Control of a conveying section facility for unit loads

Abstract

This invention relates to a method for controlling a conveyor section for individual goods (P, P1, P2, P3), wherein the conveyor section comprises multiple interconnected sub-conveyor sections (C1, C2, C3, C4, C5), each sub-conveyor section being driven by a driver. One or more sensors for detecting individual goods are located on at least some of the sub-conveyor sections. Control of the driver is performed using a machine learning model via a computing unit. The machine learning model repeatedly receives input data comprising vectors of fixed length, wherein each vector element is assigned to a segment (BIN) of the conveyor section and indicates the current occupancy percentage of the individual goods (P, P1, P2, P3) in the corresponding segment (BIN). Each sub-conveyor section (C1, C2, C3, C4, C5) is divided into multiple segments (BINs) of equal size. The invention also relates to a corresponding device or system for data processing, a computer program, a computer-readable data carrier, and a data carrier signal.

Description

Technical Field

[0001] This invention relates to the control of the conveying section for a single item of goods. Background Technology

[0002] Many logistics systems are based on the fact that single items, such as parcels or individual products, are transported along conveyor sections. In this case, a conveyor section typically consists of multiple sub-conveyor sections arranged one after another, allowing the continuous transfer of single items from one adjacent sub-conveyor section to another. Each sub-conveyor section contains a conveyor belt on which the single item moves. Alternatives to conveyor belts are track systems or multi-carrier systems. Multiple such conveyor sections can be combined in arbitrarily complex ways, such as multiple conveyor sections extending parallel to each other and terminating at a common node.

[0003] Each sub-conveyor section's conveyor belt is driven by its own driver, allowing for individual speed settings for each sub-conveyor section. This control is typically performed using a high-speed clock, enabling very precise setting of the destination for individual items, especially in the final sub-conveyor section. This requires each sub-conveyor section to be optimally accelerated or decelerated based on control signals from a computing unit, assigned to its own driver. Poor control manifests as, for example, items colliding, falling off the conveyor section, failing to reach the desired destination, or taking an excessively long time to reach it. Summary of the Invention

[0004] The objective of this invention is to demonstrate a method for controlling a transport section.

[0005] This task is solved by the method according to the invention for controlling a conveyor section of a single item. Furthermore, the subject of the invention is a corresponding device or system for data processing, a corresponding computer program, a corresponding computer-readable storage medium, preferably a non-volatile storage medium, and a corresponding data carrier signal. Advantageous designs and extensions are the subject of various embodiments.

[0006] The method according to the invention is used to control a conveyor segment for a single item. The conveyor segment comprises multiple interconnected sub-conveyor segments, each driven by a driver. One or more sensors for detecting the single item are located on at least some of the sub-conveyor segments. Control of the driver is performed using a machine learning model by means of a computing unit. For this purpose, the machine learning model repeatedly receives input data comprising a fixed-length vector, where each vector element is assigned to a segment of the conveyor segment and indicates the current occupancy percentage of the single item in the corresponding segment. In this case, the conveyor segment is divided into multiple segments of equal size. In addition to the fixed-length vector, the input data may also contain other information.

[0007] Because each sub-conveyor segment has its own drive, these sub-conveyor segments can be accelerated or decelerated individually. This acceleration or deceleration can also include a value of zero, meaning the sub-conveyor segment can also operate at a constant speed, at least temporarily. By selectively accelerating or decelerating each sub-conveyor segment, a pre-defined objective for the conveyor segment can be achieved, such as a defined distance between individual items being transported on the final sub-conveyor segment.

[0008] At least one sub-conveying section is equipped with a sensor for detecting individual items. It is possible that some or all sub-conveying sections have such sensors, and multiple sensors may be provided for each sub-conveying section. The sensor positioning related to the extension of the sub-conveying section can be the same, for example, at the beginning of each sub-conveying section, but the sensor positioning can also differ between conveying sections. The sensor is preferably a grating. Other suitable sensor examples are: 2D / 3D cameras, RFID receivers combined with RFID tags on individual items, and sensing loops.

[0009] The driver for the sub-conveyor section is controlled by a computing unit. The computing unit pre-sets the speed or, alternatively, the acceleration or deceleration that should be set for the driver.

[0010] This is calculated by a machine learning model. For this, the machine learning model needs input data describing the current state of the transport segment. Specifically, the input data should describe where a single item is currently located on the transport segment. This single item location is given using vectors of fixed length; that is, each time the machine learning model receives new input data describing the location of a single item, the vectors indicating these locations have the same number of vector elements. Therefore, the transport segment is theoretically divided into segments of equal length. Thus, in the control-related section of the transport segment, such as from the first sensor to the last sensor, all parts of the transport segment belong to one of these segments; therefore, this section is completely divided into these segments. Individual sub-transport segments can be divided into different numbers of segments in this case, because their lengths can differ from each other. Each segment corresponds to a vector element. The corresponding vector element indicates the proportion of the segment occupied by a single item. Each segment can be completely occupied, partially occupied, or not occupied at all by a single item. For example, a single item might be located in the first half of a segment and the last third of a previous segment.

[0011] The input data provided to the machine learning model is current information, that is, it describes the current state of the transport segment. Of course, this does not exclude a certain time delay, which may exist between the determination of the transport segment's state and the machine learning model's receipt of the input data.

[0012] In an extension of this invention, the machine learning model receives updated input data at a clock and outputs information about the speed to be set for each sub-transport segment at the same clock. The faster the clock, the more accurate the control of the transport segment. Modern transport segments operate with control clocks in the millisecond range.

[0013] According to one design of the invention, the occupancy ratio is determined based on sensor measurements and the speed of a sub-conveyor segment. This allows the state of the conveyor segment to be described using a combination of measurement and calculation. Specifically, the current occupancy ratio can be determined for a specific individual item over time by a sensor detecting the item and assigning the measurement to one or more corresponding segments containing the corresponding position, then a sensor using the speed of the corresponding sub-conveyor segment to calculate the position of the item and assigning the calculated result to one or more corresponding segments containing the corresponding position, and so on. During this time-varying process, the vector value indicating the occupancy of the segment thus moves from one sensor through the segment to the next. This change from using measured values ​​when detecting an item with a sensor to using calculated values ​​of the segments between sensors can be applied to the entire length of the conveyor segment. Application: When a sensor detects an item, this point is recorded in the vector. This also applies to cases where calculated tracking of items between sensors indicates different points in time when they arrive at the sensor.

[0014] Advantageously, each vector element is represented by a non-binary value indicating the current occupancy percentage, preferably a value between 0 and 1. In this case, for example, 0 can represent "no single item in the segment", 1 represents "the segment is completely occupied by single items", and 0.5 represents "half of the segment is occupied by single items".

[0015] In an extension of the present invention, the input data further includes:

[0016] - Current information relating to at least one location of at least one part on the last sub-conveyor section; this could be, for example, a zero-line marker or the target area to which a single item should be delivered. This part moves at the corresponding speed of the sub-conveyor section.

[0017] and / or

[0018] - Information related to the current speed of the sub-conveyor segment,

[0019] and / or

[0020] - The current measurement results of the sensors; this includes which sensor is currently detecting or not detecting a single item.

[0021] In this invention, the machine learning model is trained before controlling the conveyor segment. For training, the machine learning model is pre-fed with input data comprising a fixed-length vector, the vector elements of which are determined based on a simulation of the conveyor segment. The input data during training should be of the same data type as the input data during actual control. However, unlike actual control, the training data does not come from the real system but from a virtual system. This allows for interaction between the machine learning model and the simulation without damaging individual components or spending excessive time measuring the real system.

[0022] Particularly advantageous is that the training is performed as reinforcement learning, wherein an objective function is used to determine the reward or punishment, the objective function comprising:

[0023] - A single item arrives at a pre-defined location on a sub-transport segment, preferably the last sub-transport segment; such arrival or non-arrival will trigger a reward or penalty.

[0024] and / or

[0025] - Multiple individual items collided on the conveyor section; this incident will trigger penalties.

[0026] and / or

[0027] - The speeds of adjacent sub-conveyor sections are similar; in order to avoid friction effects, the similar speeds of two successive sub-conveyor sections are advantageous for transferring a single item from one sub-conveyor section to another, which will trigger a reward.

[0028] and / or

[0029] - Maintain a minimum or target distance between two individual items. This distance can be measured by one of the sensors, preferably on the last sub-conveyor segment.

[0030] The method according to the invention and / or one or more functions, features, and / or steps of the method according to the invention and / or the design of the method can be operated in a computer-aided manner. The method can be executed or implemented, for example, by means of one or more computers, processors, application-specific integrated circuits (ASICs), digital signal processors (DSPs), and / or so-called "field-programmable gate arrays" (FPGAs). The method can also be executed at least partially in the cloud and / or in edge computing environments. One or more interacting computer programs are used for the computer-aided operation. If multiple programs are used, these programs can be stored together on a single computer and executed by that computer, or they can be executed on different computers at different locations. Since this is functionally equivalent, the terms "computer program" and "computer" are expressed herein in the singular. Attached Figure Description

[0031] The invention will now be explained in more detail with reference to embodiments. Herein:

[0032] Figure 1 The conveyor section is shown.

[0033] Figure 2 The conveyor section is divided into storage boxes.

[0034] Figure 3 A flowchart is shown.

[0035] Figure 4 A control unit for controlling the movement of the conveyor section is shown.

[0036] Figure 5 A more detailed illustration of the control unit is shown. Detailed Implementation

[0037] In internal logistics or generally in a production environment, conveyor sections with multiple individually operable conveyor belt segments are typically used. Figure 1 This illustration shows a conveyor section, which can be used, for example, in a packaging machine. Here, the product must be accelerated / decelerated on the conveyor section to transport it in sync with a clock frequency to a bagging machine (not shown) at the end of the conveyor section. In this illustration, product P from the right is transported to the left from sub-conveyor sections in the form of conveyor belts or production lines C1, C2, C3, C4, and C5. Product P is any single item. Conveyor belts C1, C2, C3, C4, and C5 can have the same length or different lengths.

[0038] Each of conveyor belts C1, C2, C3, C4, and C5 is assigned a driver A1, A2, A3, A4, and A5, respectively. Conveyor belts C1, C2, C3, C4, and C5 can be individually accelerated or decelerated by means of a control unit (indicated by arrows representing the control signals from the control unit to the corresponding drivers A1, A2, A3, A4, and A5). The goal of this individual movement control of conveyor belts C1, C2, C3, C4, and C5 is to place the product at a specific target location T on the last conveyor belt C5. The transfer of individual items to the conveyor section is carried out via conveyor belt C1, which serves as a transfer unit and is also constructed as a sub-conveyor section as shown in the figure. Therefore, in general, the arriving individual items appear randomly in terms of both time and size, and these items should be placed in a pre-given position.

[0039] To enable the control unit CONTROL to generate control signals suitable for accelerating and decelerating drives A1, A2, A3, A4, and A5, conveyor belts C1, C2, C3, C4, and C5 are equipped with sensors. Figure 1 In the example, sensors S1.1, S1.2, S2, S3, S4, and S5 are distributed as follows: the first conveyor belt C1 has two sensors, sensor S1.1 at the beginning and sensor S1.2 at the end; all other conveyor belts have one sensor each for S2, S3, S4, and S5. These sensors cannot be directly located at the start / end points of each conveyor belt, but are positioned at a small distance from the start point; however, they can generally be installed anywhere. It is not necessary to equip each conveyor belt with one or more sensors. Sensors S1.1, S1.2, S2, S3, S4, and S5 are gratings. These gratings output binary values ​​LOW and HIGH, where HIGH indicates the presence of a product in the grating. Additional sensors can also be provided for determining the speed of conveyor belts C1, C2, C3, C4, and C5, such as speed sensors for detecting the rotational speed of drives A1, A2, A3, A4, and A5, and current sensors for detecting the motor current of drives A1, A2, A3, A4, and A5, etc. Alternatively, the speeds of conveyor belts C1, C2, C3, C4, and C5 can be assumed to be known from the control signals of the control unit CONTROL.

[0040] Figure 1 The conveyor section shown is a simple example. More complex conveyor sections are found, for example, in parcel sorting facilities that may include multiple such fingers, which in turn contain multiple sequentially connected conveyor belt sections. A concrete example of this is the so-called "dynamic gapper," a controlled-drive application recently used in internal logistics. Thus, each finger in a parcel sorting facility is, in principle, like... Figure 1 The structure is illustrated. Parcel sorting facilities are specifically designed to separate parcels of different or similar sizes. The individual fingers of a parcel sorting facility are conveyor sections extending parallel to each other. At the end of these conveyor sections is a combining unit, called a merger, which transfers the individual items transported by the last sub-conveyor sections in the conveying direction to this combining unit. A single exit conveyor section is located at the exit of this combining unit. The parcel flow converges on this output conveyor belt. Therefore, the products on this output belt should be sorted out by the parcel sorting facility.

[0041] The individual acceleration and deceleration of the finger-shaped conveyor belts allow for the staggered transport of individual items on different fingers to the assembly unit. This enables the assembly unit to convey individual items to the exit conveyor section, ensuring that each pair of consecutive individual items has a pre-defined minimum distance between them. To avoid collisions on the converging conveyor belt, it is helpful to separate parallel-moving packages as early as possible.

[0042] Any complex and extended conveyor section can be conceived and used. Regardless of the specific design of the conveyor section, the task of the control unit (CONTROL) is to determine and set the optimal speed for each conveyor belt at any given time and under any conditions characterized by varying quantities, sizes, and positions of the products. The optimal speed depends on the function and target settings of the corresponding conveyor section.

[0043] Determining optimal control is technically very challenging. Classic regulators are essentially based on grating signals to track product position. A particular problem in this case is that, due to friction and different speeds of the two conveyor belts, it is difficult to manually model the transition between them. To improve this, the facility can be equipped with additional sensors, such as cameras, which can measure product position very accurately at any given time. However, the downside is that such expensive additional sensor systems are quite costly. In some cases, multiple cameras are needed for the entire conveyor section to achieve accurate measurements. Furthermore, a fast clock, such as 4ms, must be used to operate the drive. The evaluation of the camera images must be performed accordingly quickly. This image processing for capturing product position (if any) can only provide such a clock using very expensive hardware.

[0044] The following explains how artificial intelligence can be used to control the facility. For this purpose, a machine learning model (hereinafter referred to as the ML model) is trained, and then this machine learning model, as part of the control unit CONTROL, takes over the movement control of conveyor belts C1, C2, C3, C4, and C5 during the actual operation of the conveyor section. The trained model can specifically be an artificial neural network, recurrent neural network, convolutional neural network, perceptron, Bayesian neural network, autoencoder, variational autoencoder, Gaussian process, deep learning architecture, support vector machine, data-driven regression model, k-nearest neighbor classifier, physical model, and / or decision tree. Methods from the field of reinforcement learning, such as policy gradient methods, are particularly suitable as training methods.

[0045] To train the ML model, a system model of the transport segment is used. This digital twin of the transport segment can be, for example, a simulation based on physical equations (especially via Unity or NX MCD) or a neurodynamic model (typically a recurrent neural network).

[0046] The ML model obtains the following state information from the system model for training:

[0047] -The position of the last conveyor belt

[0048] Each conveyor belt has markings, etc., indicating the belt's current location. This is relevant because it allows us to determine the current location of the target point T.

[0049] - Current speed of the conveyor belt

[0050] When the ML model and the system model interact for training purposes, the current speed is consistent with the speed pre-given by the ML model;

[0051] -Product location of a single item on the conveyor section

[0052] The product location is determined in the system model as follows:

[0053] Product registration and measurement: On the first conveyor belt C1, the length of a newly arriving product is measured at the first sensor S1.1. The arrival of the new product is identified by a change in state from LOW to HIGH on the sensor S1.1. While the signal is HIGH, the time-dependent speed v(C1)_t of the conveyor belt C1 is integrated over time. This integration ends with the transition from HIGH to LOW on the sensor S1.1, thus yielding the product length l(P). To avoid measurement inaccuracies, it is advantageous to have the first conveyor belt travel at a constant speed for product registration.

[0054] Product tracking: Position updates are performed computationally. The time-related position x(P)_t of the product at time point t is recalculated by adding the current speed of the corresponding conveyor belt C where the product may currently be located (i.e., based on the calculated product tracking) to each time step: x(P)_t = x(P)_{t-1} + v(C)_t.

[0055] Modeling conveyor belt transitions (i.e., products moving from one conveyor belt to the next, e.g., from C1 to C2) is challenging. The system model maps the actual physical characteristics of these transitions. These realistic physical conditions specifically include the coefficient of friction of the belt surfaces. Therefore, all physical relationships, including friction, are modeled and calculated. This allows us to determine whether the product experiences slight slippage on the slower conveyor belt when transitioning from a faster to a slower one, or conversely, whether it does not immediately move at the same speed as the conveyor belt when transitioning from a slower to a faster one.

[0056] Product cancellation: When the product reaches the last sensor S5 on the last conveyor belt C5, the tracking of the corresponding product position also stops.

[0057] - The current state of the sensor, i.e., the binary signal indicating whether the product has just passed through the grating.

[0058] The ML model obtains these four interpretations of state information from the system model. This state information is provided to the ML model in a clock (e.g., every 4 milliseconds) for manipulating the drive. This is treated as a time slice by the ML model. In this way, the artificial intelligence can determine the manipulation information for the drive within this clock. For successful training of the ML model, it is desirable that the state information obtained from the system model possess Markov properties. That is, the ML model must be able to predict subsequent states in time using the current state information and the actions taken by the ML model (i.e., setting the speeds of each conveyor belt). The following describes how to obtain the state information corresponding to this requirement based on the aforementioned system model. Therefore, suitable input features should be provided for successful use of the ML model. Here, in the problem domain under consideration, i.e., in the described conveyor section arrangement, there is the following question regarding the state information of the product positions:

[0059] It's important to recognize that ML models require a defined, constant number of input variables or input features. This can be explained by the fact that neural networks treat input data as vectors and multiply them by a matrix of weights to be set within the training scope. However, this doesn't apply to the arrangement of product positions on a conveyor belt: at different times, there might be different numbers of products simultaneously on the conveyor belt. However, a mandatory premise of ML models is having input vectors of constant length, where that length corresponds to a constant number of input variables. For this purpose, for example, the positions of the n foremost products could be passed to the ML model. This works because, as mentioned earlier, the appearance of objects on the last conveyor belt might be the objective function to be taught to the ML model. However, this doesn't lead to optimal results because other products on the conveyor belt further back might be hidden. To take this into account and thus obtain high-quality input data for the ML model, the amount of input data must vary, as the number of packages might change over time. However, neural networks cannot solve this problem.

[0060] To avoid the problems explained, the binary grating signal is converted into features that can be processed by an ML model in the form of a neural network. Based on Figure 2 To explain this process, Figure 2 A segment of a conveyor belt divided into storage boxes of the same size is shown. Figure 2The segment shown depicts conveyor belts C2, C3, C4, and C5, with the conveyor belts traveling in the right-hand direction. Product P1 is located on conveyor belt C5, product P2 on conveyor belt C4, and product P3 on conveyor belt C3. Conveyor belts C2, C3, C4, and C5 are theoretically divided into sections (BINs), which are schematically shown below the conveyor section. In this case, the sections (BINs) have the same extension in the direction of travel. This artificial grid formed by the sections (BINs) covers the entire conveyor section.

[0061] Each segment BIN is assigned a value between 0 and 1, proportional to the product's occupancy of the corresponding segment BIN. Here, 0 means the corresponding segment BIN is empty. 1 means the corresponding segment BIN is completely covered by the product. If the product only covers a portion of the corresponding segment BIN, the corresponding segment BIN receives a value corresponding to its occupancy percentage. These values ​​form the product position state vector VECTOR, whose values ​​are exemplarily entered into the segment BIN based on the positions of products P1, P2, and P3 located on conveyor belts C2, C3, C4, and C5. This product position state vector VECTOR, containing the occupancy rate of each segment BIN and thus indicating the product position, always has a constant length and can therefore be used as an input feature for product position in ML models. This enables artificial intelligence modeling using neural networks. Figure 2 The lower part is a graphical representation of the values ​​of the product position state vector VECTOR. This representation is for illustrative purposes only; there is no specification for handling this continuous change in vector values ​​through an ML model.

[0062] Corresponding to the camera image, the described product position state vector VECTOR corresponds to a complete snapshot of the conveyor segment and all products located on that segment. Therefore, a small number of time slices (approximately 10) are sufficient for the ML model to make optimal decisions regarding motion control, as a holistic view of the system is provided to the ML model in the form of all segment bins. Conversely, in alternative approaches—such as when modeling using a recurrent neural network, where only the grating signal and conveyor belt speed are considered—many time slices (approximately 60) must be provided at a high clock rate to enable the ML model to estimate the Markov state.

[0063] As already described, these segment bins have the same extension in the direction of travel, for example, 6 cm. To improve the overall representation, this extension of the segment bins is drawn significantly larger in the diagram. The size of the segment bins can be freely chosen. An important criterion is the training time required for the ML model, which increases dramatically with the number of segment bins. As a converse criterion, it should be considered that if the range of the segment bins is chosen to be very large, multiple packages may be located within a single segment bin. Markov states can be reliably obtained when the segment bins are chosen to be small enough that at most one product can be completely contained within a single segment bin.

[0064] Advantageously, the division into segmented BINs is adapted to the distance between sensors. That is, at each sensor location, there exists a boundary between two adjacent segmented BINs. This is particularly meaningful when product tracking is synchronized at the sensors, as explained below. By synchronizing at the sensor locations, local modeling or local merging is generated on the observed sub-segments. Each sub-segment provides a predefined number of features in the form of vector values ​​of the corresponding segmented BIN, which are part of the overall product location state vector (VECTOR) of the ML model.

[0065] Furthermore, the ML model can infer from the product position state vector VECTOR that multiple products collide and subsequently travel together on the conveyor section. Based on Figure 2 The lower graphic representation shows the distance between the rising and falling edges, as well as the product length known due to registration, which allows us to determine that multiple products pass through the grating sequentially without gaps. This is generally avoided because it is no longer possible to separate products in a targeted manner, or the products may be damaged in the event of a collision.

[0066] The ML model learns the control strategy for the transport section using described state information, which includes the product position state vector VECTOR. In this case, the four state information items together form a vector of constant length.

[0067] Reinforcement learning methods are particularly well-suited for training ML models. In this case, an objective function is defined for the training, the type of which depends on the desired operation of the conveyor segment. The achievement or non-achievement of the objective can then be evaluated using rewards and / or penalties, thus generating learning effects for the ML model. In the case of a parcel sorting facility, for example, parallel activation of segment bins on two conveyor belts might be penalized with corresponding negative rewards. Binning allows neural networks to, for example, also identify collisions between two products because neighborhood relationships are created through segment bins. This is especially important when collisions should be negatively penalized to avoid them in reinforcement learning-based control. Without a binning grid, it would be impossible to map collisions in an input vector of defined length. Furthermore, binning allows collision detection to be performed in linear runtime (O(N), where N = number of parcels) instead of O(Nlog(N)), whereas if parcel positions are first sorted and then adjacent parcels are checked for collisions individually, collision detection would result in O(Nlog(N)). This enables fast, real-time control of the system. Another meaningful aspect to be achieved in the objective function is that the speed difference between adjacent conveyor belts should be small; this avoids slippage and similar effects at belt transitions.

[0068] After training, the artificial intelligence can take over actual control, that is, operate on the real system. For this purpose, the AI ​​is equivalent to the control unit or a part of the control unit. Therefore, the ML model calculates the speed of each conveyor belt and the control signals required for the drivers. In other words, the output of the ML model in actual operation is the target speed of each conveyor belt for each time slice, corresponding to the clock of the control.

[0069] In actual operation, a system model is no longer needed because the data now comes from the real system. The input data that the ML model receives to control the transport segment is composed in the same way as the state information during training. This means that the ML model receives the following data with the clock speed required to control the transport segment:

[0070] -The position of the last conveyor belt

[0071] - Current speed of the conveyor belt

[0072] The current speed is not required to be consistent with the speed pre-defined by the ML model. This is because, in actual operation, small deviations may occur when the trained ML model controls the conveyor section, such as the delay in the control signals reaching the drive or the time required for the conveyor belt to decelerate / accelerate.

[0073] -Product location of a single item on the conveyor section

[0074] For this purpose, the signals actually measured by the sensors are used, and then, as explained above regarding product tracking in the system model, the position values ​​between the sensors are calculated based on the conveyor belt speed. For belt transitions, a simple solution is to use the speed v(C1)_t of C1 as long as most of the product remains on C1. Only when more than half of the product is on C2 is the speed v(C2)_t of C2 used to continue tracking the product position x(P)_t. Alternatively, however, both speeds can be used proportionally, depending on the proportion of the product length l(P) that lies on the corresponding conveyor belt.

[0075] Only the first and last sensors are needed for this type of product tracking. The downside of this approach is that errors can accumulate over time due to inaccurate estimations of product position (e.g., primarily at conveyor belt transitions). Therefore, the transition from the first conveyor belt to the second is crucial, as it's not mandatory for products on the second conveyor belt to immediately adopt its speed. This is due to the different coefficients of friction on the conveyor belt surfaces and the different speeds of the conveyor belts. For example, if the speed of the first conveyor belt is significantly higher than that of the second conveyor belt, the product will slip at least a short distance before slowing down to the second conveyor belt's speed through static friction. Therefore, it is meaningful to run product tracking locally on each conveyor belt. Figure 1 In a specific embodiment, product P on conveyor belt C1 is computationally tracked from sensor S1.2 to sensor S2 on conveyor belt C2; that is, errors in transferring the product from conveyor belt C2 are only tracked up to sensor S2. Upon reaching sensor S2, product registration is performed on conveyor belt C2 without needing to re-determine the product length. The measurement from sensor S2 is used as the product's position, and product tracking restarts from there. Product tracking then proceeds from sensor S2 to sensor S3 on conveyor belt C3, and so on, thereby generating synchronization at each grating.

[0076] For synchronization at sensor location, use Figure 2 The falling edge in the lower part of the graphic representation is also advantageous.

[0077] As already explained, the product may slip during the transition from conveyor belt to conveyor belt, then slow down after a certain time so that the product does not slip as it moves up the corresponding conveyor belt. This slippage is more likely to end when the rear end of the product leaves the grating than when the front end of the product reaches the grating.

[0078] In this way, the product position state vector VECTOR is derived by combining sensor measurements with product tracking performed computationally between sensors.

[0079] - The current state of the sensor, which indicates whether the product has passed exactly through the grating, is the actual measurement signal.

[0080] There is no pre-defined set of which interpreted input features the ML model uses and to what extent it controls them. The ML model addresses this issue during training.

[0081] The particular advantage of using ML models for the motion control of industrial machines described is that it takes only a few days for an ML model to solve the control strategy to be applied, while the creation / programming of traditional control algorithms can take up to several years.

[0082] The explained use of the ML model is also a cost-effective and hardware-intensive solution because it eliminates the need for expensive additional sensor systems (such as cameras) to track the products, as the grid formed by the segmented BINs is similar to a camera image covering the entire conveyor section. In particular, the high clock rates required for motion control can be achieved without problems, allowing control to be performed using a clock of 4 ms or even less. Alternatively, if camera images are evaluated to determine the position of all products, a clock of approximately 100 ms must be considered.

[0083] In short, Figure 3 A flowchart of the described process is shown. In the first step, SIM-MODEL, a simulation model is created that maps to the real conveyor section. In the second step, BINNING, the BIN segment is defined to enable the creation of input vectors from the simulation model suitable for the ML model. In the third step, TRAIN, the ML model learns the optimal control for the conveyor section. In the fourth step, MOTION CONTROL, the ML model controls the conveyor section based on the behavior learned in the previous step, by pre-given various conveyor belt speeds or accelerations. This is the inference phase, where the weights taught by the neural network during the TRAIN step are retained and applied to the real system.

[0084] Figure 4 This illustrates how a control unit (CONTROL) can be constructed to implement an ML model. While the components explained in more detail below exist singly in the diagram, these components can also exist in multiple implementations, such as as a distributed system. In this way, the functionality of the control unit (CONTROL) can be divided into multiple systems that can be hierarchically linked to each other if necessary. One or more systems may be located near the transport section or elsewhere.

[0085] The control unit (CONTROL) includes a computing unit or processor (PRO). The processor (PRO) is connected to a memory (MEM), in which the computer program (PROGRAM) is stored. The memory (MEM) is preferably a non-volatile computer-readable data storage medium. Storage can be performed in any manner suitable for ensuring readability by the computing unit, such as magnetic storage (e.g., by means of a floppy disk), optical storage (e.g., by means of a CD), magneto-optical storage, ROM (Read-Only Memory) storage, RAM (Random Access Memory) storage, EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), or flash memory.

[0086] The steps of the above process can be executed by executing the instructions of the program program in the processor PRO. This is particularly applicable to... Figure 3 The steps TRAIN and MOTION CONTROL are performed in advance, while the steps SIM-MODEL and BINNING are performed if necessary. For this purpose, the processor PRO is connected to the IN / OUT input / output unit, through which information can be exchanged between the control unit and other components and / or the user. In this case, the interface can be designed in a suitable manner, for example, via radio or via cable, and the communication can be performed using a suitable standard.

[0087] Figure 5 An example is shown Figure 4 The control unit (CONTROL) can be designed in several ways: it can be a general-purpose computer control unit or a mobile general-purpose computer control unit (CONTROL-MOBILE). In this case, a general-purpose computer control unit represents various types of digital computing devices, such as desktop computers, workstations, servers, blade servers, mainframes, or other suitable devices. A mobile general-purpose computer control unit (CONTROL-MOBILE) correspondingly represents various types of mobile digital computing devices, such as laptops, PDAs, mobile phones or smartphones, or other suitable devices. If based on... Figure 5 The components that can be used are shown and explained in detail, and should be understood as examples; implementation of the invention is not limited to these components.

[0088] The computer control unit includes the processor (PRO), memory (MEM), and the high-speed interface (HS-INTER) between them. Additionally, the high-speed expansion port (EXP) and the low-speed interface (LS-INTER) are connected to the high-speed interface (HS-INTER). The storage device (STORAGE) and the low-speed bus (LS-BUS) are connected to the low-speed interface (LS-INTER). Components PRO, MEM, STORAGE, HS-INTER, EXP, and LS-INTER are connected via suitable connectors / buses and can be mounted on a general-purpose motherboard.

[0089] Although Figure 5 The components of the computer control shown are implemented as a single unit, but some or all of these components can also be provided in multiple forms. The computer control can also include multiple interconnected computers located in different locations if necessary.

[0090] The processor PRO can process instructions that should be executed in the computer's control, particularly those that can be stored in memory (MEM) or storage devices (STORAGE). Information, especially the processing results in the processor PRO, can be graphically output on a display (such as a display screen) connected to a high-speed interface (HS-INTER) using a GUI.

[0091] The memory (MEM) is used to store information within the computer control. This can be volatile or non-volatile memory. The memory (MEM) may include multiple storage units. The storage device (STORAGE) is preferably a high-capacity storage device for computer-readable media. For this purpose, the storage device (STORAGE) may include, for example, a floppy disk drive, hard disk drive, optical disk drive, magnetic tape device, flash memory device, or even a series of devices such as those with a storage area network configuration. The high-speed interface (HS-INTER) handles bandwidth-intensive processes within the computer control, while the low-speed interface (LS-INTER) is used for processes with lower bandwidth requirements. For this purpose, the high-speed interface (HS-INTER) is connected to the memory (MEM), the display (DISPLAY, if necessary, via a graphics processor), and the high-speed expansion port (EXP), which can accommodate various expansion cards. The storage device (STORAGE) and the low-speed bus (LS-BUS) are connected to the low-speed interface (LS-INTER), on which a low-speed expansion port may exist. The low-speed expansion port may have various communication interfaces, such as USB, Bluetooth, Ethernet, and wireless Ethernet. These communication terminals can connect to various input and / or output devices, such as keyboards, mice, scanners, and network devices such as switches or routers. The task allocation between the two interfaces, HS-INTER and LS-INTER, is exemplary and can also be organized in other ways.

[0092] The computer control can be implemented in different ways, as shown on the right side of the figure. For example, the computer can be implemented as a personal computer (PC), a standard server (SERV), or a group of such servers, such as a server farm, or a rack server system (R-SERV), or as part of such a system.

[0093] A mobile computer (CONTROL-MOBILE) includes a processor (PRO), a memory (MEM), input and output devices (DISPLAY), a communication interface (COM-INTER), and transceivers (TX / RX). Components PRO, MEM, COM-INTER, and TX / RX are connected via suitable connections / buses and can be mounted on a common motherboard or otherwise suitably mounted. Additionally, further storage devices, such as microdrives, can be provided to create additional storage possibilities. The processor PRO executes instructions within the mobile computer (CONTROL-MOBILE), particularly those stored in the memory (MEM). The processor PRO can be implemented as a chip or chipset containing one or more analog or digital processing units. The processor PRO is particularly responsible for coordinating other components of the mobile computer (CONTROL-MOBILE), such as for controlling one or more user interfaces, applications running on the mobile computer (CONTROL-MOBILE), and wireless communication for the mobile computer (CONTROL-MOBILE). Information is transferred between the user of the mobile computer (CONTROL-MOBILE) and the processor PRO via the user interface (USER-INTER), for example, via voice input / output, and via the display interface (DISPLAY-INTER), for example, via text input / output. Input and output devices (DISPLAY) can be based on technologies such as TFT LCD (Thin Film Transistor Liquid Crystal Display) or OLED (Organic Light Emitting Diode).

[0094] In addition, an external interface EXT-INT is provided for connection to the processor PRO, through which the mobile computer CONTROL-MOBILE can perform near-field communication with other devices. Wired and / or radio communication is possible via the external interface EXT-INT.

[0095] The memory MEM is used to store information within the mobile computer (CONTROL-MOBILE) and can be implemented, for example, as a volatile or non-volatile memory comprising one or more storage cells. Furthermore, an extended memory EXT-MEM can be provided, which is connected to the mobile computer (CONTROL-MOBILE) via an extended interface MEM-INTER, such as an interface for a SIMM (Single In-line Memory Module) or SIM (Subscriber Identity Module). The extended memory EXT-MEM provides additional storage capacity to the mobile computer (CONTROL-MOBILE) and can also store various applications. For example, the extended memory EXT-MEM can be used as a security module for the mobile computer (CONTROL-MOBILE) by storing identification information there.

[0096] The mobile computer (CONTROL-MOBILE) can communicate wirelessly via a transceiver (TX / RX) equipped with means for digital signal processing. A communication interface (COM-INTER) enables communication via suitable protocols, such as GSM, SMS, MMS, CDMA, TDMA, PDC, WCDMA, CDMA2000, GPRS, EDGE, UMTS, LTE, and 5th generation or higher communication protocols. In addition to the TX / RX transceiver, a transceiver (not shown) for near-field communication such as Bluetooth and WiFi can also be provided. Finally, a GPS module (GPS) can be provided to enable location-based services. Received audio information (especially including user commands) can be converted into digital information that can be processed by the mobile computer (CONTROL-MOBILE) using an audio codec (AUDIO); correspondingly, the audio codec (AUDIO) can be used to generate sound information that can be captured by the user.

[0097] As shown on the right, the mobile computer CONTROL-MOBILE is preferably implemented as a smartphone PHONE or a laptop computer.

[0098] Interaction between a user and a computer control or a mobile computer control module can occur in various ways, and is not limited to these methods. Figure 5 Specifically, any perceptible transmission of information (visible, audible, tactile) is possible.

[0099] Although the components of computer control and mobile computer control-mobile have been described separately, a computer that includes both computer control components and mobile computer control-mobile components can also be used.

[0100] Figure 4 The computer program PROGRAM shown can be stored in Figure 5 The storage can be distributed across multiple memories, including MEM, STORAGE, and EXT-MEM. Alternatively, or additionally, storage can also be cloud-based.

[0101] The invention has been described above using examples. It is readily understood that various changes and modifications can be made without departing from the scope of the invention.

Claims

1. A method for controlling a conveyor section for a single item of goods, wherein... The conveyor section comprises multiple interconnected sub-conveyor sections, each driven by a driver. - One or more sensors for detecting individual items are located on at least some sub-conveyor sections. The control of the driver is performed using a machine learning model via a computing unit. The machine learning model repeatedly receives input data comprising vectors of fixed length, wherein each vector element is assigned to a segment of the transport segment and indicates the current occupancy percentage of a single item in the corresponding segment, wherein the transport segment is divided into multiple segments of equal size.

2. The method according to claim 1, wherein The machine learning model receives updated input data using a clock, and Output information about the speed to be set for each sub-transport segment at the same clock.

3. The method according to claim 1 or 2, wherein The current occupancy rate is determined based on the sensor measurements and the speed of the sub-conveying section.

4. The method according to claim 3, wherein Determining the current occupancy percentage of a specific individual item over time is achieved by a sensor detecting the item and assigning the measurement results to one or more corresponding segments containing the corresponding location. The next sensor uses the speed of the corresponding sub-conveyor segment to calculate the position of the single item, and assigns the calculation result to one or more corresponding segments containing the corresponding position. The next sensor detects the single item and assigns the measurement results to one or more corresponding sections containing the corresponding location.

5. The method according to claim 1 or 2, wherein Each vector element is represented by a non-binary value indicating the current occupancy percentage.

6. The method of claim 5, wherein the non-binary value is a value between 0 and 1.

7. The method according to claim 1 or 2, wherein the input data further comprises: - Current information relating to at least one position of at least one part on the last sub-transport segment. and / or - Information related to the current speed of the sub-conveyor segment, and / or - The current measurement result of the sensor.

8. The method of claim 1 or 2, wherein the machine learning model is trained before controlling the transport segment. For the purpose of training, the machine learning model is pre-fed input data including a vector of fixed length, wherein the vector elements are determined based on simulation of the transport section.

9. The method according to claim 8, wherein The training is performed as reinforcement learning, wherein an objective function is used to determine the reward or punishment, the objective function comprising: -A single item arrives at a pre-defined location on the final sub-transport segment. and / or Multiple individual items collided on the conveyor section. and / or - The speeds of adjacent sub-transport sections are similar. and / or - Maintain the minimum or target distance between two individual items.

10. An apparatus for data processing, comprising means for performing the method according to any one of claims 1 to 9.

11. A system for data processing, comprising means for performing the method according to any one of claims 1 to 9.

12. A computer program product comprising a computer program including instructions that, when executed by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 9.

13. A computer-readable storage medium having a computer program, the computer program including instructions that, when executed by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 9.

Images