Remote-controlled pipeline weld flaw detection method for narrow space of water purification plant

Abstract

The present application relates to the technical field of pipeline weld flaw detection, and discloses a remote control water purification plant narrow space pipeline weld flaw detection method. The method comprises the following steps: acquiring, through a distributed sensor network, real-time pipeline weld flaw detection data streams such as ultrasonic echo signal sequences, and environmental disturbance data streams such as equipment vibration frequency spectrum and temperature gradient data. According to a predefined defect identification specification, the ultrasonic echo signal sequences are subjected to multi-scale feature analysis, and weld defect distribution feature information is extracted. Based on an embedded flaw detection decision graph and performance compliance conditions, basic flaw detection decisions are generated by fusing the defect distribution feature information and the environmental disturbance data streams. Defect evolution and environmental disturbance prediction data are obtained through adaptive prediction calculation, based on which the basic decisions are corrected and final results are output, driving remote equipment to complete automatic flaw detection. The method improves defect identification accuracy and decision reliability, and adapts to narrow space operation requirements.

Description

Technical Field

[0001] This invention relates to the field of pipeline weld flaw detection technology, specifically a remotely controlled method for flaw detection of pipeline welds in confined spaces of a water purification plant. Background Technology

[0002] Non-destructive testing (NDT) of weld seams in confined spaces of water treatment plants is a crucial step in ensuring the safety of water supply systems, and ultrasonic testing has become the mainstream technology due to its high detection accuracy. However, current ultrasonic testing methods often rely heavily on operator experience in processing weld echo signals, lacking standardized predefined defect identification protocols. Signal analysis frequently employs a single-scale approach, focusing only on basic parameters such as echo amplitude and propagation time, without addressing the hierarchical differences in the characteristics of different defect types. In the decision-making stage, traditional techniques often rely on manual comparison with standard spectra to make judgments, failing to construct a built-in flaw detection decision map or incorporate environmental disturbances and performance compliance conditions into a comprehensive consideration, outputting conclusions based solely on single flaw detection data.

[0003] These technologies present significant challenges in practical applications. The lack of standardized guidance and multi-scale perspectives in signal analysis leads to the masking of complex defects such as microcracks and irregular slag inclusions, resulting in incomplete extraction of weld defect distribution characteristics and a high risk of missed or misjudged detections. During decision-making, manual operation is inefficient and susceptible to experience-based biases, leading to inconsistent results. Ignoring environmental disturbances on flaw detection data and failing to establish judgment criteria based on performance compliance conditions results in decisions that are out of sync with actual pipeline operation requirements, making it difficult to meet the accuracy and reliability requirements of water treatment plants for weld flaw detection. In confined space environments, these problems are amplified, increasing inspection costs and failing to provide effective guarantees for pipeline safety.

[0004] To address the aforementioned issues, a technical solution is needed that uses clear specifications to guide refined signal analysis and enables systematic decision-making based on standardized spectra and compliance conditions. This solution aims to resolve the problems of lacking unified standards for defect analysis and lacking systematic basis for decision-making processes in traditional technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a remotely controlled method for inspecting weld seams in confined spaces of water purification plants, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a remotely controlled method for inspecting weld seams in confined spaces of a water purification plant, the method comprising:

[0007] The flaw detection data stream and environmental disturbance data stream of the pipeline weld are captured in real time through a distributed sensor network. The flaw detection data stream includes ultrasonic echo signal sequences, and the environmental disturbance data stream includes equipment vibration spectrum and temperature gradient data.

[0008] According to the predefined defect identification specifications, the ultrasonic echo signal sequence is subjected to multi-scale feature analysis to extract the weld defect distribution feature information.

[0009] Based on the built-in flaw detection decision map and performance compliance conditions, the basic flaw detection decision is generated by integrating the weld defect distribution characteristics information and the environmental disturbance data stream.

[0010] Adaptive prediction calculations are performed on the weld defect distribution characteristics and the environmental disturbance data stream to obtain defect evolution characteristic data and environmental disturbance prediction data;

[0011] The basic flaw detection decision is corrected in real time using the defect evolution characteristic data and environmental disturbance prediction data, and the final flaw detection decision is output.

[0012] The final flaw detection decision drives the remote flaw detection equipment to complete the automated flaw detection process of pipeline welds.

[0013] Preferably, the step of performing multi-scale feature analysis on the ultrasonic echo signal sequence according to a predefined defect identification specification to extract weld defect distribution feature information includes:

[0014] Obtain a set of geometric parameters for the pipe weld, the set of geometric parameters including weld width, depth and radius of curvature;

[0015] A digital twin model of the weld is constructed based on the set of geometric parameters.

[0016] The ultrasonic echo signal sequence is input into the weld digital twin model to simulate and generate a three-dimensional cloud map of internal defects in the weld.

[0017] Based on the defect size threshold and shape rules in the predefined defect identification specification, pattern recognition is performed on the three-dimensional cloud map of the internal defects of the weld, the defect area is segmented and the defect density distribution is calculated, thereby obtaining the defect distribution feature information of the weld.

[0018] Preferably, the method further includes a preprocessing stage for constructing a defect feature analysis model, including:

[0019] For the aforementioned pipe weld, a set of welds of the same material is selected as a reference set, which includes multiple historical test weld samples;

[0020] Load a reference defect feature database from the reference set, the reference defect feature database storing defect echo data and defect classification labels of reference welds;

[0021] For the pipe weld currently being tested, load its dedicated defect feature record library;

[0022] A deep learning model is trained using the aforementioned reference defect feature database to establish an initial defect feature parsing model.

[0023] The initial defect feature parsing model is optimized by transfer learning using the dedicated defect feature record library to generate a high-precision defect feature parsing model for multi-scale feature parsing.

[0024] Preferably, the step of using the dedicated defect feature record library to perform transfer learning optimization on the initial defect feature parsing model to generate a high-precision defect feature parsing model includes:

[0025] The initial defect feature parsing model is cross-validated using the dedicated defect feature record library, and the model generalization error index is calculated.

[0026] Compare the generalization error index of the model with the upper limit of the allowable error;

[0027] When the model generalization error index exceeds the upper limit of the allowable error, the training data is enhanced by an adversarial generative network, and the model parameters are retrained until the model generalization error index drops to the allowable range, thereby obtaining the high-precision defect feature analysis model.

[0028] Preferably, the step of generating basic flaw detection decisions based on the built-in flaw detection decision map and performance compliance conditions, by fusing the weld defect distribution characteristic information and the environmental disturbance data stream, includes:

[0029] Based on the weld defect distribution characteristics and the environmental disturbance data stream, the flaw detection decision map is correlated and matched to obtain a matching decision subset.

[0030] The risk level of the matching decision subset is divided to identify safe decision intervals;

[0031] Generate candidate flaw detection strategies within the stated safety decision interval;

[0032] The consistency between the candidate flaw detection strategy and the performance compliance conditions is evaluated to obtain a strategy compliance score;

[0033] Determine whether the compliance score of the strategy meets the preset compliance standard;

[0034] When the compliance score of the strategy reaches the preset compliance standard, the candidate flaw detection strategy will be incorporated into the basic flaw detection decision.

[0035] Preferably, the step of performing correlation matching on the flaw detection decision map based on the weld defect distribution characteristic information and the environmental disturbance data stream to obtain a matching decision subset includes:

[0036] Traverse each decision record in the flaw detection decision map. Each decision record includes sample weld defect distribution characteristics, sample environmental disturbance data stream, and sample flaw detection strategy.

[0037] Calculate the feature similarity matrix between the weld defect distribution feature information and the sample weld defect distribution feature information;

[0038] Calculate the dynamic time-warped distance between the environmental disturbance data stream and the sample environmental disturbance data stream;

[0039] The feature similarity matrix and the dynamic time-normalized distance are combined using a weighted fusion algorithm to generate a comprehensive correlation value.

[0040] Determine whether the overall correlation value is greater than or equal to the correlation threshold;

[0041] When the comprehensive correlation value is greater than or equal to the correlation threshold, the sample flaw detection strategy is added to the matching decision subset.

[0042] Preferably, the adaptive prediction calculation of the weld defect distribution characteristic information and the environmental disturbance data stream to obtain defect evolution characteristic data and environmental disturbance prediction data includes:

[0043] A long short-term memory network model is applied to perform time series prediction on the weld defect distribution characteristics to generate defect evolution characteristic data.

[0044] The environmental disturbance data stream is state estimated using a Kalman filter to generate environmental disturbance prediction data.

[0045] Preferably, the step of using the defect evolution characteristic data and environmental disturbance prediction data to perform real-time correction on the basic flaw detection decision and output the final flaw detection decision includes:

[0046] The flaw detection decision map is dynamically updated based on the defect evolution characteristic data and environmental disturbance prediction data to obtain an optimized decision map.

[0047] The highest priority flaw detection strategy is extracted from the optimization decision graph and used as the core strategy.

[0048] The core strategy and the basic flaw detection decision are integrated and calculated, and a weighted average algorithm is used to generate transitional decisions;

[0049] The feasibility of the transitional decision is verified, including resource constraint checks and timing consistency checks. Once the checks are passed, the final flaw detection decision is output.

[0050] Preferably, the real-time capture of the flaw detection data stream and environmental disturbance data stream of the pipeline weld through a distributed sensor network includes:

[0051] An array of ultrasonic probes is deployed to scan along the pipe weld seam and collect ultrasonic echo signal sequences.

[0052] Install an infrared thermal imager to monitor the temperature change of the weld surface and generate temperature gradient data;

[0053] The vibration frequency of the equipment is measured using an accelerometer to generate a vibration spectrum.

[0054] The ultrasonic echo signal sequence, temperature gradient data, and vibration spectrum synchronization timestamp are integrated into a unified data stream.

[0055] Preferably, the automated flaw detection process of the pipeline weld seam driven by the final flaw detection decision to the remote flaw detection equipment includes:

[0056] The equipment control parameters in the final flaw detection decision are analyzed, including the probe movement path and signal transmission strength;

[0057] The device control parameters are sent to the controller of the remote flaw detection equipment via a wireless communication module.

[0058] The controller adjusts the robotic arm posture and sensor settings of the flaw detection equipment to perform the scanning operation;

[0059] The flaw detection data is received in real time and fed back to the decision-making system to form a closed-loop control.

[0060] Compared with the prior art, the beneficial effects of the present invention are:

[0061] Multi-scale feature analysis is performed on ultrasonic echo signal sequences based on predefined defect identification specifications. Standardized specifications clarify the feature extraction dimensions and judgment criteria for different defect types, avoiding the chaotic analysis standards caused by traditional reliance on manual experience. The multi-scale analysis mode can extract data value from multiple levels, including microscopic signal fluctuations, mesoscopic waveform morphology, and macroscopic propagation patterns. It can capture weak signal distortions caused by microcracks and distinguish the characteristic differences between deep slag inclusions and surface porosity, making the extracted weld defect distribution feature information more complete and targeted. This standardized, multi-dimensional analysis method filters out invalid interference components in the signal, allowing for accurate presentation of key information such as defect location, size, and morphology. This significantly reduces the risk of missed detections and misjudgments caused by insufficient feature extraction, making it particularly suitable for detection scenarios with complex weld defect types in the confined spaces of water treatment plants.

[0062] Based on a built-in flaw detection decision map and performance compliance conditions, this system integrates weld defect distribution characteristics with environmental disturbance data streams to generate basic flaw detection decisions. By embedding standardized decision logic and compliance requirements, it replaces the traditional manual comparison of maps, eliminating excessive reliance on operator expertise. Environmental data such as equipment vibration spectrum and temperature gradient are simultaneously incorporated into the decision-making process, quantifying the impact of environmental factors on flaw detection results and avoiding biased decisions caused by solely relying on flaw detection data. The built-in map provides a unified reference standard for defect judgment, while performance compliance conditions ensure that the decision results align with the actual operational safety requirements of the pipeline. The combination of these two aspects makes the basic flaw detection decisions both accurate and practical, improving decision generation efficiency and ensuring consistency and reliability of decision results across different inspection scenarios, providing stable core support for remote automated flaw detection. Attached Figure Description

[0063] Figure 1 This is a schematic diagram illustrating the working principle of the remotely controlled pipe weld flaw detection method in a confined space of a water purification plant as described in this invention.

[0064] Figure 2 This is a flowchart of multi-scale feature analysis and defect distribution feature extraction;

[0065] Figure 3 Comparison chart of performance analysis models for identifying defects in pipeline welds;

[0066] Figure 4 A flowchart for basic flaw detection decision generation and compliance assessment;

[0067] Figure 5 Vibration spectrum analysis diagram of pipeline weld flaw detection equipment. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Please see Figure 1This invention provides a remotely controlled method for flaw detection of pipe welds in confined spaces within a water treatment plant. The method includes: a distributed sensor network deployed around the pipe weld in the water treatment plant, comprising multiple ultrasonic probes, infrared thermal imagers, and accelerometers, for real-time capture of flaw detection data streams and environmental disturbance data streams from the pipe weld; wherein the flaw detection data stream mainly consists of ultrasonic echo signal sequences, and the environmental disturbance data stream includes equipment vibration spectrum and temperature gradient data. The captured data streams are transmitted wirelessly to a central processing unit for initial timestamp synchronization and format standardization. The system performs multi-scale feature analysis on the ultrasonic echo signal sequences according to predefined defect identification specifications. This process involves signal decomposition and feature extraction to identify defect patterns within the weld and output weld defect distribution characteristic information, including defect size, location, and density indices. The system integrates weld defect distribution characteristics with environmental disturbance data streams based on a built-in flaw detection decision map and performance compliance conditions. The flaw detection decision map is a pre-built database containing strategy mappings of historical flaw detection cases, while the performance compliance conditions define safety thresholds and operational standards. Basic flaw detection decisions are generated through matching and evaluation. The system performs adaptive predictive calculations on the weld defect distribution characteristics and environmental disturbance data streams, employing time series analysis algorithms to infer defect evolution trends and environmental changes, generating defect evolution characteristic data and environmental disturbance prediction data. The predicted data is used to correct the basic flaw detection decisions in real time. The correction process involves dynamic weight adjustment and feasibility verification, outputting a final flaw detection decision that encompasses equipment control parameters and scanning paths. This final flaw detection decision drives remote flaw detection equipment, such as robotic arms and sensor arrays, via a communication interface to automate the scanning and inspection of pipeline welds, forming a closed-loop control system to ensure the accuracy and real-time performance of flaw detection.

[0070] Example 1: See Figure 2 In practical implementation, obtaining the set of geometric structural parameters of the pipeline weld is the initial step. This set of parameters is obtained through non-contact measurement of the target pipeline weld area using a high-precision 3D laser scanner. The measurement output includes precise values ​​for weld width, weld depth, and radius of curvature. These values ​​are recorded as a structured data list and stored in the system cache. In practical implementation, a digital twin model of the weld is constructed based on the set of geometric structural parameters. The construction process uses a professional 3D modeling software interface, inputting the set of geometric structural parameters into a parametric modeling engine. The engine automatically generates a corresponding 3D mesh model based on the weld width, depth, and radius of curvature. The mesh model includes vertex coordinates, facet connectivity, and surface normal vector information. The generated digital twin model of the weld is saved in a standard file format and loaded into memory for subsequent calculations.

[0071] In practical implementation, the ultrasonic echo signal sequence is input into the digital twin model of the weld and a 3D cloud map of internal weld defects is generated. This process is completed by a finite element analysis solver. The solver loads the ultrasonic echo signal sequence as boundary conditions onto the digital twin model of the weld and calculates the propagation, reflection, and attenuation of ultrasonic waves within the model by solving the elastic wave equation. The simulation results are output in the form of voxel data, where each voxel stores sound pressure or displacement values. Finally, volume rendering technology is used to convert the voxel data into a visualized 3D cloud map of internal weld defects. In practical implementation, pattern recognition is performed on the 3D cloud map of internal weld defects based on the defect size threshold and shape rules in the predefined defect identification specifications. The predefined defect identification specifications are stored in the form of configuration files. The defect size threshold specifies the critical values ​​for defect volume and diameter, and the shape rules define the geometric morphological characteristics of defects such as roundness, aspect ratio, and convexity. The pattern recognition algorithm traverses each connected region in the 3D cloud map of internal weld defects, calculates the geometric attributes of each region, and matches them with the defect size threshold and shape rules. Regions that conform to the rules are marked as defect regions.

[0072] In specific implementation, defect regions are segmented and defect density distribution is calculated to obtain weld defect distribution feature information. The segmentation operation uses a region growing algorithm to extract continuous defect voxel clusters from the three-dimensional cloud map of defects inside the weld. For each segmented defect region, its geometric features such as volume, centroid coordinates, equivalent diameter, and orientation are calculated. The defect density distribution is calculated by statistically analyzing the number of defect regions per unit volume of weld and the percentage of the total volume. Finally, the weld defect distribution feature information is organized into a data matrix, where each row of the matrix represents the feature vector of a defect region. The entire matrix, along with the defect density distribution value, is used as the output. In some embodiments, the method also includes a preprocessing stage for constructing a defect feature analysis model. In the preprocessing stage, a set of welds of the same material is selected as a reference set for the pipeline weld. The selection process is completed by querying a historical inspection database. The database is filtered according to material grade, welding process, and service environment. The reference set contains multiple historical test weld samples, and each historical test weld sample is associated with a complete inspection file. In some embodiments, a reference defect feature database is loaded from a reference set. The reference defect feature database is a relational database table that stores defect echo data and defect classification labels for reference welds. The defect echo data is a feature vector of the original ultrasonic signal after preprocessing. The defect classification labels are defect type identifiers such as porosity, slag inclusion, or lack of fusion, which are manually labeled or identified by authoritative equipment.

[0073] In practice, a dedicated defect feature record library is loaded for each pipe weld currently being tested. This library, a data file stored in the local file system, records defect feature data and timestamps collected from all previous inspection cycles for that specific pipe weld. The loading process is executed through a database query interface, using the unique identifier of the current pipe weld as the query condition to retrieve relevant data from the archived records. In the implementation, a deep learning model is trained using a reference defect feature database to establish an initial defect feature parsing model. The deep learning model employs a convolutional neural network architecture. The training process uses defect echo data from the reference defect feature database as input and defect classification labels as supervision signals. The network weights are optimized through a backpropagation algorithm. The trained model can extract high-level features from new ultrasonic echo signal sequences and preliminarily determine the defect category. Optionally, a dedicated defect feature record library is used to optimize the initial defect feature parsing model through transfer learning. This optimization retains the convolutional layer structure of the initial model, replacing and retraining only the fully connected classification layer. Mini-batch gradient descent training is performed using data from the dedicated defect feature record library. The optimization process focuses on adjusting model parameters to better fit the unique defect feature patterns of the current pipeline weld. Alternatively, transfer learning optimization generates a high-precision defect feature parsing model for multi-scale feature parsing. This model is integrated into a real-time processing pipeline. The model receives real-time acquired ultrasonic echo signal sequences and outputs a calibrated defect probability distribution and enhanced feature maps. These outputs serve as direct inputs to the multi-scale feature parsing module.

[0074] It is understandable that the multi-scale feature analysis process relies on the collaborative work of the weld digital twin model and the defect feature analysis model. The weld digital twin model provides the physical background and geometric constraints, while the defect feature analysis model provides signal interpretation and pattern recognition capabilities. It is also understandable that the high-precision defect feature analysis model built in the preprocessing stage significantly improves the accuracy of subsequent feature analysis. By integrating historical experience and specific weld-specific data, the model achieves robust identification of complex defect patterns. In practice, the entire process is automated. From acquiring geometric parameters to outputting weld defect distribution feature information, no manual intervention is required. The system schedules various software modules and hardware resources through pre-written scripts and a workflow engine. In practice, the weld defect distribution feature information is ultimately encapsulated into a specific data structure and transmitted to the decision fusion module via inter-process communication, providing input for generating basic flaw detection decisions. In practice, all computational tasks can run on industrial computers or embedded edge computing devices. The specific deployment method is configured according to the computing resource conditions and real-time requirements of the water treatment plant site. In practical implementation, the accuracy of the weld digital twin model directly depends on the measurement accuracy of the geometric parameter set. Therefore, a strategy of minimizing measurement error is adopted by performing high-resolution laser scanning and averaging multiple measurements. In practical implementation, the parameters of the defect identification specification can be dynamically adjusted according to different standards such as ISO 5817 or ASME BPVC. The system provides a graphical interface for engineers to modify defect size thresholds and shape rules. In practical implementation, the reference defect feature database is updated regularly, incorporating newly discovered defect cases and classification results, thus enabling the initial defect feature analysis model to continuously evolve. In practical implementation, the frequency of transfer learning optimization can be set according to the inspection history and data accumulation of the pipeline weld, for example, triggering an optimization cycle every ten inspections or when the data volume of the dedicated defect feature record library doubles.

[0075] Example 2: In specific implementation, the first step is to cross-validate the initial defect feature parsing model using a dedicated defect feature record library. The cross-validation adopts the k-fold cross-validation method, and the value of k is usually set to 5 or 10. The system randomly and uniformly divides the defect echo data and defect classification label dataset contained in the dedicated defect feature record library into k mutually exclusive subsets. In each iteration, one subset is selected as the test set, and the remaining k-1 subsets are used as the training set. The initial defect feature parsing model performs forward and backward propagation on the training set to calculate the loss function, and makes predictions on the test set. The prediction results are compared with the real defect classification labels, and the prediction error of each iteration is calculated. Finally, the error results of k iterations are combined to generate the model generalization error index. The model generalization error index includes quantitative evaluation values ​​of multiple dimensions such as classification accuracy, precision, recall, and F1 score. In practice, comparing the model generalization error index with the upper limit of the allowable error is a key subsequent judgment. The upper limit of the allowable error is preset by the system administrator according to the actual quality requirements and industry standards of pipeline weld inspection and stored in the system configuration file. For example, if the upper limit of the allowable error of the classification accuracy is set to 95%, the system will automatically compare the corresponding value in the calculated model generalization error index with the upper limit of the allowable error and generate a Boolean logic judgment result.

[0076] In practice, when the model's generalization error exceeds the allowable error limit, the system triggers an optimization process. This process uses an adversarial generative network (GAN) to enhance the training data. The GAN consists of a generator network and a discriminator network. The generator network takes a random noise vector and a latent feature vector extracted from a dedicated defect feature record library as input. It generates synthetic defect echo data through deconvolutional layers and upsampling layers. The discriminator network receives real defect echo data and the synthetic defect echo data generated by the generator network. It judges the authenticity of the input data through convolutional layers and downsampling layers. The generator network and the discriminator network compete against each other during training. The generator network aims to generate synthetic data that is sufficient to deceive the discriminator network, while the discriminator network aims to accurately distinguish between real and synthetic data. Through this adversarial training process, the generator network can eventually generate high-quality synthetic data that closely resembles the distribution of real defect echo data. This synthetic data is then merged with the original real data in the dedicated defect feature record library to form the enhanced training dataset. In practice, retraining the model parameters using the enhanced training data is the core optimization step. The retraining process does not start from scratch with the initial defect feature parsing model; instead, it employs a fine-tuning strategy from transfer learning. The convolutional layer weights, trained on the reference defect feature database, remain largely unchanged, with only the weights of the fully connected and classification layers being adjusted. The training algorithm uses stochastic gradient descent or Adam optimization, with the cross-entropy loss function as the optimization objective. Multiple rounds of iterative training are performed on the enhanced training dataset until the loss function converges or reaches the preset maximum number of iterations. In practice, the retraining process involves repeated rounds of model generalization error metric evaluation. After each round of retraining, the system immediately recalculates the model generalization error metric using the current model parameters on the same test set and compares it again with the allowable error upper limit, forming a closed-loop feedback mechanism of training-evaluation-comparison. The iteration process terminates only when the model generalization error metric falls within the allowable range.

[0077] In some embodiments, the calculation of the model generalization error metric may involve more complex ensemble evaluation strategies. For example, the system may calculate not only the average error under k-fold cross-validation but also the standard deviation of the error to assess the model's stability. In some embodiments, the allowable error upper limit may not be a fixed value but a dynamically adjusted threshold that can be adaptively adjusted based on historical optimization results or real-time monitoring data of pipe welds. Optionally, the structure of the adversarial generative network can be customized based on the data characteristics of a dedicated defect feature record library. For example, when defect echo data has obvious time-series characteristics, the generator network and discriminator network can use long short-term memory networks or gated recurrent unit structures instead of traditional convolutional structures to better capture temporal dependencies in the data. Optionally, the fine-tuning strategy during the retraining of model parameters can have different granularities. For example, the system allows administrators to select, through configuration files, only the weights of the last few layers of the network to be fine-tuned, or to set different learning rates for all layers in the network, using smaller learning rates for layers closer to the input layer to retain general features and larger learning rates for layers closer to the output layer to quickly adapt to new data.

[0078] It is understandable that the entire transfer learning optimization process is a highly automated iterative loop. The system continuously monitors the relationship between the model's generalization error metric and the upper limit of the allowable error, autonomously deciding whether to initiate data augmentation and model retraining. It is also understandable that the application of generative adversarial networks effectively solves the problem of potentially insufficient data in the dedicated defect feature record library, improving the robustness and generalization ability of the high-precision defect feature analysis model through data augmentation. In specific implementation, the final high-precision defect feature analysis model is saved in a serialized file format and integrated into the real-time flaw detection data processing pipeline, directly used for multi-scale feature analysis of ultrasonic echo signal sequences. The model loading and inference process is implemented through an optimized inference engine to meet real-time requirements. In specific implementation, all intermediate data during the optimization process, such as the model parameters for each iteration, the augmented training dataset, and the historical records of the model generalization error metric for each evaluation, are recorded in detail by the system log module for subsequent auditing and performance analysis. In specific implementation, training the generative adversarial network may require significant computational resources; the system supports distributing the generative adversarial network training task to GPU clusters or cloud computing platforms for accelerated computation. In practice, the criteria for determining whether the model generalization error index has fallen to the allowable range are strict. Optimization is considered successful only if all preset error indices (such as accuracy, precision, and recall) are simultaneously below their respective allowable error limits. If, after multiple rounds of iterative optimization, the model generalization error index still cannot be reduced to the allowable range, the system will trigger an alarm mechanism to notify the administrator for manual intervention and inspection. In practice, the frequency of transfer learning optimization can be set to either periodic triggering or on-demand triggering based on the service life of the pipeline weld, the inspection frequency, and the accumulation of historical data.

[0079] See Figure 3This figure focuses on the performance of defect feature analysis models in pipeline weld inspection, comparing the differences in accuracy, recall, and F1 score among three types of models: the initial model, the transfer learning optimized model, and the adversarial generative network (GCN) enhanced model. The figure shows that the initial model's performance is relatively weak, while after transfer learning optimization, all indicators of the model are significantly improved, and the model with GCN enhancement achieves optimal performance. This figure aligns with the technical logic: in the defect feature analysis stage of pipeline weld inspection, an initial deep learning model is first trained using a reference defect feature database, then optimized using a dedicated defect feature record library through transfer learning. When the model's generalization error exceeds the limit, the training data is enhanced with GCN and retrained, ultimately resulting in a high-precision defect feature analysis model. This progressive improvement in model performance directly ensures the accuracy of multi-scale feature analysis of weld defects, providing reliable support for subsequent defect feature-based inspection decisions. It effectively solves the problems of missed detections and misjudgments caused by insufficient model accuracy in traditional technologies, and is particularly suitable for the complex defect types in pipeline welds within the confined spaces of water treatment plants.

[0080] Example 3: See Figure 4 In practical implementation, the initial step is to perform correlation matching on the flaw detection decision map based on the weld defect distribution characteristics and environmental disturbance data stream. The flaw detection decision map is a two-dimensional data table stored in a relational database. Each row of the data table represents a decision record, and each decision record contains three main fields: sample weld defect distribution characteristics, sample environmental disturbance data stream, and sample flaw detection strategy. The system traverses each decision record in the flaw detection decision map. For the decision record currently being processed, it calculates the feature similarity matrix between the weld defect distribution characteristics and the sample weld defect distribution characteristics in the current decision record. The feature similarity matrix is ​​obtained by calculating the cosine similarity between two feature vectors. The calculation formula is as follows:

[0081] ;

[0082] in: This represents a vector of weld defect distribution characteristics acquired in real time. This represents the vector of sample weld defect distribution characteristics in the current decision record of the flaw detection decision map. This is the calculated cosine similarity value, ranging from -1 to 1. A value closer to 1 indicates greater feature similarity. In practice, calculating the dynamic time warping distance between the environmental disturbance data stream and the sample environmental disturbance data stream is a parallel computational task. The dynamic time warping algorithm is used to align two time series data that may have length differences or local deformations. The algorithm constructs a cost matrix to find the optimal curved path between the two time series and calculates the cumulative minimum cost along this path as the dynamic time warping distance. The smaller this distance value, the closer the two environmental disturbance data streams are in shape over time.

[0083] In practice, the key data fusion step is to combine the feature similarity matrix and the dynamic time-normalized distance using a weighted fusion algorithm to generate a comprehensive correlation value. The weighted fusion algorithm assigns a weight coefficient to the feature similarity matrix. Assign a weight coefficient to the dynamic time warp distance. Weighting coefficient and weighting coefficients satisfy The constraints, the specific values ​​of which are pre-set by the system administrator based on historical matching results or domain knowledge and stored in the configuration file; the comprehensive relevance value. The calculation method is to use the feature similarity matrix values With weighting coefficients Multiply to normalize the distance value over time. After normalization and weighting coefficients Multiply them and sum the two products, that is... ,in This is a preset maximum distance constant used for distance normalization. In practice, determining whether the comprehensive correlation value is greater than or equal to the correlation threshold is the core of the screening logic. The correlation threshold is a floating-point number between 0 and 1, set by the system according to the decision accuracy requirements. The system will calculate the comprehensive correlation value... With correlation threshold When comparing, the overall correlation value Greater than or equal to the correlation threshold When the decision record being traversed is determined to be highly correlated with the real-time data, the sample flaw detection strategy stored in the decision record is added to the matching decision subset. The matching decision subset is a list structure that is dynamically maintained in memory and is used to temporarily store all candidate strategies that have been filtered out through correlation matching.

[0084] In some embodiments, risk level classification of the matching decision subset may employ a rule-based approach. The system assigns risk weights to each defect feature in the weld defect distribution feature information and each disturbance index in the environmental disturbance data stream. A comprehensive risk score is calculated for each matching decision record through weighted summation, and then the risk level is classified according to the preset interval in which the comprehensive risk score falls. In some embodiments, identifying the safe decision interval may involve more complex multi-objective optimization. The safe decision interval not only requires a low comprehensive risk score but may also require short strategy execution time or low resource consumption. The system uses Pareto front analysis to find the set of non-dominated solutions that satisfy multiple constraints as the safe decision interval. In some embodiments, generating candidate flaw detection strategies within the safe decision interval can be a combinatorial optimization process. The system extracts the basic operational units of the sample flaw detection strategies from multiple decision records contained in the safe decision interval, and then recombines and optimizes the parameters of these basic operational units using a genetic algorithm or simulated annealing algorithm to generate a series of new candidate flaw detection strategies adapted to the current specific situation.

[0085] In practice, assessing the consistency between candidate flaw detection strategies and performance compliance conditions to obtain a strategy compliance score is a rigorous verification step. Performance compliance conditions are a set of predefined business rules and physical constraints, which may include maximum allowable scanning time, minimum detection accuracy, equipment power limits, and safety regulatory requirements. The assessment process is executed through a rule engine. The rule engine substitutes the parameters of each candidate flaw detection strategy into each rule of the performance compliance conditions for logical judgment. A certain score is accumulated for each rule passed, and the scores of all rules are finally summarized into a strategy compliance score. In practice, determining whether the strategy compliance score reaches a preset compliance standard is the decision threshold. The preset compliance standard is a minimum score line, such as 80 points (out of 100). The system compares the calculated strategy compliance score with this standard line. In practice, when the strategy compliance score reaches the preset compliance standard, incorporating the candidate flaw detection strategy into the basic flaw detection decision is the final output action. Incorporation means encapsulating the parameter settings, execution steps, and expected results of the candidate flaw detection strategy into a standardized decision object and adding it to the basic flaw detection decision set. The basic flaw detection decision set serves as the input for the subsequent real-time correction module.

[0086] It is understandable that the entire process of generating basic flaw detection decisions is a multi-layered filtering and optimization pipeline, gradually selecting a limited set of strategies from massive historical decision records that are both relevant to the current situation and meet safety and performance requirements. It is also understandable that the relevance matching stage, by combining static feature similarity and dynamic sequence similarity, can more comprehensively evaluate the correlation between real-time data and historical cases, thereby improving the quality of the matched decision subset. Optionally, the weighting coefficients in the weighted fusion algorithm... and weighting coefficients It can be designed as an adaptive parameter that dynamically adjusts based on the stability of the environmental disturbance data stream. For example, when the vibration spectrum data fluctuates drastically, the weight of the dynamic time warping distance can be appropriately reduced. To mitigate the impact of unstable temporal matching on the overall correlation score, in practice, if the matching decision subset remains empty after traversing all decision records (i.e., no overall correlation score reaches the correlation threshold), the system will activate an emergency strategy generation mechanism. Based on the weld defect distribution characteristics and the fundamental attributes of the environmental disturbance data stream, it will generate a most conservative flaw detection strategy as the basic flaw detection decision according to the default rule base. In practice, the performance compliance rule base supports online updates, allowing administrators to add, modify, or delete compliance rules without system downtime to adapt to regulatory changes or process updates. In practice, the strategy compliance score may be calculated using a weighted scoring method, assigning different importance weights to different performance compliance rule sets. Core safety rules have the highest weight, and their score has the greatest impact on the total strategy compliance score.

[0087] Example 4: In specific implementation, the key prediction step is to use a Long Short-Term Memory (LSTM) network model to predict the time series of weld defect distribution characteristics and generate defect evolution characteristic data. The LTM network model has a special structure with input gates, forget gates, output gates, and cell states, which can effectively handle long-term dependencies in sequence data. The system organizes the historically collected weld defect distribution characteristics into equally spaced time series data in chronological order. The time series data contains feature vectors at multiple time steps, each representing the defect distribution state at a specific moment. The LTM network model performs forward propagation calculations on the input time series data using parameters learned during training. The input gate controls the inflow of new information, the forget gate controls the retention of historical information, and the output gate controls the output from the current cell state to the hidden state. Finally, the output layer of the LTM network model generates predicted values ​​of weld defect distribution characteristics for one or more future time steps. These predicted values ​​are encapsulated as defect evolution characteristic data, which includes key information such as the predicted trend of defect quantity changes, defect size growth rate, and defect spatial distribution evolution pattern. In practical implementation, the use of a Kalman filter to estimate the state of the environmental disturbance data stream and generate environmental disturbance prediction data is a parallel state prediction process. The Kalman filter performs recursive prediction and updating based on the linear system state equation and observation equation. The system uses the equipment vibration spectrum and temperature gradient data in the environmental disturbance data stream as an observation sequence to establish a state-space model describing the changes in the environmental disturbance state. The Kalman filter executes two main stages at each time step: the prediction stage uses the state equation to predict the prior estimate of the state and the prior estimate of the error covariance at the current time based on the state estimate at the previous time step; the update stage uses the actual observation data at the current time step to correct the prior estimate and obtain the posterior estimate of the state and the posterior estimate of the error covariance at the current time step. By continuously repeating the prediction and update process, the Kalman filter outputs an optimized estimate of the future value of the environmental disturbance data stream, i.e., the environmental disturbance prediction data. The environmental disturbance prediction data includes parameters such as the predicted equipment vibration dominant frequency, vibration acceleration amplitude, and temperature field gradient changes.

[0088] In practical implementation, referring to Table 1, the first step in decision correction is to dynamically update the flaw detection decision map based on defect evolution characteristic data and environmental disturbance prediction data to obtain an optimized decision map. The dynamic update operation is achieved by modifying the metadata of existing decision records in the flaw detection decision map or inserting new simulated decision records. The system performs similarity matching between the defect evolution pattern predicted in the defect evolution characteristic data and the disturbance trend predicted in the environmental disturbance prediction data, and the sample weld defect distribution characteristic information and sample environmental disturbance data stream corresponding to each sample decision record stored in the flaw detection decision map. For sample decision records with high matching degree, the system will dynamically adjust the association weight or confidence score of the sample decision record in the flaw detection decision map according to the degree of difference between the predicted data and the sample data, as well as the uncertainty of the prediction result. At the same time, the system will synthesize one or more virtual decision records with time foresight based on defect evolution characteristic data and environmental disturbance prediction data, and add these virtual decision records to the flaw detection decision map. The sample flaw detection strategy in the virtual decision records is obtained by adaptively adjusting similar historical strategies. The database obtained after the above dynamic update process is the optimized decision map. In practice, the highest priority flaw detection strategy is extracted from the optimized decision graph as the core strategy. The extraction process is based on the updated comprehensive score of each decision record in the optimized decision graph. The comprehensive score is calculated by factors such as the original relevance weight of the record, the newly adjusted confidence score, and the newness of the record. The system sorts the comprehensive scores of all decision records in descending order and selects the flaw detection strategy associated with the record ranked first as the core strategy. The core strategy represents the optimal action plan determined by the system based on the current and future predicted states.

[0089] Table 1: List of Feasibility Verification and Inspection Items for Transitional Decisions

[0090] ;

[0091] In practical implementation, the feasibility verification of transition decisions includes resource constraint checks and timing consistency checks. Resource constraint checks aim to confirm whether the physical and computational resources required to execute the transition decision are within available limits. Checks include whether the joint load of the robotic arm is within safe thresholds, whether the estimated power consumption for the entire detection mission cycle is within the battery capacity's support range, and whether the storage space and transmission bandwidth of the large amount of ultrasonic echo data generated by the detection meet requirements. Timing consistency checks focus on analyzing the logical correctness and timing rationality of a series of ordered operations in the transition decision. This requires verifying whether there are overlaps or conflicts in the timing of various path segments in the robotic arm's motion trajectory, ensuring a strict sequential relationship between the data acquisition command issuance time, the probe positioning completion time, and the actual data acquisition start time, and that the time intervals meet system requirements. In practical implementation, after the feasibility verification is passed, the final flaw detection decision is output. The final flaw detection decision is the transition decision that has successfully passed all verification items. Its parameters are solidified into an executable sequence of equipment control commands and sent to the underlying controller of the remote flaw detection equipment via the system bus or network interface.

[0092] It is understandable that introducing defect evolution characteristic data and environmental disturbance prediction data to correct basic flaw detection decisions essentially expands the system's decision-making perspective from the static current state to the dynamic future trend, thereby significantly improving the foresight and robustness of the final flaw detection decision. In practical implementation, if a transitional decision fails either the resource constraint check or the timing consistency check, the system will not directly output failure. Instead, it will initiate a decision adjustment loop. This loop will attempt to fine-tune specific parameters in the transitional decision, and then re-verify the feasibility of the adjusted decision scheme until a verifiable feasible solution is found, or after reaching the maximum number of adjustment iterations, a backup safe and conservative strategy will be activated. In practical implementation, the process of dynamically updating the flaw detection decision map needs to possess transactional characteristics to ensure that in the event of an unexpected interruption during the update process, the flaw detection decision map can be rolled back to a consistent state, avoiding data corruption. In practical implementation, the equipment rated parameters upon which the resource constraint check depends, such as the rated torque of each joint of the robotic arm and the total battery capacity, are typically stored in the equipment control system in the form of configuration files and support online updates to adapt to hardware replacement or upgrades. In practice, temporal consistency checks can be formally verified using models based on time Petri nets or time automata to more rigorously prove that the temporal logic of each operation in the decision-making process is correct.

[0093] Example 5: In specific implementation, deploying an ultrasonic probe array to move and scan along the pipe weld is the main means of collecting ultrasonic echo signal sequences. The ultrasonic probe array consists of multiple piezoelectric ceramic transducer units arranged in a linear or matrix form. The array is mounted on a scanning trolley or robotic arm end effector that can move along the pipe weld trajectory. During the scanning process, each probe unit in the ultrasonic probe array emits high-frequency ultrasonic pulses according to a preset timing sequence. When the ultrasonic pulses enter the pipe weld and encounter defects or interfaces, they generate reflected echoes. The same probe unit or other units in the array receive these echo signals and convert them into electrical signals. After the electrical signals are amplified by a preamplifier and sampled by an analog-to-digital converter, a digital ultrasonic echo signal sequence is formed. The ultrasonic echo signal sequence contains detailed information about the internal structure of the weld. In practical implementation, an infrared thermal imager is installed to monitor the temperature changes on the weld surface to generate temperature gradient data. The infrared thermal imager is fixedly mounted on a bracket that overlooks the scanning area of ​​the pipe weld. The infrared detector array of the infrared thermal imager captures the infrared radiation energy of the weld and its surrounding area, and converts the radiation energy intensity into temperature values. By continuously capturing thermal images and analyzing the temperature value of each pixel in the image, the system can calculate the temperature at different locations on the weld surface and its rate of change over time, thereby generating temperature gradient data. The temperature gradient data is usually represented in the form of a two-dimensional temperature field matrix or a temperature-location-time series. In practical implementation, an accelerometer is used to measure the vibration frequency of the equipment to generate a vibration spectrum. The accelerometer is usually manufactured using MEMS technology and is firmly installed on the pipe body, support structure, or key vibration parts of the scanning equipment. The accelerometer monitors the vibration acceleration at its location in real time and outputs a voltage signal proportional to the acceleration. The voltage signal is processed by a conditioning circuit and a spectrum analysis module to convert the time-domain vibration acceleration signal into a frequency-domain vibration spectrum. The vibration spectrum clearly shows the distribution of vibration energy at different frequency components. In practical implementation, synchronizing and integrating the ultrasonic echo signal sequence, temperature gradient data, and vibration spectrum with timestamps into a unified data stream is a key step in data preprocessing. The system uses a high-precision clock source to provide a unified time reference for each data acquisition node, and each data packet is marked with a timestamp accurate to the millisecond level when it is generated. The data integration gateway receives the raw data with timestamps from the ultrasonic probe array, infrared thermal imager, and accelerometer through a high-speed bus or network interface, encapsulates and arranges the data according to a predefined data format specification, and finally forms a unified data stream with strict time consistency. The unified data stream is sent to the central processing unit for subsequent feature analysis and decision analysis.

[0094] In practical implementation, parsing the equipment control parameters in the final flaw detection decision is a prerequisite for driving the remote flaw detection equipment to execute actions. These parameters include key instructions such as the probe movement path and signal transmission intensity. The probe movement path is defined by a series of three-dimensional spatial coordinate points, movement speed, and attitude angle sequences, while the signal transmission intensity specifies the voltage amplitude and energy of the ultrasonic probe's excitation pulse. The system extracts these parameter values ​​from the final flaw detection decision's data structure using an analytical algorithm and converts them into a low-level instruction format that the equipment controller can recognize. In practical implementation, sending the equipment control parameters to the remote flaw detection equipment's controller via a wireless communication module is the core of remote control. The wireless communication module can use 5G, Wi-Fi 6, or a dedicated industrial wireless network protocol to ensure low-latency and high-reliability data transmission. The equipment control parameters are packaged into data frames, with checksums added to ensure transmission integrity, and then transmitted through the transmitter of the wireless communication module. The receiver of the wireless communication module at the remote flaw detection equipment receives the data frames, performs verification and unpacking, and transmits the parsed equipment control parameters to the flaw detection equipment's controller. In practice, the execution phase involves the controller adjusting the robotic arm posture and sensor settings of the flaw detection equipment and performing the scanning operation. Based on the received probe movement path parameters, the controller inversely solves the kinematic equations of the robotic arm and calculates the rotation angle sequence of each joint motor to drive the ultrasonic probe array carried at the end of the robotic arm to move precisely along the predetermined path. Simultaneously, the controller configures the parameters of the ultrasonic transmitting circuit, converting the signal transmission intensity parameters into specific pulse voltage values ​​and outputting them to the probes. After the scanning operation is initiated, the robotic arm moves along the planned path, and the ultrasonic probes transmit and receive ultrasonic waves according to the set parameters, acquiring flaw detection data in real time. In practice, flaw detection data is received in real time and fed back to the decision-making system to form a closed-loop control. During the scanning process, newly acquired ultrasonic echo signal sequences, as well as temperature gradient data and vibration spectrum continuously generated by infrared thermal imagers and accelerometers, are transmitted back to the central decision-making system in real time. The decision-making system compares these feedback data with the expected data to determine whether the flaw detection process proceeds as expected and whether the detection results are effective. It can also dynamically fine-tune some parameters in the final flaw detection decision based on the feedback information. For example, it can appropriately increase the signal transmission intensity when poor signal quality is detected, or replan the local movement path when an obstacle is encountered, thereby realizing a closed-loop control cycle of perception, decision-making, execution, and feedback.

[0095] In some embodiments, the motion scanning path planning of the ultrasonic probe array may need to consider the complex geometry of pipe welds, such as welds at elbows or tees. The path planning algorithm needs to generate a motion trajectory that ensures the probe always maintains the optimal incident angle with the weld surface. In some embodiments, temperature monitoring by an infrared thermal imager can be used to assist in identifying abnormal heat points caused by friction or internal damage, which may indicate potential defects or faults. Optionally, if the wireless communication module encounters signal interference during data transmission leading to packet loss, it will initiate an automatic retransmission request mechanism to ensure the complete delivery of control commands and feedback data. Optionally, the feedback adjustment in the closed-loop control can be set to different sensitivity levels. A high-sensitivity mode is used for detection areas with high precision requirements, allowing for more frequent and fine parameter adjustments. Optionally, the posture adjustment of the robotic arm may include a force-position hybrid control strategy. While ensuring the probe moves along the path, the force sensor feedback control keeps the contact force between the probe and the weld surface within a constant range to avoid excessive pressure damaging the probe or insufficient pressure causing poor coupling.

[0096] It is understandable that the entire process, from data acquisition through a distributed sensor network to closed-loop control of remote flaw detection equipment, constitutes a complete automated non-destructive testing system. This reduces manual intervention and improves the efficiency and safety of pipe weld flaw detection in confined spaces. In practice, the calibration of the ultrasonic probe array needs to be performed before the scanning task begins, using a standard test block to calibrate parameters such as probe zero bias, sensitivity, and sound velocity. Vibration spectrum analysis focuses on vibration components close to the equipment's natural frequency, as resonance can severely interfere with the test results. The unified data stream typically uses a structured format similar to JSON to facilitate parsing and processing between different subsystems. Controller reliability design is crucial, often employing redundancy or watchdog timers to prevent equipment malfunction due to single-point failures. The closed-loop control cycle time is a key performance indicator; the entire cycle from data feedback to decision adjustment and actuator response needs to be controlled within milliseconds to meet real-time requirements.

[0097] See Figure 5This figure is a vibration spectrum diagram of the equipment. The horizontal axis represents frequency, and the vertical axis represents vibration amplitude, visually presenting the vibration energy distribution characteristics of the pipeline weld flaw detection equipment at different frequencies. The figure is one of the environmental disturbance data streams generated after the vibration data of the equipment is collected by the accelerometer in the distributed sensor network and processed by the spectrum analysis module. In the pipeline weld flaw detection process, equipment vibration is a key environmental disturbance factor. Ignoring its interference with the flaw detection data can easily lead to misjudgment of defects. The value of this vibration spectrum diagram lies in providing a basis for subsequent flaw detection decision generation by quantifying the vibration energy at different frequencies and fusing the environmental disturbance data stream. The system can identify the inherent vibration frequency of the equipment, eliminate interference factors such as resonance, and ensure that defect analysis based on ultrasonic echo signals is not subject to false interference from environmental vibration. Ultimately, this improves the accuracy and reliability of pipeline weld flaw detection in the confined space of a water treatment plant, and is an important visual representation of the distributed sensor network data acquisition + environmental disturbance fusion decision-making technology link.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for remote controlled NDE of pipe welds in tight spaces in a water treatment plant, characterized in that, The method comprises: Real-time capture of pipeline weld flaw detection data stream and environmental disturbance data stream through a distributed sensor network, wherein the flaw detection data stream comprises a sequence of ultrasonic echo signals, and the environmental disturbance data stream comprises device vibration frequency spectrum and temperature gradient data; Multi-scale feature analysis is performed on the sequence of ultrasonic echo signals according to a predefined defect identification specification, and weld defect distribution feature information is extracted; Based on an embedded flaw detection decision graph and performance compliance conditions, the weld defect distribution feature information and the environmental disturbance data stream are fused to generate a basic flaw detection decision; Adaptive prediction calculation is performed on the weld defect distribution feature information and the environmental disturbance data stream to obtain defect evolution feature data and environmental disturbance prediction data; The basic flaw detection decision is corrected in real time using the defect evolution feature data and environmental disturbance prediction data, and a final flaw detection decision is output; The final flaw detection decision drives a remote flaw detection device to complete an automatic flaw detection process of the pipeline weld; The fusion of the weld defect distribution feature information and the environmental disturbance data stream based on the embedded flaw detection decision graph and performance compliance conditions to generate a basic flaw detection decision comprises: Correlation matching of the flaw detection decision graph is performed according to the weld defect distribution feature information and the environmental disturbance data stream to obtain a matching decision subset; Risk level division is performed on the matching decision subset to identify a safe decision interval; A candidate flaw detection strategy is generated in the safe decision interval; The consistency degree of the candidate flaw detection strategy with the performance compliance conditions is evaluated to obtain a strategy compliance score; It is judged whether the strategy compliance score reaches a preset compliance standard; When the strategy compliance score reaches the preset compliance standard, the candidate flaw detection strategy is included in the basic flaw detection decision; The correlation matching of the flaw detection decision graph according to the weld defect distribution feature information and the environmental disturbance data stream to obtain a matching decision subset comprises: Each decision record in the flaw detection decision graph is traversed, and each decision record comprises sample weld defect distribution feature information, sample environmental disturbance data stream and sample flaw detection strategy; A feature similarity matrix of the weld defect distribution feature information and the sample weld defect distribution feature information is calculated; A dynamic time warping distance of the environmental disturbance data stream and the sample environmental disturbance data stream is calculated; The feature similarity matrix and the dynamic time warping distance are combined according to a weighted fusion algorithm to generate a comprehensive correlation value; It is judged whether the comprehensive correlation value is greater than or equal to a correlation threshold value; When the comprehensive correlation value is greater than or equal to the correlation threshold value, the sample flaw detection strategy is added to the matching decision subset.

2. A method of remote controlled inspection of a pipe weld in a confined space of a water treatment plant as claimed in claim 1, wherein, The multi-scale feature analysis of the sequence of ultrasonic echo signals according to a predefined defect identification specification to extract weld defect distribution feature information comprises: A set of geometric structure parameters of the pipeline weld is obtained, and the set of geometric structure parameters comprises weld width, depth and curvature radius; A weld digital twin model is constructed according to the set of geometric structure parameters; inputting the ultrasonic echo signal sequence into the weld digital twin model to simulate a three-dimensional cloud map of internal defects of the weld; performing pattern recognition on the three-dimensional cloud map of internal defects of the weld based on defect size thresholds and shape rules in the predefined defect identification specification, segmenting a defect region and calculating a defect density distribution, thereby obtaining the weld defect distribution feature information.

3. A method of remote controlled inspection of a pipe weld in a confined space of a water treatment plant as claimed in claim 1, wherein, The method further comprises a preprocessing stage of constructing a defect feature analysis model, including: selecting a group of welds of the same material as a reference set for the pipeline weld, the reference set containing a plurality of historical test weld samples; loading a reference defect feature database from the reference set, the reference defect feature database storing defect echo data and defect classification labels of the reference welds; loading a dedicated defect feature record database for the current test pipeline weld; training a deep learning model using the reference defect feature database to establish an initial defect feature analysis model; performing transfer learning optimization on the initial defect feature analysis model using the dedicated defect feature record database to generate a high-precision defect feature analysis model for the multi-scale feature analysis.

4. A method of remote controlled inspection of a pipe weld in a confined space of a water treatment plant as claimed in claim 3, wherein, The transfer learning optimization on the initial defect feature analysis model using the dedicated defect feature record database to generate a high-precision defect feature analysis model includes: cross-validating the initial defect feature analysis model using the dedicated defect feature record database to calculate a model generalization error indicator; comparing the model generalization error indicator with an allowed error upper limit; when the model generalization error indicator exceeds the allowed error upper limit, enhancing training data through a generative adversarial network and retraining model parameters until the model generalization error indicator is reduced to within the allowed range to obtain the high-precision defect feature analysis model.

5. A method of remote controlled inspection of a pipe weld in a confined space of a water treatment plant as claimed in claim 1, wherein, The adaptive prediction calculation on the weld defect distribution feature information and the environmental disturbance data stream to obtain defect evolution feature data and environmental disturbance prediction data includes: applying a long short-term memory network model to perform time series prediction on the weld defect distribution feature information to generate defect evolution feature data; using a Kalman filter to perform state estimation on the environmental disturbance data stream to generate environmental disturbance prediction data.

6. A method of remote controlled inspection of a pipe weld in a confined space of a water treatment plant as claimed in claim 1, wherein, The real-time correction of the basic flaw detection decision using the defect evolution feature data and environmental disturbance prediction data to output a final flaw detection decision includes: dynamically updating the flaw detection decision graph based on the defect evolution feature data and environmental disturbance prediction data to obtain an optimized decision graph; extracting the flaw detection strategy with the highest priority from the optimized decision graph as a core strategy; fusing and calculating the core strategy with the basic flaw detection decision using a weighted average algorithm to generate a transition decision; performing feasibility verification on the transition decision, including resource constraint checking and timing consistency checking, and outputting the final flaw detection decision after passing.

7. A method for remote controlled inspection of a pipe weld in a confined space of a water treatment plant as claimed in claim 1, wherein, The real-time capture of the flaw detection data stream and the environmental disturbance data stream of the pipeline weld by the distributed sensor network includes: deploying an ultrasonic probe array to move along the pipeline weld to scan and collect ultrasonic echo signal sequences; An infrared thermal imager is installed to monitor the temperature change of the weld surface and generate temperature gradient data; An acceleration sensor is used to measure the vibration frequency of the equipment and generate a vibration spectrum; The ultrasonic echo signal sequence, temperature gradient data, and vibration spectrum are time-stamped synchronously and integrated into a unified data stream.

8. A method of remote controlled inspection of a pipe weld in a confined space of a water treatment plant as claimed in claim 1, wherein, The automatic flaw detection process of the pipeline weld is completed by the remote flaw detection equipment driven by the final flaw detection decision, including: Analyzing the equipment control parameters in the final flaw detection decision, including probe movement path and signal emission intensity; Sending the equipment control parameters to the controller of the remote flaw detection equipment through the wireless communication module; The controller adjusts the mechanical arm posture and sensor settings of the flaw detection equipment and performs the scanning operation; Real-time reception of flaw detection data and feedback to the decision system to form a closed-loop control.

Images