Intelligent Governance Method for Asset Data Quality Based on Rule Engine and Deep Learning

Abstract

This invention discloses an intelligent asset data quality governance method based on a rule engine and deep learning, comprising the following steps: S1, collecting asset field data and generating a field sequence tensor; S2, calling the rule engine to generate a rule label sequence; S3, extracting contextual features and concatenating rule labels to construct a rule fusion input vector sequence; S4, generating a fusion feature representation through a rule attention gating mechanism; S5, inputting the fusion features into anomaly identification, field repair, and field classification sub-networks, outputting anomaly judgment, repair value, and category label; S6, constructing a joint loss based on the three types of outputs and performing gradient optimization; S7, applying the trained model to perform error correction and classification processing on newly added field data. This invention achieves efficient asset data governance and improves the accuracy of anomaly identification and the automation level of field repair.

Description

Technical Field

[0001] This invention relates to the field of asset data governance technology, and in particular to an intelligent asset data quality governance method based on rule engines and deep learning. Background Technology

[0002] In the process of asset management informatization, the quality of asset field data has a critical impact on the accuracy of accounting, resource allocation, statistical accounting and other business operations. In the existing technology, the governance of asset data quality issues mainly relies on manually defined logic based on rule engines. Rule engines usually judge the field format, mandatory fields, value range and cross-field logical relationships by constructing field-level validation logic, regular expression matching, and logical consistency judgment, and output the location of abnormal fields and the corresponding error type. These methods are effective in handling scenarios with a high degree of structure and unified field definition standards. However, when the types of fields are complex, the sources are diverse and the expression habits are different, the types of problems that the rule engine can cover are limited.

[0003] To improve coverage, some systems have introduced machine learning-based auxiliary anomaly detection models. These models typically construct field feature vectors and train classifiers to determine whether data anomalies exist. However, due to a lack of sufficient labeled samples, model training is prone to overfitting, and the expression of anomaly samples is irregular, resulting in unstable and uninterpretable detection results. Some existing methods attempt to introduce multi-task learning structures to model anomaly detection, field prediction, and label classification tasks in parallel. However, they lack a structural representation of rule information in the input design, and the model cannot effectively integrate existing rule resources during the learning process, resulting in the neglect of known logic, increased learning costs, and difficulty in aligning output results.

[0004] Existing technologies still suffer from the problem of disconnect between model output and actual error correction operations. Some models can output anomaly judgments but cannot provide usable field repair suggestions or field attribution information, which means that subsequent processing still needs to rely on manual confirmation. Although rule engines can construct repair suggestion logic, they are difficult to cover combinational errors in complex business scenarios. Although the model part has generalization ability, it cannot give accurate predictions for specific repair values. Currently, there is a lack of a method that can deeply integrate rule results with model structure and achieve collaborative optimization of rule-driven and context learning during the training phase.

[0005] Furthermore, existing multi-task network architectures generally use a unified feature representation to input each sub-network, which ignores the differences in feature dimensions that different tasks focus on. Inter-task interference is significant, leading to limited overall performance. In terms of loss function construction, most existing solutions use a simple weighted approach to combine the losses of multiple sub-tasks, lacking inter-task balancing mechanisms and dynamic weight adjustment strategies, which is detrimental to model training convergence and result consistency control. Current technologies have not yet established a complete technical system for asset field structure data that can combine rule labels, contextual semantics, and multi-task error correction output.

[0006] Therefore, how to provide intelligent governance methods for asset data quality based on rule engines and deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose an intelligent governance method for asset data quality based on rule engines and deep learning. This invention constructs a rule-guided multi-task error correction network to achieve integrated processing of field anomaly identification, content repair, and type classification. It has the advantages of high structural fusion accuracy, strong task output consistency, and strong adaptability to asset diversity.

[0008] The intelligent governance method for asset data quality based on rule engines and deep learning according to embodiments of the present invention includes the following steps:

[0009] S1. Collect structured field data from the asset data table, construct an asset field sample set containing field content, field order, and field attributes, and preprocess to generate an asset field sequence tensor;

[0010] S2. Based on the field content and field attributes in the asset field sample set, call the rule engine to perform rule matching and generate a rule label sequence;

[0011] S3. Input the asset field sequence tensor into the bidirectional gated recurrent unit network, extract the context feature representation of each field position, and concatenate it with the rule label sequence of the corresponding position to generate the rule fusion input vector sequence.

[0012] S4. Input the rule fusion input vector sequence into the rule attention gating unit, calculate the attention weight between the context features and the rule labels, adjust the rule fusion input vector sequence according to the attention weight, generate the rule gating fusion feature representation, and build a multi-task error correction network. The multi-task error correction network includes an anomaly identification sub-network, a field repair sub-network, and a field classification sub-network.

[0013] S5. Input the rule-gated feature representation into the anomaly recognition subnetwork and output the anomaly judgment result for each field position. Input the field repair subnetwork and output the recommended repair value for the corresponding field position. Input the field classification subnetwork and output the category label of the field.

[0014] S6. Based on the anomaly judgment result, recommended repair value and category label of each field, construct anomaly recognition loss, field repair loss and field classification loss, calculate the joint loss value, and perform joint gradient optimization to update the parameters of the multi-task error correction network.

[0015] S7. Apply the trained multi-task error correction network to the newly added asset field data input. Based on the output anomaly judgment results, recommended repair values ​​and field category labels, complete the anomaly identification, field repair and type classification of the asset field.

[0016] Optionally, the preprocessing process in S1 includes: performing character-level segmentation on the content of each field and encoding it into a fixed-length word vector; converting the field attributes of each field into attribute embedding vectors; encoding the sequential position of each field in the data table into a position vector; concatenating the word vectors, attribute embedding vectors, and position vectors by dimension to generate a field embedding representation; and combining all field embedding representations into a field embedding sequence according to the field order to form an asset field sequence tensor.

[0017] Optionally, the process of calling the rule engine to perform rule matching in S2 includes: calling the rule subset corresponding to the attribute type according to the field attribute, performing rule expression matching operation on each field content, outputting a boolean result of whether the rule expression is matched, forming a rule match boolean vector by combining the matching results of each field under all rule subsets, mapping the rule match boolean vector to a rule tag encoding vector according to the predefined rule numbering system, and arranging the rule tag encoding vector corresponding to each field in the field order to generate a rule tag sequence.

[0018] Optionally, the characteristic of step S3 is that it specifically includes:

[0019] S31. Input the asset field sequence tensor into the embedding encoding layer in the field order, and perform dimension alignment and concatenation on the content vector, attribute vector and position vector of each field position to generate an embedded feature vector sequence.

[0020] S32. Input the embedded feature vector sequence into the forward gated loop unit, recursively pass the input vector and the previous position state according to the field arrangement order, and output the forward state vector sequence.

[0021] S33. Input the embedded feature vector sequence into the reverse gated loop unit, recursively pass the input vector and the state of the next position according to the reverse order of the fields, and output the reverse state vector sequence.

[0022] S34. Concatenate the feature dimensions of the forward and backward state vectors corresponding to each field position to generate a context feature representation sequence.

[0023] S35. Concatenate the context feature representation sequence with the rule label sequence of the corresponding field according to their positions to form the rule fusion input vector sequence.

[0024] Optionally, the characteristic of step S4 is that it specifically includes:

[0025] S41. Input the rule fusion input vector sequence into the attention weight calculation unit in the order of fields, and generate the field attention weight coefficient sequence based on the vector similarity between the context feature sub-vector and the rule label sub-vector of each field.

[0026] S42. Normalize the attention weight coefficients for each field position to generate a field weight vector within a unit interval, which measures the proportion of information of the rule label in the context representation.

[0027] S43. Using the field attention weight coefficient sequence as the weight input, perform element-level weighting on the rule fusion input vector sequence to adjust the feature distribution intensity of each field position and generate a weighted fusion feature vector sequence.

[0028] S44. Combine the weighted fusion feature vector sequence according to the field order, and output the rule-gated fusion feature representation, which serves as the common input of the anomaly identification subnetwork, the field repair subnetwork, and the field classification subnetwork.

[0029] Optionally, the characteristic of step S5 is that it specifically includes:

[0030] S51. The feature vectors of each field position in the rule-gated fusion feature representation are sequentially input into the anomaly recognition sub-network. The anomaly recognition sub-network includes an input mapping layer, a discriminant coding layer and a classification output layer. The input mapping layer receives the feature vectors and performs feature compression. The discriminant coding layer performs feature transformation on the compressed vectors to generate anomaly indication representations. The classification output layer outputs the field anomaly judgment value based on the anomaly indication representation. The anomaly judgment value is 0 or 1, indicating whether there is a structural or content anomaly in the current field.

[0031] S52. Input the feature vector of each field position in the rule-gated fusion feature representation into the field repair sub-network in sequence. The field repair sub-network contains a state decoding layer and a reconstruction generation layer. The state decoding layer performs semantic restoration on the feature vector. The reconstruction generation layer outputs the recommended repair value of the current field position based on the semantic restoration result. The recommended repair value is consistent with the original field data type and retains the position index order.

[0032] S53. Input the feature vector of each field position in the rule-gated fusion feature representation into the field classification sub-network in sequence. The field classification sub-network includes a category mapping layer and a category prediction layer. The category mapping layer maps the feature vector to the category embedding space. The category prediction layer performs multi-classification operation on the category embedding vector and outputs the category label of the field corresponding to the current field position. The category label of the field comes from the preset field category set, while retaining the label sequence that is consistent with the order of the input fields.

[0033] Optionally, the characteristic of step S6 is that it specifically includes:

[0034] S61. The anomaly judgment results output by the anomaly recognition subnetwork are matched one-to-one with the standard anomaly labeling results according to the field position. The recognition error of each field position is calculated. The binary classification cross-entropy loss function is used to average the error values ​​of all field positions to generate the anomaly recognition loss value.

[0035] S62. Align the recommended repair values ​​and standard field values ​​output by the field repair subnetwork according to the field position, construct the repair residual tensor, and use the mean square error function to square and sum the repair residuals of all field positions and take the average to generate the field repair loss value.

[0036] S63. Map the field category labels output by the field classification subnetwork to the standard field category labels by position index, calculate the multi-class classification error, and generate the field classification loss value using the multi-class cross-entropy loss function;

[0037] S64. Let the anomaly identification loss value be... The field repair loss value is The field classification loss value is Let the weighting coefficients be respectively. , , Joint loss value The calculation method is as follows:

[0038] ;

[0039] S65, Combined loss values Input the training engine, perform backpropagation based on the joint loss value, calculate the gradient values ​​of all updatable parameters in the multi-task error correction network, and complete an iterative update operation of the network parameters in combination with the learning rate.

[0040] Optionally, the process of calculating the gradient values ​​of all updatable parameters in the multi-task error correction network in S65 includes: converting the joint loss value... Input the training engine, perform partial derivative calculations on all trainable parameters in the multi-task error correction network, obtain the gradient value tensor at each parameter position, scale the gradient value tensor according to the set learning rate coefficient, generate the update amount corresponding to each parameter, use tensor subtraction operation to subtract the current value of each parameter from the corresponding update amount to obtain the optimized parameter value, write the optimized parameter value into the structure of the multi-task error correction network to complete the parameter replacement, and write the joint loss value, subtask loss value, parameter change magnitude and gradient average value of the current round into the training process log.

[0041] Optionally, the process of anomaly identification, field repair, and type classification of asset fields in S7 includes: using the anomaly judgment result to mark the location of suspicious fields, using the recommended repair value to replace the content of abnormal fields or provide a reference for manual verification, and using the category label to which the field belongs to construct structured classification information of field types.

[0042] The beneficial effects of this invention are:

[0043] This invention achieves deep fusion of rules and contextual features: by constructing a rule label sequence and introducing a rule attention gating mechanism, the rule engine recognition results are used in the feature representation generation process in a structured form, solving the problem that traditional deep models cannot call rule priors and improving the model's ability to identify explicit errors and structural conflicts.

[0044] This invention improves the comprehensiveness and accuracy of field error correction output: by designing a multi-task error correction network structure, it incorporates three tasks—anomaly identification, field repair, and field classification—into a unified network architecture. Based on feature sharing, it performs task branch modeling, ensuring the consistency of field position and the semantic availability of output, and meeting the synchronous requirements for anomaly location, repair suggestions, and label classification in the scenario of batch governance of asset data.

[0045] This invention enhances the stability and controllability of model training by employing a joint loss function construction method, structurally combining the losses of each sub-task, and introducing a weight factor adjustment mechanism. This ensures that the model training process maintains a convergence balance among multiple task objectives, avoids accuracy degradation caused by task interference, and improves network training efficiency and generalization ability. Attached Figure Description

[0046] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0047] Figure 1 This is an overall flowchart of the intelligent governance method for asset data quality based on rule engine and deep learning proposed in this invention.

[0048] Figure 2 This is a diagram of the multi-task error correction network structure for the intelligent governance method of asset data quality based on rule engine and deep learning proposed in this invention, which introduces a rule attention gating mechanism.

[0049] Figure 3 This is a flowchart illustrating the multi-task output and joint loss optimization process of the intelligent governance method for asset data quality based on rule engines and deep learning proposed in this invention. Detailed Implementation

[0050] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0051] refer to Figure 1-3 The intelligent governance method for asset data quality based on rule engines and deep learning includes the following steps:

[0052] S1. Collect structured field data from the asset data table, construct an asset field sample set containing field content, field order, and field attributes, and preprocess to generate an asset field sequence tensor;

[0053] S2. Based on the field content and field attributes in the asset field sample set, call the rule engine to perform rule matching and generate a rule label sequence;

[0054] S3. Input the asset field sequence tensor into the bidirectional gated recurrent unit network, extract the context feature representation of each field position, and concatenate it with the rule label sequence of the corresponding position to generate the rule fusion input vector sequence.

[0055] S4. Input the rule fusion input vector sequence into the rule attention gating unit, calculate the attention weight between the context features and the rule labels, adjust the rule fusion input vector sequence according to the attention weight, generate the rule gating fusion feature representation, and build a multi-task error correction network. The multi-task error correction network includes an anomaly identification sub-network, a field repair sub-network, and a field classification sub-network.

[0056] S5. Input the rule-gated feature representation into the anomaly recognition subnetwork and output the anomaly judgment result for each field position. Input the field repair subnetwork and output the recommended repair value for the corresponding field position. Input the field classification subnetwork and output the category label of the field.

[0057] S6. Based on the anomaly judgment result, recommended repair value and category label corresponding to each field, construct anomaly recognition loss, field repair loss and field classification loss, calculate the joint loss value, and perform joint gradient optimization to update the parameters of the multi-task error correction network.

[0058] S7. Apply the trained multi-task error correction network to the newly added asset field data input. Based on the output anomaly judgment results, recommended repair values ​​and field category labels, complete the anomaly identification, field repair and type classification of the asset field.

[0059] This invention addresses the issue of asset field data quality by proposing a method that integrates a rule engine with a deep learning structure. This method achieves an integrated processing flow for data anomaly identification, field repair, and type classification. By constructing sequence tensor input, rule label guidance, attention control mechanism, and multi-task error correction output, a complete intelligent governance closed loop is formed. This method improves the accuracy of field anomaly detection and the usability of repair results. At the same time, it overcomes the problems of limited coverage and poor generalization ability of traditional rule-driven methods, and has the advantages of scalable structure and strong adaptability to asset diversity.

[0060] In this embodiment, the preprocessing process in S1 includes: performing character-level segmentation on the content of each field and encoding it into a fixed-length word vector; converting the field attributes of each field into attribute embedding vectors; encoding the sequential position of each field in the data table into a position vector; concatenating the word vectors, attribute embedding vectors, and position vectors by dimension to generate a field embedding representation; and combining all field embedding representations into a field embedding sequence according to the field order to form an asset field sequence tensor.

[0061] This invention designs a standardized tensor construction process for the preprocessing of field samples and proposes a joint concatenation mechanism of content vectors, attribute vectors, and position encoding vectors. This ensures that the input structure retains the relative relationships and attribute expressions between fields during context learning. Compared with the existing technology that constructs vectors based solely on field values, this invention achieves the synchronous embedding of field semantics, structural attributes, and position information, providing a stable input foundation for subsequent context feature extraction and rule fusion, and enhancing the ability of sequence models to model structured data.

[0062] In this embodiment, the operation process of calling the rule engine to perform rule matching in S2 includes: calling the rule subset corresponding to the attribute type according to the field attribute, performing rule expression matching operation on each field content, outputting a boolean result of whether the rule expression is matched, forming a rule match boolean vector by combining the matching results of each field under all rule subsets, mapping the rule match boolean vector to a rule tag encoding vector according to the predefined rule numbering system, and arranging the rule tag encoding vector corresponding to each field in the field order to generate a rule tag sequence.

[0063] In the process of generating rule labels, this invention proposes an operation flow of calling rule subsets according to field attributes, constructing Boolean hit vectors and mapping them to label codes, thereby realizing the vectorized expression of rule results. By introducing a rule numbering system, this invention realizes the structural mapping of the rule engine output results in the neural network. This breaks through the problem that traditional rule results cannot be embedded in the model structure. This mechanism makes the rule engine a part of the model training data, improving the model's ability to respond to known error patterns and the guiding effect of rules on model training.

[0064] In this embodiment, the characteristic is that S3 specifically includes:

[0065] S31. Input the asset field sequence tensor into the embedding encoding layer in the field order, and perform dimension alignment and concatenation on the content vector, attribute vector and position vector of each field position to generate an embedded feature vector sequence.

[0066] S32. Input the embedded feature vector sequence into the forward gated loop unit, recursively pass the input vector and the previous position state according to the field arrangement order, and output the forward state vector sequence.

[0067] S33. Input the embedded feature vector sequence into the reverse gated loop unit, recursively pass the input vector and the state of the next position according to the reverse order of the fields, and output the reverse state vector sequence.

[0068] S34. Concatenate the feature dimensions of the forward and backward state vectors corresponding to each field position to generate a context feature representation sequence.

[0069] S35. Concatenate the context feature representation sequence with the rule label sequence of the corresponding field according to their positions to form the rule fusion input vector sequence.

[0070] This invention addresses the structure of context feature extraction and rule label fusion by designing a bidirectional gated recurrent network combined with a rule label concatenation input mechanism. Based on the concatenation of forward and backward states, rule-dimensional features are introduced to enable the semantic representation of fields to have sequential context and rule orientation. This design differs from the traditional approach of context extraction that relies solely on sequence modeling, and can improve the sensitivity of field vectors to rule conflicts and field type anomalies. This also provides a more controllable vector input foundation for subsequent attention mechanism computation.

[0071] In this embodiment, the characteristic is that S4 specifically includes:

[0072] S41. Input the rule fusion input vector sequence into the attention weight calculation unit in the order of fields, and generate the field attention weight coefficient sequence based on the vector similarity between the context feature sub-vector and the rule label sub-vector of each field.

[0073] S42. Normalize the attention weight coefficients for each field position to generate a field weight vector within a unit interval, which measures the proportion of information of the rule label in the context representation.

[0074] S43. Using the field attention weight coefficient sequence as the weight input, perform element-level weighting on the rule fusion input vector sequence to adjust the feature distribution intensity of each field position and generate a weighted fusion feature vector sequence.

[0075] S44. Combine the weighted fusion feature vector sequence according to the field order, and output the rule-gated fusion feature representation, which serves as the common input of the anomaly identification subnetwork, the field repair subnetwork, and the field classification subnetwork.

[0076] This invention proposes a weighted fusion method for input vectors based on the similarity between rule labels and context vectors in the rule attention gating mechanism. This effectively controls the guiding role of rule information in the model structure. Compared with the problem that existing attention mechanisms cannot explain the source of rules, this invention constructs a structural path in which rule weights participate in the reconstruction of input features, so that the model gives higher attention to the fields hit by the rules, which will improve the accuracy of field recognition.

[0077] In this embodiment, the characteristic is that S5 specifically includes:

[0078] S51. The feature vectors of each field position in the rule-gated fusion feature representation are sequentially input into the anomaly recognition sub-network. The anomaly recognition sub-network includes an input mapping layer, a discriminant coding layer and a classification output layer. The input mapping layer receives the feature vectors and performs feature compression. The discriminant coding layer performs feature transformation on the compressed vectors to generate anomaly indication representations. The classification output layer outputs the field anomaly judgment value based on the anomaly indication representation. The anomaly judgment value is 0 or 1, indicating whether there is a structural or content anomaly in the current field.

[0079] S52. Input the feature vector of each field position in the rule-gated fusion feature representation into the field repair sub-network in sequence. The field repair sub-network contains a state decoding layer and a reconstruction generation layer. The state decoding layer performs semantic restoration on the feature vector. The reconstruction generation layer outputs the recommended repair value of the current field position based on the semantic restoration result. The recommended repair value is consistent with the original field data type and retains the position index order.

[0080] S53. Input the feature vector of each field position in the rule-gated fusion feature representation into the field classification sub-network in sequence. The field classification sub-network includes a category mapping layer and a category prediction layer. The category mapping layer maps the feature vector to the category embedding space. The category prediction layer performs multi-classification operation on the category embedding vector and outputs the category label of the field corresponding to the current field position. The category label of the field comes from the preset field category set, while retaining the label sequence that is consistent with the order of the input fields.

[0081] This invention designs three independent sub-networks in a multi-task output structure, respectively for anomaly identification, field repair, and field classification. The input maintains consistency with rule-gated fusion features, and the output structure is aligned according to field position, meeting the synchronous requirements for multi-dimensional error correction information in asset data structured governance scenarios. Unlike traditional single-task classification or regression models, this invention ensures correlation and consistency between identification accuracy, repair suggestions, and classification labels through a multi-output structure, improving the model's adaptability and practicality in large-scale batch data processing.

[0082] In this embodiment, the characteristic is that S6 specifically includes:

[0083] S61. The anomaly judgment results output by the anomaly recognition subnetwork are matched one-to-one with the standard anomaly labeling results according to the field position. The recognition error of each field position is calculated. The binary classification cross-entropy loss function is used to average the error values ​​of all field positions to generate the anomaly recognition loss value.

[0084] S62. Align the recommended repair values ​​and standard field values ​​output by the field repair subnetwork according to the field position, construct the repair residual tensor, and use the mean square error function to square and sum the repair residuals of all field positions and take the average to generate the field repair loss value.

[0085] S63. Map the field category labels output by the field classification subnetwork to the standard field category labels by position index, calculate the multi-class classification error, and generate the field classification loss value using the multi-class cross-entropy loss function;

[0086] S64. Let the anomaly identification loss value be... The field repair loss value is The field classification loss value is Let the weighting coefficients be respectively. , , Joint loss value The calculation method is as follows:

[0087] ;

[0088] S65, Combined loss values Input the training engine, perform backpropagation based on the joint loss value, calculate the gradient values ​​of all updatable parameters in the multi-task error correction network, and complete an iterative update operation of the network parameters in combination with the learning rate.

[0089] This invention proposes a joint loss structure based on task partitioning in the construction of the loss function. It calculates the anomaly identification loss, field repair loss, and classification loss separately, and forms the training optimization objective by weighted summation. Compared with the existing multi-task network that uses equal weight combination, this invention introduces an adjustable weight factor to effectively balance the training gradients between tasks and improve the model convergence stability. This design realizes the collaborative optimization of multiple task objectives during the model learning stage, improves training efficiency, and reduces the risk of task conflict.

[0090] In this embodiment, the process of calculating the gradient values ​​of all updatable parameters in the multi-task error correction network in S65 includes: converting the joint loss value... Input the training engine, perform partial derivative calculations on all trainable parameters in the multi-task error correction network, obtain the gradient value tensor at each parameter position, scale the gradient value tensor according to the set learning rate coefficient, generate the update amount corresponding to each parameter, use tensor subtraction operation to subtract the current value of each parameter from the corresponding update amount to obtain the optimized parameter value, write the optimized parameter value into the structure of the multi-task error correction network to complete the parameter replacement, and write the joint loss value, subtask loss value, parameter change magnitude and gradient average value of the current round into the training process log.

[0091] This invention designs a backpropagation mechanism driven by joint loss value during parameter optimization, clearly defining the calculation of loss value, gradient generation, learning rate scaling, parameter update, and optimization log recording process in each training round. It constructs a complete and controllable parameter iteration path. Unlike existing methods where the parameter update process is not fully described and cannot be traced, this invention makes the training process more controllable and auditable by precisely controlling the source of loss and the optimization trajectory, adapting to the requirements of long-term operation and version evolution.

[0092] In this embodiment, the process of anomaly identification, field repair, and type classification of asset fields in S7 includes: using the anomaly judgment result to mark the location of suspicious fields, using the recommended repair value to replace the content of abnormal fields or provide a reference for manual verification, and using the category label to which the field belongs to construct structured classification information of field types.

[0093] This invention proposes an operational process for processing newly added asset fields based on a trained multi-task network during the model application stage. By unifying the input structure of field sequence tensors, it automates the processing of three tasks: anomaly identification, repair value output, and type label generation. Unlike existing methods that only output anomaly identifiers, this invention can generate repair suggestions and field attribution while identifying problems. It supports structured field verification, asset master data correction, and label classification operations, improving the feasibility and processing efficiency in practical applications.

[0094] Example 1:

[0095] To verify the feasibility of this invention in practice, it was applied to a large-scale asset data governance scenario in a large information management platform. This scenario involves quality inspection and repair of millions of basic asset field data. The field types include equipment number, asset classification code, depreciation method, depreciation period, activation time, user department, asset status identifier, etc. Due to the heterogeneity of historical data sources, some fields have obvious format errors, missing field content, logical relationship conflicts, and missing classification tags. Although the original rule engine can cover some fixed format problems, it is powerless to deal with complex error combinations, field semantic mismatches, and missing tags. Manual review is time-consuming, labor-intensive, and has poor consistency.

[0096] In this scenario, the collected asset field data is imported into a structured sample preprocessing workflow. Field embedding vectors are generated based on field content, attributes, and order, and a unified asset field sequence tensor is constructed. Subsequently, a rule engine performs rule matching on the field content to construct a rule label sequence. The asset field sequence tensor and the rule label sequence are then input into a trained multi-task error correction network to perform anomaly detection, field repair, and classification tasks, outputting the error correction results.

[0097] In the anomaly identification subtask, the model effectively identified issues such as illegal encoding, mismatched formats, excessive values, and logical conflicts in the field content, outputting a binary anomaly judgment label for each record. In the test set, there were 35,896 manually labeled real anomaly data entries, and the model correctly identified 34,385 of them, achieving an accuracy rate of 95.79%. In the field repair task, the model generated repair suggestions for abnormal fields based on contextual semantics and rule guidance. Compared with real fields, the correct matching rate of repaired values ​​reached 91.64%. In the field classification task, the model performed type prediction for unlabeled fields, involving 10 field categories, with a classification accuracy rate of 94.08%.

[0098] The original system relied on manual or rule-based judgment to process 10,000 asset data entries, averaging approximately 186 minutes and requiring three different job roles. After applying the method of this invention, the processing time for the same scale of data was reduced to 28 minutes, the error correction accuracy increased to over 93%, and the one-time automatic repair coverage reached 88.7%. Taking some samples as examples, in the original fields, the "Depreciation Code" was D005, but the corresponding "Depreciation Period" was 5 years. The rule hit indicated a logical conflict, and the model correctly corrected the depreciation code to D004. Another record had an "Asset Classification Code" that was unknown; the model, based on the field context and rule annotations, successfully classified it as "High-Voltage Transmission and Transformation." Another field value, "Status Identifier," was "Active," but combined with the user department's records, it should actually be "Scrapped." The model successfully identified and flagged this anomaly. The following is a summary table of the key performance indicators for this batch of processed samples:

[0099] Table 1: Performance Statistics of Multi-Task Error Correction Network in Asset Field Data Governance

[0100] ;

[0101] In the performance statistics shown in the table above, the multi-task error correction network demonstrated high consistency, excellent accuracy, and significantly improved processing efficiency when applied to asset field data in different batches. The first batch of test samples totaled 10,000, and the model identified 9,583 abnormal data points, achieving an anomaly identification accuracy of 95.83%. The number of recommended repair values ​​hit was 8,761. Compared with the verification data after manual annotation, the accuracy of the repair values ​​was 91.64%. The accuracy of the field classification task in this batch was 94.08%, and the overall automatic repair coverage reached 88.70%. The processing time was 28 minutes, which is far lower than the traditional manual process.

[0102] The second batch of samples expanded to 20,000. The model still maintained high stability, with 19,117 anomaly detections and an accuracy rate of 95.59%. 18,345 repair suggestions were corrected with an accuracy rate of 91.73%, and the classification prediction accuracy rate was 93.82%. The overall processing time was 56 minutes, and the unit data processing efficiency did not decrease, indicating that the model has good scalability.

[0103] The third batch of samples increased to 30,000, with 28,645 anomaly detections, achieving an accuracy rate of 95.48%. 27,408 samples were recommended for repair, with a repair accuracy rate of 91.36%. The classification accuracy rate was 94.19%, and the automatic repair coverage rate was 89.11%. The processing time was 83 minutes. As the data scale increased, the model output accuracy and execution time remained stable, supporting deployment and operation in large-scale asset scenarios.

[0104] Compared to the original manual review process, with the same 10,000 samples, traditional manual operation takes about 186 minutes, and the automatic repair coverage is only 32.20%, which is far lower than the model's processing result. Furthermore, it lacks real-time performance and consistency. The model significantly shortens the processing time and improves field verification accuracy while completely replacing manual review. At the same time, it outputs clearly structured repair suggestions and type labels, possessing the technical advantages of being replicable, implementable, and iterative. These data verify the effectiveness and engineering practical value of the method proposed in this invention in asset data governance tasks.

[0105] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent governance method for asset data quality based on rule engines and deep learning, characterized in that, Includes the following steps: S1. Collect structured field data from the asset data table, construct an asset field sample set containing field content, field order, and field attributes, and preprocess to generate an asset field sequence tensor; S2. Based on the field content and field attributes in the asset field sample set, call the rule engine to perform rule matching and generate a rule label sequence; S3. Input the asset field sequence tensor into the bidirectional gated recurrent unit network, extract the context feature representation of each field position, and concatenate it with the rule label sequence of the corresponding position to generate the rule fusion input vector sequence. S4. Input the rule fusion input vector sequence into the rule attention gating unit, calculate the attention weight between the context features and the rule labels, adjust the rule fusion input vector sequence according to the attention weight, generate the rule gating fusion feature representation, and build a multi-task error correction network. The multi-task error correction network includes an anomaly identification sub-network, a field repair sub-network, and a field classification sub-network. S5. Input the rule-gated feature representation into the anomaly recognition subnetwork and output the anomaly judgment result for each field position. Input the field repair subnetwork and output the recommended repair value for the corresponding field position. Input the field classification subnetwork and output the category label of the field. S6. Based on the anomaly judgment result, recommended repair value and category label corresponding to each field, construct anomaly recognition loss, field repair loss and field classification loss, calculate the joint loss value, and perform joint gradient optimization to update the parameters of the multi-task error correction network. S7. Apply the trained multi-task error correction network to the newly added asset field data input. Based on the output anomaly judgment results, recommended repair values ​​and field category labels, complete the anomaly identification, field repair and type classification of the asset field.

2. The intelligent governance method for asset data quality based on rule engine and deep learning according to claim 1, characterized in that, The preprocessing process in S1 includes: performing character-level segmentation on the content of each field and encoding it into a fixed-length word vector; converting the field attributes of each field into attribute embedding vectors; encoding the sequential position of each field in the data table into a position vector; concatenating the word vectors, attribute embedding vectors, and position vectors by dimension to generate a field embedding representation; and combining all the field embedding representations into a field embedding sequence according to the field order to form an asset field sequence tensor.

3. The intelligent governance method for asset data quality based on rule engine and deep learning according to claim 2, characterized in that, The process of calling the rule engine to perform rule matching in S2 includes: calling the rule subset corresponding to the attribute type according to the field attribute, performing rule expression matching operation on the content of each field, outputting a boolean result of whether the rule expression is matched, forming a rule match boolean vector by combining the matching results of each field under all rule subsets, mapping the rule match boolean vector to a rule tag encoding vector according to the predefined rule numbering system, arranging the rule tag encoding vector corresponding to each field in the field order, and generating a rule tag sequence.

4. The intelligent governance method for asset data quality based on rule engine and deep learning according to claim 3, characterized in that, S3 specifically includes: S31. Input the asset field sequence tensor into the embedding encoding layer according to the field order, and perform dimension alignment and concatenation on the content vector, attribute vector and position vector of each field position to generate an embedded feature vector sequence. S32. Input the embedded feature vector sequence into the forward gated loop unit, recursively pass the input vector and the previous position state according to the field arrangement order, and output the forward state vector sequence. S33. Input the embedded feature vector sequence into the reverse gated loop unit, recursively pass the input vector and the state of the next position according to the reverse order of the fields, and output the reverse state vector sequence. S34. Concatenate the feature dimensions of the forward and backward state vectors corresponding to each field position to generate a context feature representation sequence. S35. Concatenate the context feature representation sequence with the rule label sequence of the corresponding field according to their positions to form the rule fusion input vector sequence.

5. The intelligent governance method for asset data quality based on rule engine and deep learning according to claim 4, characterized in that, S4 specifically includes: S41. Input the rule fusion input vector sequence into the attention weight calculation unit in the order of fields, and generate the field attention weight coefficient sequence based on the vector similarity between the context feature sub-vector and the rule label sub-vector of each field. S42. Normalize the attention weight coefficients for each field position to generate a field weight vector within a unit interval, which measures the proportion of information of the rule label in the context representation. S43. Using the field attention weight coefficient sequence as the weight input, perform element-level weighting on the rule fusion input vector sequence to adjust the feature distribution intensity of each field position and generate a weighted fusion feature vector sequence. S44. Combine the weighted fusion feature vector sequence according to the field order, and output the rule-gated fusion feature representation, which serves as the common input of the anomaly identification subnetwork, the field repair subnetwork, and the field classification subnetwork.

6. The intelligent governance method for asset data quality based on rule engine and deep learning according to claim 5, characterized in that, S5 specifically includes: S51. The feature vectors of each field position in the rule-gated fusion feature representation are sequentially input into the anomaly recognition sub-network. The anomaly recognition sub-network includes an input mapping layer, a discriminant coding layer and a classification output layer. The input mapping layer receives the feature vectors and performs feature compression. The discriminant coding layer performs feature transformation on the compressed vectors to generate anomaly indication representations. The classification output layer outputs the field anomaly judgment value based on the anomaly indication representation. The anomaly judgment value is 0 or 1, indicating whether there is a structural or content anomaly in the current field. S52. Input the feature vector of each field position in the rule-gated fusion feature representation into the field repair sub-network in sequence. The field repair sub-network contains a state decoding layer and a reconstruction generation layer. The state decoding layer performs semantic restoration on the feature vector. The reconstruction generation layer outputs the recommended repair value of the current field position based on the semantic restoration result. The recommended repair value is consistent with the original field data type and retains the position index order. S53. Input the feature vector of each field position in the rule-gated fusion feature representation into the field classification sub-network in sequence. The field classification sub-network includes a category mapping layer and a category prediction layer. The category mapping layer maps the feature vector to the category embedding space. The category prediction layer performs multi-classification operation on the category embedding vector and outputs the category label of the field corresponding to the current field position. The category label of the field comes from the preset field category set, while retaining the label sequence that is consistent with the order of the input fields.

7. The intelligent governance method for asset data quality based on rule engine and deep learning according to claim 6, characterized in that, S6 specifically includes: S61. The anomaly judgment results output by the anomaly recognition subnetwork are matched one-to-one with the standard anomaly labeling results according to the field position. The recognition error of each field position is calculated. The binary classification cross-entropy loss function is used to average the error values ​​of all field positions to generate the anomaly recognition loss value. S62. Align the recommended repair values ​​and standard field values ​​output by the field repair subnetwork according to the field position, construct the repair residual tensor, and use the mean square error function to square and sum the repair residuals of all field positions and take the average to generate the field repair loss value. S63. Map the field category labels output by the field classification subnetwork to the standard field category labels by position index, calculate the multi-class classification error, and generate the field classification loss value using the multi-class cross-entropy loss function; S64. Let the anomaly identification loss value be... The field repair loss value is The field classification loss value is Let the weighting coefficients be respectively. , , Joint loss value The calculation method is as follows: ； S65, Combined loss values Input the training engine, perform backpropagation based on the joint loss value, calculate the gradient values ​​of all updatable parameters in the multi-task error correction network, and complete an iterative update operation of the network parameters in combination with the learning rate.

8. The intelligent governance method for asset data quality based on rule engine and deep learning according to claim 7, characterized in that, The process of calculating the gradient values ​​of all updatable parameters in the multi-task error correction network in S65 includes: calculating the joint loss value... Input the training engine, perform partial derivative calculations on all trainable parameters in the multi-task error correction network, obtain the gradient value tensor at each parameter position, scale the gradient value tensor according to the set learning rate coefficient, generate the update amount corresponding to each parameter, use tensor subtraction operation to subtract the current value of each parameter from the corresponding update amount to obtain the optimized parameter value, write the optimized parameter value into the structure of the multi-task error correction network to complete the parameter replacement, and write the joint loss value, subtask loss value, parameter change magnitude and gradient average value of the current round into the training process log.

9. The intelligent governance method for asset data quality based on rule engine and deep learning according to claim 8, characterized in that, The process of anomaly identification, field repair, and type classification of asset fields in S7 includes: using the anomaly judgment result to mark the location of suspicious fields, using the recommended repair value to replace the content of abnormal fields or provide a reference for manual verification, and using the category label of the field to construct structured classification information of field types.

Images