TinyML-based smelting control method, device, equipment and medium

Abstract

The present invention relates to the field of industrial Internet of Things security technology, and discloses a smelting control method, device, equipment and medium based on TinyML. The method includes obtaining historical operating data of smelting equipment, generating abnormal behavior data of smelting equipment, constructing a TinyML model, including a parameter adjustment sub-model and an anomaly detection sub-model; training the TinyML model based on the historical operating data and abnormal behavior data to output the safe adjustment range of process parameters, output abnormal probability and abnormal type; deploying the trained TinyML model at the edge of the smelting equipment, obtaining multi-source sensor data of the smelting equipment in real time, and performing different processing based on the output results. The method provided by the present invention can realize real-time, high-precision edge-end reasoning of the abnormal probability of smelting equipment, and realize millisecond-level access decision-making of smelting equipment behavior, thereby reducing the risk of production line shutdown caused by network attacks.

Description

Technical Field

[0001] The present invention relates to the field of industrial Internet of Things security technology, and in particular to a TinyML-based smelting control method, device, equipment, and medium. Background Art

[0002] In the industrial smelting process, real-time status monitoring and abnormal warning of smelting equipment are key technical means to ensure production safety and improve energy efficiency. With the development of industrial Internet of Things and artificial intelligence technologies, technologies based on machine learning for unmanned management of smelting equipment have emerged one after another. However, the current use of traditional machine learning models for intelligent monitoring and control decision-making not only carries the risk of network attacks causing production line shutdowns, but also data uploaded to the cloud may expose industrial sensitive information. The smelting environment also has corresponding problems such as limited network bandwidth and high communication latency. In addition, centralized cloud-based decision-making is difficult to meet the real-time response requirements of milliseconds and cannot meet the needs of refined management and control of industrial equipment. Summary of the Invention

[0003] The purpose of the present invention is to provide a smelting control method, device, equipment and medium based on TinyML to address the problems of high communication delay, weak real-time response capability and potential safety hazards in the monitoring and control of smelting equipment in related technologies. The method can realize real-time and high-precision edge-end reasoning of the abnormal probability of smelting equipment, and realize millisecond-level access decision-making of smelting equipment behavior, thereby reducing the risk of production line shutdown caused by network attacks.

[0004] In order to achieve the above-mentioned purpose, in a first aspect of the present invention, a smelting control method based on TinyML is provided, comprising: S1, obtaining historical operating data of smelting equipment under normal working conditions and abnormal working conditions, and generating abnormal behavior data of smelting equipment; S2, constructing a TinyML model, comprising a parameter adjustment sub-model and an anomaly detection sub-model, the parameter adjustment sub-model being formed by serially combining a convolutional neural network and a long short-term memory network, and the anomaly detection sub-model being formed by a lightweight classifier based on Transformer-Attention; S3, training the parameter adjustment sub-model based on the historical operating data of normal working conditions to output a safe adjustment range of process parameters, and The abnormal behavior data is used to train the anomaly detection sub-model to output the anomaly probability and anomaly type; S4, the trained TinyML model is deployed on the edge of the smelting equipment, and the multi-source sensor data of the smelting equipment is obtained in real time using a sliding window. The trained TinyML model is input, and the access rights of the smelting equipment API interface are adjusted based on the safe adjustment range of the output process parameters. Based on the output abnormal probability and abnormal type, a graded response is performed for processing; S5, the abnormal probability output by the anomaly detection sub-model in real time is obtained. If the abnormal probability output within the preset time does not exceed the first threshold, the amount of data obtained by the sliding window is reduced. If the abnormal probability output exceeds the second threshold, the amount of data obtained by the sliding window is increased.

[0005] In the second aspect of the present invention, a smelting control device based on TinyML is provided, comprising: an acquisition module for acquiring historical operating data of smelting equipment under normal and abnormal conditions, and generating abnormal behavior data of the smelting equipment; a modeling module for constructing a TinyML model, comprising a parameter adjustment sub-model and an anomaly detection sub-model, wherein the parameter adjustment sub-model is formed by a series combination of a convolutional neural network and a long short-term memory network, and the anomaly detection sub-model is formed by a lightweight classifier based on Transformer-Attention; a training module for training the parameter adjustment sub-model based on the historical operating data of normal conditions to output a safe adjustment range of the process parameters, and a training module for training the parameter adjustment sub-model based on the historical operating data and the abnormal behavior data. An anomaly detection sub-model is trained for the data to output anomaly probability and anomaly type; a deployment module is used to deploy the trained TinyML model on the edge of the smelting equipment, use a sliding window to obtain multi-source sensor data of the smelting equipment in real time, input the trained TinyML model, adjust the access rights of the smelting equipment API interface based on the safe adjustment range of the output process parameters, and process the data in a graded manner based on the output anomaly probability and anomaly type; a feedback module is used to obtain the anomaly probability output by the anomaly detection sub-model in real time. If the output anomaly probability does not exceed the first threshold within the preset time, the amount of data obtained by the sliding window is reduced. If the output anomaly probability exceeds the second threshold, the amount of data obtained by the sliding window is increased.

[0006] In a third aspect of the present invention, an electronic device is provided, comprising a memory and a processor coupled to the memory, wherein the processor is configured to execute the method of the first aspect based on instructions stored in the memory.

[0007] In a fourth aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the method of the first aspect is implemented.

[0008] The technical solution of the present invention, by constructing and training a lightweight TinyML model, allows relevant operators or systems to access and operate when it is detected that the equipment is operating normally and the data complies with the rules. If abnormal data or potential security threats are found, such as abnormal fluctuations in data or non-compliance with preset rules, access is immediately restricted or prohibited, and a security alarm is issued. Through the fast reasoning capability of the TinyML model, millisecond-level access decisions are achieved, effectively preventing illegal access caused by network attacks, ensuring the safe and stable operation of the smelting production line, and reducing the risk of production line shutdown due to network attacks. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present application, a brief introduction will be given below to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0010] Figure 1 It is a flow chart of the TinyML-based smelting control method described in the embodiment.

[0011] Figure 2 This is a flow chart of monitoring the oxygen gun using the TinyML model described in the embodiment.

[0012] Figure 3 It is a structural diagram of the TinyML-based smelting control device described in the embodiment.

[0013] Figure 4 Schematic diagram of the structure of the electronic device described in the embodiment. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0015] The present application is described below with reference to specific embodiments in conjunction with the accompanying drawings.

[0016] like Figure 1 As shown, a smelting control method based on TinyML in this embodiment includes: S1, obtaining historical operating data of smelting equipment under normal working conditions and abnormal working conditions, and generating abnormal behavior data of smelting equipment; S2, constructing a TinyML model, including a parameter adjustment sub-model and an anomaly detection sub-model, the parameter adjustment sub-model is formed by combining a convolutional neural network and a long short-term memory network in series, and the anomaly detection sub-model is formed by a lightweight classifier based on Transformer-Attention; S3, training the parameter adjustment sub-model based on the historical operating data of normal working conditions to output a safe adjustment range of process parameters, and training the parameter adjustment sub-model based on the historical operating data and abnormal behavior data. Train the anomaly detection sub-model to output the anomaly probability and anomaly type; S4, deploy the trained TinyML model on the edge of the smelting equipment, use the sliding window to obtain the multi-source sensor data of the smelting equipment in real time, input the trained TinyML model, adjust the access rights of the smelting equipment API interface based on the safe adjustment range of the output process parameters, and process the data in a graded manner based on the output anomaly probability and anomaly type; S5, obtain the anomaly probability output by the anomaly detection sub-model in real time. If the anomaly probability output within the preset time does not exceed the first threshold, reduce the amount of data obtained by the sliding window. If the anomaly probability output exceeds the second threshold, increase the amount of data obtained by the sliding window.

[0017] For historical operation data, the data format can be a time series, and the sampling frequency is set according to the equipment characteristics. The acquisition also includes the processing of historical operation data. For example, various pre-processing such as data cleaning and data standardization are required for the historical operation data. Then, based on the timing characteristics of the smelting equipment, the historical operation data can be used to extract key features such as statistical features, frequency domain features or time-frequency joint features, reduce the data dimension and retain abnormal discriminant information, and then screen out historical operation data with high-discrimination features for training models. The statistical features include mean, variance, maximum value, minimum value, peak-to-peak value, root mean square, etc. The frequency domain features can be calculated through fast Fourier transform to calculate the main frequency component, spectral entropy, harmonic distortion rate, etc.

[0018] When obtaining historical operating data of abnormal operating conditions, it is necessary to collect operating data of various abnormalities occurring in smelting equipment to improve the TinyML (Tiny Machine Learning) model's ability to identify abnormal phenomena. Since abnormal behavior data is relatively rare in historical data, in order to cover various attack scenarios during training, in some optional implementations, fault injection tools (such as Chaos Monkey for IoT) can be used to simulate attack scenarios such as PLC instruction tampering and sensor signal drift to generate a large amount of abnormal behavior data for model training.

[0019] Since the operating data of many smelting equipment has a typical hierarchical structure of temporal features, with both local feature changes and long-term trends, after obtaining historical operating data under normal operating conditions, a convolutional neural network (CNN) and a long short-term memory network (LSTM) can be combined in series to form a parameter adjustment sub-model. In some optional implementations, the specific architecture of the parameter adjustment sub-model may include a CNN layer, an LSTM layer, and a fusion layer, wherein the CNN layer extracts spatial features from the input data, the LSTM layer receives the feature sequence output by the CNN, and processes the temporal dependency through a gating mechanism. The fusion layer is used to splice the spatial features of the CNN with the temporal features of the LSTM. In this way, the characteristics of CNN being good at extracting spatial or local features and LSTM being good at processing long-term temporal dependencies are fully utilized. Based on the lightweight design requirements, the serial structure is adopted to avoid the parameter expansion caused by the dual-path feature fusion in the parallel structure. The model has fewer parameters and is more suitable for edge device operation. The serial combination can optimize memory access patterns, reduce computing latency, and meet the millisecond-level decision-making requirements of smelting scenarios.

[0020] Input historical operating data under normal working conditions for training, so that the parameter adjustment sub-model can output the parameter range of process parameters that can be safely adjusted. After the TinyML model is deployed on the smelting equipment, the access rights of the smelting equipment API interface can be automatically adjusted according to the output results of the parameter adjustment sub-model to form dynamic permission management. The API interface of the smelting equipment is used to exchange data and call functions with other systems or applications. By adjusting the access rights of the smelting equipment API interface, it can not only prevent improper operations of remote control by staff, but also resist network attacks and data tampering. For example, in a certain smelting stage, after multi-source sensor data is input to the TinyML model, based on By training on historical operating data under normal operating conditions, it is determined that under the current normal operating conditions, the safe adjustment range of the rolling force of the rolling mill is 200-400KN. The access rights of the rolling mill API interface are set accordingly. For example, except for the administrator, other personnel can only adjust the rolling force within the range of 200-400KN. At this time, ordinary operators are not allowed to increase their rolling force to 450KN. If the next smelting stage is entered after a period of time, the TinyML model outputs that the safe adjustment range of the rolling force of the rolling mill is 200-500KN. The access rights of the rolling mill API interface will be adjusted accordingly. At this time, the ordinary operator will be given a valid instruction to increase the rolling force to 450KN.

[0021] After acquiring historical operating data for normal and abnormal operating conditions and generating abnormal behavior data, an anomaly detection sub-model can be formed based on a lightweight classifier using Transformer-Attention. A lightweight classifier is a classification model that maintains high classification accuracy while having a small model size and low computational requirements. The anomaly detection sub-model is a lightweight classification model. Using a lightweight classifier based on Transformer-Attention, it combines the Transformer's ability to capture temporal features with the efficiency of a lightweight architecture to achieve accurate identification and real-time detection of equipment anomalies. It directly calculates the correlation weights of features at different time points in the smelting equipment's operating data (such as the temporal correlation of signals like temperature, pressure, and current), thereby capturing subtle trend changes preceding equipment anomalies. Through a self-attention mechanism, the anomaly detection sub-model automatically selects the most critical features for anomaly detection (such as motor load fluctuations during high-temperature periods), suppresses noise interference, and improves the recognition of abnormal features.

[0022] In some optional embodiments, specific methods of lightweight processing may include: adjusting the self-attention layer, that is, reducing multi-head attention according to actual needs, realizing structured pruning, and removing redundant weights in multi-head attention, such as the number of heads of multi-head self-attention ≤4, and the convolution kernel size ≤3×3, so as to reduce the number of parameters and retain the key feature extraction capability, and the input dimension can also be reduced by linear projection; using a feedforward network composed of depthwise separable convolution to replace the standard fully connected layer to reduce the amount of computation; the classification layer can output anomaly probability distribution based on a normalized exponential function. In addition, the teacher model of a complex Transformer can be used to guide a lightweight anomaly detection sub-model to improve model capabilities and realize knowledge distillation. The lightweight design can control the inference delay to the millisecond level, meet the real-time monitoring needs of smelting equipment, and avoid detection lags caused by time-consuming computations of traditional heavy models.

[0023] For the anomaly detection sub-model, historical operating data under normal and abnormal conditions, as well as the generated abnormal behavior data, can be input for training, so that the anomaly detection sub-model can output the anomaly probability and anomaly type. Through rich data training, the anomaly detection sub-model can promptly identify changes in operating data or abnormal instructions, and even predict the occurrence of anomalies in smelting equipment. It can also promptly identify malicious attacks or misoperations and take corresponding actions to ensure the production safety of smelting equipment.

[0024] In some optional implementations, during the training phase of the parameter adjustment submodel and the anomaly detection submodel, all calculations are performed using the 32-bit floating-point format by default. After the TinyML model is deployed for use, that is, during the inference phase, it can be converted to INT8 quantization to control the size of the TinyML model. Based on the data distribution characteristics of multiple plants and multiple production lines in smelting enterprises, a federated learning framework can be used for model training. Each plant or production line only uploads parameter update information for model training while retaining data locally, and aggregates and distributes parameters through a central server to achieve collaborative training of multi-source data. This training method can not only protect the privacy and security of enterprise data, but also make full use of dispersed data resources, accelerate the model training process, and improve the generalization and adaptability of the model.

[0025] In step S4, the edge end, also known as the edge device, refers to the computing node deployed at the smelting equipment site, which has the ability to collect data, perform real-time reasoning, and make local decisions; multi-source sensor data is information collected from multiple sensors on the smelting equipment, including various operating data of the smelting equipment, such as pressure, current, vibration, etc., and can also include required data information such as ambient temperature and humidity.

[0026] After the deployment of the TinyML model, by real-time analysis of multi-source sensor data from smelting equipment, such as operating data or control instructions, the TinyML model can identify potential anomalies in advance, such as motor bearing wear, sensor drift, misoperation, network attacks, and other abnormal probabilities, and then prioritize maintenance operations in access control to avoid downtime losses caused by sudden failures; the TinyML model can also dynamically allocate permissions based on real-time production status rather than static fixed permissions. For example, during normal production, only authorized engineers are allowed to adjust key process parameters within a safe adjustment range. During equipment maintenance, temporary operation permissions are automatically opened to maintenance personnel, and high-risk instructions are restricted. At night or during non-production periods, access rights to non-essential equipment are automatically tightened to reduce the risk of misoperation.

[0027] Regarding step S5, in the TinyML model, the abnormality probability is a complex process that integrates feature analysis, model reasoning and probability calibration. For example, when detecting bearing faults in smelting equipment, after obtaining vibration signals and current signals, the TinyML model trained based on historical data will output the abnormality probability of the bearing according to the abnormalities of the vibration signals and current signals. In addition, the abnormality probability will be corrected and calibrated in combination with environmental characteristics. The amount of data input to the TinyML model is adjusted in real time based on the abnormality probability, which can effectively utilize computing resources. The use of sliding windows makes it easy to control the amount of multi-source sensor data obtained. For example, adjusting the size and step size of the sliding window can directly affect the amount of data obtained. In long-term normal working conditions, reducing the amount of data obtained by the sliding window can achieve lightweight processing of the input data of the TinyML model. For example, if the output abnormality probability does not exceed 5% within the preset eight hours, the time step of the sliding window can be appropriately increased and the frequency of window sliding can be reduced to reduce the amount of data input to the TinyML model. In occasional abnormal working conditions, such as when the abnormality probability is greater than 40%, the amount of data obtained by the sliding window can be increased accordingly, which can increase the frequency of window sliding and specifically enhance the TinyML model's ability to identify abnormalities. In addition, the first threshold and second threshold for controlling the amount of data obtained can be determined according to the on-site working conditions.

[0028] The present invention deploys a trained lightweight machine learning model at the edge, which can realize direct data processing at the edge without uploading to the cloud, avoiding the risk lag caused by network delays, that is, localized decision-making can still operate stably in harsh network environments, and can monitor smelting equipment in real time and make necessary control strategies, especially the authority control of smelting equipment under various abnormalities. It can adaptively generate the security boundary of process parameters based on historical data as the authority range for controlling smelting equipment, which is better than the traditional fixed threshold method. It can also ensure that unauthorized dangerous operations are blocked within milliseconds, such as abnormal behaviors caused by misoperation, program vulnerabilities or malicious attacks, and block physical operations and network attack risks from the source of authority. The constructed active defense system meets the smelting needs of real-time, high reliability, physical safety and network security. At the same time, based on the output feedback of the TinyML model, the lightweight processing of real-time collected data can reasonably reduce the operating burden of the TinyML model. Combined with the lightweight structural design of the TinyML model, the TinyML model can also run well on low-cost edge devices, such as Advanced RISC The microcontroller with Machine architecture does not require complex hardware upgrades and is suitable for smelting companies to carry out intelligent transformation of a large number of existing smelting equipment, reducing energy consumption and deployment costs.

[0029] In some optional embodiments, based on the output abnormality probability and abnormality type, a graded response is processed, including: different abnormality types correspond to abnormality probabilities; when the abnormality probability is <30%, a log is recorded and the adjustment range of the process parameters is optimized; when 30%≤abnormality probability<70%, an alarm is triggered and non-critical operations are restricted; when the abnormality probability is ≥70%, a forced shutdown is initiated and a maintenance notification is triggered.

[0030] A multi-level dynamic protection is formed through a hierarchical response mechanism, which can achieve risk pre-control. That is, after a certain abnormal type is identified by the TinyML model and the corresponding abnormal probability is output, when the abnormal probability is low-risk <30%, the log is recorded and the process parameter adjustment range is optimized to actively reduce the probability of the corresponding abnormal type. For example, when the TinyML model identifies an abnormal type that may cause bearing peeling, the high-frequency vibration amplitude range can be tightened accordingly, and the root mean square value of the vibration acceleration can be limited to avoid the expansion of damage caused by high-frequency impact. However, it will not actively interfere with the normal operation of the equipment at present; when the medium risk is 30%≤abnormal probability<70%, based on the abnormal type corresponding to the abnormal probability The TinyML model can isolate the corresponding operating permissions and limit some non-critical operations of the corresponding equipment to prevent human errors from exacerbating faults. For example, when the TinyML model identifies an abnormality in the current harmonic distortion rate of a motor and determines that the abnormal type of bearing wear may have occurred, and the corresponding abnormal probability is medium risk, the TinyML model will trigger an alarm to the relevant personnel and prohibit increasing the speed permission of the corresponding equipment, but will retain the core functions to maintain production continuity. When the abnormal probability is high (≥70%), a safety fuse mechanism needs to be implemented. Forced shutdown can be set to minimize the impact of the equipment abnormality. Combined with rapid response of operation and maintenance, it can effectively reduce the incidence of major accidents.

[0031] This mechanism uses an abnormal state-authority linkage model to ensure the real-time response capability of the production system and build a defense-in-depth system. In addition, it can also implement dynamic rate limits on high-frequency access interfaces to prevent edge device paralysis caused by DDoS attacks.

[0032] In some optional embodiments, step S5 further includes: obtaining the abnormality probability output by the TinyML model in real time; if the abnormality probability is greater than 30%, reducing the sliding window length by 40% to 80%; if the abnormality probability output within a preset time is less than 10%, increasing the sliding window length by 20% to 30%.

[0033] Smelting equipment failures (such as bearing spalling and gear tooth breakage) usually manifest as short-term high-frequency vibrations or current fluctuations. When the probability of an abnormality is too high, the sliding window can be reduced. For example, the time sliding window can be shortened from the default 1 second to 100ms. This can improve the ability to capture high-frequency abnormal characteristics (such as sudden mechanical vibration and current spikes), avoid the smoothing or delay of key abnormal information due to an overly large window, and improve time resolution, achieve rapid response, and adapt to the computing power limitations of edge devices.

[0034] The preset time can be set according to the needs of the factory. For example, if the abnormal probability of the output for 24 consecutive hours is less than 10%, the length of the sliding window can be expanded. That is, within the preset time, the equipment has been in normal operation and no obvious abnormal characteristics have been found. Noise interference can be smoothed through longer-period data, and the adaptability of the TinyML model to steady-state working conditions can be improved, thereby reducing the false alarm rate, reducing the frequency of TinyML model calls, and reducing the computing load.

[0035] Based on the combination of real-time monitoring of abnormal probability and historical statistics, dynamic adjustment of the sliding window length is achieved, which can adapt to changes in the equipment's smelting stages and seasonal operating conditions, forming a fast response-slow optimization closed-loop control.

[0036] In some optional embodiments, real-time acquisition of multi-source sensor data of smelting equipment further includes: when the operating parameters of the smelting equipment are close to critical values ​​or fluctuate frequently in small amplitudes, increasing the frequency and duration of corresponding data collection.

[0037] If the operating parameters of smelting equipment approach critical values ​​or experience frequent small fluctuations, these may indicate potential failures. Therefore, increasing the frequency and duration of data collection can obtain more information at critical moments, thereby increasing the probability of discovering abnormal signs. For example, when the rolling force of smelting equipment approaches the equipment's load limit multiple times in a short period of time, the corresponding sensors are immediately activated to collect data at high frequency. Collecting multiple sensor data at this stage can provide more targeted abnormal data for TinyML model reasoning, thereby effectively improving the ability to identify equipment abnormalities.

[0038] In some optional embodiments, the smelting equipment includes an oxygen lance, and a TinyML model is used to monitor the torque current of the oxygen lance and the CO concentration in the flue gas. If the slope of the change of the torque current exceeds a third threshold, or the torque current changes but the CO concentration does not change, the oxygen lance operation authority is locked; if the rate of change of the CO concentration exceeds a fourth threshold, the operation authority of the relevant valve is locked and the backup sensor is switched to detect the CO concentration.

[0039] During the smelting process, the security of oxygen lance control instructions is crucial. To prevent malicious attacks or misoperation that could cause abnormal oxygen lance operation, such as sudden acceleration or incorrect descent, a two-way verification mechanism has been designed. This mechanism ensures the authenticity and rationality of control instructions by detecting the physical consistency of the motor current waveform and flue gas analysis data. Furthermore, corresponding control methods are provided for forged instructions or data tampering. This two-way verification mechanism ensures that oxygen lance control complies with both the physical laws of the equipment and the requirements of the smelting process through cross-verification of electrical signals and process parameters. This effectively protects against cyberattacks and misoperation risks. For example, if a forged oxygen lance descent instruction occurs (such as a malicious acceleration causing a splash accident) or sensor data is tampered with (such as masking the actual flue gas concentration), a dual protection chain is established by simultaneously verifying the physical execution effect of the control instruction (i.e., torque and current changes) and the physical consistency of the process parameters (i.e., continuous changes in CO concentration). This can strongly couple network access control with the metallurgical process mechanism.

[0040] Since the current change of the oxygen gun motor shows a smooth upward trend (S-shaped curve) during normal acceleration, if a sudden drop command is received, the current will increase at a specific slope due to the inertia of the motor load. Therefore, by monitoring the three-phase current of the motor, the torque current component can be extracted and the current change slope per unit time can be analyzed. Under normal operation, the current change slope should be within a reasonable range. For example, under the operating conditions of a certain factory, the reasonable range corresponding to the current change slope is less than 500A / ms. If the torque current change slope is detected to be abnormal, for example, the preset abnormal threshold is 20% higher than the theoretical value, that is, when the torque current change slope reaches 600A / ms, the abnormality detection sub-model can determine that the oxygen gun control command is a forged command based on the input of sensor data, and will immediately lock the oxygen gun operation authority to prevent the implementation of the erroneous command.

[0041] At the same time, CO concentration changes are limited by chemical reaction kinetics and cannot jump significantly in a very short period of time. Therefore, a TinyML model is used to monitor the CO concentration sequence of the flue gas analyzer in real time and calculate the change rate K of adjacent data points. During normal smelting, due to the decarburization reaction rate, its change rate K is usually less than 5%. If a sudden change in CO concentration is detected, such as the CO concentration change rate K greater than 5%, the TinyML model can determine that a malicious attack of data tampering may have occurred. In particular, if this is not accompanied by a change in the oxygen lance action, the possibility of data tampering is higher. Accordingly, an alarm will be triggered and valve control rights will be frozen. Freezing valve control rights is an important protection mechanism. By locking the operating rights of key valves, especially the oxygen regulating valve and shut-off valve, it can avoid false operations caused by data tampering, thereby ensuring process safety, equipment stability and personnel safety. Its core logic is to prioritize blocking the risk source when the abnormality is unclear. Therefore, switching to the backup sensor to re-test the CO concentration can further verify the accuracy of the abnormality judgment, facilitating subsequent control decisions.

[0042] Therefore, when the current waveform based on the oxygen lance action conforms to the motor characteristics and the flue gas data changes smoothly, it is normal operation and the control rights of related equipment such as the oxygen lance and valves can be opened accordingly. However, when the torque current changes but the CO concentration does not change, it is determined that it is likely a forged instruction generated by a network attack, and the oxygen lance operation rights are locked to protect production and equipment. In addition, when the current suddenly changes and the CO concentration also jumps, that is, the slope of the torque current change exceeds the third threshold and the rate of change of the CO concentration exceeds the fourth threshold, it is likely that a more serious network attack has occurred and an emergency smelting interruption can be set accordingly for response protection.

[0043] In some optional implementations, after locking the oxygen lance operation authority, if the oxygen lance is in operation, the oxygen lance is controlled to complete the operation before the oxygen lance drive is cut off; if the oxygen lance is not in operation, the oxygen lance drive is directly cut off.

[0044] When the TinyML model identifies an anomaly in the oxygen lance, such as a cyber attack such as forged instructions or data tampering, it will first lock the oxygen lance operation permission, and then make further control processing based on the working status of the oxygen lance, thus forming a complete automated control chain, and then being able to respond promptly to cyber attacks and ensure the normal operation of smelting operations.

[0045] In some optional embodiments, a TinyML model is used to monitor the oxygen flow rate of the oxygen lance. When the oxygen flow rate change rate is greater than a third threshold, the molten pool surface temperature, the sound intensity of the molten pool in the 300-500 Hz frequency band, and the density of the splashing droplets in the molten pool are obtained. The TinyML model is used to identify the solution splashing in the molten pool. If the solution splashing phenomenon does not occur, the oxygen lance control authority is frozen and manual review is initiated.

[0046] During the smelting process, the molten pool is the area where liquid metal accumulates and also serves as the working area of ​​the oxygen lance. Solution splashing is a common physical phenomenon during oxygen blowing using the oxygen lance. When the oxygen lance increases its oxygen output, it is usually accompanied by violent molten pool disturbance, solution splashing, and gas release. When the oxygen flow rate change exceeds a third threshold (the third threshold is determined based on actual factory operating conditions), for example, a rate of change > 20% indicates a sudden increase in oxygen flow. This could be due to a specific process requirement, malicious instructions, or sensor failure. Therefore, a TinyML model is used to monitor multiple sensor parameters of the molten pool, and splashing characteristics are used as an auxiliary basis for determining whether the oxygen flow rate is abnormal. This clever combination of smelting physical phenomena improves the oxygen lance's ability to identify anomalies, effectively avoiding production disruptions or even unplanned shutdowns due to misjudgments, and ensuring the safe operation of smelting equipment. A dynamic permission control system with millisecond-level response and intelligent process identification has been established, providing excellent security protection for high-risk smelting operations.

[0047] In some optional embodiments, such as Figure 2 As shown in the figure, the TinyML model is used to identify the solution splashing of the molten pool, including: based on the molten pool surface temperature, the sound intensity of the molten pool in the 300-500 Hz frequency band, and the density of the molten pool splashing droplets, three characteristic parameters are obtained, namely the standard deviation of the molten pool temperature distribution, the intensity change of the molten pool sound, and the density of the molten pool splashing droplets. If at least two of the characteristic parameters do not meet the preset conditions, it is determined that the molten pool does not have the solution splashing phenomenon.

[0048] In order to accurately identify the splashing characteristics and distinguish normal process operations from abnormal attack behaviors, multimodal sensor fusion detection can be used. Specifically, the surface temperature distribution of the molten pool can be monitored in real time by an infrared thermal imager. The temperature gradient is relatively gentle during normal oxygen blowing (about 5~10℃ / cm), while the local temperature suddenly changes when splashing occurs. Therefore, the standard deviation T of the temperature distribution can be calculated. The standard deviation T corresponds to a normal reference value of 5~10℃. The preset condition of the standard deviation T can be determined according to the on-site working conditions. For example, the preset condition is that when T>30℃, it meets the phenomenon of solution splashing. Further, a high-temperature resistant microphone can be used to capture the acoustic signal generated by the bursting of bubbles in the molten pool. When the solution splashes, the sound energy will be concentrated in the 300~500Hz frequency band, and the sound pressure level (S Therefore, when a sudden increase in energy in this frequency band is detected, for example, the preset condition can be a sudden increase of 50dB in sound intensity for more than 100ms, which is consistent with the characteristics of solution splashing in the molten pool. Finally, a high-speed camera can be used to capture the trajectory of solution splashing. During normal operation, the splashing density is low, for example, the splashing droplets are less than 100 drops / frame, while the splashing density will increase sharply during abnormal splashing. For example, the preset condition can be that the splashing droplets are greater than 500 drops / frame, which is consistent with the phenomenon of solution splashing. Therefore, the TinyML model is used to monitor the parameters of multiple different sensors in the molten pool. The authenticity of the physical phenomenon can be cross-verified through multimodal perception methods, effectively avoiding misjudgment of a single sensor, and ensuring the reliability of safety decisions.

[0049] The following describes a TinyML-based smelting control device provided in an embodiment of the present application. The device described below and the method described above can refer to each other. Based on the above embodiment, Figure 3 It is a structural diagram of a TinyML-based smelting control device provided in this embodiment.

[0050] like Figure 3As shown, a smelting control device based on TinyML includes: an acquisition module 10, which is used to obtain historical operating data of smelting equipment under normal working conditions and abnormal working conditions, and generate abnormal behavior data of smelting equipment; a modeling module 20, which is used to construct a TinyML model, including a parameter adjustment sub-model and an anomaly detection sub-model, the parameter adjustment sub-model is formed by a series combination of a convolutional neural network and a long short-term memory network, and the anomaly detection sub-model is formed based on a lightweight classifier of Transformer-Attention; a training module 30, which trains the parameter adjustment sub-model based on the historical operating data of normal working conditions to output a safe adjustment range of process parameters, and trains the parameter adjustment sub-model based on the historical operating data and abnormal behavior data An anomaly detection sub-model outputs anomaly probability and anomaly type; a deployment module 40 is used to deploy the trained TinyML model on the edge of the smelting equipment, use a sliding window to obtain multi-source sensor data of the smelting equipment in real time, input the trained TinyML model, adjust the access rights of the smelting equipment API interface based on the safe adjustment range of the output process parameters, and process the output anomaly probability and anomaly type with a graded response; a feedback module 50 is used to obtain the anomaly probability output by the anomaly detection sub-model in real time. If the anomaly probability output within the preset time does not exceed the first threshold, the amount of data obtained by the sliding window is reduced. If the anomaly probability output exceeds the second threshold, the amount of data obtained by the sliding window is increased.

[0051] Based on the above method embodiment, please see Figure 4 , Figure 4 This is a structural diagram of an electronic device provided in an embodiment of the present application. An electronic device 300 provided in an embodiment of the present application includes: a processor 301 and a memory 302, the memory 302 stores machine-readable instructions executable by the processor 301, and the machine-readable instructions execute the above method when executed by the processor 301; the electronic device can be a physical device.

[0052] Based on the above method embodiments, a computer-readable storage medium is further provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method in any of the above embodiments are implemented.

[0053] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0054] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0055] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0056] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0057] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0058] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0059] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information, which can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0060] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0061] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0062] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A smelting control method based on TinyML, characterized in that: include: S1, obtaining historical operating data of smelting equipment under normal and abnormal operating conditions, and generating abnormal behavior data of smelting equipment; S2: Build a TinyML model, including a parameter adjustment sub-model and anomaly detection sub-model. The parameter adjustment sub-model is formed by combining a convolutional neural network and a long short-term memory network in series, and the anomaly detection sub-model is formed by a lightweight classifier based on Transformer-Attention. S3: Train the parameter adjustment sub-model based on historical operating data under normal conditions to output the safe adjustment range of process parameters. Train the anomaly detection sub-model based on historical operating data and abnormal behavior data to output the anomaly probability and anomaly type. S4 deploys the trained TinyML model on the edge of the smelting equipment. A sliding window is used to acquire multi-source sensor data from the smelting equipment in real time. The trained TinyML model is then input into the model. Based on the safe adjustment range of the output process parameters, the access rights to the smelting equipment API interface are adjusted. Based on the output anomaly probability and anomaly type, a hierarchical response is implemented for processing. S5, obtain the abnormal probability output by the abnormality detection sub-model in real time. If the abnormal probability output within the preset time does not exceed the first threshold, reduce the amount of data obtained by the sliding window. If the abnormal probability output exceeds the second threshold, increase the amount of data obtained by the sliding window.

2. The TinyML-based smelting control method according to claim 1, characterized in that: Based on the output abnormal probability and abnormal type, graded response processing is carried out, including: Different types of abnormalities have corresponding abnormal probabilities. When the abnormal probability is <30%, a log is recorded and the adjustment range of process parameters is optimized. When 30%≤abnormal probability<70%, an alarm is triggered and non-critical operations are restricted. When the abnormal probability is ≥70%, a forced shutdown is triggered and a maintenance notification is triggered.

3. The TinyML-based smelting control method according to claim 1, characterized in that: Step S5 further includes: Obtain the anomaly probability output by the TinyML model in real time. If the anomaly probability is greater than 30%, reduce the sliding window length by 40% to 80%. If the anomaly probability output within the preset time is less than 10%, increase the sliding window length by 20% to 30%.

4. The TinyML-based smelting control method according to claim 1, characterized in that: Use sliding windows to acquire multi-source sensor data from smelting equipment in real time, including: When the operating parameters of the smelting equipment are close to critical values ​​or fluctuate frequently in small amplitudes, increase the frequency and duration of corresponding data collection.

5. The TinyML-based smelting control method according to any one of claims 1 to 4, characterized in that: The smelting equipment includes an oxygen lance. The TinyML model is used to monitor the torque current of the oxygen lance and the CO concentration in the flue gas. If the slope of the torque current change exceeds the third threshold, or the torque current changes but the CO concentration does not change, the oxygen lance operation authority is locked. If the rate of change of the CO concentration exceeds the fourth threshold, the operation authority of the relevant valve is locked and the backup sensor is switched to detect the CO concentration.

6. The TinyML-based smelting control method according to claim 5, characterized in that: A TinyML model is used to monitor the oxygen flow rate of the oxygen lance. When the oxygen flow rate change rate is greater than the third threshold, the surface temperature of the molten pool, the sound intensity of the molten pool in the 300-500 Hz frequency range, and the density of the molten pool splashing droplets are obtained. The TinyML model is used to identify the solution splashing in the molten pool. If the solution splashing phenomenon does not occur, the oxygen lance control authority is frozen and manual review is initiated.

7. The TinyML-based smelting control method according to claim 6, characterized in that: Use the TinyML model to identify the solution splashing in the molten pool, including: Based on the molten pool surface temperature, the sound intensity of the molten pool in the 300-500 Hz frequency range, and the density of the molten pool splashing droplets, three characteristic parameters are obtained: the standard deviation of the molten pool temperature distribution, the intensity change of the molten pool sound, and the density of the molten pool splashing droplets. If at least two of the characteristic parameters do not meet the preset conditions, it is determined that no solution splashing occurs in the molten pool.

8. A smelting control device based on TinyML, characterized in that: include: An acquisition module is used to acquire historical operating data of smelting equipment under normal and abnormal operating conditions, and generate abnormal behavior data of smelting equipment; The modeling module is used to build the TinyML model, including a parameter adjustment sub-model and anomaly detection sub-model. The parameter adjustment sub-model is formed by combining a convolutional neural network and a long short-term memory network in series, and the anomaly detection sub-model is formed by a lightweight classifier based on Transformer-Attention. The training module trains the parameter adjustment sub-model based on historical operating data under normal conditions to output the safe adjustment range of process parameters, and trains the anomaly detection sub-model based on historical operating data and abnormal behavior data to output anomaly probability and anomaly type; The deployment module deploys the trained TinyML model on the edge of the smelting equipment. It uses a sliding window to acquire multi-source sensor data from the smelting equipment in real time. The module inputs the trained TinyML model and adjusts the access rights of the smelting equipment API based on the safe adjustment range of the output process parameters. The module then responds to the abnormality based on the output probability and type of the abnormality. The feedback module is used to obtain the anomaly probability output by the anomaly detection sub-model in real time. If the anomaly probability output within a preset time does not exceed a first threshold, the amount of data obtained by the sliding window is reduced. If the anomaly probability output exceeds a second threshold, the amount of data obtained by the sliding window is increased.

9. An electronic device, characterized in that: The method comprises a memory and a processor coupled to the memory, wherein the processor is configured to execute the method according to any one of claims 1 to 7 based on instructions stored in the memory.

10. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images