WO2024007476A1

WO2024007476A1 - Design concept generation network construction method and automatic concept scheme generation method

Info

Publication number: WO2024007476A1
Application number: PCT/CN2022/125347
Authority: WO
Inventors: 赵武; 于淼; 郭鑫; 张凯; 赵乾; 於辉; 李骏; 赖兵; 姜冲; 蒋亦伟; 伍博; 陈星宇
Original assignee: 四川大学
Priority date: 2022-07-04
Filing date: 2022-10-14
Publication date: 2024-01-11
Also published as: CN115169227B; CN115169227A

Abstract

Disclosed in the present invention are a design concept generation network construction method and an automatic concept scheme generation method. A design concept generation network comprises a Transformer encoder, a Transformer decoder, an importance constraint matrix generation module, an importance constraint embedding layer, a cross-attention layer, and an optimization module. According to the present invention, on the basis of an attention mechanism of Transformer, a vocabulary importance constraint is ingeniously introduced, input vocabulary constraint information comprised in a generated text sequence is recorded, and the reliability and effectiveness of a generated concept scheme can be effectively guaranteed; and potential importance information at a semantic level can be captured, and knowledge reasoning at the semantic level can be realized.

Description

Design concept generation network construction method and concept scheme automatic generation method

Technical field

The invention belongs to the technical field of product design and relates to the automatic generation of concept solutions in product design, and in particular to the construction of a design concept generation network and the automatic generation of concept solutions based on the generation network.

Background technique

Prior design data is an important source of innovation. As the core of product innovation conceptual design, concept solution generation is the process of absorbing valuable design knowledge from prior design data, and further migrating and reorganizing cross-domain design knowledge to generate creative concept solutions. With the advent of the era of big data and big knowledge, the engineering data used in conceptual design is increasing day by day, bringing a rich source of innovation to the research on conceptual scheme generation. Fully applying it to the conceptual scheme generation stage will be beneficial to Expand the design space and generate more design concepts. But it also faces more severe challenges, mainly in two aspects: on the one hand, with the explosive growth of design data, the amount of knowledge applied to conceptual design is also gradually increasing. Based on the designer’s manual experience and design heuristics, It has become increasingly difficult to reason, transfer and reorganize a large amount of design knowledge to produce creative conceptual solutions; on the other hand, design knowledge mainly comes from the description of existing product design solutions in different fields, which often presents complex Diversity, such as rich knowledge types such as functions, structures, scientific effects, cases, etc., and the relationships between knowledge are also more complex and flexible. How to screen out valuable design knowledge based on design problems or design constraints and combine multiple types of cross-domain design knowledge to generate new conceptual solutions is becoming increasingly difficult.

With the rapid development of deep learning technology, many automatic generation technologies have been developed and successfully completed various intelligent tasks, such as machine translation, image generation, speech recognition, etc. The latest deep generative models have also achieved important breakthroughs in many aspects of engineering design, such as structural optimization, material design, shape synthesis, etc. There are also studies using generative models such as topology optimization and generative adversarial networks to automatically generate design concepts in the form of images and spatial shapes. These design concepts are either too abstract to be understood or too detailed to be suitable for early stage conceptual design exploration.

Research has found that text is the most versatile and common form of design concept description, and can cover rich and valuable design knowledge. How to learn the potential combination rules of reasoning, migration and reorganization between design knowledge from massive cross-domain text data through simple and effective models to generate conceptual solutions suitable for the early stages is an important issue that needs to be solved in current product design. .

Contents of the invention

In view of the current technical status quo of the lack of automatic generation method of concept solutions in the field of product design, the purpose of the present invention is to provide a method of constructing a design concept generation network and an automatic generation method of concept solutions based on the generation network, which can generate concepts from massive texts based on design problems. Adaptively learn potential rules such as reasoning, transfer, and reorganization of cross-domain design knowledge in the data, and automatically generate conceptual solutions in text form, thereby reducing reliance on designers' manual experience and improving design efficiency.

The idea of the invention is as follows: the invention provides a Design Concept Generation Network (DCGN), further performs network training and learning, and finally automatically generates concept solutions by inputting design problems into the trained DCGN network.

In order to achieve the above objects, the present invention adopts the following technical solutions.

The design concept generation network construction method provided by the present invention is based on the self-attention mechanism of the Transformer network and cleverly introduces vocabulary importance constraints to construct a new generation network; the design concept generation network includes a Transformer encoder, Transformer decoder, importance constraint matrix generation module, importance constraint embedding layer, cross-attention layer and optimization module; the present invention uses training sample set data to train the design concept generation network; the training sample set data includes several Samples, each sample includes input vocabulary and target sequence; the design concept generation network construction method includes the following steps:

S1 uses the Transformer encoder to obtain the hidden layer features of the encoder based on the input vocabulary in the sample;

S2 uses the Transformer decoder to obtain the decoder hidden layer features based on the target sequence in the sample;

S3 uses the importance constraint matrix generation module to obtain the importance constraint matrix based on the input vocabulary and target sequence in the sample;

S4 uses the importance constraint embedding layer to map the importance constraint matrix to the distributed vector space to obtain two input vocabulary importance embedding features;

S5 uses the cross-attention layer to obtain the generated sequence based on the encoder hidden layer features, decoder hidden layer features and two input vocabulary importance embedding features;

S6 constructs a loss function based on the generation sequence and the target sequence, and uses the optimization module to adjust the network parameters based on the loss function; then repeat steps S1-S6 until the loss function meets the set requirements, and the design concept generation network is obtained.

In the above step S1, the Transformer encoder converts discrete input words into words through the self-attention layer.

(m represents the number of input words in the current sample, n represents the dimension of the input word embedding vector) mapped to the distributed feature representation, that is, the encoder hidden layer features are obtained

(d represents the number of neurons in the hidden layer. In the present invention, the number of neurons in the hidden layer of the Transformer encoder and Transformer decoder is designed to be the same):

In the formula, SA() represents spatial attention;

Represent the weight matrix of the Transformer encoder's self-attention layer respectively. x is discrete and unordered, so there is no need to incorporate the position embedding in the graph when calculating _he , and the output _he will not contain any position information. When the calculated _he vector dimension m is less than M, use 0 vectors to complete it, so that

M≥m>1, M represents the maximum number of input words contained in the samples in the entire training sample set.

In the above step S2, the Transformer decoder maps the target sequence y _:t-1 = [y ₀ , y ₁ , L, y _t-1 ] at the previous moment to the distributed feature representation through the self-attention layer, that is, we get Decoder hidden layer features

In the formula, SA() represents spatial attention;

Respectively represent the weight matrix of the Transformer encoder's self-attention layer; y _:t-1 represents the target sequence at time (t-1) during the training process.

The SA() function in the above formulas (1) and (2) can be calculated by the following formula:

For the encoder, K means

V means

Q means

For the decoder, K means

V means

Q means

In the above step S3, the importance constraint matrix in the present invention is represented by C, which is the input vocabulary information and the target sequence at different times.

The result of joint action can be expressed as:

In the formula, y ₀ is the sequence given at the initial moment, which can be generated using special characters such as <EOS>;

Represents the input vocabulary importance constraint vector contained in the target sequence y _:t , that is, C _:t ; y _:t represents the target sequence before time t (including time t) in the sample; T represents the length of the target sequence in the sample.

Can be calculated as:

In the formula, · represents the vector or matrix dot product operation;

is the relative importance vector of input x in the target sequence y _:t , which can be calculated as:

In the formula,

represents the relative importance of the i-th input vocabulary in the target sequence y _:t ; w _i represents the absolute importance of the i-th input vocabulary in the target sequence y: _t ; w _min represents the minimum value of the input vocabulary in the target sequence y _:t Absolute importance; w _max represents the maximum absolute importance of the input vocabulary in the target sequence y _:t ; [] is the rounding operation.

The relative importance value after the above regularization process

is an integer.

in addition,

Represents the input vocabulary constraints contained in the target sequence y _:t ; when the target sequence y _:t contains the i-th word of the input vocabulary, the i-th element in the vector c _t is 1, which can be calculated as follows:

Therefore, calculated according to formula (3)

is an integer vector composed of relative importance.

In the above step S4, the present invention introduces two new importance constraint embedding matrices.

and

Map the importance constraint matrix C constructed above to the distributed vector space to obtain two input vocabulary importance embedding features.

and

Therefore, at the t-th moment of generation, there is:

In the formula, t∈{1,2,…,T}. In addition, equations (7) and (8) are based on the relative importance

Index importance constraint matrix

and

The corresponding row, the default row is set to zero, and the feature is obtained

In the above step S5, across the attention layer (Cross-Attention layer, CA), the encoder hidden layer features (h _e ) and the decoder hidden layer features are fused

and two input vocabulary importance embedding features (preferably two input vocabulary importance embedding features in the present invention

), get the generated sequence at the current time t

In the formula,

Represents the weight matrix of the decoder’s self-attention layer.

In the specific implementation, the j-th element in the CA function can be expressed as:

In the formula,

As time goes by, the above steps S2-S5 are repeated. When t=T, DCGN obtains the final text generation sequence.

For the samples in the training sample set, repeat the above steps S1-S5 to obtain the generated sequences corresponding to different samples.

In step S6, for the given N samples

The loss function of DCGN constructed based on the generated sequence and the target sequence is:

In the formula,

Represents the generated sequence at time t

The error between it and the target sequence y _:t at the corresponding time is usually calculated by cross entropy.

Based on the above loss function, adjust and optimize the network parameters through the Adam optimization algorithm, and then repeat steps S1-S6 until the loss function meets the set requirements. For example, the loss function tends to be stable and basically unchanged, and the design concept generation network is completed. of construction. The network parameters here mainly refer to the weight matrix of the encoder self-attention layer used to obtain the encoder hidden layer features, the weight matrix and the importance constraint embedding matrix of the decoder self-attention layer used to obtain the decoder hidden layer features. . The initialization parameters of the importance constraint embedding matrix can be implemented through random initialization. The initialization parameters of the weight matrix of the encoder self-attention layer used to obtain the encoder hidden layer features and the weight matrix of the decoder self-attention layer used to obtain the decoder hidden layer features can be implemented by random initialization; in the preferred implementation , the weight matrix of the encoder self-attention layer and the weight matrix of the decoder self-attention layer are obtained by using the common sense text database to perform regular Transformer networks (such as T5 (Text-to-Text Transfer Transformer), GPT (Generative Pre-trained Transformer), etc.), so that the design concept generation network provided by the present invention has the ability to understand common sense knowledge and ensure the fluency of the design concept description generated by the design concept generation network DCGN. The method provided by the present invention is then used to further design the concept generation network DCGN for training, which can enable the network model to have intelligent reasoning capabilities of engineering design knowledge and ensure the rationality of the generated design concept description.

The present invention further provides a method for automatically generating concept solutions, using the constructed design concept generation network to perform operations in accordance with the following steps:

L1, based on the input vocabulary, use the Transformer encoder to obtain the hidden layer features of the encoder;

L2, based on the sequence generated at the previous moment, use the Transformer decoder to obtain the hidden layer features of the decoder at the current moment;

L3, based on the input vocabulary in the sample and the generated sequence at the previous moment, use the importance constraint matrix generation module to obtain the importance constraint matrix;

L4, use the importance constraint embedding layer to map the importance constraint matrix to the distributed vector space, and obtain two input vocabulary importance embedding features;

L5, based on the encoder hidden layer features, decoder hidden layer features and two input vocabulary importance embedding features, use cross-attention layers to obtain the generated sequence.

In the above step L1, the input vocabulary can be composed of keywords from the construction of design problems, or it can be composed of more than one design incentive, or it can be composed of keywords from the construction of design requirements, or at least a combination of the above two sources of input vocabulary. .

In the above step L2, based on the sequence generated at the previous moment, the hidden layer features of the decoder at the current moment are calculated according to the following formula

In the formula, y _:t-1 represents the input sequence of the decoder at time t during the generation process,

Represents the sequence given at the initial moment. Special characters such as <EOS> can be used to represent the generation.

Represents the sequence generated at the previous moment.

In the above step L3, in the conceptual solution generation stage, the constraint matrix is calculated in time steps based on the actual sequence generated at each moment.

Based on the input vocabulary in the sample and the sequence generated at the previous moment, the importance constraint matrix C _:t-1 is calculated according to the following formula:

In the formula, x represents the input vocabulary, y _:t-1 represents the input sequence of the decoder at time t during the generation process,

is the relative importance vector of input x in the decoder input sequence y _:t-1 , which can be calculated according to the above formula (5), where the absolute importance of the input vocabulary in the decoder input sequence y _:t-1 can be calculated according to the input The order of vocabulary importance is given in advance, and can also be set to be consistent.

In the above step L4, two input vocabulary importance embedding features are used

Calculate the importance embedding features of the two input words at the current moment according to the above formulas (7) and (8).

In the above step L5, the generated sequence at the current time is calculated according to formulas (9)-(10).

Repeat the above steps L1-L5 until the length of the generated sequence meets the set requirements or the end identifier <EOS> is encountered, and the final generated sequence, that is, the conceptual solution is obtained.

Compared with the prior art, the present invention has the following beneficial effects:

1) Based on the attention mechanism of Transformer, this invention cleverly introduces vocabulary importance constraints and constructs a new design concept generation network.

2) The vocabulary importance constraint matrix proposed by this invention records the input vocabulary constraint information contained in the generated text sequence, which can effectively ensure the reliability and effectiveness of the generated concept solution;

3) The present invention proposes an importance constraint embedding layer, which maps the constructed importance constraint matrix to a distributed vector space, and uses continuous real number vectors to represent the relative importance of input words in the generated sequence or target sequence, which is beneficial to capturing potential importance information at the semantic level to realize knowledge reasoning at the semantic level;

4) The cross-attention layer constructed by the present invention maps the input vocabulary importance embedding features to the generated sequence to supervise the generation of text sequences containing input vocabulary importance information.

Description of the drawings

Figure 1 is a schematic diagram of the construction and use framework of the design concept generation network in the embodiment of the present invention.

Figure 2 is a schematic diagram of the principle of a method for constructing a design concept generation network in an embodiment of the present invention.

Figure 3 is a schematic diagram of the concept method generation process in the embodiment of the present invention.

Detailed ways

The technical solutions of various embodiments of the present invention are clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without any creative efforts belong to the present invention.

Example 1

As shown in Figure 1, this embodiment first uses web crawler technology to obtain text data and perform preprocessing; then constructs a DCGN model to further train and learn the model; and finally inputs the design into the trained DCGN model. Keywords are used as input vocabulary to automatically generate concept plans.

This embodiment first uses conventional web crawler technology to crawl massive text data from websites, such as scientific papers, patents, etc., and selects sentences of a certain length from the collected text data as the corpus for this study. Then the text data is preprocessed, and a keyword extraction algorithm is used to obtain a certain number of keywords (excluding stop words) and their importance from each sentence. Finally, each sentence and its corresponding keyword information are combined into a sample pair, and an entire sample set composed of sample pairs is constructed for subsequent network training. Each sample uses the extracted keywords as the input sequence and the corresponding sentence as the target sequence.

(1) Construction of design concept generation network

This embodiment cleverly introduces vocabulary importance constraints based on the self-attention mechanism of the Transformer network to construct a new generation network; the design concept generation network includes a Transformer encoder, a Transformer decoder, and an importance constraint matrix generation module, importance constraint embedding layer, cross-attention layer and optimization module. The Transformer encoder is used to obtain the encoder hidden layer features; the Transformer decoder is used to obtain the decoder hidden layer features; the importance constraint matrix generation module is used to generate the importance constraint matrix; the importance constraint embedding layer is used to map the importance constraint matrix to Distributed vector space is used to obtain two input word importance embedding features; the cross-attention layer is used to obtain the generated sequence; the optimization module is used to optimize network parameters based on the loss function.

The design concept generation network construction method provided in this embodiment mainly involves training and learning on the training sample set to obtain the weight matrix of the encoder self-attention layer for obtaining the characteristics of the encoder hidden layer, and the weight matrix for obtaining the decoder hidden layer. The decoder of layer features has a weight matrix from the attention layer and two importance constraint embedding matrices.

This embodiment uses the common sense text database (selected from Wikipedia) to train the conventional Transformer network (T5 (Text-to-Text Transfer Transformer)) to obtain the weight matrix of the encoder self-attention layer used to obtain the encoder hidden layer features. and the weight matrix initialization parameters of the decoder self-attention layer used to obtain the decoder hidden layer features. The two importance constraint embedding matrices obtain their initialization parameters through random initialization.

1. Use the common sense text database to train the T5 network

Here, the common sense text database is used to train the T5 network to obtain the weight matrix of the encoder's self-attention layer.

and the weight matrix of the decoder self-attention layer

Explanation of encoders and decoders as mentioned earlier. For the specific implementation process of T5 network training, please refer to the literature Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer (Colin Raffel et al, Journal of Machine Learning Research 21 (2020) 1-67. Take the trained T5 network The weight matrix of the encoder self-attention layer in

and the weight matrix of the decoder self-attention layer

As the design concept of the present invention, corresponding initialization parameters of the network are generated.

2. Design concept generation network construction

As shown in Figure 2, the design concept generation network construction method provided by this embodiment includes the following steps:

S1, based on the input vocabulary in the sample, use the Transformer encoder to obtain the hidden layer features of the encoder.

This step is based on the input vocabulary x={x ₁ , x ₂ , L, x _m } in the sample and calculates the encoder hidden layer feature _he according to the previous formula (1).

S2, based on the target sequence in the sample, use the Transformer decoder to obtain the decoder hidden layer features.

This step is based on the target sequence y in the sample _{: t-1} = [y ₀ , y ₁ , L, y _t-1 ], and calculates the hidden layer features of the decoder at time t according to the previous formula (2)

S3, based on the input vocabulary and target sequence in the sample, use the importance constraint matrix generation module to obtain the importance constraint matrix.

The importance constraint matrix C is determined by the formula (3) given previously.

Here, the importance constraint matrix at time t-1 is obtained based on the input vocabulary and the target sequence y _:t at time t-1, that is,

It can be calculated according to the previous formulas (4)-(6).

The following uses a specific example to show the detailed calculation process of C during DCGN training. Assume that the input of DCGN is a set of three keywords {"sensor", "device", "sowing"}, and the generated target sequence is "a sensor device for determining a position of seeds while sowing." Assuming that the importance of the input vocabulary in the target sequence is w=[0.9,0.7,0.5], and M=5, the relative importance vector can be calculated according to formula (4) as

This value represents the relative importance of these three input words in the target sequence. Next, the calculation steps of C are explained, as shown in Table 1:

a) When starting to generate the start symbol <EOS>, the target sequence corresponding to this moment does not contain any input vocabulary, so c ₀ is an all-zero vector at this time. Calculated by formula (4)

It is also an all-zero vector, corresponding to the first column value in Table 1;

b) Because the second generated target word is "a", the target sequence at this moment still does not contain any input words, so c ₁ is an all-zero vector at this time. Calculated by formula (4)

It is also an all-zero vector, corresponding to the second column value in Table 1;

c) Because the third generated target word is "sensor", the target sequence at this moment only contains "sensor" in the input vocabulary, so c ₂ = [1; 0; 0], and then according to formula (4) Computable

Corresponds to the third column value in Table 1;

d) Because the fourth generated target word is "device", the target sequence at this moment contains "sensor" and "device" in the input vocabulary, so there is c ₃ = [1; 1; 0], and then according to the formula (4) Calculable

Corresponds to the fourth column value in Table 1;

e) And so on until the end character <EOS> is generated.

Table 1 The generation process of C during the construction of DCGN network

S4, use the importance constraint embedding layer to map the importance constraint matrix to the distributed vector space, and obtain two input vocabulary importance embedding features.

In this step, the importance embedding features of the two input words at time t are calculated according to formulas (7) and (8).

and

S5, based on the encoder hidden layer features, the decoder hidden layer features and the two input vocabulary importance embedding features, use the cross-attention layer to obtain the generated sequence.

In this step, the generated sequence at time t is calculated according to formulas (9) and (10)

For N samples given in the training sample set

Repeat the above steps S1-S5 to obtain the generated sequences corresponding to N samples.

S6: Construct a loss function based on the generation sequence and the target sequence, and adjust the network parameters based on the loss function; then repeat steps S1-S6 until the loss function meets the set requirements, and the design concept generation network is obtained.

In this step, for the given N samples, the loss function of DCGN is calculated according to formula (11). Based on this loss function, adjust and optimize the network parameters through the conventional Adam optimization algorithm, and then repeat steps S1-S6 until the loss function meets the set requirements. For example, the loss function tends to be stable and basically unchanged, and the design concept generation is completed. Network construction.

After the DCGN network model is fully trained, it has the ability to express knowledge and reason, and can adaptively absorb, transfer, and reorganize cross-domain design knowledge. At this stage, well-defined design problems or valuable knowledge incentives are input into the trained DCGN, and relevant design concept descriptions can be automatically generated. The DCGN network model combines design knowledge from different fields to generate design concepts that not only contain design input information, but also ensure the novelty and creativity of the generated design concepts.

(2) Testing of the design concept generation network.

Next, the validity and practicability of the proposed automatic generation method of conceptual solutions are tested by inputting design questions (i.e., keywords).

The method for automatically generating concept solutions provided in this embodiment uses the constructed design concept generation network to perform operations according to the following steps:

L1, based on the input vocabulary, uses the Transformer encoder to obtain the hidden layer features of the encoder.

In this step, the encoder hidden layer _feature he is calculated according to the above formula (1).

L2, based on the sequence generated at the previous moment, uses the Transformer decoder to obtain the decoder output hidden layer features at the current moment.

In this step, the decoder hidden layer features are calculated according to the above formula (12)

L3, based on the input vocabulary in the sample and the generated sequence at the previous moment, use the importance constraint matrix generation module to obtain the importance constraint matrix.

In this step, the importance constraint matrix C _:t-1 is calculated according to formula (13).

In this embodiment, the absolute importance of the input words in the decoder input sequence y _:t-1 is set to be consistent, and the value of w _i is 1.

L4, use the importance constraint embedding layer to map the importance constraint matrix to the distributed vector space, and obtain two input vocabulary importance embedding features.

In this step, two input vocabulary importance embedding features are used

In this step, the generated sequence at the current moment is calculated according to formulas (9)-(10).

Therefore, the specific concept plan generation stage is that the output vocabulary at the previous moment will be used as a new part of the input at the current moment, and new vocabulary will be generated in sequence until the end identifier <EOS> is encountered. The process is shown in Figure 3. Use x={drone,deliver,life,preserver} as the input vocabulary, and use special characters such as <EOS> to represent the initial moment generation sequence

Repeat the above steps L1-L5 until the end identifier <EOS> is encountered and the generated sequence is obtained

In the generation phase, C is calculated in time steps based on the actual sequence generated at each moment, regardless of the target sequence, which is completely different from the training phase.

The following are specific examples of concept generation schemes from different input vocabulary sources:

1. The design problem of this case is to provide edible water sources for residents in coastal areas. In order to express the design problem more accurately and concisely, 10 graduate students majoring in mechanical engineering were invited to define the design problem using a limited number of keywords. Taking into account the advantages of abundant sunshine and light in coastal areas, the design team unanimously agreed to use "purification" (purification or purify), "desalination" (desalination or desalinate), "solar" (solar), "seawater" (seawater) and The "drink" keyword is used to define the design problem. By combining different keywords as design input, the corresponding design concept can be automatically generated by using the constructed DCGN network according to the above-mentioned concept scheme automatic generation method. The results are shown in Table 2. The automatically generated design concepts provide more specific and feasible design concepts, such as inventing a purification system for purifying seawater into drinking water, or using solar energy to desalinate seawater for the production of canned drinking water or beverage products. These design concepts provide residents or businesses in coastal areas with early design ideas for product development.

Table 2 Automatically generated conceptual solutions using different design problem keywords as input

2. The design problems involved in the present invention can also be composed of design incentives. In the process of product innovation concept design, design incentives provide rich and valuable design inspiration. In the traditional process of artificially generating conceptual solutions, the transition from design incentives to conceptual solutions often relies on the designer's rich experience and knowledge, and generates Conceptual solutions are inefficient and the process becomes very difficult for inexperienced novice designers. Some design incentives for UAVs obtained in this embodiment are shown in Table 3. By combining different design incentives as inputs to the DCGN network model, the automatically generated conceptual scheme is shown in Table 4. Since there are many types of combinations, only some valuable conceptual solutions are shown and analyzed here. For example:

(1) By combining the design incentives "drone", "bio", "radar" and "rescue", the DCGN network automatically generates the design concept "a drone rescue radar system is disclosed that is capable of detecting the presence of an animal in the "nearby of the drone using bio".

(2) By combining the design incentives "drone", "fire", "ground" and "data", the DCGN network automatically generates the design concept "the drone may also be configured to receive ground fire data from the ground drone and to determine a location of the fire in response to detecting the resulting fire.", and "the drone may also be configured to receive ground fire data from the ground drone and to determine a location of the fire in response to determining the terrain".

Table 3 Design incentives retrieved in UAV cases (in no particular order)

Table 4 Design concepts automatically generated by combining different design incentives on the UAV case

3. Further, in order to fully supplement the design issues involved, some design issues can be defined by design requirements. In the early stages of product design, design requirements are critical in determining the design direction of a new product. Online product review data provides accurate, reliable, authentic information for analyzing design needs, and is easily accessible. Here, we used conventional crawler technology to extract 20,918 user review texts of a certain milk bottle sterilizer from an e-commerce platform. Through the data preprocessing process provided above, we analyzed the included keywords and corresponding word frequencies. The results are shown in Table 5. shown. The analysis found that users mainly expressed clear needs in terms of functions, disinfection, capacity, temperature, etc. In order to apply the design requirements to obtain the design issues, the keywords "disinfection" (sterilization), "temperature" (temperature), "function" (function), and "capacity" (capacity) are used here as the design issues of the DCGN network model, automatically The generated conceptual scheme is shown in Table 6. It is easy to find that using different input keyword combinations will generate different conceptual solutions. More importantly, all automatically generated conceptual solutions contain the entered design problem keywords, and some feasible and creative conceptual solutions have been generated, such as using ion exchangers to improve sterilization and disinfection capabilities. , which meets the design needs to a certain extent.

Table 5 The top 30 most frequent demand keywords in online user review data

Table 6 Design concepts automatically generated using different design requirement keywords as input

To sum up, if designers think about these design issues and rely solely on artificial experience to generate conceptual solutions, it will not only be difficult to create innovative conceptual solutions, but the efficiency will also be very low. Aiming at the problem of difficulty in transferring and reorganizing cross-domain design knowledge and automatically generating design concept plans in the product concept plan generation stage, which is based on manual experience, the present invention proposes a method for automatically generating concept plans based on the Design Concept Generation Network (DCGN). DCGN can adaptively learn potential rules such as reasoning, migration, and reorganization of cross-domain design knowledge from massive text data, and automatically generate product concept solutions based on design problems. This not only reduces the burden of manually generating concept solutions, but also improves efficiency. Design efficiency provides new ideas for intelligent conceptual design.

Those of ordinary skill in the art will appreciate that the embodiments described here are provided to help readers understand the principles of the present invention, and it should be understood that the scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific modifications and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and these modifications and combinations are still within the protection scope of the present invention.

Claims

A method for constructing a design concept generation network, characterized in that the design concept generation network includes a Transformer encoder, a Transformer decoder, an importance constraint matrix generation module, an importance constraint embedding layer, a cross-attention layer and an optimization module; The design concept generation network construction method includes the following steps:

S1 uses the Transformer encoder to obtain the hidden layer features of the encoder based on the input vocabulary in the sample;

S2 uses the Transformer decoder to obtain the decoder hidden layer features based on the target sequence in the sample;

S3 uses the importance constraint matrix generation module to obtain the importance constraint matrix based on the input vocabulary and target sequence in the sample;

S4 uses the importance constraint embedding layer to map the importance constraint matrix to the distributed vector space to obtain two input vocabulary importance embedding features;

S5 uses the cross-attention layer to obtain the generated sequence based on the encoder hidden layer features, decoder hidden layer features and two input vocabulary importance embedding features;

S6 constructs a loss function based on the generation sequence and the target sequence, and uses the optimization module to adjust the network parameters based on the loss function; then repeat steps S1-S6 until the loss function meets the set requirements, and the design concept generation network is obtained.
The design concept generation network construction method according to claim 1, characterized in that, in step S1, the Transformer encoder obtains the encoder hidden layer feature he according to the following formula:

In the formula, x represents the input vocabulary; SA() represents spatial attention;
Represent the weight matrix of the Transformer encoder's self-attention layer respectively.
The design concept generation network construction method according to claim 1, characterized in that, in step S2, the Transformer decoder converts the target sequence y at the previous moment through the self-attention layer : t-1 = [y 0 , y 1 , L,y t-1 ] is mapped to the distributed feature representation, that is, the decoder hidden layer features are obtained

In the formula, SA() represents spatial attention;
Represent the weight matrix of the Transformer encoder's self-attention layer respectively.
The design concept generation network construction method according to claim 3, characterized in that, in step S3, using
Represents the input vocabulary importance constraint vector contained in the target sequence y :t , that is, C :t ;

Can be calculated as:

In the formula, · represents the vector or matrix dot product operation;
is the relative importance vector of input x in the target sequence y :t , which can be calculated as:

In the formula,
represents the relative importance of the i-th input vocabulary in the target sequence y :t ; w i represents the absolute importance of the i-th input vocabulary in the target sequence y: t ; w min represents the minimum value of the input vocabulary in the target sequence y :t Absolute importance; w max represents the maximum absolute importance of the input vocabulary in the target sequence y :t ; [] is the rounding operation; M≥m>1, M represents the maximum number of input vocabulary contained in the sample in the entire training sample set;

Indicates the input vocabulary constraints contained in the target sequence y :t ; when the target sequence y :t contains the i-th word of the input vocabulary, the i-th element in the vector c t is 1, which can be calculated as follows:
The design concept generation network construction method according to claim 1 or 4, characterized in that, in step S4, two importance constraint embedding matrices are introduced
and
Map the importance constraint matrix constructed above to the distributed vector space to obtain two input vocabulary importance embedding features.
and
At the t-th moment of generation, there is:
The design concept generation network construction method according to claim 5, characterized in that, in step S5, the encoder hidden layer features he and the decoder hidden layer features are fused across the attention layer
and two input word importance embedding features
Get the generated sequence at the current time t

In the formula,
Represents the weight matrix of the decoder’s self-attention layer;

In the specific implementation, the j-th element in the CA function can be expressed as:

In the formula,

i＝0,1,…,M-1; j, l＝0,1,…,d-1; () T represents the transpose operation.
The design concept generation network construction method according to claim 1, characterized in that, in step S6, for the given N samples
The loss function constructed based on the generated sequence and the target sequence is:

In the formula,
Represents the generated sequence at time t
and the error between the target sequence y :t at the corresponding time.
A method for automatically generating concept plans, characterized in that the design concept generation network constructed using any method of claims 1 to 7 performs operations in accordance with the following steps:

L1, based on the input vocabulary, use the Transformer encoder to obtain the hidden layer features of the encoder;

L2, based on the sequence generated at the previous moment, use the Transformer decoder to obtain the hidden layer features of the decoder at the current moment;

L3, based on the input vocabulary in the sample and the generated sequence at the previous moment, use the importance constraint matrix generation module to obtain the importance constraint matrix;

L4, use the importance constraint embedding layer to map the importance constraint matrix to the distributed vector space, and obtain two input vocabulary importance embedding features;

L5, based on the encoder hidden layer features, decoder hidden layer features and two input vocabulary importance embedding features, use cross-attention layers to obtain the generated sequence.
The method for automatically generating conceptual solutions according to claim 8, characterized in that, in step L1, the input vocabulary consists of keywords constructed from design problems, or consists of more than one design incentive, or keywords constructed from design requirements. composition, or a combination of at least two sources of input vocabulary.
The method for automatically generating conceptual solutions according to claim 8 or 9, characterized in that, in step L2, the decoder hidden layer characteristics at the current moment are calculated according to the following formula based on the generated sequence at the previous moment.

In the formula, y :t-1 represents the input sequence of the decoder at time t during the generation process,
Represents the sequence given at the initial moment,
Represents the sequence generated at the previous moment.