Unmanned production control method and system for building stainless steel pipe fittings

Abstract

The present application relates to the technical field of stainless steel pipe manufacturing, in particular to a kind of unmanned production control method and system of building stainless steel pipe fittings.The method includes obtaining target product model and reference process parameters, and loading quality potential energy function parameters from cloud model library;Multi-window scanning is carried out on the cut pipe blank, defect features are extracted and potential energy value is calculated, the pipe blank is divided into high potential energy, medium potential energy or low potential energy according to double threshold, and low potential energy pipe blank is removed;For high potential energy pipe blank, directly use reference process parameters for processing, for medium potential energy pipe blank, generate correction amount by gauss process regression model to modify reference process parameters for processing.The present application realizes fine evaluation of pipe blank by quality potential energy function, and generates individualized correction amount for medium potential energy pipe blank by gauss process regression model, which solves the problem that traditional production cannot execute differentiated regulation and control according to the actual quality state of single pipe blank.

Description

Technical Field

[0001] This invention relates to the field of stainless steel pipe fitting manufacturing technology, and more specifically, to an unmanned production control method and system for building stainless steel pipe fittings. Background Technology

[0002] Stainless steel pipe fittings for construction are widely used in fluid transport systems such as building water supply and drainage, fire protection, and compressed air. Their production mainly includes core processes such as fixed-length cutting, pipe end forming, CNC bending, and finished product inspection. They are important basic components of building engineering.

[0003] As the construction industry upgrades, the market demands increasing diversity in pipe fitting specifications, production flexibility, and consistent quality. Traditional production relies heavily on manual debugging and inspection, resulting in low changeover efficiency, poor quality stability, and fragmented data from independent equipment operation, making fully automated production impossible.

[0004] Currently, the industry mostly adopts a production mode of single-machine automated processing combined with manual sampling inspection. Each process equipment independently completes automated operations, while manual labor is used to complete specification changes, parameter settings, and defect screening. Standardized batch production of pipe fittings is achieved by using fixed and unified process parameters, which can meet the basic large-scale production needs of conventional pipe fittings.

[0005] However, this production model still has certain limitations: it can only make a binary qualitative judgment of tube blanks as qualified or unqualified, and cannot quantitatively identify and evaluate the subtle initial quality differences in qualified tube blanks. Therefore, in the production process, all tube blanks can only be processed with uniform fixed process parameters, and it is impossible to make differentiated process control based on the actual quality status of individual tube blanks. As a result, some edge tube blanks with slight defects cannot be repaired and used, which not only wastes materials, but also makes it difficult to ensure the overall quality consistency of the finished pipe fittings.

[0006] Therefore, there is an urgent need for a method and system for unmanned production control of stainless steel pipe fittings for building construction that can achieve refined quality assessment of pipe blanks and differentiated process control. Summary of the Invention

[0007] One objective of this invention is to provide an unmanned production control method for stainless steel pipe fittings in construction. This method quantitatively evaluates and grades each pipe blank by establishing a mass potential energy function. Based on the potential energy level, high-potential-energy pipe blanks are directly processed using benchmark parameters, while medium-potential-energy pipe blanks are processed differently by using a Gaussian process regression model to generate personalized correction values. This solves the problems mentioned in the background art, namely: Traditional production methods make it difficult to precisely quantify the initial quality state of tube blanks, and also make it impossible to implement differentiated process control based on the actual differences of individual tube blanks.

[0008] To achieve the above objective, the method includes the following steps: S1. Obtain the target product model and corresponding baseline process parameter vector in the production task. Load the parameters of the mass potential energy function from the cloud model library according to the target product model. The mass potential energy function is a logistic regression model with a four-dimensional feature vector as input. The four feature dimensions correspond to ellipticity deviation, pit depth, straightness deviation and surface roughness change, respectively. S2. Cut the raw material pipe to length to obtain the cut pipe blank. Perform multi-window scanning on the cut pipe blank, extract the ellipticity deviation, pit depth, straightness deviation, surface roughness change and confidence weight of each window, and obtain the normalized feature vector after weighted fusion and normalization. Substitute it into the mass potential energy function to calculate the potential energy value. Based on the preset high potential energy threshold and low potential energy threshold, the tube blank is divided into high potential energy tube blank, medium potential energy tube blank and low potential energy tube blank. The low potential energy tube blank is discarded and the high potential energy tube blank and medium potential energy tube blank are stored in the three-dimensional material silo. S3. Take out the tube blank from the three-dimensional silo and read the potential energy value and normalized eigenvector of the tube blank; For high-potential-energy tube blanks, the reference process parameter vector is directly used for tube end forming and bending. For a medium-potential tube blank, the normalized feature vector of the tube blank is concatenated with the reference process parameter vector and then input into a Gaussian process regression model. The reference process parameters are corrected by the correction amount output by the model to obtain the actual processing parameters. The actual processing parameters are then used for tube end forming and bending to obtain the finished tube fitting.

[0009] In the above technical solution, the traditional detection results are transformed into continuous and quantifiable potential energy values ​​by introducing a mass potential energy function, establishing a comparable quality benchmark for each tube blank. If decisions are made solely based on the original defect characteristics, the different dimensions of each characteristic and their coupling prevent the formation of a unified evaluation standard. Based on this, tube blanks are divided into three levels according to the potential energy value using a dual threshold system. This allows tube blanks in the intermediate state to be identified rather than simply categorized as qualified or unqualified. If this grading is omitted and all qualified tube blanks are processed using uniform parameters, tube blanks with minor defects will become scrap due to a lack of targeted adjustments. For medium-potential-energy tube blanks, a Gaussian process regression model is further used to generate personalized correction values. This model can output the prediction variance to characterize the reliability of the correction. Compared to fixed-parameter correction or single neural network correction, it utilizes the process patterns in historical data and avoids blind extrapolation in areas lacking training samples, which could lead to processing failures. This allows for differentiated process adaptation for each tube blank while ensuring safety.

[0010] Based on this, when extracting the confidence weight of each window, it is calculated in real time according to the ambient light intensity, the reflectivity of the tube blank surface and the quality of the sensor signal at the scanning time. During weighted fusion, the confidence weight of each window is used as the weighting coefficient to calculate the weighted average of similar features.

[0011] In another technical solution, the Gaussian process regression model adopts... The kernel function serves as the covariance function, and the predicted variance output by the model is used to characterize the degree of uncertainty of the current prediction. When the predicted variance exceeds a preset safety threshold, the correction is abandoned and the sample is recorded for subsequent model updates.

[0012] This technical solution introduces confidence weights to quantify the detection quality during multi-window scanning. This automatically reduces the weight of windows affected by illumination or reflection during fusion. If a simple equal-weighted average is used, the detection error of any window will be distributed across the entire billet feature set, leading to distorted potential energy assessment. The Gaussian process regression model is employed... The kernel function assumes that the mapping relationship between process parameters and quality has second-order differentiable smoothness, and compared with a fixed kernel function, it can better fit the actual characteristics of continuous physical processes such as pipe bending. The introduction of prediction variance solves the problem of blind correction in sparse data regions. When the input feature combination deviates from the distribution of existing training samples, the variance automatically increases. The system then abandons the correction and records the sample to expand the training set. Without this judgment mechanism, the model would forcibly output the correction amount in the parameter space lacking historical data, which may turn the originally acceptable billet into a scrap product. At the same time, it cannot identify data blind spots, causing the model iteration to stagnate.

[0013] The second objective of this invention is to provide an unmanned production control system for stainless steel pipe fittings for construction, including a central scheduling server, a cloud model library, a laser cutting unit, an online inspection unit, a three-dimensional material silo, an integrated pipe end forming and bending machine, a finished product three-dimensional precision inspection unit, and a finished product sorting robot. The central scheduling server is used to obtain the target product model and the baseline process parameter vector, load the mass potential energy function parameters from the cloud model library, and send each process parameter to the corresponding execution unit. The cloud model library is used to store the baseline process parameter vector, mass potential energy function parameter, historical best process parameter vector, and similarity association edge set between models for each pipe fitting model. The laser cutting unit is used to perform fixed-length cutting on raw material pipes; The online detection unit is used to perform multi-window scanning on the cut tube blank, extract defect features and confidence weights, calculate potential energy values, and classify the tube blank into high potential energy, medium potential energy and low potential energy according to the potential energy values, and discard low potential energy tube blanks. The three-dimensional silo is used to store high-potential-energy tube blanks and medium-potential-energy tube blanks, and each tube blank is bound with a potential energy value, grade label and normalized feature vector. The tube end forming and bending integrated machine is used to take out tube blanks from the three-dimensional silo. According to the potential energy level, for high potential energy tube blanks, a reference process parameter vector is used for tube end forming and bending. For medium potential energy tube blanks, the actual processing parameters after correction by the Gaussian process regression model are used for tube end forming and bending. The finished product three-dimensional precision inspection unit is used to perform three-dimensional scanning on the processed pipe fittings and output the comprehensive quality index and defect category; The finished product sorting robot is used to sort pipe fittings according to the comprehensive quality index.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a mass potential energy function with the normalized feature vector of the tube blank as input, transforming traditional qualitative detection results into continuous and quantifiable potential energy values. Based on dual thresholds, the tube blank is classified into high, medium, and low grades. This design enables the accurate identification of edge tube blanks in the intermediate state, rather than simply categorizing them as qualified or unqualified. It provides a clear target for the differentiated correction of medium-potential-energy tube blanks, solving the problem that existing technologies cannot perform refined quantitative evaluation of the initial quality of tube blanks.

[0015] 2. This invention introduces a Gaussian process regression model for generating personalized process parameter corrections for medium-potential tube blanks. This model employs... The kernel function fits the continuous mapping relationship between process parameters and finished product quality, and outputs the predicted variance to characterize the reliability of the correction. When the predicted variance exceeds the safety threshold, the system automatically abandons the correction and records the sample for model update. This design avoids processing failure caused by blind extrapolation in the parameter space where training samples are lacking. At the same time, it realizes the one-to-one adaptation of edge billets, breaking through the technical bottleneck that the existing uniform fixed parameter processing mode cannot respond to the quality differences of individual billets.

[0016] 3. This invention integrates quality potential energy assessment, graded flow distribution, and Gaussian process regression correction to form a closed-loop control architecture covering the entire process from billet inspection to processing execution. By completing potential energy calculation and grade classification during the inspection stage, and invoking differentiated decision logic based on potential energy levels during the processing stage, the system has the ability to evaluate each billet in real time and dynamically adjust process parameters. Combined with a self-learning update mechanism to continuously optimize the potential energy function threshold and regression model parameters, it realizes unmanned intelligent control of stainless steel pipe fitting production, moving from passive sampling to proactive sensing, and from uniform processing to piece-by-piece adaptation. Attached Figure Description

[0017] Figure 1 This is an overall flowchart of the unmanned production control method for stainless steel pipe fittings in buildings according to the present invention; Figure 2 This is a flowchart of step S2 of the present invention; Figure 3 This is a flowchart of step S3 of the present invention; Figure 4 This is a schematic diagram of the unmanned production control system for stainless steel pipe fittings for buildings according to the present invention. Detailed Implementation

[0018] 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.

[0019] Here are some explanations of technical terms: In this application, pipe diameter refers to the outer diameter of the pipe fitting, that is, the radial distance between the outer surfaces of the pipe wall.

[0020] Currently, traditional production methods struggle to precisely quantify the initial quality state of tube blanks and cannot implement differentiated process control based on the actual differences in individual tube blanks. One objective of this invention is to provide an unmanned production control method for stainless steel pipe fittings used in construction. See [link to relevant documentation]. Figure 1 As shown, it includes the following steps: S1. Obtain the target product model and corresponding baseline process parameter vector in the production task. Load the parameters of the mass potential energy function from the cloud model library according to the target product model. The mass potential energy function is a logistic regression model with a four-dimensional feature vector as input. The four feature dimensions correspond to ellipticity deviation, pit depth, straightness deviation and surface roughness change, respectively. S2. Cut the raw material pipe to length to obtain the cut pipe blank. Perform multi-window scanning on the cut pipe blank, extract the ellipticity deviation, pit depth, straightness deviation, surface roughness change and confidence weight of each window, and obtain the normalized feature vector after weighted fusion and normalization. Substitute it into the mass potential energy function to calculate the potential energy value. Based on the preset high potential energy threshold and low potential energy threshold, the tube blank is divided into high potential energy tube blank, medium potential energy tube blank and low potential energy tube blank. The low potential energy tube blank is discarded and the high potential energy tube blank and medium potential energy tube blank are stored in the three-dimensional material silo. S3. Take out the tube blank from the three-dimensional silo and read the potential energy value and normalized eigenvector of the tube blank; For high-potential-energy tube blanks, the reference process parameter vector is directly used for tube end forming and bending. For a medium-potential tube blank, the normalized feature vector of the tube blank is concatenated with the reference process parameter vector and then input into a Gaussian process regression model. The reference process parameters are corrected by the correction amount output by the model to obtain the actual processing parameters. The actual processing parameters are then used for tube end forming and bending to obtain the finished tube fitting.

[0021] By establishing a tube blank quality potential energy assessment model, the processing potential of each tube blank is quantitatively characterized, and differentiated parameter decisions and self-learning feedback are executed accordingly to achieve efficient and highly consistent control of tube fitting production under unmanned conditions.

[0022] Step S1 of this invention involves acquiring the production task and completing the initial parameter configuration. The central scheduling server receives the target product model and production quantity, retrieves data from the cloud model library according to the model code, generates a baseline process parameter package and mass potential energy function model parameters, and then distributes the parameters to each device. Specifically, it includes the following steps: S101: Deploy a central dispatch server in the production line monitoring room. After the system starts, the server automatically initializes the operating environment and loads the device communication protocol stack and database connection parameters.

[0023] The production equipment involved in this method includes a laser cutting unit, an online inspection unit, a three-dimensional silo, a pipe end forming and bending integrated machine, and a finished product sorting robot. The central scheduling server establishes communication links with each of the above devices through industrial Ethernet, and at the same time maintains a real-time connection with a graph database-type cloud model library deployed in a remote cloud.

[0024] Production tasks are issued by the upper-level Manufacturing Execution System (MES). The central scheduling server periodically accesses the task interface of the MES to obtain information on currently pending work orders. Each work order includes at least the target product model, planned production quantity, scheduled delivery time, and corresponding set of process parameters. The set of process parameters defines key values ​​such as cutting length parameters, pipe end forming stroke and holding time, bending speed, and clamping force.

[0025] After parsing the work order, the central dispatch server directly reads the target product model code and its corresponding set of process parameters from the work order, and temporarily stores the model code and work order parameters in the local cache, waiting for subsequent steps to call them.

[0026] S102: The central dispatch server reads the target product model code and the set of process parameters issued by the work order that were temporarily stored in step S101, and then initiates a retrieval request to the cloud model library based on the model code.

[0027] It should be noted that the cloud model library uses a graph database as its core architecture. For each type of pipe fitting, a dedicated data node is independently built. Each node uniformly stores the baseline process parameter vector, mass potential energy function parameters, and historical best process parameter vector. At the same time, it records the set of similarity association edges between each model and other pipe fitting models, which can provide complete data support for process parameter adaptation, model transfer learning, and self-learning iteration.

[0028] During the retrieval process, the central dispatch server uses the target product model code as the unique retrieval key to match the corresponding model data nodes within the cloud model library, and completes parameter loading and status marking according to the retrieval results and working conditions.

[0029] Specifically, if a matching model node is found, and the node has stored the historical best process parameter vector that has been iterated, the central scheduling server will simultaneously load two sets of parameter systems: the baseline process parameter vector issued by the manufacturing execution system work order, and the historical best process parameter vector stored in the cloud node.

[0030] The baseline process parameter vector is used for the initial process configuration of the current production batch to ensure normal production, while the historical best process parameter vector is stored locally as a benchmark for parameter comparison and process iteration optimization in the subsequent self-learning stage.

[0031] Correspondingly, if a matching model node is found, but this node only stores the basic baseline process parameter vector and does not retain the converged historical optimal process parameter vector, then it is determined that the model has not completed self-learning iterative optimization. In this case, the central scheduling server only loads the baseline process parameter vector corresponding to the work order and marks the model as pending optimization.

[0032] If no data node matching the target model code is found in the cloud model library, the fitting model is determined to be a newly launched production model. In this case, the central scheduling server only loads the baseline process parameter vector issued by the manufacturing execution system as the initial production parameters. At the same time, it automatically creates an independent data node for this new model in the cloud model library and marks it as a newly launched production model, preparing to generate the initial mass potential energy function parameters through transfer learning in subsequent steps.

[0033] After completing all parameter retrieval and loading, the central scheduling server caches the final confirmed baseline process parameter vector locally. Simultaneously, based on the operational needs of each production unit, the parameters are precisely broken down and distributed in a targeted manner: the cutting length parameter is distributed to the laser cutting unit, the mass potential energy function parameter is distributed to the online detection unit, the forming and bending-related process parameter components are distributed to the pipe end forming and bending integrated machine, and the grading threshold parameter is distributed to the finished product sorting robot.

[0034] S103: Based on the model determination results of S102, when the central scheduling server confirms that the target fitting model is being produced for the first time and there is no corresponding model node in the cloud model library, the system automatically starts the migration learning process. This process aims to match suitable initial process and model parameters for the newly launched model based on the production data of historical mature models, effectively solving the problem of blank parameters in the initial production of new models, avoiding batch production quality fluctuations caused by initializing parameters from zero, and ensuring production stability in the early stages of new model production.

[0035] Specifically, the central dispatch server first retrieves the node data of all pipe fitting models that have completed mass production from the cloud model library and extracts the standardized feature vectors corresponding to each model.

[0036] It should be noted that this standardized feature vector contains five core dimensions: material grade code, pipe diameter, pipe wall thickness, bending radius to pipe diameter ratio, and end shape type code.

[0037] Among them, the material grade code and the end shape type code are non-numerical features, which are converted into standardized values ​​through serial number coding or unique thermal coding; the pipe diameter, wall thickness and bend ratio are all directly adopted from the actual physical values.

[0038] Based on this, the central scheduling server performs unified normalization on all extracted model feature vectors, mapping the values ​​of each dimension to the [0, 1] interval to eliminate dimensional differences and provide a unified calculation benchmark for subsequent similarity calculations.

[0039] Furthermore, the central scheduling server constructs the feature vector of the new model of the current target according to the same dimensional rules. And ensure that its dimension is completely consistent with the feature vectors of all existing models. Then, the central scheduling server calculates the matching similarity between the target model and each produced model sequentially using a cosine similarity algorithm. The formula for calculating the cosine similarity is: ; In the formula, The feature vectors are for existing manufactured models.

[0040] The numerator of the formula is the dot product of the two sets of feature vectors, and the denominator is the product of the magnitudes of the two sets of feature vectors. The higher the cosine similarity value of the final output, the closer the process characteristics and production rules of the two models are.

[0041] After completing all similarity calculations, the central scheduling server automatically selects the existing model with the highest cosine similarity as the most similar reference model for the target new model. If multiple models have the same maximum similarity value, the system prioritizes the model with the largest total production batch or the best historical average quality index as the final reference benchmark, ensuring the reliability and universality of the migration data.

[0042] After identifying the most similar model, the central scheduling server retrieves the mature mass potential energy function parameters stored for that model from the cloud model library. This set of parameters includes two core model parameters: weight vectors and bias terms. The central scheduling server directly replicates this set of parameters and assigns them as the initial mass potential energy function parameters for the new model, thus completing the rapid initialization of the new model's potential energy model.

[0043] At the same time, the central scheduling server, based on the benchmark process parameters of the most similar model and combined with the physical size differences between the old and new models, completes the linear proportional mapping adaptation of key process parameters.

[0044] Specifically, using the pipe diameter ratio between the old and new models as the correction benchmark, core process parameters such as cutting length, bending speed, clamping force, and forming stroke are converted and corrected to obtain initial benchmark process parameters adapted to the new model. The core mapping formula is as follows: ; ; In the formula, The initial cutting length for the target new model; The cutting length for the most similar model; The target is the diameter of the new model pipe; The diameter of the pipe of the most similar model; The initial bending speed for the target new model; The bending speed is the most similar model.

[0045] The cutting length is linearly mapped to the pipe diameter to ensure that the blanking size matches the pipe specifications; the bending speed is inversely mapped to the square root of the pipe diameter ratio, so that larger diameter pipes are matched with lower bending speeds, reducing bending stress and wrinkle defects, and improving the stability of the initial forming quality of the new model.

[0046] All other forming and bending process parameters are mapped to the same logical proportions according to the corresponding physical relationships, ensuring that all migration parameters conform to the physical structure and production characteristics of the new model.

[0047] After parameter mapping and model initialization are completed, the central scheduling server improves the data node information of the current new model in the cloud model library, stores the generated initial baseline process parameter vector and initial mass potential energy function parameters into the node, and marks the node status of the new model as migration initialization completed.

[0048] At this point, the new model migration learning process is complete. The system will use the complete set of initialization parameters generated this time as the formal production baseline parameters, replacing the blank parameter configuration, for intelligent production control in subsequent processes.

[0049] S104: After completing the loading and generation of parameters for the current model, the central scheduling server constructs the mass potential energy function required for the current production batch based on the matching parameters of the current model.

[0050] It should be noted that the mass potential energy function is a logistic regression model that takes the initial inspection characteristics of the tube blank as input and constrains the output range to 0~1. It is used to quantitatively evaluate the forming potential and potential processing quality of a single tube blank, providing accurate model support for subsequent tube blank grading and parameter adaptive correction.

[0051] Specifically, the central dispatch server first retrieves the quality potential energy function parameters corresponding to the currently produced model from the cloud model library. The parameter set includes two types of core model parameters: weight vector w and bias term b. The model weight vector w is used to characterize the influence weight of each detection feature on the quality of the finished product, and the model bias term b is used to correct the overall output benchmark of the model.

[0052] At the same time, the online detection normalization parameters for this model are retrieved, covering the historical mean μ and historical standard deviation σ of each detection feature dimension, which are used to perform real-time normalization processing on the original feature data of the tube blank in subsequent steps.

[0053] The specific mathematical expression for the mass potential energy function is as follows: ; In the formula, This represents the mass potential energy value corresponding to the tube blank. The closer the value is to 1, the higher the forming quality potential of the tube blank. The normalized four-dimensional tube blank initial inspection feature vector specifically includes the normalized ellipticity deviation, maximum pit depth, straightness deviation, and surface roughness variation value.

[0054] The physical meaning of each dimension is as follows: Ellipticity deviation is used to characterize the degree to which the cross-section of the tube blank deviates from a standard circle; The maximum pit depth is used to characterize the defect depth of scratches and indentations on the surface of the tube blank; Straightness deviation is used to characterize the degree of axial bending deformation of the tube blank as a whole; The surface roughness variation value is used to characterize the fluctuation range of the tube blank surface roughness relative to the standard working condition, and comprehensively covers the appearance and morphological defect characteristics of the tube blank.

[0055] After completing the construction of the mass potential energy function and parameter initialization, the central scheduling server, based on the operational needs of each production unit, completes the precise breakdown and targeted pre-distribution of all process parameters. The distribution rules are as follows: ① The fixed-length cutting parameters are sent to the laser cutting unit to control the precise fixed-length cutting operation of long tubes; ② The weights of the mass potential energy function, the bias parameters, and the normalization parameters of each feature are sent to the online detection unit, so that the equipment can complete real-time feature calculation, mass potential energy assessment, and grade classification based on the original feature data of the billet. ③ The core process parameters such as pipe end forming stroke, pressure holding time, bending speed, clamping force, and mandrel extension are uniformly distributed to the pipe end forming and bending integrated machine to accurately control the subsequent forming and bending processes. ④ The quality grading threshold parameters of finished products are sent to the finished product sorting robot to control the judgment criteria for subsequent intelligent sorting operations of finished products.

[0056] At this point, the central dispatch server has completed all pre-production parameter configuration and enters the online production execution phase.

[0057] like Figure 2 As shown, step S2 of this invention is online pipe segment detection and dynamic potential energy assessment, used to collect multi-dimensional features of each cut pipe blank, and calculate its potential energy value based on the mass potential energy function constructed in step S1, thereby completing the grading and screening of pipe blanks and material diversion. Specifically, it includes the following steps: S201: The laser cutting unit receives the fixed-length cutting parameters from the central dispatch server in real time, which serve as the baseline control parameters for the cutting of this batch of pipes. The raw material pipes are automatically and orderly fed by the feeding mechanism and smoothly transported to the laser cutting station. The laser cutting unit detects the pipe arrival signal at the station in real time, and automatically starts the preset cutting execution program after confirming that the pipe clamping and positioning are complete.

[0058] The laser cutting unit performs fully automatic fixed-length cutting operations strictly according to the received fixed-length cutting parameters, controlling the pipe cutting size accuracy within ±0.5mm throughout the process to ensure the consistency of the size of each pipe section.

[0059] During the cutting process, the laser cutting unit has a built-in residual material optimization algorithm that links the remaining length of the current raw material with the data of the work orders to be cut in the system background in real time, and dynamically adjusts the execution order and cutting ratio of each cut.

[0060] Specifically, the scrap optimization algorithm takes minimizing the length of raw material scrap as its core optimization objective, traverses and calculates various cutting combination schemes, selects the cutting strategy with the best material utilization rate, and implements it. After a single raw material cutting is completed, if the remaining pipe length cannot meet the cutting requirements of the next standard pipe segment, the laser cutting unit automatically marks the remaining pipe segment as scrap, and simultaneously records information such as the scrap size and pipe batch number, caches and retains it for use in subsequent production batches.

[0061] After all the cutting operations on a single raw material pipe are completed, the unloading mechanism at the laser cutting station automatically pushes the cut, qualified pipe segments to the downstream conveyor belt. The conveyor belt starts operating simultaneously, smoothly transporting the cut pipe segments to the subsequent online inspection unit, awaiting subsequent defect detection and quality assessment.

[0062] At the same time, the laser cutting unit uploads data such as the actual cutting length, residual material information, and pipe batch number to the central dispatch server to complete data archiving and material traceability records.

[0063] S202: The cut pipe segments move smoothly along the conveyor belt and sequentially enter the inspection work area of ​​the online inspection unit. The online inspection unit has four line laser profilometers evenly spaced above the conveyor belt, and each line laser profilometer independently completes the surface scanning operation of its corresponding pipe segment.

[0064] During the inspection process, the online inspection unit automatically matches and calculates the scanning window length of a single profilometer based on the pipe diameter of the pipe section to be tested. Specifically, the scanning window length adopts a dynamic adaptive calculation method, and the calculation formula is as follows: ; In the formula, This is the length of the profilometer's real-time scanning window; α is a fixed proportionality coefficient with a value range of 3 to 5. This value is determined based on engineering experience in pipe fitting processing and inspection. It can be adapted to the contour inspection needs of most pipe specifications. It can ensure the scanning coverage density of small-diameter pipe sections and guarantee the detection accuracy, while limiting the scanning range of large-diameter pipe sections and avoiding invalid data collection. The diameter of the pipe section to be tested.

[0065] This dynamic window configuration method ensures sufficient detection resolution for small-diameter pipe sections while avoiding redundant data generated during the detection of large-diameter pipe sections, thus balancing detection accuracy and efficiency.

[0066] The online detection unit controls four line laser profilometers to simultaneously start scanning operations. Each profilometer performs 360° full-area profilometry acquisition on the pipe section surface within its corresponding dynamic scanning window. After scanning, five types of detection values ​​are extracted from each window: ellipticity deviation, maximum pit depth, straightness deviation, surface roughness change, and the confidence weight corresponding to the current window. All the acquired values ​​together constitute the original feature data of the pipe blank.

[0067] Specifically, the online detection unit dynamically calculates the confidence weight of each window by combining the ambient light intensity, the reflectivity of the pipe section surface, and the real-time signal quality of the sensor at the scanning time. This parameter is used to characterize the reliability and effectiveness of the single-window detection results. After all windows have completed the acquisition of the original feature data of the pipe blank, the online detection unit temporarily stores all feature data in the local cache of the device, providing complete data support for subsequent comprehensive feature calculations and quality potential assessment of the entire pipe section.

[0068] S203: The online detection unit reads the raw feature data collected from four windows from the local cache, each window containing ellipticity deviation. Maximum pit depth Straightness deviation Surface roughness variation and the confidence weight of the window. subscript This represents the window number, with a value ranging from 1 to 4.

[0069] Furthermore, the online detection unit uses a weighted average method to fuse similar features from the four windows and calculate the comprehensive original feature value of the entire tube blank. The weighted average formula is as follows: , , , ; The denominator is the sum of the confidence weights of the four windows, and the numerator is the sum of the products of the feature value of each window and its corresponding confidence weight. This weighting method allows the detection results of high-confidence windows to occupy a higher proportion in the comprehensive feature, while the detection results of low-confidence windows are suppressed accordingly, thereby improving the reliability and accuracy of the comprehensive feature.

[0070] After completing the feature fusion calculation, the comprehensive original feature vector of the tube blank is obtained. .

[0071] Based on this, the online detection unit calls the normalization parameters issued in step S104, including the historical mean μ1, μ2, μ3, μ4 and historical standard deviation σ1, σ2, σ3, σ4 of each feature dimension, and performs normalization calculation on each component in the comprehensive original feature vector. The normalization formula is: ; Subscript =1, 2, 3, 4 correspond to the four feature dimensions: ellipticity deviation, maximum pit depth, straightness deviation, and surface roughness variation, respectively.

[0072] After normalization, we obtain a dimensionless input feature vector. Each feature component is mapped to a uniform data distribution range, eliminating the impact of original dimensional differences on subsequent model calculations.

[0073] The online detection unit will normalize the feature vector. It is temporarily stored in the local cache and bound to the unique identifier of the pipe segment for use in the next step of mass potential energy calculation and grade determination.

[0074] S204: The online detection unit obtains the feature vector after normalization in step S203. And the mass potential energy function parameter weight vector issued in step S104 And the bias term b.

[0075] The online detection unit will use feature vectors Substitute into the logistic regression model to calculate the current mass potential energy value of the billet. The calculation formula is as follows: ; In the formula, .

[0076] Calculated mass potential energy value The value is a continuous value between 0 and 1. The closer the value is to 1, the higher the probability that the tube blank will become a high-quality finished product after processing with the reference process parameters; the closer the value is to 0, the lower the probability that the tube blank has serious defects and will become a qualified product after processing.

[0077] Furthermore, the online detection unit, based on The numerical values ​​are used to determine the grade of the tube blank. The system presets two thresholds: a high potential energy threshold... and low potential threshold .in The initial setting is 0.75. The initial setting is 0.40.

[0078] It should be noted that this threshold was determined based on tube blanks collected from historical production batches. The statistical analysis of the distribution relationship between numerical values ​​and corresponding finished product quality indices revealed the following: when When the value is greater than 0.75, the finished product qualification rate of the tube blank processed using the benchmark process parameters exceeds 95%; when When the value is less than 0.4, regardless of the parameter correction scheme used, the finished product qualification rate of the tube blank is less than 30%. For tube blanks falling between these two categories, the pass rate can be increased to over 80% after parameter correction.

[0079] Based on the above statistical patterns, 0.75 and 0.4 were determined as the dividing thresholds between high potential energy and low potential energy, respectively.

[0080] The two thresholds will be dynamically adjusted in subsequent self-learning steps based on the quantile distribution of the latest production data to continuously adapt to the actual quality distribution characteristics of the current production line.

[0081] Based on this, the judgment rules are as follows: ①If If so, the tube blank is marked as a high-potential-energy tube blank; ②If If so, the tube blank is marked as a medium potential energy tube blank; ③If If so, the tube blank is marked as a low potential energy tube blank.

[0082] The online detection unit will calculate the results. The value, grade determination result, and the billet's unique identifier are associated and bound, and stored in a local cache. Then, this information, along with the normalized feature vector, is... They are then transmitted to the automated material storage management system for subsequent diversion and material caching operations.

[0083] S205: After the online inspection unit completes the billet grade determination, it performs the corresponding diversion operation based on the determination result. Specifically, a pneumatic pusher assembly is installed at the end of the conveyor belt. This assembly communicates with the online inspection unit in real time and receives the grade determination signal for each billet.

[0084] When a low-potential-energy tube blank reaches the end of the conveyor belt, the online detection unit sends a rejection command to the pneumatic pusher. The pneumatic pusher immediately moves, pushing the tube blank from the side of the conveyor belt into the scrap bin, preventing it from entering subsequent processing steps. The scrap bin is equipped with a counting sensor that automatically records the number of rejected tube blanks and uploads the data to the central dispatch server.

[0085] When the high-potential-energy tube blank or the medium-potential-energy tube blank reaches the end of the conveyor belt, the pneumatic pusher remains stationary, and the tube blank continues to be conveyed forward along the conveyor belt until it reaches the entrance position of the three-dimensional silo.

[0086] The automated storage and retrieval system (AS / RS) employs a multi-layer racking structure, with each layer divided into sections based on the diameter and length of the tube blanks. Stacking robots are deployed at the AS / RS entrance, responsible for grabbing the tube blanks arriving at the entrance and stacking them to their designated locations. The stacking robots obtain the location allocation instruction corresponding to the current tube blank from the central scheduling server. This instruction includes the layer number, column number, and depth coordinates of the target location.

[0087] After receiving the storage location instruction, the stacking robot grabs the tube blank and transports it to the target storage location according to the storage location allocation instruction, completing the stacking operation. After stacking, the stacking robot sends an inbound confirmation signal to the storage management system. Upon receiving the confirmation signal, the storage management system updates the inventory information in real time, marks the storage location as occupied, and records the unique identifier of the tube blank occupying the storage location.

[0088] Meanwhile, the silo management system establishes a data binding record for each incoming tube blank, and the associated information includes: the tube blank's unique identifier and potential energy value. Level labels (high potential energy or medium potential energy), normalized feature vectors The data includes original feature data and the timestamp of entry into the warehouse. This data is stored in the local database of the warehousing management system and synchronized to the central scheduling server via industrial Ethernet for use in subsequent forming and bending processes.

[0089] The central dispatch server dynamically schedules the stacking robot to retrieve tube blanks from the silo based on the production cycle and buffer status of the downstream forming and bending unit. When the tube end forming and bending integrated machine issues a material retrieval request, the central dispatch server queries the silo management system, selects the optimal tube blank for retrieval based on the tube blank grade, warehousing time, and silo distribution, and sends the retrieval instruction and target location coordinates to the stacking robot. The stacking robot performs the material retrieval operation, removes the tube blank from the location, places it on the conveyor device, and sends it to the loading station of the tube end forming and bending integrated machine.

[0090] like Figure 3 As shown, step S3 of this invention is a potential energy-driven adaptive correction and processing of process parameters. This involves retrieving a tube blank from the three-dimensional silo that has undergone potential energy assessment, executing differentiated process parameter decisions based on its potential energy level, and performing personalized parameter corrections for medium-potential-energy tube blanks, ultimately completing tube end forming and bending. Specifically, it includes the following steps: S301: After the pipe end forming and bending integrated machine completes the processing of the current pipe fitting, it sends a material picking request to the central dispatch server. After receiving the material picking request, the central dispatch server immediately queries the material storage management system and selects the pipe blank to be processed from the three-dimensional material storage.

[0091] It should be noted that the selection rules for the billet by the central dispatch server are as follows: Prioritize selecting high-potential-energy tube blanks with the longest storage time; If there are no high-potential-energy tube blanks in the silo, then select the medium-potential-energy tube blank with the longest storage time.

[0092] The billet selection rule aims to balance the buffering time of the billet in the silo and avoid surface quality deterioration caused by long-term storage.

[0093] After the central dispatch server selects the tube blank, it reads the storage location coordinates of the tube blank in the silo and sends the coordinates to the stacking robot. Upon receiving the coordinate command, the stacking robot automatically moves to the target location and performs the material handling operation. The stacking robot removes the tube blank from the location and places it on the conveyor at the silo outlet. The conveyor then starts, transporting the tube blank to the loading station of the tube end forming and bending integrated machine.

[0094] After the tube blank arrives at the loading station, the tube end forming and bending integrated machine sends a reading command to the barcode reader at the loading station. The barcode reader scans the unique identification code pasted or sprayed on the surface of the tube blank to obtain the identification code information of the tube blank. The tube end forming and bending integrated machine sends this identification code to the material storage management system to request the retrieval of the complete production record of the tube blank.

[0095] After receiving the request, the silo management system retrieves the local database based on the identification code and returns the information binding record for the tube blank. This information binding record must contain at least the following: the potential energy value of the tube blank. Level labels (high potential energy or medium potential energy), normalized feature vectors Original feature data and entry timestamp.

[0096] The pipe end forming and pipe bending integrated machine will read the potential energy value. Level labels and normalized feature vectors Stored in a local cache as a basis for subsequent process parameter decisions.

[0097] S302: The pipe end forming and bending integrated machine reads the grade label of the pipe blank obtained in step S301, and performs the following differentiated process parameter decisions based on the grade label: ① If the tube blank grade is high potential energy, the tube end forming and bending integrated machine directly uses the reference process parameter vector issued by the central scheduling server in step S104 as the process parameters for this processing, without performing any additional parameter corrections.

[0098] Here, the reference process parameter vector is set as... It includes parameters such as cutting length, pipe end forming stroke, pressure holding time, bending speed, clamping force, and mandrel extension.

[0099] The basis for choosing this branch is: the normalized eigenvector of the high-potential tube blank. It has been shown that its initial quality is excellent, and a high finished product qualification rate can be obtained by processing with the benchmark parameters. The marginal benefits brought by additional corrections are limited. Therefore, the benchmark parameters are directly used to simplify the control process.

[0100] ② If the tube blank grade is medium potential energy, the trigger parameter correction process for the tube end forming and bending integrated machine is as follows: Normalized feature vector of medium potential energy tube blank. This indicates that its initial quality state is in the critical range. Directly using the benchmark parameters for processing may result in the finished product not meeting the quality standards. It is necessary to make personalized parameter adjustments based on the specific defect characteristics of the current tube blank.

[0101] The pipe end forming and pipe bending integrated machine first normalizes the feature vector. Compared with the reference process parameter vector The vectors are concatenated to form the input vector of the modified model. Input vector The dimension is the sum of the feature vector dimension and the process parameter dimension, where the feature vector dimension is 4, and the process parameter dimension is determined according to the actual configuration.

[0102] After the splicing is completed, the pipe end forming and bending integrated machine will input the vector. The data is passed to the Gaussian process regression model, and we wait for the model to output the parameter correction values.

[0103] S303: After the potential energy tube blank triggers the parameter correction process, the tube end forming and bending integrated machine calls the built-in Gaussian process regression model to correct and calculate the process parameters of the current tube blank.

[0104] Specifically, the Gaussian process regression model is deployed in the edge controller of the pipe end forming and bending integrated machine. This model has been trained offline using historical production data. The vector... Input a Gaussian process regression model and output the process parameter correction. The dimension of the correction amount and the baseline process parameter vector Maintain consistency.

[0105] It should be noted that the Gaussian process regression model uses The kernel function serves as the covariance function. The kernel function, based on the assumption of double-differentiability and smoothness, can effectively describe the continuous mapping relationship between pipe fitting processing parameters and finished product quality. For the input vector... Correction amount of model prediction output It follows a Gaussian distribution, and its predicted mean is and prediction variance The calculation formula is as follows: ; ; In the formula, The covariance matrix of the training samples; This is the covariance vector between the test input and each training sample; The correction labels corresponding to the training samples; The covariance value of the input is used for testing.

[0106] Forecast Mean The representative value of the correction that the model considers most likely is the prediction variance. The variance represents the degree of uncertainty of the model regarding the current prediction result. The larger the variance, the more likely the model lacks sufficient training samples to support it in that region of the input space.

[0107] Based on this, the prediction variance is compared with a preset safety threshold. The safety threshold is set to 0.1, a value determined by statistical analysis of the relationship between model prediction accuracy and variance distribution in historical production data. Since the prediction variance of the model is typically below 0.05 in regions with dense training data, but often exceeds 0.15 in regions with sparse data, 0.1 is chosen as the cutoff value to distinguish between reliable and unreliable predictions.

[0108] If the prediction variance If the prediction result is deemed reliable, the model prediction is considered credible. The pipe end forming and bending integrated machine uses the predicted mean as the actual correction value to calculate the actual processing parameters. : ; If the prediction variance If the model's prediction for the current tube blank is deemed too uncertain, this situation typically occurs when the feature combination of the current tube blank differs significantly from the samples in the training sample set. In this case, directly applying model correction may introduce incorrect parameter adjustments, leading to a decrease in processing quality. To avoid the risks associated with model extrapolation, the tube end forming and bending integrated machine abandons correction and directly uses the baseline process parameters for processing. ; Simultaneously, the pipe end forming and bending integrated machine will input the blank's vector... Prediction variance The trigger records for abandoning corrections are packaged and uploaded to the central scheduling server. This data will be included in the candidate sample set for the next round of model updates to expand the training coverage of the model in sparse data regions and improve the model's subsequent prediction ability for this type of billet.

[0109] Complete actual processing parameters After calculation, the pipe end forming and bending integrated machine stores the results in a local cache and proceeds to the subsequent pipe end forming processing stage.

[0110] S304: After the pipe end forming and bending integrated machine completes the process parameter decision of step S303, it positions the pipe blank to the first station of the two-station turntable, namely the pipe end forming station. This station is equipped with a hydraulically driven flaring mold set and a shrinking mold set. The two sets of mold sets are installed independently and can be automatically switched according to processing requirements.

[0111] It should be noted that the mold assembly is installed on the mold base at the molding station using electromagnetic adsorption. When a mold needs to be replaced, the system cuts off the power supply to the electromagnetic adsorption device, the mold base loses its magnetism, and the operator or automatic mold changing device can remove the current mold. When installing a new mold, after placing the mold in the corresponding position on the mold base, the system restores the power supply to the electromagnetic adsorption device, and the mold base generates magnetic force to firmly adsorb and fix the mold. This electromagnetic adsorption structure keeps the mold replacement time within 30 seconds, significantly reducing the time spent on mold changeover operations.

[0112] Pipe end forming and pipe bending integrated machine according to actual process parameters The forming parameters in the code are used to perform the tube end forming operation. The forming parameters include the flaring or shrinking stroke and the holding time, with the forming stroke control accuracy being ±0.05 mm.

[0113] The hydraulic drive unit is activated, pushing the mold towards the end of the tube blank. After the mold contacts the end of the tube blank, it continues to apply pressure to the end of the tube blank according to the set stroke, causing plastic deformation of the material at the end of the tube blank, gradually forming the desired flared or constricted shape. During the forming process, a displacement sensor installed in the hydraulic cylinder monitors the mold stroke in real time. When the stroke reaches the preset position, the displacement sensor sends a position signal to the controller.

[0114] Upon receiving the arrival signal, the controller immediately sends a pressure-holding command to the hydraulic drive unit. The hydraulic drive unit stops feeding but maintains the current pressure, and the system enters the pressure-holding phase. The pressure-holding time can be set from 0.5 to 1 second, with the specific value determined by the process parameters. The holding time parameter is determined. The purpose of holding pressure is to keep the material at the end of the tube blank stable after forming, reduce the amount of springback after forming, and ensure the accuracy of the end dimensions.

[0115] After the holding pressure time is reached, the controller sends a retraction command to the hydraulic drive unit. The hydraulic drive unit drives the mold to move in the opposite direction, gradually separating it from the end of the tube blank and returning it to its initial position. At this point, the cold forming process of the tube end is completed, and the end of the tube blank has formed a flared or narrowed shape that meets the process requirements. The tube end forming and bending integrated machine records the actual stroke, actual holding pressure time, and other data of this forming process to the local log, and then prepares to enter the turntable switching stage.

[0116] S305: After the pipe end cold forming process is completed, the pipe end forming and bending integrated machine starts the turntable switching program. The two-station turntable carries the pipe blank with the completed end forming, rotates 180 degrees in the preset direction, and sends the pipe blank from the first station to the second station, that is, the CNC bending station.

[0117] During the rotation of the turntable, an angle encoder installed at the bottom of the turntable monitors the rotation angle in real time. When the rotation angle reaches 180 degrees, the angle encoder sends a position signal to the controller, which then issues a stop command, and the turntable stops rotating.

[0118] After the turntable stops, the positioning and locking mechanism automatically starts to lock the relative position between the turntable and the machine tool base, ensuring that the position reference of the tube blank is consistent with the tube end forming process in the subsequent tube bending process, and avoiding processing deviations caused by the turntable loosening.

[0119] After the tube blank reaches the CNC tube bending station, the clamping mechanism of the tube bending machine clamps according to the actual process parameters. The clamping force parameter in the settings is used to perform the clamping operation. The clamping mechanism consists of two clamping molds, which move towards the tube blank under hydraulic drive, firmly fixing the tube blank in the cavity of the bending die.

[0120] The clamping force parameter can be set from 1 to 5 kN, and the specific value is determined by the process parameters. If the clamping force is too small, the tube blank will slip during the bending process. If the clamping force is too large, it may cause indentation or deformation on the surface of the tube blank.

[0121] After the clamping mechanism completes the fixing of the tube blank, the mandrel drive device is started, and the mandrel is moved according to the actual process parameters. The mandrel extension parameter is fed into the tube blank.

[0122] It should be noted that the mandrel has a flexible multi-ball head structure, and its outer diameter matches the inner diameter of the tube blank. It is used to support the inner wall of the tube blank during bending, preventing the inner wall from collapsing or wrinkling defects during bending deformation. The mandrel extension parameter is automatically calculated based on the tube blank diameter and bending radius, and is usually set to a position where the mandrel tip reaches 2 to 5 mm before the bending start point.

[0123] Furthermore, the pipe bending machine initiates the bending action, with the bending arm rotating around the bending center under the drive of a servo motor. The bending speed is determined according to the actual process parameters. The bending speed parameters are set within a range of 3 to 25 degrees per second. During bending, an angle encoder mounted on the rotating shaft of the bending arm acquires the bending angle in real time and feeds the measured angle value back to the controller. The controller performs a closed-loop comparison between the measured angle and the target angle. When the measured angle approaches the target angle, the system automatically reduces the bending speed to minimize inertial overshoot; when the measured angle reaches the target angle, the controller immediately issues a stop command, and the bending action stops. The angle control accuracy is ±0.5 degrees.

[0124] During the bending process, pressure sensors mounted on the mandrel tie rod monitor the axial tensile force and radial pressure on the mandrel in real time. When abnormal resistance is detected, it indicates that there may be interference or poor lubrication between the mandrel and the inner wall of the tube blank. The system will automatically fine-tune the mandrel extension or pause bending and issue an alarm signal according to the preset adjustment strategy, waiting for manual intervention or automatic troubleshooting.

[0125] Once the bending angle reaches the target value, the bending arm stops moving and remains in its current position, holding the pressure for 0.5 seconds to release residual stress after bending. After the pressure holding period, the clamping mechanism releases, and the left and right clamps retract to both sides, releasing the fixation on the tube blank. The mandrel drive device moves in the opposite direction, pulling the mandrel out of the tube blank and returning it to its initial position.

[0126] The pipe end forming and pipe bending integrated machine packages and uploads the actual bending speed, actual bending angle, clamping force curve, mandrel position and pressure sensor data of this pipe bending process to the central dispatch server and stores them in the production database.

[0127] S306: After the pipe bending process is completed, the pipe end forming and pipe bending integrated machine ejects the pipe from the mold, and the pipe slides down onto the discharge conveyor belt by gravity. The conveyor belt starts running and transports the pipe to the local inspection station.

[0128] This workstation is equipped with a single-line laser rangefinder, which is installed above and on both sides of the conveyor belt to measure the bending angle and end diameter of the pipe fittings. The single-line laser rangefinder uses the triangulation principle, emitting a laser beam and receiving the reflected signal from the surface of the pipe fitting to calculate the distance between the pipe fitting profile and the rangefinder.

[0129] As the pipe fitting passes through the local verification station on the conveyor belt, a single-line laser rangefinder automatically triggers a scan. The rangefinder, installed above the conveyor belt, moves axially along the pipe fitting, collecting contour point cloud data of the bend area. By fitting the spatial coordinates of the start point, apex, and end point of the bend, the actual bending angle is calculated. Distance measuring instruments installed on both sides of the conveyor belt measure the outer diameter of both ends of the pipe fitting, and the average value of multiple measurements is taken as the actual diameter of the end. .

[0130] After measurement, the controller at the local calibration station compares the measured value with the target value and calculates the deviation. The formula for calculating the bending angle deviation is as follows: The formula for calculating the end diameter deviation is as follows: .in The target bending angle specified in the drawing. The end target diameter is specified in the drawing.

[0131] The controller compares the calculated deviation with the preset allowable deviation. The allowable deviation for the bending angle is set to 1.5 degrees, and the allowable deviation for the end diameter is set to 0.2 mm. The determination of these two allowable deviations is based on the highest precision requirements for finished product dimensional tolerances in the relevant national standards for stainless steel pipe fittings used in construction, while also taking into account the actual needs of downstream installation operations for pipe fitting connections.

[0132] like and If the reading is within millimeters, the fitting is deemed to have passed the local inspection. The controller marks the fitting as normal, and the conveyor belt continues to run, sending the fitting to the finished product precision inspection station for subsequent 3D scanning inspection.

[0133] like or If the deviation is less than a millimeter, the fitting is deemed to have failed partial inspection. The controller marks the fitting as requiring re-inspection and records the deviation value and the type of deviation (bending angle deviation or end diameter deviation) in the fitting's production data. Fittings requiring re-inspection will be prioritized for delivery to the finished product precision inspection station for a more comprehensive scan and evaluation during subsequent 3D precision inspection to determine whether they can be repaired or scrapped.

[0134] After the local verification is completed, the pipe end forming and bending integrated machine uploads the local verification results of the pipe fitting to the central dispatch server. The specific data includes the unique identification code of the pipe fitting and the actual bending angle. Actual end diameter Bending angle deviation End diameter deviation 1. Check the status label (normal or requires review).

[0135] The above data is stored in the production database and associated with the initial inspection feature vector collected in step S2 and the actual processing parameters recorded in step S3 to form a complete single-piece production data chain.

[0136] Step S4 of this invention involves finished product fine inspection, sorting, and closed-loop updating. This is used to perform high-precision 3D scanning and quality grading of the pipe fittings that have completed partial verification, and to automatically sort them based on the grading results. Simultaneously, the production data is fed back to the self-learning engine to update the process parameters and potential energy function model in the cloud model library. Specifically, it includes the following steps: S401: After completing the partial verification in step S306, both normal and re-verified pipe fittings are transported to the finished product precision inspection station by conveyor belt. This station is equipped with a structured light 3D scanner, which is installed above and around the station and can perform 360-degree scanning of the pipe fittings without blind spots.

[0137] A structured light 3D scanner projects an coded structured light pattern onto the surface of the pipe fitting, while multiple high-speed industrial cameras simultaneously capture the structured light images reflected from the pipe fitting surface. The scanner's built-in processing unit calculates the spatial coordinates of each point on the pipe fitting surface based on the deformation pattern generated by the structured light on the pipe fitting surface using triangulation principles, generating high-density 3D point cloud data. This high-density 3D point cloud data has a density of no less than 8000 points and a scanning resolution of 0.02 mm.

[0138] The scanner transmits the generated point cloud data to the AI ​​inference server, which is deployed with... The architecture of the deep learning hierarchical model, The training process of the grading model is as follows: A total of 20,000 3D point cloud samples of pipe fittings that have undergone quality assessment in the past are collected, and professional quality inspectors label each sample with a comprehensive quality index. (Continuous values ​​between 0 and 1) and defect categories (Six categories in total: elliptical deformation, wall thinning, outer wall wrinkling, inner wall wrinkling, cracks, and qualified). The labeled samples are divided into training, validation, and test sets in an 8:1:1 ratio. The model adopts a joint classification and regression training strategy, and the loss function is the weighted sum of the classification cross-entropy loss and the regression mean squared error loss.

[0139] It should be noted that... The specific network structure of the hierarchical model is as follows: the input layer receives data in the following dimensions. Point cloud data, of which represents the number of point clouds, and 3 corresponds to spatial coordinates. Then, it passes through three collection abstraction layers in sequence, each containing a sampling layer, a grouping layer, and... The sampling layer uses the farthest point sampling algorithm, while the grouping layer uses the ball query method. The layer consists of a shared multilayer perceptron and a max-pooling layer. The output feature dimensions of the three ensemble abstraction layers are 512, 256, and 128, respectively. Finally, the point cloud resolution is restored through a feature propagation layer, and a global feature vector is output through a fully connected layer. This feature vector is then fed into the classification and regression branches, respectively. The classification branch uses a fully connected layer plus... The activation function outputs the probability distribution of six types of defects; the regression branch uses a fully connected layer. Activation function, outputting the overall quality index .

[0140] The AI ​​inference server preprocesses the input point cloud data, including downsampling and noise reduction, coordinate normalization, and point cloud alignment. After preprocessing, the point cloud data is input... A hierarchical model. The model aggregates local neighborhood features layer by layer through a multi-layer point feature extraction module, and finally outputs a global feature vector through a global pooling layer. This feature vector is then mapped by a fully connected layer to output two results: a comprehensive quality index. and defect categories .

[0141] Overall Quality Index A continuous value between 0 and 1, with values ​​closer to 1 indicating better pipe quality. Defect Category The output is in discrete categories, specifically including six categories: elliptic deformation, wall thinning, external wall wrinkling, internal wall wrinkling, cracks, and qualified. When the model determines that the pipe fitting has defects, it also outputs the quantitative values ​​of each defect, including the maximum wrinkle depth, minimum wall thickness, and ellipticity deviation.

[0142] S402: The AI ​​inference server will calculate the overall quality index. The grading thresholds are compared with preset thresholds. These thresholds are set based on the industry standards for stainless steel pipe fittings used in construction, which define the quality grades of finished products, and are determined in conjunction with downstream customers' acceptance standards for pipe fitting quality. The specific grading rules are as follows: like If the value is within the acceptable range, the pipe fitting is considered a qualified product. The threshold of 0.92 corresponds to the lower limit of the quality of superior products in the industry standard, and pipe fittings that meet this standard can be directly shipped from the factory.

[0143] like If the condition is not met, the fitting is deemed repairable. Fittings within this range have minor defects but are not considered unusable and can be repaired through localized repairs or secondary processing.

[0144] like If the quality threshold of 0.8 is not met, the fitting is deemed scrap. The threshold corresponds to the minimum acceptable quality level in industry standards. Fittings below this standard cannot be repaired to meet the qualification requirements and are directly scrapped.

[0145] The AI ​​inference server sends the grading results and the corresponding unique identification code of the pipe to the controller of the finished product sorting robot. The finished product sorting robot is a six-axis collaborative robot with a pneumatic flexible gripper at its end, which can automatically adjust its gripping posture according to the shape and size of the pipe.

[0146] Based on the received grading results, the sorting robot controller controls the robot to perform corresponding sorting actions. For qualified products, the robot picks up the pipes and places them in the qualified product bin; for reworkable products, the robot picks up the pipes and places them in the rework product bin; for defective products, the robot picks up the pipes and places them in the defective product bin. Each bin is equipped with a position sensor, and the system automatically issues a replacement prompt when a bin is full.

[0147] S403: After sorting is completed, the central dispatch server generates a unique lifecycle traceability code for each pipe fitting that has completed the entire processing. The traceability code uses a QR code format and contains the following information: unique pipe fitting identifier, raw material batch number, cutting length and actual cutting length, and the potential energy value from the online initial inspection. Including grade labels, actual stroke and holding time of pipe end forming, actual speed and angle of pipe bending, and comprehensive quality index of finished product inspection. And the defect category, sorting result (qualified products, rework products or waste products) and production timestamp.

[0148] The central dispatch server stores the traceability code information in the production record table of the cloud model library, and simultaneously sends the traceability code to the marking equipment. The marking equipment laser-engraves a QR code on the surface of the pipe fitting, completing the physical marking of the traceability code. Users or quality inspectors can subsequently scan the QR code to retrieve all production and testing data for the pipe fitting, achieving end-to-end quality traceability.

[0149] S404: The central dispatch server collects complete production data for all pipe fittings in this production batch. Each pipe fitting corresponds to a 5-tuple data record, in the form of... ,in For normalized feature vectors, As the baseline process parameter vector, This is the actual correction amount used. This is the overall quality index of the finished product, when the tube blank does not trigger the correction process. Take the zero vector.

[0150] The central dispatch server stores the above data records into the historical production record set of the current model node in the cloud model library. When the cumulative production quantity of this model reaches the preset update threshold, the system triggers the self-learning update process. The update threshold is set at 50 units. This value is determined to balance the timeliness of model updates with statistical significance. Too small a sample size will lead to model overfitting, while too large a sample size will cause the model response to lag.

[0151] The self-learning update process includes two parallel tasks: updating the Gaussian process regression model and updating the mass potential energy function.

[0152] The Gaussian process regression model is updated using incremental learning. The central scheduling server selects medium-potential energy tube billet data that have triggered parameter adjustments from the historical production record set and inputs them into the vector. With actual correction amount As new training samples, the covariance matrix and hyperparameters of the model are updated using the sparse Gaussian process method. An incremental update is performed every 20 new samples.

[0153] The mass potential energy function is updated via parameter refitting. The central scheduling server extracts the feature vectors of the 100 most recent billets from the historical production record set. With the overall quality index ,by As the target variables, the weight vector w and bias term b of the logistic regression model are retrained. The training employs maximum likelihood estimation, with the objective function being to minimize the cross-entropy loss between the predicted potential energy value and the actual mass index. After refitting, the central scheduling server stores the updated w and b in the cloud model library, replacing the original potential function parameters for this model.

[0154] S405: After completing the update of the mass potential energy function, the central scheduling server, based on the potential energy distribution of the latest batch of billets, adjusts the high potential energy threshold. and low potential threshold Make adaptive adjustments.

[0155] The central dispatch server extracts the potential energy values ​​of the most recent 100 tube blanks from the historical production record set. and the corresponding finished product quality index Plot the distribution relationship between potential energy values ​​and the mass index. The system calculates the percentiles of the potential energy value sequence: Update to the 75th percentile of the potential energy value sequence, Updated to the 15th percentile of the potential energy value sequence.

[0156] The percentile selection criteria are as follows: the 75th percentile corresponds to the top 25 tube blanks with better quality in historical data, and the 15th percentile corresponds to the bottom 15 tube blanks with poorer quality. This percentile selection strategy can match the grading rules with the actual quality distribution of the current production line.

[0157] The central dispatch server will be updated. and The model is stored in the cloud model library and simultaneously distributed to the online testing unit for use in determining the grade of billets in subsequent batches.

[0158] S406: When a certain model completes self-learning updates and its cumulative production quantity reaches 200 units or more, the central scheduling server initiates the knowledge migration process for similar models. The system traverses all models in the cloud model library that are similar to that model. For neighboring model nodes with a similarity greater than 0.7, the similarity is calculated using the cosine similarity algorithm in step S103.

[0159] For each neighboring model that meets the similarity criteria, the central scheduling server weights and merges the updated mass potential energy function parameters w and b, as well as the process parameter package, of the current model according to the similarity weight, and then migrates them to the queue of neighboring models to be updated. The migration formula is as follows: ; In the formula, The parameters are updated from the original model. Current parameters for the target model; The cosine similarity between the two models.

[0160] This migration strategy enables similar models to share process optimization results, reduces the sample requirements for independent learning, and significantly improves the quality of initial parameters for new or small-batch models.

[0161] The central scheduling server stores the parameters to be updated in the migration cache of each neighboring model node, which will be automatically loaded and used when these models are produced again.

[0162] The second objective of this invention is to provide an unmanned production control system for building stainless steel pipe fittings, used to execute the unmanned production control method for building stainless steel pipe fittings described in the above method embodiments.

[0163] The system includes a central dispatch server, a cloud model library, a laser cutting unit, an online inspection unit, a three-dimensional material silo, an integrated pipe end forming and bending machine, a finished product three-dimensional precision inspection unit, and a finished product sorting robot.

[0164] The central dispatch server is deployed inside the production line monitoring room, establishing communication links with each of the aforementioned units via industrial Ethernet, while maintaining a real-time connection with a cloud model library deployed remotely in the cloud. The server periodically accesses the task interface of the manufacturing execution system to obtain work order information, queries the cloud model library based on the target product model, completes the loading of baseline process parameters, comparison of historical best parameters, generation of transfer learning parameters for the first production model, and construction of the mass potential energy function, and sends the cutting length parameters to the laser cutting unit, the mass potential energy function parameters and normalized parameters to the online detection unit, the forming and bending process parameters to the pipe end forming and bending integrated machine, and the grading threshold parameters to the finished product sorting robot.

[0165] The cloud model library adopts a graph database architecture, with data nodes built independently for each type of pipe fitting. Each node stores a vector of baseline process parameters, mass potential energy function parameters, historical best process parameter vectors, and a set of similarity-related edges with other types.

[0166] The laser cutting unit receives the cutting length parameters and performs automatic length-based cutting on the raw material pipe, with cutting accuracy controlled within ±0.5 mm. The unit has a built-in residual material optimization algorithm that dynamically adjusts the cutting sequence based on the matching relationship between the remaining raw material length and the length of subsequent cutting tasks. After cutting, the pipe segment is conveyed to the online inspection unit by a conveyor belt.

[0167] The online inspection unit comprises four equally spaced line laser profilometers, a high-speed industrial camera, and a pneumatic pusher mounted at the end of the conveyor belt. The profilometers automatically calculate the scanning window length based on the billet diameter, extracting raw features such as ellipticity deviation, maximum pit depth, straightness deviation, surface roughness variation, and window confidence weights. The unit's embedded processor performs weighted fusion of similar features from the four windows, calls normalization parameters from the server to normalize the raw features, substitutes them into the mass potential energy function to calculate the potential energy value, and determines the billet grade based on preset high and low potential energy thresholds. The pneumatic pusher pushes low-potential-energy billets into the scrap bin, while high- and medium-potential-energy billets continue to be conveyed to the automated storage and retrieval system.

[0168] The automated storage and retrieval system (AS / RS) comprises a multi-layer racking structure and a stacking robot. Each layer stores tube blanks in sections based on their diameter and length. The stacking robot places the tube blanks into designated locations according to the retrieval location numbers assigned by the server. The warehousing management system generates a unique identifier for each tube blank and binds it to its potential energy value, grade label, and feature vector. When downstream equipment issues a material handling request, the stacking robot performs the material handling operation and transports the tube blank to the tube end forming and bending integrated machine.

[0169] The integrated pipe end forming and bending machine includes a two-station turntable, a hydraulically driven flaring and shrinking die assembly, a CNC bending system, an angle encoder, a pressure sensor, and an edge controller. The die assembly is installed using electromagnetic adsorption, with a replacement time of less than 30 seconds. The edge controller incorporates a Gaussian process regression model to generate personalized process parameter corrections for the tube blank with central potential energy, and determines whether to apply the correction value based on the comparison between the predicted variance of the model output and the safety threshold. The pipe end forming station performs flaring or shrinking operations according to the actual stroke and holding time, with a stroke control accuracy of ±0.05 mm. After the turntable rotates 180 degrees to feed the tube blank into the bending station, the CNC bending system performs bending processing according to the actual bending speed, clamping force, and mandrel extension. The angle encoder provides closed-loop control of the bending angle, and the pressure sensor provides real-time feedback on the mandrel position.

[0170] The finished product 3D precision inspection unit includes a structured light 3D scanner and an AI inference server. The scanner generates high-density 3D point cloud data by projecting coded structured light patterns, with a point cloud density of no less than 8000 points and a scanning resolution of 0.02 mm. The AI ​​inference server is deployed based on… The architecture employs a deep learning-based hierarchical model, outputting a comprehensive quality index and defect categories. The comprehensive quality index is a continuous value between 0 and 1, and the defect categories include six types: elliptical deformation, wall thinning, external wall wrinkling, internal wall wrinkling, cracks, and acceptable.

[0171] The finished product sorting robot is a 6-axis collaborative robot with a pneumatic flexible gripper at its end. The sorting robot receives the grading results sent by the AI ​​inference server, and based on the comparison between the comprehensive quality index and the preset threshold, it picks up the pipes and places them into the qualified product box, the rework box, or the scrap box.

[0172] All the above units operate collaboratively under the unified coordination of the central dispatch server, forming a fully automated production control system that spans from raw material cutting to finished product sorting.

[0173] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for unmanned production control of stainless steel pipe fittings for buildings, characterized in that, Includes the following steps: S1. Obtain the target product model and corresponding baseline process parameter vector in the production task. Load the parameters of the mass potential energy function from the cloud model library according to the target product model. The mass potential energy function is a logistic regression model with a four-dimensional feature vector as input. The four feature dimensions correspond to ellipticity deviation, pit depth, straightness deviation and surface roughness change, respectively. S2. Cut the raw material pipe to length to obtain the cut pipe blank. Perform multi-window scanning on the cut pipe blank, extract the ellipticity deviation, pit depth, straightness deviation, surface roughness change and confidence weight of each window, and obtain the normalized feature vector after weighted fusion and normalization. Substitute it into the mass potential energy function to calculate the potential energy value. Based on the preset high potential energy threshold and low potential energy threshold, the tube blank is divided into high potential energy tube blank, medium potential energy tube blank and low potential energy tube blank. The low potential energy tube blank is discarded and the high potential energy tube blank and medium potential energy tube blank are stored in the three-dimensional material silo. S3. Take out the tube blank from the three-dimensional silo and read the potential energy value and normalized eigenvector of the tube blank; For high-potential-energy tube blanks, the reference process parameter vector is directly used for tube end forming and bending. For a medium-potential tube blank, the normalized feature vector of the tube blank is concatenated with the reference process parameter vector and then input into a Gaussian process regression model. The reference process parameters are corrected by the correction amount output by the model to obtain the actual processing parameters. The actual processing parameters are then used for tube end forming and bending to obtain the finished tube fitting.

2. The unmanned production control method for stainless steel pipe fittings in construction according to claim 1, characterized in that: When the target product model does not have a historical record in the cloud model library, the model most similar to the target product model is selected from the produced models using the cosine similarity algorithm. The mass potential energy function parameters of the most similar model are copied as the initial mass potential energy function parameters of the target product model. The reference process parameter vector of the most similar model is mapped according to the outer diameter ratio and used as the initial reference process parameter vector of the target product model.

3. The unmanned production control method for stainless steel pipe fittings in construction according to claim 1, characterized in that: The parameters of the mass potential energy function include a weight vector and a bias term. The dimension of the weight vector is the same as the dimension of the normalized eigenvector. The four weight components correspond to the contribution of ellipticity deviation, maximum pit depth, straightness deviation, and surface roughness variation to the potential energy value, respectively. The bias term is used to adjust the overall output reference of the potential energy function.

4. The unmanned production control method for stainless steel pipe fittings in construction according to claim 1, characterized in that: When performing multi-window scanning on the cut tube blank, the length of each scanning window is dynamically calculated based on the outer diameter of the tube blank. The window length is equal to the outer diameter multiplied by a preset proportional coefficient, which is pre-calibrated according to the diameter range of the tube blank.

5. The unmanned production control method for stainless steel pipe fittings in construction according to claim 1, characterized in that: When extracting the confidence weight of each window, it is calculated in real time based on the ambient light intensity, the reflectivity of the tube blank surface, and the sensor signal quality at the scanning time. During weighted fusion, the confidence weight of each window is used as the weighting coefficient to calculate the weighted average of similar features.

6. The unmanned production control method for stainless steel pipe fittings in construction according to claim 1, characterized in that: The Gaussian process regression model adopts The kernel function serves as the covariance function, and the predicted variance output by the model is used to characterize the degree of uncertainty of the current prediction. When the predicted variance exceeds a preset safety threshold, the correction is abandoned and the sample is recorded for subsequent model updates.

7. The unmanned production control method for stainless steel pipe fittings in construction according to claim 1, characterized in that: When performing tube end forming and bending on medium potential energy tube blanks, the stroke control accuracy of tube end forming is controlled within the preset tolerance range by a displacement sensor closed loop, and the holding time is executed according to the process parameters; during tube bending, the bending angle is controlled by an angle encoder closed loop, and the angle control accuracy reaches the preset allowable deviation range.

8. The unmanned production control method for stainless steel pipe fittings in construction according to claim 1, characterized in that: After completing the pipe end forming and bending process, the pipe fitting is quickly checked locally. The bending angle and end diameter are measured by a single-line laser rangefinder. When the end diameter deviation exceeds the corresponding preset allowable deviation, the pipe fitting is marked as a pipe fitting that needs to be checked.

9. The unmanned production control method for stainless steel pipe fittings in construction according to claim 1, characterized in that: After all pipe fittings have been processed, the normalized feature vector, baseline process parameter vector, actual correction amount, and finished product quality index of each pipe fitting are collected to form a data record. When the cumulative production quantity of the same model reaches the preset threshold, the Gaussian process regression model is incrementally updated and the parameters of the mass potential energy function are refitted.

10. An unmanned production control system for building stainless steel pipe fittings, used to execute the unmanned production control method for building stainless steel pipe fittings as described in any one of claims 1-9, characterized in that, It includes a central dispatch server, cloud model library, laser cutting unit, online inspection unit, three-dimensional silo, pipe end forming and bending integrated machine, finished product three-dimensional precision inspection unit, and finished product sorting robot; The central scheduling server is used to obtain the target product model and the baseline process parameter vector, load the mass potential energy function parameters from the cloud model library, and send each process parameter to the corresponding execution unit. The cloud model library is used to store the baseline process parameter vector, mass potential energy function parameter, historical best process parameter vector, and similarity association edge set between models for each pipe fitting model. The laser cutting unit is used to perform fixed-length cutting on raw material pipes; The online detection unit is used to perform multi-window scanning on the cut tube blank, extract defect features and confidence weights, calculate potential energy values, and classify the tube blank into high potential energy, medium potential energy and low potential energy according to the potential energy values, and discard low potential energy tube blanks. The three-dimensional silo is used to store high-potential-energy tube blanks and medium-potential-energy tube blanks, and each tube blank is bound with a potential energy value, grade label and normalized feature vector. The tube end forming and bending integrated machine is used to take out tube blanks from the three-dimensional silo. According to the potential energy level, for high potential energy tube blanks, a reference process parameter vector is used for tube end forming and bending. For medium potential energy tube blanks, the actual processing parameters after correction by the Gaussian process regression model are used for tube end forming and bending. The finished product three-dimensional precision inspection unit is used to perform three-dimensional scanning on the processed pipe fittings and output the comprehensive quality index and defect category; The finished product sorting robot is used to sort pipe fittings according to the comprehensive quality index.

Images