Computer-implemented method and system for optimizing a clustering of a plurality of input data

A computer-implemented method for checking a clustering result generated by a machine-learning algorithm for a plurality of input data to be checked, the method comprising the following steps:

• capturing, by an imaging sensor system comprising one or more imaging sensors, a plurality of images representing a manufacturing component associated with the plurality of input data to be checked, the plurality of images forming a dataset to be checked;
• extracting, by a computing device that includes at least one processor and for each respective one of the plurality of images, a respective multi-dimensional numerical feature vector comprising a plurality of numerical components that represent features of the respective image in a latent feature space, thereby forming a plurality of feature vectors;
• reducing, by the computing device, a dimensionality of the plurality of feature vectors by:
• computing from the plurality of feature vectors a dimensionality-reduction transformation that maps the feature vectors into a reduced-dimension feature space whose dimensionality is lower than a dimensionality of the latent feature space;
• and
• applying the dimensionality-reduction transformation to the feature vectors, thereby generating, for respective ones of the plurality of images, reduced-dimension feature representations that have fewer numerical components than the extracted plurality of feature vectors;
• clustering, by the computing device and in the reduced-dimension feature space, the reduced-dimension feature representations via a clustering algorithm that groups unlabeled data based on inter-data similarity, so as to obtain a check result comprising a plurality of clusters, each cluster including reduced-dimension feature representations corresponding to respective ones of the plurality of images;
• evaluating, by the computing device for each of the plurality of clusters of the check result, a cluster-quality metric based on distances among the reduced-dimension feature representations assigned to that cluster, the cluster-quality metric indicating a quality of clustering of the respective cluster;
• performing sorting by removing, from the dataset for subsequent processing, the images whose reduced-dimension feature representations are assigned to at least one of the clusters whose respective cluster-quality metric is determined, according to a predefined comparison rule, to exceed the cluster-quality metrics of one or more other ones of the clusters;
• iteratively repeating, by the computing device, the dimensionality-reduction, clustering, evaluating, and sorting steps, each iteration, subsequent to a first one of the iterations, being performed exclusively for remaining, unsorted images of the dataset obtained after a preceding one of the sorting steps, (i) until all of the plurality of images are clustered or (ii) until a predetermined termination condition is satisfied;
• and
• comparing, by the computing device, the check result obtained from the iteratively repeated clustering with the clustering result generated by the machine-learning algorithm, to determine whether the clustering result generated by the machine-learning algorithm is valid for the plurality of input data to be checked.

A system for inspecting a manufacturing component, the system comprising:

• an imaging sensor system, wherein the imaging sensor system is configured to capture a plurality of images representing the manufacturing component to be inspected, the plurality of images forming a dataset;
• and
• a computing device that includes at least one processor, wherein the computing device is configured to:
• extract, for each respective one of the plurality of images, a respective multi-dimensional numerical feature vector comprising a plurality of numerical components that represent features of the respective image in a latent feature space, thereby forming a plurality of feature vectors;
• reduce a dimensionality of the plurality of feature vectors by:
• computing from the plurality of feature vectors a dimensionality-reduction transformation that maps the feature vectors into a reduced-dimension feature space whose dimensionality is lower than a dimensionality of the latent feature space;
• and
• applying the dimensionality-reduction transformation to the feature vectors, thereby generating, for respective ones of the plurality of images, reduced-dimension feature representations that have fewer numerical components than the extracted plurality of feature vectors;
• cluster, in the reduced-dimension feature space, the reduced-dimension feature representations via a clustering algorithm that groups unlabeled data based on inter-data similarity, so as to obtain a plurality of clusters, each cluster including reduced-dimension feature representations corresponding to respective ones of the plurality of images;
• evaluate, for each of the plurality of clusters, a cluster-quality metric based on distances among the reduced-dimension feature representations assigned to that cluster, the cluster-quality metric indicating a quality of clustering of the respective cluster;
• perform sorting by removing, from the dataset for subsequent processing, the images whose reduced-dimension feature representations are assigned to at least one of the clusters whose respective cluster-quality metric is determined, according to a predefined comparison rule, to exceed the cluster-quality metrics of one or more other ones of the clusters;
• iteratively repeat the dimensionality-reduction, clustering, evaluating, and sorting, wherein each iteration, subsequent to a first one of the iterations, is performed exclusively for remaining, unsorted images of the dataset obtained after a preceding one of the sorting steps, until all of the plurality of images are clustered or until a predetermined termination condition is satisfied;
• and
• provide, based on one or more clusters to which the reduced-dimension feature representations corresponding to the images are assigned after completion of all of the iterations, an inspection output indicating a defect classification or inspection status for the manufacturing component represented by the images.

A non-transitory computer-readable data carrier on which are stored program code of a computer program for inspecting a manufacturing component, the program code being executable by a computer and, when executed by the computer, causing the computer to perform the following steps:

• capturing, by an imaging sensor system comprising one or more imaging sensors, a plurality of images representing the manufacturing component to be inspected, the plurality of images forming a dataset;
• extracting, by a computing device that includes at least one processor and for each respective one of the plurality of images, a respective multi-dimensional numerical feature vector comprising a plurality of numerical components that represent ing features of the respective image in a latent feature space, thereby forming a plurality of feature vectors;
• reducing, by the computing device, a dimensionality of the plurality of feature vectors by:
• computing from the plurality of feature vectors a dimensionality-reduction transformation that maps the feature vectors into a reduced-dimension feature space whose dimensionality is lower than a dimensionality of the latent feature space;
• and
• applying the dimensionality-reduction transformation to the feature vectors, thereby generating, for respective ones of the plurality of images, reduced-dimension feature representations that have fewer numerical components than the extracted plurality of feature vectors;
• clustering, by the computing device and in the reduced-dimension feature space, the reduced-dimension feature representations via a clustering algorithm that groups unlabeled data based on inter-data similarity, so as to obtain a plurality of clusters, each cluster including reduced-dimension feature representations corresponding to respective ones of the plurality of images;
• evaluating, by the computing device for each of the plurality of clusters, a cluster-quality metric based on distances among the reduced-dimension feature representations assigned to that cluster, the cluster-quality metric indicating a quality of clustering of the respective cluster;
• performing sorting by removing, from the dataset for subsequent processing, the images whose reduced-dimension feature representations are assigned to at least one of the clusters whose respective cluster-quality metric is determined, according to a predefined comparison rule, to exceed the cluster-quality metrics of one or more other ones of the clusters;
• iteratively repeating, by the computing device, the dimensionality-reduction, clustering, evaluating, and sorting steps, each iteration, subsequent to a first one of the iterations, being performed exclusively for remaining, unsorted images of the dataset obtained after a preceding one of the sorting steps, (i) until all of the plurality of images are clustered or (ii) until a predetermined termination condition is satisfied;
• and
• providing, by the computing device and based on one or more clusters to which the reduced-dimension feature representations corresponding to the images are assigned after completion of all of the iterations, an inspection output indicating a defect classification or inspection status for the manufacturing component represented by the images.