CN116805160A - Method for training a neural network to determine a feature vector representing the wear state of a device - Google Patents

Abstract

A method for training a neural network to determine a feature vector representative of a wear state of a device. According to various embodiments, a method for training a neural network to determine a feature vector representative of a wear state of a device is described, the method comprising detecting sensor data elements; forming pairs of sensor data elements; determining a time difference between the pair of sensor data elements; processing the pair of sensor data elements into a feature vector; and processing the feature vectors by a regression network for estimating time differences between the sensor data elements of the pair and adapting the neural network and the regression network to reduce losses, the larger the difference between the determined time differences and the estimated time differences of the pair.

Description

Method for training a neural network to determine a feature vector representing the wear state of a device

Technical Field

The present disclosure relates to a method for training a neural network to determine feature vectors representative of a wear state of a device.

Background

In order to operate a technical system having one or more devices, such as a production line, efficiently and reliably, it is desirable to monitor the wear of one or more devices, for example, in order to replace the components in time, but also not early. A foreseeable solution is to use artificial intelligence. However, training of the corresponding machine learning model often requires the use of expert knowledge, e.g. the training data has to be provided with labels, which means high costs.

Thus, an efficient scheme for training a machine learning model for monitoring wear of equipment is desirable.

Disclosure of Invention

According to various embodiments, there is provided a method for training a neural network to determine a feature vector representing a state of wear of a device, the method comprising, for each point in time in a sequence of points in time, detecting a sensor data element belonging to the point in time, the sensor data element containing information about the state of the device at the point in time or about the state of a product produced by the device at the point in time; and pairs of sensor data elements, wherein each pair has a first sensor data element of the sequence of sensor data elements belonging to a point in time at a first point in time and a second sensor data element of the sequence of sensor data elements belonging to a point in time at a second point in time. The method further comprises determining, for each pair, a time difference between a point in time to which the first sensor data element belongs and a point in time to which the second sensor data element belongs; processing the first sensor data element into a first feature vector through a neural network; processing the second sensor data element into a second feature vector by the neural network; and processing the first feature vector and the second feature vector together by a regression network for estimating a time difference between a point in time to which the first sensor data element belongs and a point in time to which the second sensor data element belongs from the first feature vector and the second feature vector. The method comprises adapting the neural network and the regression network for reducing losses, the larger the difference between the determined time difference and the estimated time difference of the pair, the larger the losses.

Since this is penalized for training the determined loss when the regression network incorrectly estimates the relative aging (time difference of the sensor data elements), the output of the neural network, which can be regarded as a feature vector, by co-training the neural network and the regression network contains information about the wear of the device, which can be regarded as the actual chronological aging of the device. Thereby enabling training of a neural network that can be used for estimating wear of the device (or even the whole system consisting of the device). This is achieved without having to pretrain the neural network or the expert having to equip the training data with tags.

Embodiment 2 is the method of embodiment 1, comprising selecting, for each of at least some of the pairs, one of the sensor data elements belonging to a point in time between the first point in time and the second point in time, wherein, at

The sum of the time difference estimated by the regression network for the first sensor data element and the selected sensor data element (i.e. for the pair consisting of the first sensor data element and the selected sensor data element) and the time difference estimated by the regression network for the selected sensor data element and the second sensor data element (i.e. for the pair consisting of the selected sensor data element and the second sensor data element) and

time difference estimated by the regression network for the pair

The larger the deviation between them, the larger the loss.

Thus, the neural network is trained such that the neural network is in a time relationship between sensor data elements. Such intermediate points may be selected for many such pairs and considered accordingly in the loss. This ensures a consistent estimation of the regression network and also ensures consistent eigenvectors produced by the trained neural network in terms of wear represented by the estimation. In the above example, for example, the first point in time is located before the second point in time, and the selected point in time is located between the two points in time.

It should be noted that in the above expression "the higher the certain value, the greater the loss", this means that if all other values on which the loss depends remain the same, the loss rises together with the value (i.e. irrespective of even the value rising, a certain relation may lead to a total drop in loss, for example because the estimation in total on average corresponds better to the tag, even for a pair of the above-mentioned deviations increases).

Embodiment 3 is a method for determining a wear state of a device, the method comprising training a neural network according to embodiment 1 or 2; detecting another sensor data element containing information about the current state of the device or about the current state of a product produced by the device, delivering the other sensor data element to a trained neural network and deriving the wear state of the device from a feature vector output by the neural network in response to the other sensor data element.

The other sensor data element is, for example, the current sensor data element, and thus the trained neural network may be used to determine (estimate) the current wear state.

Embodiment 4 is the method of embodiment 3, comprising deriving the wear state from a distance between a feature vector output by the neural network in response to the further sensor data element and a reference feature vector for a reference wear state of the device.

Thus, the regression network no longer needs to be used to determine the wear state and can operate directly on the generated feature vectors.

Embodiment 5 is the method of embodiment 4, the method comprising detecting a reference sensor data element, delivering the reference sensor data element to a trained neural network, and determining a reference feature vector from feature vectors output by the neural network in response to the reference sensor data element.

For example, the reference feature vector or even a plurality of reference feature vectors may be formed by averaging or clustering of feature vectors output by the neural network in response to the reference sensor data elements. This improves the robustness of the estimation, as a plurality of reference sensor data elements (e.g. measurements) are used as a basis for the reference.

Embodiment 6 is a wear monitoring system set up for performing the method according to any one of embodiments 1 to 5.

Embodiment 7 is a computer program having instructions which, when executed by a processor, cause the processor to perform the method according to any one of embodiments 1 to 5.

Embodiment 8 is a computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform the method of any one of embodiments 1 to 5.

Various embodiments are described below.

Embodiment 1 is a method for controlling a robot as described above.

Drawings

In the drawings, like reference numerals generally refer to the same parts throughout the different views. The drawings are not necessarily to scale, emphasis instead generally being placed upon illustrating the principles of the application. In the following description, various aspects are described with reference to the following drawings.

Fig. 1 shows a production device according to an embodiment.

FIG. 2 illustrates a machine learning model according to one embodiment.

Fig. 3 shows an example of a wear curve.

Fig. 4 shows a flow chart illustrating a method for training a neural network to determine feature vectors representative of a wear state of a device.

Detailed Description

The following detailed description refers to the accompanying drawings, which illustrate specific details and aspects of the disclosure in which the application may be practiced. Other aspects may be utilized and structural, logical, and electrical changes may be made without departing from the scope of the present application. Because some aspects of the disclosure may be combined with one or more other aspects of the disclosure to form new aspects, the different aspects of the disclosure are not necessarily mutually exclusive.

Various examples are described in more detail below.

Fig. 1 shows a production apparatus 100 according to an embodiment.

One or more components 101 are fed to a production line 102 and processed into a final product 104 by means of one or more machines or tools 103 (here symbolically by means of a robotic arm). The apparatus of one or more machines or tools 103 is referred to below as a manufacturing apparatus, which may have one or more manufacturing devices (machines, robots, etc.).

As the manufacturing apparatus 103 ages (e.g., edging wear or tool dulling) during its operation, the quality of the end product 104 being manufactured generally also decreases over time. In order to replace the manufacturing device or its components before the quality of the manufactured end product declines too strongly, a prediction-based maintenance (english: predictive maintenance (predictive maintenance)) may be used, i.e. to observe how the quality of the manufactured product 104 deteriorates over time due to ageing of the manufacturing device, and to predict when the quality will become too low and then intervene in time.

For this purpose, it would be desirable to determine a slowly rising curve by observing the quality (e.g. a plurality of quality features) of the end product 104, wherein a higher value describes a greater wear or lower quality of the end product, and then possibly trigger an alarm from a specific value (quality limit) which causes the manufacturing device or a respective component of the manufacturing device, which is responsible for the respectively observed quality feature of the end product, to be replaced.

Such a curve is also called a wear curve (english is often called "drift curve"). Similarly, a decreasing "quality curve" may also be determined.

Such wear curves may be determined in various ways. A model can be used as a basis (however, the model is usually not very accurate and must be built first if necessary). Alternatively, training can be performed based on wear data so far (i.e. observations of reduced quality) by means of supervised learning of a machine learning model (ML model) that predicts wear at a later time.

However, the wear data is typically only present in sufficient amounts for the beginning (early) of the wear curve, since it is common in production to actively and thus very early replace the manufactured components in order not to produce a final product with defects. Furthermore, before using the wear data for prediction by means of the ML model, it has to be specified (at least) up to which point in time the wear data are collected.

There are also solutions based on supervised learning in order to generate wear curves by means of ML models. One example is the use of image data with a pre-trained convolutional network (e.g., imagenet). Thus, the ML model can learn the distribution (e.g., represented by clustering or Gaussian mixture model) of features (English: features) of the initial image (showing the final product in a "good state") and determine the distance of the features of the later recorded image for this purpose. However, such pre-trained networks do not exist for all types of data, and it is costly to collect sufficient training data and tasks for pre-training itself.

According to various embodiments, means are provided for monitoring technical systems, for example manufacturing devices (in particular for determining wear curves), which enable the use of existing data, avoid the expense of equipping training input data with target data (i.e. tags for supervised learning, i.e. labels), for example even in the case of self-supervised learning, and achieve a good generalization (in particular compared to solutions requiring explicit tags).

According to various embodiments, the quality of the end product 104 is monitored by means of images 106, which are provided by one or more cameras 105 and evaluated by a monitoring system 107 (e.g. a computer or a control unit).

For example, the camera records one or more images 106 for each manufactured end product 104 after its manufacture (before the end product is further processed or delivered), i.e., one or more images 106 showing the state of the end product 104 in which the end product has been manufactured (i.e., immediately after its manufacture, for example).

The camera 105 recording the images 106 or the monitoring system 107 itself is provided with a time stamp for each image 106. The timestamp indicates the time at which the image 106 was recorded and thus the time at which the product 104 was manufactured (i.e., made).

The monitoring system may determine the temporal distance between two images 105 (e.g., from a common time series) by comparing the time stamps, one in about seconds.

The monitoring system 107 uses this information for training the ML model 108.

FIG. 2 illustrates a machine learning model 200, for example implemented as the ML model 108 of the monitoring system 107, according to one embodiment.

The ML model 108 has a twin neural network 201, i.e. the same neural network 202 (in the sense that the same architecture is used twice with the same parameters, in particular weights) is applied to a first image 203 (time stamp t) and a second image 204 (time stamp t+n, e.g. n seconds). Thus, the neural network 202 generates the associated feature vectors 205, 206 for the two images 203, 204, respectively. This is done not only in training but also when using the trained ML model 200. Typically, neural networks are feature extraction models. For example, a convolutional network may be used, in particular for processing images.

In training, the second neural network 207 (i.e., not the twin neural network 201, which will be referred to hereinafter as the first neural network for clarity) is trained to estimate a time difference (i.e., n) from the two feature vectors 205, 206 (e.g., a cascade thereof). The second neural network 207 is accordingly also referred to as a regression network (for time differences). For this training, the labels n are available for each (training) pair of images (the difference in the time stamps of the two images and the regression network are trained according to the monitored learning with these labels.

This results in the ML model 200 learning abstract representations (feature vectors) based on the time stamps of the data (i.e. images) in training and by comparing feature vectors belonging to images with different time stamps (i.e. generated by the first neural network 201 from images with different time stamps) can obtain information about the time differences or estimates of the expected time differences between the images (i.e. between the time stamps of the images).

Training may be improved by using characteristics or relationships of time distances, such as

D(X2，X0)＝D(X1，X0)+D(X1，X2)。

Where D (,) represents the temporal distance between two arguments, and X0, X1 and X2 are the images (or generally (sensor) data elements) in that temporal order. Then, the machine learning model is trained such that it estimates D (X2, X0), D (X1, X0) and D (X1, X2) for the group consisting of three pairs (X2, X0), (X1, X2) and the loss function used for training has a term that penalizes D (besrafen) if D (X2, X0) is different from D (X1, X0) +d (X1, X2).

Based on the trained ML model 200, the monitoring system 107 may calculate the wear curve in various ways. In the simplest case, the monitoring system estimates for the current image 106 the time difference between the current image 106 and the initial image (of the end product 104 manufactured when the manufacturing apparatus has not worn) from the feature vectors provided by the first neural network 201 for the current image and the initial image. The time difference thus determined accounts for wear as reflected in the feature vector and can be considered (in view of the observed wear) as the expected time difference. If the expected time difference is higher than the actual time difference (the difference in the time stamps), the monitoring system 107 may interpret this as an indication of a dramatic increase in wear (aging) and react accordingly (e.g., facilitate replacement of the component). In other words, the ratio of the actual past time difference between the two images to the time difference estimated by the ML model 200 can be analyzed. If more time than estimated has elapsed, the manufacturing apparatus ages to a degree below the average level. If less time than estimated has elapsed, the manufacturing apparatus ages very strongly (and must therefore be replaced earlier if necessary). These results may be of interest to the data scientist or even engineers at the production line, but may also be used directly by the monitoring system automatically for control.

The monitoring system 107 may also implement anomaly detection. This indicates an anomaly if, for example, the expected time difference for the end product (at the initial or reference time) is, for example, much greater than the expected time difference for other end products (manufactured at similar times). Likewise, if the end product manufactured later has a smaller expected time difference than the end product (strongly), this also indicates an anomaly. The monitoring system 107 may then react accordingly (e.g., pick out the end product).

Instead of determining the time difference between the current image 106 and the initial image, the monitoring system 107 may calculate for multiple or one cluster (e.g. one or more average feature vectors) and combine the results if necessary.

The ML model 200 may be trained without the need for a pre-trained neural network for this purpose. In addition, no expert is required to mark the data. The feature vector provides an abstract representation and enables calculation of an expected time difference (i.e., a time difference for expected wear over time).

The feature vectors 205, 206 are abstract representations that are particularly sensitive to things that change from image to image over time. Such a representation may be used for many tasks, i.e. for example also for classification, or for detecting abnormal measurements, clusters, etc.

After training, the monitoring system 107 may determine the temporal distance of the two images, i.e., as output by the regression network 207, using the complete ML model 200. The monitoring system can thus determine the wear curve, in particular, for the image sequences recorded for the respective production device 103. To this end, the monitoring system may set the wear of the first image of the sequence to a value of zero. For each other image, the monitoring system feeds the feature vectors of the first image and the other images to the regression network 207 and uses the time distance estimated by the regression network 207 as the Y value of the wear curve (relative to the X value given by the time stamp of the other images).

Alternatively, after training, the monitoring system 107 may operate even without the regression network 207 by calculating distances between feature vectors, which are elements of the feature space, from the metrics of the feature space.

For example, the monitoring system 107 determines a model (e.g. a gaussian mixture model) describing the initial state (e.g. the distribution of feature vectors of the first five images of the sequence of images recorded for the manufacturing apparatus) and calculates the feature vector of the other recorded (current) image, e.g. the distance from the average of the model describing the initial state, purely based on the feature vector (e.g. by means of cosine distances in feature space). These distances (from the time series of images) are arranged in rows to derive a wear curve. The monitoring system 107 can in turn use the wear profile for detecting abnormal measurements or for triggering maintenance (e.g. replacement of components of the corresponding technical system, e.g. manufacturing equipment).

Fig. 3 shows an example of a wear curve.

Each point represents an image 106, where the x-coordinate of the point describes the timestamp of the image and the y-coordinate of the point describes the wear reflected by the image. This wear may be (see also description above):

the distance of the feature vector determined for the image from the representation of the "good state" (typically the reference state) in the feature space. The good state may be an average of the feature vectors of the recorded first image. The monitoring system may also for example apply a method for clustering, such as a DBSCAN (Density-Based Spatial Clustering of Applications with Noise (spatial clustering of noisy applications based on Density)) algorithm, in order to obtain a well-conditioned cluster, then calculate the cosine distance of the feature vector of the current image from each core point and select the minimum of these distances as the wear of the current image.

The wear of the first image of the sequence is set to zero and for each other image the wear is set as the distance of the feature vector relative to the first image, e.g. the distance estimated by the regression model 207 (illustrated by arrow 201 in fig. 2)

The distance used as a feature vector for wear may consist of a plurality of distances, for example according to the above mentioned formula D (X2, X0) =d (X1, X0) +d (X1, X2). For example, the monitoring system may perform this calculation for the current image (feature vector X2) for many intermediate images (feature vector X1 respectively) and average over all results D (X2, X0) and set the wear of the current image to this average.

From the wear curve, the engineer may then deduce from which distance level from the original image he/she wants to be notified, for example in order to get an impression of wear by himself or in order to change the tool.

In the above example, the wear of the technical system 107 is checked indirectly, i.e. by means of the image 106 of the manufactured end product 104. Another example for such an indirect check would be, for example, also the state of the camera (for example in a vehicle) which is derived by checking the quality of the images recorded by means of the camera. However, it is also possible to check the state of the technical system directly, i.e. by means of the technical system itself by means of images (or generally sensor data), for example measurements of currents in the installation, images of the surfaces of the devices, etc.

In summary, according to various embodiments, a method is provided, as shown in fig. 4.

Fig. 4 shows a flowchart 400 illustrating a method for training a neural network to determine feature vectors representative of a wear state of a device.

In 401, for each point in time in the sequence of points in time, a sensor data element belonging to the point in time is detected, said sensor data element containing information about the state of the device at the point in time or about the state of a product produced by the device at the point in time.

In 402, pairs of sensor data elements are formed, wherein each pair has a first one of the sensor data elements belonging to a first time point of the sequence of time points and a second one of the sensor data elements belonging to a second time point of the sequence of time points.

In 403, for each pair,

in 404, determining a time difference between the point in time to which the first sensor data element belongs and the point in time to which the second sensor data element belongs (i.e. determining the determined time difference);

processing the first sensor data element into a feature vector of the first sensor data element (this feature vector is referred to as a first feature vector) by a (first) neural network in 405;

processing the second sensor data element into a feature vector of the second sensor data element (this feature vector is referred to as a second feature vector) by the (first) neural network in 406; and

in 407, the first feature vector is processed together with the second feature vector by a (neural) regression network (i.e. by a second neural network) for estimating from the first and second feature vectors a time difference between the point in time the first sensor data element belongs to and the point in time the second sensor data element belongs to (i.e. determining the estimated time difference (time difference estimate)).

In 404, the neural network and the regression network are adapted to reduce the losses, the larger the difference between the determined and estimated time differences of the pair (i.e. the difference between the determined and estimated time differences for the respective pair, respectively, i.e. the respective estimates of the time differences; the difference may be determined for each pair).

As is usual, a batch of training data elements (here pairs of sensor data elements) can be formed for this purpose, by means of which a loss function containing terms corresponding to the losses is evaluated, and the parameters (in particular the weights) of the neural network (the neural network determining the feature vectors and the regression network) are adapted by means of gradient descent in order to reduce the losses.

The neural network may be regarded as a feature detector or a feature detector network.

The method of fig. 4 may be performed by one or more computers having one or more data processing units. The term "data processing unit" may be understood as any type of entity that enables processing of data or signals. For example, data or signals may be processed in accordance with at least one (i.e., one or more) particular function performed by the data processing unit. The data processing unit may comprise or be constructed from analog circuitry, digital circuitry, logic circuitry, a microprocessor, a microcontroller, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), a Digital Signal Processor (DSP), a programmable gate array (FPGA) integrated circuit, or any combination thereof. Any other party for implementing the functions described in more detail herein may also be understood as a data processing unit or a logic circuit arrangement. One or more of the method steps described in detail herein may be performed (e.g., implemented) by a data processing unit through one or more specific functions performed by the data processing unit.

The sensor data (contained in the sensor data elements) contains information about the status of the device (directly or indirectly, for example about the status of the end product). Instead of the image 106 as in the above embodiments, sensor signals from different sensors, such as video, radar, liDAR (laser radar), ultrasound, motion, thermal imaging, audio, etc., may be received and processed in order to obtain data regarding the status of the device.

In particular, data from the time series (from sensors, for example, measurements of the flow) can be analyzed and the corresponding technical system operated accordingly. Here, each (sensor) data element (e.g. each measurement or each image) is provided with a time stamp, which describes the point in time at which the data element was detected. To train a machine learning model (composed of a neural network and a regression network), training data records are first collected, which are composed of a number of example data elements (obtained from a row, series, or sequence of recorded data samples). According to various embodiments, it is assumed that the first data element (having an earlier time in the timestamp) represents a "good" or "new" state (reference state), while the last data element (having a later time in the timestamp) is already shortly before the "bad" or "dangerous" point in time. Each training data element of the training data record consists of such an example data element and a label given by the difference of the time stamps.

The sensor data may be sensor data from the manufactured product (and thus indirect information about the device or system) or sensor data from the device or system itself (i.e. direct information about this). For example, the analysis provides anomaly detection. In response to an anomaly, the technical system may be placed in a secure mode, or the cause of the anomaly may be identified and countermeasures taken. For example, the user may be notified, for example, as a measure to change a tool or to adapt a parameter of the machine.

The time difference estimated by means of the machine learning model can also be used to reduce the data traffic, for example from the production line to the back-end computer, in such a way that the sensor data element is only forwarded if it is sufficiently different from the previous sensor data element (thus expressed as the expected time difference being above the limit). The back-end system may then use the forwarded sensor data elements, for example, as training data for the ML model.

Control signals for the robotic device may be generated from the output of the neural network. The term "robotic device" may be understood to relate to any technical system (with a mechanical part whose movement is controlled), such as a manufacturing apparatus (production line), a computer-controlled machine, a vehicle, a household appliance, an electric tool, a production machine, a personal assistant or an access control system or a system for providing information, such as a monitoring system. Accordingly, the monitoring system 107 from the above-described embodiments can be understood as an example and the above-described functionality can be provided by any control device in the respective application context (if necessary using sensor data different or additional from the image 106 and with respect to a different technical system than the production device 100).

Although specific embodiments have been illustrated and described herein, it will be appreciated by those of ordinary skill in the art that a wide variety of alternate and/or equivalent implementations may be substituted for the specific embodiments shown and described without departing from the scope of the present application. Any adaptations or variations of the specific embodiments discussed herein are intended to be encompassed by the present application. Therefore, it is intended that this application be limited only by the claims and the equivalents thereof.

Claims (8)

1. A method for training a neural network to determine feature vectors representative of a wear state of a device, the method comprising:

for each point in time in the sequence of points in time, detecting a sensor data element belonging to the point in time, the sensor data element containing information about the state of the device at the point in time or about the state of a product produced by the device at the point in time;

forming pairs of sensor data elements, wherein each pair has a first sensor data element of the sequence of sensor data elements belonging to a time point at a first time point and a second sensor data element of the sequence of sensor data elements belonging to a time point at a second time point,

for each pair of the two pairs,

determining a time difference between a point in time to which the first sensor data element belongs and a point in time to which the second sensor data element belongs;

processing the first sensor data element into a first feature vector through a neural network;

processing the second sensor data element into a second feature vector by the neural network; and

processing the first feature vector and the second feature vector together by a regression network for estimating a time difference between a point in time to which the first sensor data element belongs and a point in time to which the second sensor data element belongs from the first feature vector and the second feature vector;

the neural network and the regression network are adapted for reducing losses, the greater the difference between the determined time difference and the estimated time difference of the pair.

2. The method of claim 1, comprising selecting for each of at least some of the pairs one of the sensor data elements belonging to a point in time between the first point in time and the second point in time,

wherein the larger the deviation between the sum of the time differences estimated by the regression network for the first sensor data element and the selected sensor data element and the time differences estimated by the regression network for the selected sensor data element and the second sensor data element and the time differences estimated by the regression network for the pair, the larger the loss.

3. A method for determining a wear state of a device, the method comprising:

training a neural network according to claim 1 or 2;

detecting a further sensor data element, the further data element containing information about the current state of the device or about the current state of a product produced by the device,

the further sensor data element is transmitted to a trained neural network and the wear state of the device is derived from a feature vector output by the neural network in response to the further sensor data element.

4. A method according to claim 3, comprising deriving the wear state from a distance between a feature vector output by the neural network in response to the further sensor data element and a reference feature vector for a reference wear state of the device.

5. The method of claim 4, comprising detecting a reference sensor data element, delivering the reference sensor data element to a trained neural network, and determining a reference feature vector from feature vectors output by the neural network in response to the reference sensor data element.

6. A wear monitoring system set up for performing the method according to any one of claims 1 to 5.

7. A computer program having instructions which, when executed by a processor, cause the processor to perform the method of any one of claims 1 to 5.

8. A computer readable medium storing instructions which, when executed by a processor, cause the processor to perform the method of any one of claims 1 to 5.