Intelligent energy consumption management optimization system for welded pipe production process

Abstract

The present application relates to the technical field of metal pipe production, and discloses an intelligent energy consumption management optimization system for welded pipe production process, which aims to solve the problem that product quality and energy consumption are difficult to be optimized simultaneously in the production process.The system comprises: an acquisition unit for synchronously acquiring multi-modal associated signals; a state representation unit for mapping high-dimensional signals into low-dimensional hidden space vectors representing process states by using a variational autoencoder; a target calibration unit for correlating quality and energy consumption data to dynamically calibrate an optimal state region; and a control unit for generating control instructions according to the current state and the optimal target by using a model predictive control strategy.The present application can significantly reduce the energy consumption per unit product while ensuring product quality and improving the stability and economy of the production process by constructing a data-driven closed-loop adaptive control loop.

Description

Technical Field

[0001] This invention relates to the field of metal pipe manufacturing technology, specifically to an intelligent energy consumption management and optimization system for welded pipe production. Background Technology

[0002] Welded pipe production is a continuous and high-speed industrial process. Its core lies in fusing the edges of steel strips together through high-frequency induction or contact welding to form continuous pipes. Ensuring stable and reliable weld quality and controlling energy consumption during production are two key objectives that determine product competitiveness and production efficiency. However, in actual production, these two aspects are often difficult to optimize in a coordinated manner.

[0003] Currently, process control in welded pipe production lines largely relies on operator experience and pre-set static process parameters, such as welding power and production line speed. This control method has inherent limitations. A major challenge is that conventional process parameters (such as voltage and current) provide a coarse-grained reflection of the macroscopic state of the production process; their changes often lag behind the transient changes in the actual physical process, failing to capture early, subtle signs leading to weld defects or energy waste. Therefore, the control system lacks the ability to accurately perceive the deeper levels of the production process.

[0004] This lack of perception directly leads to conservative and inefficient control strategies. To ensure product qualification rates, operators often set large safety margins for energy-related parameters, such as excessively increasing welding power, sacrificing energy efficiency to avoid potential quality risks. This separation of quality control and energy management keeps the production process in an uneconomical operating state for extended periods. Furthermore, when raw material batches, environmental temperature and humidity, and other operating conditions change, the fixed process parameters cannot be adaptively adjusted, causing fluctuations in product quality or abnormal energy consumption, requiring manual intervention. This further limits the overall stability and intelligence level of the production process.

[0005] Therefore, existing technologies urgently need a new technical solution that can break through the dependence on traditional process parameters, perform a deeper and more comprehensive real-time status characterization of the production process, and on this basis, achieve closed-loop, adaptive, and collaborative optimization of quality and energy consumption. Summary of the Invention

[0006] The technical problem to be solved by this invention is that the existing energy consumption management and optimization methods for welded pipe production processes have problems such as relying on human experience to set control targets, difficulty in balancing production quality and energy consumption efficiency, and insufficient adaptability to batch fluctuations of raw materials and changes in equipment status.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] The first aspect of this invention provides an intelligent energy consumption management and optimization system for welded pipe production processes, the system comprising:

[0009] One acquisition unit is used to synchronously acquire multimodal associated signals during the welded pipe production process and process them into high-dimensional feature vectors;

[0010] A state representation unit, connected to the acquisition unit, is used to map the high-dimensional feature vector into a low-dimensional latent space state vector representing the overall state of the current production process.

[0011] A target calibration unit is used to receive the quality signal and the system energy consumption signal provided by the downstream online non-destructive testing unit, associate the quality signal and energy consumption signal with the low-dimensional latent space state vector, and calibrate an optimal state region that takes into account the preset quality and energy consumption benefits based on the association result.

[0012] A control unit, connected to the state characterization unit and the target calibration unit, is used to generate and output control commands to the production equipment based on the low-dimensional latent space state vector and the optimal state region, which aim to guide the production process state to the optimal state region.

[0013] Preferably, the acquisition unit is specifically used to acquire multimodal associated signals, including at least the high-frequency electromagnetic harmonic signals of the high-frequency welding machine power supply and the mechanical vibration signals of key mechanical components.

[0014] Preferably, the state representation unit includes a pre-trained variational autoencoder model. The variational autoencoder model consists of an encoder and a decoder, where the encoder is used to map from the high-dimensional feature vector x to the low-dimensional latent space state vector z. The training process of the variational autoencoder model involves optimizing a loss function. Implementation, which is defined as:

[0015]

[0016] in, To reconstruct the loss term, D KL For encoder output distribution q φ The KL divergence term between (z|x) and the standard normal prior distribution p(z), where β is the weighting coefficient, and φ and θ are the network parameters of the encoder and decoder, respectively. This is the reconstructed high-dimensional feature vector.

[0017] In one specific embodiment, the target calibration unit performs the following operations:

[0018] 1. Based on the production line delay, the quality signal is... and energy consumption signal E tWith the corresponding low-dimensional latent space state vector z t Perform time alignment to form data association pairs

[0019] 2. Based on the data association pairs, firstly, select the state points whose quality signals are qualified to form a qualified state set. Subsequently, from the qualified state set, low-dimensional latent space state vectors with corresponding energy consumption signal values ​​not exceeding a preset quantile are further selected to form the elite state set.

[0020] 3. Determine the optimal state region based on the elite state set.

[0021] Preferably, the target calibration unit determines the center target point z of the optimal state region by calculating the arithmetic mean of all low-dimensional latent space state vectors in the elite state set. golden :

[0022]

[0023] Preferably, the control unit includes a model prediction controller.

[0024] In one specific embodiment, the model predictive controller includes a state transition model, which is used to determine the current low-dimensional latent space state vector z. t and control command u t Predict the latent space state trajectory at one or more future time steps. The model predictive controller determines the optimal control command sequence by solving an optimization problem. The optimization problem aims to minimize a comprehensive cost function J, which has the following form:

[0025]

[0026] The first term is the deviation between the latent space state trajectory of the future multiple time steps and the optimal state region; the second term is the energy consumption term associated with the optimal control command sequence; and the third term is the rate of change term of the optimal control command sequence. N p To predict the time domain, W z W u W Δu This is the corresponding weight matrix.

[0027] Preferably, the target calibration unit is configured to periodically or dynamically update the optimal state region when the production conditions change by a preset time, and provide the updated optimal state region to the control unit as its optimization target.

[0028] A second aspect of this invention provides an intelligent energy consumption management and optimization method for welded pipe production processes, the method comprising the following steps:

[0029] Multimodal associated signals during the welded pipe production process are acquired synchronously and processed into high-dimensional feature vectors;

[0030] The high-dimensional feature vector is mapped to a low-dimensional latent space state vector that represents the overall state of the current production process.

[0031] The system receives the quality signal and the system's energy consumption signal from the downstream online non-destructive testing unit, and calibrates the optimal state region based on the correlation results between the quality signal, the energy consumption signal and the low-dimensional latent space state vector.

[0032] Based on the low-dimensional latent space state vector and the optimal state region, a model predictive control strategy is used to generate control commands aimed at guiding the production process state to the optimal state region, and these commands are applied to the production equipment.

[0033] This invention provides an intelligent energy consumption management and optimization system for welded pipe production. It has the following beneficial effects:

[0034] 1. This invention establishes a target calibration unit to correlate the low-dimensional latent space state vector of the upstream production process with the downstream online quality and energy consumption signals in real time, and autonomously calibrates the optimal state region based on the correlation results. This mechanism enables the system's control target to be dynamically updated according to changes in raw material batches or equipment status, avoiding the problems of poor adaptability and inadequate control effects caused by relying on fixed parameters or manual experience to set control targets in traditional methods.

[0035] 2. This invention synchronously acquires multi-modal accompanying signals such as high-frequency electromagnetic harmonics and mechanical vibrations through an acquisition unit, and maps them into low-dimensional latent space state vectors using a state characterization unit. This approach can capture and quantify deep-level information reflecting the comprehensive state of the production process that is difficult to measure directly using traditional methods. Compared with technical solutions that rely on only a single or a few process parameters for control, it provides a more comprehensive and accurate description of the process state, thus providing a reliable data foundation for subsequent precise control and optimization decisions.

[0036] 3. The invention uses a model-predictive control unit, which includes penalty terms for state deviation, control energy consumption, and control rate of change in its comprehensive cost function, and solves for the optimal control command within a unified framework. This method can proactively plan control actions, maintaining the production process state within the optimal state region defined by quality and energy consumption, while also considering the economy of the control command itself and the stability of the process. It achieves synergistic optimization of multiple, even conflicting, production objectives (such as high quality, low energy consumption, and high stability). Attached Figure Description

[0037] Figure 1 This is a structural block diagram of an intelligent energy consumption management and optimization system for welded pipe production process according to an embodiment of the present invention;

[0038] Figure 2 A flowchart of an intelligent energy consumption management optimization method for welded pipe production process according to an embodiment of the present invention;

[0039] Figure 3 This is a low-dimensional hidden space state distribution diagram according to an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram illustrating the optimal state region calibration according to an embodiment of the present invention;

[0041] Figure 5 This is a schematic diagram of the trajectory of the model predictive control process in an embodiment of the present invention;

[0042] Figure 6 This is a time series comparison chart of the application effects of embodiments of the present invention.

[0043] Among them, 10 is the acquisition unit; 20 is the state characterization unit; 30 is the target calibration unit; and 40 is the control unit. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the scope of protection of the invention.

[0045] See attached document Figure 1 , Figure 1 This is a structural block diagram of an intelligent energy consumption management and optimization system for welded pipe production process according to an embodiment of the present invention. The intelligent energy consumption management and optimization system for welded pipe production process provided by the present invention may include: a data acquisition unit 10, a status characterization unit 20, a target calibration unit 30, and a control unit 40.

[0046] The acquisition unit 10 is used to synchronously acquire multimodal associated signals during the welded pipe production process and process them into high-dimensional feature vectors. Multimodal associated signals include, but are not limited to, high-frequency electromagnetic harmonic signals from the high-frequency welding machine power supply, mechanical vibration signals from key mechanical components, acoustic signals, and conventional process parameters. The acquisition unit 10 processes the high-dimensional feature vector x representing time t. t Output to state representation unit 20.

[0047] State representation unit 20 is connected to the output of acquisition unit 10, and it receives high-dimensional feature vector x. tThis unit is used to process high-dimensional feature vectors x t The mapping is a low-dimensional latent space state vector z that represents the overall state of the current production process. t In one embodiment, a pre-trained variational autoencoder model is deployed within the state representation unit 20. This model maps high-dimensional feature vectors to low-dimensional latent space state vectors through its encoder part, and its training is accomplished by optimizing a loss function that includes a reconstruction loss term and a KL divergence regularization term. The state representation unit 20 generates the low-dimensional latent space state vector z. t The output is sent to the target calibration unit 30 and the control unit 40.

[0048] The target calibration unit 30 is connected to the output of the state characterization unit 20, and simultaneously receives the quality signal Q from the online non-destructive testing unit downstream of the welded pipe production line. t and the system's energy consumption signal E t This unit is used to correlate the quality signal and energy consumption signal with the low-dimensional latent space state vector, and to calibrate an optimal state region that balances preset quality and energy consumption benefits based on the correlation result. The center target point of this optimal state region is numerically determined by calculating the arithmetic mean of all member vectors in an elite state set. The elite state set consists of all low-dimensional latent space state vectors whose correlated quality signals are qualified and whose correlated energy consumption signal values ​​are not higher than a preset threshold. The target calibration unit 30 outputs the calculated center target point of the optimal state region to the control unit 40.

[0049] The control unit 40 has its input terminals connected to the output terminals of the state characterization unit 20 and the target calibration unit 30, respectively. This unit is used to determine the current low-dimensional latent space state vector z based on the received data. t Based on the target point at the center of the optimal state region, generate and output control commands u designed to guide the production process state to the optimal state region. t To the production equipment. In one embodiment, the control unit 40 employs a model predictive control (MPC) strategy. This strategy determines the optimal control command by solving a finite-time optimization problem in each control cycle. The optimization problem aims to minimize a comprehensive cost function that includes penalties for deviations from the target region in the future state trajectory, control command energy consumption, and the rate of change of the control command.

[0050] The workflow of this system is as follows: The acquisition unit 10 continuously acquires multimodal associated signals from the welded pipe production line and processes them into a high-dimensional feature vector x. t The state representation unit 20 receives x t It is mapped to a low-dimensional latent space state vector z t Simultaneously, the target calibration unit 30 receives the downstream quality signal Q. t and energy consumption signal Et , with history z t The system performs correlation and dynamically calculates the center target point of the optimal state region. The control unit 40 receives the current z-axis value. t By solving the model predictive control optimization problem and finding the central target point, the optimal control command u is calculated. t It is then sent to the actuator of the welded pipe production equipment (such as high-frequency welding machine, rolling mill drive motor, etc.), thus forming a closed-loop adaptive optimization control process.

[0051] See attached document Figure 1 The acquisition unit 10 in this embodiment of the invention will now be described in detail. The function of the acquisition unit 10 is to synchronously acquire multimodal associated signals during the welded pipe production process and process them into a unified high-dimensional feature vector x. t This serves as the basis for subsequent analysis and control.

[0052] In one specific embodiment, the multimodal associated signals include high-frequency electromagnetic signals, mechanical vibration signals, acoustic signals, and conventional process parameters. High-frequency electromagnetic signals are acquired by deploying high-precision voltage transformers and broadband current transformers on the main power input circuit of the induction welding machine. Mechanical vibration signals are acquired by installing industrial-grade piezoelectric accelerometers on key mechanical components, such as the final-stage finishing roll stand in the forming section, the extrusion roll bearing housing at the welding point, and the first-stage sizing roll stand in the sizing section. Acoustic signals are acquired by installing acoustically shielded industrial-grade broadband microphones in the vicinity of the welding point to collect acoustic radiation information generated during the welding process.

[0053] All sensor signals are synchronously acquired via a data acquisition card (DAQ) at a preset high sampling frequency and assigned a high-precision timestamp generated by a unified clock source. Before feature extraction, the acquisition unit 10 performs preprocessing operations on the raw time-domain signal. For example, a bandpass filter is applied to the raw signal to filter out power frequency interference and high-frequency random noise that are irrelevant to the production process status, retaining only the effective frequency band containing key status information.

[0054] After preprocessing, the signal enters the feature engineering module. For high-frequency electromagnetic signals, the acquisition unit 10 applies a window function (such as a Hanning window) and performs a short-time Fourier transform (STFT). The length of the window function is set to cover multiple periods of the fundamental frequency, and an overlap rate of not less than 50% is set to achieve a balance between time-domain and frequency-domain resolution. Subsequently, the amplitude and phase information of the fundamental frequency and a preset number (e.g., the first 50) of each harmonic are extracted. For mechanical vibration signals, their time-domain statistical characteristics (root mean square, kurtosis, peak factor, etc.) and frequency-domain characteristics (power spectral density, band energy, etc.) are calculated. For acoustic signals, Mel-frequency cepstral coefficients (MFCCs) that reflect their timbre characteristics are extracted.

[0055] At each timett, the acquisition unit 10 arranges all the features extracted in the preceding steps, along with routine process parameters (such as mill line speed, set power, cooling water temperature, etc.) obtained from the production line programmable logic controller (PLC) via the data bus, in a fixed order to form an initial high-dimensional feature vector. Since the physical units and numerical ranges of each feature are different, the acquisition unit 10 performs normalization processing on this initial feature vector before output, for example, using the Z-score normalization method, so that each feature component in the vector has zero mean and unit variance. After normalization, the final high-dimensional feature vector x with dimension D is formed. t As a standardized output of the acquisition unit 10, it is transmitted to the state characterization unit 20.

[0056] See attached document Figure 1 The state characterization unit 20 in this embodiment of the invention will now be described in detail. The function of the state characterization unit 20 is to receive a high-dimensional feature vector x from the acquisition unit 10. t It is nonlinearly mapped into a low-dimensional latent space state vector z that can compactly represent the overall state of the production process. t In one specific embodiment, this functionality is implemented by a pre-trained variational autoencoder (VAE) model.

[0057] The variational autoencoder model is structurally composed of two interconnected neural network modules: an encoder and a decoder.

[0058] The encoder's network structure is a multi-layer feedforward neural network. Its input layer receives a high-dimensional feature vector x of dimension D. tThe network contains several hidden layers, each consisting of multiple neurons and a non-linear activation function (e.g., the rectified linear unit ReLU) to extract abstract representations of the input features layer by layer. The encoder's output layer is special; instead of directly outputting a single vector, it outputs two independent vectors: a mean vector μ of dimension J and a log-variance vector logσ of dimension J. 2 In this context, the dimension J of the latent space is set to be much smaller than the dimension D of the input features (i.e., J << D).

[0059] Between the encoder and decoder, there exists a sampling operation step. The low-dimensional latent space state vector z... t It is not directly output by the encoder, but is generated through a reparameterization process. Specifically, it is first generated from a standard normal distribution. We sample a random noise vector ∈ with the same dimension J as the latent space, and then calculate z using the following formula. t :

[0060] z t =μ+σ⊙∈;

[0061] Where σ is obtained by taking the log-variance vector logσ 2 The standard deviation vector is obtained by taking the exponent and square root, and ⊙ represents element-wise multiplication. This structural design makes the sampling process itself differentiable, thus allowing gradient backpropagation during model training.

[0062] The decoder's network structure is also a multi-layer feedforward neural network, typically designed as a mirror image of the encoder. Its input layer receives a low-dimensional latent space state vector z of dimension J. t The decoder decodes and upscales the latent space state vector layer by layer through several hidden layers containing non-linear activation functions (such as ReLU). Its output layer has a dimension of D, the same as the encoder's input dimension, and typically employs linear activation functions to reconstruct a vector with the same numerical range as the normalized input feature vector.

[0063] The parameters of the variational autoencoder model are not preset, but learned through an offline training process. This training process is separate from the online operation phase of the system.

[0064] During the offline training phase, a large-scale, representative historical dataset is first required. This dataset consists of a large number of high-dimensional feature vectors {x1, x2, ..., x...} collected and processed by the acquisition unit 10 during the production process over a past period. N The sample consists of N, where N is the total number of samples.

[0065] The goal of training is to adjust the network parameters φ of the encoder and θ of the decoder to minimize a predefined combined loss function. The loss function consists of a weighted sum of two parts:

[0066]

[0067] The first part is the reconstruction of the loss term. It calculates the original input vector x and the reconstructed vector output by the decoder. The difference between them. In one embodiment, this loss term is specifically calculated using the Mean Squared Error (MSE) function. This term drives the model to learn how to effectively compress information and recover the original information losslessly or with minimal loss from the compressed representation.

[0068] The second part is the KL divergence term D. KL (q φ (z|x)||p(z)). This term measures the posterior probability distribution q output by the encoder. φ (z|x) and a standard normal prior distribution The distance between them. This term, as a regularization term, constrains the latent space generated by the encoder to have continuous and smooth structural characteristics. The parameter β is a weight coefficient greater than zero, used to adjust the relative importance of reconstruction loss and KL divergence loss in the total loss.

[0069] The model is trained using a mini-batch-based stochastic gradient descent algorithm (e.g., the Adam optimizer). In each iteration, a small batch of samples is randomly selected from the historical dataset, its average loss is calculated, and the parameters are updated based on the gradient of this loss with respect to the network parameters φ and θ. This process is repeated until the total loss function converges to a preset low value or a preset number of training epochs is reached.

[0070] After the model training is completed and deployed in the online system, its operating mode changes. During the online operation phase of the system, only the trained encoder part is used, and the decoder part is no longer used. The state representation unit 20 receives the real-time high-dimensional feature vector x from the acquisition unit 10. t Then, a forward propagation calculation is performed through the encoder network to obtain the mean vector μ. t Sum of logarithmic variance vector To ensure the stability of the control system, the final output low-dimensional latent space state vector z t It is determined to be the mean vector, i.e., z t =μ t The determined vector z t It is transmitted to the target calibration unit 30 and the control unit 40.

[0071] See attached document Figure 1 The target calibration unit 30 in this embodiment of the invention will now be described in detail. The function of the target calibration unit 30 is to receive the low-dimensional latent space state vector z generated by the state characterization unit 20. t This data is then correlated with downstream feedback of quality and energy consumption signals to pinpoint the optimal state region. The first step in this process is to achieve precise correlation and alignment of the data over time.

[0072] Because the sensors used for condition sensing on the production line (e.g., sensors mounted on the extrusion rollers) are physically separated from the online non-destructive testing (NDT) units used for quality inspection, the quality information of a specific welded pipe cross-section generated at time t will be delayed by a time Δt. delay Only then can it be acquired by the NDT unit. In order to achieve the correct causal relationship between state and result, the target calibration unit 30 must compensate for this delay.

[0073] In one specific embodiment, the time delay Δt delay It is a dynamic variable whose value depends on the production line speed. The target calibration unit 30 obtains the current production line speed v in real time from the production line programmable logic controller (PLC). line (t). Physical distance L NDT , that is, the pipeline travel distance from the state-aware measuring point to the NDT measuring point, is a fixed parameter pre-determined and configured in the system. Therefore, for the latent space state vector z generated at time t... t The corresponding quality signal is expected to arrive at the NDT cell at time t. arrival for:

[0074]

[0075] To perform this alignment operation, the target calibration unit 30 maintains a data buffer internally, which employs a first-in-first-out (FIFO) queue structure. At each time t, the state representation unit 20 generates a low-dimensional latent space state vector z. t Its corresponding timestamp t and the energy consumption signal E at that time t As a data tuple (z t E t ,t) is stored in the buffer.

[0076] When downstream t arrival At a certain time, the NDT unit outputs a quality signal. At that time, the target calibration unit 30 performs a retrieval in the buffer based on the timestamp. It searches for and retrieves the timestamp t that satisfies t = t arrival -LNDT / v line (t) data tuples (z t E t Through this operation, the target calibration unit 30 successfully determined the upstream production status z. t The instantaneous energy consumption E generated in this state t And the downstream product quality Q ultimately resulting from this state. tarrival These three elements, when combined, form a complete, time-aligned data association pair. These data associations will serve as input for subsequently determining the optimal state region.

[0077] After completing the data alignment, the target calibration unit 30 bases its data on the formed data association pairs. The set of states is used to perform a multi-stage screening process to determine the optimal state region. This process aims to identify a subset of states from a large number of historical production states that can both stably produce qualified products and achieve low energy consumption.

[0078] First, the target calibration unit 30 operates on a continuously updated database containing all data pairs from a recent period (e.g., the most recent production shift or the most recent batch of raw materials). This unit performs an initial screening on each record in the database, using quality signals as the screening criterion. The value of the quality signal. In one embodiment, the quality signal is a binary signal, where one value indicates that the product is qualified and the other value indicates that it is unqualified. The first screening operation retains all records where the quality signal is qualified, forming a qualified status set. This set contains all known low-dimensional latent space state vectors capable of producing qualified products and their corresponding energy consumption values, in the form of:

[0079] Next, the target calibration unit 30 is in the qualified state set Based on this, a second screening is performed, this time targeting energy consumption. This unit first extracts the set of qualified states. All energy consumption signal values ​​E in t The distribution of these energy consumption values ​​is statistically analyzed to calculate a preset lower quantile. This preset quantile is a configurable system parameter, for example, set to 25%, i.e., the lower quartile. The energy consumption value corresponding to this quantile is determined as the energy consumption threshold E. thteshold .

[0080] The second screening operation will iterate through the set of qualified states. Of all records, only those energy consumption signal values ​​E are retained. t Less than or equal to the energy consumption threshold E thresholdThe records obtained through this screening are defined as the elite state set. This set represents the optimal subset of operating conditions in historical production, namely those production states that produced qualified products with the lowest energy consumption cost.

[0081] Finally, the target calibration unit 30 determines the target based on the elite state set. To determine the specific parameters of the optimal state region, the central target point z in the latent space of this region is determined. golden We obtain the following by calculating the arithmetic mean of all low-dimensional latent space state vectors in the elite state set:

[0082]

[0083] in, It is the total number of vectors in the elite state set. This is the central target point z. golden This will serve as the core optimization objective for control unit 40. Furthermore, to fully describe the morphology of this region, an elite state set can also be calculated. The covariance matrix describes the distribution range of the optimal state in the latent space and the correlation between its dimensions, and can be used for weighting the cost function in the subsequent control unit.

[0084] The target point z at the center of the optimal state region calculated by the target calibration unit 30 golden It is not static and unchanging, but rather recalibrated periodically or through event-driven updates via a dynamic update mechanism. This mechanism ensures that the system's control objectives can continuously adapt to slowly changing production conditions such as raw material batch fluctuations, equipment wear and tear, and environmental changes.

[0085] In one embodiment, the dynamic update mechanism includes two triggering methods: periodic triggering and event triggering.

[0086] The periodic triggering method is based on a preset, configurable time interval or production quantity. For example, the system can be configured to automatically trigger an update process at the end of each production shift (e.g., every 8 hours) or after the production of a certain length (e.g., 10,000 meters) of welded pipe. When the triggering condition is met, the target calibration unit 30 will use all data association pairs collected and stored during this period as a new data basis to re-execute the complete screening and calculation process described above, thereby obtaining a new central target point z. golden .

[0087] Event-triggered methods provide immediate responses to changes in specific operating conditions. One type of event is a definite external event; for example, when the production line changes to a new steel strip coil, the Manufacturing Execution System (MES) sends a signal to this system, which directly triggers the update process of the target calibration unit 30. Another type of event is a process drift event detected internally by the system. The target calibration unit 30 continuously monitors the real-time low-dimensional latent space state vector z output by the state representation unit 20. t The statistical distribution. It calculates z over a sliding time window (e.g., the most recent 1000 state vectors). t The mean and covariance of z are compared with those of the previous calculation. golden Elite state set used at that time The statistical distributions of the two are compared. If the Mahalanobis distance or KL divergence between the two exceeds a preset threshold, it is determined that the production process has drifted significantly, thereby triggering an update process.

[0088] Regardless of the triggering method, once the update process is initiated, the target calibration unit 30 will recalculate the center target point z using a defined, up-to-date dataset (e.g., all data collected since the last update). golden After the calculation is completed, the newly generated target point will replace the old target point and be immediately transmitted to the control unit 40. At the beginning of the next control cycle, the control unit 40 will seamlessly adopt this new target point as the core objective of its model predictive control optimization problem, thereby guiding the production process to a new optimal state region adapted to the current operating conditions.

[0089] See attached document Figure 1 The control unit 40 in this embodiment of the invention will now be described in detail. The control unit 40 implements its functions through a Model Predictive Controller (MPC). In one specific embodiment, the MPC logically consists of three core parts: a state transition model for predicting the future behavior of the system; a numerical optimization solver for solving constrained optimization problems online; and execution logic for implementing rolling time-domain optimization. This execution logic repeats the steps of "state measurement - optimization solution - application of the first control command" in each control cycle.

[0090] The state transition model is mathematically defined as a function f(·), whose input is the low-dimensional latent space state vector z at the current time t. t and the control command vector u applied to the production equipment at the same time t Its output is a prediction of the latent space state at the next time step t+1. The relationship is as follows:

[0091]

[0092] Wherein, the control command vector u t It is a vector containing all adjustable process parameter settings, such as the output power setting of the high-frequency welding machine, the pressing amount setting of the extrusion roll, and the mill line speed setting.

[0093] In one specific embodiment, the function f(·) is implemented using a pre-trained recurrent neural network (RNN), specifically a long short-term memory (LSTM) network structure. LSTM networks, due to their internal gating mechanisms (input gate, forget gate, output gate) and cell states, are capable of effectively learning and memorizing long-term dependencies in time-series data, which is essential for modeling the dynamic characteristics of industrial processes.

[0094] The input layer of this LSTM model receives a vector z from the current latent space state at each time step t. t and control command vector u t An augmented vector is constructed by concatenating elements along the feature dimension. This augmented vector is then passed through one or more LSTM layers. The output of the LSTM layer, i.e., its hidden state, is fed into a standard feedforward fully connected layer (also known as a dense layer). The output dimension of this fully connected layer is set to be the same as the hidden space dimension J, and it does not use a non-linear activation function (or uses a linear activation function). Its output is the prediction of the state at the next time step.

[0095] The network parameters of the state transition model are also determined through offline training. The training dataset consists of a series of time-series segments extracted from historical production data. Each data segment contains a continuous latent space state vector and its corresponding actual applied control command vector, i.e., in the form {(z... t ,u t ),(z t+1 ,u t+1 The sequence is ), ...}. The goal of training is to minimize the error of the model's prediction in one step, and its loss function is usually calculated using the mean squared error (MSE) to calculate the predicted value. Compared with the actual observed value z t+1 The difference between them. The training process uses the backpropagation over time (BPTT) algorithm and a stochastic gradient descent optimizer (such as Adam) to update the network weights until the model's prediction error on the validation set converges to a preset range.

[0096] The core task of the control unit 40 in each control cycle t is to solve a finite-time optimization problem based on the trained state transition model in order to calculate the optimal control command to be applied at the current moment.

[0097] The optimization problem is conceived over two predefined time scales: a prediction time domain N. p and a control time domain N c , where N c ≤N p Predicting the time domain N p This refers to the number of steps the model takes to predict the system state forward; N in the control time domain c It refers to the length of the future control instruction sequence that the optimization problem needs to calculate.

[0098] At each time tt, the goal of control unit 40 is to find an optimal sequence of control commands. This sequence can minimize a comprehensive cost function J. The specific form of the comprehensive cost function J is as follows:

[0099]

[0100] in, This is the control sequence to be optimized.

[0101] The cost function J consists of three weighted terms:

[0102] State trajectory tracking item: This calculation computes the future state trajectory predicted by the state transition model within the prediction time domain. The optimal state center point z provided by the target calibration unit 30 golden The weighted quadratic error between them. Weight matrix W z It is a positive definite or semi-positive definite diagonal matrix. The values ​​of its diagonal elements reflect the importance of each dimension component of the latent space state vector. The larger the value, the higher the cost will be due to deviation in that dimension.

[0103] Energy consumption control items: This calculation measures the weighted norm of the control command vector to be applied within the control time domain. Its physical meaning lies in penalizing the absolute amplitude of the control command, aiming to achieve the control objective with a smaller control cost, thereby indirectly optimizing the energy consumption of the actuator. The weight matrix W... u The setting is related to the physical unit and economic cost of the specific control quantity.

[0104] Control of rate of change: Where Δu t+k|t =u t+k|t -u t+k-1|t This calculation measures the weighted norm of the change in control commands between adjacent time steps. Its purpose is to suppress drastic fluctuations in control commands, ensure the smoothness of the control process, and avoid excessive mechanical shock and wear on physical actuators (such as motors and cylinders) on the production line.

[0105] Solving this optimization problem also requires satisfying a series of constraints:

[0106] Dynamic constraints: in The actual state z at the current moment t This constraint ensures that the predicted state trajectory follows the established state transition model.

[0107] Control input constraints: u min ≤u t+k|t ≤u max This constraint ensures that the calculated control command values ​​are within the operating range allowed by the physical actuator.

[0108] Control of rate of change constraint: Δu min ≤Δu t+k|t ≤Δu max This constraint limits the maximum adjustment range of the control command within a single step.

[0109] Since the state transition model f(·) is a nonlinear neural network model, the above optimization problem is a nonlinear and constrained optimization problem. The control unit 40 integrates a numerical optimization solver, employing iterative algorithms such as Sequential Quadratic Programming (SQP) or the interior-point method to efficiently solve this problem within each control cycle, thereby obtaining the optimal control command sequence.

[0110] After the nonlinear optimization problem is solved in each control cycle t, the control unit 40 will obtain a value in the current control time domain N. c Intra-optimal control command sequence However, according to the core strategy of model predictive control, namely the Receding Horizon Control strategy, this sequence will not be fully executed.

[0111] In one specific embodiment, the control unit 40 selects only from this optimal control command sequence. Extract the first element, i.e. This will be taken as the actual control command u that needs to be applied to the production equipment at the current time t. t The vector u t It includes specific settings for all adjustable process parameters such as the output power of the high-frequency welding machine and the mill line speed. The control unit 40 sends these settings to the programmable logic controller (PLC) at the bottom of the production line or directly to the corresponding actuators via an industrial communication bus (e.g., via the OPC UA protocol).

[0112] The rest of the sequence Then it will be discarded.

[0113] In the next control cycle, i.e., at time t+1, the control unit 40 will receive a new low-dimensional latent space state vector z from the state representation unit 20, which has been obtained through actual measurement and mapping. t+1 At this point, the control unit 40 does not use the previously calculated control plan, but instead uses this new state measurement value z. t+1 As initial conditions, we reconstruct and solve a completely new time-domain N based on time t+1. p The optimization problem.

[0114] This cyclical process of measurement-optimization-execute the first step-discard margin-measure again is repeated continuously in each control cycle. This rolling optimization method constitutes a closed-loop feedback mechanism. It enables the controller to continuously use the latest system state feedback to correct its control decisions, thereby exhibiting good robustness to system dynamics not fully captured by the model, external disturbances, and measurement noise, ensuring the stability and adaptability of the entire intelligent energy management optimization system in actual production environments.

[0115] See attached document Figure 2 , Figure 2 This is a flowchart of an intelligent energy consumption management and optimization method for welded pipe production process according to an embodiment of the present invention. The method corresponds to the intelligent energy consumption management and optimization system for welded pipe production process in the foregoing embodiments, and specifically includes the following steps:

[0116] Step S201: Synchronously acquire multimodal associated signals during the welded pipe production process and process them into high-dimensional feature vectors.

[0117] In one embodiment, the step includes: acquiring high-frequency electromagnetic signals, mechanical vibration signals, etc., by using sensors deployed at locations such as induction welding machines and key mechanical components, and combining them with conventional process parameters obtained from the PLC; and fusing them into a unified high-dimensional feature vector at each timestamp through processing such as filtering, feature extraction, and normalization.

[0118] Step S202: Map the high-dimensional feature vector to a low-dimensional latent space state vector that represents the overall state of the current production process.

[0119] In one embodiment, this step uses a pre-trained nonlinear dimensionality reduction model (such as a variational autoencoder) to compress and map high-dimensional, redundant input features into a low-dimensional vector that can compactly represent the macroscopic state of the production process.

[0120] Step S203: Receive the quality signal and the system energy consumption signal provided by the downstream online non-destructive testing unit, and determine the optimal state region based on the correlation results between the quality signal, the energy consumption signal and the low-dimensional latent space state vector.

[0121] In one embodiment, the step includes: first, compensating for the time delay of the quality signal based on the production line speed to achieve precise alignment between the state and the result; then, identifying a subset of states that can stably produce qualified products and whose energy consumption is within a preset low quantile range through multi-stage screening; and finally, determining the center target point of the optimal state region based on the statistical characteristics (such as the mean) of the subset of states and dynamically updating the target point.

[0122] Step S204: Based on the low-dimensional latent space state vector and the optimal state region, a control command aimed at guiding the production process state to the optimal state region is generated through model prediction control strategy, and then applied to the production equipment.

[0123] In one embodiment, the step includes: in each control cycle, taking the current state as the initial condition and the center target point of the optimal state region as the optimization objective, predicting the future state trajectory through a state transition model, and solving an optimization problem of a comprehensive cost function aimed at minimizing state tracking error, control energy consumption, and control change rate; then, according to the rolling time domain control principle, issuing only the first instruction of the optimal control instruction sequence obtained by the solution to the relevant actuators of the production line.

[0124] Example:

[0125] To further clarify the content of this invention, a specific application example is provided below. This example describes the application of the system and method of this invention to an actual high-frequency straight seam welded pipe (ERW) production line to solve the problems of unstable welding quality and high unit energy consumption when dealing with different batches of steel strip coils.

[0126] In this embodiment, the target production line has extremely high requirements for weld quality consistency. The adjustable control command u of the control unit 40... t Limited to two key process parameters: the output power P of the high-frequency welding machine weld The extrusion pressure setting value F of the extrusion roller squeeze .

[0127] The acquisition unit 10 integrates the following data sources:

[0128] 1. Install a high-frequency electromagnetic probe downstream of the induction coil and upstream of the extrusion roller to collect electromagnetic field leakage signals in the weld area at a sampling rate of 20kHz.

[0129] 2. Install a piezoelectric accelerometer on the bearing housing of the extrusion roller to collect its radial vibration signal.

[0130] 3. Install an infrared thermometer in the weld cooling zone to measure the surface temperature of the weld after it has formed.

[0131] 4. The high-frequency welding machine output power P is read in real time from the main PLC on the production line via the OPC UA communication protocol. weld Extrusion pressure F squeeze and production line speed v line .

[0132] In each control cycle (200 milliseconds), the acquisition unit 10 performs a fast Fourier transform on the acquired electromagnetic signal to obtain 256 frequency point amplitudes, calculates the root mean square value of the vibration signal, and normalizes these features together with temperature and PLC readings, then fuses them into a high-dimensional feature vector x with dimension D = 260. t .

[0133] The state representation unit 20 employs a pre-trained variational autoencoder (VAE) with its latent space dimension set to J=2 for ease of visualization and analysis. This VAE model is trained using historical production data and can represent x with a latent space dimension of D=260. t Mapped to a two-dimensional latent space state vector z t =[z1,z2] T .

[0134] See attached document Figure 3 , Figure 3 This is a low-dimensional latent space state distribution diagram according to an embodiment of the present invention. The diagram is a two-dimensional scatter plot showing the distribution of different production conditions in the two-dimensional latent space. Each point in the diagram represents a production state at a historical moment. By correlating with historical quality data, it can be found that states of normal, stable production and qualified products (represented by blue dots in the diagram) form a dense and clearly defined main cluster in the latent space. States with historically occurring burn-through defects (represented by red crosses in the diagram) are clustered in the upper left corner of the main cluster, while states with incomplete solder joint defects (represented by green plus signs in the diagram) are distributed in the lower right corner of the main cluster. This diagram shows that the two-dimensional latent space constructed by the present invention can effectively distinguish different macroscopic states of the production process, and the "distance" between states reflects, to some extent, the similarity of the operating conditions.

[0135] Target calibration unit 30 continuously receives and buffers z generated by state characterization unit 20 t and the corresponding energy consumption signal E t (This is the actual output power of the high-frequency welding machine). Simultaneously, it also receives binary quality signals from the online eddy current testing (ECT) unit at the end of the production line. (Pass / Fail)

[0136] See attached document Figure 4 , Figure 4 This is a schematic diagram of the optimal state region calibration according to an embodiment of the present invention. This diagram is also a two-dimensional latent space diagram. The target calibration unit 30 first determines the optimal state region based on the linear velocity v. line And physical distance compensation for time delay, will (z t E t ) and its corresponding The data is correlated. In the diagram, gray dots represent all historical data points that can produce qualified products, forming a set of qualified states. Next, the unit calculates the lower quartile (i.e., the 25th percentile) of the energy consumption value corresponding to these gray dots, and sets this value as the energy consumption threshold E. threshold In the diagram, qualified states with energy consumption below the threshold are highlighted as green dots, forming the elite state set. Finally, the geometric center (mean) of all green dots is calculated to obtain the central target point z of the optimal state region. golden This is indicated by a red star in the diagram. This calibration process is automatically triggered and updated when the steel strip coil is changed to accommodate the characteristics of the new material.

[0137] The control unit 40 loads a pre-trained LSTM-based state transition model and uses a calibrated z-axis. golden As an optimization target.

[0138] See attached document Figure 5 , Figure 5 This is a schematic diagram of the trajectory of the model predictive control process according to an embodiment of the present invention. The diagram illustrates the current state z when the production process is disturbed (power grid voltage fluctuations cause a momentary drop in welding power). t The system's response behavior when deviating from the optimal region. In the diagram, the current state point (blue solid circle) is located to the lower right of the optimal target point (red star), in a region historically associated with solder joint defects. At the current moment, the Model Predictive Controller (MPC), based on the state transition model and optimization solution, plans an optimal state trajectory (represented by the red dashed line in the diagram) for the future prediction time domain. This trajectory returns to the target point in the fastest and smoothest way. Based on this, the controller calculates the optimal control command sequence to be applied and sets the first command... (i.e., the adjusted P) weld and F squeeze The value is then sent to the actuator. In the following control cycles, the actual state trajectory of the system (represented by the solid blue line in the figure) can be seen to be effectively guided, gradually approaching and eventually stabilizing around the target point marked by the red star.

[0139] See attached document Figure 6 , Figure 6 This is a time-series comparison chart of the application effects according to an embodiment of the present invention. The chart, through two sub-charts sharing a time axis, specifically illustrates the evolution of key performance indicators (KPIs) during two roughly equal observation periods before and after the intelligent energy consumption management and optimization system of the present invention was implemented on an actual production line for producing welded pipes of the same specification. A vertical dashed line marks the system's activation time, dividing the entire observation period into two stages: "before application" and "after application."

[0140] The subplot above illustrates the change in the hourly defect rate of the product. Before the system was implemented, the defect rate curve showed a high mean and significant fluctuations, reflecting instability in the production process. After the system was implemented, the defect rate was consistently suppressed to an extremely low level, and its volatility was significantly reduced, indicating a significant improvement in product quality consistency.

[0141] The subplot below illustrates the change in production energy consumption per unit length of welded pipe (in kJ / m). Similar to the defect rate, before the system was implemented, the energy consumption curve was also at a high level and fluctuated greatly. After the system was implemented, the average energy consumption per unit length was significantly reduced, and the curve became smoother and more stable. This demonstrates that the present invention not only achieves the goal of energy saving and consumption reduction, but also makes energy consumption control more precise and predictable by stabilizing the production process, thereby achieving stable and economical operation of the production process.

[0142] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent energy consumption management and optimization system for welded pipe production process, characterized in that, include: The acquisition unit is used to synchronously acquire multimodal associated signals during the welded pipe production process and process them into high-dimensional feature vectors; The state representation unit, connected to the acquisition unit, includes a pre-trained variational autoencoder model for mapping the high-dimensional feature vector to a low-dimensional latent space state vector representing the overall state of the current production process. The encoder outputs a mean vector μ and a log-variance vector from the variational autoencoder model, and these are expressed using a reparameterization formula. Generate the low-dimensional latent space state vector, where, Let σ be a random noise vector sampled from a standard normal distribution, σ be the standard deviation vector, and ⊙ denote element-wise multiplication; The target calibration unit receives quality signals and system energy consumption signals from the downstream online non-destructive testing unit, based on the production line speed v. line (t) and physical distance L NDT Calculate the estimated arrival time , where Δt delay Given a time delay, the low-dimensional latent space state vector z at time t... t Compared with the expected arrival time t arrival The quality signal and energy consumption signal are time-aligned; based on the aligned data, low-dimensional latent space state vectors with qualified quality signals and energy consumption signal values ​​meeting preset conditions are selected to form an elite state set, and the center of this elite state set is calculated as the central target point z of the optimal state region. golden ; The control unit, connected to the state characterization unit and the target calibration unit, includes a model prediction controller for generating and outputting control commands to the production equipment that aim to guide the production process state to the optimal state region based on the low-dimensional latent space state vector and the optimal state region. The model predictive controller includes a state transition model based on a recurrent neural network, used to determine the state vector z in the low-dimensional latent space at the current time t. t and control command u t According to the formula Predict the hidden space state at the next moment; The model predictive controller, at each time t, seeks an optimal sequence of control commands that aims to minimize the overall cost function J by solving a finite-time optimization problem. The first instruction of the sequence is output to the production equipment; the comprehensive cost function J includes a state trajectory tracking term, a control energy consumption term, and a control change term, wherein the state trajectory tracking term is... Energy consumption control items are ; Control change items are ; The formula for the comprehensive cost function J is: ； in, The sequence of control instructions to be optimized; N p For prediction in the time domain; N c To control the time domain; To control the amount of change in instructions between adjacent time steps, W z W u W Δu These are the corresponding weight matrices.

2. The intelligent energy consumption management and optimization system for welded pipe production process according to claim 1, characterized in that, The acquisition unit is specifically used to acquire multimodal associated signals, including at least the high-frequency electromagnetic harmonic signals of the high-frequency welding machine power supply and the mechanical vibration signals of key mechanical components.

3. The intelligent energy consumption management and optimization system for welded pipe production process according to claim 1 or 2, characterized in that, The target calibration unit is configured to periodically or dynamically update the optimal state region when production conditions change, and provide the updated optimal state region to the control unit as its optimization target.

4. A method for intelligent energy consumption management and optimization in welded pipe production, implemented based on the intelligent energy consumption management and optimization system for welded pipe production as described in claim 1 or 2, characterized in that, Includes the following steps: Multimodal associated signals during the welded pipe production process are acquired synchronously and processed into high-dimensional feature vectors; The high-dimensional feature vector is mapped to a low-dimensional latent space state vector that represents the overall state of the current production process. The system receives the quality signal and the system's energy consumption signal from the downstream online non-destructive testing unit, and calibrates the optimal state region based on the correlation results between the quality signal, the energy consumption signal and the low-dimensional latent space state vector. Based on the low-dimensional latent space state vector and the optimal state region, a model predictive control strategy is used to generate control commands aimed at guiding the production process state to the optimal state region, and these commands are applied to the production equipment.

Images