An online quality detection and control system for intelligent boiler manufacturing and artificial intelligence

Abstract

This invention relates to the field of intelligent boiler manufacturing and industrial quality inspection technology, and discloses an online quality inspection and control system for intelligent boiler manufacturing and artificial intelligence. The system includes a multi-source heterogeneous data real-time sensing and fusion acquisition module, an improved deep learning-based online defect detection module, a multi-factor coupling analysis module, an adaptive closed-loop control module, and a full-process data traceability and visualization management module. These modules are connected via data links to form a closed-loop feedback path. Quality deviation attribution and trend prediction are achieved through grey relational analysis and LSTM time series prediction, and parameter adaptive adjustment is driven by a fuzzy PID control algorithm, realizing online quality control throughout the entire boiler production process and operation cycle.

Description

Technical Field

[0001] This invention relates to a system for online quality detection and adaptive control in the intelligent manufacturing process of boilers. Background Technology

[0002] As a core pressure-bearing device for industrial production and energy conversion, the manufacturing quality and operational safety of boilers directly affect the stability of industrial production and personnel safety. The boiler manufacturing process involves multiple steps, including welding, forming, and assembly. Each step carries the risk of quality defects such as cracks, porosity, wall thickness deviations, and incomplete penetration. These defects often exhibit significant dimensional differences, random distribution, and varied shapes, placing high demands on detection methods.

[0003] In traditional boiler manufacturing quality control practices, quality inspection mainly relies on manual visual inspection and offline sampling. Manual visual inspection is heavily influenced by the experience and subjective judgment of the inspectors. In complex industrial environments such as high temperatures, dust, strong light reflection, and continuous vibration, the inspectors' perception is significantly limited, making it difficult to effectively control the missed detection rate of minute cracks and hidden defects. Offline sampling, on the other hand, suffers from insufficient timeliness; defects are often only discovered after batch production is complete, leading to high rework costs and scrap rates, and making real-time intervention in the production process impossible.

[0004] At the level of process parameter control, the parameter adjustments in existing boiler production lines mainly rely on the experience and judgment of operators, lacking quantitative analysis support for the nonlinear coupling relationship between quality parameters and process parameters. Since boiler manufacturing involves the synergistic effect of multiple parameters such as welding current, welding speed, heating temperature, air volume ratio, and fuel ratio, empirical adjustments of a single parameter often fail to accurately pinpoint the root cause of quality deviations, resulting in problems such as inaccurate control direction and delayed control response.

[0005] At the quality data management level, existing systems generally lack the ability to systematically archive and trace quality data throughout the entire boiler lifecycle. Inspection records, process parameter records, and control operation records for each process are stored in a scattered manner, making it difficult to form a complete quality traceability chain. This is not conducive to meeting the requirements of digital archiving for special equipment quality supervision, nor is it conducive to continuously optimizing inspection and control models through the accumulation of historical data.

[0006] Given the shortcomings of the existing technologies in terms of real-time detection capabilities, parameter correlation analysis, closed-loop control response, and data traceability management, it is necessary to provide an online quality control system that can organically integrate multi-source sensing, deep learning detection, coupled analysis, and adaptive control to achieve intelligent quality assurance throughout the entire boiler manufacturing and operation process. Summary of the Invention

[0007] This invention proposes measures for online quality detection and control in the intelligent manufacturing process of boilers.

[0008] According to the present invention, this is achieved by the following method: multi-source heterogeneous data is collected synchronously from the boiler production process and the operation site by a distributed sensing unit and preprocessed. Then, the data is sequentially processed through an improved deep learning detection network, a multi-factor coupling analysis module, and an adaptive closed-loop control module to complete the entire process of defect identification, correlation analysis, and closed-loop control. The modules are connected in series by data links to form a closed-loop feedback path.

[0009] According to the present invention, boiler manufacturing sites are characterized by concurrent multi-source heterogeneous data, diverse defect types, and complex coupling relationships between process parameters and quality fluctuations. A single sensing method or algorithm is difficult to simultaneously meet the requirements of real-time detection accuracy and control response speed. By systematically integrating attention-enhanced deep learning detection networks, multi-factor coupling analysis combining grey relational analysis and LSTM prediction, and PID control algorithms based on fuzzy inference, real-time perception, trend prediction, and adaptive control of quality status can be achieved within a unified closed-loop framework.

[0010] This invention proposes an online quality inspection and control system for intelligent boiler manufacturing and artificial intelligence. Preprocessed image and sensor data are input into an improved deep learning detection network for real-time inference. This network includes a backbone feature extraction unit, a channel spatial attention unit, a multi-scale feature fusion unit, and a detection output unit connected in sequence. The type, location, size, and grade of various quality defects are identified and output. Defect diagnosis data, real-time process parameter data, and operating condition data are input into a multi-factor coupling analysis module. A grey relational analysis algorithm is used to quantify the correlation weights between quality parameters, process parameters, and operating condition parameters. An LSTM network is used to predict the quality fluctuation trend within a preset time window. Defect classification results, correlation weight data, and trend prediction data are input into an adaptive closed-loop control module. A fuzzy PID control algorithm dynamically generates incremental adjustment commands for each controlled parameter and performs automatic correction or pushes manual handling according to the hierarchical control logic.

[0011] The channel spatial attention unit calculates attention weights for each level of feature map output by the backbone feature extraction unit along both the channel and spatial dimensions. The feature map after attention weighting... It is given by the following formula: ; in, For the input feature map, Attention weights are calculated along the channel dimension. Attention weights are calculated along the spatial dimension. This represents element-wise multiplication. By applying dual attention weights to both the channel and spatial dimensions, the network's response strength to discriminative features related to defects is enhanced, while its response to background interference regions is correspondingly suppressed.

[0012] In a preferred embodiment of the invention, the multi-scale feature fusion unit receives the attention-weighted shallow feature map. With deep feature maps Cross-scale fusion is achieved through bidirectional feature transfer via top-down and bottom-up paths, resulting in fused feature maps. It is given by the following formula: ; in, This indicates that the deep feature map is upsampled to the same spatial resolution as the shallow feature map, and then channel-aligned by a convolutional layer. This indicates that the shallow feature map is downsampled to the same spatial resolution as the deep feature map through stride convolution, and then channel aligned by a convolutional layer. Bidirectional feature propagation preserves the fine-grained spatial information of the shallow layer and the high-level semantic information of the deep layer in a unified feature representation, thereby improving the detection coverage of defects of different sizes.

[0013] In a preferred embodiment of the present invention, the input side of the improved deep learning detection network is configured with an adaptive anti-interference processing unit. This unit sequentially performs the following operations: decomposing the input image into structural and texture components based on morphological filtering, with only the structural component used as the primary input for subsequent detection, and the texture component used as an auxiliary reference channel input; performing adaptive histogram equalization on the structural component to eliminate local overexposure; and performing inverse displacement compensation on the image frame sequence based on the synchronous vibration signal collected by the vibration sensor to correct image blurring and displacement deviation caused by equipment vibration. These processing steps effectively ensure the input image quality of the detection network in complex industrial environments.

[0014] In a preferred embodiment of the present invention, the positioning bounding box coordinates output by the detection output unit are relative coordinates normalized to the image size. After inverse normalization transformation and mapping to actual pixel coordinates, they are then converted into actual physical dimensions in millimeters by combining sensor calibration parameters. The defect grading and evaluation unit receives the defect type, defect size, and defect location information output by the detection output unit. Based on a pre-written industry standard rule base, it automatically classifies each detected defect into four levels: minor, moderate, severe, and fatal. The accurate restoration of physical dimensions and standardized grading and evaluation provide an objective quantitative basis for subsequent control decisions.

[0015] In a preferred embodiment of the present invention, the grey relational analysis algorithm uses defect detection rate, defect severity level, and wall thickness deviation as reference sequences, and welding current, welding speed, heating temperature, furnace negative pressure, air volume ratio, fuel ratio, and load fluctuation amplitude as comparison sequences. It outputs the correlation degree value of each comparison sequence relative to the reference sequence. A higher correlation degree value indicates a greater weight of the parameter's influence on quality fluctuation. Each parameter in the reference and comparison sequences is Z-score standardized before being input into the grey relational analysis algorithm. Standardization eliminates the numerical magnitude differences between parameters of different dimensions, making the correlation degree calculation results comparable.

[0016] In a preferred embodiment of the present invention, the input of the LSTM network is a multidimensional time series extracted from real-time data streams and historical quality inspection databases, and the output is a future preset time window. The predicted values ​​of the quality parameters at each time point and their corresponding confidence intervals, where The prediction step size is preset according to the response delay of the specific process. The predicted value of the quality parameter output by the LSTM network is a normalized value after standardization. After being denormalized and restored to the actual dimensional value of the corresponding physical quantity, it is compared with the preset quality acceptance domain boundary. When the predicted value exceeds the quality acceptance domain boundary, an early warning signal is generated and transmitted to the adaptive closed-loop control module.

[0017] In a preferred embodiment of the present invention, the fuzzy PID control algorithm uses the current quality deviation value, the deviation change rate, and the parameter weight information output by the multi-factor coupling analysis module as inputs to control the output quantity. It is given by the following formula: ; in, Let be the mass deviation value at time t. Let be the rate of change of the deviation at time t. This is the proportionality coefficient. The integral coefficient is... The differential coefficients are... , , All are based on fuzzy reasoning rules and The membership degree is dynamically given. The incremental adjustment output of the fuzzy PID control algorithm is a normalized value, which is restored to the actual physical quantity increment of the corresponding controlled parameter after inverse normalization transformation and then sent to the production line controller for execution. By introducing parameter weight information into the fuzzy inference input, the generation of control commands has an adaptive response capability to the actual influence of each process parameter.

[0018] In a preferred embodiment of the present invention, the hierarchical control logic is divided into four levels based on the defect severity and executed sequentially: when a defect is determined to be minor, the adjustment instruction is automatically generated by the fuzzy PID control algorithm and directly sent to the production line controller for execution; when a defect is determined to be moderate, the adjustment instruction is simultaneously pushed to the operation terminal for confirmation after generation, and the warning information is displayed on the visualization platform; when a defect is determined to be severe, the operation of the corresponding process is automatically locked, and the adjustment plan along with the attribution analysis report is pushed to the management terminal for handling decisions; when a defect is determined to be critical, the process and its downstream related processes are automatically isolated and shut down, and the alarm signal is broadcast throughout the entire system. The four-level control hierarchy ensures that the response measures triggered by defects of different severity levels match their potential risk levels.

[0019] In a preferred embodiment of the present invention, during continuous operation, newly generated defect detection samples and control feedback data are automatically collected and incorporated into the training dataset. The improved deep learning detection network and the fuzzy PID control algorithm perform incremental training and parameter updates at a preset update cycle. All process data is automatically recorded in the full lifecycle quality database and indexed and archived using unique equipment identifiers and process numbers. Incremental training enables the detection and control capabilities to continuously evolve with the accumulation of operational data, and the full lifecycle data archiving provides complete data support for quality traceability and compliance review.

[0020] [System]: This invention also proposes an online quality detection and control system for intelligent boiler manufacturing and artificial intelligence, used to execute the above-mentioned methods. The system includes a multi-source heterogeneous data real-time sensing and fusion acquisition module, an improved deep learning online defect detection module, a multi-factor coupling analysis module, an adaptive closed-loop control module, and a full-process data traceability and visualization management module. The modules are connected in series via data links to form a closed-loop feedback path.

[0021] The multi-source heterogeneous data real-time sensing and fusion acquisition module synchronously collects multi-dimensional raw data from the boiler production process and operation site through distributed sensing units, and outputs fused data after multi-level filtering and noise reduction, heterogeneous format unification, and timestamp alignment. The improved deep learning defect online detection module performs real-time inference on the fused data based on the improved deep learning detection network to identify and output the type, location, size, and level of various quality defects. The multi-factor coupling analysis module receives defect diagnosis data, real-time process parameter data, and operating condition data, quantifies the correlation weights through a grey relational analysis algorithm, and predicts the quality fluctuation trend within a preset time window through an LSTM network. The adaptive closed-loop control module dynamically generates incremental adjustment commands based on defect classification results, correlation weight data, and trend prediction data using a fuzzy PID control algorithm, and executes control according to the hierarchical management logic. The full-process data traceability and visualization management module receives and archives all process data from upstream modules and presents the system operation status and quality situation in a visual form.

[0022] The online quality detection and control system provided by this invention has the following technical effects: Firstly, the multi-source heterogeneous data real-time sensing and fusion acquisition module integrates multiple types of sensing units and is configured with multi-level filtering and noise reduction processing units, so that the noise level of the pre-processed data is controlled within a preset threshold, thereby providing input data with satisfactory quality for subsequent detection models, and suppressing the impact of industrial environmental factors such as high temperature dust and electromagnetic interference on data quality.

[0023] Secondly, the improved deep learning detection network introduces channel spatial attention units and multi-scale feature fusion units on the basis of the lightweight YOLO detection network, which enhances the response of defect-related features and suppresses the response of background noise. After cross-scale fusion of shallow detail features and deep semantic features, targets with significant scale differences, such as microcracks, shallow pores and irregular hidden defects, can be effectively detected. The configuration of the adaptive anti-interference processing unit enables the detection module to achieve full coverage detection of the production line without stopping the machine in complex industrial environments.

[0024] Third, the multi-factor coupling analysis module uses the grey relational analysis algorithm to quantify and evaluate the nonlinear coupling relationship between quality parameters and process parameters, transforming the fuzzy parameter correlation in traditional experience into a quantifiable weight ranking; the LSTM network predicts the quality fluctuation trend within a preset time window, enabling quality control to shift from reactive response to proactive prevention, and the formation of batch quality defects and safety hazards can be avoided in advance.

[0025] Fourth, the adaptive closed-loop control module adopts a fuzzy PID control algorithm, which dynamically adjusts the proportional coefficient, integral coefficient, and derivative coefficient according to the membership degree of the deviation value and the rate of change of deviation. The control response can adaptively match with the changes in operating conditions, and the overshoot and oscillation amplitude during the adjustment process are controlled within the preset stability range. The hierarchical closed-loop control logic enables most operating condition deviations to be automatically corrected without human intervention, and only major defects are submitted to the manual handling channel.

[0026] Fifth, the full-process data traceability and visualization management module indexes and archives the quality data of the entire life cycle using the unique equipment identifier and process number, so that the quality data of a single piece of equipment and a single process can be completely traced, meeting the requirements of special equipment quality supervision and digital archiving; the system has self-learning capabilities, and the improved deep learning detection network and fuzzy PID control algorithm perform incremental training and parameter optimization with a preset update cycle, and the recognition accuracy and control adaptability continue to improve with the accumulation of operating data. Attached Figure Description

[0027] Figure 1 This is a flowchart of the online quality detection and control method for intelligent boiler manufacturing and artificial intelligence provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the overall flowchart of the online monitoring system for boiler welding quality provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the internal processing flow of the online defect detection network provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the ranking of the correlation degree of each process parameter in the grey relational analysis provided in the embodiment of the present invention; Figure 5 This is a schematic diagram comparing the LSTM quality trend prediction results with the qualified domain boundary provided in this embodiment of the invention; Figure 6 This is a schematic diagram comparing welding process parameters before and after adjustment, provided in an embodiment of the present invention. Figure 7 This is a schematic diagram of the distribution of normalized values ​​of multi-source sensing data provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the archiving and traceability structure of the full lifecycle quality database provided in this embodiment of the invention; Figure 9 This is a schematic diagram illustrating the trend of defect detection rate changes before and after the implementation of the control measures provided in this embodiment of the invention. Detailed Implementation

[0028] Example 1: The present invention will be further described below with reference to the accompanying drawings and specific embodiments. In a preferred embodiment of the present invention, a quality online detection and control system for intelligent boiler manufacturing and artificial intelligence is provided. This quality online detection and control system includes a multi-source heterogeneous data real-time sensing and fusion acquisition module, an improved deep learning defect online detection module, a multi-factor coupling analysis module, an adaptive closed-loop control module, and a full-process data traceability and visualization management module. The above modules are connected in series by data links to form a closed-loop feedback path, thereby realizing online quality control and dynamic parameter adjustment of the entire boiler production process and the entire operation cycle.

[0029] I. Real-time Sensing and Fusion Acquisition Module for Multi-Source Heterogeneous Data In a preferred improvement of the method according to the present invention, the multi-source heterogeneous data real-time sensing and fusion acquisition module is configured as a distributed sensing layer for the intelligent manufacturing production line and operation site of boilers, for synchronously acquiring full-dimensional raw data covering production processes and operating conditions.

[0030] In some embodiments, the sensing units integrated by the multi-source heterogeneous data real-time sensing and fusion acquisition module include: a high-definition industrial camera, a laser thickness sensor, an ultrasonic flaw detector, a temperature and pressure sensor, a furnace flame sensor, a flue gas composition sensor, and a vibration sensor. These sensing units are deployed at key locations in the welding, forming, assembly, and boiler operating sections, enabling the simultaneous acquisition of multi-dimensional data such as weld images, wall thickness dimensions, internal crack signals, surface defect images, furnace temperature field, combustion flame state, flue gas composition, and equipment vibration amplitude. In various embodiments, the multi-source heterogeneous data real-time sensing and fusion acquisition module is also compatible with LIBS spectral composition analysis data, thereby achieving comprehensive detection of the base material composition, weld composition, and heated surface material condition.

[0031] In a preferred embodiment of the present invention, the multi-source heterogeneous data real-time sensing and fusion acquisition module incorporates a multi-level filtering and noise reduction processing unit to address environmental factors commonly found in industrial environments, such as high-temperature heat radiation, high-concentration dust, and electromagnetic interference from multiple devices. This multi-level filtering and noise reduction processing unit performs dual denoising operations in the spatial and frequency domains on the image data and baseline drift correction and missing value interpolation operations on the analog sensor data. Advantageously, this ensures that the noise level of the preprocessed data is controlled within a preset threshold, and the data quality meets the input requirements of subsequent deep learning models.

[0032] In some embodiments, the multi-source heterogeneous data real-time sensing and fusion acquisition module is further configured with a standardized data interface layer to unify the communication protocols and storage formats of various heterogeneous data. Production line equipment, testing equipment, and control systems achieve data interconnection through the standardized data interface layer, and the data acquisition frequency is set to be no less than a preset frequency threshold, thereby ensuring the real-time performance and integrity of the data. This can promote the alignment and fusion of multi-source heterogeneous data in a unified data space. For example, sensor data with different sampling rates are synchronized to a unified time base through timestamp alignment, thus providing a complete and consistent data foundation for subsequent detection and analysis.

[0033] Since the data collected by each sensing unit encompasses physical quantities with varying dimensions, such as image pixel values, temperature, pressure, vibration amplitude, wall thickness, and flue gas concentration, this multidimensional heterogeneous data must be normalized before being input into the subsequent model to eliminate the influence of dimensional differences on the calculation. Z-score normalization is applied to various numerical sensing data, mean normalization based on the range is applied to image pixel data, and decimal scaling normalization is applied to proportional flue gas composition data, ensuring that all input features are mapped to a uniform numerical range.

[0034] II. Improved Deep Learning-Based Online Defect Detection Module; In a preferred improvement of the method according to the present invention, the improved deep learning online defect detection module is configured to receive preprocessed data from the multi-source heterogeneous data real-time perception and fusion acquisition module, and perform real-time identification and hierarchical diagnosis of various quality defects of the boiler based on the improved deep learning detection network.

[0035] 2.1 Structure of the improved deep learning detection network; The improved deep learning detection network used in this invention is an improved network formed by introducing a channel spatial attention submodule and a multi-scale feature fusion submodule on the basis of a lightweight YOLO detection network. Its constituent units and the data transmission relationship between each unit are as follows.

[0036] The improved deep learning detection network includes a backbone feature extraction unit, a channel spatial attention unit, a multi-scale feature fusion unit, and a detection output unit.

[0037] The backbone feature extraction unit takes preprocessed image data as input and extracts multi-level feature maps from shallow to deep through layer-by-layer convolution operations. Among them, the shallow feature maps retain detailed information such as defect edges and textures, while the deep feature maps encode the semantic category information of the defects.

[0038] The channel spatial attention unit receives feature maps from each level output by the backbone feature extraction unit and calculates attention weights along both the channel and spatial dimensions. and ,in This represents the response weight of each channel to the defect feature. This represents the response weight of each spatial location to the defect region. The feature map after attention weighting. It is given by the following formula: ; in, For the input feature map, Attention weights are calculated along the channel dimension. Attention weights are calculated along the spatial dimension. This represents element-wise multiplication. This operation enhances feature responses related to defects while suppressing feature responses related to background noise and process textures. The attention-weighted feature maps at each level are then passed to the multi-scale feature fusion unit.

[0039] The multi-scale feature fusion unit receives the attention-weighted shallow feature map. With deep feature maps By employing bidirectional feature transfer via top-down and bottom-up paths, shallow detail features and deep semantic features are fused across scales, resulting in a fused feature map. It is given by the following formula: ; in, This represents a top-down upsampling path operation, specifically for deep feature maps. After performing bilinear interpolation upsampling to the same spatial resolution as the shallow feature map, channel alignment is performed through a convolutional layer. This represents a bottom-up downsampling path operation, specifically for shallow feature maps. After performing stride convolution downsampling to the same spatial resolution as the deep feature map, channel alignment is performed through convolutional layers. This allows the improved deep learning detection network to maintain high detection sensitivity for targets with significant scale differences, such as microcracks, shallow pores, and irregular hidden defects.

[0040] The detection output unit receives the fused feature map. After mapping by a fully connected layer, the system outputs the category label, confidence score, localization bounding box, and estimated defect size for each defect target. The localization bounding box coordinates output by the detection output unit are relative coordinates normalized to the image size. They must be inversely normalized and mapped to actual pixel coordinates before being converted into actual physical dimensions in millimeters by combining sensor calibration parameters. This yields defect location and size information that can be directly used for defect grading and control decisions.

[0041] The improved deep learning detection network described above uses an annotated defect image dataset as training input and a weighted sum of class cross-entropy loss and bounding box regression loss as the total loss function. It is trained using transfer learning: first, the network is pre-trained with a massive amount of historical boiler quality inspection samples to obtain general defect feature representations; then, the parameters are fine-tuned with annotated data from the target scene, using the Adam optimization algorithm. This allows the improved deep learning detection network to converge quickly and achieve high recognition accuracy in new scenes with limited training samples.

[0042] 2.2 Adaptive anti-interference processing for industrial scenarios; In a preferred embodiment of the present invention, the improved deep learning defect online detection module is equipped with an adaptive anti-interference processing unit on the input side of the improved deep learning detection network, which is used to perform pre-adaptive processing on the input image data for industrial scenarios during the inference stage.

[0043] The adaptive anti-interference processing unit includes the following operation steps performed in sequence: Step 1: Image texture separation. The input image is decomposed into structural and texture components based on morphological filtering. The structural component retains the geometric contour information of defects, while the texture component contains processing traces and environmental interference information. Only the structural component is used as the primary input for subsequent detection, while the texture component serves as an auxiliary reference channel input.

[0044] Step two, background noise reduction and strong light suppression. Adaptive histogram equalization is performed on the structural components to eliminate local overexposure caused by high-temperature thermal radiation and strong light reflection areas, thereby improving the contrast between defective and non-defective areas to a distinguishable level.

[0045] Step 3, vibration error compensation operation. Based on the synchronous vibration signal collected by the vibration sensor, reverse displacement compensation is performed on the image frame sequence to correct the image blurring and displacement deviation caused by equipment vibration.

[0046] Advantageously, this enables the improved deep learning online defect detection module to maintain stable detection performance in complex industrial environments such as high temperature dust, strong light reflection, and continuous equipment vibration, and to achieve full coverage detection of the production line without shutdown.

[0047] 2.3 Defect grading and diagnosis; In various embodiments, the improved deep learning-based online defect detection module also includes a defect grading and evaluation unit. This unit receives defect type, size, and location information from the detection output unit and automatically classifies each detected defect into four levels: minor, moderate, severe, and critical, based on a pre-written industry standard rule base. The grading results, along with a defect diagnosis report and preliminary defect cause analysis results, are simultaneously output to the multi-factor coupling analysis module and the adaptive closed-loop control module. This provides a standardized defect information format for subsequent precise control and rework decisions.

[0048] III. Multi-factor Coupling Analysis Module; In a preferred improvement of the method according to the present invention, the multi-factor coupling analysis module is configured to receive defect diagnosis data from the improved deep learning online defect detection module and real-time process parameter data and operating condition data from the multi-source heterogeneous data real-time perception and fusion acquisition module, and realize the quantitative attribution and trend prediction of quality deviation through correlation analysis and time series prediction.

[0049] Because parameters such as defect detection rate, defect severity level, wall thickness deviation, welding current, welding speed, heating temperature, furnace negative pressure, air volume ratio, fuel ratio, and load fluctuation amplitude have different dimensions, these parameters must be standardized by Z-score before being input into the correlation analysis and time series prediction model. This is to eliminate the influence of dimensional differences on correlation calculation and time series prediction operation, and to ensure that the parameter sequences are within a comparable numerical range.

[0050] 3.1 Nonlinear coupling correlation analysis; In a preferred embodiment of the present invention, a grey relational analysis algorithm is used to quantitatively evaluate the nonlinear coupling relationship between boiler quality parameters, production process parameters, and operating condition parameters. The grey relational analysis algorithm uses defect detection rate, defect severity level, and wall thickness deviation as reference sequences, and process parameters and operating condition parameters such as welding current, welding speed, heating temperature, furnace negative pressure, air volume ratio, fuel ratio, and load fluctuation amplitude as comparison sequences. It outputs the correlation degree value of each comparison sequence relative to the reference sequence; a higher correlation degree value indicates a greater weight of the parameter's influence on quality fluctuations.

[0051] This allows us to transform the vague parameter relationships in traditional experience into quantifiable weight rankings, providing an objective basis for subsequent regulation.

[0052] 3.2 Quality trend time series prediction; In some embodiments, the multi-factor coupling analysis module further incorporates an LSTM network for time-series prediction of quality parameters. The input to the LSTM network is a multi-dimensional time series extracted from real-time data streams and historical quality inspection databases, including defect detection rate sequences, wall thickness variation sequences, key process parameter sequences, and operating condition parameter sequences; the output layer is a fully connected layer, and the output is a future preset time window. The predicted values ​​of the quality parameters at each time point and their corresponding confidence intervals, where The prediction step size, pre-set based on the response delay of specific processes, represents the length of time to predict backward from the current moment. The predicted quality parameters output by the LSTM network are normalized values ​​after standardization. These must be de-standardized to restore the actual dimensional values ​​of the corresponding physical quantities before comparison with the preset quality acceptance boundary. The LSTM network uses historical quality inspection time-series data as training samples, with the mean squared error between the predicted and actual values ​​as the loss function, and is trained using the Adam optimization algorithm.

[0053] When the predicted value exceeds the preset quality compliance range boundary, an early warning signal is generated and transmitted to the adaptive closed-loop control module. This can promote the transformation of quality control from reactive response to proactive prevention. For example, the early prediction of welding deformation trends, wall thickness deviation evolution direction, and aging process of heated surfaces can be achieved, and the formation of batch quality defects and safety hazards can be avoided in advance.

[0054] 3.3 Intelligent attribution and tracing; In various embodiments, the multi-factor coupling analysis module also includes an attribution analysis unit. When a quality deviation or operational anomaly is detected, the attribution analysis unit automatically distinguishes the categories of causes leading to the deviation based on the parameter weight ranking output by the grey relational analysis algorithm and the abnormal time period location output by the LSTM network. These causes include equipment status deviations, process parameter deviations, environmental disturbance fluctuations, and human operation errors. The attribution analysis results are appended to the defect diagnosis report and transmitted to the adaptive closed-loop control module to drive targeted parameter corrections. Advantageously, this avoids the problem of blindly taking corrective measures due to unclear attributions in traditional quality control.

[0055] IV. Adaptive Closed-Loop Control Module; In a preferred improvement of the method according to the present invention, the adaptive closed-loop control module is configured to receive defect classification results from the improved deep learning online defect detection module, correlation weight data and trend prediction data from the multi-factor coupling analysis module, and dynamically generate and execute parameter adjustment instructions accordingly to achieve adaptive optimization of production process parameters and boiler operating parameters.

[0056] 4.1 Adaptive fuzzy PID control; In a preferred embodiment of the present invention, the adaptive closed-loop control module employs a fuzzy PID control algorithm as the execution algorithm for parameter adjustment. The fuzzy PID control algorithm takes the current quality deviation value (i.e., the difference between the measured quality parameter and the target quality parameter), the deviation change rate, and the parameter weight information output by the multi-factor coupling analysis module as inputs, and outputs the incremental adjustment amount for each controlled process parameter and operating condition parameter.

[0057] Among them, the fuzzy inference step is based on the deviation value. and the rate of change of deviation The proportional gain of the PID controller is dynamically adjusted based on the fuzzy range in which it operates. Integral coefficient With differential coefficients Regulate output It is given by the following formula: ; Where t is the current time, For integration variables, Let be the mass deviation value at time t. Let be the rate of change of the deviation at time t. This is the proportionality coefficient. The integral coefficient is... The differential coefficients are... , , All are based on fuzzy reasoning rules and The membership degree is given dynamically. Let be the control output at time t, enabling the control response to adaptively match changes in operating conditions. The input fuzzy PID control algorithm's quality deviation value is also considered. and the rate of change of deviation All values ​​are dimensionless values ​​after Z-score standardization, representing the incremental adjustment output of the fuzzy PID control algorithm. Similarly, normalized values ​​must undergo denormalization to be restored to the actual physical quantity increment of the corresponding controlled parameter before they can be sent to the production line controller for execution. In various embodiments, the method includes the following adjustment scenarios: for welding quality deviations, welding speed and current / voltage parameters are automatically fine-tuned; for combustion uniformity deviations, the air-coal ratio and furnace temperature setpoints are automatically matched; for wall thickness deformation deviations, forming process parameters are automatically corrected.

[0058] This can improve the real-time performance and stability of parameter adjustment. For example, the overshoot and oscillation amplitude during the adjustment process are controlled within the preset stability range.

[0059] 4.2 Hierarchical closed-loop control logic; In a preferred embodiment of the present invention, the control logic of the adaptive closed-loop control module is divided into four levels and executed sequentially according to the defect level: When a defect is determined to be minor, the corresponding process parameter adjustment instructions are automatically generated by the fuzzy PID control algorithm and directly sent to the production line controller for execution, without the need for operator intervention.

[0060] When a defect is determined to be of a general level, the adjustment instruction is pushed to the operation terminal for confirmation after it is generated, and the warning information is displayed on the visualization platform of the full-process data traceability and visualization management module.

[0061] When a defect is determined to be severe, the operation of the corresponding process is automatically locked, and the adjustment plan, along with the attribution analysis report, is pushed to the management terminal for handling decisions.

[0062] When a defect is determined to be critical, the process and its downstream related processes are automatically isolated and shut down, and the alarm signal is broadcast throughout the entire process area.

[0063] Advantageously, this allows the vast majority of operating condition deviations to be automatically corrected without human intervention, with only major defects being submitted to the manual handling channel, thus achieving a fully automated closed-loop operation from detection to correction.

[0064] 4.3 Multi-scenario adaptive adaptation; In some embodiments, the adaptive closed-loop control module supports dynamic updates of control parameters and model weights through incremental learning. When the online quality monitoring and control system is deployed to boiler production lines of different tonnages, materials, or process types, it collects operational data under that scenario and incrementally updates model parameters to adapt the control strategy to the specific operating characteristics of that scenario. This adaptation process also covers different operating modes such as stable operation, deep peak shaving, and variable load operation, thereby significantly enhancing the generalization ability of the online quality monitoring and control system across different boiler types and operating conditions.

[0065] V. Full-process data traceability and visualization management module; In a preferred improvement of the method according to the present invention, the full-process data traceability and visualization management module is configured to receive and archive all process data from upstream modules, while presenting the system operation status and quality status in a visual form.

[0066] 5.1 Full lifecycle data traceability and archiving; In a preferred embodiment of the present invention, the full-process data traceability and visualization management module incorporates a full lifecycle quality database to automatically store production parameter records, inspection data records, defect information records, control command records, operating condition records, and rectification operation records for each boiler unit. These records are indexed by a unique equipment identifier and process number, enabling complete traceability of quality data for individual equipment and processes. This provides a standardized traceability format to meet the requirements of special equipment quality supervision and digital archiving.

[0067] 5.2 Visualization; In some embodiments, the end-to-end data traceability and visualization management module is configured with a dual-channel visualization platform on both web and mobile devices. Production line quality status, defect spatial distribution, process parameter variation curves, equipment operating status, and risk warning information are presented in real time. In various embodiments, defect distribution is overlaid on a boiler structure diagram in the form of a heat map, the temporal changes of quality parameters are continuously plotted in the form of a trend chart, and the data flow relationships between various processes are presented in the form of a topology diagram. This allows managers to achieve a comprehensive and intuitive grasp of the production quality status and operational safety status.

[0068] 5.3 Model self-learning and iterative optimization; In a preferred embodiment of the present invention, the online quality detection and control system possesses self-learning capabilities. During continuous operation, newly generated defect detection samples and control feedback data are automatically collected and incorporated into the training dataset. An improved deep learning detection network and a fuzzy PID control algorithm perform incremental training and parameter optimization at a preset update cycle, enabling continuous improvement in recognition accuracy and control adaptability as operational data accumulates. This self-learning pathway can be flexibly combined to adapt to the performance optimization needs of different production lines and boiler models at different operational stages.

[0069] VI. Overall System Operation Flow; In a preferred embodiment of the present invention, the above modules operate collaboratively according to the following standardized process: Step S1, Multi-source Sensing Acquisition. After the online quality inspection and control system is started, the sensing units and vision acquisition units deployed at each process location collect real-time raw data of the boiler production and operation process from all dimensions. After multi-level filtering and noise reduction, heterogeneous format unification and timestamp alignment, the fused data is output to the improved deep learning online defect detection module and multi-factor coupling analysis module.

[0070] Step S2, online defect detection. The improved deep learning detection network performs real-time inference on the image data and sensor data in the fused data, and identifies and outputs the type, location, size and grade of various quality defects.

[0071] Step S3, Coupling Analysis and Trend Prediction. The grey relational analysis algorithm quantifies and evaluates the correlation weights between quality parameters, process parameters, and operating condition parameters. The LSTM network predicts future trends within a preset time window. The system predicts the quality fluctuation trend within the system, and the attribution analysis unit automatically traces the causes of detected deviations.

[0072] Step S4, Adaptive Closed-Loop Control. The fuzzy PID control algorithm dynamically generates incremental adjustment commands for each controlled parameter based on the defect classification results, associated weight data, and trend prediction data, and performs automatic correction or pushes manual handling according to the hierarchical control logic.

[0073] Step S5, Data Archiving and Traceability. All process data generated in steps S1 to S4 above are automatically recorded in the full lifecycle quality database and archived using equipment identifiers and process numbers as indexes.

[0074] Step S6: Model Iterative Update. Based on the newly added detection samples and control feedback data in Step S5, the improved deep learning detection network and fuzzy PID control algorithm perform incremental training and parameter updates at a preset cycle. The overall performance of the online quality detection and control system continues to improve as the running time increases.

[0075] The above steps S1 to S4 are continuously executed in a loop during the operation of the online quality detection and control system, while steps S5 and S6 are executed synchronously and in parallel, thus forming a complete closed-loop operation path of "sensing and acquisition, defect detection, coupling analysis, closed-loop control, data traceability, and model iteration".

[0076] As an example of such a multi-process collaborative quality control scenario in intelligent manufacturing of industrial boilers, the following should be mentioned: In the welding process, weld images and welding parameters are collected simultaneously. An improved deep learning detection network identifies cracks and incomplete penetration defects in real time. A multi-factor coupling analysis module quantifies the influence weights of welding current and welding speed on the defect incidence rate. Based on this, an adaptive closed-loop control module automatically fine-tunes the welding parameters. The control results and defect data are completely archived, and this batch of data is included as incremental training samples during subsequent model updates. This closed-loop process runs through all quality control links in the forming process, assembly process, and boiler operation stage.

[0077] The steps of this implementation method are as follows: Step S1: Multi-source sensing and data acquisition; After the multi-source heterogeneous data real-time sensing and fusion acquisition module is started, various sensing units synchronously collect raw data from all dimensions of the boiler production site. After multi-level filtering and noise reduction, heterogeneous format unification and timestamp alignment, the fused data is output to the downstream module.

[0078] In a certain batch of production in 20XX, an industrial boiler production line (equipment number BLR-07) ​​was performing automated welding on the circumferential seam of the boiler drum. A high-definition industrial camera continuously acquired images of the weld area at a preset frame rate, a laser thickness sensor scanned the pipe wall thickness point by point, an ultrasonic flaw detector monitored the internal structure of the weld in real time, and temperature, pressure, and vibration sensors simultaneously collected operating data. Due to the high-temperature heat radiation and welding spatter present at the welding site, the image data underwent dual denoising in both the spatial and frequency domains, ensuring the noise level was controlled within a preset threshold. The analog sensor data underwent baseline drift correction and missing value interpolation to guarantee data integrity. All data streams were synchronized to a unified time base after timestamp alignment, forming a fused data stream that could be directly used by subsequent modules.

[0079] Z-score normalization is applied to various types of sensor data, including numerical sensor data such as welding current, wall thickness, and vibration amplitude. The normalization formula is as follows: Where x is the original observation value, This is the historical average. The historical standard deviation; image pixel data is mapped to a uniform numerical range after mean normalization.

[0080] Table 1. Raw data and preprocessing results of multi-source sensing:

[0081] Step S2, online defect detection; An improved deep learning detection network performs real-time inference on image and sensor data in fused data, and identifies and outputs the type, location, size and grade of various quality defects.

[0082] After the data fusion output in step S1, the preprocessed image of frame F-031 is fed into the improved deep learning detection network. The adaptive anti-interference processing unit first performs image texture separation on the frame, stripping the welding process texture from the structural components. Then, it performs adaptive histogram equalization on the structural components to suppress local overexposure caused by high-temperature radiation. Based on the synchronous vibration signal (amplitude 0.38 mm / s), it performs inverse displacement compensation on the image frame to correct the blurring deviation introduced by vibration.

[0083] The backbone feature extraction unit extracts feature maps layer by layer from the processed image. Shallow feature maps preserve weld edge details, while deep feature maps encode defect semantics. The channel spatial attention unit calculates channel attention weights for each level of feature map. Spatial attention weights ,through After weighting, the feature response of the defect area is enhanced, and the interference of the welding texture background is suppressed, where F is the original feature map. This is an element-wise multiplication method. The multi-scale feature fusion unit fuses shallow detail features and deep semantic features across scales, ensuring that micro-cracks and irregular pores are all within the effective receptive field. The detection output unit ultimately outputs the defect target's category, confidence level, localization bounding box, and size estimate. After inverse normalization transformation and sensor calibration parameter conversion, the bounding box yields the actual physical size of the defect in millimeters. The defect grading and evaluation unit classifies the detected defects according to an industry standard rule base.

[0084] Table 2 Defect Detection Output Results:

[0085] Step S3, Coupling Analysis and Trend Prediction; Grey relational analysis algorithm quantifies and evaluates the correlation weights between quality parameters, process parameters, and operating condition parameters. LSTM network calculates the correlation weights within a preset time window. The system predicts the quality fluctuation trend within the system, and the attribution analysis unit automatically traces the causes of detected deviations.

[0086] The defect diagnosis data (surface cracks, general grade) output in step S2 and the real-time process parameter data collected in step S1 are synchronously input into the multi-factor coupling analysis module. After Z-score standardization, the grey relational analysis algorithm uses the defect detection rate as the reference sequence and welding current, welding speed, welding zone temperature, vibration amplitude, and wall thickness deviation as comparison sequences to calculate the correlation degree of each comparison sequence relative to the reference sequence. The results show that the correlation degree of welding speed is the highest, followed by welding current, indicating that the main cause of crack defects in this batch is insufficient fusion due to excessively high welding speed.

[0087] The LSTM network takes multi-dimensional time-series data from the current batch and 12 historical time periods as input, and calculates the data within a preset time window. (Set to be 4 inspection cycles in the future) Output the predicted value of the defect detection rate. After the predicted value is denormalized and restored to the actual physical quantity, the predicted value of the defect detection rate in the 3rd prediction cycle exceeds the upper boundary of the quality qualified domain. The warning signal is triggered and transmitted to the adaptive closed-loop control module.

[0088] The attribution analysis unit combines the correlation weight ranking with the LSTM abnormal period location to attribute this deviation to the process parameter deviation (welding speed exceeding the set range), rather than equipment status deviation or environmental interference fluctuation.

[0089] Table 3: Ranking of Grey Relational Analysis Parameter Weights

[0090] Table 4: LSTM Quality Trend Prediction Results

[0091] Step S4, adaptive closed-loop control; The fuzzy PID control algorithm dynamically generates incremental adjustment instructions for each controlled parameter based on the defect classification results, associated weight data, and trend prediction data, and performs automatic correction or pushes manual handling according to the hierarchical control logic.

[0092] The attribution results in step S3 confirm that welding speed is the primary controlling factor, and the defect level is classified as general. Based on the hierarchical control logic, adjustment instructions are simultaneously pushed to the operation terminal for confirmation after generation, and early warning information is simultaneously displayed on the visualization platform.

[0093] Fuzzy PID control algorithm based on the current welding speed deviation value (The difference between the measured welding speed and the target welding speed, after Z-score standardization, is...) and the rate of change of deviation As input, the proportional coefficient is dynamically given by the fuzzy inference rules based on the membership degree between the two. Integral coefficient Differential coefficients Substitute into the control formula: ; in To normalize the control quantity, This is the integral variable. In the approximate calculation at the current time, the integral term takes the cumulative approximation of the current deviation value, and the output is the normalized control quantity. After de-standardization, the corresponding welding speed reduction is 12 mm / min. This instruction, confirmed by the operator, is then sent to the production line controller for execution. The welding current is automatically and finely adjusted, reduced by 8 A to match the adjusted heat input requirements. After the adjustment is executed, the defect detection rate in the next inspection cycle falls back to within the acceptable quality range, and the warning signal is lifted.

[0094] Table 5. Results of Control Command Generation and Execution

[0095] Step S5, Data Archiving and Traceability: All process data generated in steps S1 to S4 above are automatically recorded in the full lifecycle quality database and archived using equipment identification and process number indexes.

[0096] All process data for the circumferential weld of the BLR-07 boiler drum in this batch, including multi-source sensing raw data and preprocessing records, defect detection reports (DEF-001, DEF-002), correlation weight ranking table, trend prediction results, attribution analysis report, and control command execution records, are written into the full life cycle quality database using the equipment unique identifier BLR-07 and process number WLD-03 as indexes. Operation records of operators confirming control commands at the operating terminal are also archived simultaneously, meeting the traceability requirements for special equipment quality supervision. On the visualization platform, a heat map of weld defect distribution is overlaid on the boiler drum structural schematic diagram, with the crack location of DEF-001 highlighted, and the time-series change trend chart of quality parameters updated in real time, displaying the warning cancellation status.

[0097] Table 6. Full Lifecycle Quality Database Archive Index:

[0098] Step S6, model iterative update: Based on the newly added detection samples and control feedback data in step S5, the improved deep learning detection network and fuzzy PID control algorithm perform incremental training and parameter updates at a preset cycle, and the overall system performance continues to improve as the running data accumulates.

[0099] The two defect samples (DEF-001 surface crack and DEF-002 shallow porosity) generated in this batch, along with their corresponding annotation information and control feedback data, were automatically included in the training dataset. After the preset update cycle, the improved deep learning detection network underwent incremental training with the newly added samples. The network parameters were fine-tuned using the Adam optimization algorithm, with the weighted sum of class cross-entropy loss and bounding box regression loss as the optimization objective. This further improved the detection sensitivity for similar weld cracks in subsequent batches. The fuzzy PID control algorithm incrementally updated the weights of the fuzzy inference rules under the welding speed deviation scenario based on the actual decline in the defect detection rate after this control, making the control response more accurate and stable under similar deviation scenarios in subsequent batches. The above iterative update process was executed in parallel with the online operation of steps S1 to S4, without affecting the continuous operation of the production line.

[0100] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A quality online detection and control system for intelligent boiler manufacturing and artificial intelligence, wherein multi-source heterogeneous data is synchronously collected from boiler production processes and operating sites by distributed sensing units, and output to subsequent detection and analysis processes after preprocessing, characterized in that, Preprocessed image data and sensor data are input into an improved deep learning detection network for real-time inference. The improved deep learning detection network includes a backbone feature extraction unit, a channel spatial attention unit, a multi-scale feature fusion unit, and a detection output unit connected in sequence. The types, locations, sizes, and levels of various quality defects are identified and output. Defect diagnosis data, real-time process parameter data, and operating condition data are input into a multi-factor coupling analysis module. The grey relational analysis algorithm is used to quantify and evaluate the correlation weights between quality parameters, process parameters, and operating condition parameters. An LSTM network is used to predict the quality fluctuation trend within a preset time window. The defect classification results, associated weight data, and trend prediction data are input into the adaptive closed-loop control module. The fuzzy PID control algorithm dynamically generates incremental adjustment commands for each controlled parameter and executes automatic corrections or pushes manual intervention according to the hierarchical control logic. Specifically, the channel spatial attention unit calculates attention weights for each level of feature map output by the backbone feature extraction unit along both the channel and spatial dimensions. The attention-weighted feature map F is given by the following formula: F = αs αc F, where F is the input feature map, αc is the attention weight calculated along the channel dimension, and αs is the attention weight calculated along the spatial dimension. This indicates element-wise multiplication.

2. The online quality inspection and control system for intelligent boiler manufacturing and artificial intelligence as described in claim 1, characterized in that, The multi-scale feature fusion unit receives the attention-weighted shallow feature map Flow and the deep feature map Fhigh, and performs cross-scale fusion through bidirectional feature transfer via top-down and bottom-up paths. The fused feature map Ffuse is given by the following formula: Ffuse = upFhigh+ downFlow, where This indicates that the deep feature map is upsampled to the same spatial resolution as the shallow feature map, and then channel-aligned by a convolutional layer. This indicates that the shallow feature map is downsampled to the same spatial resolution as the deep feature map by stride convolution, and then channel alignment is performed by a convolutional layer.

3. The online quality detection and control system for intelligent boiler manufacturing and artificial intelligence as described in claim 1, characterized in that, The improved deep learning detection network is equipped with an adaptive anti-interference processing unit on its input side, and the adaptive anti-interference processing unit performs the following operations in sequence: Based on morphological filtering, the input image is decomposed into structural components and texture components. Only the structural components are used as the main input for subsequent detection, while the texture components are used as the input of the auxiliary reference channel. Adaptive histogram equalization is performed on the structural components to eliminate local overexposure. Based on the synchronous vibration signal collected by the vibration sensor, reverse displacement compensation is performed on the image frame sequence to correct the image blurring and displacement deviation caused by equipment vibration.

4. The online quality detection and control system for intelligent boiler manufacturing and artificial intelligence as described in claim 1, characterized in that, The positioning bounding box coordinates output by the detection output unit are relative coordinates normalized to the image size. After being inversely normalized and mapped to the actual pixel coordinates, they are then converted into actual physical dimensions in millimeters by combining the sensor calibration parameters. The defect grading and evaluation unit receives the defect type, defect size and defect location information output by the detection output unit. Based on the pre-written industry standard rule library, it automatically classifies each detected defect into four levels: minor, general, severe and fatal.

5. The online quality inspection and control system for intelligent boiler manufacturing and artificial intelligence as described in claim 1, characterized in that, The grey relational analysis algorithm uses defect detection rate, defect severity level, and wall thickness deviation as reference sequences, and welding current, welding speed, heating temperature, furnace negative pressure, air volume ratio, fuel ratio, and load fluctuation amplitude as comparison sequences. It outputs the correlation degree value of each comparison sequence relative to the reference sequence. The higher the correlation degree value, the greater the influence weight of the parameter on quality fluctuation. Each parameter in the reference sequence and comparison sequence is Z-score standardized before being input into the grey relational analysis algorithm.

6. The online quality detection and control system for intelligent boiler manufacturing and artificial intelligence as described in claim 1, characterized in that, The input to the LSTM network is a multidimensional time series extracted from real-time data streams and historical quality inspection databases. The output is the predicted quality parameter values ​​and corresponding confidence intervals for each time point within a preset future time window Tw, where Tw is the prediction step size preset according to the response delay of a specific process. The predicted quality parameter values ​​output by the LSTM network are normalized values ​​after standardization. After inverse standardization transformation, they are restored to the actual dimensional values ​​of the corresponding physical quantities and compared with the preset quality compliance domain boundary. When the predicted value exceeds the quality compliance domain boundary, an early warning signal is generated and transmitted to the adaptive closed-loop control module.

7. The online quality inspection and control system for intelligent boiler manufacturing and artificial intelligence as described in claim 1, characterized in that, The fuzzy PID control algorithm takes the current quality deviation value, the deviation change rate, and the parameter weight information output by the multi-factor coupling analysis module as inputs. The control output ut is given by the following formula: ut = Kp et + Ki0te d +Kd et, where et is the quality deviation value at time t, et is the deviation change rate at time t, Kp is the proportional coefficient, Ki is the integral coefficient, Kd is the differential coefficient, and Kp, Ki, and Kd are all dynamically given by the fuzzy inference rules based on the membership degree between et and et. The incremental adjustment output by the fuzzy PID control algorithm is a normalized value. After being denormalized and transformed back to the actual physical quantity increment of the corresponding controlled parameter, it is sent to the production line controller for execution.

8. The online quality detection and control system for intelligent boiler manufacturing and artificial intelligence as described in claim 4, characterized in that, The hierarchical control logic is divided into four levels based on the defect level and executed sequentially: When a defect is determined to be minor, the adjustment command is automatically generated by the fuzzy PID control algorithm and directly sent to the production line controller for execution. When a defect is determined to be of a general level, the adjustment instruction is pushed to the operation terminal for confirmation after it is generated, and the warning information is displayed on the visualization platform. When a defect is determined to be severe, the operation of the corresponding process is automatically locked, and the adjustment plan, along with the attribution analysis report, is pushed to the management terminal for handling decisions. When a defect is determined to be critical, the process and its downstream related processes are automatically isolated and shut down, and the alarm signal is broadcast throughout the entire process area.

9. The online quality inspection and control system for intelligent boiler manufacturing and artificial intelligence as described in claim 1, characterized in that, During continuous operation, newly generated defect detection samples and control feedback data are automatically collected and incorporated into the training dataset. The improved deep learning detection network and the fuzzy PID control algorithm perform incremental training and parameter updates at a preset update cycle. All process data is automatically recorded in the full lifecycle quality database and indexed and archived using unique equipment identifiers and process numbers.

10. A quality online detection and control system for intelligent boiler manufacturing and artificial intelligence, used to execute the quality online detection and control system for intelligent boiler manufacturing and artificial intelligence as described in any one of claims 1 to 9, characterized in that, include: The multi-source heterogeneous data real-time sensing and fusion acquisition module is designed to synchronously collect multi-dimensional raw data from boiler production processes and operation sites through distributed sensing units, and output fused data after multi-level filtering and noise reduction, heterogeneous format unification and timestamp alignment. An improved deep learning-based online defect detection module is designed to perform real-time inference on the fused data based on the improved deep learning detection network to identify and output the type, location, size, and grade of various quality defects. The multi-factor coupling analysis module is designed to receive defect diagnosis data, real-time process parameter data, and operating condition data, quantify the correlation weights through grey relational analysis algorithm, and predict the quality fluctuation trend within a preset time window through LSTM network. The adaptive closed-loop control module is designed to dynamically generate incremental control instructions based on defect classification results, associated weight data, and trend prediction data using a fuzzy PID control algorithm, and execute control according to the hierarchical management logic. The full-process data traceability and visualization management module is designed to receive and archive all process data from upstream modules and present the system's operating status and quality situation in a visual form. The modules are connected by data links to form a closed-loop feedback path.

Images