Sulfuric acid liquid level fluctuation early warning method, system and medium

Abstract

This invention belongs to the field of modern sulfuric acid preparation, specifically involving a method, system, and medium for early warning of sulfuric acid level fluctuations. The method includes: real-time acquisition of multivariate time-series data corresponding to sulfuric acid level fluctuations in a drying circulation tank; inputting the multivariate time-series data into a pre-trained sulfuric acid level fluctuation early warning model; performing feature extraction and classification on the multivariate time-series data through the model; and outputting a predicted level fluctuation result for the drying circulation tank within a preset future time window. If the predicted level fluctuation result indicates fluctuation, an early warning is issued. By constructing a mapping relationship between "historical observation window multivariate data" and "future level fluctuation state corresponding to the control response time lag period (e.g., 15 to 30 minutes in the future)," accurate fluctuation prediction results can be output in advance before the level substantially exceeds the limit, thus providing sufficient intervention and control time for the control system or operators, effectively preventing drastic fluctuations in the drying circulation tank level or exceeding the safe operating range.

Description

Technical Field

[0001] This application belongs to the field of modern sulfuric acid preparation, specifically involving sulfuric acid level fluctuation early warning methods, systems, and media. Background Technology

[0002] In the dry absorption process of modern sulfuric acid production, the stability of the liquid level in the drying circulation tank directly affects the safe operation of the production system and product quality. Because the distributed control system in the sulfuric acid dry absorption process generally exhibits significant response time lag, traditional passive feedback control strategies struggle to promptly adjust abnormal liquid levels back to the process setpoint, easily leading to drastic liquid level fluctuations that may even exceed the safe operating range. Therefore, by utilizing high-frequency time-series data of multiple variables such as liquid level, sulfuric acid concentration, and flow rate collected on-site, a proactive classification model is constructed to predict liquid level fluctuation trends over a future period, thus shifting from passive regulation to proactive early warning and intervention.

[0003] To achieve accurate prediction of complex multivariate time series data, existing technologies have proposed a multivariate time series classification architecture that combines local sequence morphological feature extraction with global long-range dependency modeling. This architecture captures the differences between categories by extracting highly discriminative local morphological subsequences from the input time series, and uses an encoder based on a self-attention mechanism to capture the dynamic coupling relationship and global time delay effect between multiple process variables. It has demonstrated excellent accuracy in conventional time series classification tasks.

[0004] However, directly applying the aforementioned advanced multivariate time series classification architecture to the early warning task of sulfuric acid dry absorption process faces severe technical bottlenecks. In the actual construction of the liquid level warning model, in order to fully capture the potential correlation between historical dynamic evolution characteristics and future fluctuation risks, it is necessary to use a sliding observation window to segment continuously collected multidimensional high-frequency multivariate sensor data to generate the input sequence time series samples required for supervised learning. This time step-based sliding window mechanism inevitably leads to extremely long overlapping sampling regions between adjacent time series samples, resulting in an extremely large and highly similar redundant original training time series sample set. When a complex classification model with a local morphological subsequence extraction mechanism attempts to continuously optimize within this massive and redundant original training time series sample set of sulfuric acid process time series to extract the most recognizable local morphological features, it requires extremely dense sequence point-to-point distance calculations and feature matching. This leads to extremely high memory resource consumption and exponentially expanding computational time overhead, severely restricting the large-scale training and rapid iterative deployment of complex warning models in chemical plants.

[0005] To alleviate the problem of excessive computational overhead, those skilled in the art typically employ traditional time-series sample simplification strategies based on nearest neighbor rules or spatial clustering before training the early warning model. However, in the actual operation of sulfuric acid drying circulation tanks, the key time-series segments characterizing the transition from a stable to a state of violent fluctuation in liquid level often exhibit local irregularities. These segments are usually located in the boundary regions of overlapping categories in the feature space, and their local neighborhood structures are extremely mixed. If traditional screening methods based on spatial neighborhood consistency are used to compress the original training time-series sample set of sulfuric acid liquid level, these high-value boundary time-series samples containing rich information for discriminating state transitions are easily discarded as outliers, while a large number of internally stable and redundant time-series samples far from the decision boundary are retained. This results in the classification model not only failing to effectively reduce computational consumption but also losing its ability to keenly discriminate sudden liquid level fluctuations. Summary of the Invention

[0006] Firstly, in view of the shortcomings of the prior art, the purpose of this application is to provide a sulfuric acid level fluctuation early warning method, which solves the technical problem in the prior art that the sulfuric acid level fluctuation cannot be predicted in advance due to the lag in control response and is difficult to achieve effective early warning.

[0007] The objective of this application can be achieved through the following technical solutions: A method for early warning of sulfuric acid level fluctuations includes: Real-time acquisition of multivariate time series data corresponding to sulfuric acid level fluctuations in the drying circulation tank; The multivariate time series data is input into a pre-trained sulfuric acid level fluctuation early warning model. The sulfuric acid level fluctuation early warning model extracts features and classifies the multivariate time series data, and outputs the predicted liquid level fluctuation of the drying circulation tank within a preset future time window. If the predicted liquid level fluctuation indicates that fluctuation has occurred, an early warning will be issued.

[0008] Secondly, in view of the shortcomings of the prior art, the purpose of this application is to provide a sulfuric acid level fluctuation early warning system, which solves the technical problem in the prior art that the sulfuric acid level fluctuation cannot be predicted in advance due to the lag in control response and is difficult to achieve effective early warning.

[0009] The objective of this application can be achieved through the following technical solutions: A sulfuric acid level fluctuation early warning system includes: The input module is used to acquire multivariate time series data corresponding to the sulfuric acid level fluctuation in the drying circulation tank in real time; The processing module is used to input the multivariate time series data into a pre-trained sulfuric acid level fluctuation early warning model, extract features and classify the multivariate time series data through the sulfuric acid level fluctuation early warning model, and output the predicted liquid level fluctuation of the drying circulation tank within a preset future time window. The output module is used to issue an early warning if the predicted liquid level fluctuation indicates that fluctuation has occurred.

[0010] Thirdly, in view of the shortcomings of the prior art, the purpose of this application is to provide a computer-readable storage medium that solves the technical problem in the prior art that the inability to predict sulfuric acid level fluctuations in advance due to control response lag makes it difficult to achieve effective early warning.

[0011] The objective of this application can be achieved through the following technical solutions: A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of the first aspect.

[0012] The beneficial effects of this application are: This invention transforms the traditional sulfuric acid level control optimization problem into a multivariate time series predictive classification problem. By constructing supervised learning samples of "historical 1-hour multivariate time series data" and "level fluctuation status in the next 15 to 30 minutes," the model can learn the potential correlation between historical dynamic characteristics and future fluctuation patterns. This achieves a fundamental shift from "passive response regulation" dependent on deviation feedback to "active predictive regulation" based on predictive information. This breaks through the limitation of traditional advanced process control systems that can only correct deviations after they occur, shifting the control window from post-correction to pre-warning, reserving sufficient intervention and control time for controllers or operators, and effectively avoiding drastic fluctuations in the level of the drying circulation tank or exceeding the safe operating range.

[0013] The proposed information gain-based sample selection method does not rely on a specific classifier or nearest neighbor rule, and exhibits stronger versatility and stability compared to classic algorithms such as CNN, ENN, and DROP3. This method uses CID as the distance metric to divide the original dataset into nearest and far neighbor subsets. It quantifies the discriminative information content by evaluating the degree to which each sample reduces the overall uncertainty before and after the dataset partitioning, and prioritizes retaining boundary samples with low information gain while eliminating redundant samples within each category. This method can significantly compress the training set size while maintaining the model's generalization performance, reducing the computational and memory overhead of the Shapeformer model in the Shapelet feature extraction and self-attention mechanism calculation processes, thus making the algorithm highly applicable to large-scale industrial data.

[0014] In the sample selection and model training stages, this invention utilizes Complexity Invariant Distance (CID) instead of traditional Euclidean distance or dynamic time warping. By introducing a complexity correction factor composed of the sum of squared differences of the first order into the distance calculation, this invention can amplify the complexity difference between fluctuating and stationary operating condition sequences, significantly widening the distance between dissimilar time series samples in the feature space. This not only effectively improves the class separability of subsequent entropy calculation and information gain evaluation, but also ensures that the time complexity of CID only increases linearly, eliminating the need for complex parameter tuning and sequence alignment, making it highly suitable for the characteristics of large-scale, high-frequency time series data collected in chemical plants. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the overall structure of an embodiment of this application; Figure 2 This is a schematic diagram illustrating the construction of training time series sample labels according to an embodiment of this application; Figure 3 This is a schematic diagram of the reasoning process of the sulfuric acid level fluctuation early warning model in an embodiment of this application; Figure 4 This is a comparison diagram of the inference results and real data of the IG-IS time series sample screening method according to an embodiment of this application; Figure 5 This application embodiment shows the inference verification effect of four classic algorithms—OSC, DROP3, ENN, and CNN—on the same synthetic test set. Figure 6 This is a graph showing the results of screening fluctuating time series samples (red) and stationary time series samples (blue) using the IG-IS method in this application embodiment; Figure 7 This is an embodiment of the present application. Figure showing the inference results of the liquid level in November; Figure 8 This is an illustration of an embodiment of this application. The inference result diagram at that time; Figure 9 This is a diagram showing the inference results of the CNN method in an embodiment of this application; Figure 10 This is a diagram illustrating the inference results of the ENN method in an embodiment of this application; Figure 11 This is a diagram illustrating the reasoning results of the DROP3 method in an embodiment of this application. Figure 12 This is a diagram showing the reasoning result of the OSC method in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In chemical production, especially in the drying-absorption process of sulfuric acid preparation, stable control of the sulfuric acid level in the drying circulation tank is crucial for ensuring continuous, safe, and high-quality production. Due to the inherent strong coupling, high inertia, and external disturbances within the production process system, relying on traditional advanced process control (APC) systems for real-time adjustment results in inherent control response lags, making it difficult to predict the risk of level instability in a timely manner. This often leads to level fluctuations being detected only after they occur, missing the optimal intervention window. Therefore, this invention aims to provide a method for predicting level fluctuations in advance, transforming reactive post-event adjustments into proactive early warnings, thus gaining valuable response time for operators. This invention does not rely on a single level measurement value but comprehensively utilizes multivariate process time-series data closely related to the operating state of the drying circulation tank. Through a well-trained sulfuric acid level fluctuation early warning model, it mines the dynamic patterns that predict future level changes, achieving high-precision prediction and early warning of level fluctuations within a preset time window.

[0019] Figure 1 A flowchart illustrating the early warning method for sulfuric acid level fluctuations is shown. Figure 1As shown, this embodiment of the invention provides a method for early warning of sulfuric acid level fluctuations, which includes steps S1, S2, and S3. In practice, step S1 is responsible for real-time acquisition of multi-dimensional, synchronous process variable data related to the operation of the drying circulation tank from the industrial control system (DCS / APC system), forming multivariate time series data. Step S2 inputs the real-time multivariate time series data acquired in step S1 into a pre-trained sulfuric acid level fluctuation early warning model. The sulfuric acid level fluctuation early warning model uses its internal neural network structure (such as the Transformer module and Shapelet module) to perform deep feature extraction and pattern recognition on the input time series data, and finally outputs a classification result, that is, predicting whether the drying circulation tank will transition from the current state to a level fluctuation state within a preset future time window (e.g., 15 to 30 minutes). Step S3 is based on the liquid level fluctuation prediction result output in step S2. When the result is "fluctuation has occurred", an early warning mechanism (such as audible and visual alarm, operation screen pop-up, SMS push, etc.) is immediately triggered. This effectively solves the technical problem in the background technology that the fluctuation cannot be predicted in advance due to the lag in control response and it is difficult to intervene in time, thus achieving the technical effect of proactive and early warning.

[0020] In this embodiment, step S1 involves acquiring multivariate time series data corresponding to the sulfuric acid level fluctuation in the drying circulation tank in real time. The multivariate time series data includes the measured value of the sulfuric acid level in the drying circulation tank and multiple synchronously acquired process variable data.

[0021] First, a distributed control system (DCS) or advanced process control (APC) system in the industrial field is used to acquire multivariate time-series data that is highly coupled with the liquid level status of the drying circulation tank in real time. To comprehensively characterize the extremely complex dynamic evolution process in the dry absorption step of sulfuric acid preparation, this step collected 30 key process variables covering nine dimensions, including sulfuric acid concentration, liquid level, pressure, flow rate, temperature, regulation feedback, and valve opening, as shown in Table 1. The multivariate time series data... This includes not only the measured sulfuric acid level in the drying circulation tank, but also data from multiple process variables that are physically or controllably related to this level. Specifically, these process variable data may include, but are not limited to, combinations of data from the following list: sulfuric acid concentration data in the drying circulation tank, sulfuric acid concentration data in the primary suction circulation tank, sulfuric acid level data in the drying circulation tank, level data in the primary suction circulation tank, sulfuric acid concentration adjustment feedback data in the drying circulation tank, sulfuric acid concentration adjustment feedback data in the primary suction circulation tank, sulfuric acid level adjustment feedback data in the drying circulation tank, finished acid flow rate adjustment feedback data to the finished product tank area, finished acid flow rate data in the outlet area, acid flow rate data on the drying tower, acid flow rate data on the primary suction tower, opening data of the dilute acid to the primary suction circulation tank regulating valve, opening data of the finished acid flow rate regulating valve, and secondary suction level adjustment. Data includes: valve opening degree, secondary nicotinic acid level regulating valve opening degree, primary nicotinic acid sulfuric acid concentration regulating valve opening degree, primary acid absorption supply to low-temperature heat recovery flow control valve opening degree, low-temperature heat recovery return to primary acid absorption flow control valve opening degree, drying tower outlet acid temperature, drying tower inlet acid temperature, primary absorption tower outlet acid temperature, primary absorption tower inlet acid temperature, drying tower inlet air temperature, drying tower inlet acid pressure, drying tower inlet air pressure, drying tower inlet and outlet pressure difference, primary absorption tower inlet and outlet pressure difference, primary acid absorption supply to low-temperature heat recovery flow rate, and low-temperature heat recovery return to primary acid absorption flow rate.

[0022] Table 1. 30 key variables, including sulfuric acid level in the drying tank. The system collects a data point every 4 seconds and inputs the collected multivariate time series data into a pre-trained sulfuric acid level fluctuation early warning model. The sulfuric acid level fluctuation early warning model performs feature extraction and classification on the multivariate time series data and outputs the predicted fluctuation results of the drying circulation tank level within a preset future time window.

[0023] Specifically, before conducting online inference after acquiring real-time data, a high-performance sulfuric acid level fluctuation early warning model is pre-built and trained to establish a nonlinear mapping relationship between historical dynamic evolution characteristics and future fluctuation risks.

[0024] Step S2.1: Construction of the original training time series sample set for the sulfuric acid level fluctuation early warning model. To fully capture the target variable With the remaining 29 process variables ( The potential coupling relationship between ) and, for the target variable of this application The data includes sulfuric acid level in the drying circulation tank, while the remaining 29 process variables... ( The remaining 29 process variables in Table 1 are used in this application; all 30-dimensional multivariate time series data are used. For example, when the system sampling frequency is set to 4 seconds / point, the historical observation window length is... It can be configured within the range of 600 to 1200 to predict the future window length. It can be configured within the range of 150 to 300. Preferably, such as Figure 2 As shown, this invention will adjust the historical observation window length. The optimal value is set to 900 (corresponding to 1 hour of historical data), and the preset future time window length is set. The optimal value is set to 225 (corresponding to the interval of 15 to 30 minutes in the future). This length not only fully covers the slow-changing dynamic evolution cycle of thermodynamics and fluid dynamics inside the drying circulation tank, but also perfectly spans the control response dead zone of about 15 minutes. This allows the sulfuric acid level fluctuation early warning model to extract sufficient precursor features before the level actually exceeds the limit, thus ensuring the foresight and effectiveness of the early warning in engineering practice.

[0025] Time series samples Therefore Starting point A continuous observation window with all 30 process variables as the endpoint. Its tags According to the interval Internal target variable status Sure: That is, if the interval Internal target variable If the time series sample is always in a fluctuating condition, then The tag is Otherwise, its label is .

[0026] Similarly, to divide continuous historical time-series data into discrete computational units, a sliding window mechanism is required. The sliding step size S is also a selectable variable, for example, it can be configured to 10, 50, 90, or 150, etc. Preferably, this invention sets the optimal value of the sliding step size S to 90, so that adjacent generated windows retain exactly 90% (i.e., 810 sampling points) of time-series overlap. This can both capture the transient evolution of the working conditions in a fine-grained manner to prevent feature omission and limit the disorderly expansion of the total amount of the original training time-series sample set at the source. Based on this, the constructed original time-series sample set It contains 1343 fluctuating time series samples and 11254 stationary time series samples, with a total time series sample size of 1343. .

[0027] Because industrial control systems typically have a control response time lag of approximately 15 minutes, the sulfuric acid level fluctuation early warning model needs to predict the level fluctuation within the next 15 to 30 minutes based on historical 1-hour data. Specifically, let the current time be... Historical observation window Multivariate time series Constructing inference time series samples Sulfuric acid level fluctuation early warning model output The predicted liquid level status within the interval, such as Figure 3 As shown.

[0028] The moment progressed to At that time, inference time series samples Updated to The corresponding prediction interval becomes If the sliding window step size is set to 1, the sulfuric acid level fluctuation early warning model will update with a new level prediction result each time a new inference time series sample is input. Therefore, this application uses inference time series samples... The output after inputting the sulfuric acid level fluctuation early warning model is considered as time. Liquid level state prediction value .

[0029] Original training time series sample set Generated using a sliding window method, total time series sample size There are 1343 fluctuating time series samples and 11254 stationary time series samples. The ratio of fluctuating time series samples to stationary time series samples is approximately... The original training time series sample set is significantly imbalanced. If this entire set is used to train the sulfuric acid level fluctuation early warning model, the following key challenges will be faced: (1) Computational overhead and memory usage: Time-series sulfuric acid level fluctuation early warning models, represented by Transformer, rely on a self-attention mechanism, and their key-value (KV) cache consumes a lot of GPU memory. In addition, some time-series feature extraction methods (such as Shapelet) need to calculate the point-to-point distance in each dimension, and the complexity increases sharply with the size of the time series samples. Massive redundant time series samples will lead to a longer training cycle and a surge in hardware resource requirements, which will seriously restrict the iterative efficiency of sulfuric acid level fluctuation early warning models.

[0030] (2) High redundancy of time series samples: In order to make full use of time series data while taking into account the efficiency of time series sample generation, the training time series samples are generated with a sliding step size of 90 sampling points, and there is an overlap region of up to 810 sampling points between adjacent time series samples. This sampling method results in extremely high similarity between time series samples and low information density. The large number of repetitive patterns not only increases the computational burden, but also easily induces the sulfuric acid level fluctuation early warning model to overfit to the common noise in the overlapping region, thus impairing the generalization performance.

[0031] (3) Label noise interference: The labels of time series samples are manually labeled by domain experts based on their experience and knowledge. Although expert judgment has a certain degree of reliability, the subjective labeling process inevitably introduces random errors and individual biases. Such noisy labels will distort the classifier's learning of the real data distribution and weaken the robustness and discrimination accuracy of the sulfuric acid level fluctuation early warning model.

[0032] (4) Uneven quality of time series samples: Existing research shows that not all time series samples contribute equally to the construction of classification boundaries. "Boundary time series samples" located in the class overlap region or near the decision boundary contain the richest discriminative information, while "internal time series samples" far from the boundary have little impact on the classifier. Removing the latter can significantly reduce the original training time series sample set with almost no loss in classification performance. Therefore, identifying and retaining boundary time series samples with high information concentration is the key to achieving efficient screening.

[0033] This application, from an information theory perspective, proposes an IG-IS time series sample selection method using complexity-invariant distance (CID) as the distance metric between time series samples. This method quantifies the amount of discriminative information contained in a time series sample by evaluating the degree of change in the purity of the original training time series sample set after each time series sample partitions the original training time series sample set around itself. The information gain selection process for time series samples is as follows: (1) Global entropy calculation: Calculate the original training time series sample set entropy .

[0034] (2) Distance matrix construction: For each time series sample Calculate its relationship with all other time series samples ( CID distance .

[0035] (3) Divide the neighborhood according to the distance threshold: use the distance threshold With radius, Divide into nearest neighbor subsets distant neighbor subsets .

[0036] (4) Calculate the weighted entropy: Calculate the weighted entropy based on the time series samples. Center Nearest neighbor subsets partitioned by radius and distant neighbor subsets Weighted entropy .

[0037] (5) Calculate information gain: distance threshold Information gain below .

[0038] (6) Retain the class with the lowest information gain. For each time series sample: calculate the optimal threshold for each time series sample. The information gain scores are then calculated, and the score with the lowest information gain is retained. A number of time series samples.

[0039] (5) Smaller, indicating that The original training time series sample set, after being partitioned around the center, still maintains a high degree of heterogeneity, with its local structure closely resembling the overall distribution. In extreme cases, This indicates that the time series sample cannot effectively reduce the uncertainty of the original training time series sample set. Such time series samples are more likely to be located in the class boundary region, and therefore have higher discriminative value. Conversely, if A larger information gain indicates that a high-purity local region can be formed centered on this time series sample. This time series sample is more likely to belong to a time series sample within the same category, and its contribution to the construction of the decision boundary is limited. Therefore, by selecting the sample with the lowest information gain in each class... By using a number of time series samples, redundant time series samples within a category can be effectively removed, while retaining boundary time series samples with strong discriminative ability, thereby significantly compressing the size of the training data while ensuring classification performance.

[0040] The specific details regarding the calculation of medium entropy and weighted entropy, the distance threshold selection strategy, and the selection of distance metric methods are as follows: (1) Entropy, weighted entropy and information gain Information entropy measures the uncertainty of the original training time series sample set; a higher value indicates a more uniform class distribution and higher uncertainty. For a binary classification task, the original training time series sample set... The entropy is defined as: in and Stationary and fluctuating time series samples, respectively. The proportion in, and The entropy reaches its maximum value of 1 when the number of samples in the two time series classes is equal; if the original training time series sample set contains only a single class, the entropy is 0.

[0041] Based on time series samples Distance threshold can Divide into nearest neighbor subsets distant neighbor subsets : in for and The CID distance. After partitioning, define the weighted entropy. The weighted average of the entropies of the two subsets: In the formula The weighted entropy represents the number of time-series samples in the set. The average uncertainty of the entire original training time series sample set after dividing the sample into neighborhoods around the center. Time series samples At the threshold Information gain below Defined as the decrease in entropy before and after the partition, i.e. .

[0042] (2) Time series distance metric The core of the sulfuric acid level fluctuation early warning model lies in accurately identifying fluctuating conditions to allow for early intervention and prevent energy consumption increases and safety risks caused by abnormal level fluctuations. This problem is essentially a multivariate time series advance classification task, the key being the effective differentiation between fluctuating and stationary time series samples. Distance metrics, as the basis for characterizing the dissimilarity between time series samples, directly affect class separability and thus determine the performance of the sulfuric acid level fluctuation early warning model. Therefore, choosing an appropriate distance metric is crucial. Based on the above analysis, this application selects the CID distance as the dissimilarity metric between time series samples. While maintaining computational lightweightness, it effectively highlights the essential differences between fluctuating and stationary time series samples. The Complexity-Invariant Distance (CID) introduces a complexity correction factor based on the Euclidean distance, amplifying the contribution of the complexity difference between sequences to the distance. The complexity correction factor CF is defined as the larger of the ratios of the complexity estimates of the two sequences: sequence Complexity estimation Using the first-order difference sum of squares: .

[0043] (3) Selection of the optimal distance threshold For each time series sample The set of distances between it and all other time series samples Sort in ascending order to obtain distinct candidate threshold sequences. ,in In turn, with each To determine the threshold, the original training time series sample set is... Divided into nearest neighbor subsets and distant neighbor subsets The corresponding information gain is Optimal distance threshold Defined as the candidate distance threshold that maximizes information gain, i.e.: Time series samples Information gain Ultimately, it is defined as the maximum information gain across all distance thresholds, i.e. This score reflects the... The maximum reduction in uncertainty that can be achieved by the optimal partition centered on [the target element]. But if [the target element is] centered on [the target element], then [the target element is] centered on [the target element]. If the information gain is still low compared to other time series samples of the same type, it means that even Efforts were made to improve the purity of the original training time series sample set, but compared to other similar time series samples, it still falls short. The original training time series sample set, after being partitioned by center, still maintains a high degree of heterogeneity. Its local structure is close to the overall distribution. Such time series samples are more likely to be located in the class boundary region, and therefore have high discriminative value. Therefore, subsequent steps will be based on... Within each category, sort in ascending order and prioritize retaining time series samples with lower scores.

[0044] Original training time series sample set Number of retained time series samples for each class Filter subsets The IG-IS time series sample selection method is adopted, and the specific algorithm is as follows: Requirement: Original training time series sample set Number of retained time series samples for each category Ensure: Select subset 1: Calculate global entropy: 2: for to do 3: Calculate time series samples CID distance to all time series samples: 4: Will Sort in ascending order 5: Initialize maximum information gain 6: for each candidate threshold (Determined by the sorted distance) do 7: Divide the original training time series sample set: 8: Calculate the weighted entropy: 9: Calculate information gain: 10: Update: 11: end for 12: end for 13: for each category do 14: Obtain the category with the lowest information gain. Time series samples 15: Add a subset 16: end for 17: return Regarding the above algorithm, the IG-IS time series sample selection method is as follows: First, based on the original training time series sample set The time series sample number distribution of the stationary time series sample set and the fluctuating time series sample set is used to calculate the original training time series sample set. The global entropy value H(D) is calculated using the following formula: ,in for or , and These represent the proportions of time series samples in the original training time series sample set D for the stationary operating condition category (labeled 0) and the fluctuating operating condition category (labeled 1).

[0045] Then, for each time series sample in the original training time series sample set... It is necessary to calculate the dynamic evolution characteristics of the liquid level embodied in the time series sample and compare them with the original training time series sample set. All other time series samples Time series distance between To effectively measure the differences between time series with different dynamic complexities, this embodiment uses complexity-invariant distance (CID) as the distance metric. The time series distance is the complexity-invariant distance, and its calculation formula is as follows: ,in, Time series samples and The Euclidean distance between them; The complexity correction factor is calculated using the following formula: ,in, Let X be the complexity estimate of the time series sample. It is calculated by calculating the root of the first-order sum of squared differences of the time series sample, i.e. The complexity correction factor (CF) amplifies the distance between sequences with large differences in dynamic pattern complexity, making the distance metric more sensitive to morphological changes in the sequence.

[0046] Next, for each time series sample The corresponding time series distance set After sorting the distinct values ​​in the data, multiple candidate distance thresholds were determined. For each candidate distance threshold With the current time series samples Centered on the candidate distance threshold The original training time series sample set D is divided into nearest neighbor subsets based on the radius. distant neighbor subsets Then, based on the category distributions in the nearest and far neighbor subsets, the weighted entropy at the candidate distance threshold is calculated. The weighted entropy is the weighted average of the entropies of the nearest neighbor subset and the far neighbor subset obtained based on the candidate distance threshold. ,in, This indicates the number of time series samples in the corresponding set. The information entropy of a set.

[0047] Subsequently, the information gain at the candidate distance threshold is determined based on the entropy reduction of the global entropy value relative to the weighted entropy. ,Right now The greater the information gain, the better the information is perceived by the current threshold. The greater the reduction in uncertainty in the category distribution after classification, the better the time series samples are. The greater the information contribution to class distinction below this threshold, the higher the information content for class differentiation. This applies to samples from the same time series. The calculated multiple information gains In the process, the maximum value is selected as the final information gain score of the time series sample. ,Right now .

[0048] Finally, for the stationary time series sample set and the fluctuating time series sample set, the final information gain score is applied respectively. The samples are sorted from low to high. Time series samples with low information gain scores mean that it is difficult to find a distance threshold centered on them that can significantly reduce the class uncertainty of the original training time series sample set. These time series samples are usually located in the core region inside their respective classes, and their information is highly redundant with the surrounding time series samples of the same class. Conversely, time series samples with high information gain scores are usually "boundary time series samples" located near the boundary between the two classes, and they are crucial for distinguishing between the two types of conditions. Therefore, a predetermined number of K boundary time series samples (i.e., the K with the lowest scores) with the highest ranking are selected from the time series samples corresponding to each class to form a simplified training set S. The sulfuric acid level fluctuation early warning model (such as Shapeformer) is trained using the simplified training set S. This method effectively eliminates a large number of redundant internal time series samples, significantly compressing the size of the training set (e.g., from 12597 time series samples to K=40 time series samples selected for each class, for a total of 80 time series samples) while retaining the most discriminative data. As a result, the sulfuric acid level fluctuation early warning model trained is optimized in terms of both computational efficiency and generalization performance.

[0049] In this embodiment, step S3 can be executed as an early warning triggering and execution step. When the sulfuric acid level fluctuation early warning model outputs a level fluctuation prediction result of "fluctuation occurred" (i.e., the corresponding fluctuation condition category) in step S2, the system will immediately execute an early warning action. The specific form of the early warning can be configured according to the needs of the actual production environment. For example, a prominent warning window can pop up on the operator station screen in the central control room, displaying the predicted fluctuation time and possible impact; an alarm can be triggered by a sound and light alarm device; or the early warning information can be pushed to the mobile terminals of relevant technical or management personnel through the enterprise messaging platform or SMS. This step transforms the prediction result of the sulfuric acid level fluctuation early warning model into direct operation instructions, completing a closed loop from data perception, intelligent analysis to decision execution, enabling the early warning information to be conveyed to relevant personnel in a timely and effective manner, thus gaining valuable time for intervention measures such as adjusting process parameters, inspecting equipment, or activating emergency plans.

[0050] This invention also provides an electronic device, which includes a processor and a memory. The memory stores computer programs, such as program code for executing the sulfuric acid level fluctuation early warning method described above. The memory can be random access memory (RAM), read-only memory (ROM), or other non-volatile storage media. The processor executes the computer program stored in the memory to implement the steps in the sulfuric acid level fluctuation early warning method described above, including real-time acquisition of multivariate time series data, calling and running a trained sulfuric acid level fluctuation early warning model for prediction, and triggering an early warning based on the prediction results. The processor can be an integrated circuit chip with signal processing capabilities, such as a central processing unit (CPU), digital signal processor (DSP), application-specific integrated circuit (ASIC), or graphics processing unit (GPU). The processor and memory can be connected via a bus. The electronic device may also include a data interface for communicating with an industrial control system and input / output devices (such as a display screen, alarm light, speaker, etc.) for human-machine interaction.

[0051] This invention also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps of the sulfuric acid level fluctuation early warning method described in the above embodiments. The computer-readable storage medium can be any medium capable of storing program code, such as a cloud storage device, USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0052] Verification 1: The effectiveness and generalization ability of the IG-IS time series sample screening method proposed in this application in the sulfuric acid level fluctuation early warning task are verified by simulation: Data synthesis mechanism: To characterize the inertial characteristics and closed-loop regulation behavior of the liquid level object, a second-order inertial industrial object model was used, combined with a proportional-integral (PI) control strategy, to simulate the dynamic change process of sulfuric acid liquid level under the action of a PID control system.

[0053] Assume the liquid level process setting value is And add a drift amount The valid setpoint is obtained: The liquid level at time t is obtained based on the second-order inertial model. Output The recurrence relation: in It is the second-order inertia coefficient. Indicates external disturbance. This is the control quantity. Under a proportional-integral (PI) control strategy, the control quantity is calculated as follows: in and These are the proportional and integral coefficients, respectively, and the control error. .

[0054] To simulate the two common operating conditions in industrial processes—steady and fluctuating—different drift patterns were designed for each. With disturbance Generation Mechanism. Under stable operating conditions, a small Gaussian perturbation and a weak Gaussian drift are superimposed, causing the liquid level sequence to exhibit low-variance oscillations and a slow trend drift. Under fluctuating operating conditions, a stronger Gaussian perturbation is superimposed, and a low-frequency sinusoidal drift is introduced. in The length of the current working section. Normalized time within the segment, This represents the drift amplitude. This mechanism generates structural low-frequency fluctuation characteristics.

[0055] Through the aforementioned closed-loop control structure and segmented disturbance mechanism, the generated liquid level sequence possesses a long-range dependency structure caused by second-order inertia and covers both steady-state and fluctuating operating conditions. This allows for a relatively realistic depiction of the dynamic behavior of industrial liquid levels under PID control conditions, providing a physically interpretable data foundation for subsequent simulation experiments. The specific values ​​of the relevant parameters are detailed in the experimental setup section.

[0056] Simulation parameter settings This application details the key parameters used in the data generation process. The training set duration is 60 days, and the inference set duration is 30 days. Both use the same generation mechanism and parameter configuration. Sampling period. Process setpoint The safe operating range is The hard limit boundary is Second-order inertia coefficient The proportional and integral coefficients are respectively set as , The integral term is limited to... To prevent divergence, when the liquid level exceeds the hard limit, the current value is forcibly pulled back within the boundary and the integral term is decayed by 30%.

[0057] The operating conditions are randomly generated with the following probabilities: a stable period has a probability of 0.8 and a length between 4,000 and 20,000 sampling points (corresponding to approximately 4.4 to 22.2 hours); a fluctuating period has a probability of 0.2 and a length between 1,000 and 12,000 sampling points (corresponding to approximately 1.1 to 13.3 hours). The disturbance term for the stable operating condition is... Drift item Obey random walk The initial value is 0. Fluctuation condition disturbance term. The drift term is generated by the sine function. ,in The current fluctuation segment length, Normalized time within the segment, amplitude .

[0058] All random processes used a fixed random seed (seed = 42) to ensure the reproducibility of the experiments. These parameters collectively determine the dynamic characteristics of the generated data, providing a unified experimental basis for subsequent time series sample selection and model inference.

[0059] Simulation Result Analysis Based on the IG-IS time series sample selection method, 40 time series samples with the lowest information gain values ​​were selected from both stationary and fluctuating time series sample sets for training the Shapeformer model. The inference performance of the sulfuric acid level fluctuation early warning model on the synthetic test set is shown below. Figure 4 As shown: Among them, sub Figure 4 In the diagram, (a) represents the actual distribution of synthetic inference data. Figure 4 (b) in the figure shows the inference results of the model after training with the IG-IS time series sample selection method. It can be seen that the inference results are in high agreement with the actual situation, and the overall prediction accuracy reaches 99.1%, which verifies the effectiveness of the method proposed in this application.

[0060] To verify the superiority of the IG-IS time series sample selection method, four classic algorithms—OSC, DROP3, ENN, and CNN—were selected as comparison methods. Forty time series samples were chosen from each category to train the Shapeformer model, and inference verification was performed on the same synthetic test set. The results are as follows: Figure 5 As shown in the comparison results, it is clear that the OSC inference results, as shown in the comparison results, are significantly better. Figure 5 As shown in (a); the DROP3 inference result is as follows: Figure 5 As shown in (b); the ENN inference result is as follows: Figure 5 As shown in (c); CNN inference results, as shown in Figure 5 As shown in (d); (1) The typical time series sample selection strategy based on clustering of the OSC algorithm completely failed in this experimental scenario. The model misclassified almost all stationary data as fluctuating data and could not effectively distinguish the data categories. (2) Although the DROP3 algorithm can identify most fluctuating sections, it has a serious misjudgment problem - a large number of stable sections are incorrectly marked as fluctuating sections, which is inconsistent with the core requirement of only identifying sections with severe fluctuations (or even beyond the safe operating range); (3) The performance of ENN and CNN algorithms is better than DROP3, and the degree of misjudgment is significantly reduced. However, there are still a few cases where stable segments are misjudged as fluctuating segments, which does not fully meet the requirements of practical applications.

[0061] In summary, the proposed IG-IS time series sample selection method can accurately select key time series samples with strong class discrimination power on the original training time series sample set. This helps the Shapeformer model learn clearer and more robust classification decision boundaries. Compared with traditional classic algorithms, it not only effectively reduces the misclassification rate of stationary segments but also ensures the recognition accuracy of fluctuating segments, ultimately achieving a high inference accuracy of 99.1%. This fully demonstrates the advantages of this method in time series sample selection tasks.

[0062] Experiments show that the IG-IS time series sample selection method proposed in this application accurately selects key time series samples with strong class discrimination power, helping the Shapeformer model learn clearer and more robust classification decision boundaries. Compared with traditional classic algorithms, it not only effectively reduces the misclassification rate of stationary segments but also ensures the recognition accuracy of fluctuating segments, ultimately achieving a high inference accuracy of 99.1%. Its performance is better than models trained by traditional time series sample selection methods such as CNN, ENN, DROP3, and OSC, fully demonstrating the advantages of this method in time series sample selection tasks.

[0063] Verification 2: Empirical verification of the effectiveness and generalization ability of the IG-IS time series sample screening method proposed in this application in the sulfuric acid level fluctuation early warning task: Training the original training time series sample set with the sulfuric acid liquid level fluctuation early warning model Using the sliding window method (window length of 900 time points, sliding step size of 90 time points), a total of 11254 original time series samples were generated, including 1343 fluctuating time series samples and 9911 stationary time series samples. Given that the number of stationary time series samples far exceeds that of fluctuating time series samples, to avoid class imbalance interfering with the training of the sulfuric acid level fluctuation early warning model, 1343 time series samples were sampled from the stationary time series samples to maintain a consistent sample size for both classes. Considering the large number of generated time series samples and the high redundancy of time series samples generated in adjacent windows, to reduce the sample size and thus lower the computational and memory overhead of the Shapelet feature extraction stage, the IG-IS method described in Chapter 2 was used to select the 40 time series samples with the lowest information gain from both the stationary and fluctuating classes, forming a simplified training set of size 80. Figure 6 The visualization results of some of the filtered time series samples are shown, where red represents fluctuating time series samples and blue represents stationary time series samples.

[0064] Subsequently, the simplified training set was randomly divided into a training set (56 time-series samples) and a validation set (24 time-series samples) in a 7:3 ratio. The training set was used to train the Shapeformer model, and the core hyperparameter configurations involved in training the sulfuric acid level fluctuation early warning model are shown in Table 2. Then, the optimal model with the highest accuracy on the validation set was selected for online inference.

[0065] Table 2. Shapeformer Model Hyperparameter Configuration Reasoning Results and Analysis when At that time, the IG-IS time series sample selection method selects the 40 time series samples with the lowest information gain values ​​from the original time series samples of both stationary and fluctuating classes, forming a simplified training set of size 80. The inference results of the Shapeformer model trained on this training set are as follows: Figure 7As shown, red indicates that the model predicts the sulfuric acid level will fluctuate drastically or even exceed the threshold at the corresponding time, while blue indicates that the model predicts the level will remain stable at the corresponding time. Overall, the sulfuric acid level fluctuation early warning model can accurately identify drastic fluctuations in the level and their turning points, and can also correctly determine periods of stable level, demonstrating good generalization performance and inference accuracy.

[0066] Given that this application focuses on the early warning capability for drastic liquid level fluctuations and events exceeding the threshold, the statistical performance of the model is primarily analyzed near the lower limit of the safe operating range (73) and the upper limit (79): liquid level A total of 278 alerts were issued, with 270 correct alerts, achieving an accuracy rate of 97.12%; liquid level... A total of 43 alerts were issued, with 42 correct alerts, achieving an accuracy rate of 97.67%; the liquid level was at... The interval (excluding 73) was tested 342 times, with 329 correct predictions, achieving an accuracy rate of 96.19%; the liquid level was within... A total of 613 predictions were made within the specified interval (excluding 79), with 594 correct predictions, achieving an accuracy rate of 96.90%. Overall, the sulfuric acid level fluctuation early warning model maintains a prediction accuracy exceeding 96% even in high-risk threshold areas, demonstrating strong risk prediction capabilities and stability.

[0067] Value sensitivity analysis The IG-IS time series sample selection method retains the sample with the minimum information gain in each class. A simplified training set is constructed from a set of time series samples, with parameters... It directly determines the size of the training set. If the value is too small, the number of time series samples in the validation set will be insufficient, resulting in a large variance in the accuracy assessment and unstable model selection; if... If the value is too large, it may introduce redundant or low-quality time series samples, increasing computational overhead and weakening the discriminative ability. For analysis... Impact on model performance, settings Four sets of comparative experiments. Figure 8 Showing The reasoning results at that time were all inferior to Figure 7 middle Performance during the calculation. Considering both reasoning accuracy and computational cost, 40 was ultimately selected as the performance level. The optimal value.

[0068] Comparison of different time series sample selection methods To systematically evaluate the effectiveness of the proposed method, this application selects several classic time series sample selection algorithms such as CNN, ENN, DROP3 and OSC, and conducts comparative analysis on a class-balanced sulfuric acid level original training time series sample set of size 2686, focusing on four aspects: noise filtering capability, time series sample reduction rate, total time consumption and model generalization ability maintenance.

[0069] Table 3 shows the comparison results of the above methods in terms of noise filtering, reduction rate, and time consumption. As can be seen from the table, CNN and ENN have relatively weak time series sample compression capabilities, with screening rates of 58.56% and 43.26%, respectively. This indicates that they retain a large number of original time series samples. Since both are based on kNN neighborhood structures for discrimination, the computation process relies on frequent neighborhood searches, resulting in high requirements for memory and computing resources. In contrast, DROP3 further implements a decreasing time series sample deletion strategy based on the ENN noise filtering mechanism, achieving a compression rate of 3.50%, but its computational complexity is high. The OSC method compresses time series samples through clustering, enabling screening to be completed in a shorter time, achieving a certain balance between compression rate and running efficiency.

[0070] To maintain consistency with the IG-IS time series sample selection method, this application randomly sampled 40 fluctuating time series samples and 40 stationary time series samples from the time series samples filtered by each comparison method to construct a training set. Based on the accuracy of the validation set, the optimal model was selected, and inference verification was performed on the November liquid level data. The analysis of the inference results is as follows: The inference results of the CNN method on data from a specified month, such as November, are as follows: Figure 9 As shown in the results, while the model correctly predicted most of the drastic fluctuations, it misclassified many stable time points as fluctuations. This may be because CNNs are difficult to apply to... The -NN rule correctly divides noisy time series samples, preserving both boundary time series samples and noisy time series samples.

[0071] Figure 10 The results of the ENN method inference show that the prediction results of ENN and CNN are very similar, and they correctly predict most cases of sharp fluctuations. However, they misclassify many stable time points as fluctuations. Although ENN can effectively remove isolated noise points based on NN rules, in this case, the liquid level fluctuation itself has local irregularities. While removing noise, ENN may inadvertently delete true boundary time series samples, thus weakening its sensitivity to sudden fluctuations.

[0072] The DROP3 forecast for November is as follows: Figure 11 As shown, the time series sample selection methods are largely the same as those using CNN and ENN rules. DROP3 prioritizes deleting internal time series samples based on the concept of "correlation" and incorporates an ENN noise filtering step. Existing research has shown that when using methods different from... When using a classification model with a neural network (NN), the generalization ability of the subset selection may decrease.

[0073] Figure 12 The OSC prediction results are generally poor. The OSC method emphasizes time series sample center compression, selecting representative time series samples through clustering or centrality measures. It tends to retain "typical time series samples" while ignoring difficult time series samples near the decision boundary, which is not conducive to threshold area early warning tasks.

[0074] In summary, the inference accuracy of the sulfuric acid level fluctuation early warning model trained on time series samples selected using the method described in this application is comprehensively superior to classic time series sample selection methods such as CNN, ENN, DROP3, and OSC. Most of the aforementioned traditional methods are based on... While neighborhood structures (NNs) are used for time series sample compression or noise removal, they tend to over-rely on local neighborhood consistency, resulting in insufficient mining of discriminative information for time series samples near class boundaries and difficulty in capturing subtle differences at key decision boundaries. This study, however, focuses on fine-grained discrimination tasks near liquid level thresholds. The information gain method proposed in this application directly measures the amount of discriminative information from time series samples for class classification. Its selection objective is not simply to pursue "neighborhood consistency," but rather to maximize the information representation capability of decision boundaries. Therefore, compared to traditional methods based on... Compared with the time series sample selection method of -NN, the information gain screening strategy proposed in this application has better performance in terms of boundary discrimination accuracy and generalization stability.

[0075] In the description of this specification, the references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0076] The foregoing has shown and described the basic principles, main features, and advantages of this application. Those skilled in the art should understand that this application is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this application. Various changes and modifications can be made to this application without departing from the spirit and scope thereof, and all such changes and modifications fall within the scope of the claims of this application.

Claims

1. A method for early warning of sulfuric acid level fluctuations, characterized in that, include: Real-time acquisition of multivariate time series data corresponding to sulfuric acid level fluctuations in the drying circulation tank; The multivariate time series data is input into a pre-trained sulfuric acid level fluctuation early warning model. The sulfuric acid level fluctuation early warning model extracts features and classifies the multivariate time series data, and outputs the predicted liquid level fluctuation of the drying circulation tank within a preset future time window. If the predicted liquid level fluctuation indicates that fluctuation has occurred, an early warning will be issued.

2. The sulfuric acid level fluctuation early warning method according to claim 1, characterized in that, Multivariate time series data includes sulfuric acid concentration data in the drying circulating tank, sulfuric acid concentration data in the primary suction circulating tank, sulfuric acid level data in the drying circulating tank, sulfuric acid level data in the primary suction circulating tank, sulfuric acid concentration adjustment feedback data in the drying circulating tank, sulfuric acid concentration adjustment feedback data in the primary suction circulating tank, sulfuric acid level adjustment feedback data in the drying circulating tank, finished acid flow rate adjustment feedback data to the finished product tank area, finished acid flow rate data in the outlet area, acid flow rate data on the drying tower, acid flow rate data on the primary suction tower, opening data of the dilute acid to the primary suction circulating tank regulating valve, opening data of the finished acid flow regulating valve, opening data of the secondary suction level regulating valve, and secondary nicotinic acid level adjustment. This includes various data such as valve opening degree, primary nicotinic acid and sulfuric acid concentration regulating valve opening degree, primary acid absorption supply to low-temperature heat recovery flow control valve opening degree, low-temperature heat recovery return primary acid absorption flow control valve opening degree, drying tower outlet acid temperature, drying tower on-top acid temperature, primary absorption tower outlet acid temperature, primary absorption tower on-top acid temperature, drying tower inlet air temperature, drying tower on-top acid pressure, drying tower inlet air pressure, drying tower inlet and outlet pressure difference, primary absorption tower inlet and outlet pressure difference, primary acid absorption supply to low-temperature heat recovery flow rate, and low-temperature heat recovery return primary acid absorption flow rate.

3. The sulfuric acid level fluctuation early warning method according to claim 2, characterized in that, The pre-trained sulfuric acid level fluctuation early warning model was obtained through the following training steps: Historical time series data of the multivariate time series data are obtained, and a preset number of time series samples are obtained by sequentially sampling the historical time series data through a sliding window. The time series samples are used to construct an original training time series sample set. Based on the fluctuation state of the sulfuric acid liquid level measurement value of the drying circulation tank in the historical time series data within a preset future time period corresponding to the control response lag after each historical observation window, the original training time series sample set is divided into a stationary time series sample set and a fluctuating time series sample set. Based on the distribution of the number of time series samples in the stationary and fluctuating time series sample sets of the original training time series sample set, the global entropy value of the original training time series sample set is calculated. For each time series sample in the original training time series sample set, the time series distance between the liquid level dynamic evolution characteristics embodied by the time series sample and all other time series samples in the original training time series sample set is calculated. For each time series sample, the distinct values ​​in the corresponding time series distance set are determined as multiple candidate distance thresholds. For each candidate distance threshold, the original training time series sample is processed with the current time series sample as the center and the candidate distance threshold as the radius. The time series sample set is divided into corresponding nearest neighbor subsets and far neighbor subsets. Based on the category distribution in the nearest and far neighbor subsets, the weighted entropy under the candidate distance threshold is calculated. The information gain under the candidate distance threshold is determined based on the entropy reduction of the global entropy value relative to the weighted entropy. From the multiple information gains calculated for the same time series sample, the maximum value is selected as the final information gain score for that time series sample. For the stationary time series sample set and the fluctuating time series sample set, they are sorted from low to high according to the final information gain score. A predetermined number of boundary time series samples with the highest ranking are selected from the time series samples corresponding to each category to form a simplified training set. The simplified training set was used to train the sulfuric acid level fluctuation early warning model.

4. The sulfuric acid level fluctuation early warning method according to claim 3, characterized in that, The process involves sequentially sampling the historical time-series data using a sliding window to obtain a preset number of time-series samples. These time-series samples are then used to construct an original training time-series sample set, including: Set the sampling frequency and determine the sequence length corresponding to the historical observation window. and the sequence length corresponding to the preset future time window. ; For any current time Extracting historical time series data containing multivariate time series data The continuous observation window within the interval is used as the original training time series sample; Based on the sulfuric acid level data of the aforementioned drying circulation tank, in the future State within the interval Determine the working condition labels of the training time series samples. The judgment condition is: in, This indicates that the training time series sample belongs to the fluctuating operating condition category. This indicates that the training time series sample belongs to the stationary operating condition category; Characterizing liquid level at It is constantly fluctuating; Operating condition label The original training time series samples are divided into stationary time series samples and fluctuating time series samples, and stationary time series sample sets and fluctuating time series sample sets are constructed according to the corresponding stationary time series samples and fluctuating time series samples.

5. The sulfuric acid level fluctuation early warning method according to claim 3, characterized in that, The time series distance is a complexity-invariant distance. The calculation formula is as follows: in, The Euclidean distance between time series samples; The complexity correction factor is calculated using the following formula: in, The complexity estimate, Time series samples The complexity estimate.

6. The sulfuric acid level fluctuation early warning method according to claim 3, characterized in that, The global entropy value The calculation formula is: in, and These are the original training time series sample sets. The proportion of stationary time series samples to fluctuating time series samples.

7. The sulfuric acid level fluctuation early warning method according to claim 3, characterized in that, The weighted entropy Based on candidate distance threshold The nearest neighbor subset obtained by partitioning With the distant neighbor subset The weighted average of entropy: in, This indicates the number of time series samples in the corresponding set. Represents a set Information entropy.

8. The sulfuric acid level fluctuation early warning method according to claim 1, characterized in that, The sulfuric acid level fluctuation early warning model includes the Shapeforme model.

9. A sulfuric acid liquid level fluctuation early warning system, characterized in that, include: The input module is used to acquire multivariate time series data corresponding to the sulfuric acid level fluctuation in the drying circulation tank in real time; The processing module is used to input the multivariate time series data into a pre-trained sulfuric acid level fluctuation early warning model, extract features and classify the multivariate time series data through the sulfuric acid level fluctuation early warning model, and output the predicted liquid level fluctuation of the drying circulation tank within a preset future time window. The output module is used to issue an early warning if the predicted liquid level fluctuation indicates that fluctuation has occurred.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images