Industrial text retrieval method and system based on multiple gate fusion and hybrid metric

Abstract

This invention relates to an industrial text retrieval method and system based on multi-gated fusion and hybrid metrics. The method includes: splitting multiple heterogeneous feature fields and constructing a global joint concatenated vector; calculating routing weights based on a multi-gated decision network and a zero-sum game competition mechanism; constructing a triplet supervised sample set and using a contrastive loss function to guide the multi-gated decision network in fine-tuning training to construct a mechanism decision boundary, thereby obtaining a global fusion representation vector; calculating the dense semantic similarity score of the global fusion representation vector and introducing a sparse matching score, and using a hybrid metric mechanism containing a composite balance factor to weight and calibrate the score. This invention overcomes the technical bottlenecks of core feature annihilation caused by indiscriminate concatenation of multiple heterogeneous fields, semantic overgeneralization and mismatch caused by pure dense representation, and sampling collisions occurring in general contrastive learning, significantly improving the retrieval accuracy of multi-field industrial text in multi-mechanism mutually exclusive scenarios.

Description

Technical Field

[0001] This invention relates to the fields of intelligent equipment operation and maintenance, industrial big data analysis and industrial information retrieval technology, specifically to an industrial text retrieval method and system based on multi-gating fusion and hybrid measurement. Background Technology

[0002] In the process of industrial digital transformation and smart factory construction, process industry enterprises (such as petrochemicals, metallurgy, and power) have accumulated massive amounts of unstructured equipment maintenance work orders, repair logs, and troubleshooting cases in their asset management systems and digital operation and maintenance platforms. These short text records, entered by front-line operators or maintenance personnel in columns on the system's interactive interface, highly encapsulate the failure evolution mechanisms, core anomaly characteristics, and expert maintenance knowledge of underlying industrial equipment. However, such corpora generally exhibit typical characteristics such as non-standardized terminology, heavy colloquialism, mismatched tag numbers, and fragmented distribution of key information across multiple fields.

[0003] In existing technologies, retrieval and recommendation solutions for this type of semi-structured industrial text mainly extract discrete text features and perform cascaded or global fusion mapping with static structured data features to achieve fallback retrieval based on the similarity of historical cases. However, traditional industrial retrieval and fusion matching technologies face the following deep algorithmic bottlenecks when practically applied to heavy industrial operation and maintenance scenarios with high homogeneity and high mechanistic resolution requirements: First, core fault characteristics are easily obscured by complex background noise. For example, in the petrochemical industry, industrial texts typically consist of multiple heterogeneous fields such as "specialty," "equipment description," "defect description," and "defect phenomenon." Traditional concatenation paradigms crudely merge these fields into the network, lacking fine-grained routing control over information flow. Equipment descriptions or tag numbers are often dozens of words long, while the most crucial description of the defect phenomenon is often only a few words, making core characteristics easily diluted by lengthy background noise.

[0004] Secondly, purely dense semantic vector representations are highly susceptible to "semantic overgeneralization" and false positive mismatches. Industrial work order databases contain numerous samples with highly overlapping prefixes but diametrically opposed actual fault mechanisms. Traditional semantic similarity models based on deep pre-trained language models, lacking explicit physical literal constraints, are prone to latent space representation collapse. This results in extremely high semantic similarity scores for such samples with highly similar prefixes but different mechanisms, triggering numerous false positives and severely compromising the robustness of the maintenance decision-making system.

[0005] Furthermore, the contrastive learning fine-tuning strategy for general text suffers from severe "sampling collisions" in the context of industrial verticals. Existing model fine-tuning typically employs random negative sampling or is based on a general unsupervised contrastive learning architecture. If, during sampling, records with extremely similar texts within the same professional domain but different core fault categories are mistakenly classified as negative samples, contradictory supervisory signals will be transmitted to the model, resulting in extremely blurred classification decision boundaries, making it difficult to meet the deterministic requirements of precise troubleshooting in industry. Summary of the Invention

[0006] This invention provides an industrial text retrieval method and system based on multi-gated fusion and hybrid metrics to solve at least one of the above-mentioned technical problems.

[0007] The technical solution of this invention to solve the above-mentioned technical problems is as follows: An industrial text retrieval method based on multi-gated fusion and hybrid metrics, comprising: The industrial query text and the standard document base are split into multiple heterogeneous feature fields, and the local spatial dense features of each heterogeneous feature field are extracted using a parameter-shared and frozen dual-tower pre-trained language model. Then, the query global joint concatenation vector and the base global joint concatenation vector are constructed accordingly. Based on a multi-gated decision network and a zero-sum game competition mechanism, the routing weights of each feature channel in the global semantic expression of the query global joint concatenation vector and the base database global joint concatenation vector are dynamically calculated respectively, and then the initial query global fusion representation vector and the initial base database global fusion representation vector are constructed accordingly. Offline construction of a triplet supervised sample set based on vertical domain business mechanism constraints, and combined with the contrastive loss function to guide the fine-tuning training of the multi-gated decision network, so as to construct the mechanism decision boundary, and then determine the latest weight of the multi-gated decision network in the preset context, thereby calibrating the initial query global fusion representation vector and the initial base database global fusion representation vector respectively, and obtaining the query global fusion representation vector and the base database global fusion representation vector accordingly; Calculate the dense semantic similarity score between the global fusion representation vector of the query and the global fusion representation vector of the base library in the latent space, and simultaneously introduce the sparse matching score between the industrial query text and the standard document base library based on the full text statistical features of the inverted index. Then, use a hybrid measurement mechanism with a composite balancing factor to perform weighted calibration on the dense semantic similarity score and the sparse matching score to obtain a consistent retrieval ranking list.

[0008] Based on the above-mentioned industrial text retrieval method based on multi-gated fusion and hybrid metrics, the present invention also provides an industrial text retrieval system based on multi-gated fusion and hybrid metrics.

[0009] An industrial text retrieval system based on multi-gated fusion and hybrid metrics, applied to the industrial text retrieval method based on multi-gated fusion and hybrid metrics as described above, includes: The local feature independent encoding module is used to split the industrial query text and the standard document base into multiple heterogeneous feature fields respectively, and use the parameter-shared and frozen dual-tower pre-trained language model to extract the local spatial dense features of each heterogeneous feature field, and then construct the query global joint concatenation vector and the base global joint concatenation vector accordingly. The gating interaction dynamic fusion module is used to dynamically calculate the routing weights of each feature channel in the global semantic expression of the query global joint concatenation vector and the base database global joint concatenation vector based on a multi-gating decision network and a zero-sum game competition mechanism, and then construct the initial query global fusion representation vector and the initial base database global fusion representation vector accordingly. The fine-tuning and calibration module is used to offline construct a triplet supervision sample set based on vertical domain business mechanism constraints, and combine it with the contrastive loss function to guide the fine-tuning training of the multi-gated decision network to construct the mechanism decision boundary, thereby determining the latest weights of the multi-gated decision network in the preset context, and thus calibrating the initial query global fusion representation vector and the initial base database global fusion representation vector respectively, and obtaining the query global fusion representation vector and the base database global fusion representation vector accordingly; The cross-space hybrid metric reasoning module is used to calculate the dense semantic similarity score between the global fusion representation vector of the query and the global fusion representation vector of the base library in the latent space, and simultaneously introduce the sparse matching score between the industrial query text and the standard document base library based on the full text statistical features of the inverted index. Furthermore, the module uses a hybrid metric mechanism with a composite balancing factor to perform weighted calibration on the dense semantic similarity score and the sparse matching score to obtain a consistent retrieval ranking list.

[0010] The beneficial effects of this invention are as follows: This invention provides an industrial text retrieval method and system based on multi-gated fusion and hybrid metrics. The multi-gated decision network can adaptively identify and suppress background noise caused by lengthy texts such as redundant equipment tag numbers, preventing the dilution of core short text defective feature streams. Simultaneously, a triplet supervised sample set based on vertical domain business mechanism constraints is constructed to fine-tune the multi-gated decision network, ultimately calibrating the global fusion representation vector. This effectively prevents the misclassification of work orders with similar key semantics but completely different processing mechanisms as negative samples, greatly improving the model's performance in handling highly homogeneous industrial texts. The invention improves the mechanistic resolution of industrial texts. Furthermore, by explicitly embedding a cross-space dense / sparse hybrid measurement mechanism in the inference retrieval module, discrete literal word frequency statistical scores are used as absolute calibration anchors. Hard penalties are applied to artificially high semantic cosine scores caused by fixed prefixes and long text inclusion, significantly reducing the false positive recall rate. Therefore, this invention overcomes the technical bottlenecks in existing technologies, such as the annihilation of core features due to indiscriminate cascading of multiple heterogeneous fields, semantic overgeneralization mismatch due to purely dense representations, and sampling collisions in general contrastive learning. This significantly improves the retrieval accuracy of multi-field industrial texts in scenarios with mutually exclusive mechanisms. Attached Figure Description

[0011] Figure 1 This is a flowchart of an industrial text retrieval method based on multi-gated fusion and hybrid metrics according to the present invention. Figure 2 This is a structural block diagram of an industrial text retrieval system based on multi-gated fusion and hybrid metrics according to the present invention. Detailed Implementation

[0012] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0013] Example 1: like Figure 1 As shown, an industrial text retrieval method based on multi-gated fusion and hybrid metrics includes: S1, the industrial query text and the standard document base library are split into multiple heterogeneous feature fields respectively, and the local spatial dense features of each heterogeneous feature field are extracted using a parameter-shared and frozen dual-tower pre-trained language model, and then the query global joint concatenation vector and the base library global joint concatenation vector are constructed accordingly. S2, based on a multi-gated decision network and a zero-sum game competition mechanism, dynamically calculates the routing weights of each feature channel in the global semantic expression of the query global joint concatenation vector and the base database global joint concatenation vector, and then constructs the initial query global fusion representation vector and the initial base database global fusion representation vector accordingly. S3. Offline construction of a triplet supervised sample set based on vertical domain business mechanism constraints, and combined with the contrastive loss function to guide the fine-tuning training of the multi-gated decision network to construct the mechanism decision boundary, thereby determining the latest weight of the multi-gated decision network in the preset context, thereby calibrating the initial query global fusion representation vector and the initial base database global fusion representation vector respectively, and obtaining the query global fusion representation vector and the base database global fusion representation vector accordingly. S4, calculate the dense semantic similarity score between the query global fusion representation vector and the base database global fusion representation vector in the latent space, and simultaneously introduce the sparse matching score between the industrial query text and the standard document base database based on the inverted index full-text literal statistical features. Furthermore, use a hybrid measurement mechanism containing a composite balancing factor to perform weighted calibration on the dense semantic similarity score and the sparse matching score to obtain a consistent retrieval ranking list.

[0014] To make the objectives, technical solutions, and beneficial effects of this invention clearer and more thorough, the following detailed explanation of the technical solution transfer process of this invention is provided in conjunction with specific experimental simulation environments and real industrial scenario data.

[0015] The algorithm verification and system operating environment configuration in this embodiment are as follows: The computing hardware platform uses a single NVIDIA GeForce GTX 1650 graphics card (equipped with 4GB of dedicated video memory), the underlying deep learning framework is based on PyTorch 2.0, and the vector indexing acceleration uses the open-source FAISS (Facebook AI Similarity Search) high-dimensional vector retrieval library. The following is a detailed description of each step in the method of this invention.

[0016] The above S1 specifically refers to the steps of multi-field structured recombination and independent encoding of local latent space features.

[0017] In S1 above, this embodiment specifically splits the industrial query text and standard document base into multiple heterogeneous feature fields according to the business logic structure; the multiple heterogeneous feature fields specifically include professional field fields, equipment description fields, defect description fields, and defect phenomenon fields.

[0018] Specifically, this embodiment extracts historical equipment defect maintenance work orders accumulated in the asset management system of a large petrochemical enterprise as a standard document base. This standard document base contains 42,000 structured and semi-structured maintenance records. When a frontline maintenance engineer submits a new equipment defect (i.e., industrial query text) in the user interface at the system's front end, the system receives the input text sequence in real time and dynamically segments and distributes it to four independent feature channels according to preset business logic rules, defined as follows: Professional Field For example, "statics", "electrical engineering", "power engineering" or "instrumentation engineering"; Device description field : Typically contains complex plant location numbers and physical entity names, such as "P-101A / B power pump for No. 1 high-pressure hydrogenation unit"; Defect description field Record the specific spatial location where the defect occurred, such as "the mating surface between the non-drive end bearing housing and the sealing gland of the pump body"; Defect phenomenon field It carries specific fault characterization patterns, such as "the occurrence of continuous oil leakage accompanied by abnormal intermittent frictional metallic noise".

[0019] It should be noted that this embodiment uses industrial text retrieval in the petrochemical industry as an example. This industry uses four feature channels—professional field fields, equipment description fields, defect description fields, and defect phenomenon fields—to split industrial query text. For industrial text retrieval in other industries (such as metallurgy, power, etc.), the feature channels can be adapted and reconstructed according to the industry's own process flow, equipment system, and operation and maintenance pain points. The splitting rules and the number of feature channels can be adaptively adjusted so that this invention can quickly adapt to the work order templates of different enterprises.

[0020] In S1 above, the local spatially dense features of each heterogeneous feature field are extracted using a parameter-shared and frozen dual-tower pre-trained language model. This is then used to construct the query global joint concatenation vector and the base database global joint concatenation vector, specifically including: S11, perform word-based processing on each heterogeneous feature field to explicitly insert control tags and sentence break tags at the beginning and end of each heterogeneous feature field; S12 involves performing multi-layer self-attention context interaction on the dual-tower pre-trained language model (with shared and frozen parameters) after the lexicalized heterogeneous feature fields are input into the model. This process extracts the final hidden state vectors corresponding to the control tags at the beginning and end of each heterogeneous feature field, thus obtaining the local spatial dense features of each heterogeneous feature field. The formula for bidirectional self-attention context interaction is as follows: ; In the formula, For the first Local spatial density features of heterogeneous feature fields in road For a pre-trained language model with shared parameters and already frozen, These are the shared parameters in the dual-tower pre-trained language model that are completely frozen and do not participate in gradient updates. For the control flag, This is a function for word segmentation. For the first heterogeneous feature fields of roads The sentence segmentation marker, This is the final hidden state vector corresponding to [CLS] extracted after the multi-layer self-attention context interaction. Let be the set of all real numbers. The dimension representing the heterogeneous feature field; S13, perform tensor concatenation on the local spatially dense features of each heterogeneous feature field in the industrial query text to obtain the global joint concatenated vector of the query; perform tensor concatenation on the local spatially dense features of each heterogeneous feature field in the standard document base library to obtain the global joint concatenated vector of the base library; wherein, the formula for tensor concatenation is: ; In the formula, For the query global federated concatenation vector or the base database global federated concatenation vector, corresponding to, For the industrial query text or the first document in the standard document base library Local spatial density features of heterogeneous feature fields in road , The total number of heterogeneous feature fields in the industrial query text or the standard document base library.

[0021] Specifically, based on the aforementioned splitting, this embodiment selects the pre-trained language model paraphrase-multilingual-MiniLM-L12-v2 as the shared base encoder of the dual-tower architecture (i.e., the dual-tower pre-trained language model). This dual-tower pre-trained language model has a total of only 33 million (33M) parameters, possesses strong semantic discrimination capabilities for short Chinese texts, and is suitable for deployment on localized industrial servers. This invention utilizes heterogeneous feature fields from four feature channels. (i.e., the aforementioned) , , , Independent tokenization is performed on each heterogeneous feature field, and special control markers [CLS] and sentence segmentation markers [SEP] are explicitly inserted at the beginning and end of each heterogeneous feature field. The segmented heterogeneous feature fields are then fed into the base encoder for 12 layers of standard bidirectional self-attention context interaction. After the multi-layer Transformer interaction, the final hidden state vector corresponding to the precisely extracted first [CLS] of the sequence is used as a 384-dimensional locally high-dimensional semantically dense vector for the corresponding feature channel. By processing the four feature channels in parallel, a fixed-length locally joint concatenated vector is finally constructed for each industrial query text. This embodiment will analyze the local spatial density features of each feature channel. The superposition performs a horizontal tensor concatenation, integrating them into a one-dimensional global joint concatenation vector of length 1536. .

[0022] The above S2 is the step of channel-level adaptive gating interaction fusion.

[0023] The above S2 specifically includes: S21, a multi-gated decision network based on nonlinear gated projection is used to perform nonlinear projection on the query global joint concatenated vector and the base database global joint concatenated vector, respectively, to extract the global multi-channel context feature vectors of the query global joint concatenated vector and the base database global joint concatenated vector; wherein, the formula for performing nonlinear projection is: ; In the formula, For the global multi-channel context feature vector of the query global joint concatenation vector or the base database global joint concatenation vector, corresponding to, This refers to either the global join concatenated vector of the query or the global join concatenated vector of the base database. for Nonlinear activation function For the learnable gated first mapping weight matrix, To and The corresponding bias vector; in this embodiment, The dimensions are 512×1536. The dimension is 512; S22, using the Softmax activation operator based on a zero-sum game competition mechanism, dynamically and competitively allocate weights to the global multi-channel context feature vectors of the query global joint concatenation vector and the base database global joint concatenation vector, respectively, to obtain the routing weights of each feature channel in the global semantic expression of the query global joint concatenation vector and the base database global joint concatenation vector; wherein, the formula for competitive weight allocation calculation is: ; In the formula, ,and For the query global joint concatenation vector or the base database global joint concatenation vector, the first... The routing weights of path feature channels in the global semantic representation. , The total number of heterogeneous feature fields in the industrial query text or the standard document base library; For the Softmax activation operator, For learnable gated second mapping weight matrix, To and The corresponding bias vector; in this embodiment, These represent the dynamic weights of the four feature channels, and naturally satisfy... Normalized hard constraints, Its dimensions are 4×512. The dimension is 4. S23, using the routing weights of each feature channel in the global semantic expression of the query global joint concatenation vector and the base database global joint concatenation vector, weighted summation is performed on the local spatially dense features of each heterogeneous feature field in the industrial query text and the standard document base database, respectively. Normalization yields the initial global fusion representation vector for the query and the initial global fusion representation vector for the base database; wherein, weighted summation and... The normalization formula is: ; In the formula, For the initial query global fusion representation vector or the initial base database global fusion representation vector, corresponding to, For the industrial query text or the first document in the standard document base library Local spatial density features of heterogeneous feature fields in road for Norm.

[0024] Specifically, to prevent long and frequently repeated device description fields (Background noise in long text) This issue highlights the flawed phenomenon of fields with fewer characters but the highest information density during global cascading. To mitigate the flooding effect, this invention constructs a multi-gated decision network. The invention first integrates the globally joint concatenated vectors... The data is fed into a multi-gated decision network and mapped to a low-dimensional latent space using the Tanh nonlinear activation function to extract the global multi-channel contextual feature vector at the current sample granularity. Subsequently, to impose a forced "zero-sum game" competition mechanism among the feature channels, this invention introduces the Softmax activation operator to dynamically predict the importance routing weights of each feature channel in the global semantic representation. Finally, the dynamically generated routing weights are used... Local spatially dense features of heterogeneous feature fields Weighted summation and normalized projection are performed to obtain the initial global fusion representation vector of the query and the initial global fusion representation vector of the base database.

[0025] The above-mentioned S3 specifically refers to the steps of reshaping the latent space decision boundary.

[0026] In S3 above, an offline triplet-supervised sample set based on vertical domain business mechanism constraints is constructed, specifically including: Use any industrial query text as an anchor point sample; Retrieve standard documents from the standard document base that are completely consistent with the defect phenomenon subcategories of the anchor sample, and use them as positive samples. Using a same-professional-different-mechanism mining mechanism, standard documents belonging to the same professional field as the anchor sample but whose corresponding defect phenomenon subcategories are mutually exclusive are retrieved from the standard document base library and used as difficult negative samples. Based on the anchor point samples, the positive samples, and the difficult negative samples, a triplet supervised sample set based on vertical domain business mechanism constraints is constructed.

[0027] Specifically, during the offline fine-tuning training phase, in order to force the model to recognize confusing work orders with highly similar prefixes but completely opposite underlying fault mechanisms, this invention, based on the aforementioned gating fusion, activates a "same specialty, different mechanism" negative sample mining mechanism to reshape the classification decision boundary of the latent space. For any sample used as an anchor point... For query work orders (i.e., any industrial query text), this invention first selects anchor point samples from the standard document base library. Historical work orders (i.e., standard documents) that perfectly match the defect phenomenon subcategories and labels (fault truth labels) are used as positive samples. Then, hard constraint sampling is performed: samples matching the anchor points are retrieved from the standard document base library. Belonging to the exact same "professional field field" (e.g., moving equipment specialty), but its historical work orders with completely different and mutually exclusive "defect phenomenon subcategories labels" are used as difficult negative samples. .

[0028] During model iteration, the base parameters of the underlying MiniLM language encoder... Maintaining a global lock-and-freeze, gradient backpropagation updates are performed only on the mapping matrix parameters of the multi-gated network. The fine-tuned contrastive loss function employs an improved InfoNCE contrastive loss function, which minimizes the latent space geometric distance between the anchor sample and the positive sample, and maximizes the latent space geometric distance between the anchor sample and the hard negative sample. This improved InfoNCE contrastive loss function is expressed as: ; In the formula, For InfoNCE comparison loss, The cosine similarity function is used. For temperature hyperparameters, For the anchor point sample, This is the representation vector of the anchor point sample. For the positive sample, Let be the representation vector of the positive sample. For the aforementioned difficult negative samples, For a difficult negative sample set, is the representation vector of the difficult negative sample.

[0029] In this embodiment, the temperature hyperparameter is fixed at τ=0.05, the optimizer is AdamW, and the initial learning rate is set to... When the model training iterations reach a preset number of iterations or the contrastive loss value on the validation set converges to a preset threshold, an early stopping mechanism is triggered. This locks and saves the current gating weight checkpoint file, successfully establishing a high-resolution partition boundary for heterogeneous faults in the latent space, and outputting the final optimized query global fusion representation vector and base database global fusion representation vector. The overall parameter size of the aforementioned multi-gating decision network is only 591,000 (591K), accounting for less than 1.8% of the total system parameters, ensuring the high efficiency and lightweight nature of the feature fusion layer.

[0030] The above S4 specifically refers to the steps of cross-space dense semantics and sparse literal mixed metric calibration inference.

[0031] The above-mentioned S4 specifically includes: S41, using the FAISS library, a high-dimensional Euclidean distance vector inverted index is built in memory in the global fusion representation vector of the base library; S42, perform K-nearest neighbor search on the query global fusion representation vector in the high-dimensional Euclidean distance vector inverted index to recall the top multiple global fusion representation vectors with the closest semantic distance from the global fusion representation vector in the base database, and take the standard document corresponding to the recalled global fusion representation vector as the candidate document. S43, calculate the dense semantic similarity score between the query global fusion representation vector and the candidate document global fusion representation vector based on dense high-dimensional latent space cosine similarity; wherein, the formula for calculating the dense semantic similarity score is: ; In the formula, The dense semantic similarity score is given. The query global fusion representation vector, This is the global fusion representation vector of the candidate document; S44, based on the traditional inverted index mechanism, calculate the sparse matching score between the industrial query text and the candidate documents based on sparse literal word frequency statistical matching; wherein, the formula for calculating the sparse matching score is: ; In the formula, The score for the sparse matching is given. The industrial query text. The candidate document, The BM25 algorithm model is used; in this embodiment, the core hyperparameters of the BM25 algorithm model are fixed as follows: word frequency saturation parameter. Document length normalization parameter ; S45, the dense semantic similarity score and the sparse matching score are weighted and calibrated using a hybrid metric mechanism containing a composite balance factor to obtain a consistent retrieval ranking list; wherein, the formula for calculating the consistent retrieval ranking list is: ; In the formula, The consistency retrieval sort list. For dense semantic similarity scores, The score for the sparse matching is given. This refers to the composite balance factor.

[0032] Specifically, during the online retrieval and inference phase in the production environment, this invention utilizes a pre-trained multi-gated decision network to offline convert all standard documents in the standard document base into corresponding global fusion representation vectors, and uses the FAISS library to build a high-dimensional Euclidean distance vector inverted index in memory. When a new work order text (i.e., industrial query text) is input, the system calculates its 384-dimensional global fusion representation vector in real time through the pre-trained multi-gated decision network. (i.e., query the global fusion representation vector) and perform a fast K-nearest neighbor (KNN) search in the FAISS index space to recall the top 200 standard documents with the closest semantic distance as candidate documents (Top-200 coarse-ranked candidate document set).

[0033] Subsequently, a fine-ranking hybrid metric engine intervened in parallel to perform cross-space joint scoring and calibration on these 200 coarse-ranking candidate documents: First approach: Calculate the global fusion representation vector for the query. Globally fused representation vectors with candidate documents Dense semantic similarity scores between them are used to capture deep unstructured semantic similarity; The second approach: Based on the traditional inverted index mechanism, calculate the sparse matching score between industrial query text and candidate documents.

[0034] Finally, the fine-ranking scoring engine uses a hybrid metric formula that includes a composite balance factor to linearly weight and fuse the two scores.

[0035] In this invention, the composite balance factor As a key adjustable hyperparameter, the contribution weights of the two scores are balanced through a global mechanism or manual adjustment during system deployment and online inference. Increasing or decreasing the composite balance factor α allows for adaptive calibration of the literal constraints on different categories of short texts. Increasing the composite balance factor... When the system is adjusted, it will be more inclined to capture deep, cross-lexical semantic generalization associations through deep neural networks; when the composite balance factor is reduced... In this case, the system tends to impose strict physical literal matching constraints, thereby explicitly correcting and penalizing the semantically inflated scores caused by identical device references or fixed text prefixes.

[0036] Through this cross-spatial hybrid metric calibration mechanism, the system effectively retains its deep unstructured semantic generalization capabilities while achieving precise defense against false positive mismatches caused by traditional dense retrieval. During online inference, the system's end-to-end P99 quantile response latency remains stable at 196.17 milliseconds (ms), far below the 500ms latency upper limit performance threshold set for real-time interactive systems in industrial automation. The system then uses the final calculated composite score... The candidate document set is sorted in descending order, and the top-ranked historical case with absolute consistency and the corresponding standard maintenance measures are output to the system front end, thus achieving high-resolution and robust intelligent retrieval and sorting of multi-field industrial short texts in a closed loop.

[0037] Example 2: Based on the above-mentioned industrial text retrieval method based on multi-gated fusion and hybrid metrics, the present invention also provides an industrial text retrieval system based on multi-gated fusion and hybrid metrics.

[0038] like Figure 2 As shown, an industrial text retrieval system based on multi-gated fusion and hybrid metrics is applied to the industrial text retrieval method based on multi-gated fusion and hybrid metrics described above, including: The local feature independent encoding module is used to split the industrial query text and the standard document base into multiple heterogeneous feature fields respectively, and use the parameter-shared and frozen dual-tower pre-trained language model to extract the local spatial dense features of each heterogeneous feature field, and then construct the query global joint concatenation vector and the base global joint concatenation vector accordingly. The gating interaction dynamic fusion module is used to dynamically calculate the routing weights of each feature channel in the global semantic expression of the query global joint concatenation vector and the base database global joint concatenation vector based on a multi-gating decision network and a zero-sum game competition mechanism, and then construct the initial query global fusion representation vector and the initial base database global fusion representation vector accordingly. The fine-tuning and calibration module is used to offline construct a triplet supervision sample set based on vertical domain business mechanism constraints, and combine it with the contrastive loss function to guide the fine-tuning training of the multi-gated decision network to construct the mechanism decision boundary, thereby determining the latest weights of the multi-gated decision network in the preset context, and thus calibrating the initial query global fusion representation vector and the initial base database global fusion representation vector respectively, and obtaining the query global fusion representation vector and the base database global fusion representation vector accordingly; The cross-space hybrid metric reasoning module is used to calculate the dense semantic similarity score between the global fusion representation vector of the query and the global fusion representation vector of the base library in the latent space, and simultaneously introduce the sparse matching score between the industrial query text and the standard document base library based on the full text statistical features of the inverted index. Furthermore, the module uses a hybrid metric mechanism with a composite balancing factor to perform weighted calibration on the dense semantic similarity score and the sparse matching score to obtain a consistent retrieval ranking list.

[0039] It should be noted that the specific functions of each module in the industrial text retrieval system based on multi-gated fusion and hybrid metric of the present invention can be found in the specific steps of the industrial text retrieval method based on multi-gated fusion and hybrid metric of the present invention, and will not be repeated here.

[0040] In summary, the industrial text retrieval method and system based on multi-gated fusion and hybrid metrics of the present invention have the following beneficial effects: (1) Fine-grained dynamic routing and adaptive noise reduction for multiple heterogeneous fields in industry have been achieved. This invention changes the traditional hard splicing or manual fixed weight paradigm. By using a gated projection network with Softmax competitive operator, it can adaptively identify and suppress background noise caused by long texts such as lengthy equipment tag numbers according to the specific work order text, and prevent the core short text defect feature stream from being diluted.

[0041] (2) Effectively overcomes the problems of false positive mismatch and semantic generalization collapse in vertical domains. By explicitly embedding a dense and sparse hybrid measurement mechanism across spaces in the inference retrieval module, the discrete literal word frequency statistical score is used as the absolute calibration anchor point. Hard penalties are imposed on the inflated semantic cosine score caused by fixed prefixes and long text inclusion, which greatly reduces the false positive recall rate of the system.

[0042] (3) Eliminates sampling collisions in industrial vertical contrastive learning and broadens the decision boundary. This invention innovatively proposes a difficult negative sample construction strategy of "same profession, different mechanism", which effectively prevents the conflict of mistakenly treating work orders with similar key semantics but completely different processing mechanisms as negative samples, and greatly improves the mechanism resolution of the model when processing highly homogeneous industrial text.

[0043] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An industrial text retrieval method based on multiple gate fusion and hybrid metrics, characterized in that, include: The industrial query text and the standard document base are split into multiple heterogeneous feature fields, and the local spatial dense features of each heterogeneous feature field are extracted using a parameter-shared and frozen dual-tower pre-trained language model. Then, the query global joint concatenation vector and the base database global joint concatenation vector are constructed accordingly. Based on a multi-gated decision network and a zero-sum game competition mechanism, the routing weights of each feature channel in the global semantic expression of the query global joint concatenation vector and the base database global joint concatenation vector are dynamically calculated respectively, and then the initial query global fusion representation vector and the initial base database global fusion representation vector are constructed accordingly. Offline construction of a triplet-based supervised sample set constrained by vertical domain business mechanisms is performed, and the contrastive loss function is used to guide the fine-tuning training of the multi-gated decision network to construct the mechanism decision boundary. This determines the latest weights of the multi-gated decision network under a preset context, thereby calibrating the initial query global fusion representation vector and the initial base database global fusion representation vector, resulting in the corresponding query global fusion representation vector and base database global fusion representation vector. The offline construction of the triplet-based supervised sample set constrained by vertical domain business mechanisms specifically includes: using arbitrary industrial query text as anchor samples; retrieving samples from the standard document base database that match the... Standard documents with identical defect subcategories labels to the anchor samples are used as positive samples. A same-profession, different-mechanism mining mechanism is used to retrieve standard documents from the standard document database that belong to the same professional field as the anchor samples but have mutually exclusive defect subcategories labels, which are then used as difficult negative samples. Based on the anchor samples, the positive samples, and the difficult negative samples, a triplet supervised sample set constrained by vertical domain business mechanisms is constructed. During fine-tuning training, the base parameters of the dual-tower pre-trained language model are globally locked and frozen, and gradient backpropagation updates are performed only on the mapping matrix parameters of the multi-gated decision network. Calculate the dense semantic similarity score between the global fusion representation vector of the query and the global fusion representation vector of the base library in the latent space, and simultaneously introduce the sparse matching score between the industrial query text and the standard document base library based on the full text statistical features of the inverted index. Then, use a hybrid measurement mechanism with a composite balancing factor to perform weighted calibration on the dense semantic similarity score and the sparse matching score to obtain a consistent retrieval ranking list. 2.The industrial text retrieval method based on multiple gating fusion and hybrid metrics according to claim 1, wherein, The local spatially dense features of various heterogeneous feature fields are extracted using a parameter-shared and frozen dual-tower pre-trained language model, specifically including: Each heterogeneous feature field is processed into a term, so that control tags and sentence break tags are explicitly inserted at the beginning and end of each heterogeneous feature field; After the heterogeneous feature fields of each path are processed into terms, they are input into a dual-tower pre-trained language model with shared parameters and frozen, and multi-layer self-attention context interaction is performed to extract the final hidden state vector corresponding to the first and second control tags of each heterogeneous feature field, and the local spatial dense features of each heterogeneous feature field are obtained accordingly. The formula for bidirectional self-attention context interaction is as follows: ； In the formula, For the first Local spatial density features of heterogeneous feature fields in road For a pre-trained language model with shared parameters and already frozen, These are the shared parameters in the dual-tower pre-trained language model that are completely frozen and do not participate in gradient updates. For the control flag, This is a function for word segmentation. For the first heterogeneous feature fields of roads The sentence segmentation marker, This is the final hidden state vector corresponding to [CLS] extracted after the multi-layer self-attention context interaction. Let be the set of all real numbers. Indicates the dimension of the heterogeneous feature field. 3.The industrial text retrieval method based on multiple gate fusion and hybrid metrics of claim 1, wherein, Constructing the query global joint concatenation vector and the base database global joint concatenation vector specifically includes: The local spatial dense features of each heterogeneous feature field in the industrial query text are concatenated by tensors to obtain the global joint concatenated vector of the query. The local spatially dense features of each heterogeneous feature field in the standard document base library are concatenated by tensors to obtain the global joint concatenated vector of the base library; The formula for tensor cascading is as follows: ； In the formula, is the corresponding, is the total number of paths of the heterogeneous feature fields in the industrial query text or the standard document corpus. is the local spatial dense feature of the heterogeneous feature field, , is the total number of paths of the heterogeneous feature fields in the industrial query text or the standard document corpus. 4.The industrial text retrieval method based on multiple-gated fusion and hybrid metrics of claim 1, wherein, Based on a multi-gated decision network and a zero-sum game competition mechanism, the routing weights of each feature channel in the global semantic representation of the query global joint concatenation vector and the base database global joint concatenation vector are dynamically calculated, thereby constructing the initial query global fusion representation vector and the initial base database global fusion representation vector, specifically including: A multi-gated decision network based on nonlinear gated projection is used to perform nonlinear projection on the query global joint concatenation vector and the base database global joint concatenation vector, respectively, so as to extract the global multi-channel context feature vectors of the query global joint concatenation vector and the base database global joint concatenation vector. The Softmax activation operator, based on a zero-sum game competition mechanism, is used to dynamically and competitively allocate weights to the global multi-channel context feature vectors of the query global joint concatenation vector and the base database global joint concatenation vector, respectively, thereby obtaining the routing weights of each feature channel in the global semantic expression of the query global joint concatenation vector and the base database global joint concatenation vector. Using the routing weights of each feature channel in the global semantic representation of the query global joint concatenation vector and the base database global joint concatenation vector, weighted summations are performed on the local spatially dense features of each heterogeneous feature field in the industrial query text and the standard document base database, respectively. Normalization yields the initial query global fusion representation vector and the initial base database global fusion representation vector.

5. The industrial text retrieval method based on multi-gated fusion and hybrid metrics according to claim 4, characterized in that, The formula for performing nonlinear projection is: ； In the formula, For the global multi-channel context feature vector of the query global joint concatenation vector or the base database global joint concatenation vector, corresponding to, This refers to either the global join concatenated vector of the query or the global join concatenated vector of the base database. for Nonlinear activation function For the learnable gated first mapping weight matrix, To and The corresponding bias vector; The formula for calculating the competitive allocation weights is: ； In the formula, ,and For the query global joint concatenation vector or the base database global joint concatenation vector, the first... The routing weights of path feature channels in the global semantic representation. , The total number of heterogeneous feature fields in the industrial query text or the standard document base library; For the Softmax activation operator, For learnable gated second mapping weight matrix, To and The corresponding bias vector; Perform weighted summation and The normalization formula is: ； In the formula, For the initial query global fusion representation vector or the initial base database global fusion representation vector, corresponding to, For the industrial query text or the first document in the standard document base library Local spatial density features of heterogeneous feature fields in road for Norm.

6. The industrial text retrieval method based on multi-gated fusion and hybrid metric as described in claim 1, characterized in that, The triplet supervision sample set includes anchor samples, positive samples, and difficult negative samples; The contrastive loss function is specifically an improved InfoNCE contrastive loss function, which is used to minimize the latent space geometric distance between the anchor sample and the positive sample, and to maximize the latent space geometric distance between the anchor sample and the hard negative sample. The improved InfoNCE contrastive loss function is expressed as follows: ； In the formula, For InfoNCE comparison loss, The cosine similarity function is used. For temperature hyperparameters, For the anchor point sample, This is the representation vector of the anchor point sample. For the positive sample, Let be the representation vector of the positive sample. For the aforementioned difficult negative samples, For a difficult negative sample set, is the representation vector of the difficult negative sample.

7. The industrial text retrieval method based on multi-gated fusion and hybrid metrics according to claim 1, characterized in that, Calculate the dense semantic similarity score in the latent space between the query global fusion representation vector and the base database global fusion representation vector, and simultaneously introduce the sparse matching score between the industrial query text and the standard document base database based on the inverted index full-text literal statistical features, specifically including: A high-dimensional Euclidean distance vector inverted index is built in memory in the global fusion representation vector of the base library using the FAISS library; The query global fusion representation vector is subjected to K-nearest neighbor search in the high-dimensional Euclidean distance vector inverted index to recall the top few global fusion representation vectors with the closest semantic distance from the base global fusion representation vectors, and the standard documents corresponding to the recalled global fusion representation vectors are used as candidate documents. Calculate the dense semantic similarity score based on dense high-dimensional latent space cosine similarity between the query global fusion representation vector and the candidate document global fusion representation vector; Based on the traditional inverted index mechanism, the sparse matching score between the industrial query text and the candidate documents is calculated based on sparse literal word frequency statistical matching. The formula for calculating the dense semantic similarity score is as follows: ； In the formula, The dense semantic similarity score is given. The global fusion representation vector for the query. This is the global fusion representation vector of the candidate document; The formula for calculating the sparse matching score is as follows: ； In the formula, The score for the sparse matching is given. The industrial query text. The candidate document, This is the BM25 algorithm model.

8. The industrial text retrieval method based on multi-gated fusion and hybrid metric as described in claim 1, characterized in that, The formula for weighted calibration of the dense semantic similarity score and the sparse matching score using a hybrid metric mechanism with a composite balance factor is as follows: ； In the formula, The consistency retrieval sort list. For dense semantic similarity scores, The score for the sparse matching is given. This refers to the composite balance factor.

9. An industrial text retrieval system based on multi-gating fusion and hybrid metrics, characterized in that, The industrial text retrieval method based on multi-gated fusion and hybrid metrics as described in any one of claims 1 to 8 includes: The local feature independent encoding module is used to split the industrial query text and the standard document base into multiple heterogeneous feature fields respectively, and use the parameter-shared and frozen dual-tower pre-trained language model to extract the local spatial dense features of each heterogeneous feature field, and then construct the query global joint concatenation vector and the base global joint concatenation vector accordingly. The gating interaction dynamic fusion module is used to dynamically calculate the routing weights of each feature channel in the global semantic expression of the query global joint concatenation vector and the base database global joint concatenation vector based on a multi-gating decision network and a zero-sum game competition mechanism, and then construct the initial query global fusion representation vector and the initial base database global fusion representation vector accordingly. The fine-tuning and calibration module is used to offline construct a triplet supervision sample set based on vertical domain business mechanism constraints, and combine it with the contrastive loss function to guide the fine-tuning training of the multi-gated decision network to construct the mechanism decision boundary, thereby determining the latest weights of the multi-gated decision network in the preset context, and thus calibrating the initial query global fusion representation vector and the initial base database global fusion representation vector respectively, and obtaining the query global fusion representation vector and the base database global fusion representation vector accordingly; The cross-space hybrid metric reasoning module is used to calculate the dense semantic similarity score between the global fusion representation vector of the query and the global fusion representation vector of the base library in the latent space, and simultaneously introduce the sparse matching score between the industrial query text and the standard document base library based on the full text statistical features of the inverted index. Furthermore, the module uses a hybrid metric mechanism with a composite balancing factor to perform weighted calibration on the dense semantic similarity score and the sparse matching score to obtain a consistent retrieval ranking list.

Images