Cleaning system, computer program and data carrier signal

EP3950159C0Active Publication Date: 2026-10-07FRAUNHOFER GESELLSCHAFT ZUR FORDERUNG DER ANGEWANDTEN FORSCHUNG EV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
EP2021190229
Authority / Receiving Office
EP · EP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2020-08-07
Filing Date
2021-08-06
Publication Date
2026-10-07
Estimated Expiration
2041-08-06

AI Technical Summary

Technical Problem

Existing automated cleaning systems for complex systems lack sensors for detecting local contamination and are inflexible, leading to inefficient cleaning processes that waste resources and fail to adapt to varying contamination levels, often resulting in incomplete cleaning and contamination.

Method used

A cleaning system with a targeted jet cleaner and integrated sensor device, utilizing a rotating head and telescopic arm, combined with a control device and machine learning, to autonomously detect and adapt cleaning processes to specific contamination levels.

Benefits of technology

The system efficiently reduces cleaning time and resource consumption by intelligently directing cleaning agents only where needed, ensuring thorough and optimized cleaning based on real-time contamination detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure IMGF0001
    Figure IMGF0001
  • Figure IMGF0002
    Figure IMGF0002
  • Figure IMGF0003
    Figure IMGF0003
Patent Text Reader
Need to check novelty before this filing date? Find Prior Art

Description

[0001] The invention relates to a cleaning system comprising a targeted jet cleaner, which in turn includes at least one nozzle for dispensing cleaning agents in the form of at least one jet, wherein the nozzle is arranged on a rotating head and connected thereto to a manipulation device that can move the nozzle so that the jet is directed in all directions, wherein the rotating head is further arranged at a first end of a telescopic arm. The invention further relates to a computer program and a data carrier signal.

[0002] Cleaning-in-place (CIP) systems are used for automated cleaning to create reproducible and verifiable cleaning processes. However, the automated cleaning of technically complex systems presents a significant challenge. Often, a large portion of the systems is difficult or impossible to inspect, and there is no way to automatically monitor the cleaning success. Only sensors exist for indirect monitoring of cleaning success or the degree of soiling based on the dirt load in the cleaning agent, for example, using turbidity sensors.

[0003] Currently available tank cleaning technologies also prevent the flexible consideration of product-specific cleaning needs, as they employ a fixed, predefined spray pattern. Due to a lack of sensors and flexible cleaning systems, the automated cleaning of machines and equipment currently occurs regardless of the local degree of contamination. Therefore, these systems are designed for worst-case scenarios (e.g., localized drying on the fill level ring) with additional safety margins and are generally significantly oversized. As a result, large quantities of drinking water, cleaning chemicals, energy, and production time are wasted.

[0004] Sensors for detecting local contamination and adaptive cleaning systems are typically lacking. Studies on multi-storey storage tanks have shown that the majority of tanks are considered clean after only 10-40% of the cleaning time. This means that approximately 60-90% of the cleaning time is spent on the unnecessary and therefore inefficient removal of local residual contamination (e.g., dried-on deposits and connection fittings). The typically high proportion of cleaning time (15-20% of production time), combined with the industry's characteristically low margins of 1-2%, would result in significant economic and environmental benefits if the cleaning process were optimized.

[0005] From German patent application DE 10 2018 114 476 A1, a cleaning system is known that comprises a supply robot and cleaning satellites, as well as a cleaning method. The surface cleaning system is designed for automatic and autonomous operation, with a control system linking the supply robot and the cleaning satellite for cooperative operation. The supply robot can hold an auxiliary substance for continuous or discontinuous dispensing from an auxiliary substance reservoir to the cleaning satellite. The supply robot can autonomously retrieve the auxiliary substance and energy from a service station and return to the cleaning task location on the surface to transfer the auxiliary substance to the cleaning satellite.Furthermore, the detection of existing contamination is possible if a contamination sensor device, which may include multiple contamination sensors, is intended for use with the supply robot or the cleaning satellite. The supply robot or the cleaning satellite can examine surfaces for contamination, at least at problem areas such as on equipment and pipelines, under equipment, and in other hard-to-reach areas, and in particular determine the type of contamination (e.g., grease or starch residues), its condition (e.g., dried, spoiled), the quantity or thickness of the dirt layer, and its other properties.

[0006] German patent application DE 10 2018 114 476 B3 describes a cleaning system that includes a supply robot to provide cleaning fluid to other system components. This is not necessary for a stationary targeted jet cleaner, which is already permanently connected to a supply line and thus receives the cleaning fluid directly. Specifically, a cleaning drone is proposed that can approach the contamination directly and from any direction. Furthermore, position data and the degree of contamination are used. However, no control device is provided, nor is a computer program designed to run such a device included.

[0007] Document WO 2020 / 139839 A1 describes a programmable, hydraulically controlled tank cleaning system with a nozzle that has one or more degrees of freedom of movement. This system can be operated as a standalone unit and is designed for cleaning tanks containing hazardous liquids or waste, such as oil, chemicals, or radioactive material. It provides a safe and efficient method for dissolving and removing deposits in tanks that are otherwise difficult to access, using high-pressure jets of liquid. The system is suitable for use in potentially explosive atmospheres. The described tank cleaning system uses a programmable, hydraulically controlled nozzle that is manually controlled by an operator via a control station.The human-machine interface (HMI) allows the operator to control the nozzle movements in real time, define cleaning paths, and adjust parameters such as pressure and flow rate. There is no autonomous or automatic control; all movements and cleaning processes are directly monitored and manually controlled by the operator. Thus, a cleaning system according to the preamble of claim 1 is disclosed.

[0008] Further cleaning systems are known from publications KR 101467341 B1 and US 2010 0 050 366 A1 and describe a supply robot with satellites and an autonomous vacuum cleaner system, respectively. Controllable jet cleaners with a rotatable rotary tube and a jet device arranged in a swiveling mechanism are known from publications DE 10 2017 115 837 A1 and EP 3 517 226 A3.

[0009] These and other known automated cleaning systems, however, do not take into account varying levels of contamination in process plants or can only analyze them inadequately. Due to unsuitable control systems and the lack of available sensor and actuator technologies, cleaning is therefore carried out with excessively high resource expenditure (time, energy, cleaning agents) based on worst-case scenarios. Despite this systematic over-engineering, serious food contamination repeatedly occurs due to critical points or irregularities in the preceding production process. Furthermore, cleaning control is only performed on a random sample basis and, in closed systems such as tanks, even only indirectly.

[0010] Studies have shown that in dairy processing plants, cleaning times can be reduced by an average of over 50% by appropriately interrupting the cleaning process when the cleaning path is "clean," and by a further 50% with intelligent parameterization of the cleaning lanes. This translates to realistically achievable savings of an average of 75% in cleaning time. Simple rule-based controls are not effective due to the wide variety of soil types and conditions (composition, moisture content, drying, etc.) and potential cleaning parameters (pressure, temperature, chemical composition of cleaning agents, lane profile, etc.).

[0011] The object of the present invention is therefore to offer an intelligent cleaning system including associated hardware and software components. The cleaning system should react completely autonomously to widely varying levels of soiling.

[0012] The problem is solved by a cleaning system according to claim 1, comprising a targeted jet cleaner which in turn includes at least one nozzle for dispensing cleaning agent in the form of at least one jet. The nozzle is arranged on a rotating head and connected to a manipulation device that can move the nozzle so that the jet is directed in all directions. The rotating head is arranged at a first end of a telescopic arm. When using the cleaning system according to the invention, the first end of the telescopic arm projects into the area to be cleaned, e.g., a tank. The telescopic arm can be rigid, with a fixed length, as a movable lance, or in multiple sections. Various embodiments are explained in more detail below.In any case, the telescopic arm bridges the distance between its first end, where the rotary head is located, and its second end, into which the necessary operating fluids, such as the cleaning agent, as well as energy for operating the rotary head and control signals for its operation, are fed, and sensor signals, as explained in more detail below, are picked up. The distance between the first end, which regularly immerses itself in the area to be cleaned during operation, and the second end outside this area is bridged by the telescopic arm.

[0013] According to the invention, the cleaning system comprises at least one control device for condition-based control of the cleaning process and at least one sensor device, wherein the sensor device comprises a contamination sensor and a light source. The contamination sensor, in particular designed as an optical sensor, for example a camera, scans the environment for contaminated areas.

[0014] Preferably, the function of this optical sensor is based on a luminescence effect. For example, a light spectrum is emitted that, in the presence of the expected contamination (e.g., bacterial growth), produces an easily detectable luminescence effect. Alternatively, it is also possible to apply a substance to the surface to be inspected that, under the intended illumination, interacts with the contamination to be removed and produces a luminescence effect. Processing images taken under different light spectra has also proven advantageous. In this way, only the areas that are actually contaminated can be cleaned as needed. Furthermore, the cleaning process can be monitored and the cleaning success verified.

[0015] It is therefore particularly advantageous if the device for detecting adhesions is designed as an optical detector utilizing a luminescence effect of the adhesion caused by light radiation emitted from the cleaning device or in the application area. The wavelength can, for example, lie outside the visible spectrum and include at least part of the UV or IR spectrum.

[0016] In an alternative embodiment, the device for detecting adhesions is designed in such a way that NIR or MIR spectroscopy can be performed using the reflection method.

[0017] In each embodiment, the wavelength emitted by the illumination device is matched to the type of adhesion in such a way that a luminescence effect of the adhesion is utilized. For example, certain bacterial deposits can exhibit a luminescence effect under UV light and are therefore easily detectable with an optical detector provided by the invention. Most importantly, the resulting image is readily evaluable and the result can be incorporated into a control sequence.

[0018] The sensor device is connected to the control device for transmitting sensor data. During cleaning operation, one or more computer programs run within the control device, specifically two programs: one that controls the cleaning process and another that performs reinforcement learning, a form of machine learning, to further optimize the cleaning process. For this purpose, the control device has sufficiently powerful computer components, at least a processor and main memory.

[0019] The computer program(s) running in the control unit, and especially the machine learning used to optimize the cleaning process (as described in more detail below), can also be outsourced to a separate server with the necessary or greater computing capacity. In this case, the control unit only needs to perform simple control tasks and collect data from the contamination sensor to send to the server. According to the present definition, execution on the separate server is also considered execution within the control unit.

[0020] The invention thus proposes a cleaning system capable of performing tank cleaning in a situation-specific and therefore needs-based manner. A key component of the system is a freely configurable and, at least when preferably inserted into a tank from the top, vertically movable target jet cleaner, in combination with an integrated sensor device, hereinafter also referred to as contamination sensor, and self-learning software, which is explained in more detail in the computer-implemented aspect of the invention.

[0021] The cleaning system comprises a targeted jet cleaner which, through two independently controllable drive axes, is capable of individually reaching any point in the tank and thus realizing any desired cleaning path. The targeted jet cleaner is combined with an integrated sensor unit for dirt detection. The drive technology of the targeted jet cleaner (also referred to as an adaptive jet cleaner) forms the basis for the cleaning system according to the invention. Furthermore, the system according to the invention is capable of moving freely within the tank (especially vertically) and, by means of the integrated lighting and camera system, detecting and selectively cleaning contaminants. It is also provided for the targeted jet cleaner to be inserted into the tank through a lateral opening, in which case movement is horizontal or otherwise directional rather than vertical.

[0022] According to a further advantageous embodiment of the telescopic arm, it is designed as a long, undivided spray lance which can be moved down or up in the tank by means of an external sliding guide, a sliding frame running therein which is connected to the telescopic arm, and the drive device.

[0023] According to a first embodiment, at least one sensor device is arranged in a movable sensor arm. Alternatively or additionally, according to a second embodiment, at least one sensor device is arranged in the rotating head, wherein the rotating head has a housing that is at least partially transparent (at least for the selected radiation spectrum). Alternatively, or furthermore, the at least one sensor device is arranged in the telescopic arm, spaced apart from the rotating head. In addition, the components of the sensor device, namely the contamination sensor and the light source, can be positioned differently, analogous to the aforementioned embodiments.

[0024] A light source and / or a contamination sensor that incorporates active and / or passive cooling is particularly advantageous. This minimizes the installation space required for the components integrated into the rotating head (e.g., camera, light source). Furthermore, it ensures low weight and protection against the aggressive environmental conditions found in a tank being cleaned. Advantageously, resistance to temperatures exceeding 120 °C and to cleaning agents is achieved.

[0025] Furthermore, the contamination sensor system must be integrated, and its electronic power supply and connection must be ensured. Since the contamination sensor system, preferably a camera, is integrated into the lower part of the cleaner, all its power and signal lines must be safely routed through the system. Vertical movements originating from the z-axis must also be compensated for, unless the undivided spray lance is used. The sensor system is located in or on the rotating head. It is preferably housed in a nozzle arm, allowing it to be aligned as needed.

[0026] This also involves the use of innovative sensor components, the miniaturization of which presented particular challenges. These include heat dissipation from the components through a small housing surface area, a sufficiently high UV illumination intensity to enable rapid detection of contamination, and high light sensitivity of the camera sensor despite its small size. According to a preferred embodiment, the intended size is approximately 175 cm³, in particular a cuboid measuring 100 x 50 x 35 mm.

[0027] Components housed in a casing require significantly more space, partly due to the necessary cooling, while conventionally offering temperature resistance up to 85°C. In contrast, the invention provides resistance to temperatures exceeding 120°C and to cleaning agents. This is achieved primarily through active cooling using a fluid (water or air). Furthermore, cooling is supported by a Peltier element or a heat pipe. These measures ensure that even with increased permissible operating temperatures, the installed components achieve a longer service life or can be operated at even higher ambient temperatures.

[0028] The cleaning system according to the invention thus comprises an innovative lighting system that meets the requirements for compactness and heat resistance while simultaneously providing very high luminous intensity. This is preferably achieved by using a single LED, in particular a high-power LED, instead of many smaller LEDs distributed over a larger area. It has been shown that this also requires active cooling of the lighting. For this, the development of a suitable passive or active cooling strategy as a highly integrable solution was necessary, among other things. Several different cooling strategies for air and water cooling have proven suitable: Air cooling: Air, including compressed air, is passed over a heat sink connected to the lighting and camera. The air can either be routed back through the shaft of the target cleaner or directly into the tank. Water cooling is possible as a closed or open circuit, with water being routed into the tank. The water can be routed either through a hose in a closed loop or directly into the tank. The water is also routed over a heat sink.

[0029] In addition to the constructive measures, switching-technically suitable strategies for e.g. pulse-pause operation are also preferably provided, which are integrated into the intelligent process control via a qualified interface.

[0030] Furthermore, the preferred light source is designed as a switchable hybrid solution that enables the emission of different light spectra. The particularly preferred light source comprises at least one LED or high-power LED. The light source is switchable for at least UV light (preferably ≈ 365 nm), daylight, or white light. Other light colors are also provided if they offer high contrast to the substrate or the expected contamination. The color tone of the camera image provides information about the type of contamination and should be recognizable with the highest possible contrast to the surroundings. Alternatively or additionally to contrast, it is also provided that the surface, the surroundings, and / or the contamination are excited to emit light (luminescence). In a first embodiment, a single-chip LED is to be used to achieve the desired compactness.

[0031] If a UV bandpass filter is used before the UV illumination, preventing white light from passing through, a second embodiment incorporates a UV LED and a separate white light LED. In both cases, the hybrid function with UV and white light provides an ideal combination for geometry and contamination detection.

[0032] Another aspect of the present invention is the distance-dependent focal length of the illumination, implemented in a particularly compact manner. This is achieved by using an optical component, preferably a compact zoom lens. This enables automatic adjustment of the focal length within a small installation space. The focal length and focus can be adjusted by electrical signals.

[0033] This is particularly necessary when the targeted jet cleaner generates a changing distance to the observed surface for the sensor system with every change in angle. The main challenge for the illumination here is achieving the optimal illumination diameter. At the beginning of the cleaning process, with a high degree of soiling, the soiling is easy to detect. Here, a large area can be illuminated or observed with low light intensity per unit area. Towards the end of the cleaning process, however, when optical detection limits become increasingly important due to the decreasing layer thickness, sufficiently high detection accuracy requires more focused illumination or excitation over a smaller area.

[0034] The preferred light source therefore comprises a light control device which, by means of controllable optical components, in particular a controllable optic with a variable focal length (zoom), e.g., implemented as a motorized zoom, enables a variable optimal illumination diameter. At the beginning of the cleaning process, when the degree of soiling is high and the soiling is still easily detectable, a shorter focal length illuminates a large area with correspondingly low light intensity per unit area. Conversely, towards the end of the cleaning process, a longer focal length with strong focus illuminates a small area. Advantageously, the light control device is linked to or integrated into the control device. Alternatively, low detection limits can be compensated for by increasing the exposure time.

[0035] The contamination sensor is preferably designed as a camera device that includes controllable components and can be robustly operated on a multi-axis mechanism. The particularly advantageous controllable optical components comprise a zoom lens or autofocus zoom lens, preferably a lens with freely adjustable focus. Liquid Lenses, executed. This applies equally to the focusing of the lighting, however Liquid LensesThe camera image is converted into a digital representation of the contamination level in an evaluation unit. This evaluation unit can also be integrated into the control unit. The control of the optical component or zoom lens can also be handled by the control unit, which decides, based on the situation, which area should be captured and evaluated by the camera. Criteria for this decision can lie within the cleaning process itself, in the area illuminated according to the method described above, or in the computing capacity for processing the resulting data. These factors are considered both in the control of the cleaning process and in the machine learning.

[0036] It has proven advantageous if the sensor device includes a device for removing deposits and / or condensate, which cleans the openings of the light source and / or the contamination sensor.

[0037] To achieve a rough positioning of the target jet cleaner in the area to be cleaned, e.g. a tank, the telescopic arm changes its length extending into the interior of the tank and thus specifically positions the rotating head.

[0038] Another advantageous aspect of the invention relates to the completely encapsulated fluid flow, meaning that the cleaning fluid does not flow through mechanical transmission parts and does not come into contact with electronic components. Furthermore, the traversing movements of the z-axis are compensated, for example, by a telescopic piping or flexible hose, unless a one-piece spray lance is used. A cleaning agent line is provided that supplies the cleaning agent to the nozzle without coming into contact with moving parts of the rotating head, thus forming a multi-channel system. Likewise, a supply cable, a control line, and / or a coolant channel are routed through the interior of the telescopic arm and rotating head.

[0039] The invention further comprises a cleaning control device that includes or stores software modules for execution, which are designed or programmed for image evaluation of the images supplied by the optical contamination sensor. The optical contamination sensor consists of a camera and a light source. The evaluation primarily concerns the algorithms for contamination, geometry, and object recognition, and their mutual correlation. These are necessary as information sources for the self-learning control system. During the evaluation, the system will detect the location and amount of contamination within the tank to be cleaned and select spatially resolved, efficient parameter sets for cleaning.

[0040] Geometry and object recognition support the integration of expert knowledge. For example, if the system detects a process nozzle or other internal components, it can automatically suggest cleaning this geometry with an efficient spiral path.

[0041] Another effective alternative for individual critical areas are meandering paths, which can already be programmed manually into the cleaner.

[0042] To implement the cyber-physical system—a system in which information and software technology are linked with mechanical components, and where data transfer, exchange, and control occur in real time via an infrastructure—a number of data interfaces must be defined and provided. This includes connecting the data interfaces of sensors and actuators (e.g., contamination sensors, temperature and pressure of the cleaning media, and the angular position of the drive motors for positioning the target jet cleaner). Furthermore, additional interfaces for data transfer are provided, such as for importing CAD models of the tank geometry and for exporting data and files that need to be stored for documentation purposes. Data connectivity to higher-level systems within the production process is also implemented.For internal data processing, a connection to a database will be set up.

[0043] A user interface will be designed for subsequent interaction between the user and the cleaning system. This will include, for example, the manual adjustment of various cleaning process parameters (such as cleaning path parameters), the visualization of the manually set or automatically calculated cleaning path in a 3D model, and the display of documentation information. The aspect of calculating, displaying, and creating an interaction option for the user regarding adaptive cleaning paths in conjunction with the flexible cleaning system is particularly novel and is not yet addressed in any way by the current state of the art. Special attention will be paid to the traceability of the software system's decisions during interface design. The user interface will preferably be designed according to Ecological Interface Design (EID) principles.

[0044] EID is an interface design approach specifically developed for complex socio-technical, real-time, and dynamic systems. It has been applied in a wide range of fields, including process control (e.g., nuclear power plants, petrochemical plants), aviation, and medicine. EID differs from some interface design methods, such as User-Centered Design (UCD), in that its analysis focuses on the work domain or environment, rather than the end user or a specific task.

[0045] The goal of EID is to make constraints and complex relationships in the work environment perceptible to the user (e.g., visible, audible). This allows users to allocate more of their cognitive resources to higher cognitive processes such as problem-solving and decision-making. EID is based on two key concepts from cognitive engineering research: the abstraction hierarchy (AH) and the skills, rules, knowledge (SRK) framework. By reducing mental workload and supporting knowledge-based reasoning, EID aims to improve user performance and overall system reliability for both expected and unexpected events in a complex system.

[0046] The object of the invention is also achieved by a computer program for state-based control of a cleaning process according to claim 10, for implementation in a computer control of a cleaning system as described above, comprising the steps: Transfer of optimized parameters to the control system or the control device; output of varied control instructions from the control process to the target jet cleaner; inclusion of the contamination state based on sensor data from the contamination sensor; transfer of a contamination characteristic value; creation of a digital image of the contamination state; determination of the cleaning progress and inclusion of the current cleaning progress, determined from the jet movements already carried out, and transfer of the digital image of the contamination state and the cleaning progress as well as the currently valid process parameters to the control system (and to the evaluation of the cleaning process and the parameter optimization for machine learning).

[0047] This allows the control-related influence on the target jet cleaner to be minimized, a predetermined cleaning program to run initially unaffected, and the current cleaning progress, taking into account the contamination value, to be used to initiate the termination of the cleaning program when the cleaning target is reached.

[0048] Another aspect of the solution according to claim 10 is a computer program for self-learning optimization of the start parameters by reinforcement learning. This results in optimization between processes. The program for execution in a computer controller of a cleaning system, as described above, comprises the following steps: Incorporating results from a previously performed simulation, i.e., before the first use of the cleaning system or before a later use of the cleaning system to be optimized; incorporating past events and data stored in a database; incorporating requirements from a higher-level control system of a processing plant where the cleaning process is carried out, which are not related to the cleaning process itself but are determined by the processing process; incorporating stored expert knowledge about the cleaning process; incorporating a digital representation of the contamination state; incorporating process parameters of the cleaning process and outputting optimized start parameters to the state-based control system of a cleaning process.

[0049] The aforementioned computer programs are specifically tailored to the requirements of food storage tanks, but are fundamentally programmed to be applicable to a wide range of applications. Using machine learning (ML) methods – specifically reinforcement learning – the intelligent process control system generates a continuously self-optimizing cleaning system that also reacts inline to inhomogeneous levels of contamination. For the wide variety of contamination levels and cleaning parameters, simple rule-based control has proven insufficient. Instead, self-learning models have proven to be more suitable.

[0050] A key component of the self-learning model is the implementation of existing specialist and expert knowledge from the operators. This includes information from upstream and downstream processes, such as the degree and location of contamination, derived from filling time, filling volume, and cleanliness requirements, as well as simulation results. The implementation in the data preprocessing methodology and the control algorithms is primarily intended to support path planning for the target jet cleaner. Due to initially missing and successively built-up training datasets, these strategies enable a starting parameter selection that is equivalent to, or even superior to, the current, quasi-manual standard, and the realization of an industrially acceptable learning rate during the subsequent autonomous, situation-specific optimization.

[0051] The basis for the control system's solution approach is a database enriched with both external knowledge (experiential knowledge, simulations, etc.) and the continuous learning effects of machine learning. In a first optimization level, the system learns from process to process and successively improves the selection of starting parameters through partial, knowledge-based variation and the learning effects from a second optimization level. This additional second optimization level continuously optimizes the process inline and also reacts to unforeseen (not yet learned) cleaning situations. This ensures process reliability even during the early training phases and quickly generates a large number of data sets.

[0052] The cleaning system according to the invention, for example a tank cleaning robot with an integrated contamination sensor, can target and navigate to any point in the tank or other internal area of ​​a system for cleaning and contamination detection due to independently motor-controlled x, y, and z-axis adjustments. This allows the issued control instructions to be implemented flexibly.

[0053] Reinforcement learning, also known as repetitive learning, is a set of machine learning methods in which an agent independently learns a strategy to maximize rewards. The agent is not shown which action is best in a given situation, but rather receives a reward at specific times, which can also be negative. Based on these rewards, it approximates a utility function that describes the value of a particular state or action. Within the scope of the present invention, reinforcement learning is used to optimize parameters (e.g., water pressure, water temperature, cleaning additives, waiting times, cleaning paths, etc.). This means that, based on data from the contamination sensor, an optimization using machine learning is performed to generate an optimized cleaning program.This process uses a cost function to minimize water consumption and time required, while simultaneously delivering a clean tank. Furthermore, the cleaning parameters can be adjusted during the cleaning process. Unlike deep learning, reinforcement learning avoids the significant effort required for deep learning, which necessitates hundreds of large-scale trials.

[0054] Data-driven systems are capable of identifying patterns from the diversity of real-world situations. Using reinforcement learning, as applied in the present invention, machine learning software can independently recognize patterns in the cleaning parameters. It can thus be improved through training, ultimately delivering optimized cleaning paths. The recognized patterns are used for control and regulation.

[0055] The optimization of cleaning parameters, the machine learning, should preferably be performed on a separate server to ensure the necessary resources are available. For this purpose, the data from the contamination sensor(s) and the process data are sent to the server.

[0056] Developments in memory and processor technology are rapidly expanding the capabilities of these systems. As a result, their use in industrial processes is becoming increasingly important. Deep learning, in particular, but also reinforcement learning, as employed in the present invention, has led to significant successes in recent years, partly due to the large, domain-specific datasets available. Examples from other fields include the mastery of the game of Go achieved solely through self-training computers using deep learning algorithms, and the training of hand-eye coordination in robots with 800,000 grasping trials. These successes were achieved using deep learning methods, specifically convolutional neural networks.

[0057] In the field of cleaning, however, it is not possible to generate large amounts of data in a reasonable timeframe, thus hindering the use of deep learning. Surprisingly, it has been shown that it is also possible to work with small datasets. This requires sufficient preparation of the features from which the algorithm generates the model.

[0058] The software requires input data from the contamination sensor, which initially consists of many individual images. These images are then combined into a single, larger image, for example, depicting the interior of an entire tank. This image is analyzed using a software algorithm, contaminated areas are identified, and the results are displayed as an image or matrix. The output clearly shows which areas of a tank, for example, are contaminated and to what extent.

[0059] Furthermore, the optimization algorithm should also be fed an image, array, or matrix representing the cleaning difficulty of individual areas within the tank. This means that critical areas within the tank, particularly those that are difficult to clean, are marked separately. A tank nozzle or other internal components would be highlighted so that they can be cleaned more thoroughly and specifically.

[0060] This image, array, or matrix can be generated in several different ways: 1. Manual identification of the relevant areas in the tank by an expert: For example, the various critical areas could be color-coded in an image of the cut-open tank. 2. Using machine learning: For example, areas that are difficult to clean are identified in a white light image using object recognition and segmentation. 3. Using a CAD model: The tank geometry is converted into an image or array using a CAD model.

[0061] The CAD model is "fed" with all available information to serve as a learning aid. This includes strategies for integrating experiential knowledge or information from upstream processes. Relevant information can include the type of contamination, the drying time, the ambient temperature, etc. This information can be retrieved by a higher-level control system or by an operator after inputting the data. The information allows conclusions to be drawn about the contamination characteristics. The contamination characteristics can also be determined within the framework of a simulation. In this way, even with just a few data sets, good results can be achieved, enabling the use of the previously described methods in an industrial context.

[0062] Another aspect of the invention relates to a data carrier signal of a computer program, which can be used to distribute the program. The data carrier signal can, for example, be stored on a data carrier or offered for download on the internet.

[0063] The data processing component of the software, the computer-implemented invention, is the core of the software, realized using a suitable software architecture that controls the overall process and the interaction of control, regulation, and optimization of the cleaning process, as well as all data handling. Furthermore, the essential process knowledge is integrated at this stage. A model is created using machine learning (ML) to serve as the basis for optimizing the cleaning process. The input data is derived from the understanding of the process, e.g., process parameters such as temperature, pressure, and cleaning agents, but also from calculated characteristic values ​​(features). A reward function is provided as the target variable for the ML model, enabling an overall evaluation of the cleaning success achieved and the actual resources required (time, energy, water, cleaning chemicals).Methods for generating new process parameters are implemented through targeted parameter variation based on evolutionary algorithms and metaheuristic optimization.

[0064] An algorithm according to the invention for the state-based control of the cleaning process comprises the following steps: Transfer of optimized parameters to the control system, transfer of varied process parameters to the target jet cleaner from the control system, detection of the contamination state by the contamination sensor and transfer of a contamination indicator, creation of a digital image of the contamination state, determination of the cleaning progress, transfer of the digital image of the contamination state and the cleaning progress as well as the currently valid process parameters to the control system (and to the evaluation of the cleaning process and the parameter optimization for machine learning).

[0065] A preferred algorithm for self-learning optimization of the starting parameters as an optimization between processes includes the following steps: Data transfer for the evaluation of the cleaning process and parameter optimization for machine learning from: ∘ a simulation, ∘ a database, ∘ expert knowledge of the cleaning process, ∘ currently valid process parameters, ∘ currently valid digital representation of the contamination state and cleaning progress, ∘ data from the higher-level control system of the processing plant; output of optimized parameters to the control system, output of data to the database.

[0066] The computer program and the data carrier signal for the state-based control of the cleaning process and for the self-learning optimization of the start parameters can be implemented as hardware or as software, e.g. running on a computer.

[0067] Retrofitting existing cleaning systems with the system according to the invention is also possible. It has been found that this requires consideration of a limited size, which is associated with the challenges discussed above. For example, if an existing system fits through a DN 125 connection and is also mounted on one, then when replacing the existing system with a new system according to the present invention, this new system must also fit through the DN 125 connection.

[0068] The present invention provides a combined system consisting of a freely configurable actuator (tank cleaner) with integrated, innovative contamination sensors that detect the current contamination level in the tank in real time. Together with the accumulating collection of historical data and information from upstream and downstream processes, the intelligent process control is enabled to continuously optimize the cleaning processes in relation to the specific situation. The cleaning system according to the invention reacts to the highly variable contamination levels by using an implemented self-learning algorithm to preselect the optimal cleaning procedure after a learning phase and readjust it inline as needed.

[0069] The invention will be explained in more detail below with reference to the description of exemplary embodiments and their representation in the accompanying drawings.

[0070] Fig. 1Figure 1 schematically shows a perspective view of an embodiment of a cleaning system 1 according to the invention, comprising a telescopic arm 2, a rotary head 4 arranged at a first end of the telescopic arm 2, and a nozzle arm 8 rotatably mounted at its first end on the rotary head 4. The second end of the nozzle arm 8 has a nozzle 6 designed to dispense a cleaning agent, preferably in liquid form. The cleaning agent can also be gaseous or powdery and can also be dispensed as a jet.

[0071] The telescopic arm 2 is rotatable (see arrow direction) and is connected to a drive (not shown) at its other end. Furthermore, the height of the rotating head 4 can be changed vertically by changing the length of the telescopic arm 2.

[0072] The drive for telescopic arm 2 and nozzle arm 8, which is movable relative to the rotating head 4 (see arrow direction there), can be fully or partially motorized, with a fluid drive also being an option. This allows each individual point in the room, tank, or container to be precisely targeted with the cleaning agent jet. Any desired cleaning path can be configured, as the controllable axes can be adjusted independently.

[0073] Fig. 2 Figure 1 schematically shows another embodiment of a cleaning system 1 according to the invention, in particular the rotary head 4, which has the nozzle arm 8 combined with the sensor arm 22. Thus, the nozzle 6, a light source 24, and a contamination sensor 26 are arranged together in a single unit that is movable relative to the rotary head 4. The light source 24 and the contamination sensor 26 together form the sensor device 20.

[0074] Another embodiment of the cleaning system 1 according to the invention shows Fig. 3 , where the nozzle arm 8 with the nozzle 6 and the sensor arm 22 with the sensor device 20 (comprising the light source 24 and the contamination sensor 26) are arranged at different locations on the rotating head 4. According to a first embodiment, both can be driven synchronously by a common shaft, or according to a second embodiment, each can be moved into a different position, for which purpose different drives are provided for the nozzle arm 8 and the sensor arm 22.

[0075] In Fig. 4Figure 1 shows an embodiment in which the sensor device 20 is arranged within the rotary head 4, protected from the influences during cleaning (especially high temperatures, aggressive chemicals). For this purpose, the housing of the rotary head 4 must be at least partially transparent to allow light from the light source 24 to pass through, and partially transparent to allow the contamination sensor 26, particularly if it is designed as a camera, to see outside the rotary head 4.

[0076] The sensor device 20 can be carried along when the nozzle arm 8 is driven, or it can perform its own movement using its own drive.

[0077] Fig. 5Figure 1 shows a further embodiment of the cleaning system 1 according to the invention, in which the sensor device 20 is arranged at a distance from the rotating head 4 in the telescopic arm 2. This sensor device 20 can be provided as an additional sensor device, for example, in addition to a sensor device 20 arranged in the rotating head 4. This would make it possible, during the forward movement of the rotating head 4 in a vertical downward direction, to inspect the previously cleaned area after the cleaning fluid has drained away and to immediately document the cleaning success, transmit this information to the control device, and, if necessary, trigger a subsequent cleaning.

[0078] Fig. 6Figure 2 shows the telescopic arm 2 with the rotating head 4 at its first end and the drive unit 3 at its second end. Two positions that the telescopic arm 2 can assume in the vertical direction are shown. In the left illustration, the telescopic arm 2 has reached its maximum length, while the illustration on the right shows the minimum vertical extension.

[0079] Fig. 7 schematically shows a partially cutaway representation of an alternative embodiment of a telescopic arm 2 of the cleaning system 1 according to the invention, which, in contrast to the embodiment according to Fig. 6 without further subdivision, attached to a tank cap 30 and equipped with the rotary head 4.

[0080] A separate motor of the drive unit 3 generates a vertical traversing movement of the rotary head 4 into the tank (see direction of arrow), through the tank lid 30. The drive force of the motor can be transmitted in the drive unit 3 via a spindle, belt drive, chain drive, or rack and pinion. Linear guidance is provided by a corresponding sliding frame 32, which is longitudinally mounted in a sliding guide 34.

[0081] The basis for the control engineering solution approach according to Fig. 8This forms a database that is enriched with both external knowledge (experiential knowledge, simulations, etc.) and continuous learning effects. In optimization level 1, the system learns from process to process and successively improves the selection of starting parameters through their partial, knowledge-based variation and the learning effects from optimization level 2. This additional level 2 tracks continuous optimization inline within the process and also reacts to unforeseen (not yet learned) cleaning situations. This ensures process reliability even during the early training phases and quickly leads to a large number of data sets.

[0082] Two processes are being executed: a state-based control of the cleaning process and a self-learning optimization of the start parameters as an optimization between processes.

[0083] Machine learning involves evaluating and comparing the cleaning process, as well as optimizing its parameters. The following factors are incorporated into the machine learning process: Results of a simulation; events and data stored in a database; requirements of a higher-level control system for a processing plant; expert knowledge of the cleaning process; digital representation of the contamination state; process parameters and outputs optimized start parameters to the control and regulation process.

[0084] The control system provides continuous regulation within the process, whereby the target jet cleaner is controlled by varying process parameters, in particular its jet movement and thus the progress of the cleaning process (cleaning progress). This leads to a reduction in the level of contamination, which is detected by a contamination sensor. The measurement result (a digital representation of the contamination level) is incorporated into the control system. The jet movements are determined by documented control commands to the manipulators, such as stepper motors, which effect the required nozzle position, and / or by position sensors in the nozzles, which detect their position. This data, insofar as it is collected and available, is also incorporated into both control processes: the control of the cleaning process and machine learning. The position sensors are also part of the sensor system.

[0085] The control and regulation process forwards the process parameters to the machine learning process for evaluation, comparison, and parameter optimization.

[0086] The software concept can be implemented as hardware, in the control and regulation device, or run as software on a computer. Reference symbol list

[0087] 1 Cleaning system 2 Telescopic arm 3 Drive unit 4 Rotating head 6 Nozzle 8 Nozzle arm 10 Cleaning agent jet 20 Sensor unit, camera 22 Sensor arm 24 Light source 26 Dirt sensor 30 Tank cap 32 Sliding frame 34 Sliding guide

Claims

1. A cleaning system comprising a targeted jet cleaner, which in itself comprises at least one nozzle (6) for dispensing cleaning agent in the form of at least one jet, wherein the nozzle (6) is arranged on a rotary head (4) and is connected there to a manoeuvring device capable of moving the nozzle (6) such that the jet is directed in all directions, wherein, furthermore, the rotary head (4) is arranged at a first end of a telescopic arm (2), wherein the cleaning system (1) comprises at least one control device for state-based control of the cleaning process and at least one sensor device (20), wherein the sensor device (20) comprises at least one contamination sensor (26) and at least one light source (24) which serve to detect a contamination state and to transmit a contamination parameter, and are thus configured to generate a digital representation of the contamination state, and wherein the sensor means (20) is connected to the control means at least for the transmission of sensor data, wherein the control means is configured to execute at least one computer programme, characterised in that the computer programme comprises an implemented self-learning algorithm which is configured to respond to the widely varying contamination states on a situation-specific basis by always preselecting the optimal cleaning procedure following a training phase and readjusting it inline as required, so that the control device is configured to manage the cleaning process on a situation-specific basis close to the respective optimum.

2. Cleaning system according to claim 1, wherein, in a first embodiment, the at least one sensor device (20) is arranged in a movable sensor arm (22) or / and wherein, in a second embodiment, the at least one sensor device (20) is arranged in the rotary head (4) or / and wherein the at least one sensor device (20) is arranged in the telescopic arm (2), spaced apart from the rotary head (4).

3. Cleaning system according to claim 1 or 2, wherein the light source (24) and / or the contamination sensor (26) comprises an active and / or a passive cooling device.

4. Cleaning system according to any one of the preceding claims, wherein the light source (24) is designed as a switchable hybrid solution for emitting different light spectra.

5. Cleaning system according to claim 4, wherein the light source (24) comprises at least one LED and / or is switchable at least between UV light, daylight or white light.

6. Cleaning system according to any one of the preceding claims, wherein the light source (24) comprises a light control device which includes controllable optical components and enables a variable illumination diameter, wherein at the start of cleaning, when the degree of soiling is high and the soiling is easy to detect, a shorter focal length is set, thereby illuminating a large area with low light intensity per unit area; and towards the end of the cleaning process, when optical detection limits play an increasingly significant role due to a decrease in layer thickness, the area is illuminated with strong focusing on a small area by increasing the focal length.

7. Cleaning system according to any one of the preceding claims, wherein the contamination sensor (26) is designed as a camera device comprising controllable optical components.

8. Cleaning system according to claim 6 or 7, wherein the controllable optical components are designed as a zoom lens with a controllably variable focal length.

9. A cleaning system according to any one of the preceding claims, wherein the telescopic arm (2) can vary its length extending into a room to be cleaned and thereby variably position the rotary head (4) within the room in a longitudinal direction.

10. A computer programme comprising instructions which cause a control device of a cleaning system according to any one of claims 1 to 9 to execute a state-based control method comprising the following steps: • Transferring optimised parameters to the control device, • Outputting varied control instructions from the control and regulation process to the target beam cleaner; • taking into account the contamination status based on sensor data from the contamination sensor; • transmitting a contamination parameter; • creating a digital representation of the contamination status; • determining the cleaning progress and taking into account the current cleaning progress, determined at least from the jet movements already carried out; • Transferring the digital representation of the contamination status and the cleaning progress, as well as the currently valid process parameters, to the control unit; and • Transferring the digital representation of the contamination status and the cleaning progress, as well as the currently valid process parameters, for the evaluation of the cleaning process to the parameter optimisation module for the self-learning optimisation of the start parameters.

11. A computer program comprising instructions which cause a control unit of a cleaning system according to any one of claims 1 to 9 to execute a self-learning optimisation method for start parameters of a cleaning process, which comprises the following steps: • Incorporating results from a previously performed simulation; • Incorporating events and data from a previous cleaning process stored in a database; • Incorporating requirements from a higher-level control system of a processing plant in which the cleaning process is carried out; • Incorporating stored expert knowledge regarding the cleaning process; • Incorporating a digital representation of the contamination status; • Incorporating process parameters of the cleaning process; • Outputting optimised start-up parameters to the state-based control and regulation of the cleaning process; and • Outputting data to the database.

12. A data carrier signal of a computer program according to any one of claims 10 or 11.