Method and device for determining a geotechnical parameter of an excavated material

Abstract

The invention relates to a method for determining at least one geotechnical parameter of an excavated material from a tunnel-boring machine, wherein: the excavated material is fed onto a conveyor device (2) for transporting the excavated material; an impact body (5) is arranged in a conveyor flow (3) of the excavated material so that it is at least partly immersed in the excavated material; the excavated material in the conveyor flow (3) has a flow speed which is either predetermined or measured; a measured variable corresponding to a force exerted by the excavated material of the conveyor flow (3) on the impact body (5) is measured; a time sequence of measured values ​​of the measured variable is recorded and the at least one geotechnical parameter is determined from a portion of the sequence of measured values ​​by inputting the portion of the time sequence of measured values ​​into a data processing apparatus and processing it using this apparatus such that at least one output value corresponding to the at least one geotechnical parameter is output; and the processing of the measured values ​​is based either on the predetermined flow speed or a value for the measured flow speed is input and taken into account when the measured values are processed.

Description

[0001] Method and device for determining a geotechnical parameter of an excavated material

[0002] The invention relates to a method for determining at least one geotechnical parameter of an excavated material of a tunnel boring machine.

[0003] The type and composition of the excavated material from tunnel boring machines (TBMs) naturally depend on the nature of the rock (rock or soil) at the face, as well as on the type of advance and the substances added to the excavated material.

[0004] In soft, cohesive soils, tunnel boring machines with earth pressure support, so-called earth pressure shield tunnel boring machines, are preferred. In earth pressure shields, a slurry made of excavated material and water serves as the plastic support medium. A tool-equipped, rotating cutting wheel of the TBM is pressed against the face and loosens the in-situ soil. This excavated material enters the excavation chamber through openings, where it mixes with the existing earth pressure. A screw conveyor transports the excavated material from the floor of the excavation chamber onto a conveyor belt. The interaction of the screw conveyor rate and the advance speed allows the support pressure of the earth pressure to be precisely controlled. Earth pressure sensors in the excavation chamber continuously monitor the pressure conditions. This allows the machine operator to optimally coordinate all excavation parameters, even under changing geological conditions.This enables the necessary equalization of the pressure conditions at the tunnel face, prevents uncontrolled penetration of the soil into the part of the tunnel boring machine that is under atmospheric pressure and thus creates the conditions for rapid and low-settlement advance.

[0005] Not all soils in their natural state possess ideal properties for earth pressure balance shield excavation. However, the method's application range can be expanded through soil conditioning. This involves modifying the plastic deformability, consistency, and water permeability of the geology by adding various conditioning agents, such as water, clay, or foam. Earth pressure balance shields can thus achieve good excavation performance even in heterogeneous soils with varying gravel, sand, or water content, or in inherently unstable rock formations.

[0006] However, if too much water, clay, and / or foam is added to the excavated material, the excavated material, for example, transported from the screw conveyor to the conveyor belt, is difficult to dispose of due to its fluid consistency; it would have to be temporarily stored in pits or settling tanks, for example, or further processed. This may lead to increased costs. Furthermore, the use of landfills is often subject to a requirement that the excavated material to be disposed of must have a minimum shear strength of 10 kN / m. 2 must not be undercut.

[0007] In these and other applications, it is therefore desirable to determine geotechnical, particularly rheological, parameters of the excavated material extracted from the excavation chamber during ongoing operations, such as viscosity and yield point, (wing) shear strength, or slump (according to DIN EN 12350-2). For example, determining the slump (according to DIN EN 12350-2) requires taking a sample from the excavated stream and – as part of a slump test – filling a truncated cone-shaped metal mold of specified dimensions, lifting the metal mold at a defined height, and measuring the difference between the height of the metal mold and the height of the slumped excavated cone after lifting. This takes time and cannot be fully automated; in particular, continuous data acquisition is not possible.The same applies to the determination of alternative geotechnical parameters, such as slump, vane shear strength or dynamic viscosity and yield point.

[0008] The object of the invention is therefore to automatically determine geotechnical parameters of the excavated material extracted from the excavation chamber during ongoing operation. This object is achieved according to the invention by a method for determining at least one geotechnical parameter of an excavated material of a tunnel boring machine, having the features of claim 1, or by a device for determining at least one geotechnical parameter of an excavated material produced in a tunnel boring machine, having the features of claim 14.

[0009] The determination of at least one geotechnical parameter is preferably carried out automatically.

[0010] In the method according to the invention for determining at least one geotechnical parameter of excavated material from a tunnel boring machine, the excavated material is placed onto a conveyor device for transporting the excavated material. An impact body is arranged in a conveying stream of excavated material such that it is at least partially immersed in the excavated material, wherein the excavated material in the conveying stream has a flow velocity that is either predetermined or measured.Then, a measured variable corresponding to a force exerted by the excavated material of the conveying flow on the impact body is measured, wherein a temporal sequence of measured values ​​of the measured variable is recorded, and the at least one geotechnical parameter is determined from a section of the sequence of measured values ​​by inputting the section of the temporal sequence of measured values ​​to a data processing device and processing it by the latter in such a way that at least one output value corresponding to the at least one geotechnical parameter is output, wherein the processing of the measured values ​​is either based on the predetermined flow velocity or a value for the measured flow velocity is input and taken into account when processing the measured values.

[0011] The data processing device preferably implements a trained neural network comprising multiple layers of nodes, wherein all connections between the nodes are assigned weights and the nodes are assigned predefined activation functions, wherein the weights have been set in advance using a training method. A matrix of input data is obtained from the segment of the sequence of measured values, and the matrix of input data is input to an input layer of nodes of the neural network. The at least one output value corresponding to the at least one geotechnical parameter is output by at least one node of an output layer of the neural network.The advantage here is that measurements can be taken directly at the existing flow rate and that the desired geotechnical parameter, such as the slump, is automatically provided after a very short data processing time, which makes it possible, for example, to adjust and thus regulate the amount(s) of water and / or foam and / or other additives influencing the rheological properties of the excavated material depending on the specific geotechnical parameter(s) or to classify the excavated material itself according to the geotechnical parameter.

[0012] In one embodiment, the predefined activation function is, for example, a reLll activation function (reLll(x) = max(0,x)) or, alternatively—as a differentiable approximation of the reLU activation function—the function f(x) = ln(1 + ex). Initial values ​​are assigned to the weights before the start of the training procedure, preferably small randomly chosen initial values, for example, random numbers uniformly distributed in the interval (-0.05, 0.05) or normally distributed around zero. The training procedure uses, for example, the squared error as the loss function and a gradient descent method to adjust the weights.

[0013] In one embodiment of the method for determining at least one geotechnical parameter, the excavated material flow could be guided through a pipe in which the impact body, for example, a rod extending into the pipe, is mounted. However, in a preferred embodiment of the method, the impact body is arranged in an open flow of excavated material. This is preferably the open-top flow on a conveyor belt.

[0014] In a preferred embodiment of the method for determining at least one geotechnical parameter, a predetermined flow velocity of the conveyed stream is set and kept constant during the acquisition of the temporal sequence of measured values ​​of the measured variable. This simplifies data processing, since no measured value-time pairs need to be processed, but only measured value sequences. The prerequisite is that the set flow velocity is equal to the flow velocity that was set when obtaining the pairings of measured value sequence and measured geotechnical parameter used in the training process.

[0015] Preferably, the temporal sequence of measured values ​​is recorded with a constant sampling frequency. This also simplifies data processing, since only one value for the sampling frequency or the time intervals between the measuring points needs to be processed or assumed during processing. In a preferred embodiment of the method for determining at least one geotechnical parameter, the section of the temporal sequence of measured values ​​corresponds to a measurement duration of 0.25 s to 60 s, preferably from 0.5 s to 1 s. Practical tests with typical compositions and flow velocities of the excavated material have shown that the essential movement sequences of the impact body relevant for determining the geotechnical parameter, in particular the slump, and thus the essential force curves to be recorded take place within these time intervals.

[0016] In one embodiment of the method for determining at least one geotechnical parameter, a predetermined number of characteristic features are calculated from the segment of the temporal sequence of measured values, and these features are used as a matrix of the input data. From the segment of the temporal sequence of measured values, also referred to as the measurement sequence, characteristic features such as mean, median, maximum, minimum, and / or standard deviation, or even Fourier coefficients of a discrete Fourier transform, are calculated, arranged in a multidimensional matrix, and transferred to the input layer in a standardized form.

[0017] In another simple embodiment of the method for determining at least one geotechnical parameter, the section of the temporal sequence of measured values, the measurement sequence, is used in a normalized manner as a one-dimensional matrix of the input data.

[0018] It is conceivable for the data processing device to implement the neural network using special hardware comprising a plurality of circuit elements emulating neurons. However, a preferred embodiment of the method for determining at least one geotechnical parameter is characterized in that the neural network is simulated by the data processing device under program control (a so-called artificial neural network). This means that the data processing device performs mathematical operations that emulate the function of the neuron layers of a neural network.

[0019] The data processing unit can simulate different types of neural networks, for example, feedforward networks (including multi-layer perceptrons), convolutional networks, or recurrent neural networks. In a preferred development of this method for determining at least one geotechnical parameter, the data processing unit simulates a feedforward neural network with a one-dimensional convolutional neural network. By using convolutional neural networks (CNNs), the number of computational operations to be performed is significantly reduced compared to a fully linked network. At the same time, locally close neurons—in this case, temporally close measured values ​​of a measurement sequence—have a greater influence on the neurons of a subsequent layer.

[0020] In a currently preferred embodiment of the method for determining at least one geotechnical parameter, the data processing device simulates a neural convolutional network with a first convolutional layer followed by a pooling layer and a second convolutional layer followed by several fully linked layers. It has been shown that such a neural network design can, on the one hand, avoid overfitting and, on the other hand, achieve good predictions.

[0021] Preferred embodiments of the method for determining at least one geotechnical parameter are characterized in that, before entering the section of the temporal sequence of measured values ​​for determining the at least one geotechnical parameter, the weights in the training method are adjusted by a) for a predetermined selection of a plurality of excavated material samples of different compositions with different geotechnical parameters: a1) the at least one geotechnical parameter, for example the slump, of the excavated material sample is measured or determined and stored as a target parameter value, a2) the impact body is arranged in a conveying flow of the excavated material sample in such a way that it is at least partially immersed in the excavated material, wherein the excavated material in the conveying flow has a predetermined flow velocity, and a measured variable,which corresponds to a force exerted by the excavated material of the flow of the excavated material sample on the impact body, is measured, wherein a temporal sequence of measured values ​​of the measured variable is recorded, a3) pairs of a section of the sequence of measured values ​​are stored in association with the target parameter value, b) initial values ​​are assigned to the weights in the neural network implemented by the data processing device, c) for a pairing of a section of the sequence of measured values ​​and the associated target parameter value: c1) the at least one geotechnical parameter is determined from the section of the sequence of measured values ​​by processing the section of the temporal sequence of measured values ​​by the data processing device implementing the neural network in such a way that at least one output value corresponding to the at least one geotechnical parameter is output,c2) then the difference between the output value and the associated target parameter value is determined and c3) the weights are changed depending on the difference, wherein steps c1) to c3) are repeated for a next pairing from a section of the sequence of measured values ​​and associated target parameter value, wherein the selection of the pairings is carried out in random order, wherein steps c1) to c3) are carried out for a predetermined proportion of the pairings stored in step a), which serves as the training set.

[0022] An application of the said method for determining at least one geotechnical parameter is, according to the invention, a method for adjusting the consistency of a conditioned excavated material of an earth pressure shield tunnel boring machine, in which a conditioning agent, for example a surfactant foam, is added to the excavated material for its conditioning in the excavation chamber, the excavated material conditioned in this way is conveyed out of the excavation chamber by conveying devices, wherein in an open conveying stream formed on one of the conveying devices, at least one geotechnical parameter, preferably a slump, is determined automatically according to one of the above-mentioned methods, and the quantity of conditioning agent to be added is regulated depending on the geotechnical parameter thus determined,preferably by increasing or decreasing the amount of conditioning agent to be added depending on a difference between the geotechnical parameter thus determined and a target value of the geotechnical parameter.

[0023] One application of the said method for determining at least one geotechnical parameter is, according to the invention, a method for classifying excavated material from a tunnel boring machine, wherein the conditioned excavated material is conveyed out of the excavation chamber by means of at least one conveying device, wherein in an open conveying flow formed on one of the conveying devices at least one geotechnical parameter, preferably a slump, is determined automatically according to one of the above-mentioned methods, and the excavated material is classified as a function of the determined geotechnical parameter, in that the excavated material is preferably classified as a function of a difference between the geotechnical parameter thus determined and a target value of the geotechnical parameter.

[0024] The device according to the invention for determining at least one geotechnical parameter of an excavated material accumulating in a tunnel boring machine comprises (a) a conveying device for transporting the excavated material in such a way that a conveying flow of the excavated material is formed in an open conveying channel at a flow velocity, wherein the conveying device has a device for setting a predetermined flow velocity and / or a device for measuring the flow velocity, (b) an impact body arranged in the open conveying channel, which is at least partially immersed in the conveying flow of the excavated material, (c) a measuring device connected to the impact body for measuring a measured variable corresponding to a force exerted by the excavated material of the conveying flow on the impact body, (d) a device coupled to the measuring device for recording sample values ​​of the measured variable,for converting the sample values ​​of the measured variable into digital measured values ​​and for outputting a temporal sequence of digital measured values, (e) a storage device for temporarily storing the sequence of digital measured values, and (f) a data processing device coupled to the storage device, which is configured to process the portion of the sequence of digital measured values ​​and output at least one output value corresponding to the at least one geotechnical parameter.

[0025] It is advantageous that the data processing device is configured in such a way that it generates and processes a matrix of input data from the section of the sequence of digital measured values ​​and outputs at least one output value corresponding to the at least one geotechnical parameter and takes into account either the set predetermined flow velocity or a measured value of the flow velocity during the processing, wherein the data processing device implements a trained neural network having several layers of nodes, wherein all links between the nodes are assigned weights and the nodes are assigned predetermined activation functions, wherein the weights have been set in advance by means of a training method,wherein the matrix of input data is input to an input layer of nodes of the neural network and the at least one output value corresponding to the at least one geotechnical parameter is output to at least one node of an output layer of the neural network.

[0026] The impact body can be of any shape and is introduced into the flow of excavated material in such a way that a force dependent on the rheological properties of the excavated material acts on the impact body. The impact body can, for example, be a rod with a round cross-section extending into the flow. In a preferred embodiment of the device for determining at least one geotechnical parameter, the impact body is designed as a sphere. Such an impact body has the advantage that it does not need to be arranged in a specific orientation in the flow. A plow-like body is also an advantageous embodiment.

[0027] In a preferred device for determining at least one geotechnical parameter, the impact body is a steel ball weighing 500 g to 2000 g, preferably 800 g to 1000 g. Such a steel ball has, for example, a diameter of approximately 6 cm and a weight of approximately 900 g. At typical flow velocities (in the range between 1 m / s and 3 m / s) and compositions of the conveyed flow, a steel ball dimensioned in this way is, on the one hand, sufficiently heavy so that it does not "bounce" excessively on the surface of the conveyed flow of excavated material, but, on the other hand, not too heavy so that it remains in a constant position virtually unaffected by the conveyed flow.

[0028] The impact body, for example in the form of a sphere, can be mounted on a fixed support, such as a rod, and extend into the conveying stream. For example, a strain gauge of the measuring device could record the force-dependent deflection of such a rod.

[0029] However, in a preferred embodiment of the device for determining at least one geotechnical parameter, a spherical impact body is attached to a pivotably suspended rod or a traction cable, and a sensor for detecting the measured variable is coupled to the rod or the traction cable, wherein the measured variable corresponds to a tensile force absorbed by the rod or the traction cable. Suspending the impact body on a pivotably suspended rod or a traction cable is advantageous because it allows the impact body to deflect when it encounters a relatively large body (e.g., stone) carried in the excavated material flow.

[0030] In a preferred device for determining at least one geotechnical parameter, the data processing device is configured to simulate a feedforward neural network in the form of a one-dimensional convolutional network. Preferably, the data processing device is configured to simulate a one-dimensional convolutional neural network with a first convolutional layer followed by a pooling layer and a subsequent second convolutional layer with several subsequent fully connected layers. This has the advantages mentioned above in the context of the corresponding preferred embodiment of the method.

[0031] Advantageous and / or preferred developments of the invention are characterized in the subclaims.

[0032] The invention is described in more detail below with reference to a preferred embodiment illustrated in the drawings. In the drawings:

[0033] Figure 1 is a schematic diagram of the measuring arrangement of a preferred device for determining at least one geotechnical parameter,

[0034] Figure 2 shows an exemplary representation of a series of measurements and a section of the series of measurements representing a measurement sequence, and

[0035] Figure 3 is a schematic diagram illustrating the neural network simulated by a data processing device.

[0036] Figure 1 schematically shows a rheological measuring arrangement 1 used in the device according to the invention for determining at least one geotechnical parameter, in particular an index parameter such as the slump.

[0037] The measuring arrangement 1 is used to determine a geotechnical parameter, such as in particular the slump, in a conveying stream 3, wherein the conveying stream 3 moves at a predetermined speed in a conveying channel 2. The movement of the conveying stream 3 is indicated by the arrow 4.

[0038] For example, the conveyor channel 2 is located on a conveyor belt of a tunnel boring machine. The velocity of the conveyor flow 3 is in the range between 1 and 3 m / s, preferably in the range of 1.8 m / s to 2.5 m / s.

[0039] An impact body suspended from a traction cable 6, here preferably in the form of a steel ball 5, is immersed at least partially in the conveying stream 3. A force transducer 7, which measures the force acting on the traction cable 6, is located on the suspension of the traction cable 6 on a support with a crossbeam 8. For typical excavated materials that move at a speed between 1.8 m / s and 2.5 m / s in the conveying channel 2, a steel ball with a diameter of 6 cm and a weight of approximately 900 g is preferably used. The force transducer 7 is connected to an evaluation device (not shown in Figure 1) for evaluating the sensor output signal, which evaluation device samples the output signal of the force transducer 7 at a predetermined sampling frequency and outputs and / or temporarily stores the sampled values ​​thus acquired. In the experimental setup, for example, the sampling frequency is 200 Hz, so that a force sample value is recorded and stored at intervals of 5 ms.

[0040] Figure 2 shows in the upper diagram the series of measurements recorded in a test setup, in which the sample values ​​were recorded for a duration of, for example, about 60 seconds, whereby, for example, the sample values ​​recorded in the first 20 seconds, which were obtained during the start-up of the test setup, were discarded.

[0041] The series of force measurements obtained at intervals of 5 ms was divided into 61 measurement sequences of, for example, 128 samples each and thus of approximately 0.64 s duration.

[0042] The lower diagram in Figure 2 shows an example of a sensor signal curve during the time interval of 0.64 s in which the 128 samples were acquired. The choice of 128 samples per sequence is also advantageous because a power of two facilitates computational processing, for example, in a discrete Fourier transform to extract characteristic features from the measurement sequence.

[0043] The 128 samples of a measurement sequence were input to the nodes of an input layer of a neural network after the neural network had been trained in a manner described in more detail below. A correctly trained neural network then outputs a value for a geotechnical parameter corresponding to the input measurement sequence of 128 samples; in the preferred embodiment, this is a value for the slump value. In alternative embodiments, other and multiple geotechnical parameter values ​​can also be output.

[0044] For training, i.e., supervised learning of the neural network, it is necessary to provide training data that includes pairs of a measurement sequence of 128 force values ​​and a corresponding value of the measured geotechnical parameter, i.e., in this case, the slump. To generate the training data for the neural network, a preferred embodiment proceeded as follows. First, samples of excavated materials of varying composition and consistency were provided—here, for example, 13 samples—which were characterized by a range of different slump values ​​in the relevant range.

[0045] For each sample, the geotechnical parameter, in this case, the slump, was measured. The samples were successively placed in a test setup in which the flow rate of the sample was adjusted to a specified conveying speed of 1.8 m / s or 2.5 m / s.

[0046] An impact body in the form of a steel ball attached to a traction cable was introduced into the conveying stream and, after the arrangement had settled down at the specified speed, a series of force sample values ​​was recorded and stored, as shown as an example in the upper illustration of Figure 2.

[0047] For each of the 13 samples, for example, measurement series of the type shown were generated, preferably three measurement series here. Each of the three measurement series was in turn divided into 61 measurement sequences of 128 samples each, so that 183 measurement sequences, a total of 13 x 183 = 2379 measurement sequences, were generated and stored for each of the 13 sample excavated materials. The measured geotechnical parameter, the slump, of a sample was stored in association with the measurement sequence recorded for that sample. Thus, 2379 pairs of measurement sequences with 128 force values ​​and a corresponding slump value were stored.

[0048] From the total of 2,379 measurement sequence-slump pairings, approximately 2,000 pairings were randomly selected as training pairings for the neural network. A remaining set of approximately 300 randomly selected pairings served as validation datasets to test the quality and suitability of the trained neural network.

[0049] The neural network was trained using the squared error and gradient descent as the loss function.

[0050] The weights of the neural network were initially preset to an arbitrarily small value. Then, the 128 sample values ​​representing the measured force curve within a time interval of 0.64 s were input to an input layer of the neural network. Based on the preset weights, an output value was obtained at an output layer that was intended to correspond to the geotechnical parameter. The difference between the geotechnical parameter determined by the neural network and the actual value of the geotechnical parameter measured for the sample was then determined. Depending on these differences between the parameter values ​​output by the neural network and the actual measured values ​​of the samples, the weights were adjusted in a known manner. This process was repeated for all pairings of force sample values ​​of a measurement sequence and the geotechnical parameter (slump) measured for the corresponding sample.The training pairings were entered in random order.

[0051] The details of supervised learning performed using a squared error loss function and the gradient descent method are known to a person skilled in the art of neural networks and need not be described in detail here.

[0052] The following describes the structure of an exemplary neural network, which was also used in the experimental setup. Figure 3 shows a schematic representation of the neural network simulated by the calculations of the data processing device, in which a value for a geotechnical parameter (here, a slump) is determined from a measurement sequence of 128 force samples. This artificial neural network initially comprises the 128 nodes of an input layer shown in the top row of Figure 3. The calculations shown below simulate a convolution layer.

[0053] In a convolution, a convolution window (also called a filter, kernel, or convolution kernel) of a given width of nine weight values, shown hatched in Figure 3, is linked to the values ​​of nine input nodes. Each weight value is multiplied by the value of a node, and the nine resulting products are summed (vector scalar product). The result is evaluated with an activation function and stored in a memory location of a so-called feature map.

[0054] The convolution window is then advanced by one node of the input layer (to the right in Figure 3), so that the nine weight values ​​of the convolution window are then linked to the values ​​of nodes 2–10 of the input layer (multiplied in each case, and the results summed). In Figure 3, in the row below the input layer, a first convolution window of width 9 is shown, initially below nodes 1–9 of the input layer. The result of this scalar product is stored in the first field in the first row of the feature map displayed below the convolution windows. This is represented by an arrow connecting the first convolution window to the result field.

[0055] To the right of the first convolution window, a first convolution window is shown in dashed lines, which has been moved eleven steps further. This illustrates that in this position the nine weights of the first convolution window are linked to the values ​​of nodes 11 - 19 of the input layer, with the result being stored in the eleventh field of the first row of the feature map.

[0056] Finally, on the far right, another dashed representation of the first convolution window is shown, which is positioned so that its weight values ​​are linked to nodes 120 - 128 of the input layer, with the result being stored in the 120th field of the first row of the feature map, which is again indicated by a corresponding arrow.

[0057] In the preferred embodiment shown in Figure 3, a total of 64 different convolution windows or filters are used in the first convolution layer, as indicated by the numbering 1, 2, and 64 to the left of the filter blocks. Accordingly, all convolution operations generate results that are stored in a feature map with 120 fields in 64 rows each. These operations simulate a first one-dimensional convolution layer of the neural network. The convolution layer is one-dimensional, with a one-dimensional convolution window being moved over a one-dimensional input layer.

[0058] Before further processing, the values ​​stored in the feature map are preferably evaluated using a so-called activation function. This evaluation occurs either during their calculation (dot product) before storage in the feature map or subsequently for the entire calculated feature map. Both approaches are equivalent for a simulated neural network.

[0059] In the preferred embodiment shown here, a so-called reLll activation function is used, i.e. reLll(x) = Max(0,x). This means that all negative values ​​are replaced by zero, while the positive values ​​remain unchanged. In order to reduce the calculation time and the computational effort, the data volume of 120 x 64 elements stored in the feature map is preferably reduced in a so-called pooling layer by obtaining a new element from a predetermined number of elements of the feature map. In the preferred embodiment shown in Figure 3, an average is calculated from 10 elements in each row of the feature map and this average is temporarily stored (in a further node layer). This averaging of 10 elements in each row is represented by a curly bracket with an arrow as an example for the first 10 elements of the first 1st, 2nd and 64th row.

[0060] This results in a total of 12 x 64 = 768 averages. Due to the averaging process, this operation is also called average pooling. It would also be conceivable to have a neural network that uses so-called max pooling in its pooling layer instead of average pooling, in which only the maximum value is selected from the given number of output values ​​(also known as subsampling).

[0061] This pooling is followed by another convolution layer, as shown in Figure 3. During the subsequent convolution, 64 different convolution windows, each with a width of 9, are moved sequentially across the rows of pooling results of 12 elements each. Each time a convolution window with a width of 9 is moved across a row of 12 elements, four result elements are formed. The 64 different convolution windows or filters are moved sequentially (or in parallel) across a row of 12 elements, with the first, second, third, and fourth result elements of the 64 filters being summed.

[0062] This is repeated for each of the 64 rows of pooling results, resulting in a total matrix of 4 x 64 elements, as shown in Figure 3. The results of these convolution operations are also evaluated with an activation function, preferably again a reLU activation function. The 4 x 64 evaluated convolution results are then flattened into a vector of 256 elements. This vector is represented in Figure 3 as a node layer of 256 nodes.

[0063] This is followed by three further layers of nodes of the neural network, each with 64, 32, and 16 nodes, respectively. These four hidden layers with 256, 64, 32, and 16 nodes, shown at the bottom of Figure 3, are fully connected layers, meaning that each node is linked to every node in the next layer, although these links are not fully shown in Figure 3. Each link between two nodes is assigned a weight. Each of the nodes shown in Figure 3 also implements an activation function, preferably the reLU activation function. The last layer with 16 nodes is followed by an output layer, which in the example shown in Figure 3 only comprises one output node for outputting a value, here, for example, the slump. No activation function is assigned to this node.

[0064] During supervised learning, initial values ​​must be assigned to each of the nine weights of the 64 convolutional windows of the first convolutional layer, each of the nine weights of the 64 windows of the second convolutional layer, as well as to the 256 x 64 + 64 x 32 + 32 x 16 + 16 = 18,960 weights of the fully connected layers and the weights of additional bias neurons, and then these values ​​must be adjusted using a suitable method.

[0065] Of course, many alternative configurations for suitable neural networks are conceivable. For example, instead of 64 filters in the convolutional layers, only 16 filters could be used, while simultaneously omitting pooling after the first convolutional layer. When configuring the network, it is always important to balance the number of learnable parameters (weights) between being too small to achieve good prediction results and too high, as this could lead to overfitting of the neural network. The neural network structure described here using Figure 3 is the result of numerous experiments with different neural networks.

Claims

Patent claims 1. A method for determining at least one geotechnical parameter of excavated material from a tunnel boring machine, wherein the excavated material is fed onto a conveyor device (2) for transporting the excavated material, an impact body (5) is arranged in a conveying stream (3) of the excavated material such that it is at least partially immersed in the excavated material, wherein the excavated material in the conveying stream (3) has a flow velocity that is either predetermined or measured, a measured variable corresponding to a force exerted by the excavated material of the conveying stream (3) on the impact body (5) is measured, wherein a temporal sequence of measured values ​​of the measured variable is recorded, and the at least one geotechnical parameter is determined from a section of the sequence of measured values ​​by inputting the section of the temporal sequence of measured values ​​to a data processing device and processing it by the latter in such a way,that at least one output value corresponding to the at least one geotechnical parameter is output, whereby the processing of the measured values ​​is based either on the specified flow velocity or a value for the measured flow velocity is entered and taken into account when processing the measured values.

2. Method for determining at least one geotechnical parameter according to claim 1, characterized in that the impact body is arranged in an open conveying flow of the excavated material.

3. Method for determining at least one geotechnical parameter according to claim 1 or 2, characterized in that a predetermined flow velocity of the conveying stream is set and kept constant during the recording of the temporal sequence of measured values ​​of the measured variable.

4. Method for determining at least one geotechnical parameter according to one of claims 1 to 3, characterized in that the temporal sequence of measured values ​​is recorded with a constant sampling frequency.

5. Method for determining at least one geotechnical parameter according to one of claims 1 to 4, characterized in that the section of the temporal sequence of measured values ​​corresponds to a measuring duration of 0.25 s to 60 s, preferably of 0.5 s to 1 s.

6. Method for determining at least one geotechnical parameter according to one of claims 1 - 5, characterized in that the data processing device implements a trained neural network which has a plurality of layers of nodes, wherein all links between the nodes are assigned weights and the nodes are assigned predetermined activation functions, wherein the weights have been set in advance by means of a training method, wherein a matrix of input data is obtained from the section of the sequence of measured values ​​and the matrix of input data is input to an input layer of nodes of the neural network and wherein the at least one output value corresponding to the at least one geotechnical parameter is output by at least one node of an output layer of the neural network.

7. A method for determining at least one geotechnical parameter according to claim 6, characterized in that a predetermined number of characteristic features are calculated from the section of the temporal sequence of measured values ​​and these features are used as a matrix of the input data.

8. A method for determining at least one geotechnical parameter according to claim 6, characterized in that the section of the temporal sequence of measured values ​​is used in a standardized one-dimensional matrix of the input data.

9. Method for determining at least one geotechnical parameter according to one of claims 6 - 8, characterized in that a neural network is simulated by the data processing device in a program-controlled manner, or in that a neural feedforward network with a convolutional network is simulated by the data processing device, or in that a neural convolutional network with a first convolutional layer with a subsequent pooling layer and a subsequent second convolutional layer with a plurality of subsequent fully linked layers is simulated by the data processing device.

10. A method for determining at least one geotechnical parameter according to one of claims 6 - 9, characterized in that before entering the section of the temporal sequence of measured values ​​for determining the at least one geotechnical parameter, the weights are set in the training method by a) for a predetermined selection of a plurality of excavated material samples of different compositions with different geotechnical parameters: a1) the at least one geotechnical parameter of the excavated material sample is measured or determined and stored as a target parameter value, a2) the impact body is arranged in a conveying flow of the excavated material sample in such a way that it is at least partially immersed in the excavated material, wherein the excavated material in the conveying flow has a predetermined flow velocity, and a measured variable,which corresponds to a force exerted by the excavated material of the conveying flow of the excavated material sample on the impact body, is measured, wherein a temporal sequence of measured values ​​of the measured variable is recorded, a3) pairs of a section of the sequence of measured values ​​are stored in association with the target parameter value, b) initial values ​​are assigned to the weights in the neural network implemented by the data processing device, c) for a pairing of a section of the sequence of measured values ​​and associated target parameter value: c1) the at least one geotechnical parameter is determined from the section of the sequence of measured values ​​by processing the section of the temporal sequence of measured values ​​by the data processing device implementing the neural network in such a way that, at least one output value corresponding to the at least one geotechnical parameter is output, c2) the difference between the output value and the associated target parameter value is then determined and c3) the weights are changed as a function of the difference, wherein steps c1) to c3) are repeated for a next pairing from a section of the sequence of measured values ​​and associated target parameter value, wherein the selection of the pairings takes place in a random order, wherein steps c1) to c3) are carried out for a predetermined proportion of the pairings stored in step a), which serves as a training set.

11. Method for determining at least one geotechnical parameter according to one of claims 1 - 10, characterized in that a settlement of the excavated material is determined as the geotechnical parameter.

12. A method for adjusting the consistency of a conditioned excavated material of an earth pressure shield tunnel boring machine, wherein a conditioning agent is mixed into the excavated material for its conditioning in the excavation chamber, the conditioned excavated material is conveyed out of the excavation chamber by conveying devices, wherein in an open conveying stream formed on one of the conveying devices at least one geotechnical parameter, preferably a slump, is determined automatically according to a method according to one of claims 1 - 11, and the amount of conditioning agent to be added is regulated as a function of the geotechnical parameter thus determined.

13. A method for classifying excavated material from a tunnel boring machine, wherein the conditioned excavated material is conveyed from the excavation chamber by at least one conveying device, wherein in an open conveying stream formed on one of the conveying devices at least one geotechnical parameter, preferably a slump, is determined automatically according to a method according to one of claims 1-11, and the excavated material is classified depending on the determined geotechnical parameter.

14. Device for determining at least one geotechnical parameter of an excavated material produced in a tunnel boring machine, in particular for applying a method according to one of claims 1 to 11, comprising: a conveying device for transporting the excavated material in such a way that in an open conveying channel (2) a conveying flow (3) of the excavated material is produced at a flow velocity is formed, wherein the conveying device has a device for setting a predetermined flow velocity and / or a device for measuring the flow velocity, an impact body (5) arranged in the open conveying channel (2), which is at least partially immersed in the conveying flow (3) of the excavated material, a measuring device (7) connected to the impact body (5) for measuring a measured variable which corresponds to a force exerted by the excavated material of the conveying flow (3) on the impact body (5), a device coupled to the measuring device (7) for acquiring sample values ​​of the measured variable, for converting the sample values ​​of the measured variable into digital measured values ​​and for outputting a temporal sequence of digital measured values, a storage device for temporarily storing the sequence of digital measured values ​​and a data processing device coupled to the storage device, which is configured in such a waythat it processes a portion of the sequence of digital measured values ​​and outputs at least one output value corresponding to the at least one geotechnical parameter.

15. Device for determining at least one geotechnical parameter according to claim 14, characterized in that the data processing device is configured to generate and process a matrix of input data from the section of the sequence of digital measured values ​​and to output at least one output value corresponding to the at least one geotechnical parameter and to take into account either the set predetermined flow velocity or a measured value of the flow velocity during the processing, wherein the data processing device implements a trained neural network having several layers of nodes, wherein all links between the nodes are assigned weights and the nodes are assigned predetermined activation functions, wherein the weights have been set in advance by means of a training method,wherein the matrix of input data is input to an input layer of nodes of the neural network and the at least one output value corresponding to the at least one geotechnical parameter is output to at least one node of an output layer of the neural network.

16. Device for determining at least one geotechnical parameter according to claim 14 or 15, characterized in that the impact body (5) is designed as a sphere, preferably as a steel sphere with a weight of 500 g to 2000 g, preferably 800 g to 1000 g.

17. Device for determining at least one geotechnical parameter according to claim 16, characterized in that the ball (5) is attached to a pivotably suspended rod or a traction cable (6) and a sensor (7) for detecting the measured variable is coupled to the rod or the traction cable (6), wherein the measured variable corresponds to a tensile force absorbed by the rod or the traction cable (6).

18. Device for determining at least one geotechnical parameter according to claim 15, characterized in that the data processing device is configured to simulate a feedforward neural network with a one-dimensional convolutional network, wherein preferably the data processing device is configured to simulate a convolutional neural network with a first convolutional layer with a subsequent pooling layer and a subsequent second convolutional layer with a plurality of subsequent fully linked layers.

Images