on a sonogram (Sona) classified as normal, generate an explanation output (EA) for the predicted anomaly (ISA3) (D5); • Based on the Explanatory Output (EA), display to a human evaluator (Expert) the areas of the sonogram (Sona) that are responsible for the predicted anomaly (IFA3) (D6). [13] Computer program for monitoring an artificial intelligence (AI)-based anomaly detection during an end-of-line (EOL) acoustic test of a gearbox (G), comprising instructions that cause a computer to perform the steps of the method of claim 12 when the computer program is running on the computer.

Images