A defect root cause identification method based on special equipment report, electronic device, medium and program product

Abstract

The application provides a defect root cause identification method based on special equipment reports, an electronic device, a medium and a program product, and belongs to the technical field of special equipment detection. The method comprises the following steps: acquiring a plurality of special equipment reports; extracting detection data of all modalities in each special equipment report; fusing the detection data of all modalities to obtain fused data; outputting defect categories of all the special equipment reports through a defect detection model; determining all N-item item sets based on the number of each defect category; determining an initial causal chain based on the N-item item sets; and obtaining a defect root cause of special equipment based on a pre-established causal knowledge graph and the initial causal chain. The technical scheme provided by the application meets the large-scale report batch processing demand, eliminates artificial subjective bias, and ensures consistent and reliable results.

Description

Technical Field

[0001] This invention relates to the field of special equipment inspection technology, and more specifically, to a method, electronic equipment, medium, and program product for identifying the root cause of defects based on special equipment reports. Background Technology

[0002] With the rapid development of the special equipment industry, the number of special equipment inspection reports has increased exponentially. These reports contain multimodal information, including basic equipment information, inspection process records, sensor measurement data such as ultrasonic thickness measurement and hydrostatic testing, defect descriptions, and evaluation conclusions. They are the core basis for identifying equipment safety hazards and conducting root cause analysis of defects. Currently, defect root cause identification based on special equipment reports mainly relies on manual review and traditional data analysis methods. Some automated solutions use natural language processing technology to extract defect information from the report text, and then use traditional association rule algorithms (such as Apriori) to mine the co-occurrence relationships between defects, thereby inferring possible root causes. However, existing technologies have the following limitations: most only perform single-modal analysis on the report text, ignoring the objective verification role of sensor data, making it difficult to detect hidden defects where the text and measured data are inconsistent, resulting in insufficient defect identification accuracy; traditional association rule algorithms can only mine statistical co-occurrence relationships, failing to distinguish between true causality and accidental pseudo-associations, which can easily lead to misleading conclusions; moreover, they are mostly qualitative analyses, unable to quantify the contribution of each factor to the defect, and difficult to generate a complete root cause transmission chain, resulting in improvement suggestions lacking specificity and priority; at the same time, the degree of automation is low, with many steps relying on human experience, which is not only difficult to meet the batch processing needs of millions of reports, but the results are also easily affected by subjective factors and inconsistent. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a method, electronic device, medium and program product for identifying the root cause of defects based on special equipment reports, which can improve the problem of insufficient accuracy of defect identification in the prior art.

[0004] To achieve the above technical objectives, the technical solution adopted in this application is as follows:

[0005] In a first aspect, embodiments of this application provide a method for identifying the root cause of defects based on special equipment reports, the method comprising:

[0006] Obtain multiple special equipment reports;

[0007] Based on a preset extraction tool, detection data for all modalities are extracted from each of the special equipment reports;

[0008] According to the preset fusion strategy, the detection data of all the modalities are fused to obtain fused data;

[0009] The fused data of all the special equipment reports are input into the defect detection model, and the defect detection model outputs the defect categories of all the special equipment reports.

[0010] Based on the number of each defect category, all N-item sets are determined, wherein each N-item set includes N defect categories, and the number of special equipment reports corresponding to the N defect categories is greater than or equal to a preset number, where N is an integer greater than or equal to 2;

[0011] Based on the N-item set, an initial causal chain is determined, wherein any defect category of the N-item set represents the antecedent of the initial causal chain, and other defect categories of the N-item set represent the consequent of the initial causal chain;

[0012] Based on the pre-established causal knowledge graph, the initial causal chain, and the equipment characteristics corresponding to the detection data, the root causes of defects in special equipment are obtained.

[0013] Secondly, embodiments of this application also provide an electronic device, the electronic device including a processor and a memory coupled to each other, the memory storing a computer program, and when the computer program is executed by the processor, causing the electronic device to perform the method described in the first aspect.

[0014] Thirdly, embodiments of this application also provide a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when run on a computer, causes the computer to perform the method described in the first aspect.

[0015] Fourthly, embodiments of this application also provide a program product, characterized in that it includes a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0016] The invention employing the above technical solution has the following advantages:

[0017] The technical solution provided in this application first extracts full-modal detection data from the report, and generates fused data through semantic space alignment, similarity weighting, and gating filtering to achieve cross-validation of text and sensor data, effectively detecting hidden defects where the text and actual measurements are inconsistent. Then, the fused data is input into a defect detection model to obtain defect categories, generating N-item sets and initial causal chains. The intermediate causal chains are obtained by filtering pseudo-associations through a causal knowledge graph, avoiding misleading conclusions from traditional algorithms. Subsequently, the contribution of entity nodes is quantified through a pre-trained XGBoost model, and a complete root cause transmission chain is generated by combining the knowledge graph, realizing the transformation from qualitative to quantitative analysis. Finally, improvement suggestions are generated by combining the maintenance knowledge base and sorted by priority. The entire process is automated, which not only meets the needs of large-scale batch processing of reports, but also eliminates human subjective bias and ensures consistent and reliable results. Attached Figure Description

[0018] This application can be further illustrated by the non-limiting embodiments given in the accompanying drawings. It should be understood that the following drawings only illustrate some embodiments of this application and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained from these drawings without any inventive effort.

[0019] Figure 1 A flowchart of a defect root cause identification method based on special equipment reports provided in this application embodiment.

[0020] Figure 2 This is a sub-flowchart of S130 provided in an embodiment of this application.

[0021] Figure 3 This is a sub-flowchart of S170 provided in an embodiment of this application. Detailed Implementation

[0022] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that similar or identical parts are referred to by the same reference numerals in the drawings or description. Implementations not shown or described in the drawings are forms known to those skilled in the art. In the description of this application, terms such as "first" and "second" are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] Please refer to Figure 1 This application provides a method for identifying the root causes of defects based on special equipment reports. This method can be applied to electronic devices, and the steps of the method can be executed or implemented by the electronic device. The electronic device can be, but is not limited to, personal computers, smartphones, etc. The method for identifying the root causes of defects based on special equipment reports may include the following steps:

[0024] S110, obtain multiple special equipment reports;

[0025] S120, based on a preset extraction tool, extract detection data for all modalities from each of the special equipment reports;

[0026] S130, according to the preset fusion strategy, fuse the detection data of all modalities to obtain fused data;

[0027] S140, input the fused data of all the special equipment reports into the defect detection model, and output the defect categories of all the special equipment reports through the defect detection model;

[0028] S150, based on the number of each defect category, determine all N-item sets, wherein the N-item set includes N defect categories, the number of special equipment reports corresponding to the N defect categories is greater than or equal to a preset number, and N is an integer greater than or equal to 2;

[0029] S160, Based on the N-item set, determine an initial causal chain, wherein any defect category of the N-item set represents the antecedent of the initial causal chain, and other defect categories of the N-item set represent the consequent of the initial causal chain;

[0030] S170, based on the pre-established causal knowledge graph, the initial causal chain, and the equipment characteristics corresponding to the detection data, the root cause of the defect in the special equipment is obtained.

[0031] In the above implementation, this solution first acquires multiple special equipment reports, extracts detection data for all modalities from each report using a preset extraction tool, and then fuses the detection data for all modalities according to a preset fusion strategy to obtain fused data. The fused data is then input into a defect detection model to output the defect categories of all reports, solving the problem that existing technologies only perform single-modal analysis on the report text and have insufficient defect identification accuracy. Next, based on the number of each defect category, a preset number of N itemsets is determined. An initial causal chain is generated from the N itemsets, with the antecedent being a defect category and the consequent being other defect categories. Then, combined with a pre-established causal knowledge graph, the root cause of the special equipment defect is obtained, solving the problem that traditional association rule algorithms can only mine statistical co-occurrence relationships and cannot distinguish between true causality and accidental pseudo-associations, which can easily mislead root cause localization. The entire process is executed according to preset steps without manual intervention, solving the problems of low automation, difficulty in meeting the needs of batch report processing, and inconsistent results easily affected by subjective factors in existing technologies.

[0032] The following is a detailed explanation of each step in the defect root cause identification method based on special equipment reports:

[0033] In S110, multiple special equipment reports can be randomly extracted using a BS-based extraction model. For example, multiple special equipment reports can be obtained through an extraction model based on a browser / server (BS) architecture. This model adopts a mature deployment method that separates the client and server. Users access the extraction system interface through any standard browser, input filter conditions such as the number of reports to be extracted, time range, and equipment type, and then submit a request. After receiving the request, the server connects to the structured database storing all special equipment reports, first filters the reports according to the filter conditions to obtain the complete set of reports that meet the requirements, and then calls the system's built-in Mason swirl pseudo-random number generation algorithm to generate a non-repeating random index that matches the number of reports to be extracted. Based on the random index, the corresponding number of special equipment reports are extracted from the complete set of reports, thus completing the random acquisition of multiple reports.

[0034] Understandably, other existing random sampling models can also be used to randomly sample a large number of special equipment reports. For example, a stratified random sampling model can be used, which first divides all reports into non-overlapping strata based on dimensions such as equipment type, region of use, and inspection cycle, and then performs independent random sampling within each stratum, ensuring that the distribution of the sampled report is consistent with the total number of reports. This is suitable for root cause analysis scenarios where the sample needs to be fully representative. Another example is a systematic sampling model, which first sorts all reports by unique identifier, calculates a fixed sampling interval, and then randomly selects a starting point from the first interval, and so on, sampling every other interval. Retrieving a single report requires no generation of numerous random numbers, resulting in extremely high computational efficiency, making it suitable for rapid batch extraction of reports exceeding one million. Similarly, the reservoir sampling model eliminates the need for prior knowledge of the total number of reports. By dynamically maintaining a fixed-size sample pool, it performs real-time random sampling of the incoming report stream, making it suitable for extracting continuously generated incremental report data. Furthermore, a weighted random sampling model can be employed, assigning different sampling weights to each report based on factors such as equipment risk level and historical defect records. Reports with higher weights have a greater probability of being selected, making it suitable for risk-oriented targeted sampling and root cause analysis.

[0035] In S120, the special equipment report in this embodiment includes detection data in three modalities, including text data, sensor data, and structural data. In this embodiment, sensor data and structural data are defined as physical feature data. Sensor data is time-series quantitative detection data collected during the inspection process, and structural data is structured parameters of equipment attributes. Together, they constitute multi-source feature inputs in the physical dimension and participate in subsequent cross-modal fusion.

[0036] Text data refers to unstructured semantic information recorded in natural language in special equipment reports. It is mainly distributed in the inspection instructions, defect descriptions, evaluation conclusions, and rectification requirements of the report. It carries the subjective judgment and qualitative description of the equipment status by the inspectors. It is rich in semantic information but has a flexible and diverse format. For example, "The measured wall thickness at the 3rd inspection point of the pressure vessel cylinder is 11.2 mm, which is less than the minimum design wall thickness of 12.0 mm. It is judged to be a wall thickness deficiency defect and a re-inspection is required within 3 months."

[0037] Sensor data are objective, quantitative data automatically collected by various testing instruments during the inspection of special equipment. They mainly include wall thickness values ​​collected by ultrasonic thickness gauges, pressure-time curves recorded by hydrostatic testing systems, detection signals from magnetic particle flaw detectors, and medium temperatures collected by temperature sensors. They are characterized by numerical data, time-series data, and strong objectivity, and are the core basis for verifying the true state of the equipment. For example, "Pressure values ​​collected at 120 time points during the hydrostatic test: [0,0.5,1.0,1.5,1.6,1.6,...,1.6,1.0,0.5,0] MPa, with no significant pressure drop during the 30-minute pressure holding period."

[0038] Structured data, presented in fixed tables and forms in special equipment reports, mainly includes basic equipment information tables, inspection item record tables, defect statistics tables, and standard reference tables. It features a unified format, clearly defined fields, and ease of machine parsing, serving as a bridge between text data and sensor data. For example, "Basic Equipment Information Table: Equipment No. EQ20260605001, Equipment Type: Fixed Pressure Vessel, Design Pressure: 1.6MPa, Service Life: 12 years, Material: Q245R, Inspection Date: 2026-06-05". The structured data extracted from the report serves a dual purpose: firstly, it works alongside sensor data as physical feature data for cross-modal fusion, supporting defect category detection; secondly, it directly serves as the equipment feature data source for root cause analysis, and after standardized numerical mapping, it is used for subsequent entity node contribution calculations. The entire process is based on the same extraction result, eliminating the need for repeated report parsing.

[0039] In this embodiment, the preset extraction tool can be a dedicated large language model based on the DistilBERT-6L-768D pre-trained model and fine-tuned through a large number of historical reports of special equipment, or it can be replaced by open-source large language models such as Qwen-7B and Llama-3-8B after fine-tuning with data from the same domain. The model first identifies fixed tables and form areas in the report, extracts the table headers and corresponding cell contents and maps them to predefined structured fields to complete the standardized extraction of structured data. Secondly, it locates unstructured text paragraphs such as inspection instructions, defect descriptions, and evaluation conclusions in the report, and extracts semantic information such as defect category, location, severity, and rectification requirements through named entity recognition and relation extraction technology to obtain text data. Finally, it identifies the instrument test results presented in the form of numerical lists and curve annotations in the report, extracts quantitative information such as ultrasonic thickness measurement, hydrostatic test pressure sequence, and medium temperature and converts them into calculable numerical formats to complete the extraction of sensor data.

[0040] In S130, because the text data, sensor data, and structural data have different modalities, it is necessary to fuse them and analyze the defect categories based on the fused data. Based on this, as follows... Figure 2 As shown, S130 includes the following steps:

[0041] S131: Input the text data into the transformation model, and the transformation model adds multi-head self-attention to the text data to obtain a global text feature matrix;

[0042] S132: Project the physical feature data onto the semantic space of the text data to obtain a physical feature matrix with the same dimension as the global text feature matrix;

[0043] S133: Perform global average pooling along the sequence dimension on the global text feature matrix to obtain a global text semantic vector. Calculate the inner product of the feature vector at each position in the physical feature matrix and the global text semantic vector to obtain a similarity weight. The similarity weight is used to characterize the degree of correlation between physical feature data and text data.

[0044] S134: The physical feature matrix is ​​weighted by the similarity weight to obtain a cross-modal feature matrix that represents the matching part between text data and physical feature data;

[0045] S135: Input the cross-modal feature matrix into the sigmoid gating function, and obtain the gating weights through the sigmoid gating function;

[0046] S136: Obtain the fused data based on the gating weights and the cross-modal feature matrix.

[0047] The specific solution in S131 is as follows:

[0048] The text data is input into a conversion model, which determines the semantic features of each character in the text data in a specified dimension based on a pre-trained semantic sub-model. The conversion model matches the corresponding token to each character based on the semantics of each clause in the text data to obtain a tag sequence. There are M types of tokens, and the semantic rules corresponding to each token are pre-stored in the conversion model, where M is an integer greater than 1. Each clause includes several characters. The conversion model uses the tag sequence as an index and extracts and divides the semantic features of all characters according to the type of token to obtain M first sub-matrices. Each first sub-matrice contains the semantic features of the character corresponding to the token type. Each first sub-matrice is then split into Q second sub-matrices along the feature dimension. Q second submatrices corresponding to the first submatrix are respectively input into the analysis head corresponding to the multi-head self-attention. The analysis head obtains the similarity of each second submatrix within the same first submatrix with respect to other second submatrices. The analysis head is pre-deployed in the transformation model. Based on the similarity of each second submatrix within the same first submatrix with respect to other second submatrices, the attention weight of each second submatrix is ​​obtained. The analysis head weights all second submatrices within the corresponding range based on the attention weights to obtain the weighted matrix corresponding to each analysis head. The transformation model first concatenates all weighted matrices corresponding to each first submatrix to obtain the attention enhancement feature matrix of each token type. All token type attention enhancement feature matrices are backfilled and concatenated according to the original character position order to obtain the global text feature matrix.

[0049] In this embodiment, the transformation model deeply integrates character-level fine-grained semantic features with predefined domain rule token labels. Semantic fields are divided using token labels as masks, and a multi-head self-attention mechanism with independent fields is used to jointly model text semantics and structural information. This preserves the original semantic details of each character while introducing domain knowledge constraints from special equipment reports through token labels to achieve field-level semantic isolation. At the same time, multi-head self-attention is used to capture the global dependencies between different semantic positions within each field. This solves the problems of semantic and structural separation and insufficient integration of domain knowledge in traditional text feature extraction methods, providing a reliable text feature foundation for subsequent multimodal data fusion and defect detection.

[0050] In this embodiment, the conversion model first calls the internally pre-trained semantic sub-model to semantically encode each character in the text data, generating semantic features of a specified dimension for each character, which are then arranged in character order to form a global semantic feature matrix. Simultaneously, based on the semantic content of each clause in the text data and according to the semantic rules corresponding to M pre-stored tokens, the conversion model matches the corresponding token type for each character, generating a label sequence with the same length as the text. This label sequence is only used as a positional mask for field partitioning and does not participate in feature value calculations. Subsequently, using the label sequence as an index, the conversion model extracts feature vectors for all positions of the corresponding category from the global semantic feature matrix according to the token type, splitting it into M first sub-matrices. The pure semantic features corresponding to each type of token are independently distributed in the corresponding first sub-matrices. Each first submatrix is ​​then uniformly split into Q second submatrices along the feature dimension, corresponding to the Q analysis heads of the multi-head self-attention system. The Q second submatrices for each field are independently input into the field-specific multi-head self-attention analysis head. Each analysis head calculates the similarity between all feature positions within the current field, normalizes based on similarity to obtain the attention weight for each position, and then performs a weighted summation of all positional features according to the attention weights to obtain the weighted matrix output by that analysis head. Finally, the weighted matrices output by the Q analysis heads are concatenated within each field to obtain the attention-enhanced feature matrix for that field. The transformation model then backfills and concatenates the attention-enhanced feature matrices of all fields according to the original character position order to obtain the final global text feature matrix.

[0051] For example, this embodiment takes the text data "The measured wall thickness of the pressure vessel cylinder is 11.2mm, which is determined to be insufficient wall thickness" as an example. The conversion model first splits the text into 18 characters, calls the pre-trained semantic sub-model to generate 512-dimensional semantic features for each character, and forms a global semantic feature matrix with dimensions [18, 512] in sequence. For example, the semantic features of the character "wall" include core semantic information such as "equipment structural component" and "thickness measurement object", and the semantic features of "11.2" include semantic information such as "numerical parameter" and "measurement result". Then, based on the semantics of the clause and the five pre-stored Token semantic rules, the conversion model matches the corresponding Token type for each character and generates a label sequence of length 18 [0,0,0,0,0,1,1,1,1,1,1,1,1,1,1,2,2,2,2], where 0 represents the test object Token, 1 represents the parameter record Token, and 2 represents the evaluation conclusion Token. This sequence is only used as a position partitioning mask.

[0052] The transformation model uses the above label sequence as an index to extract the feature vectors at the corresponding positions in the global semantic feature matrix according to the category, and splits them into 3 first sub-matrices (only 3 token types appear in this example): the first sub-matrix of the test object contains the pure semantic features of the first 5 characters, with dimensions [5, 512]; the first sub-matrix of the parameter record contains the pure semantic features of the middle 9 characters, with dimensions [9, 512]; the first sub-matrix of the evaluation conclusion contains the pure semantic features of the last 4 characters, with dimensions [4, 512].

[0053] The transformation model uniformly splits each first submatrix into eight second submatrices along the feature dimension, with each second submatrix having a dimension of [L_i, 64] (L_i being the number of characters in the corresponding field). The eight second submatrices for each field are then input into eight dedicated analysis heads for that field. Taking the first analysis head of the parameter record field as an example, it calculates the similarity between all character features within that field. The feature similarity between the positions corresponding to "11.2mm" and "wall thickness" is found to be the highest, thus assigning it the highest attention weight. The outputs of the eight analysis heads within each field are concatenated to obtain the attention-enhanced features for that field. Finally, the transformation model backfills and concatenates the features from the three fields according to the original character order, resulting in a global text feature matrix with dimensions [18, 512]. This matrix simultaneously integrates the semantic information of each character, the field structure information, and the semantic association information within the field. For example, the semantic rules for a token can be shown in Table 1.

[0054] Table 1 Semantic Rules for Tokens

[0055]

[0056] The semantic understanding model used in the above example is a dedicated semantic understanding model based on the DistilBERT-6L-768D base model and specifically trained with special equipment domain data. Alternatively, open-source large language models such as Qwen-7B and Llama-3-8B can be used as the underlying architecture to obtain stronger long-text semantic capture capabilities. The training process of this model is as follows: First, a large number of historical inspection reports of special equipment are collected. After preprocessing such as garbled character filtering, format unification, and professional terminology correction, a domain pre-training corpus is constructed. The base model is then further pre-trained to learn the professional terminology of the special equipment domain. The model first establishes fixed sentence structures and industry semantic patterns. Then, it manually annotates multiple reports with token-by-token semantic role labels (corresponding to 5 token types), constructing a semantic annotation dataset containing categories such as inspection objects, parameter records, and evaluation conclusions. The pre-trained model is then supervised and fine-tuned using the cross-entropy loss function to optimize the model's recognition accuracy for different semantic roles. Finally, the semantic rules corresponding to the 5 tokens are embedded into the model's reasoning process as hard constraints, and the labels output by the model are automatically verified and corrected. Ultimately, a dedicated semantic understanding model is obtained that can accurately understand the semantics of special equipment reports and match the 5 token types.

[0057] In S132, since the physical feature data includes structural data and sensor data, it is first necessary to perform categorized preprocessing and feature concatenation on the physical feature data. Specifically:

[0058] For structural data, a combination of one-hot encoding and numerical normalization is used to convert discrete attributes such as equipment type and material category into one-hot vectors of corresponding dimensions. Continuous numerical attributes such as service life, design pressure, and design wall thickness are normalized using min-max normalization and then concatenated to obtain a fixed-dimensional global structural attribute feature vector. For sensor data, its time series / measuring point sequence structure is preserved to form the original sensor feature sequence.

[0059] The global structural attribute feature vector is broadcast along the time series / measurement point dimension and concatenated with the original sensor feature sequence along the feature dimension to obtain the complete original physical feature sequence matrix.

[0060] Finally, the original physical feature sequence matrix is ​​input into a learnable linear projection layer. The output dimension of this projection layer is completely consistent with the last dimension of the global text feature matrix. Through linear transformation, the physical feature data is mapped from the original numerical measurement and attribute space to the semantic space of the text data, resulting in a physical feature matrix that perfectly matches the dimension of the global text feature matrix. This achieves spatial alignment of the three-modal data and lays the foundation for subsequent similarity calculation.

[0061] The learnable linear projection layer proposed in this embodiment is a fully connected neural network layer. Its parameters (weight matrix and bias vector) are automatically learned during the end-to-end training of the entire defect detection model, without the need for manual design of mapping rules. The input dimension of the projection layer is the sum of the sensor feature dimension and the structural attribute feature dimension (e.g., 10-dimensional ultrasonic thickness measurement sequence spliced ​​with 32-dimensional structural attribute features, corresponding to an input dimension of 42). The output dimension is exactly the same as the last dimension of the global text feature matrix (e.g., 512-dimensional). During training, the weight matrix is ​​initialized with a uniform distribution of Xavier and the bias vector is initialized with a zero vector. In the model training phase, paired text data and corresponding physical feature data are input. After projection layer mapping, similarity calculation, and gated fusion, backpropagation is finally performed with the cross-entropy loss function of defect classification as the target. The weights and biases of the projection layer are continuously updated iteratively through the gradient descent algorithm, so that it learns the transformation relationship of optimally mapping physical features from numerical measurement and attribute space to text semantic space, so that semantically related physical features and text features have higher similarity in the same space. After training, the parameters of the projection layer are fixed, and the fixed parameters are directly used to complete the spatial alignment of physical features in the inference phase.

[0062] In S133, since the global text feature matrix corresponds to the length of the text character sequence and the physical feature matrix corresponds to the length of the measurement point / time sequence, the sequence dimension lengths of the two are inconsistent, and element-by-element matching operations cannot be directly performed. Therefore, global average pooling is first performed on the global text feature matrix along the sequence dimension to aggregate the semantic features of all character positions into a single global text semantic vector. The dimension of the global text semantic vector is consistent with the feature dimension, which fully represents the overall semantics of a single report text.

[0063] The global text semantic vector is then multiplied by the feature vector of each sequence position in the physical feature matrix to obtain a similarity weight vector with the same length as the sequence of the physical feature matrix. The weight value of each position in the similarity weight vector is used to characterize the degree of correlation between the physical feature data of the corresponding position and the overall semantics of the text. The closer the weight value is to 1, the higher the matching degree between the physical feature of that position and the core content of the text description. The closer the weight value is to 0, the lower the correlation between the two.

[0064] In S134, the similarity weight vector is broadcast and expanded along the feature dimension into a weight matrix with the same dimension as the physical feature matrix. Then, it is multiplied element-wise with the physical feature matrix to perform weighted processing on each position of the physical feature matrix. Physical features at high-weight positions are preserved and enhanced, while physical features at low-weight positions are suppressed. Finally, a cross-modal feature matrix containing only the semantic matching part of the text data is obtained, effectively filtering out noise information in the physical features that is irrelevant to the text description.

[0065] In S135, the cross-modal feature matrix is ​​input into the sigmoid activation function. The sigmoid function nonlinearly maps each element of the cross-modal feature matrix to a value between 0 and 1, resulting in a gated weight matrix with the same dimension as the cross-modal feature matrix. The gated weights are used to control the contribution of cross-modal features in the final fused data. The closer the value is to 1, the more reliable the cross-modal feature at that position is, and the higher the fusion weight should be assigned.

[0066] In step S136, the gated weight matrix and the cross-modal feature matrix are multiplied element-wise to perform final gating and filtering of the cross-modal features, resulting in a two-dimensional sequence-level fusion feature matrix. Subsequently, global average pooling and global max pooling operations are performed along the sequence dimension of this two-dimensional matrix to extract global statistical features along the sequence dimension. The two pooling results are concatenated to obtain a 512-dimensional sequence-level global fusion feature vector. Simultaneously, a 256-dimensional global structural attribute feature vector obtained from the structural data encoding is extracted, and these two vectors are concatenated along the feature dimension to obtain a 768-dimensional fixed-length one-dimensional fusion feature vector, which serves as the input to the defect detection model. This fusion vector simultaneously encompasses three types of information: textual semantics, sensor measurements, and device attributes. With fixed dimensions, it can be directly input into a classification model, effectively detecting hidden defects where textual descriptions differ from actual measurement data.

[0067] For example, taking the text data "The measured wall thickness of the pressure vessel cylinder is 11.2 mm, which is determined to be insufficient" and the corresponding physical characteristic data as an example, the sensor data is a sequence of 10 cylinder wall thickness measurement values ​​collected by an ultrasonic thickness gauge: [11.5, 11.3, 11.2, 11.1, 11.4, 11.6, 11.3, 11.2, 11.0, 11.2], and the structural data is the equipment attribute parameters: equipment type is fixed pressure vessel, material is Q245R, service life is 12 years, and design pressure is 1.6 MPa. First, the text data is input into the transformation model to obtain a global text feature matrix of dimensions [18, 512]. The feature vectors corresponding to "11.2mm" and "insufficient wall thickness" contain the semantics of "wall thickness measurement result" and "defect judgment." Next, the physical feature data is preprocessed: the structural data is encoded, with equipment type and material converted into unique heat vectors, and service life and design pressure normalized into numerical features. After a fully connected transformation layer, a 256-dimensional global structural attribute feature vector is obtained. This vector is then broadcast along the temporal dimension of 10 measurement points and concatenated with a 10-dimensional wall thickness measurement value sequence to form an original physical feature sequence matrix of dimensions [10, 266]. A linear projection layer then maps this matrix to a physical feature matrix of dimensions [10, 512], achieving spatial alignment with the text features. Finally, the inner product of the feature vector at each position in the physical feature matrix and the global text semantic vector is calculated to obtain a 10-dimensional similarity weight vector, where the weight corresponds to the measurement of 11.2mm. The positional weight reached 0.89, and the semantic features of "Q245R material" and "12-year service life" in the structural data were associated with the semantic features of "insufficient wall thickness" in the text, which simultaneously improved the similarity weight of the corresponding position. Then, the physical feature matrix was weighted with the similarity weight to obtain a cross-modal feature matrix of [10, 512], which filtered out noise that was irrelevant to the wall thickness measurement and equipment attributes. The cross-modal feature matrix was then input into the sigmoid gating function to obtain a gating weight matrix between 0 and 1, where the positional gating weight corresponding to the 11.2mm measurement value was 0.92. Finally, the gating weight was multiplied with the cross-modal feature matrix to obtain the final fused data. This data simultaneously contained the semantic judgment of "insufficient wall thickness" in the text, the 11.2mm wall thickness value measured by the sensor, and basic attribute information such as equipment material and service life. It can support defect judgment from three dimensions: semantic description, measured data, and equipment attributes, effectively improving the accuracy of defect identification.

[0068] Based on this, the preset fusion strategy is a semantic alignment-driven trimodal data fusion method. First, the text is converted into a global feature matrix that integrates semantics, structure, and domain rules. Simultaneously, the sensor time-series data and structural attribute data in the physical feature data are encoded and concatenated, and then mapped to the text semantic space through a learnable linear projection layer to achieve dimensional alignment. Next, the similarity between modes is calculated through inner product and irrelevant noise is filtered by weighting. Finally, the reliability is filtered by sigmoid gating, and the output is fused data that has text semantic understanding ability, objective quantitative information from sensors, and structural attribute features of equipment. It can effectively detect hidden defects that are inconsistent between text description and measured data, and can also help improve the accuracy of defect classification through equipment attribute information.

[0069] In S140, the defect detection model can be a multilayer perceptron (MLP) deep classification model. The principle of the MLP defect detection model is to fit the complex implicit mapping relationship between high-dimensional fusion features and defect categories through the nonlinear transformation capability of multilayer fully connected neural networks. Its input is a 768-dimensional cross-modal fusion feature vector. Through stacked fully connected layers, features at different levels of abstraction are extracted layer by layer. Nonlinear activation functions such as ReLU are used to break the limitations of linear transformation, thereby uncovering the deep correlation between textual semantic information and sensor quantization data (such as the causal relationship between the semantic description of "insufficient wall thickness" and the measurement value of low wall thickness). The output layer uses the sigmoid activation function, so that each neuron independently outputs the probability of occurrence of the corresponding defect category (0-1 interval), realizing multi-label defect detection. The entire model is trained end-to-end with the binary cross-entropy loss function as the target. The weight parameters of each layer are automatically updated through the backpropagation algorithm. Without the need for manual feature engineering, it can automatically learn the optimal defect recognition pattern.

[0070] The defect detection process of this model is as follows: First, the 768-dimensional fused feature vector of a single special equipment report is input into the input layer of the model. Then, the features are transformed through multiple hidden layers in sequence. Taking a typical 3-layer MLP as an example, the first hidden layer contains 256 neurons, each fully connected to the 768 input features. After linear combination and bias superposition through the weight matrix, the ReLU activation function is used to output 256-dimensional primary abstract features. The second hidden layer contains 128 neurons, which perform the same fully connected transformation and nonlinear activation on the 256-dimensional primary features to output 128-dimensional high-level abstract features. Finally, the output layer is entered. The number of neurons in the output layer is consistent with the total number of predefined defect categories (e.g., 23 categories). Each neuron performs a linear transformation on the 128-dimensional high-level features and outputs the probability of occurrence of the corresponding defect category through the sigmoid activation function. The probabilities of all defect categories are compared with a preset threshold (e.g., 0.5) one by one. Defect categories with probabilities greater than or equal to the threshold are retained, and finally, a list of all defect categories present in the report is output.

[0071] In this embodiment, through steps S110-S140, the defect category of each special equipment report can be obtained. It is understood that each special equipment report has more than one defect category. For example, after a pressure vessel inspection report is processed by the aforementioned steps, the fused data is sent to the MLP defect detection model. The model calculates that the probabilities of two defects, insufficient wall thickness and excessive medium corrosion, are both higher than the 0.5 judgment threshold. Finally, the report outputs two defect categories, insufficient wall thickness and excessive medium corrosion, simultaneously. Some old equipment reports may also detect three types of defects at the same time: shell corrosion, weld cracks, and substandard pressure resistance. This reflects the actual inspection status of multiple defects coexisting in a single report and provides a data foundation for subsequent mining of N frequent itemsets composed of multiple defects.

[0072] In this embodiment, the entire model adopts a phased, progressive training scheme. The training order, data requirements, and loss function of each module are clearly defined and can be completely reproduced. The specific process is as follows:

[0073] Phase 1: Domain Pre-training of Semantic Sub-models

[0074] Using a large corpus of unlabeled historical reports on special equipment, the data was preprocessed and then further pre-trained based on the DistilBERT-6L-768D model. A masked language model loss function was employed to enable the model to learn domain-specific terminology, fixed sentence structures, and semantic patterns. After pre-training, five token semantic rules were embedded into the inference process as hard constraints, resulting in a usable semantic sub-model. This stage is unsupervised pre-training and does not require defect-labeled data.

[0075] Phase 2: Cross-modal fusion + end-to-end joint training for defect detection

[0076] Using special equipment reports labeled with defect categories as the training dataset, each sample contains paired text data, sensor data, and structural data, with labels representing the corresponding report's multi-defect category binary tags. During training, the underlying Transformer parameters of the semantic sub-model are fixed. The upper-level transformation model, linear projection layer, similarity weighting module, sigmoid gating module, dimension transformation module, and MLP defect detection model are jointly trained end-to-end. The overall loss function uses multi-label binary cross-entropy loss, and the parameters of each learnable module are updated synchronously using gradient descent until the AUC of defect recognition on the validation set converges. After this stage, the entire chain of text feature extraction, cross-modal fusion, and defect classification can be directly used for inference.

[0077] Phase 3: Independent training of the XGBoost contribution model

[0078] Using historical defect detection reports as samples, the device feature vector mapped to the entity nodes of each report and the corresponding defect occurrence label are extracted. A separate binary classification training dataset is constructed for each predefined defect category, and the corresponding XGBoost binary classification model is trained accordingly. After training, the model can output the contribution of each entity node to the corresponding defect through feature importance calculation. This stage is trained independently of the preceding multimodal model and does not rely on gradient backpropagation from the preceding model.

[0079] In S150, an N-item set is defined as an N-item set containing a number of special equipment reports for all N defect categories that are greater than or equal to a preset minimum support threshold (i.e., a preset number), where N is an integer greater than or equal to 2. For example, when N=2, if the number of reports for both "insufficient wall thickness" and "excessive media corrosion" simultaneously meets the preset requirement, then {insufficient wall thickness, excessive media corrosion} is a 2-item set. When N=3, if the number of reports for "cylinder corrosion," "weld crack," and "pressure resistance failure" simultaneously meets the requirement, then {cylinder corrosion, weld crack, pressure resistance failure} is a 3-item set. The N-item set essentially reflects the statistical significance of multiple defect categories occurring together during the operation of special equipment. It can filter out defect combinations that coexist only by chance, providing basic data support for constructing initial causal chains and distinguishing between true causal relationships and false relationships.

[0080] Based on this, S150 specifically includes the following steps:

[0081] S151: Based on the number of each defect category, determine the number of special equipment reports containing the defect categories;

[0082] S152: Determine the frequency of the defect category based on the quotient of the number of special equipment reports containing the defect category divided by the total number of special equipment reports;

[0083] S153: Filter out all defect categories with a frequency greater than or equal to the preset frequency threshold;

[0084] S154: Select any two different defect categories from all defect categories whose frequency is greater than or equal to the preset frequency threshold, and generate a second-order candidate defect combination according to the predefined sorting rules;

[0085] S155: For all k-order candidate defect combinations, count the number of special equipment reports that simultaneously include all the defect categories within the k-order candidate defect combination, and obtain the frequency of the k-order candidate defect combination based on the total number of special equipment reports.

[0086] S156: Determine whether the iteration condition is met. The iteration condition is: among all k-order candidate defect combinations, there exists a k-order defect combination with a frequency greater than or equal to a preset frequency threshold, and among all k-order defect combinations with a frequency greater than or equal to the preset frequency threshold, there exists at least one pair of k-order defect combinations whose first k-1 defect categories are completely identical. If the iteration condition is met, then the pair of k-order defect combinations are concatenated and deduplicated, and arranged according to a predefined sorting rule to generate a k+1-order candidate defect combination.

[0087] S157: When the iteration condition is met, assign k+1 to k and execute S155-S156 again. When the iteration condition is not met, the k-order candidate defect combination is the N-item set, k=N, where k is an integer greater than or equal to 2, and the initial value of k is 2.

[0088] In S151 of this embodiment, the defect detection results of all special equipment reports are traversed, each predefined defect category is counted independently, and the total number of special equipment reports containing the defect category is calculated, that is, the absolute number of occurrences of the defect category, which provides basic statistical data for subsequent calculation of the frequency of single defects.

[0089] In S152 of this embodiment, the number of reports containing the defect corresponding to each defect category is divided by the total number of special equipment reports involved in this analysis to obtain the frequency of the defect category. This value is between 0 and 1 and is used to measure the prevalence of a single defect category in all reports.

[0090] In step S153 of this embodiment, all defect categories with a frequency greater than or equal to a preset frequency threshold are selected as the basic elements for generating higher-order candidate defect combinations. Only defect categories that meet the frequency threshold can participate in the subsequent mining of multiple defect combinations. In this embodiment, the preset frequency threshold is determined based on the ratio of the preset quantity to the total number of special equipment reports.

[0091] In S154 of this embodiment, any two different defect categories are selected from the frequency-compliant defect categories obtained through screening, and arranged according to a predefined sorting rule to generate an initial second-order candidate defect combination. At this time, the initial value of the iteration order k is set to 2.

[0092] The sorting rule is an ascending order rule for defect category coding, which is a fundamental prerequisite rule to ensure the representation of defect combinations. Specifically, it is defined as follows: During system initialization, a unique and non-repeating integer category code is assigned to all predefined defect categories. The codes can be assigned in segments according to the major technical category to which the defect belongs. All defect categories within any defect combination are arranged in a fixed order according to the numerical value of their corresponding codes, from smallest to largest. This rule eliminates differences in the arrangement order of defect combinations, ensuring that the same group of defect categories corresponds to only one combination representation, avoiding duplicate statistics and redundant candidate combinations. Furthermore, it provides a unified comparison benchmark for judging the splicing conditions of subsequent higher-order candidate combinations, ensuring that the splicing logic is reproducible.

[0093] In S155 of this embodiment, for all current k-order candidate defect combinations, all special equipment reports are re-traversed, the number of reports that simultaneously contain all defect categories within the candidate combination is counted, and then the number is divided by the total number of reports to obtain the frequency of the corresponding k-order candidate defect combination, which is used to determine whether the multi-defect combination has statistical co-occurrence significance.

[0094] In this embodiment, S156 is used to determine whether the conditions for iteratively generating higher-order candidate combinations are met, and to complete the generation of higher-order candidate combinations when the conditions are met: First, all k-order candidate defect combinations are traversed, and k-order defect combinations with a frequency greater than or equal to a preset frequency threshold are selected; then, it is determined whether the iteration conditions are met. The iteration conditions include two levels of determination: first, there are k-order defect combinations with a frequency greater than or equal to a preset frequency threshold; second, among all k-order defect combinations with the required frequency, there is at least one pair of combinations whose first k-1 defect categories are completely identical after being arranged according to a predefined sorting rule; if both levels of determination are met, it is determined that the iteration conditions are met, and the pair of k-order defect combinations that meet the splicing requirements are spliced ​​and deduplicated, and arranged according to a predefined sorting rule to generate k+1-order candidate defect combinations; if any level of determination is not met, it is determined that the iteration conditions are not met, and higher-order candidate combinations cannot be generated.

[0095] In this embodiment, S157 is used to control the flow and termination of the iteration loop: when the iteration condition is met, k+1 is assigned to the iteration order k, and steps S155 to S156 are executed again to perform frequency calculation and iteration judgment for the new order candidate combination; when the iteration condition is not met, the iteration terminates, and all k-order defect combinations that meet the frequency standard are now called k-item sets. All item sets containing more than or equal to 2 defect categories are collectively called N-item sets. Here, k is an integer greater than or equal to 2, and the initial value of k is 2, corresponding to the second-order candidate defect combination in the first iteration.

[0096] For example, suppose this analysis involves 1000 pressure vessel inspection reports, with a preset frequency threshold of 0.05 (meaning at least 50 reports simultaneously contain the corresponding defect combination). The predefined defect categories and corresponding codes are: insufficient wall thickness (code 01), excessive media corrosion (code 02), shell corrosion (code 03), weld cracks (code 04), and pressure resistance failure (code 05). All defect combinations are arranged in ascending order of codes to ensure that the combination description is unique.

[0097] First, execute steps S151, S152, and S153 to iterate through all reports, count the occurrence of each single defect, and calculate the frequency: insufficient wall thickness 0.12 (120 reports), excessive medium corrosion 0.10 (100 reports), cylinder corrosion 0.08 (80 reports), weld cracks 0.06 (60 reports), and pressure resistance failure 0.04 (40 reports). Among them, the frequency of pressure resistance failure is lower than the preset threshold and is removed. The remaining 4 defect categories are used as the basic elements for subsequent combination generation.

[0098] Next, execute step S154, select any two different categories from the four defect categories that meet the frequency criteria, arrange them according to the coding ascending order rule to generate 6 groups of second-order candidate defect combinations, and set the initial value of the iteration order k to 2.

[0099] Perform step S155. For all second-order candidate defect combinations, count the number of reports that simultaneously contain the corresponding two defects and calculate the frequency. The results show that {insufficient wall thickness, excessive medium corrosion} is 0.07 (70 reports) and {cylinder corrosion, weld crack} is 0.06 (60 reports). The frequencies of other candidate combinations such as {insufficient wall thickness, cylinder corrosion} are 0.03 (30 reports), {insufficient wall thickness, weld crack} is 0.02 (20 reports), and {excessive medium corrosion, cylinder corrosion} is 0.04 (40 reports) are all below the threshold.

[0100] Then, step S156 is executed to determine the iteration condition: Since k=2, two sets of second-order combinations with sufficient frequency are first selected. Then, it is checked whether there exists at least one pair of combinations where the first k-1=1 defect categories are completely identical. In the case of {insufficient wall thickness, excessive medium corrosion}, the first term is insufficient wall thickness; in the case of {cylinder corrosion, weld crack}, the first term is cylinder corrosion. Since the first terms of these two combinations are inconsistent, there is no combination pair that satisfies the splicing condition, therefore the iteration condition is not met.

[0101] Finally, step S157 was executed, but the iteration stopped because the iteration conditions were not met. The frequency-compliant defect combinations obtained in this mining all contain two defect categories, i.e., two-item sets. The final results are {insufficient wall thickness, excessive medium corrosion} and {cylinder corrosion, weld cracks}. Both belong to N-item sets with N≥2, reflecting that these two sets of defects have a statistically significant co-occurrence pattern in pressure vessels.

[0102] Assuming this analysis involves 1000 pressure vessel inspection reports, with a preset frequency threshold of 0.05 (meaning at least 50 reports simultaneously contain the corresponding defect combination), the predefined defect categories are arranged in ascending order of codes as follows: insufficient wall thickness (code 01), excessive medium corrosion (code 02), shell corrosion (code 03), weld cracks (code 04), and pressure resistance failure (code 05). All defect combinations follow this sorting rule.

[0103] First, execute steps S151-S154: After calculating the frequency of single defects, those that fail to meet the withstand voltage standard are removed because their frequency is 0.04, which is below the threshold. The remaining 4 defect categories generate 6 groups of second-order candidate defect combinations, and the iteration order k is set to an initial value of 2.

[0104] Perform step S155 to calculate the frequency of all second-order candidate combinations group by group:

[0105] {Insufficient wall thickness, excessive media corrosion}: 70 samples, frequency 0.07, meets standard;

[0106] Insufficient wall thickness, cylinder corrosion: 60 samples, frequency 0.06, meets standard;

[0107] Insufficient wall thickness, weld cracks: 20 samples, frequency 0.02, not up to standard;

[0108] {Excessive corrosion of the medium, corrosion of the cylinder}: 40 samples, frequency 0.04, not up to standard;

[0109] {Excessive corrosion from the medium, weld cracks}: 30 samples, frequency 0.03, not up to standard;

[0110] {Cylinder corrosion, weld cracks}: 65 samples, frequency 0.065, meets standard;

[0111] Next, step S156 is executed to determine the iteration conditions: Since k=2, three sets of second-order combinations with sufficient frequency are first selected. Then, it is checked whether there exists at least one pair of combinations that satisfy the condition that the first k-1=1 defect categories are completely identical. In the cases {insufficient wall thickness, excessive medium corrosion} and {insufficient wall thickness, cylinder corrosion}, the first term is both insufficient wall thickness, satisfying the requirement that the first term is completely identical. Therefore, the iteration conditions are satisfied.

[0112] Then execute step S157: splice the pair of 2nd-order combinations that meet the conditions and remove duplicates, arrange them in ascending order of codes to generate 3rd-order candidate defect combinations {insufficient wall thickness, excessive medium corrosion, cylinder corrosion}; assign k+1=3 to k, and return to execute steps S155-S156 again.

[0113] Entering the second iteration, step S155 is executed: the number of reports corresponding to the 3rd-order candidate combination {insufficient wall thickness, excessive medium corrosion, cylinder corrosion} is 45, and the frequency is calculated to be 0.045, which is lower than the preset threshold, and there is no 3rd-order combination with the required frequency.

[0114] Execute step S156 and determine the iteration condition: Currently k=3, there is no 3rd order combination with sufficient frequency, and there is no combination pair that can be used for splicing. Therefore, it is determined that the iteration condition is not met.

[0115] Finally, step S157 was executed: the iteration was terminated because the iteration conditions were not met. The frequency-compliant defect combinations obtained in this mining all contain 2 defect categories, i.e., 2-item sets. The final results are {insufficient wall thickness, excessive medium corrosion}, {insufficient wall thickness, cylinder corrosion}, and {cylinder corrosion, weld cracks}. All three belong to N-item sets with N≥2, reflecting that the corresponding defect combinations have a statistically significant co-occurrence pattern in pressure vessels.

[0116] In S160, all N-item sets generated in the preceding steps are first obtained. Each N-item set is processed independently. Each defect category in the N-item set is taken as the antecedent of the initial causal chain, and all other defect categories in the N-item set except the antecedent are taken as the consequents of the corresponding initial causal chain. This generates N different initial causal chains for each N-item set. Each chain is represented in a directed form of "antecedent defect category → consequent defect category set". This process generates all potential causal hypotheses by enumerating all possible defect causal directions in the N-item set.

[0117] Using the two-item set {insufficient wall thickness, excessive medium corrosion} obtained in the previous section as an example, this N-item set N=2, and the defects within the items are used in turn to construct the initial causal chain: first, with insufficient wall thickness as the antecedent and excessive medium corrosion as the consequent, the causal chain "insufficient wall thickness → excessive medium corrosion" is generated; then, with excessive medium corrosion as the antecedent and insufficient wall thickness as the consequent, the causal chain "excessive medium corrosion → insufficient wall thickness" is generated. Finally, two initial causal chains with different directions are generated from a single two-item set, and all potential causal points of this defect combination are completely enumerated, forming two sets of causal hypotheses to be verified by the subsequent knowledge graph.

[0118] Taking a three-item set {insufficient wall thickness, excessive medium corrosion, weld crack} as an example, with N set to 3, the three types of defects within the set are alternately used as antecedents, and the remaining two types of defects form consequents, generating three initial causal chains: "insufficient wall thickness → excessive medium corrosion, weld crack", "excessive medium corrosion → insufficient wall thickness, weld crack", and "weld crack → insufficient wall thickness, excessive medium corrosion". By completely enumerating all potential causal points of the three defect combinations, three initial causal hypotheses awaiting verification by the causal knowledge graph are obtained.

[0119] In S170, such as Figure 3 As shown, it may include the following steps:

[0120] S171: Based on the causal knowledge graph, delete the initial causal chain that does not conform to the standard causal chain to obtain the intermediate causal chain;

[0121] S172: Based on the causal knowledge graph, obtain the entity nodes in all standard causal chains that contain the defect categories involved in the intermediate causal chains;

[0122] S173: Based on the pre-trained XGBoost model, the contribution of each entity node to the defect category of the successor located in the intermediate causal chain is calculated according to the device characteristics.

[0123] S174: Sort all the contribution scores in descending order to obtain the entity nodes corresponding to the first P contribution scores, where P is an integer greater than or equal to 1;

[0124] S175: Based on the entity nodes corresponding to the first P contribution values ​​and the causal knowledge graph, determine the transmission chain from the entity node to the defect category in the intermediate causal chain, and the transmission chain serves as the root cause of the defect in the special equipment.

[0125] In this embodiment, the causal knowledge graph is a structured causal knowledge base pre-built based on multi-source data from the special equipment field, containing multiple verified standard causal chains. Its construction process is as follows:

[0126] The first step is to determine the sources of multi-source data: First, collect four types of core data to build an original corpus, such as the causes and transmission patterns of defects clearly recorded in national and industry standards such as the "Special Equipment Safety Technical Specifications" and the "Pressure Vessel Safety Technical Supervision Regulations"; second, special equipment failure analysis reports, accident investigation reports, and defect repair records accumulated in the industry over the past 10 years; third, experience summaries and review opinions from senior engineers and experts in the fields of special equipment inspection and structural safety; and fourth, research literature related to the defect mechanisms of special equipment included in publicly available academic databases, to achieve multi-dimensional data coverage of standards, rules, engineering cases, and academic mechanisms.

[0127] The second step is to build a standardized node system: all nodes are uniformly divided into two categories: physical nodes and defect nodes. Physical nodes are further subdivided into four subcategories: equipment attribute (e.g., material type, service life, design pressure), media environment (e.g., corrosive media, high-temperature operating environment), process manufacturing (e.g., unqualified welding process, missing heat treatment), and operation and maintenance management (e.g., overdue inspection, insufficient maintenance level). All nodes are named using industry standard terminology to avoid synonyms. Defect nodes correspond one-to-one with the predefined defect categories output by the aforementioned defect detection model, and the naming is completely consistent to ensure uniform terminology across the entire chain.

[0128] The third step is the extraction and initial construction of causal edges: For structured data such as standard clauses and rework records, causal triples of "inducing factor-intermediate phenomenon-final defect" are extracted through rule matching; for unstructured text data such as fault reports and documents, three types of directed causal relationships are automatically extracted based on the domain-fine-tuned BERT relationship extraction model: entity-entity, entity-defect, and defect-defect; all extracted causal edges are labeled with an initial confidence level, where the confidence level of causal relationships explicitly specified in the standard is set to 1.0, the confidence level of causal relationships labeled by expert experience is set to 0.9, and the confidence level of causal relationships automatically extracted by the model is set to 0.6-0.8.

[0129] The fourth step is causal edge screening and graph verification: a confidence threshold of 0.7 is set to filter out weak confidence relationships below the threshold; then, a review group composed of more than 3 experts in the field of special equipment with senior engineer titles verifies all the retained causal relationships one by one, eliminating accidental associations, reverse causality, and redundant erroneous edges, and finally forming accurate directed causal edges; multiple complete standard causal chains are formed by entity nodes, defect nodes and verified directed edges, and each standard causal chain represents a confirmed defect cause transmission path.

[0130] The fifth step is to establish a dynamic update mechanism: each quarter, newly added defect repair records, accident reports and new standard content are synchronized and new nodes and causal edges are added according to the above extraction-verification process; a full graph confidence review is carried out once a year, and the confidence of the corresponding causal edges is adjusted according to the newly added engineering data to ensure the accuracy and timeliness of the causal knowledge graph.

[0131] In the causal knowledge graph constructed above, the standard causal chain is composed of entity nodes and defect nodes connected by directed edges. The entity nodes are used to represent objective things that may cause defects, and the defect nodes are used to represent the defect category. The direction of the directed edge represents the direction of the causal relationship.

[0132] For example, this causal knowledge graph, taking the pressure vessel field as an example, includes two types of core nodes: entity nodes and defect nodes. Entity nodes include objective things that characterize the causes of defects, such as "corrosive media," "Q245R low-strength material," "service life exceeding 10 years," "unqualified welding process," and "material aging." Defect nodes include predefined defect categories such as "excessive media corrosion," "insufficient wall thickness," "weld cracks," and "pressure resistance failure." Nodes are connected by directed edges to form multiple standard causal chains, such as "corrosive media → excessive media corrosion → insufficient wall thickness," "service life exceeding 10 years → material aging → insufficient wall thickness," "unqualified welding process → weld cracks," "insufficient wall thickness → pressure resistance failure," and "weld cracks → pressure resistance failure." The direction of the directed edges represents the direction of causal relationship transmission, clearly demonstrating the complete causal logic from objective causes to intermediate defects and finally to severe defects. This provides authoritative domain knowledge basis for subsequent verification of the authenticity of the initial causal chain and elimination of accidental pseudo-associations.

[0133] Based on this, the process of S171 is as follows: each of the generated initial causal chains is matched and verified with the standard causal chain in the causal knowledge graph. The directed relationship of "predecessor defect category → consequent defect category" in the initial causal chain is checked to see if it exists in the directed edge of the standard causal chain. If it exists, the initial causal chain is retained. If it does not exist, it is determined to be an accidental pseudo-association and deleted. Finally, all the retained initial causal chains are the intermediate causal chains.

[0134] The process of S172 is as follows: extract all defect categories contained in all intermediate causal chains, search all standard causal chains containing these defect categories in the causal knowledge graph, traverse all nodes of these standard causal chains, filter out the nodes that belong to the entity node type, and obtain the set of all potential root cause entity nodes related to the current intermediate causal chain after deduplication.

[0135] In S173, before calculating the contribution of entity nodes, all entity nodes in the causal knowledge graph are first standardized and numerically mapped, converting them into equipment feature vectors that can be input into the model. The specific mapping rules are as follows: For discrete category entity nodes (such as corrosive media, Q245R low-strength material, unqualified welding process, and overdue inspection), a binary encoding method is used, with each entity node corresponding to one feature dimension. A value of 1 indicates that the equipment has the corresponding attribute and the potential cause exists, while a value of 0 indicates that it does not have it. For continuous threshold entity nodes (such as service life exceeding 10 years and insufficient design pressure), there are two feature dimensions: one is the normalized original continuous value (such as the min-max normalized value of service life and design pressure), and the other is a binary threshold label used to indicate whether the judgment condition of the entity node is met. The features of all entity nodes after mapping are concatenated in a fixed order to form a fixed-dimensional equipment feature vector, which is then input into the pre-trained XGBoost model for contribution calculation.

[0136] Based on this, the process of S173 is as follows: all entity nodes related to the intermediate causal chain are standardized numerically mapped based on the equipment features extracted from the report to obtain the numerical feature vectors corresponding to each entity node; among which the equipment features include attributes such as service life, material type, media corrosivity, and welding process, and after mapping, a feature vector with fixed dimension is formed, which is input into the XGBoost model pre-trained for the defect category of the downstream component of the intermediate causal chain.

[0137] The model traverses all its decision trees, outputting the original contribution value of each entity node based on the information gain generated when the tree splits according to the features of each entity node. Then, for each entity node, its original contribution values ​​across all decision trees are summed to obtain the first sum for that entity node. Simultaneously, the original contribution values ​​of all entity nodes across all decision trees are summed to obtain the second sum of all original contributions. Finally, the first sum of each entity node is divided by the second sum to obtain the normalized contribution of that entity node to the consequent defect category. This value ranges from 0 to 1; a higher value indicates a greater influence of the entity node on the occurrence of the defect.

[0138] The pre-training process of the XGBoost model in this embodiment is as follows: First, collect equipment feature data and corresponding defect detection results from historical special equipment inspection reports. For each predefined defect category, construct an independent binary classification training dataset. The input features are the numerical equipment features mapped from all potential root cause entity nodes (e.g., for the "insufficient wall thickness" defect, the input features include service life, media corrosivity, material type, etc.), and the label is whether the defect has occurred (1 for occurrence, 0 for non-occurrence). Divide the dataset into training and validation sets in an 8:2 ratio, initialize the XGBoost model, and set hyperparameters such as the number of decision trees to 100, the maximum depth to 6, and the learning rate to 0.1. Use binary cross-entropy as the loss function and gradient descent algorithm for training. At the same time, apply an early stopping mechanism on the validation set. Stop training when the validation set loss no longer decreases for 10 consecutive rounds to prevent overfitting. After training, the model learns the mapping relationship between the features of each entity node and the corresponding defect occurrence probability, and can output the contribution value of each entity node to the defect category during the inference stage.

[0139] The process of S174 is as follows: Sort all entity nodes by their contribution in descending order, select the entity nodes corresponding to the top P-ranked contribution as core root cause candidates, and then find the shortest directed transmission path from each core root cause entity node to the corresponding defect category in the intermediate causal chain in the causal knowledge graph. Use this path as the final root cause transmission chain for the defect category, and fully demonstrate the causal transmission process from the root cause entity to the occurrence of the defect.

[0140] For example, taking a pressure vessel report that detected the intermediate causal chain of "excessive media corrosion → insufficient wall thickness" as an example, the contribution of the four relevant entity nodes calculated by the XGBoost model is first sorted from largest to smallest. The results are: corrosive media 0.62, service life exceeding 10 years 0.28, Q245R low-strength material 0.07, and welding process non-compliance 0.03. Setting P=2, the top two "corrosive media" and "service life exceeding 10 years" are selected as core root cause candidates. Then, the shortest directed transmission path from these two entity nodes to the "insufficient wall thickness" defect node is searched in the causal knowledge graph. Two complete root cause transmission chains are obtained: "corrosive media → excessive media corrosion → insufficient wall thickness" and "service life exceeding 10 years → material aging → insufficient wall thickness". This shows the two main fundamental causes of the insufficient wall thickness defect of the pressure vessel and its causal transmission process.

[0141] This embodiment also includes S180, which can be specifically:

[0142] The transmission chain is input into the maintenance knowledge base. Based on the entity nodes and defect categories in the transmission chain, the maintenance knowledge base outputs several maintenance strategies and a difficulty score for each maintenance strategy. The maintenance strategies are related to the entity nodes. Based on the difficulty score and the contribution of the corresponding entity node, the priority score of the maintenance strategy is determined. The maintenance strategies are arranged in descending order according to the priority score to obtain improvement suggestions.

[0143] In this embodiment, the identified defect root cause transmission chain is input into the maintenance knowledge base. The maintenance knowledge base retrieves all executable maintenance strategies corresponding to each node by matching each entity node and defect node in the transmission chain, and outputs a standardized difficulty score (1-5 points, the higher the score, the greater the difficulty) for each maintenance strategy.

[0144] The priority scoring rules are as follows: for maintenance strategies of corresponding entity nodes, the contribution of that entity node is directly used in the calculation; for maintenance strategies of defective nodes, the maximum contribution of all directly upstream entity nodes in the causal knowledge graph is used. A unified formula is adopted: Priority Score = Corresponding Contribution × (1 - Difficulty Score / 5), which takes into account both the importance of the root cause and the feasibility of maintenance. Finally, all maintenance strategies are sorted in descending order of priority score, generating a structured list of improvement suggestions that includes priority, maintenance content, difficulty, and expected results.

[0145] For example, taking the previously obtained two transmission chains, "corrosive medium → excessive medium corrosion → insufficient wall thickness" and "service life exceeding 10 years → material aging → insufficient wall thickness," after inputting them into the maintenance knowledge base, the corresponding maintenance strategies and difficulty scores are retrieved: For the physical node "corrosive medium," "replace with 316L stainless steel" (difficulty 4 points) and "add corrosion inhibitor" (difficulty 2 points); For the defect node "excessive medium corrosion," "regularly test the medium composition" (difficulty 1 point), taking the contribution of the directly upstream physical node "corrosive medium" of 0.62 as the calculation basis; For the physical node "material aging," "partially replace the aged cylinder" (difficulty 3 points) and "replace the entire equipment" (difficulty 5 points).

[0146] The priority scores were calculated based on the corresponding contribution: "Adding corrosion inhibitor" was 0.62×(1-2 / 5)=0.372, "Regularly testing media composition" was 0.62×(1-1 / 5)=0.496, "Partially replacing the aged cylinder" was 0.28×(1-3 / 5)=0.112, "Replacing with 316L stainless steel" was 0.62×(1-4 / 5)=0.124, and "Completely replacing the equipment" was 0.28×(1-5 / 5)=0. After sorting by priority in descending order, the final improvement recommendations were: Regularly testing media composition, Adding corrosion inhibitor, Replacing with 316L stainless steel, and Partially replacing the aged cylinder.

[0147] In this embodiment, the process of establishing the maintenance knowledge base is as follows: First, the maintenance requirements in the "Special Equipment Safety Technical Specifications" are collected, along with a large amount of historical maintenance records accumulated in the industry, experience summaries from senior inspection engineers and maintenance experts, and advanced maintenance technology literature from home and abroad. Then, the collected information is structured to establish a triplet mapping relationship of "entity node - defect node - maintenance strategy," and each maintenance strategy is labeled with information such as applicable scenarios, operation steps, required resources, and expected effects. Next, industry experts are organized to score the difficulty of all maintenance strategies and formulate a unified scoring standard (1 point: no downtime required, can be completed by a single person; 5 points: requires downtime for major repairs, with the cooperation of a professional team). Finally, a dynamic update mechanism for the knowledge base is established, updating the maintenance strategies and difficulty scores quarterly based on new maintenance cases and technological developments to ensure the timeliness and accuracy of the knowledge base.

[0148] Based on this, maintenance strategies are specific and actionable technical and management measures formulated for each link in the defect root cause transmission chain. Their core objectives are to eliminate the root cause of the defect, block the causal transmission path, and prevent the defect from further deteriorating or recurring. Maintenance strategies can be divided into three categories: preventive maintenance strategies (such as regular inspection and adding corrosion inhibitors), used to eliminate potential risks before defects occur; corrective maintenance strategies (such as partial replacement of the cylinder and repair of welds), used to repair defects that have already occurred; and improvement maintenance strategies (such as replacing with corrosion-resistant materials and upgrading welding processes), used to fundamentally solve design or material defects and improve the long-term safety of the equipment.

[0149] This application provides an electronic device that may include a processing module and a memory. The memory stores a computer program, which, when executed by the processor, enables the electronic device to perform the corresponding steps in the aforementioned defect root cause identification method based on special equipment reports.

[0150] In this embodiment, the processor can be an integrated circuit chip with signal processing capabilities. For example, the processor can be a central processing unit (CPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0151] The memory can be, but is not limited to, random access memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, etc. In this embodiment, the memory can be used to store preset numbers, etc. Of course, the memory can also be used to store programs, which the processor executes after receiving an execution instruction.

[0152] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the electronic device described above can be referred to the corresponding steps in the aforementioned method, and will not be elaborated further here.

[0153] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the defect root cause identification method based on special equipment reports as described in the above embodiments.

[0154] Computer-readable storage media may be magnetic disks, optical disks, read-only memory, random access memory, flash memory, USB flash drives, hard disks, or solid-state drives, etc., and may also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implement the methods shown in the above embodiments.

[0155] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned defect root cause identification method based on special equipment reports. The computer program product may exist in a computer-readable storage medium in forms including, but not limited to, source files, executable files, and installation package files.

[0156] Based on the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, electronic device, or network device, etc.) to execute the methods described in the various implementation scenarios of this application.

[0157] In the embodiments provided in this application, it should be understood that the disclosed methods can also be implemented in other ways. The method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, program segment, or part of code, which includes one or more executable instructions for implementing a specified logical function. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0158] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for identifying the root causes of defects based on special equipment reports, characterized in that, The method includes: Obtain multiple special equipment reports; Based on the preset extraction tool, detection data of all modalities are extracted from each special equipment report. The detection data includes text data and physical feature data. Sensor data and structural data are defined as physical feature data, where sensor data is time-series quantitative detection data collected during the inspection process, and structural data is structured parameters of equipment attributes. According to the preset fusion strategy, the detection data of all the modalities are fused to obtain fused data; The fused data of all the special equipment reports are input into the defect detection model, and the defect detection model outputs the defect categories of all the special equipment reports. Based on the number of each defect category, all N-item sets are determined, wherein each N-item set includes N defect categories, and the number of special equipment reports corresponding to the N defect categories is greater than or equal to a preset number, where N is an integer greater than or equal to 2; Based on the N-item set, an initial causal chain is determined, wherein any defect category of the N-item set represents the antecedent of the initial causal chain, and other defect categories of the N-item set represent the consequent of the initial causal chain; Based on the pre-established causal knowledge graph, the initial causal chain, and the equipment characteristics corresponding to the detection data, the root causes of defects in special equipment are obtained; The causal knowledge graph includes multiple standard causal chains, each including entity nodes and defect nodes. Entity nodes are connected to each other, defect nodes are connected to each other, and entity nodes and defect nodes are connected by directed edges. Entity nodes are used to represent objective things that may cause defects, and defect nodes are used to represent the defect categories. The method of obtaining the root causes of defects in special equipment based on a pre-established causal knowledge graph, the initial causal chain, and the equipment features corresponding to the detection data includes: Based on the causal knowledge graph, the initial causal chain that does not conform to the standard causal chain is deleted to obtain the intermediate causal chain; Based on the causal knowledge graph, obtain the entity nodes in all standard causal chains that contain the defect categories involved in the intermediate causal chains; Based on the pre-trained XGBoost model, the contribution of each entity node to the defect category of the successor located in the intermediate causal chain is calculated according to the device characteristics. Sort all the contribution scores in descending order to obtain the entity nodes corresponding to the first P contribution scores, where P is an integer greater than or equal to 1; Based on the entity nodes corresponding to the first P contribution values ​​and the causal knowledge graph, the transmission chain from the entity node to the defect category in the intermediate causal chain is determined, and the transmission chain is used as the root cause of the defect in the special equipment. The pre-trained XGBoost model, based on the device features, calculates the contribution of each entity node to the defect category of the consequent located in the intermediate causal chain, including: All entity nodes related to the intermediate causal chain are standardized numerically mapped according to the equipment characteristics to obtain numerical feature vectors corresponding to all entity nodes. The equipment characteristics include service life, material type, media corrosivity, and welding process. All numerical feature vectors are input into the pre-trained XGBoost model to obtain the original contribution values ​​of each entity node output by each decision tree in the XGBoost model. For each entity node, its original contribution value in all decision trees is summed to obtain a first sum; the original contribution values ​​of all entity nodes in all decision trees are summed to obtain a second sum. The contribution of the entity node is obtained by dividing the first sum by the second sum.

2. The method according to claim 1, characterized in that, The step of fusing the detection data of all modalities according to a preset fusion strategy to obtain fused data includes: The text data is input into the transformation model, which adds multi-head self-attention to the text data to obtain a global text feature matrix; The physical feature data is projected onto the semantic space of the text data to obtain a physical feature matrix with the same dimension as the global text feature matrix; The global text feature matrix is ​​subjected to global average pooling along the sequence dimension to obtain a global text semantic vector. The inner product of the feature vector at each position in the physical feature matrix and the global text semantic vector is calculated to obtain the similarity weight. The similarity weight is used to characterize the degree of correlation between physical feature data and text data. The physical feature matrix is ​​weighted by the similarity weights to obtain a cross-modal feature matrix that represents the matching part between text data and physical feature data; The cross-modal feature matrix is ​​input into the sigmoid gating function, and the gating weights are obtained through the sigmoid gating function. The fused data is obtained based on the gating weights and the cross-modal feature matrix.

3. The method according to claim 2, characterized in that, The text data is input into a transformation model, which adds multi-head self-attention to the text data to obtain a global text feature matrix, including: The text data is input into a conversion model, which determines the semantic features of each character in the text data in a specified dimension based on a pre-trained semantic sub-model. The transformation model matches the corresponding tokens for each character based on the semantics of each clause in the text data to obtain a tag sequence. There are M types of tokens, and the semantic rules corresponding to each token are pre-stored in the transformation model. M is an integer greater than 1, and each clause includes several characters. The conversion model uses the tag sequence as an index and extracts and divides the semantic features of all characters according to the type of the token to obtain M first sub-matrices. Each first sub-matrice contains the semantic features of the characters corresponding to the token type. Each of the first submatrixes is split into Q second submatrixes along the feature dimension; The Q second submatrices corresponding to each first submatrix are respectively input into the analysis head corresponding to the multi-head self-attention. The similarity of each second submatrix within the same first submatrix with respect to other second submatrices is obtained through the analysis head. The analysis head is pre-deployed in the transformation model. Based on the similarity between each second submatrix within the same first submatrix and other second submatrixes, the attention weight of each second submatrix is ​​obtained; Based on the attention weights, the analysis head weights all the second sub-matrices within the corresponding range to obtain the weighted matrix corresponding to each analysis head. The transformation model first concatenates all weighted matrices corresponding to each first sub-matrix to obtain attention-enhanced feature matrices for each token type; The global text feature matrix is ​​obtained by backfilling and concatenating all attention-enhanced feature matrices of all token types according to the original character position order.

4. The method according to claim 1, characterized in that, The determination of all N itemsets based on the number of each defect category includes: S151: Based on the number of each defect category, determine the number of special equipment reports containing the defect categories; S152: Determine the frequency of the defect category based on the quotient of the number of special equipment reports containing the defect category divided by the total number of special equipment reports; S153: Filter out all defect categories with a frequency greater than or equal to the preset frequency threshold; S154: Select any two different defect categories from all defect categories whose frequency is greater than or equal to the preset frequency threshold, and generate a second-order candidate defect combination according to the predefined sorting rules; S155: For all k-order candidate defect combinations, count the number of special equipment reports that simultaneously include all the defect categories within the k-order candidate defect combination, and obtain the frequency of the k-order candidate defect combination based on the total number of special equipment reports. S156: Determine whether the iteration condition is met. The iteration condition is: among all k-order candidate defect combinations, there exists a k-order defect combination with a frequency greater than or equal to a preset frequency threshold, and among all k-order defect combinations with a frequency greater than or equal to the preset frequency threshold, there exists at least one pair of k-order defect combinations whose first k-1 defect categories are completely identical. If the iteration condition is met, then the pair of k-order defect combinations are concatenated and deduplicated, and arranged according to a predefined sorting rule to generate a k+1-order candidate defect combination. S157: When the iteration condition is met, assign k+1 to k and execute S155-S156 again. When the iteration condition is not met, the k-order candidate defect combination is the N-item set, k=N, where k is an integer greater than or equal to 2, and the initial value of k is 2.

5. The method according to claim 1, characterized in that, After determining the transmission chain from the entity node to the defect category in the intermediate causal chain based on the entity nodes corresponding to the first P contributions and the causal knowledge graph, the method further includes: The transmission chain is input into the maintenance knowledge base, which outputs several maintenance strategies and a difficulty score for each maintenance strategy based on the entity nodes and defect categories in the transmission chain. The maintenance strategies are related to the entity nodes. Based on the difficulty score and the contribution of the corresponding entity node, the priority score of the maintenance strategy is determined. The maintenance strategies are sorted in descending order of priority scores to obtain improvement suggestions.

6. An electronic device, characterized in that, The electronic device includes a processor and a memory coupled together, the memory storing a computer program that, when executed by the processor, causes the electronic device to perform the method as described in any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the method as described in any one of claims 1 to 5.

8. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 5.

Images