Equipment defect identification method based on multi-modal information fusion and adaptive information optimization

Abstract

The invention discloses an equipment defect identification method based on multi-modal information fusion and self-adaptive information optimization, and aims to solve the problems that a traditional single-modal defect identification method is difficult to effectively fuse multi-source heterogeneous information in a substation inspection log, and the model efficiency is low due to information redundancy. The method comprises the following steps: firstly, constructing a gating level attention multi-modal feature fusion network, and realizing deep integration and cross-type interaction modeling of multi-modal information; secondly, designing an information attribute importance analysis mechanism, and quantifying the contribution degree and redundancy of each information attribute to defect identification; then, constructing a self-adaptive information attribute selection framework, and realizing accurate identification through joint optimization of information attribute screening and model training; and finally, an intelligent identification system covering data processing, feature fusion, attribute optimization and defect prediction is established.

Description

Technical Field

[0001] This invention belongs to the field of substation equipment operation and maintenance, and particularly relates to a method for identifying equipment defects through multimodal information fusion and adaptive information optimization. Background Technology

[0002] As a crucial link in energy supply, the stability of substation equipment directly affects the efficiency of power transmission and distribution. Accurate identification of equipment defects is a core element in ensuring the reliable operation of substations. With the increasing intelligence of power grids, substation inspection logs have accumulated rich, multi-source, heterogeneous information, including text descriptions, numerical parameters, and category attributes. This information contains key diagnostic value and serves as an important data foundation for achieving intelligent defect identification.

[0003] However, existing technologies have many limitations in substation equipment defect identification. First, traditional methods often rely on single-modal information processing, such as semantic analysis based solely on text logs, statistical modeling based solely on sensor values, or rule matching based solely on category attributes. These methods lack the collaborative fusion and in-depth mining of multi-source heterogeneous data, making it difficult to capture the full picture of equipment defects. This single-dimensional approach has significantly insufficient generalization ability when facing scenarios with diverse equipment types and complex fault modes, failing to meet the accuracy requirements of practical engineering. Second, existing multimodal fusion methods often employ simple splicing or shallow feature fusion strategies, failing to effectively model deep semantic relationships between modalities. Furthermore, the feature redundancy problem brought about by high-dimensional heterogeneous data further restricts model performance. In addition, traditional information attribute selection methods are independent of the model training process, making it difficult to balance information effectiveness and redundancy, resulting in high model computational complexity and poor interpretability.

[0004] These shortcomings not only reduce the accuracy and efficiency of equipment defect identification but also hinder the advancement of intelligent substation operation and maintenance. Therefore, there is an urgent need for a substation equipment defect identification method that integrates multimodal information fusion and adaptive information optimization. By integrating multi-source heterogeneous data, constructing a deep fusion mechanism, and employing dynamic attribute selection strategies, this method can achieve accurate defect identification and provide reliable technical support for substation operation and maintenance decisions. Summary of the Invention

[0005] The main objective of this invention is to overcome the shortcomings and deficiencies of existing technologies and provide a method for equipment defect identification based on multimodal information fusion and adaptive information optimization. This method achieves deep integration and cross-type interactive modeling of textual, numerical, and categorical multimodal information by constructing a Gated Hierarchical Attention Network (GHAN); it designs an Information Attribute Importance Analysis Mechanism (IAIAM) to quantify the importance and redundancy of each information attribute using multi-dimensional indicators; and it constructs an Adaptive Information Attribute Selection Framework (AIASF) to jointly optimize information attribute selection and model training. Ultimately, this method achieves accurate identification of substation equipment defects, improves model efficiency and generalization ability, and provides reliable technical support for intelligent substation operation and maintenance.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for identifying equipment defects using multimodal information fusion and adaptive information optimization, characterized by comprising the following steps:

[0008] S1: Retrieves multimodal data from substation equipment to provide data support;

[0009] S2: Construct a gated hierarchical attention-based multimodal feature fusion network based on multimodal data to achieve deep fusion of multimodal information in substation equipment inspection logs;

[0010] S3: Design an information attribute importance analysis mechanism to quantify the importance and redundancy of each information attribute for defect identification;

[0011] S4: Construct an adaptive information attribute selection framework to achieve accurate defect identification through joint optimization of information attribute filtering and model training;

[0012] S5: Establish a substation equipment defect identification system based on multimodal information fusion and adaptive attribute optimization.

[0013] As a preferred technical solution of the present invention: in step S1, the multimodal data includes information modality, equipment inspection log, and information type.

[0014] Information modalities include text modalities, numerical modalities, and categorical modalities.

[0015] Equipment inspection logs include defect descriptions, defect locations, remarks, voltage levels, manufacturers, equipment models, and the source of the defect.

[0016] Information types include descriptive information, equipment information, and management information.

[0017] As a preferred technical solution of the present invention, step S2 is specifically as follows:

[0018] S21. Text modality feature processing, the specific process is as follows:

[0019] S211, Text Sequence Encoding and Context Embedding Generation

[0020] Given a text sequence ,in For words in the text, To determine the number of lexical units, a 12-layer Transformer encoder of a domain-adaptive BERT model is used for semantic feature extraction to generate context-aware embeddings. The calculation formula is as follows:

[0021] (1);

[0022] In the formula, The context embedding matrix output by BERT, with dimension . , As an embedding dimension, this context embedding matrix can capture the contextual dependencies and domain-specific semantics of lexical units in the text;

[0023] S212, Attention Pooling and Fixed-Dimension Feature Aggregation

[0024] To aggregate variable-length sequence representations into fixed-dimensional vectors, attention pooling is applied to the word embeddings output from the last four layers of the BERT model, as follows:

[0025] First, calculate the attention weight for each word, using the following formula:

[0026] (2);

[0027] In the formula, For learnable parameters, For the first Embedding vectors of 1 word element This represents the attention weight of the word, reflecting its importance in the semantic expression of the text;

[0028] Subsequently, the fixed-dimensional text modal features are obtained by weighted summation, as shown in the formula:

[0029] (3);

[0030] In the formula, This is the final feature representation of the text modality, which can focus on the key semantic information in the text, providing a unified dimensional text feature foundation for subsequent multimodal fusion.

[0031] Through the above processing, deep semantic mining and effective compression of unstructured text information are achieved, ensuring that text modal features can retain key diagnostic information and adapt to the dimensional requirements of multimodal fusion.

[0032] S22. Numerical modal feature processing, the specific process is as follows:

[0033] S221, Business Attribute Derivation Calculation

[0034] Based on the raw numerical information, derived attributes with greater diagnostic value are extracted through time difference calculation, including equipment age and service time:

[0035] Equipment age: Characterizes the time span from equipment production to defect discovery, calculated using the following formula:

[0036] (4);

[0037] In the formula, Age represents the age of the device. This is a timestamp indicating when the defect was discovered. This is the manufacturing date and timestamp.

[0038] Service life: Characterizes the time span from when equipment is put into operation to when a defect is discovered. The calculation formula is:

[0039] (5);

[0040] In the formula, ServiceTime represents the service time. A timestamp for the equipment commissioning date;

[0041] S222, Numerical Characteristic Linear Transformation

[0042] The derived equipment age and service time features are linearly transformed to map them into feature vectors suitable for model processing. The calculation formula is as follows:

[0043] Equipment age characteristics: ;

[0044] Service life characteristics: ;

[0045] In the formula, , These are the linear transformation weights corresponding to the equipment age and service time characteristics, respectively. , These are the corresponding bias terms, all of which are learnable parameters; , These are characteristics representing equipment age and service time, respectively.

[0046] Through the above processing, the original numerical information is transformed into quantitative characteristics that can directly reflect the aging degree and operating history of the equipment;

[0047] S23. Categorical modality feature processing, the specific process is as follows:

[0048] S231, Category Attribute Embedding Transformation

[0049] For each discrete category attribute, an embedding layer maps it to a low-dimensional continuous space, generating a corresponding feature vector. Let there exist... The category feature, the first Each category attribute is Embedding dimension is Its characteristics are expressed as:

[0050] (6);

[0051] In the formula, The embedding layer function transforms discrete category values ​​into a dimensionless array using a learnable embedding matrix. A continuous vector that can capture the potential semantic associations between different categories;

[0052] S232, Category Feature Aggregation

[0053] The embedded features of all category attributes are aggregated to form a unified category modality feature representation. The calculation formula is as follows:

[0054] (7);

[0055] In the formula, For feature splicing operations, The embedded features of each category attribute are integrated into a comprehensive category modality feature vector. This provides a unified dimensional category feature foundation for subsequent multimodal fusion;

[0056] Through the above processing, discrete and isolated category information is transformed into continuous and related feature representations;

[0057] S24. Gating fusion mechanism, the specific process is as follows:

[0058] S241, Construction of Information Type Features

[0059] By comprehensively utilizing the feature processing methods in S11, S12, and S13, and based on the independent processing results of text, numerical, and categorical modalities, three types of information features are constructed accordingly:

[0060] Descriptive information features: Converting plain text descriptions of information attributes into feature representations. ;

[0061] Equipment information characteristics: Converting information attributes related to the equipment itself into feature representations. ;

[0062] Management information characteristics: Converted into characteristic representation ;

[0063] S242, Unify all feature dimensions

[0064] The three types of information feature dimensions are projected onto a unified dimension through linear transformation and nonlinear activation. To ensure dimensional matching during fusion, the calculation formula is as follows:

[0065] (8);

[0066] (9);

[0067] (10);

[0068] In the formula, , , The projection matrix is ​​learnable. , , For the corresponding bias term, It is a non-linear activation function. , , These are the projected feature vectors;

[0069] S243, Gating Weight Calculation

[0070] Dynamic weights are generated through a gating unit to adjust the fusion contribution of the three types of information, as detailed below:

[0071] First, the projected features are concatenated, and then a gate value between 0 and 1 is output using the Sigmoid function. The calculation formula is as follows:

[0072] (11);

[0073] In the formula, For the Sigmoid function, and For learnable parameters, Split into , and These correspond to the gating weights of description information, equipment information, and management information, respectively.

[0074] S244, Dynamic Weighted Fusion

[0075] The gating weights are multiplied element-wise with the projected features, and then summed to obtain the preliminary fused features. The calculation formula is as follows:

[0076] (12);

[0077] In the formula, Represents element-wise multiplication. For the initial feature representation of fusion, it is possible to dynamically enhance key information and suppress noise based on the characteristics of the input data;

[0078] S25. Cross-type interaction enhancement and defect prediction, the specific process is as follows:

[0079] S251, Construction of Cross-Attention Mechanism

[0080] Using descriptive information features as the query, and the fusion features of device information and management information as the key and value, cross-attention is constructed to capture cross-type interactions:

[0081] Query: Generated based on descriptive information projection features, the formula is:

[0082] (13);

[0083] Key and Value: Generated based on the concatenation features of device information and management information, using the following formula:

[0084] (14);

[0085] (15);

[0086] In the formula, , , This is a learnable projection matrix used to map features to the attention head dimension; , , These are the features projected during the gated fusion stage;

[0087] S252, Scaling Dot Product Attention Calculation

[0088] Interaction weights are calculated using a scaled dot product attention mechanism to quantify the strength of associations between different information types. The calculation formula is as follows:

[0089] (16);

[0090] In the formula, For attention head dimension; The function ensures that the sum of weights is 1, highlighting key interactions.

[0091] S253, Residual Connectivity and Final Feature Generation

[0092] The initial features obtained by fusing cross-attention output with gating are combined through residual connections. This process preserves the original information while introducing cross-type enhancement information, resulting in the final features. The calculation formula is as follows:

[0093] (17);

[0094] In the formula, Preliminary features generated for gating fusion This is the final feature representation that includes cross-type interaction information;

[0095] S254, Defect Category Prediction

[0096] The final features are input into the fully connected layer, and then... The activation function outputs the probability distribution of each defect category, thus enabling defect prediction.

[0097] (18);

[0098] In the formula, and Here are the weight matrix and bias terms for the output layer; the model training objective is to minimize the prediction distribution. With real labels Cross-entropy loss between;

[0099] Through the above process, the model not only integrates multi-source heterogeneous information, but also captures deep cross-type associations through cross-attention.

[0100] As a preferred technical solution of the present invention, step S3 is as follows:

[0101] S31. The method for assessing the importance of information attributes, the specific process is as follows:

[0102] S311, Calculation of Information Attribute Threshold Value

[0103] For each information attribute group, its intrinsic importance is quantified by learning independent gating weights, as follows:

[0104] Given the first Representation of each information attribute The gate control value is calculated through the gate control unit. The calculation formula is:

[0105] (19);

[0106] In the formula, For the Sigmoid function, and In this mechanism, learnable parameters are used to generate the gating weights. The value range is [0,1], and the larger the value, the higher the intrinsic importance of the information attribute group;

[0107] S312, Calculation of Information Attribute Gradient Sensitivity

[0108] Using the gradient backpropagation method, we measure the direct impact of changes in information attributes on the model's loss function, given information attributes. For information attributes Its gradient sensitivity The calculation formula is:

[0109] (20);

[0110] In the formula, The model loss function, The larger the value, the more significant the impact of the information attribute on the prediction result;

[0111] S313, Importance Assessment of Information Attribute Ranking

[0112] A permutation importance method independent of model structure is used to evaluate the global contribution of information attributes on the validation set, given information attributes. For information attributes By randomly shuffling the input of this attribute, we can obtain... Then, calculate the degree of performance degradation of the model caused by disrupting this attribute, using the following formula:

[0113] (twenty one)

[0114] In the formula, The expected performance of the model under the original input. The expected performance of the model after shuffling the attributes. A larger value indicates a higher level of information attribute. The more significant the contribution to the overall performance of the model;

[0115] The importance of information attributes is comprehensively evaluated from three aspects: intrinsic attention, loss sensitivity, and global performance impact, using the above three dimensions of indicators.

[0116] S32. Information attribute correlation and redundancy analysis, the specific process is as follows:

[0117] S321, Mutual Information Calculation

[0118] Mutual information (MI) is used to measure the non-linear dependencies between information attributes and target variables, as detailed below:

[0119] First, there is the mutual information between the information attributes and the target variable, which measures the discriminative value of the information attributes for defect categories. The calculation formula is as follows:

[0120] (twenty two);

[0121] In the formula, For the first Each information attribute is a random variable. For the defect category random variable, for and The joint probability distribution, , They are respectively and The marginal probability distribution;

[0122] Secondly, there is mutual information between information attributes, which measures the degree of cross-redundancy between different information attributes. The calculation formula is as follows:

[0123] (twenty three);

[0124] In the formula, For the first Each information attribute is a random variable. for and The joint probability distribution, , These are the marginal probability distributions of the two, respectively;

[0125] S322, Redundancy Measurement

[0126] Define information attributes Redundancy The formula for calculating this value is: This comprehensively measures the degree to which it is repeatedly expressed by other attributes.

[0127] (twenty four);

[0128] In the formula, the numerator is the first... The sum of the mutual information of an information attribute with all other information attributes represents the amount of information about that attribute that is covered by other attributes; the denominator is the mutual information between that attribute and the defect category, representing its effective diagnostic value. The larger the value, the higher the redundancy of the information attribute, and the more likely it is to be simplified.

[0129] S323. Obtain the importance score of the comprehensive information attributes.

[0130] Integrating multidimensional assessment results , , With feature redundancy The comprehensive score for each information attribute is obtained, calculated using the following formula:

[0131] (25);

[0132] In the formula, , , , These are the weighting coefficients;

[0133] The above analysis can identify redundant information attributes that contribute little to the target but are highly correlated with other attributes.

[0134] As a preferred technical solution of the present invention, step S4 is specifically as follows:

[0135] S41. Multi-threshold information attribute selection strategy, the specific process is as follows:

[0136] S411, Design of Information Attribute Selection Indicator Function

[0137] First, based on information attributes Overall rating Design information attribute selection indicator function This is used to determine whether to include the attribute in the candidate subset, and the calculation formula is:

[0138] (26);

[0139] In the formula: To select the threshold, when the overall score of the information attribute is not lower than the threshold, the indicator function is set to 1, indicating selection; otherwise, it is 0, indicating rejection.

[0140] S412, Multi-threshold set preset

[0141] To explore the optimal subset of information attributes, a pre-defined ordered set of multiple thresholds is used:

[0142] (27);

[0143] In the formula, The thresholds are arranged in ascending order to cover different levels of stringency in the screening criteria, so as to comprehensively examine the performance of different subsets.

[0144] S413. Information Attribute Subset Filtering and Performance Evaluation

[0145] First, for each threshold in the multi-threshold set Filter out the corresponding information attribute subset based on the indicator function:

[0146] (28);

[0147] Subsequently, the model performance metrics corresponding to this subset are evaluated on the validation set, and the threshold corresponding to the optimal performance is finally selected. Determine the initial subset of information attributes ,

[0148] By exploring multiple thresholds, we can capture the performance differences of different combinations of information attributes and ensure that the selected subset has the best discriminative power.

[0149] S42. Joint optimization of information attribute selection and model training, the specific process is as follows:

[0150] S421, Gating value initialization

[0151] To ensure that all information attributes are treated equally in the initial state, the gating values ​​of all information attributes are initialized. This serves as the starting point for iterative optimization.

[0152] S422, Iterative Optimization Process

[0153] In each training cycle The following two steps are executed alternately to achieve dynamic coordination between information attribute selection and model training:

[0154] S4221, Fixed Information Attribute Gating Value Update model parameters :

[0155] Based on the current gate value, the information attributes are weighted, and the loss function is minimized using gradient descent to update the model parameters. The calculation formula is as follows:

[0156] (29);

[0157] In the formula, For learning rate, For loss function For model parameters The gradient ensures that the model optimizes in the optimal direction under the current information attribute weights;

[0158] S4222, Fixed Model Parameters Update the importance score and gating value of information attributes:

[0159] Using the information attribute importance analysis mechanism, the comprehensive score of each information attribute is re-evaluated based on the updated model parameters. The score is then mapped to a new gate value via a gating unit, calculated using the following formula:

[0160] (30);

[0161] (31);

[0162] In the formula, This is a scoring function in the information attribute importance analysis mechanism. and For learnable parameters, The Sigmoid function ensures that the gate value range is [0,1], dynamically reflecting the real-time importance of information attributes;

[0163] S423, Implementation of soft selection mechanism

[0164] By dynamically weighting the contribution of each information attribute using a gating value, soft selection of information attributes is achieved. The calculation formula is as follows:

[0165] (32);

[0166] In the formula, For weighted information attributes, a higher gate value means a greater impact of that attribute on the model, enabling the model to automatically focus on key attributes and suppress redundant attributes.

[0167] Through the above iterative optimization, when the model training is completed, the final comprehensive score of information attributes and the gating value can be output, realizing deep collaboration between information attribute selection and model training.

[0168] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0169] 1. This invention proposes a multimodal feature fusion network (GHAN) based on gated hierarchical attention. It dynamically fuses text, numerical, and category multimodal information through a gating mechanism and captures deep cross-type correlations with the help of cross-attention. This achieves unified representation and deep collaboration of multi-source heterogeneous information in substation equipment inspection logs, overcoming the problem that traditional single-modal or shallow fusion methods are insufficient in characterizing complex defect features.

[0170] 2. This invention designs an Information Attribute Importance Analysis (IAIAM) mechanism, which combines multiple dimensions such as gating value, gradient sensitivity, permutation importance and mutual information to quantify the importance and redundancy of each information attribute for defect identification. This solves the limitations of existing technologies in the single-dimensional evaluation of information attribute value and insufficient redundancy processing, and provides a scientific basis for the screening of key information.

[0171] 3. This invention constructs an adaptive information attribute selection framework (AIASF). Through multi-threshold strategy and joint optimization of information attribute selection and model training, it realizes dynamic screening of information attributes with high discriminative power and low redundancy. While improving the accuracy of defect identification, it reduces the computational complexity of the model and enhances the generalization ability and engineering practicality of the model. Attached Figure Description

[0172] Figure 1 This is an architectural diagram of the method of the present invention;

[0173] Figure 2 This is a flowchart of the method of the present invention;

[0174] Figure 3 This is a chart showing some of the equipment inspection log data;

[0175] Figure 4 Continued display of some equipment inspection log data;

[0176] Figure 5 This is a comparison chart of the recognition effects of the present invention and existing methods;

[0177] Figure 6 This is a table diagram of an ablation test for information attributes. Detailed Implementation

[0178] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0179] like Figure 1-2 As shown, the device defect identification method based on multimodal information fusion and adaptive information optimization proposed in this invention includes the following steps:

[0180] S1: Retrieves multimodal data from substation equipment to provide data support;

[0181] Specifically, multimodal data includes information modalities, equipment inspection logs, and information types.

[0182] Information modalities include text modalities, numerical modalities, and categorical modalities.

[0183] Equipment inspection logs include defect descriptions, defect locations, remarks, voltage levels, manufacturers, equipment models, and the source of the defect.

[0184] Information types include descriptive information, equipment information, and management information.

[0185] S2: Construct a gated hierarchical attention-based multimodal feature fusion network based on multimodal data to achieve deep fusion of multimodal information in substation equipment inspection logs;

[0186] Specifically, constructing a gated hierarchical attention multimodal feature fusion network (GHAN) involves the following steps:

[0187] S21. Text Modal Feature Processing

[0188] This section aims to extract deep semantic features from the textual modal information in substation equipment inspection logs using a domain-adaptive model and attention mechanism. It then converts variable-length text sequences into fixed-dimensional feature vectors, providing accurate textual feature support for multimodal information fusion. The specific process is as follows:

[0189] S211, Text Sequence Encoding and Context Embedding Generation

[0190] Given a text sequence (in For words in the text, (where is the number of lexical units), a 12-layer Transformer encoder of the domain-adaptive BERT model is used for semantic feature extraction to generate context-aware embedding representations. The calculation formula is as follows:

[0191] (1)

[0192] In the formula, The context embedding matrix output by BERT, with dimension . ( (As an embedding dimension), this matrix can capture the contextual dependencies and domain-specific semantics of lexical units in the text.

[0193] S212, Attention Pooling and Fixed-Dimension Feature Aggregation

[0194] To aggregate variable-length sequence representations into fixed-dimensional vectors, attention pooling is applied to the word embeddings output from the last four layers of the BERT model. First, the attention weight for each word is calculated using the following formula:

[0195] (2)

[0196] In the formula, For learnable parameters, For the first Embedding vectors of 1 word element This represents the attention weight of the lexical unit, reflecting its importance in the semantic expression of the text.

[0197] Subsequently, the fixed-dimensional text modal features are obtained by weighted summation, as shown in the formula:

[0198] (3)

[0199] In the formula, This is the final feature representation of the text modality, which can focus on the key semantic information in the text and provide a unified dimensional text feature foundation for subsequent multimodal fusion.

[0200] Through the above processing, deep semantic mining and effective compression of unstructured text information are achieved, ensuring that text modal features can retain key diagnostic information and adapt to the dimensional requirements of multimodal fusion.

[0201] S22, Numerical Modal Feature Processing

[0202] This section aims to extract quantitative features from the numerical information in substation equipment inspection logs through business derivation and linear transformation. These features directly reflect the equipment's operating status and historical information, providing numerically meaningful features to support multimodal information fusion. The specific process is as follows:

[0203] S221, Business Attribute Derivation Calculation

[0204] Based on the raw numerical information, derived attributes with greater diagnostic value are extracted through time difference calculation, including equipment age and service time:

[0205] Equipment Age: Characterizes the time span from equipment production to defect discovery, calculated using the following formula:

[0206] (4)

[0207] In the formula, This is a timestamp indicating when the defect was discovered. This is the manufacturing date and time stamp.

[0208] Service Time: Characterizes the time span from when equipment is put into operation to when a defect is discovered. The calculation formula is:

[0209] (5)

[0210] In the formula, This is a timestamp indicating the date the equipment was put into operation.

[0211] S222, Numerical Characteristic Linear Transformation

[0212] The derived equipment age and service time features are linearly transformed to map them into feature vectors suitable for model processing. The calculation formula is as follows:

[0213] Equipment age characteristics:

[0214] Service life characteristics:

[0215] In the formula, , These are the linear transformation weights corresponding to the equipment age and service time characteristics, respectively. , These are the corresponding bias terms, all of which are learnable parameters; , These are characteristics representing the equipment's age and service time, respectively.

[0216] Through the above processing, the original numerical information is transformed into quantitative features that can directly reflect the aging degree and operating history of the equipment, providing numerical dimension support with clear business meaning for multimodal fusion.

[0217] S23, Category Modal Feature Processing

[0218] This section aims to transform the categorical information (such as voltage level, manufacturer, equipment model, etc.) in substation equipment inspection logs into continuous, dense feature vectors through an embedding layer. This captures the potential correlations between categories, providing low-dimensional and semantically rich categorical feature support for multimodal information fusion. The specific process is as follows:

[0219] S231, Category Attribute Embedding Transformation

[0220] For each discrete category attribute, an embedding layer maps it to a low-dimensional continuous space, generating the corresponding feature vector. Suppose there exists... The category feature, the first Each category attribute is Embedding dimension is Its characteristics are expressed as:

[0221] (6)

[0222] In the formula, The embedding layer function transforms discrete category values ​​into a dimensionless array using a learnable embedding matrix. A continuous vector that can capture the potential semantic associations between different categories.

[0223] S232, Category Feature Aggregation

[0224] The embedded features of all category attributes are aggregated to form a unified category modality feature representation. The calculation formula is as follows:

[0225] (7)

[0226] In the formula, For feature splicing operations, The embedded features of each category attribute are integrated into a comprehensive category modality feature vector. This provides a unified dimensional category feature foundation for subsequent multimodal fusion.

[0227] Through the above processing, discrete and isolated category information is transformed into continuous and related feature representations, effectively uncovering the potential relationships between category attributes and enhancing the auxiliary value of category information in equipment defect identification.

[0228] S24, Gating Fusion Mechanism

[0229] This section aims to achieve adaptive fusion of three heterogeneous feature types—descriptive information, equipment information, and management information—through dynamic gating and weight adjustment. This addresses the challenge of simple feature concatenation or fixed weighting methods failing to adapt to the dynamic differences in contribution between different information types, effectively capturing the complex dependencies between multimodal data and generating a preliminary fused feature representation. The specific process is as follows:

[0230] S241, Construction of Information Type Features

[0231] By comprehensively utilizing the feature processing methods mentioned in S1.1, S1.2, and S1.3, and based on the independent processing results of text, numerical, and categorical modalities, three types of information features are constructed:

[0232] Descriptive information features: Convert plain text descriptions such as defect descriptions and defect locations into feature representations. ;

[0233] Equipment information features: Converting information attributes related to the equipment itself, such as voltage level and manufacturing date, into feature representations. ;

[0234] Management information characteristics: Converted into characteristic representation .

[0235] S242, Unify all feature dimensions

[0236] Since the three types of information features may have different dimensions, they need to be projected to a unified dimension through linear transformation and nonlinear activation. To ensure dimensional matching during fusion, the calculation formula is as follows:

[0237] (8)

[0238] (9)

[0239] (10)

[0240] In the formula, , , The projection matrix is ​​learnable. , , For the corresponding bias term, It is a non-linear activation function. , , is the projected feature vector.

[0241] S243, Gating Weight Calculation

[0242] Dynamic weights are generated through a gating unit to adjust the fusion contribution of the three types of information. First, the projected features are concatenated, and then a gating value between 0 and 1 is output using the Sigmoid function. The calculation formula is as follows:

[0243] (11)

[0244] In the formula, For the Sigmoid function, and For learnable parameters, Split into , and These correspond to the gating weights for description information, equipment information, and management information, respectively.

[0245] S244, Dynamic Weighted Fusion

[0246] The gating weights are multiplied element-wise with the projected features, and then summed to obtain the preliminary fused features. The calculation formula is as follows:

[0247] (12)

[0248] In the formula, Represents element-wise multiplication. The initial feature representation can dynamically enhance key information and suppress noise based on the characteristics of the input data, laying the foundation for subsequent cross-type interaction enhancement.

[0249] S25, Cross-type Interaction Enhancement and Defect Prediction

[0250] This section aims to capture fine-grained semantic relationships and complementary cues between different information types through a cross-attention mechanism, particularly the deep contextual dependencies between descriptive information and equipment / management information. It also combines residual connections to enhance feature discriminative power, ultimately achieving accurate prediction of equipment defect categories. The specific process is as follows:

[0251] S251, Construction of Cross-Attention Mechanism

[0252] Using descriptive information features as queries and the fusion features of device and management information as keys and values, cross-attention is constructed to capture cross-type interactions:

[0253] Query (Q): Generated based on projected features of descriptive information, with the following formula:

[0254] (13)

[0255] Key (K) and Value (V): Generated based on the concatenation features of device information and management information, using the following formula:

[0256] (14)

[0257] (15)

[0258] In the formula, , , This is a learnable projection matrix used to map features to the attention head dimension; , , These are the features projected during the gating fusion stage.

[0259] S252, Scaling Dot Product Attention Calculation

[0260] Interaction weights are calculated using a scaled dot product attention mechanism to quantify the strength of associations between different information types. The calculation formula is as follows:

[0261] (16)

[0262] In the formula, To improve the attention head dimension, scaling operations avoid gradient instability caused by excessively high dimensions; The function ensures that the total weight is 1, highlighting key interactions.

[0263] S253, Residual Connectivity and Final Feature Generation

[0264] The initial features obtained by fusing cross-attention output with gating are combined through residual connections. This process preserves the original information while introducing cross-type enhancement information, resulting in the final features. The calculation formula is as follows:

[0265] (17)

[0266] In the formula, Preliminary features generated for gating fusion It is the final feature representation that contains cross-type interaction information and has stronger discriminative power.

[0267] S254, Defect Category Prediction

[0268] The final features are input into the fully connected layer, and then... The activation function outputs the probability distribution of each defect category, thus enabling defect prediction.

[0269] (18)

[0270] In the formula, and Here are the weight matrix and bias terms for the output layer; the model training objective is to minimize the prediction distribution. With real labels The cross-entropy loss between the two sides ensures the accuracy of the prediction results.

[0271] Through the above process, the model not only integrates multi-source heterogeneous information, but also captures deep cross-type correlations through cross-attention, significantly improving the ability to identify complex defect patterns.

[0272] S3: Design an information attribute importance analysis mechanism to quantify the importance and redundancy of each information attribute for defect identification;

[0273] Specifically, the constructed Information Attribute Importance Analysis (IAIAM) mechanism includes:

[0274] S31. Method for assessing the importance of information attributes

[0275] This section aims to comprehensively and accurately assess the importance of various information attributes for equipment defect identification through multi-dimensional indicators, avoiding the limitations of single indicators and providing a quantitative basis for subsequent information attribute screening. The specific process is as follows:

[0276] S311, Calculation of Information Attribute Threshold Value

[0277] For each information attribute group, its intrinsic importance is quantified by learning independent gating weights. Given the first... Representation of each information attribute The gate control value is calculated through the gate control unit. The calculation formula is:

[0278] (19)

[0279] In the formula, For the Sigmoid function, and The learnable parameters used in this mechanism to generate the gating weights The value range is [0,1]. The larger the value, the higher the intrinsic importance of the information attribute group.

[0280] S312, Calculation of Information Attribute Gradient Sensitivity

[0281] We use gradient backpropagation to measure the direct impact of small changes in information attributes on the model's loss function. Given information attributes... For information attributes Its gradient sensitivity The calculation formula is:

[0282] (20)

[0283] In the formula, The model loss function, The larger the value, the more significant the impact of the information attribute on the prediction result.

[0284] S313, Importance Assessment of Information Attribute Ranking

[0285] An order importance method, independent of model structure, is used to evaluate the global contribution of information attributes on the validation set. For information attributes... By randomly shuffling the input of this attribute, we can obtain... Then, calculate the degree of performance degradation of the model caused by disrupting this attribute, using the following formula:

[0286] (twenty one)

[0287] In the formula, The expected performance of the model under the original input. The expected performance of the model after shuffling the attributes. A larger value indicates that the information attribute contributes more significantly to the overall performance of the model.

[0288] By using the above three dimensions of indicators, the importance of information attributes is comprehensively evaluated from three levels: intrinsic attention, loss sensitivity, and global performance impact, laying the foundation for subsequent redundancy analysis and attribute selection.

[0289] S32. Information Attribute Correlation and Redundancy Analysis

[0290] This section aims to identify redundant information attributes by quantifying the dependencies between information attributes and between information attributes and the target variable. This avoids increased model complexity and decreased generalization ability caused by information overlap or duplication, providing a basis for simplifying the information attribute set. The specific process is as follows:

[0291] S321, Mutual Information Calculation

[0292] Mutual information (MI) is used to measure the non-linear dependencies between information attributes and the target variable, as well as between information attributes themselves.

[0293] First, there is the mutual information between the information attributes and the target variable, which measures the discriminative value of the information attributes for defect categories. The calculation formula is as follows:

[0294] (twenty two)

[0295] In the formula, For the first Each information attribute is a random variable. For the defect category random variable, for and The joint probability distribution, , They are respectively and The marginal probability distribution.

[0296] Secondly, there is mutual information between information attributes, which measures the degree of cross-redundancy between different information attributes. The calculation formula is as follows:

[0297] (twenty three)

[0298] In the formula, For the first Each information attribute is a random variable. for and The joint probability distribution, , These are the marginal probability distributions of the two, respectively.

[0299] S322, Redundancy Measurement

[0300] Define information attributes Redundancy The formula for calculating this value is: This comprehensively measures the degree to which it is repeatedly expressed by other attributes.

[0301] (twenty four)

[0302] In the formula, the numerator is the first... The sum of the mutual information of an information attribute with all other information attributes represents the amount of information about that attribute that is covered by other attributes; the denominator is the mutual information between that attribute and the defect category, representing its effective diagnostic value. The larger the value, the higher the redundancy of the information attribute, and the more likely it is to be simplified.

[0303] S323. Obtain the importance score of the comprehensive information attributes.

[0304] Finally, the multidimensional evaluation results were integrated. , , With feature redundancy The comprehensive score for each information attribute is obtained, calculated using the following formula:

[0305] (25)

[0306] In the formula, , , , The weighting coefficients are used. Through the above analysis, redundant information attributes that contribute little to the target but are highly correlated with other attributes can be identified, providing a quantitative basis for subsequent information attribute selection and achieving efficient simplification of the information attribute set.

[0307] S4: Construct an adaptive information attribute selection framework to achieve accurate defect identification through joint optimization of information attribute filtering and model training;

[0308] The constructed Adaptive Information Attribute Selection Framework (AIASF) includes:

[0309] S41. Multi-threshold information attribute selection strategy

[0310] This section aims to systematically select the subset of information attributes with the strongest discriminative power for defect identification based on a comprehensive scoring of information attributes, through dynamic threshold control and multi-threshold exploration. This avoids the limitations of a single threshold and ensures the adaptability and optimality of information attribute selection. The specific process is as follows:

[0311] S411, Design of Information Attribute Selection Indicator Function

[0312] First, based on information attributes Overall rating Design information attribute selection indicator function This is used to determine whether to include the attribute in the candidate subset, and the calculation formula is:

[0313] (26)

[0314] In the formula: To select the threshold, the indicator function is set to 1 when the overall score of the information attribute is not lower than the threshold, indicating selection; otherwise, it is 0, indicating rejection.

[0315] S412, Multi-threshold set preset

[0316] To enable the system to explore the optimal subset of information attributes, a pre-defined ordered set of multiple thresholds is established:

[0317] (27)

[0318] In the formula, The thresholds are arranged in ascending order to cover different levels of stringency in the screening criteria, so as to comprehensively examine the performance of different subsets.

[0319] S413. Information Attribute Subset Filtering and Performance Evaluation

[0320] For each threshold in the multi-threshold set Filter out the corresponding information attribute subset based on the indicator function:

[0321] (28)

[0322] Subsequently, the model performance metrics corresponding to this subset are evaluated on the validation set, and the threshold corresponding to the optimal performance is finally selected. Determine the initial subset of information attributes This strategy, through multi-threshold exploration, can effectively capture the performance differences of different combinations of information attributes, ensuring that the selected subset has optimal discriminative power.

[0323] S42. Joint Optimization of Information Attribute Selection and Model Training

[0324] This section aims to address the problem of traditional independent information attribute selection and model training processes, resulting in a disconnect between the two objectives. It utilizes an iterative optimization algorithm to achieve synergistic optimization, ensuring a high degree of alignment between the information attribute selection results and the model optimization goal. This ultimately leads to a more compact and generalized defect identification model. The specific process is as follows:

[0325] S421, Gating value initialization

[0326] To ensure that all information attributes are treated equally in the initial state, the gating values ​​of all information attributes are initialized. This serves as the starting point for iterative optimization.

[0327] S422, Iterative Optimization Process

[0328] In each training cycle The following two steps are executed alternately to achieve dynamic coordination between information attribute selection and model training:

[0329] S4221, Fixed Information Attribute Gating Value Update model parameters :

[0330] Based on the current gate value, the information attributes are weighted, and the loss function is minimized using gradient descent to update the model parameters. The calculation formula is as follows:

[0331] (29)

[0332] In the formula, For learning rate, For loss function For model parameters The gradient ensures that the model optimizes in the optimal direction under the current information attribute weights.

[0333] S4222, Fixed Model Parameters Update the importance score and gating value of information attributes:

[0334] Using the Information Attribute Importance Analysis (IAIAM) mechanism, the comprehensive score of each information attribute is re-evaluated based on the updated model parameters. The score is then mapped to a new gate value via a gating unit, calculated using the following formula:

[0335] (30)

[0336] (31)

[0337] In the formula, For the scoring function in IAIAM, and For learnable parameters, The Sigmoid function ensures that the gate value range is [0,1], dynamically reflecting the real-time importance of information attributes.

[0338] S423, Implementation of soft selection mechanism

[0339] By dynamically weighting the contribution of each information attribute using a gating value, soft selection of information attributes is achieved. The calculation formula is as follows:

[0340] (32)

[0341] In the formula, For weighted information attributes, the higher the gate value, the greater the impact of the attribute on the model, enabling the model to automatically focus on key attributes and suppress redundant attributes.

[0342] Through the above iterative optimization, when the model training is completed, the final comprehensive score of information attributes and the gating value can be output, realizing deep collaboration between information attribute selection and model training, and significantly reducing model complexity while ensuring recognition accuracy.

[0343] S5: Establish a substation equipment defect identification system based on multimodal information fusion and adaptive attribute optimization.

[0344] To verify the effectiveness of the proposed substation equipment defect identification method based on multimodal information fusion and adaptive information optimization, this invention selects a dataset consisting of equipment inspection logs from a substation in Central China and utilizes the information in these logs for equipment defect identification. Some example data (after anonymization) is shown below. Figure 3 , Figure 4 As shown, Figure 5 and Figure 6The paper presents a comparison of the proposed method with traditional methods in terms of recognition performance, and demonstrates the improvement in computational efficiency that the proposed method can bring.

[0345] Depend on Figure 3 As shown, the recognition performance of this method is significantly better than that of traditional methods.

[0346] Depend on Figure 4 As shown, on the optimal subset of information attributes selected by the AIASF mentioned in this method, the model almost simultaneously maintains the accuracy of the full information attribute model and the computational efficiency of the single information attribute model, fully demonstrating the improvement in computational efficiency brought about by this method.

[0347] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for identifying equipment defects using multimodal information fusion and adaptive information optimization, characterized in that, Includes the following steps: S1: Retrieves multimodal data from substation equipment to provide data support; S2: Construct a gated hierarchical attention-based multimodal feature fusion network based on multimodal data to achieve deep fusion of multimodal information in substation equipment inspection logs; S3: Design an information attribute importance analysis mechanism to quantify the importance and redundancy of each information attribute for defect identification; S4: Construct an adaptive information attribute selection framework to achieve accurate defect identification through joint optimization of information attribute filtering and model training; S5: Establish a substation equipment defect identification system based on multimodal information fusion and adaptive attribute optimization.

2. The equipment defect identification method based on multimodal information fusion and adaptive information optimization according to claim 1, characterized in that, In step S1, the multimodal data includes information modalities, equipment inspection logs, and information types. Information modalities include text modalities, numerical modalities, and categorical modalities. Equipment inspection logs include defect descriptions, defect locations, remarks, voltage levels, manufacturers, equipment models, and the source of the defect. Information types include descriptive information, equipment information, and management information.

3. The equipment defect identification method based on multimodal information fusion and adaptive information optimization according to claim 1, characterized in that, Step S2 is as follows: S21. Text modality feature processing, the specific process is as follows: S211, Text Sequence Encoding and Context Embedding Generation Given a text sequence ,in For words in the text, To determine the number of lexical units, a 12-layer Transformer encoder of a domain-adaptive BERT model is used for semantic feature extraction to generate context-aware embeddings. The calculation formula is as follows: （1）； In the formula, The context embedding matrix output by BERT, with dimension . , As an embedding dimension, this context embedding matrix can capture the contextual dependencies and domain-specific semantics of lexical units in the text; S212, Attention Pooling and Fixed-Dimension Feature Aggregation To aggregate variable-length sequence representations into fixed-dimensional vectors, attention pooling is applied to the word embeddings output from the last four layers of the BERT model, as follows: First, calculate the attention weight for each word, using the following formula: （2）； In the formula, For learnable parameters, For the first Embedding vectors of 1 word element This represents the attention weight of the word, reflecting its importance in the semantic expression of the text; Subsequently, the fixed-dimensional text modal features are obtained by weighted summation, as shown in the formula: （3）； In the formula, This is the final feature representation of the text modality, which can focus on the key semantic information in the text, providing a unified dimensional text feature foundation for subsequent multimodal fusion. pass The above processing enables deep semantic mining and effective compression of unstructured text information, ensuring that text modal features can retain key diagnostic information while adapting to the dimensional requirements of multimodal fusion. S22. Numerical modal feature processing, the specific process is as follows: S221, Business Attribute Derivation Calculation Based on the raw numerical information, derived attributes with greater diagnostic value are extracted through time difference calculation, including equipment age and service time: Equipment age: Characterizes the time span from equipment production to defect discovery, calculated using the following formula: （4）； In the formula, Age represents the age of the device. This is a timestamp indicating when the defect was discovered. This is the manufacturing date and timestamp. Service life: Characterizes the time span from when equipment is put into operation to when a defect is discovered. The calculation formula is: （5）； In the formula, ServiceTime represents the service time. A timestamp for the date the equipment was put into operation; S222, Numerical Characteristic Linear Transformation The derived equipment age and service time features are linearly transformed to map them into feature vectors suitable for model processing. The calculation formula is as follows: Equipment age characteristics: ; Service life characteristics: ; In the formula, , These are the linear transformation weights corresponding to the equipment age and service time characteristics, respectively. , These are the corresponding bias terms, all of which are learnable parameters; , These are characteristics representing equipment age and service time, respectively. Through the above processing, the original numerical information is transformed into quantitative characteristics that can directly reflect the aging degree and operating history of the equipment; S23. Categorical modality feature processing, the specific process is as follows: S231, Category Attribute Embedding Transformation For each discrete category attribute, an embedding layer maps it to a low-dimensional continuous space, generating a corresponding feature vector. Let there exist... The category feature, the first Each category attribute is Embedding dimension is Its characteristics are expressed as: （6）； In the formula, The embedding layer function transforms discrete category values ​​into a dimensionless array using a learnable embedding matrix. A continuous vector that can capture the potential semantic associations between different categories; S232, Category Feature Aggregation The embedded features of all category attributes are aggregated to form a unified category modality feature representation. The calculation formula is as follows: （7）； In the formula, For feature splicing operations, The embedded features of each category attribute are integrated into a comprehensive category modality feature vector. This provides a unified dimensional category feature foundation for subsequent multimodal fusion; Through the above processing, discrete and isolated category information is transformed into continuous and related feature representations; S24. Gating fusion mechanism, the specific process is as follows: S241, Construction of Information Type Features By comprehensively utilizing the feature processing methods in S11, S12, and S13, and based on the independent processing results of text, numerical, and categorical modalities, three types of information features are constructed accordingly: Descriptive information features: Converting plain text descriptions of information attributes into feature representations. ; Equipment information characteristics: Converting information attributes related to the equipment itself into feature representations. ; Management information characteristics: Converted into characteristic representation ; S242, Unify all feature dimensions The three types of information feature dimensions are projected onto a unified dimension through linear transformation and nonlinear activation. To ensure dimensional matching during fusion, the calculation formula is as follows: （8）； （9）； （10）； In the formula, , , For a learnable projection matrix, , , For the corresponding bias term, It is a non-linear activation function. , , These are the projected feature vectors; S243, Gating Weight Calculation Dynamic weights are generated through a gating unit to adjust the fusion contribution of the three types of information, as detailed below: First, the projected features are concatenated, and then a gate value between 0 and 1 is output using the Sigmoid function. The calculation formula is as follows: （11）； In the formula, For the Sigmoid function, and For learnable parameters, Split into , and These correspond to the gating weights of description information, equipment information, and management information, respectively. S244, Dynamic Weighted Fusion The gating weights are multiplied element-wise with the projected features, and then summed to obtain the preliminary fused features. The calculation formula is as follows: （12）； In the formula, Represents element-wise multiplication. For the initial feature representation of fusion, it is possible to dynamically enhance key information and suppress noise based on the characteristics of the input data; S25. Cross-type interaction enhancement and defect prediction, the specific process is as follows: S251, Construction of Cross-Attention Mechanism Using descriptive information features as the query, and the fusion features of device information and management information as the key and value, cross-attention is constructed to capture cross-type interactions: Query: Generated based on descriptive information projection features, the formula is: （13）； Key and Value: Generated based on the concatenation features of device information and management information, using the following formula: （14）； （15）； In the formula, , , This is a learnable projection matrix used to map features to the attention head dimension; , , Features projected during the gated fusion stage; S252, Scaling Dot Product Attention Calculation Interaction weights are calculated using a scaled dot product attention mechanism to quantify the strength of associations between different information types. The calculation formula is as follows: （16）； In the formula, For attention head dimension; The function ensures that the sum of weights is 1, highlighting key interactions. S253, Residual Connectivity and Final Feature Generation The initial features obtained by fusing cross-attention output with gating are combined through residual connections. This process preserves the original information while introducing cross-type enhancement information, resulting in the final features. The calculation formula is as follows: （17）； In the formula, Preliminary features generated for gating fusion This is the final feature representation that includes cross-type interaction information; S254, Defect Category Prediction The final features are input into the fully connected layer, and then... The activation function outputs the probability distribution of each defect category, thus enabling defect prediction. （18）； In the formula, and Here are the weight matrix and bias terms for the output layer; the model training objective is to minimize the prediction distribution. With real labels Cross-entropy loss between; Through the above process, the model not only integrates multi-source heterogeneous information, but also captures deep cross-type associations through cross-attention.

4. The equipment defect identification method based on multimodal information fusion and adaptive information optimization according to claim 1, characterized in that, Step S3 is as follows: S31. The method for assessing the importance of information attributes, the specific process is as follows: S311, Calculation of Information Attribute Threshold Value For each information attribute group, its intrinsic importance is quantified by learning independent gating weights, as follows: Given the first Representation of each information attribute The gate control value is calculated through the gate control unit. The calculation formula is: （19）； In the formula, For the Sigmoid function, and In this mechanism, learnable parameters are used to generate the gating weights. The value range is [0,1], and the larger the value, the higher the intrinsic importance of the information attribute group; S312, Calculation of Information Attribute Gradient Sensitivity Using the gradient backpropagation method, we measure the direct impact of changes in information attributes on the model's loss function, given information attributes. For information attributes Its gradient sensitivity The calculation formula is: （20）； In the formula, The model loss function, The larger the value, the more significant the impact of the information attribute on the prediction result; S313, Evaluation of the Importance of Information Attribute Ranking A permutation importance method independent of model structure is used to evaluate the global contribution of information attributes on the validation set, given information attributes. For information attributes By randomly shuffling the input of this attribute, we can obtain... Then, calculate the degree of performance degradation of the model caused by disrupting this attribute, using the following formula: （21） In the formula, The expected performance of the model under the original input. The expected performance of the model after shuffling the attributes. A larger value indicates a higher level of information attribute. The more significant the contribution to the overall performance of the model; The importance of information attributes is comprehensively evaluated from three aspects: intrinsic attention, loss sensitivity, and global performance impact, using the above three dimensions of indicators. S32. Information attribute correlation and redundancy analysis, the specific process is as follows: S321, Mutual Information Calculation Mutual information (MI) is used to measure the non-linear dependencies between information attributes and target variables, as detailed below: First, there is the mutual information between the information attributes and the target variable, which measures the discriminative value of the information attributes for defect categories. The calculation formula is as follows: （22）； In the formula, For the first Each information attribute is a random variable. For the defect category random variable, for and The joint probability distribution, , They are respectively and The marginal probability distribution; Secondly, there is mutual information between information attributes, which measures the degree of cross-redundancy between different information attributes. The calculation formula is as follows: （23）； In the formula, For the first Each information attribute is a random variable. for and The joint probability distribution, , These are the marginal probability distributions of the two, respectively; S322, Redundancy Measurement Define information attributes Redundancy The formula for calculating this value is: This comprehensively measures the degree to which it is repeatedly expressed by other attributes. （24）； In the formula, the numerator is the first... The sum of the mutual information of an information attribute with all other information attributes represents the amount of information about that attribute that is covered by other attributes; the denominator is the mutual information between that attribute and the defect category, representing its effective diagnostic value. The larger the value, the higher the redundancy of the information attribute, and the more likely it is to be simplified. S323. Obtain the importance score of the comprehensive information attributes. Integrating multidimensional assessment results , , With feature redundancy The comprehensive score for each information attribute is obtained, calculated using the following formula: （25）； In the formula, , , , These are the weighting coefficients; The above analysis can identify redundant information attributes that contribute little to the target but are highly correlated with other attributes.

5. The equipment defect identification method based on multimodal information fusion and adaptive information optimization according to claim 1 or 4, characterized in that, Step S4 is as follows: S41. Multi-threshold information attribute selection strategy, the specific process is as follows: S411, Design of Information Attribute Selection Indicator Function First, based on information attributes Overall rating Design information attribute selection indicator function This is used to determine whether to include the attribute in the candidate subset, and the calculation formula is: （26）； In the formula: To select the threshold, when the overall score of the information attribute is not lower than the threshold, the indicator function is set to 1, indicating selection; otherwise, it is 0, indicating rejection. S412, Multi-threshold set preset To explore the optimal subset of information attributes, a pre-defined ordered set of multiple thresholds is used: （27）； In the formula, The thresholds are arranged in ascending order to cover different levels of stringency in the screening criteria, so as to comprehensively examine the performance of different subsets. S413. Information Attribute Subset Filtering and Performance Evaluation First, for each threshold in the multi-threshold set Filter out the corresponding information attribute subset based on the indicator function: （28）； Subsequently, the model performance metrics corresponding to this subset are evaluated on the validation set, and the threshold corresponding to the optimal performance is finally selected. Determine the initial subset of information attributes , By exploring multiple thresholds, we can capture the performance differences of different combinations of information attributes and ensure that the selected subset has the best discriminative power. S42. Joint optimization of information attribute selection and model training, the specific process is as follows: S421, Gating value initialization To ensure that all information attributes are treated equally in the initial state, the gating values ​​of all information attributes are initialized. This serves as the starting point for iterative optimization. S422, Iterative Optimization Process In each training cycle The following two steps are executed alternately to achieve dynamic coordination between information attribute selection and model training: S4221, Fixed Information Attribute Gating Value Update model parameters : Based on the current gate value, the information attributes are weighted, and the loss function is minimized using gradient descent to update the model parameters. The calculation formula is as follows: （29）； In the formula, For learning rate, loss function For model parameters The gradient ensures that the model optimizes in the optimal direction under the current information attribute weights; S4222, Fixed Model Parameters Update the importance score and gating value of information attributes: Using the information attribute importance analysis mechanism, the comprehensive score of each information attribute is re-evaluated based on the updated model parameters. The score is then mapped to a new gate value via a gating unit, calculated using the following formula: （30）； （31）； In the formula, This is a scoring function in the information attribute importance analysis mechanism. and For learnable parameters, The Sigmoid function ensures that the gate value range is [0,1], dynamically reflecting the real-time importance of information attributes; S423, Implementation of soft selection mechanism By dynamically weighting the contribution of each information attribute using a gating value, soft selection of information attributes is achieved. The calculation formula is as follows: （32）； In the formula, For weighted information attributes, a higher gate value means a greater impact of that attribute on the model, enabling the model to automatically focus on key attributes and suppress redundant attributes. Through the above iterative optimization, when the model training is completed, the final comprehensive score of information attributes and the gating value can be output, realizing deep collaboration between information attribute selection and model training.

Images