Intelligent control system for pickling steam based on edge computing and lstm neural network

Abstract

The present application relates to the technical field of metal surface treatment, and relates to an intelligent control system for pickling steam based on edge computing and an LSTM neural network. The technical problem to be solved is that existing pickling steam control relies on lagging feedback control, steam supply is disconnected from actual heat load, and active intervention is impossible, leading to steam waste and unstable pickling quality. The technical solution comprises a multi-source sensing module, an edge computing unit, a steam supply adjusting device, and a heat energy buffering device. The edge computing unit is deployed on site to run an LSTM model, taking the visual grade, thickness, plate temperature, acid temperature, width, incoming speed, and valve opening of the strip steel oxide scale as input, directly outputting future multi-time-step steam flow prediction values and valve pre-opening signals, and driving the steam supply adjusting device to pre-act in advance before the prediction time step, and simultaneously starting the heat energy buffering device to store heat, so as to realize advanced prediction and active control of steam heat demand.

Description

Technical Field

[0001] This application relates to the field of metal surface treatment technology, and more specifically to an intelligent control system for pickling steam based on edge computing and LSTM neural network. Background Technology

[0002] In continuous pickling production lines for cold-rolled strip steel, the stable control of acid solution temperature directly affects pickling efficiency and strip surface quality. Currently, mainstream processes generally use steam heating combined with traditional PID feedback control or manual valve adjustment to maintain acid solution temperature. The system monitors the acid solution temperature in real time through temperature sensors placed in the pickling tank. When the detected value deviates from the set target, the controller adjusts the steam valve opening based on the error signal to achieve temperature compensation. Although some improved solutions introduce artificial neural network models, such as Chinese patent CN104942015A which uses a neural network to link acid solution temperature, concentration, and acid pump flow rate, or CN119087931A which uses strip thickness, plate temperature, and iron oxide scale thickness as inputs to predict pickling solution concentration and pickling speed, their core remains limited to feedback adjustment or indirect control of pickling process parameters, without directly addressing the critical energy consumption aspect of steam energy consumption through forward-looking modeling.

[0003] In the aforementioned traditional technologies, because the control logic relies on a serial response mechanism of "detection-feedback-adjustment," the system response exhibits significant lag when the thickness of the incoming strip steel changes abruptly, the feed speed increases rapidly, or the surface iron oxide scale distribution is uneven, leading to a sharp rise in heat load. Often, steam supply is only increased after the acid solution temperature has already fallen significantly, resulting in insufficient pickling in certain areas or excessive steam emissions. Furthermore, existing control schemes primarily focus on predicting indirect process parameters such as pickling solution concentration or pickling speed, rather than directly predicting the time-series changes in steam heat demand, causing a disconnect between steam supply and actual heat load. In addition, most neural network models are deployed on host computers or in the cloud, and due to industrial network latency and bandwidth limitations, they cannot meet the millisecond-level real-time control requirements; the models are mostly trained offline and run in a fixed manner, lacking the ability to adaptively optimize based on continuously accumulated production data. Crucially, none of the existing solutions implement "pre-opening valve" action—that is, pre-setting the steam valve opening 2-3 minutes before heat load changes occur, thereby transforming control from a passive response to proactive intervention. Meanwhile, existing systems generally lack multimodal visual perception capabilities for the thickness of iron oxide scale and the degree of residual oil on the surface of strip steel, and fail to incorporate this key thermal resistance variable into the temperature control model, resulting in a significant decrease in control accuracy under extreme incoming material conditions, serious steam waste, and a persistently high pickling defect rate. Summary of the Invention

[0004] To address the technical deficiencies in the prior art, this application provides an intelligent control system for pickling steam based on edge computing and LSTM neural networks, comprising: a multi-source sensing module, an edge computing unit, a steam supply regulating device, and a thermal energy buffering device; the multi-source sensing module is used to collect the visual grade of iron oxide scale on the strip, the strip thickness, the strip temperature, and the acid temperature; the steam supply regulating device is used to regulate the steam supply in response to the valve pre-opening signal; and the thermal energy buffering device is used to buffer instantaneous heat load fluctuations.

[0005] The edge computing unit is deployed on the pickling production line and runs an LSTM neural network model. The LSTM neural network model is configured to take multimodal feature vectors, including the visual grade of iron oxide scale, strip thickness, strip temperature, acid temperature, strip width, feed speed and current steam valve opening, as input and directly output the predicted steam flow rate and the corresponding valve pre-opening signal for multiple time steps within a preset time period.

[0006] The edge computing unit is further configured to: drive the steam supply regulating device to perform pre-action according to the valve pre-opening signal, and simultaneously instruct the thermal energy buffer device to start heat storage, before the time step corresponding to the predicted steam flow value by a preset advance period.

[0007] Optionally, in one possible implementation of the technical solution of this application, the LSTM neural network model includes three LSTM hidden layers, each with 64-128 neurons, the activation function is tanh, and the output layer is a fully connected layer; the LSTM neural network model outputs the predicted steam flow rate and the corresponding valve pre-opening percentage for 11 time points in 1-minute increments within the next 5-15 minutes.

[0008] Each LSTM hidden layer is equipped with a gating mechanism, including an input gate, a forget gate, and an output gate. The weight parameters of the LSTM neural network model are updated during the training phase through the backpropagation algorithm. The optimizer is Adam, the initial learning rate is 0.001, and it decays by 10% every 50 training epochs. The batch size is 32.

[0009] Optionally, in one possible implementation of the technical solution of this application, the edge computing unit calls the LSTM neural network model for inference at a frequency of 1Hz, and drives the steam supply regulating device to perform pre-action 2-3 minutes in advance based on the valve pre-opening signal corresponding to the 3rd minute in the predicted output sequence; after each inference, the edge computing unit records the current timestamp, input feature vector, predicted output sequence and actual valve response delay as training samples and stores them in a local circular buffer.

[0010] When a sudden change in pickling speed or a drastic change in incoming material thickness is detected, the edge computing unit automatically increases the inference frequency of the LSTM neural network model to 2Hz and enables a dynamic attention mechanism: by calculating the Euclidean distance between the current input feature vector and the historical mean feature, when the Euclidean distance exceeds a preset threshold, the attention scores for the feed speed feature and the strip thickness feature are multiplied by a gain coefficient of 1.5.

[0011] Optionally, in one possible implementation of the technical solution of this application, after each batch of strip steel pickling is completed, the edge computing unit uses the actual steam consumption and pickling quality assessment data of the batch as a supervision signal to trigger the online incremental learning of the LSTM neural network model. The online incremental learning adopts the mini-batch gradient descent method, and each update uses 5 sets of input and output samples of the current batch. The learning rate is set to 1 / 10 of the initial learning rate of offline training, and the weight update is achieved by freezing the first two LSTM hidden layers and only fine-tuning the third LSTM hidden layer and output layer.

[0012] Optionally, in one possible implementation of the technical solution of this application, the edge computing unit runs multiple LSTM sub-models simultaneously. The multiple LSTM sub-models are independently trained on three datasets: cold-rolled low-carbon steel, high-strength steel, and stainless steel. Each sub-model has the same network structure, only the output layer weight parameters are different. The edge computing unit automatically switches to the LSTM sub-model of the corresponding material category through the built-in rule engine based on the strip material identification result at the inlet, so as to realize material-adaptive steam demand prediction.

[0013] Optionally, in one possible implementation of the technical solution of this application, the system further includes a fault diagnosis module; the fault diagnosis module judges the abnormal state by verifying the reasonable range of data from each sensor in the multi-source sensing module, the data update frequency and the communication heartbeat packet status. When the edge computing unit's inference delay exceeds 200ms for three consecutive times or any sensor data is zero for 5 consecutive frames, a fault flag is triggered, the steam control mode is automatically switched from LSTM predictive control to traditional PID feedback control mode, and an alarm is triggered.

[0014] Compared with the prior art, the technical solution of this application has the following beneficial effects:

[0015] Applying the technical solution of this application, a multi-spectral visual sensor, an incoming material thickness gauge, an infrared thermal imager, and a distributed temperature sensor array collaboratively collect multi-modal real-time data such as the visual grade of iron oxide scale on the strip surface, strip thickness, current strip temperature, and current acid temperature. Combined with process parameters such as strip width and feed speed, a comprehensive input feature set is constructed and input into an LSTM neural network model deployed on an edge computing unit. This enables accurate time-series prediction of steam heat demand per minute for the next 5-15 minutes, and simultaneously generates corresponding valve pre-opening signals. The system is further configured to... The electric proportional steam control valve is actively driven to pre-act within 2-3 minutes, and the backup energy storage tank is preheated and activated to store heat. This allows for dynamic pre-configuration of heat supply before the actual heat demand of the pickling process occurs, completely breaking through the traditional control mode that relies on delayed feedback regulation. It realizes the advanced prediction and execution closed loop of steam heat demand based on multi-source sensing and deep time-series modeling, effectively solving the problems of insufficient or excessive steam supply, energy waste, and pickling quality fluctuations caused by response delay in existing technologies. This achieves comprehensive technical effects such as improving steam utilization efficiency, stabilizing the temperature field of the pickling process, and reducing operating energy consumption and carbon emissions. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a system structure of an intelligent control system for pickling steam based on edge computing and LSTM neural network in an embodiment of this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. 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 should fall within the scope of protection of the present invention.

[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] As described in the background section, existing pickling steam control systems rely on a lag control mode of "detection-feedback-regulation". Although some neural networks are introduced to predict pickling solution concentration or pickling speed, they still fail to directly model the steam heat demand. This results in a slow response of steam supply when there are sudden changes in incoming materials, causing steam waste and fluctuations in pickling quality. Furthermore, the model is deployed on a host computer or in the cloud and lacks local real-time inference capabilities, making it impossible to achieve forward-looking intervention in changes in heat load.

[0020] After extensive research, the inventors discovered that the most fundamental reason for the aforementioned predicament lies in the fact that existing technologies treat steam regulation as a passive compensation for "already occurring temperature deviations," rather than an active prediction of "future changes in heat load." The fundamental contradiction is that the action of steam valves always lags behind the actual changes in heat load. These changes (such as a sudden increase in strip thickness, faster feed speed, or uneven distribution of iron oxide scale) have clear time-series evolution characteristics. If this evolution trend could be perceived and predicted in advance, steam supply adjustments could be completed before temperature deviations occur. However, existing technologies neither use steam flow rate as a direct prediction target, nor construct a low-latency prediction-execution closed loop at the edge, nor implement a pre-intervention logic that "triggers valve action 2-3 minutes before the prediction time point." This results in the system always being in a passive state of "chasing after temperature," unable to overcome the physical bottleneck of response delay.

[0021] Based on the above findings, the inventors proposed a control concept that uses steam heat demand as a direct predictive output and implements "pre-opening valves 2-3 minutes in advance" at the edge. The core of this concept lies in directly predicting the required steam flow rate and corresponding valve opening per minute for the next 5-15 minutes, instead of using acid temperature or pickling rate as the prediction endpoint. Simultaneously, an LSTM neural network model is deployed on the field edge computing unit, enabling local inference with millisecond-level response capabilities. This forces the control action to be executed 2-3 minutes before the predicted future time point, thus completely transforming "feedback regulation" into "feedforward intervention." This concept breaks through the traditional control paradigm, binding time series prediction and pre-action execution into an inseparable closed-loop logic for the first time. It transforms steam supply actions from "remedies after temperature deviation" to "energy storage preparation before heat load increases," fundamentally eliminating the problem of system response lag.

[0022] Guided by this creative concept, the inventors proposed an intelligent control system for pickling steam based on edge computing and LSTM neural networks.

[0023] Example 1: In a continuous pickling production line for cold-rolled strip steel, the pickling process requires high stability of the acid solution temperature. Traditional steam control relies on temperature sensors to collect the acid solution temperature and compare it with the set value to generate a deviation signal, which is then used by a PID controller to adjust the steam valve opening. This "detection-feedback-adjustment" serial control mode has inherent response lag. When the thickness of the incoming strip steel changes abruptly, the feed speed changes drastically, or the surface iron oxide scale is unevenly distributed, the system cannot predict and intervene before the heat load changes, leading to a disconnect between steam supply and actual heat demand, resulting in problems such as acid solution temperature fluctuations, excessive steam consumption, or insufficient pickling in certain areas. Furthermore, while some existing solutions introduce neural networks for prediction, their prediction targets are mostly indirect process parameters such as acid solution concentration or pickling speed, not steam heat demand as a direct prediction target. Moreover, the models are mostly deployed on host computers or in the cloud, and are limited by industrial network latency and bandwidth, making it difficult to meet the millisecond-level real-time control requirements. Therefore, a system solution is needed that can directly predict steam heat demand, perform real-time inference at the edge, and proactively pre-adjust the steam supply before heat load changes.

[0024] Example 1: As Figure 1 As shown, the intelligent control system for pickling steam based on edge computing and LSTM neural network includes: a multi-source sensing module, an edge computing unit, a steam supply regulating device, and a thermal energy buffer device; the multi-source sensing module is used to collect the visual grade of iron oxide scale on the strip, the strip thickness, the strip temperature, and the acid temperature; the steam supply regulating device is used to regulate the steam supply in response to the valve pre-opening signal; and the thermal energy buffer device is used to buffer instantaneous heat load fluctuations.

[0025] The edge computing unit is deployed on the pickling production line and runs an LSTM neural network model. The LSTM neural network model is configured to take multimodal feature vectors, including the visual grade of iron oxide scale, strip thickness, strip temperature, acid temperature, strip width, feed speed, and current steam valve opening, as input, and directly output the predicted steam flow rate and the corresponding valve pre-opening signal for multiple time steps within a preset time period. The edge computing unit is further configured to drive the steam supply regulating device to perform pre-action based on the valve pre-opening signal, one preset lead time period before the time step corresponding to the predicted steam flow rate, and simultaneously instruct the thermal energy buffer device to start heat storage.

[0026] The following provides a detailed explanation of each component of the above technical solution and its specific implementation method.

[0027] In this embodiment, the multi-source sensing module includes a multispectral visual sensor, an incoming material thickness gauge, an infrared thermal imager, and a distributed temperature sensor array. The multispectral visual sensor is used to acquire the visual grade of iron oxide scale on the strip surface. This sensor is installed 1.5m from the pickling tank inlet, with a viewing angle covering the full width of the strip, and a sampling frequency of 10 frames / second. It is used to identify the visual grade of iron oxide scale from 1 to 5, where grade 1 indicates no iron oxide scale and grade 5 indicates severe iron oxide scale accumulation. The multispectral visual sensor includes a visible light channel, a near-infrared channel, and a short-wave infrared channel. Each channel is equipped with an independent CMOS image sensor and a bandpass filter, with a spectral response range covering 450nm-1700nm. The visible light channel identifies surface morphology and color changes, the near-infrared channel captures the moisture and structural characteristics of the iron oxide scale, and the short-wave infrared channel is sensitive to differences in the crystal phase composition and thermal radiation of iron oxides. The synergy of these three channels significantly improves the accuracy of iron oxide scale visual grade identification in complex industrial environments such as oil contamination, light fluctuations, and background interference. In practical applications, the sensor can also be selected from other types of multispectral imaging devices, and this application embodiment does not limit this.

[0028] The incoming material thickness gauge is used to obtain the thickness of the strip steel. Through a non-contact laser thickness measurement method, it measures the physical dimensions of the strip steel entering the pickling production line, directly obtaining the thickness value of the strip steel before entering the pickling tank. This value is collected as a process parameter and used for subsequent steam heat demand prediction, providing the system with quantitative input on the physical properties of the strip steel to support the perception and response to changes in heat load.

[0029] An infrared thermal imager is used to acquire the current temperature of the strip steel. This infrared thermal imager is installed 0.3m downstream of the incoming thickness gauge and directly captures the thermal radiation signal emitted from the strip steel surface using non-contact infrared radiation detection technology, converting it into a corresponding temperature value. The infrared thermal imager has a temperature measurement range of 0-300℃, an accuracy of ±1℃, and employs an uncooled microbolometer focal plane array with a pixel resolution of 320×240, a frame rate of no less than 30Hz, a field of view of 25°×20°, and a focal length of 12mm. This parameter configuration enables the infrared thermal imager to stably and completely capture the surface temperature distribution of the strip steel even under high-speed strip steel feeding conditions, avoiding local temperature field loss due to insufficient resolution, dynamic temperature distortion due to excessively low frame rate, or missed detection in edge areas due to mismatched field of view. In practical applications, other models of thermal imaging equipment can also be selected for this infrared thermal imager; this embodiment does not limit this choice.

[0030] A distributed temperature sensor array is used to acquire the current acid solution temperature. This array deploys 6-8 PT100 temperature sensors every 2 meters along the length of the pickling tank, with a measurement accuracy of ±0.1℃. These sensors are used to acquire the acid solution temperature distribution at the inlet, middle, and outlet sections of the pickling tank. The PT100 temperature sensors are connected to the field instrument module using a four-wire connection, effectively eliminating systematic errors caused by lead resistance and ensuring the initial accuracy of the acid solution temperature acquisition. The weak analog signal acquired is then processed by a differential amplifier circuit to suppress common-mode interference, followed by high-resolution digitization by a 24-bit ΣΔ analog-to-digital converter, significantly improving the signal-to-noise ratio and dynamic range, and avoiding signal distortion in strong industrial electromagnetic environments. The processed digital signal is then transmitted differentially to the edge computing unit via an RS485 bus. The RS485 bus offers long-distance (up to 1200 meters), common-mode noise immunity, and multi-point networking capabilities, ensuring data integrity and real-time performance during long-distance transmission. This arrangement ensures that temperature information covers the three key areas of the pickling process: the inlet, middle, and outlet sections, providing the LSTM neural network model with acid temperature distribution data that has clear spatial dimensions. In practical applications, other platinum resistance thermometers or thermocouple sensors can also be selected for this temperature sensor; this application does not limit this choice.

[0031] In this embodiment, the visual grade of iron oxide scale is output after preprocessing visible and near-infrared images acquired by a multispectral vision sensor using a convolutional neural network. This convolutional neural network employs a lightweight MobileNetV3 architecture, containing 12 depthwise separable convolutional layers. During the training phase, supervised learning is performed using a dataset of labeled iron oxide scale thickness images. The training dataset contains 5000 labeled images, each labeled with an iron oxide scale thickness grade of 1–5 and a pixel-level mask. This preprocessing is performed at the image acquisition end, directly mapping the original multimodal images to structured grade features. This avoids the computational burden and latency caused by complex image parsing in edge computing units, ensuring that the visual grade of iron oxide scale input to the LSTM model has high consistency and low noise characteristics. In practical applications, other lightweight network structures (such as ShuffleNet, EfficientNet-Lite, etc.) can also be selected for this convolutional neural network; this embodiment does not limit this choice.

[0032] The edge computing unit is deployed on the pickling production line and employs an industrial-grade AI inference module. It is equipped with at least a quad-core ARM processor, 16GB of memory, and an NPU (Neural Processing Unit) acceleration unit supporting INT8 quantization, achieving millisecond-level inference latency. The NPU acceleration unit performs INT8 fixed-point quantization on the LSTM neural network model through quantization-aware training, mapping model weights and activation values ​​to an 8-bit integer space. The inference process uses a lookup table method instead of floating-point multiplication and addition operations, significantly reducing computational complexity and power consumption while preserving the model's high accuracy in pickling steam demand forecasting. This hardware configuration and quantization strategy enable the model to maintain millisecond-level response even in high-frequency scenarios with multiple rolling predictions per second, effectively overcoming the inference latency problem caused by traditional floating-point operations. In practical applications, this edge computing unit can also be selected from other embedded AI computing platforms (such as NVIDIA Jetson series, Huawei Atlas series, etc.), which are not limited in this embodiment.

[0033] The LSTM neural network model receives seven input features: strip thickness, strip width, feed speed, visual grade of iron oxide scale, current acid temperature, current steam valve opening, and current strip temperature. These input features constitute the model's real-time process status representation. Based on this input sequence, the model, through its internal time-dependent modeling capabilities, outputs the predicted steam flow rate and corresponding valve pre-opening signal for each minute within the next 5–15 minutes. The predicted steam flow rate is in kg / min, with a prediction range of 0–150 kg / min, and the corresponding valve pre-opening range is 0%–100%, achieving the time-series extrapolation of steam heat demand and the synchronous generation of control commands. The LSTM model's input feature vector also includes strip material and strip width, where the strip material is one of cold-rolled low-carbon steel, high-strength steel, or stainless steel, and the strip width ranges from 800 mm to 1800 mm. By introducing the material of the strip steel as a key input variable, the model can accurately capture the dynamic influence of different materials on steam heat demand based on the differences in thermal conductivity, specific heat capacity and iron oxide scale peeling characteristics, combined with the heat capacity per unit time and surface heat exchange efficiency affected by the strip steel width.

[0034] When the steam flow prediction value output by the LSTM model is converted into a steam valve opening signal, a saturation limiting mechanism is introduced to prevent the opening from exceeding the safe operating range of 10%–95%. This saturation limiting mechanism is implemented in the signal processing module of the edge computing unit: when the calculated opening is less than 10%, a forced output of 10% is applied; when it is greater than 95%, a forced output of 95% is applied, and the over-limit event is recorded to the log file. This limiting mechanism directly acts on the control signal output stage, completing signal safety calibration before driving the electric steam proportional regulating valve, ensuring the mechanical safety and response stability of the valve under long-term high-frequency regulation, and preventing equipment overload, seal failure, or fluid shock caused by over-opening or over-closing.

[0035] The steam supply regulating device is an electric steam proportional control valve. This electric steam proportional control valve connects to the edge computing unit using Modbus TCP or PROFINET industrial communication protocols, with a response delay of less than 50ms and an opening control accuracy of ±1%. Its actuator is a butterfly valve structure driven by a stepper motor, with the valve stem position fed back in real time by an encoder, and a closed-loop control cycle of 20ms. This high-precision closed-loop feedback mechanism enables the actual valve opening to accurately follow changes in the predicted value, effectively suppressing opening deviations caused by nonlinear mechanical response, transmission backlash, or position drift due to long-term operation. In practical applications, other actuator types (such as pneumatic diaphragm actuators, electric ball valves, etc.) and communication protocols (such as EtherCAT, EtherNet / IP, etc.) can also be selected for this electric control valve; this application embodiment does not limit these options.

[0036] The thermal energy buffer device is a preheating backup energy storage tank. This tank is made of acid-resistant steel and features an insulation layer. It is equipped with an electric auxiliary heating element and a temperature sensor to preheat the tank to 5–10°C above the target temperature before a predicted high heat load. The insulation layer is an aerogel composite material with a thickness of 50 mm and a thermal conductivity of less than 0.02 W / (m·K). The outer wall of the tank has a double-layer vacuum isolation structure. The insulation layer and the vacuum isolation structure work together to significantly suppress the three major heat loss paths: heat conduction, heat convection, and radiation, enabling the tank to maintain high thermal stability for a long period even without continuous heating. The heating power of the electric auxiliary heating element is dynamically adjusted by the edge computing unit based on the predicted peak steam flow rate. The heating command is controlled by a PWM (Pulse Width Modulation) signal with a duty cycle range of 10%–100%, achieving real-time matching between the heat storage temperature and the instantaneous heat load demand. In practical applications, other low thermal conductivity composite materials (such as vacuum insulation panels, nano-aerogel felts, etc.) can also be selected for thermal insulation materials, and this application does not limit this.

[0037] The technical solution operates as follows: First, a multispectral visual sensor, incoming material thickness gauge, infrared thermal imager, and distributed temperature sensor array collaboratively collect multimodal real-time data, including the visual grade of iron oxide scale on the strip surface, strip thickness, current strip temperature, and current acid solution temperature. This data, combined with process parameters such as strip width, feed speed, and current steam valve opening, forms a dynamic sensing input vector targeting the heat demand of the pickling process. Then, an LSTM neural network model deployed in the field edge computing unit uses this multimodal feature vector as input. Through joint modeling of multidimensional time-series data, it outputs a predicted steam flow rate per minute for the next 5–15 minutes and the corresponding valve pre-opening signal. Finally, 2–3 minutes before the predicted time step, the edge computing unit drives the electric steam proportional control valve to perform a pre-action based on the output valve pre-opening signal, while simultaneously sending a synchronization command to the preheating backup energy storage tank to initiate the heat storage process. This process advances the steam supply regulation behavior to before the actual heat load occurs, realizing a paradigm shift from "feedback regulation" to "feedforward intervention."

[0038] By employing the above technical solution, and through the collaborative acquisition of data from multi-source sensing modules and the local real-time inference of edge computing units, proactive prediction and control of steam heat demand are achieved, completely breaking through the traditional control mode that relies on delayed feedback. The opening adjustment of the steam valve no longer depends on real-time feedback of the acid temperature, but directly responds to the predicted results of future steam demand; the activation of the preheating backup energy storage tank is synchronized with the valve action, ensuring that the energy storage device has begun accumulating heat energy before the predicted peak heat demand arrives, providing immediate response capability for subsequent steam supply. This technical solution effectively solves the problems of delayed steam control response, heat supply and demand mismatch, energy waste, and pickling quality fluctuations in existing technologies, achieving a comprehensive technical effect of improving pickling efficiency, stabilizing the process temperature field, reducing steam energy consumption, and enhancing system response robustness.

[0039] Example 2: Based on the system described in Example 1, further consideration needs to be given to the specific structural design of the LSTM neural network model. Since the prediction of steam heat demand in the pickling process involves complex nonlinear coupling relationships among multiple time-series parameters such as strip thickness, width, feed rate, iron oxide scale visual grade, acid temperature, steam valve opening, and strip temperature, the depth, width, and training strategy of the model structure directly affect its ability to capture long-term dependencies and its prediction accuracy. If the model has too few layers or insufficient neurons, it cannot fully learn the deep temporal correlations between multimodal input features; if the training strategy is inappropriate, it may lead to slow model convergence or overfitting. Therefore, it is necessary to clearly define the hierarchical structure, neuron configuration, activation function selection, and training parameter settings of the LSTM model to ensure the model's prediction performance and stability under edge computing resource constraints.

[0040] The specific implementation of this embodiment is as follows: The LSTM neural network model includes three LSTM hidden layers, each with 64-128 neurons, the activation function is tanh, and the output layer is a fully connected layer; the LSTM neural network model outputs the predicted steam flow rate and the corresponding valve pre-opening percentage for 11 time points in 1-minute increments within the next 5-15 minutes; each LSTM hidden layer is equipped with a gating mechanism, including an input gate, a forget gate, and an output gate; the weight parameters of the LSTM neural network model are updated during the training phase using the backpropagation algorithm, the optimizer is Adam, the initial learning rate is 0.001, decays by 10% every 50 training cycles, and the batch size is 32.

[0041] In this embodiment, the LSTM neural network model adopts a stacked architecture of three LSTM hidden layers. The number of neurons in each LSTM hidden layer is between 64 and 128, and the activation function is the hyperbolic tangent function (tanh). The three-layer stacked design enables the model to extract the dependencies of different time scales in the multi-dimensional time-series features of the input, such as strip thickness, width, feed speed, iron oxide scale visual grade, acid temperature, steam valve opening, and strip temperature: the first LSTM layer captures short-term local changes, such as fluctuations in single feed speed or jumps in iron oxide scale grade; the second LSTM layer integrates medium-time-scale process evolution, such as the gradual temperature rise of acid as the strip moves continuously; the third LSTM layer models long-term global trends, such as the overall shift in steam heat demand when switching between different strip materials. The tanh activation function has an output range of (-1, 1), which can effectively avoid the gradient vanishing problem during gradient backpropagation while maintaining a strong expressive ability for nonlinear transformations of the input.

[0042] Each LSTM hidden layer is equipped with a gating mechanism, including an input gate, a forget gate, and an output gate. The input gate controls the proportion of new information written into the cell state at the current moment, the forget gate determines which information from historical cell states needs to be discarded, and the output gate determines the output of the hidden state based on the cell state. These three mechanisms work together to enable the model to dynamically and selectively retain or discard historical state information, accurately capturing the nonlinear evolution of steam heat demand during pickling and preventing long-term dependencies from being overwhelmed by noise. The model's weight parameters are updated during the training phase using a backpropagation algorithm, and the loss function uses mean squared error (MSE) to accurately measure the deviation between the predicted steam flow rate and the actual value. The optimizer chosen is Adam, which combines the advantages of momentum optimization and adaptive learning rate. The initial learning rate is set to 0.001, and a learning rate decay strategy of 10% every 50 epochs is adopted to ensure rapid convergence in the early stages of training and fine-tuning in the later stages. The batch size is set to 32 to achieve a balance between statistical accuracy and computational efficiency in gradient updates.

[0043] The output layer is a fully connected layer that receives the hidden state vector from the third LSTM hidden layer at the last time step. Through a linear transformation, it maps this hidden state vector into 11 output nodes, each corresponding to a predicted steam flow rate (in kg / min) at 11 discrete time points within the next 5–15 minutes, with a step size of 1 minute. Simultaneously, it outputs the valve pre-opening percentage (0%–100%) for each time point. This output structure transforms the originally ambiguous time-series heat demand prediction into a structurally stable, high-resolution, and directly executable control signal sequence.

[0044] The LSTM neural network model is trained offline based on historical production data from the company over the past 6–12 months. This historical production data includes multi-dimensional collaborative features such as strip thickness, width, speed, visual grade of iron oxide scale, acid temperature, steam valve opening, strip temperature, steam flow rate, and surface quality score after pickling. The data is normalized before being input into the model, with a normalization range of [0,1]. This effectively eliminates the interference of different physical dimensions and orders of magnitude on the LSTM neural network training process, ensuring stable convergence of the model during offline training and avoiding gradient explosion or vanishing. The training and test sets are divided in an 8:2 ratio to ensure the representativeness of the data distribution and the reliability of the generalization ability assessment.

[0045] The technical solution works as follows: In the offline training phase, the LSTM neural network model is fully trained using historical production data from the enterprise. Through the combination of the Adam optimizer and the MSE loss function, along with a learning rate decay strategy, the model learns the temporal mapping relationship between multi-dimensional process parameters and steam heat demand. The model's generalization ability is then verified using a test set. In the online inference phase, the trained model is deployed to the NPU acceleration unit of the edge computing unit and runs in INT8 fixed-point mode. It receives real-time multimodal input feature vectors and processes them layer by layer through the three-layer LSTM hidden layers. Finally, the fully connected output layer generates predicted steam flow rates and valve pre-opening percentages for 11 time steps within the next 5–15 minutes, providing accurate temporal decision-making basis for the edge computing unit's proactive control.

[0046] The above technical solution, employing a three-layer LSTM hidden layer combined with a tanh activation function and a fully connected output layer, adapts to the high-precision time-series prediction requirements under edge computing resource constraints. This enables the model to possess both long-term trend capture capabilities and minute-level dynamic response accuracy. The combination of gating mechanisms and backpropagation training ensures stable convergence capabilities as the model continuously receives on-site feedback data for incremental learning. Standardized training procedures (MSE loss, Adam optimizer, learning rate decay) and normalization ensure the stability and convergence speed of the model training process. This solution provides the steam control system with highly reliable prediction data, effectively resolving prediction uncertainties and control command ambiguities caused by unclear model structures, ultimately achieving intelligent and refined control of steam heat energy supply in the pickling process.

[0047] Example 3: Based on the system described in Example 1, further consideration needs to be given to the inference scheduling strategy of the edge computing unit and the prediction robustness under extreme conditions. In actual pickling production, the feed speed and thickness of the strip are not constant. When there are sudden changes in speed or thickness, the heat load of the strip changes drastically in a short period of time. If the LSTM model still infers at a fixed frequency and assigns the same attention weight to all input features, the predicted output will lag behind the actual heat demand changes, and the steam valve will not be able to respond at the appropriate time. At the same time, there may be instantaneous fluctuations between the discrete time steps of the predicted output. If this is directly used to drive the valve, it will cause the valve to start and stop frequently or the opening degree to oscillate, increasing mechanical wear. Therefore, it is necessary to clarify the inference frequency, action triggering timing, extreme condition adaptive mechanism, and output signal smoothing processing strategy to ensure the consistency and smoothness of predictive control.

[0048] The specific implementation of this embodiment is as follows: The edge computing unit repeatedly calls the LSTM neural network model for inference at a frequency of 1Hz, and drives the steam supply regulating device to perform pre-action 2-3 minutes in advance based on the valve pre-opening signal corresponding to the 3rd minute in the predicted output sequence; After each inference, the edge computing unit records the current timestamp, input feature vector, predicted output sequence and actual valve response delay as training samples and stores them in a local circular buffer; When a sudden change in pickling speed or a sudden change in incoming material thickness is detected, the edge computing unit automatically increases the inference frequency of the LSTM neural network model to 2Hz and enables a dynamic attention mechanism: By calculating the Euclidean distance between the current input feature vector and the historical mean feature, when the Euclidean distance exceeds a preset threshold, the attention score of the feed rate feature and the strip thickness feature is multiplied by a gain coefficient of 1.5.

[0049] In this embodiment, the edge computing unit uses a 1Hz frequency to continuously call the LSTM neural network model for inference, meaning it performs a complete forward inference calculation once per second. Each inference uses the latest acquired multimodal input feature vector as input and outputs a steam flow prediction sequence for 11 time steps within the next 5–15 minutes. This 1Hz inference frequency matches the rate of change of the production line process parameters, ensuring both the timeliness of the prediction and avoiding unnecessary computational resource consumption caused by excessively high inference frequencies. Based on this, the system specifically selects the valve pre-opening signal corresponding to the 3rd minute in the predicted output sequence as the control trigger, and accordingly drives the electric steam proportional regulating valve to operate 2–3 minutes in advance. The reason for choosing the 3rd minute as the action trigger point is that in the pickling process, there is a thermal inertia delay of about 2–3 minutes between the steam adjusting from the valve and the actual impact on the acid temperature. The predicted value at the 3rd minute corresponds precisely to the optimal time matching window between the valve action and the actual arrival of the heat load, ensuring that the valve opening adjustment behavior and the actual rise window of the heat load in the pickling process are strictly aligned on the time axis, realizing the core logic of "pre-opening valve" control.

[0050] After each inference iteration, the edge computing unit records the current timestamp, the input feature vector (including strip thickness, strip width, feed speed, iron oxide scale visual level, current acid temperature, current steam valve opening, and current strip temperature), the predicted output sequence (predicted steam flow rate and valve pre-opening percentage for the next 11 time steps), and the actual valve response delay as a training sample, storing them in a local ring buffer. The ring buffer uses a fixed-capacity circular storage structure; when the buffer is full, new sample data automatically overwrites the oldest data, ensuring that the training samples always reflect the most recent process operation status. By continuously accumulating the deviation information between actual valve actions and predictive control during strip steel production, the model can perform incremental self-optimization based on local closed-loop feedback data.

[0051] When a sudden change in pickling speed or a drastic change in incoming material thickness is detected, the edge computing unit determines whether the operating condition has entered an abnormal state by calculating the Euclidean distance between the current input feature vector and the historical mean feature in real time. When the Euclidean distance exceeds a preset threshold, the system automatically increases the inference frequency of the LSTM neural network model from 1Hz to 2Hz and activates the attention weight adjustment module. This module applies a 1.5x gain coefficient to the attention scores of the two key abrupt change features, namely the feed rate feature and the strip thickness feature, thereby significantly enhancing the model's sensitivity and response to dynamic heat load changes, ensuring that the predicted steam flow rate and valve pre-opening signal remain highly accurate under extreme operating conditions within the next 5–15 minutes. The historical mean feature is obtained by calculating the moving average of the input feature vectors over the past N time steps. The preset threshold is set according to the statistical distribution of historical data, typically taking the 95th percentile of the historical Euclidean distance.

[0052] Furthermore, in this embodiment, to avoid frequent valve start-stops or opening oscillations caused by instantaneous fluctuations between discrete time steps in the predicted output, the system performs cubic spline interpolation smoothing on the output steam flow prediction curve. This processing is executed in the signal smoothing module of the edge computing unit, with interpolation points at minutes 1, 3, 6, 9, and 11 of the prediction sequence, selecting five key time points as interpolation references. Through the cubic spline algorithm, a continuous and smooth transition curve is generated between adjacent interpolation points, ultimately generating 20 equally spaced intermediate points. The interpolated valve control signal is output to the control interface of the electric steam proportional regulating valve at a high-frequency update frequency of 5Hz, enabling the valve to gradually adjust its opening along a smooth trajectory during execution. This effectively suppresses mechanical shocks and high-frequency jitter caused by sudden changes in predicted values, significantly improving the stability and lifespan of valve action, while ensuring accurate matching of steam supply and heat demand predictions in the time dimension.

[0053] The technical solution operates as follows: the edge computing unit continuously receives multimodal real-time data at a frequency of 1Hz and calls the LSTM model for rolling inference to generate a future steam demand prediction sequence; after each inference, the runtime data is stored in a circular buffer to accumulate training data for subsequent model optimization; based on the valve pre-opening signal corresponding to the 3rd minute of the prediction sequence, the system drives the valve to pre-act 2-3 minutes in advance. When a sudden change occurs in the operating condition, the system automatically increases the inference frequency and activates the dynamic attention mechanism to enhance the weighted response to the sudden change features. The predicted steam flow curve is smoothed by cubic spline interpolation and output to the valve control interface at a frequency of 5Hz, driving the valve to perform opening adjustment along the smooth trajectory.

[0054] By employing the above technical solution, and combining 1Hz rolling inference with a 3-minute pre-opening valve triggering strategy, the response fluctuations and control instability caused by fuzzy prediction timing and unclear action trigger points in traditional control are eliminated, achieving high precision and high consistency in "pre-opening valve" control. The accumulation of training data in the annular buffer enables the model to possess local closed-loop self-optimization capabilities. The combination of dynamic attention mechanism and increased inference frequency allows the model to maintain high-precision prediction and rapid response even under extreme operating conditions. Cubic spline interpolation smoothing ensures smooth valve action, reduces mechanical shock, and extends equipment life. This solution upgrades steam regulation from "passive response" to "active prediction + smooth execution," significantly improving the system's robustness and control reliability under complex operating conditions.

[0055] Example 4: Based on the system described in Example 1, the continuous adaptability of the LSTM neural network model needs further consideration. In actual pickling production, due to factors such as seasonal changes in ambient temperature, equipment aging, pipeline thermal inertia drift, and fluctuations in the quality of incoming strip steel, the offline-trained LSTM model may experience a "model drift" phenomenon, where the prediction accuracy gradually decreases after long-term operation. If the model parameters remain fixed and cannot be adjusted according to the latest production data, the prediction deviation of steam flow will gradually accumulate, ultimately affecting the control accuracy and pickling quality. Therefore, an online incremental learning mechanism that can dynamically update the model parameters during production is needed, enabling the model to continuously adapt to changes in the current production conditions without interrupting production line operation or relying on an external training platform.

[0056] The specific implementation method of this embodiment is as follows: After each batch of strip steel pickling is completed, the edge computing unit uses the actual steam consumption and pickling quality assessment data of the batch as a supervision signal to trigger the online incremental learning of the LSTM neural network model. The online incremental learning adopts the mini-batch gradient descent method. Each update uses 5 sets of input and output samples of the current batch. The learning rate is set to 1 / 10 of the initial learning rate of offline training. The weight update is achieved by freezing the first two LSTM hidden layers and only fine-tuning the third LSTM hidden layer and output layer.

[0057] In this embodiment, after each batch of strip steel pickling is completed (typically a batch involves 30–60 minutes of continuous production), the edge computing unit actively collects the actual steam consumption and pickling quality assessment data for that batch. This data is then used as a monitoring signal under real-world operating conditions, triggering the locally deployed LSTM neural network model to perform online incremental learning. The actual steam consumption is obtained by cumulative statistics from the steam flow meter within the batch cycle, and the pickling quality assessment data is provided by the pickling quality assessment module.

[0058] In this embodiment, the pickling quality assessment module uses deep learning algorithms to automatically identify key quality indicators such as the residual oil area ratio, iron oxide scale removal rate, and pickling spot number on the strip surface, based on image data collected by a multispectral vision sensor. This module employs a YOLOv8 instance segmentation network. Its input is an RGB-NIR fused image (i.e., a registered and fused image of the visible red, green, and blue channels and the near-infrared channel) collected by the multispectral vision sensor. The output is a pixel-level mask image of the oil area, residual iron oxide scale area, and pickling spots, which is then used to calculate three quantitative quality indicators: the residual oil area ratio, the iron oxide scale removal rate, and the number of pickling spots. This assessment process does not rely on manual judgment or external detection equipment, achieving real-time and objective quantitative evaluation of the pickling effect. In practical applications, this instance segmentation network can also use other semantic segmentation frameworks (such as Mask R-CNN, SegFormer, etc.), which are not limited in this embodiment.

[0059] The online incremental learning process is as follows: After each batch of pickling is completed, five input-output sample pairs are randomly sampled from the current batch. Each sample pair includes: an input feature vector (strip thickness, strip width, feed rate, iron oxide scale visual grade, acid temperature, steam valve opening, and strip temperature) and a target output label (actual steam flow sequence and corresponding pickling quality score). Mini-batch gradient descent is used to calculate the gradient of the loss function in mini-batch (batch size=5) mode. The learning rate is set to 1 / 10 of the initial offline training learning rate of 0.001, i.e., 0.0001, to ensure that the magnitude of weight updates is sufficiently small, avoiding significant disturbances to the parameters of the converged model. A layered freezing strategy is used for weight updates: all weight parameters of the first and second LSTM hidden layers are frozen, and only the weights and biases of the third LSTM hidden layer and the fully connected output layer are fine-tuned. The principle behind this strategy is as follows: the first two LSTM hidden layers are mainly responsible for extracting general temporal features (such as trend recognition and periodic capture). These features have good transferability between different production batches, so freezing their weights can preserve the general knowledge that the model has learned. The third LSTM hidden layer and output layer are responsible for mapping abstract features to specific predicted values, which are most closely related to the current specific working conditions. Fine-tuning them can effectively adapt to local changes in production conditions.

[0060] The working process of this technical solution is as follows: After each batch of pickling is completed, the system automatically collects the actual steam consumption data and pickling quality assessment data of that batch as monitoring signals; using the mini-batch gradient descent method and layered freezing strategy, only the weights of the third hidden layer and the output layer of the LSTM model are slightly adjusted; after the update, the model is immediately used to predict the steam demand for the next batch of pickling. This process is automatically executed in a closed loop during production, without manual intervention or production line shutdown, achieving continuous and incremental optimization of model parameters.

[0061] By employing the above technical solution, through a batch-level online incremental learning mechanism and a layered freeze-tuning strategy, the LSTM neural network model can continuously absorb the nonlinear patterns and heat demand coupling characteristics from the latest production data without interrupting production line operation, dynamically correcting the prediction logic for future steam flow and valve action requirements. The layered freeze-tuning strategy ensures that the model maintains high stability and real-time performance even in low-computing-power edge environments, avoiding prediction bias, overfitting, or computational resource overload problems caused by full parameter updates. The introduction of the pickling quality assessment module ensures that the incremental learning supervision signal includes not only steam consumption but also pickling quality feedback, forming a complete closed loop of "prediction-execution-evaluation-optimization." This enables steam regulation to shift from "passive response" to "active prediction + self-evolution" closed-loop optimization, significantly improving steam utilization efficiency and pickling quality consistency.

[0062] Example 5: Based on the system described in Example 1, it is further necessary to consider the impact of different strip steel materials on the prediction of steam heat demand. In actual continuous pickling production lines for cold-rolled strip steel, the processed strip steel materials typically include cold-rolled low-carbon steel, high-strength steel, and stainless steel. Different materials have significantly different thermal conductivity, specific heat capacity, and iron oxide scale removal heat requirements: for example, the thermal conductivity of stainless steel is about 1 / 3 that of low-carbon steel, and under the same thickness and feed rate conditions, the heat input required for stainless steel strip to reach the target pickling temperature is significantly different; high-strength steel has a dense iron oxide scale structure on its surface, requiring higher acid temperature and reaction time for removal. If a single LSTM model is used to uniformly predict all materials, the model needs to simultaneously fit the different distribution patterns of heat demand for multiple materials, which not only increases the learning difficulty but may also introduce prediction bias when switching materials. Therefore, an adaptive control mechanism that can automatically switch the prediction model according to the strip steel material is needed to improve the prediction accuracy under different material conditions.

[0063] The specific implementation method of this embodiment is as follows: the edge computing unit runs multiple LSTM sub-models simultaneously. The multiple LSTM sub-models are trained independently on three datasets: cold-rolled low-carbon steel, high-strength steel and stainless steel. Each sub-model has the same network structure, only the output layer weight parameters are different. The edge computing unit automatically switches to the LSTM sub-model of the corresponding material category through the built-in rule engine according to the strip material identification result at the entrance, so as to realize material-adaptive steam demand prediction.

[0064] In this embodiment, the strip steel material is classified using a spectral reflectance identification module at the entrance. This module is installed at the entrance of the pickling production line and collects reflectance spectral data of the strip steel surface at different wavelengths based on the reflectance spectral characteristics in the visible and near-infrared bands. Since cold-rolled low-carbon steel, high-strength steel, and stainless steel have different reflectance spectral characteristic curves in the visible-near-infrared band (reflected in differences in absolute reflectance, spectral slope, and characteristic absorption peaks), the module's built-in Support Vector Machine (SVM) classifier learns the spectral feature distribution boundaries of various materials during the training phase, enabling online and real-time classification of the current strip steel material and outputting one of three category labels: cold-rolled low-carbon steel, high-strength steel, or stainless steel. In practical applications, this material identification module can also select other classifiers (such as random forests, K-nearest neighbors, etc.), which is not limited in this embodiment.

[0065] The edge computing unit simultaneously runs three LSTM sub-models, corresponding to cold-rolled low-carbon steel, high-strength steel, and stainless steel, respectively. The three sub-models have identical network structures: all have three LSTM hidden layers with 64–128 neurons per layer, using the tanh activation function, and a fully connected output layer. This identical model structure ensures that each sub-model shares the same inference graph within the edge computing unit, facilitating batch scheduling and memory management by the NPU acceleration unit. The only difference between the three sub-models is the weight parameters of the output layer, achieved through independent offline training on their respective material datasets: the cold-rolled low-carbon steel sub-model is trained using historical data containing only cold-rolled low-carbon steel production batches, the high-strength steel sub-model uses historical data containing only high-strength steel production batches, and the stainless steel sub-model uses historical data containing only stainless steel production batches. The training process for each sub-model is consistent with the training strategy described in claim 2.

[0066] The switching logic is implemented by the rule engine built into the edge computing unit. The rule engine's workflow is as follows: it receives the material category label output by the spectral reflectance recognition module at the inlet in real time; it determines the corresponding target sub-model ID by looking up the table based on the recognition result; it routes the current multimodal input feature vector to the target sub-model for inference; the target sub-model outputs the predicted steam flow rate and valve pre-opening signal for the next 5–15 minutes under the corresponding material category. When the incoming material material of the production line has not changed, the rule engine continues to use the current sub-model for inference; when a change in the incoming material category is detected, the rule engine automatically switches to the corresponding sub-model in the next inference cycle (within 1 second), and the switching process does not require machine downtime or manual intervention.

[0067] The working process of this technical solution is as follows: When a new strip steel coil enters the pickling production line, the spectral reflectance recognition module at the entrance first collects the reflectance spectrum data of the strip steel surface and outputs the material category in real time through the SVM classifier; the rule engine of the edge computing unit automatically switches to the corresponding LSTM sub-model according to the material classification result; subsequently, the multimodal perception data is input into the selected sub-model to generate a steam demand prediction sequence under the material working condition; the edge computing unit drives the valve to pre-act and store energy and heat before the predetermined time according to the prediction result, so as to realize differentiated and precise steam control for the pickling process of strip steel of different materials.

[0068] By employing the above technical solution, and combining a spectral reflectance recognition module with a multi-LSTM sub-model switching mechanism, adaptive regulation of steam heat demand prediction under different strip steel material conditions is achieved. Each material sub-model is trained independently on its own proprietary dataset, avoiding the accuracy trade-off caused by a single model simultaneously fitting multiple material distributions. The automatic switching mechanism of the rule engine ensures seamless transitions when materials change, requiring no manual intervention. This solution significantly improves the response accuracy and energy efficiency matching of steam supply during pickling of different strip steel materials such as cold-rolled low-carbon steel, high-strength steel, and stainless steel, effectively solving the problems of inaccurate steam prediction and control lag caused by material differences, and realizing an intelligent upgrade of steam control based on material adaptation.

[0069] Example 6: Based on the system described in Example 1, further consideration needs to be given to the fault tolerance and safety degradation issues of the system under abnormal operating conditions. The industrial environment of the pickling production line is complex, with various factors that may lead to system component failure, such as electromagnetic interference, sensor aging, loose communication lines, or overheating and frequency reduction of the edge computing unit. When the sensor data of the multi-source sensing module is abnormal (e.g., data loss, out-of-bounds values, or communication interruption), or when the edge computing unit experiences accumulated LSTM model inference timeouts due to fluctuations in computational load, if the system still relies on unreliable prediction results for steam regulation, it may lead to valve malfunction, uncontrolled acid temperature, or safety hazards. Therefore, a fault diagnosis and fault tolerance mechanism is needed that can monitor the system's health status in real time and automatically degrade to a safe control mode when an anomaly occurs, to ensure the continuous and safe operation of the pickling process under extreme conditions.

[0070] The specific implementation of this embodiment is as follows: The system also includes a fault diagnosis module; the fault diagnosis module judges the abnormal state by verifying the reasonable range of data of each sensor in the multi-source sensing module, the data update frequency and the communication heartbeat packet status. When the edge computing unit's inference delay exceeds 200ms for three consecutive times or any sensor data is zero for 5 consecutive frames, a fault flag is triggered, the steam control mode is automatically switched from LSTM predictive control to traditional PID feedback control mode, and an alarm is triggered.

[0071] In this embodiment, the fault diagnosis module is integrated into the edge computing unit and runs continuously as an independent background monitoring process. This module verifies the status of each sensor's data in real time through the following three dimensions:

[0072] First, data reasonableness range verification. A reasonable physical value range is preset for each sensor. For example: the iron oxide scale vision level of the multispectral vision sensor should be within the range of 1–5; the strip thickness of the incoming material thickness gauge should be within the range of 0.3mm–6.0mm (corresponding to common specifications of cold-rolled strip steel); the strip temperature of the infrared thermal imager should be within the range of 0–300℃; and the acid temperature of the distributed temperature sensor array should be within the range of 40–80℃. When any sensor data exceeds its preset reasonable range, the fault diagnosis module records the abnormal event and increments the corresponding sensor's abnormality counter.

[0073] Second, data update frequency verification. Each sensor has a specified update frequency for data acquisition and uploading: 10 frames / second for the multispectral vision sensor, 10Hz for the incoming material thickness gauge, 30Hz for the infrared thermal imager, and 1Hz for the distributed temperature sensor array. The fault diagnosis module counts the actual number of data frames arriving at each sensor within each monitoring cycle (1 second). If the actual update frequency is lower than 50% of the rated frequency (i.e., lower than 5 frames / second for the multispectral vision sensor, lower than 5Hz for the incoming material thickness gauge, etc.) and persists for more than 3 seconds, it is determined that the sensor has a communication delay or data loss anomaly.

[0074] Third, communication heartbeat status verification. Each sensor communicates with the edge computing unit via an industrial fieldbus (such as RS485, Modbus TCP, or PROFINET). The fault diagnosis module sends heartbeat request packets to each sensor at fixed intervals (e.g., 500ms) and waits for a response from the device. If no response is received from the sensor for three consecutive heartbeat cycles, the communication link for that sensor is considered interrupted.

[0075] The fault diagnosis module triggers a fault flag when any of the following fault conditions are met: Condition 1: The edge computing unit experiences three consecutive LSTM model inference delays exceeding 200ms—this indicates that the NPU acceleration unit or ARM processor may have a hardware fault, overheating and frequency reduction, or resource deadlock, making it unable to complete the prediction calculation within the specified time; Condition 2: Any sensor data is zero for five consecutive frames—this indicates that the sensor may have completely failed, lost power, or experienced a complete communication interruption, and cannot provide effective sensing data. After the fault flag is triggered, the fault diagnosis module immediately performs two actions: First, it automatically switches the steam control mode from LSTM predictive control to traditional PID feedback control mode, keeping the PID controller's temperature setpoint consistent with the current acid target temperature, and using the sensor data from the still-operating sensors in the distributed temperature sensor array as the PID control input; Second, it triggers the local audible and visual alarm device and sends an alarm message to the host computer system, prompting maintenance personnel to intervene and investigate.

[0076] The technical solution operates as follows: The fault diagnosis module, as an independent background process, continuously monitors the data rationality, update frequency, and heartbeat status of each sensor in the multi-source sensing module, while simultaneously monitoring the inference latency of the LSTM model in the edge computing unit. When abnormal sensor data is detected or the inference timeout reaches a preset threshold, a fault flag is triggered, and the control mode is automatically switched: steam regulation switches from LSTM predictive control to PID feedback control, and an alarm is activated. After the fault is resolved (e.g., after replacing the faulty sensor or restarting the computing unit), maintenance personnel can manually reset the system back to LSTM predictive control mode.

[0077] By employing the above technical solution, a multi-dimensional fault detection mechanism and automatic fallback switching strategy are used to build the system's safety and fault tolerance capabilities under abnormal operating conditions. The fault diagnosis module simultaneously covers anomaly detection at both the perception layer (sensor data rationality, update frequency, and communication heartbeat) and the computing layer (inference delay monitoring), ensuring that all types of faults can be identified in a timely manner. Automatic switching to PID feedback control mode ensures that even in extreme cases where the AI ​​prediction function fails, the pickling process can still maintain the acid temperature within a safe threshold range using a stable and reliable classical control strategy, avoiding uneven strip pickling, soaring energy consumption, or equipment damage caused by delayed or excessive steam supply. This solution achieves a seamless transition from intelligent predictive control to safe fallback control without relying on manual intervention, significantly improving the system's robustness and continuous operation capability in complex industrial environments.

[0078] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart control system for pickling steam based on edge computing and LSTM neural network, comprising: Multi-source sensing module, edge computing unit, steam supply regulating device and thermal energy buffer device; The multi-source sensing module is used to collect visual data on the iron oxide scale level of the strip steel, the strip steel thickness, the strip steel temperature, and the acid solution temperature; the steam supply regulating device is used to regulate the steam supply in response to the valve pre-opening signal; and the thermal energy buffering device is used to buffer instantaneous heat load fluctuations. Its features are as follows: The edge computing unit is deployed on the pickling production line and runs an LSTM neural network model on it. The LSTM neural network model is configured to take a multimodal feature vector, including the visual grade of the iron oxide scale, the strip thickness, the strip temperature, the acid temperature, the strip width, the feed speed and the current steam valve opening, as input, and directly output the predicted steam flow rate and the corresponding valve pre-opening signal for multiple time steps within a preset time period. The edge computing unit is further configured to: drive the steam supply regulating device to perform a pre-action according to the valve pre-opening signal, and simultaneously instruct the thermal energy buffer device to start heat storage, before the time step corresponding to the predicted steam flow value by a preset advance period.

2. The intelligent pickling steam control system according to claim 1, characterized in that, The LSTM neural network model includes three LSTM hidden layers, each with 64-128 neurons, an activation function of tanh, and a fully connected output layer. The LSTM neural network model outputs the predicted steam flow rate and the corresponding valve pre-opening percentage for 11 time points within the next 5-15 minutes in 1-minute increments. Each LSTM hidden layer is equipped with a gating mechanism, including an input gate, a forget gate, and an output gate. The weight parameters of the LSTM neural network model are updated during the training phase through the backpropagation algorithm. The optimizer is Adam, the initial learning rate is 0.001, the learning rate decays by 10% every 50 training epochs, and the batch size is 32.

3. The intelligent control system for pickling steam according to claim 1, characterized in that, The edge computing unit repeatedly calls the LSTM neural network model at a frequency of 1Hz for inference, and drives the steam supply regulating device to perform pre-action 2-3 minutes in advance based on the valve pre-opening signal corresponding to the 3rd minute in the predicted output sequence. After each inference, the edge computing unit records the current timestamp, input feature vector, predicted output sequence and actual valve response delay as training samples and stores them in a local circular buffer. When a sudden change in pickling speed or a drastic change in incoming material thickness is detected, the edge computing unit automatically increases the inference frequency of the LSTM neural network model to 2Hz and enables a dynamic attention mechanism: by calculating the Euclidean distance between the current input feature vector and the historical mean feature, when the Euclidean distance exceeds a preset threshold, the attention score of the feed speed feature and the strip thickness feature is multiplied by a gain coefficient of 1.

5.

4. The intelligent control system for pickling steam according to claim 1, characterized in that, After each batch of strip steel pickling is completed, the edge computing unit uses the actual steam consumption and pickling quality assessment data of that batch as a supervision signal to trigger the online incremental learning of the LSTM neural network model. The online incremental learning adopts the mini-batch gradient descent method. Each update uses 5 sets of input and output samples of the current batch. The learning rate is set to 1 / 10 of the initial learning rate of offline training. The weight update is achieved by freezing the first two LSTM hidden layers and only fine-tuning the third LSTM hidden layer and output layer.

5. The intelligent control system for pickling steam according to claim 1, characterized in that, The edge computing unit runs multiple LSTM sub-models simultaneously. These sub-models are trained independently on three datasets: cold-rolled low-carbon steel, high-strength steel, and stainless steel. Each sub-model has the same network structure, with only the output layer weight parameters differing. Based on the strip material identification results at the entry point, the edge computing unit automatically switches to the corresponding LSTM sub-model using its built-in rule engine, achieving material-adaptive steam demand prediction.

6. The intelligent control system for pickling steam according to claim 1, characterized in that, The system also includes a fault diagnosis module; the fault diagnosis module determines abnormal states by verifying the reasonable range of data from each sensor in the multi-source sensing module, the data update frequency, and the communication heartbeat packet status. When the edge computing unit's inference delay exceeds 200ms for three consecutive times or any sensor data is zero for five consecutive frames, a fault flag is triggered, automatically switching the steam control mode from LSTM predictive control to traditional PID feedback control mode and triggering an alarm.

Images