Improved Similarity Search for Industrial Component Models

Abstract

To provide a method for improving similarity search of an industrial component model.SOLUTION: A method includes following steps: each of a set of industrial component models (CM) has associated attributes, and embedding of similarity is embedding of at least some of the associated attributes; a particular CM as input (ICM) is input, and based on comparison between embedding of similarity and embedding of similarity of the ICM, output of a similarity request is a first subset of a model selected from the set of the CMs (ICMS1s); based on compatibility criteria of the ICM and a second subset of the CMs (ICMS2s), the ICMS2 selected by a user is received from ICMS1s of the ICs; a similarity attribute is associated with the ICM, and the attribute is a unique ID defined by the similarity request and / or the ICMS2 of the CM; and a new set of the embedding of similarity is calculated.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a computer-implemented method for improving the search for industrial component models. [Background technology]

[0002] When performing design work, whether on a mechanical object or an electronic circuit, it is generally preferable to use parts that exhibit interchangeability. Interchangeability means that a particular part can be replaced with another part in the design without changing the specifications required for the final product.

[0003] In that regard, identifying interchangeable parts is a challenging but important task to streamline procurement and improve design reliability by reusing parts with known operation.

[0004] For electronic components, this can be done by the manufacturer themselves by defining standardized data related to the part's fit, form, and function. For mechanical components, this is typically done by creating a signature for each part based on a topological analysis of the part's shape. Non-Patent Document 1 gives an example of one such method.

[0005] These methods all rely on theoretical models, which means that the way they work is more or less fixed unless you change the model itself. As a result, these methods offer little flexibility and little adaptability to situations that were not considered during modeling.

[0006] On an entirely different level, there are methods that take user feedback into account, and some methods use some form of clustering when performing a search. These techniques are performed as a "polishing step" that sorts or reshuffles certain search results based on user feedback. However, they are not embedded in the search itself; they only modify the result list. This means that the feedback is never actually taken into account in the search itself, and user feedback modifies part of the display process, not the object representation itself. As a result, the changes do not affect other applications, and in particular the subsequent clustering of the components that are the subject of the search. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] Kazdhan et al., article "Harmonic 3D Shape Matching," SIGGRAPH Sketches and Applications (2002) Summary of the Invention

[0008] The present invention aims to improve the situation. To this end, the applicant proposes a computer-implemented method for improving similarity searching for industrial component models, which includes: Providing a set of industrial component models, each of which has associated attributes, and a similarity embedding is an embedding of at least some of said associated attributes. receiving a similarity request using a particular industrial component model as input, the output of which is a first subset of industrial component models selected from the set of industrial component models based on a comparison between a similarity embedding and a similarity embedding of the input industrial component model; receiving a second subset of industrial component models selected by a user from the first subset of industrial component models based on a compatibility criterion between the input industrial component model and any industrial component model of the second subset of industrial component models; Associating a similarity attribute with the input industry component model, said attribute being a unique ID defined by said similarity requirement and / or said second subset of industry component models. Compute a new set of similarity embeddings.

[0009] This method is advantageous because it allows for the definition of new signatures that take into account not only traditional exchangeability data but also user feedback. This allows for the creation of compatibility classes, which effectively combine existing theoretical models with empirically induced clustering. Furthermore, because this method uses existing industrial component model attributes that were previously used as a foundational layer, it is easy to initialize the method, achieving the same performance level as existing methods from the start and a steep improvement curve as soon as user feedback begins to be incorporated. Furthermore, user feedback allows for information that is impossible to model to be taken into account, providing a more robust approach to fuzzy logic-type compatibility. In other words, the concept of exchangeability, which is necessarily limited to strict contexts, can be generalized to compatibility, which is exchangeability in more ambiguous contexts, without sacrificing the quality of the results.

[0010] In various embodiments, the method may exhibit one or more of the following features.

[0011] -Similarity embedding is achieved by vectorizing the attributes of industrial component models and embedding the resulting vectorized data.

[0012] The embedding is performed by a context-sensitive autoencoder, which includes an encoder and a decoder, which are neural networks:

[0013] -The input of the autoencoder is vectorized data, which constitutes the input of the encoder, and the output of the encoder constitutes the similarity embedding.

[0014] -The decoder input is the similarity embedding, and the decoder output is a term frequency-inverse document frequency of the vectorized data.

[0015] The encoder and decoder are adjusted so that the term frequency-inverse document frequency of the vectorized data best approximates the vectorized data.

[0016] The embedding is performed by performing a principal component analysis on the concatenation of the L1-regularized attributes with selected weights multiplied by the L1-regularized similarity attributes.

[0017] -Vectorization is performed by doc2vec vectorization of text attributes, and vectorization of similarity attributes, which involves adding a column for each unique ID, and for each industrial component model, filling this column with 1 if an industrial component model is associated with this unique ID, otherwise filling with 0.

[0018] -Vectorization is performed by applying the bidirectional encoder representation from the Transformer technique to the industrial component model.

[0019] The invention also relates to a computer program comprising instructions for carrying out the method according to the invention, a data storage medium having such a computer program recorded thereon, and a computer system comprising a processor coupled to a memory, and a memory having such a computer program recorded thereon.

[0020] Other features and advantages of the present invention are readily apparent from the following description of the drawings, which illustrate exemplary embodiments of the invention. [Brief explanation of the drawings]

[0021] [Figure 1] 1 is an overall view of a system according to the present invention; [Figure 2] FIG. 2 is an illustration of an exemplary embodiment of functions performed by the system of FIG. 1. [Figure 3] 1 is a table containing industrial component models of exemplary machine parts. [Figure 4] 1 is a table containing industry component models of exemplary electronic components. [Figure 5] FIG. 10 is an illustration of an exemplary embodiment of a function for computing similarity embedding. [Figure 6] FIG. 6 is an explanatory diagram showing an example of the structure of an autoencoder used in the function of FIG. 5; [Figure 7] 7A and 7B are diagrams illustrating exemplary results of vectors obtained after applying the functions of FIG. 5 to the industrial component models of FIG. 3 and FIG. 4, respectively, before feeding the autoencoder of FIG. 6. [Figure 8] 7A and 7B are diagrams illustrating exemplary results of vectors obtained after applying the functions of FIG. 5 to the industrial component models of FIG. 3 and FIG. 4, respectively, before feeding the autoencoder of FIG. 6.

[0022] The drawings and the following description are largely composed of positive and clearly defined features, and as a result, they are not only useful in understanding the invention, but can also be used to contribute to its definition if the need arises. DETAILED DESCRIPTION OF THE INVENTION

[0023] The description may refer to or use copyrighted or copyrightable material. Applicant has no objection to the reproduction of these materials, provided that such reproduction is limited to the required legal publications, but this should not be construed as a waiver of rights or any form of license.

[0024] 1 shows an overview of a system according to the present invention. The system 2 comprises a memory 4, an encoder 6 and a similarity search engine 8.

[0025] The memory 4 stores industrial component model data. The industrial component model data includes data that can define a mechanical or electronic component and attributes that can describe the industrial component model to a designer and allow the industrial component model to be searched. Furthermore, an industrial component model may be an assembly of parts that may or may not themselves have industrial component model data.

[0026] In the examples described herein, memory 4 may be realized in any suitable manner, i.e., by a hard disk drive, a solid state drive, flash memory, memory embedded in the processor, remote storage accessible from the cloud, etc.

[0027] In the example described herein, memory 4 associates a similarity embedding 10 with each industrial component model. This similarity embedding, described further below, allows for defining a measure of compatibility between the industrial component models under consideration. For a given industrial component model, the signature is not simply an aggregation or condensation of the industrial component model's attributes, but rather an element computed using user feedback information that allows for relating industrial component models in ways not possible with conventional techniques.

[0028] To compute the similarity embedding 10 according to the present invention, the system 2 uses an encoder 6 which is part of an autoencoder and which will be described in more detail with reference to Figure 6. As will be seen shortly, the similarity embedding constitutes an embedding of at least some of the attributes of the industrial component model.

[0029] In the examples described herein, the encoder 6 and the similarity search engine 8 are computer programs running on one or more processors, including any means known for performing automated computations, such as a CPU, a GPU, a CPU and / or GPU grid, a remote computation grid, a specially configured FPGA, a specially configured ASIC, a specialized chip such as a SOC or NOC, a dedicated AI chip, etc.

[0030] Based on the similarity embeddings calculated by the encoder 6, the similarity search engine 8 allows a user to input an industrial component model 12 and in return receive a set of industrial component models 14 that are deemed compatible with the input industrial component model 12.

[0031] The return also allows users to further improve the quality of requests by allowing them to specify a subset of industrial component models that they consider more relevant than other models within the returned set of industrial component models.

[0032] 2 depicts an exemplary embodiment of the functions performed by System 2. This exemplary embodiment represents all of the operations that System 2 may perform to implement the present invention.

[0033] As a result, the operation includes both a run-time portion, as shown in steps 210 to 260, and a training portion, as shown in step 270. The function begins with operation 200, where a function Inp() is executed. The function Inp() is an input function that receives the industrial component model ICM as an argument that is used as a global variable in the other steps.

[0034] This can be done by a human-machine interface (HMI), where the user specifies or uploads a file containing the input industrial component model attributes, or specifies an industrial component model entry ICM in memory 4. Any type of HMI can be used as long as it provides an interface where suggestions are presented to the user and the user can filter out irrelevant suggestions, verify relevant suggestions, and add industrial components. For example, the user can identify common characteristics in related suggestions, search for more industrial components with these characteristics, and select relevant results to add to the set.

[0035] Figures 3 and 4 show example tables containing attributes associated with industrial component models. The table in Figure 3 contains industrial component models for mechanical parts, while the table in Figure 4 contains industrial component models for electronic parts.

[0036] As shown in Figures 3 and 4, each row consists of the attributes of one part, and the attributes are divided into three types of columns: -A single referenced part ID column 30 that is a unique identifier for each part.

[0037] - A set of referenced feature columns 32 that describe predetermined features of each part.

[0038] - A single referenced similarity data column 34. are distributed to

[0039] For example, in the case of a mechanical part (Figure 3), these columns might include a description of the part, one or more columns detailing the material used in this part, one or more columns for storing the part's geometric signature, one or more columns for storing the dimensions of this part, one or more columns for storing relevant measurements of this part (e.g., hardness), etc. Similarly, in the case of an electronic part (Figure 4), these columns might include the part type (to describe what type of electronic function is achieved), a measurement column to describe nominal and operating values ​​and limit values, etc. All of these columns can be considered "objective features" because they describe general measurements or features that characterize the physical properties of the part.

[0040] The Similarity Data column may be empty. When filled, it contains a list of hash codes to which parts can be associated. That is, when two or more parts are deemed similar, a unique hash code is generated to represent the association between these parts, and this hash code is added to the "Similarity Data" column for each part. For example, looking at Figure 3, parts HF059210 and GJ154003 are both part of a set of parts deemed similar that received hash code 99089de4. Part HF059210 was further deemed similar in two other sets, designated by hash codes f74f0310 and d45c714e, respectively.

[0041] Operation 200 is followed by operation 210, in which a similarity search engine 8 is queried with the input industrial component model ICM of operation 200. The similarity search engine 8 performs a search of the input industrial component model and uses its similarity embeddings to determine several industrial component models that have close similarity embeddings. These industrial component models are returned in an initial set of industrial component models ICMS1. The calculation of a signature for a particular industrial component model is described in more detail with reference to FIGS. 5 and 6.

[0042] To compare the similarity embeddings of various industrial component models, the similarity search engine 8 may preferably use various measures, such as cosine similarity, or Euclidian similarity (or variations thereof, for example, after L1 or L2 normalization of the similarity embedding), or Manhattan similarity. The similarity search engine 8 is configured to return all industrial component models whose similarity embeddings and the above similarity measures applied to the similarity embedding of the input industrial component model exceed a given threshold. Alternatively, the similarity search engine 8 is configured to return a fixed number of industrial component models whose similarity embeddings are closest to the similarity embedding of the input industrial component model. The similarity search engine 8 can mix both approaches.

[0043] Note that for clarity, the similarity embedding is different from the similarity data of FIGS. 3 and 4, which will be described in more detail with respect to FIGS. 5 through 8.

[0044] In operation 220, a first set of industrial component models is presented to the user in a function Fdb() that receives a set of industrial component models ICMS1 for the user and returns a subset of industrial component models ICMS2 selected by the user because of their potential compatibility with the input industrial component model ICM.

[0045] In practice, via the function Fdb(), the user can provide feedback information by selecting a second set of industrial component models ICMS2 which are a subset of the first set of industrial component models ICMS1 and which contain elements which are deemed compatible from a design point of view with the input industrial component model ICM.

[0046] Steps 210 and 220 may be repeated to allow a user to enrich an existing query to add new industrial component models to the feedback. For example, a user may remove certain components from the results of a similarity query (as a result of a previous execution of operation 210), add other components to the results via a keyword query, and then validate the components remaining in the result set in a similar manner to the input.

[0047] In this case, these iterations are interpreted within the scope of the present invention as establishing a first set of industrial component models ICMS1 and reducing it to a second set of industrial component models ICMS2 that are a subset of the first set of industrial component models ICMS1.

[0048] 2 ends with the first set of industrial component models ICMS1 being returned as a result of the similarity search in operation 230. Otherwise, in operation 240, the function Hash() is executed. The function Hash() receives as input the second set of industrial component models ICMS2 and outputs a unique value, which is one of the hash codes described with reference to FIGS. 3 and 4, that is added to the "similarity data" column 34 of each of the industrial component models that are part of the second set of industrial component models ICMS2. The fact that this hash is unique is an important part of the definition of the similarity embedding of each industrial component model.

[0049] In operation 250, the function Upd_SS() is executed. The function Upd_SS() receives the current set of similarity embeddings, determines whether the similarity embeddings for the industrial component model in memory 4 need to be updated, and outputs this information. The decision that the similarity embeddings need to be updated may be user triggered, time triggered, dependent on the amount of new similarity data calculated since the last update, or based on another condition. In an alternative embodiment, this operation is completely separate from operations 200 through 240.

[0050] This operation has the advantage of allowing us to adjust the moment at which the autoencoder training is repeated. Such adjustment is important because training always takes a significant amount of time. As a result, it cannot be repeated too frequently, but it should be performed to take into account user feedback in the form of new similarity data added to the industrial component model.

[0051] If operation 250 indicates that the similarity embeddings do not need to be updated, then the functionality of FIG. 2 ends with returning a second set of industrial component models ICMS2 as the result of the similarity search in operation 260. In this case, the only change made is that the industrial component models of the second set of industrial component models ICMS2 have been modified to add the new similarity data calculated in operation 240. At this stage, the similarity embeddings attached to each industrial component model have not changed. This means that new queries based on the similarity embeddings do not show changes introduced by user feedback. Only when function Autoenc() is executed will the similarity embeddings, and consequently the query results, change.

[0052] If operation 250 indicates that the similarity embeddings need to be updated, the encoder 6 is called by the function Autoenc() to recalculate the similarity embeddings in operation 270, and the function of FIG. 2 ends with the return of a second set of industrial component models as a result of the similarity search in operation 260. The function Autoenc() receives all industrial component models (including the hash added in operation 240 as new similarity data) and calculates new similarity embeddings for each industrial component model. As will be seen in the description of FIGS. 5 through 8 below, the nature of the function Autoenc() ensures that the calculated similarity embeddings take into account the similarity data to provide similarity embeddings that reflect user feedback of compatibility.

[0053] Figure 5 discloses an example embodiment of the function Aotenc() of Figure 2 for computing affinity embeddings. This function includes two separate sets of operations for formatting and vectorizing attribute data prior to the machine learning operations that output affinity embeddings.

[0054] A first set of operations 500a is tailored to attributes such as text, where the attributes of a given industrial component model may be text fields 500 or category names 505.

[0055] In both cases, the data is normalized in operation 510 or 515. This normalization includes standard procedures such as stopword, accent or hyphenation removal, lemmatization, stemming, etc. Operations 510 and 515 are optional and can be omitted, especially for the category name 505 attribute. Optionally, phrase detection can also be used. The concept of phrase should be understood as the detection of multi-word expressions, where expressions replace words and become units of meaning. Phrase detection is very useful when vectorization is performed on a bag-of-words representation, where each expression becomes a dimension of the bag-of-words representation. Phrase detection may be omitted when vectorization is performed using doc2vec representation.

[0056] According to the first embodiment, vectorization is performed using a bag-of-words approach. In this case, after all text data is normalized, it undergoes a vocabulary reduction operation 520 to retain only the most frequent terms. Finally, the resulting tokens are vectorized using a bag-of-words representation. As is known, in a bag-of-words representation, each document is mapped to a space where each dimension of the space corresponds to a word in the vocabulary. The coordinates of a document in each dimension are typically the TFIDF (term frequency-inverse document frequency) of the words in the document. Alternatively, a count of the number of times a word occurs in a document can be used instead of the TFIDF. Similarity data is vectorized by creating a column that is added to the attributes of each hash code and filling it with 1 if a particular component contains a hash code in the similarity data column and 0 otherwise.

[0057] According to another embodiment, vectorization is performed using a neural network-based document embedding method.

[0058] The first neural network-based document embedding method that can be used is the Word2vec or Doc2vec method. In this case, operations 510 and 515 remain optional, and operation 520 is omitted. Word2vec is a neural network-based word vectorization that learns to generate word representations by predicting words from their neighbors (or vice versa) and fine-tuning the representation for each word in the process. Doc2vec is a similar algorithm, but adds a document representation to the input and also fine-tunes this representation in the process.

[0059] A second neural network-based document embedding method that can be used is the BERT (Bidirectional Encoder Representations from Transformers) technique. In this case, operations 510 and 515 remain optional, and operation 520 is omitted. BERT is an algorithm based on deep network construction, where the network learns which other words are important to look for in the context of the word for which it is building a representation. For example, if it represents "fire," it learns to look for "forest" or "engine" because the presence of these words changes the interpretation of "fire." BERT is typically trained on large text corpora, such as Wikipedia, to predict hidden words or predict the next word. Document vectorization is not the primary goal of BERT, but the output of certain layers of the BERT model can provide useful document vectors.

[0060] The second set of operations is tailored to attributes such as numeric values. In this case, a given industrial component model has a date field 540 or a number field 545. In the former case, the data undergoes vectorization in operation 550. After all numeric data has been vectorized, it undergoes a normalization operation 560, for example, based on a value divided by the variance of each vectorized numeric data.

[0061] Considering the above, it seems that the vectorization operation can be adjusted as needed: for example, text data can be vectorized using Doc2vec or BERT, category names can be vectorized using bag of words, etc.

[0062] 7 and 8 show exemplary results of the vectors obtained after applying these two sets of operations to the industrial component models of FIGS. 3 and 4, respectively.

[0063] After operations 530 and 560 are performed, the resulting vectorized data is fed into a CSAE (Context-Sensitive Autoencoder) to perform machine learning operation 570. Figure 6 shows an example structure of a CSAE that can be used for that purpose.

[0064] A CSAE is a network in the class of autoencoders or deep autoencoders. Deep autoencoders have at least three layers of neurons, each interconnected to the previous and next layers. Typically, the hidden layers are smaller than the input, and the input and output layers are the same size. An autoencoder is trained to reproduce the input. The autoencoder learns to retain as much information as possible from the hidden layer input in order to reconstruct the input as accurately as possible. Because the hidden layers are small, the input layer is a compressed representation of the input. CSAEs are unique in that they distinguish between main data and context data by assigning different weights to the reconstruction of the main data and context data in an error function. While the applicant has found CSAEs to be particularly advantageous in terms of performance for implementing the present invention, the use of CSAEs is not essential, and other types of autoencoders may be used.

[0065] CSAE6 includes an encoder and a decoder. The encoder and decoder are as follows: the input of the encoder is the vectorized data 50 output by operations 530 and 560, and the output of the encoder is the similarity embedding 10 for each industrial component model; The decoder input 60 is the similarity embedding 10 of each industrial component model, and the decoder output is an approximation 70 of the vectorized data 50; As such, it is a tuned neural network. example: If the vectorization method is based on a bag of words using TFIDF, the approximation 70 should be as close as possible to the input TFIDF of the input vector.

[0066] If the vectorization method is based on Doc2vec, the approximation 70 should be as close as possible to the input vector, which is the output of the Doc2vec neural network.

[0067] If the vectorization technique is based on BERT, the approximation 70 should be as close as possible to the input vector, which is the output of the BERT neural network.

[0068] The result of the CSAE6 used in the query is a similarity embedding 10. It is easy to see that the similarity embedding 10 is an abstract embedding that represents the dimensionality reduction of the input vector. As explained above, the CSAE6 could be replaced with another type of autoencoder that is not context sensitive. Alternatively, the autoencoder could be replaced with other means that can provide dimensionality reduction for vectorized data, such as PCA (Principal Component Analysis) or SVD (Singular Value Discrimination), which the applicants have determined to be particularly efficient and cost-effective.

[0069] Feedback introduced into the similarity data as part of this dimensionality reduction can gradually skew the similarity embedding toward empirical results. In other words, the more similarity data entered by the user through feedback, the more the similarity embedding will consider the compatibility of industrial component models. This allows the system of the present invention to embed legacy industrial component model attributes with compatibility information to provide a new similarity search capability that is tuned not by the model used to establish the similarity, but rather by the data itself that influences the resulting similarity embedding. Furthermore, the concept of an industrial component model may include the association of independent industrial component models.

Claims

1. - providing a set of industrial component models, each having associated attributes, and wherein the similarity embedding (10) is an embedding of at least some of the associated attributes; receiving a similarity request using a particular industrial component model (ICM) as input, the output of the similarity request being a first subset (ICMS1) of industrial component models selected from the set of industrial component models based on a comparison between the similarity embedding (10) and the similarity embedding of the input industrial component model; receiving a second subset (ICMS2) of industrial component models selected by a user from the first subset (ICMS1) of industrial component models based on compatibility criteria between the input industrial component model (ICM) and any industrial component model of the second subset (ICMS2) of industrial component models; - generating a unique hash code based on a second subset (ICMS2) of the industrial component models; - adding the unique hash code to the associated attribute of each of the industrial component models of the second subset of industrial component models as similarity data; Computing a new set of similarity embeddings (10); 1. A computer-implemented method for improving similarity searching of industrial component models (12), comprising:

2. 2. The computer-implemented method for improving similarity searching of industrial component models of claim 1, wherein the similarity embedding (10) is embedded by vectorizing attributes of the industrial component model and embedding the resulting vectorized data (50).

3. The embedding is performed by a context-sensitive autoencoder (6) that includes an encoder and a decoder that are both neural networks, the neural network the input of said context-sensitive autoencoder (6) is said vectorized data (50), said vectorized data (50) further constituting the input of said encoder, and the output of said encoder constituting said similarity embedding; - the input of the decoder is the affinity embedding and the output of the decoder is the term frequency-inverse document frequency (TFIDF) of the vectorized data (50); - the encoder and decoder are adjusted so that the term frequency-inverse document frequency (TFIDF) of the vectorized data best approximates the vectorized data (50); 3. The computer-implemented method for improving similarity searching of industrial component models of claim 2, comprising:

4. 3. The computer-implemented method for improving similarity searching of industrial component models (ICMs) of claim 2, wherein the embedding is performed by performing principal component analysis on a concatenation of an L1-normalized version of the attributes with selected weights multiplied by an L1-normalized version of the similarity data.

5. A computer-implemented method for improving similarity searching of industrial component models (ICMs) as described in any one of claims 2 to 4, characterized in that the attributes of the industrial component models comprise text attributes, and the vectorization is performed by doc2vec vectorization of the text attributes and by vectorization of the similarity data including adding a column for each unique hash code, and for each industrial component model, populating the column with a 1 if the industrial component model is associated with the unique hash code, and populating the column with a 0 otherwise.

6. 5. The computer-implemented method for improving similarity search of industrial component models (ICMs) according to any one of claims 2 to 4, characterized in that the vectorization is performed by applying a Bidirectional Encoder Representation from Transformer (BERT) technique to the industrial component models.

7. A computer program comprising instructions for carrying out the method of any one of claims 1 to 6.

8. 8. A data storage medium having the computer program according to claim 7 recorded thereon.

9. A computer system comprising a processor coupled to a memory (4), said memory (4) storing a computer program according to claim 7.

Images