Intelligent monitoring method and system for sulfur dioxide flue gas in waste lead-acid battery processing process

Abstract

The application discloses a kind of waste lead storage battery processing process sulfur dioxide flue gas intelligent monitoring method and system, it is related to industrial flue gas monitoring technical field, it is obtained time series monitoring data set by clock synchronization and fusion splicing to multiple channel sensing data through unified time reference, it is obtained multi-dimensional feature matrix by pre-processing and window feature extraction to it and carries out working condition marking according to concentration and temperature combination condition, after splicing feature matrix and working condition label, time series reasoning prediction is carried out to obtain concentration prediction sequence and synchronous evaluation equipment health degree, concentration prediction and health degree are combined to generate graded early warning signal and combine working condition label query decision rule base mapping control strategy, finally according to early warning level determination execution rate control instruction stepwise issue.The present application realizes the trend prediction and working condition self-adaptive accurate control of sulfur dioxide flue gas, improves the active control ability of desulfurization process.

Description

Technical Field

[0001] This application relates to the field of industrial flue gas monitoring technology, and more specifically, to an intelligent monitoring method and system for sulfur dioxide flue gas in the process of treating waste lead-acid batteries. Background Technology

[0002] In the recycling and smelting process of waste lead-acid batteries, sulfur-containing materials generate a large amount of sulfur dioxide flue gas during high-temperature smelting and oxidation reactions. If the emission concentration exceeds the standard, it will cause serious pollution to the surrounding atmospheric environment. Therefore, the stable operation of the flue gas desulfurization system and the real-time monitoring of sulfur dioxide concentration are the core links in the environmental compliance of the recycled lead industry. At present, the industry generally uses online continuous monitoring systems to detect the sulfur dioxide concentration in flue gas at fixed points and trigger alarms based on fixed thresholds. However, this method has obvious shortcomings: on the one hand, a single sensor data source is prone to data deviation due to clock drift and signal loss between channels, resulting in a decrease in the reliability of monitoring results; on the other hand, fixed threshold alarms can only respond after the concentration has exceeded the limit, lacking the ability to predict the trend of concentration changes in advance, making it difficult to gain enough time for early intervention in the process control of the desulfurization system.

[0003] Furthermore, the processing conditions of waste lead-acid batteries are complex and variable. Factors such as fluctuations in feed composition, changes in furnace temperature, and the circulation status of desulfurization liquid all significantly affect the characteristics of sulfur dioxide emissions. The evolution of flue gas concentration varies greatly under different operating conditions. Existing monitoring methods typically do not combine operating condition identification with concentration prediction, nor have they established a quantitative assessment mechanism for the operational health status of desulfurization equipment, resulting in a lack of targeted control decisions. When an early warning is triggered, operators often rely on experience to manually adjust desulfurization process parameters, which results in slow response speed and limited control accuracy, making it difficult to achieve a precise match between the warning level and the control intensity. Therefore, there is an urgent need for an intelligent monitoring method that can integrate multi-channel sensor data, identify real-time operating conditions, predict concentration trends, and automatically generate graded control strategies to improve the initiative and accuracy of sulfur dioxide flue gas treatment during the processing of waste lead-acid batteries. Summary of the Invention

[0004] This application provides an intelligent monitoring method and system for sulfur dioxide flue gas in the process of treating waste lead-acid batteries, in order to solve the technical problems existing in the prior art.

[0005] According to the first aspect of this application, a method for intelligent monitoring of sulfur dioxide flue gas in the waste lead-acid battery treatment process is provided, comprising: S1, synchronizing and fusing multi-channel sensor data collected using a unified time reference to obtain a sulfur dioxide flue gas time-series monitoring dataset; S2, preprocessing and extracting window features from the sulfur dioxide flue gas time-series monitoring dataset to obtain a multi-dimensional feature matrix for flue gas desulfurization, and labeling each window with operating conditions based on a combination of concentration and temperature ranges to obtain a desulfurization operating condition label sequence; S3, splicing the multi-dimensional feature matrix for flue gas desulfurization and the desulfurization operating condition label sequence and performing time-series inference prediction to obtain a sulfur dioxide concentration prediction sequence, and simultaneously evaluating the anomaly of each window to obtain a desulfurization equipment health vector; S4, jointly classifying the sulfur dioxide concentration prediction sequence and the desulfurization equipment health vector to obtain a graded early warning signal, and mapping the control strategy by querying the decision rule base using the desulfurization operating condition label sequence to obtain a desulfurization process control instruction set; S5, determining the control execution rate based on the graded early warning signal, and issuing the desulfurization process control instruction set in steps.

[0006] According to a second aspect of this application, an intelligent monitoring system for sulfur dioxide flue gas in the waste lead-acid battery processing process is provided, comprising: a data synchronization module, used to synchronize and fuse multi-channel sensor data collected using a unified time reference to obtain a sulfur dioxide flue gas time-series monitoring dataset; a feature annotation module, used to preprocess and extract window features from the sulfur dioxide flue gas time-series monitoring dataset to obtain a multi-dimensional feature matrix for flue gas desulfurization, and to annotate each window with operating conditions based on a combination of concentration and temperature ranges to obtain a desulfurization operating condition label sequence; and a prediction and diagnosis module, used to monitor the flue gas... The desulfurization multidimensional feature matrix and the desulfurization operating condition label sequence are concatenated and time-series inference predictions are performed to obtain the sulfur dioxide concentration prediction sequence. At the same time, the anomaly degree of each window is evaluated to obtain the desulfurization equipment health vector. The decision mapping module is used to jointly classify the sulfur dioxide concentration prediction sequence and the desulfurization equipment health vector to obtain the graded early warning signal. It also combines the desulfurization operating condition label sequence to query the decision rule base to perform control strategy mapping to obtain the desulfurization process control instruction set. The instruction issuance module is used to determine the control execution rate according to the graded early warning signal and issue the desulfurization process control instruction set in steps.

[0007] Compared with existing technologies, the intelligent monitoring method and system for sulfur dioxide flue gas in the waste lead-acid battery treatment process provided in this application has the following technical effects: By synchronizing and fusing multi-channel sensor data, the data synchronization deviation caused by inconsistent sampling time bases of various sensors is eliminated, providing a reliable input with strict time dimension alignment for subsequent feature extraction and model inference; through sliding window feature extraction and a working condition labeling mechanism based on concentration-temperature combination conditions, the system can identify the current load state of desulfurization operation in real time, providing a working condition context basis for differentiated selection of control strategies; by fusing working condition coding information with a multi-dimensional feature matrix for time series prediction, the system can accurately monitor sulfur dioxide concentration. The system can detect rising trends and issue early warnings before exceeding limits, significantly shortening the time window from the emergence of anomalies to system response. By jointly classifying concentration prediction results with equipment health scores, it avoids missed or false alarms caused by single-dimensional criteria, improving the overall reliability of warning signals. By jointly indexing the decision rule base with warning levels and operating condition categories and issuing control commands step by step at an adaptive rate, it achieves a precise match between control intensity and urgency. This prevents over-adjustment in low-risk scenarios from disturbing the system's steady state and ensures rapid response capabilities in high-risk scenarios. Overall, it transforms the treatment of sulfur dioxide flue gas in the process of waste lead-acid battery processing from passive response monitoring to prediction-driven proactive control. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a schematic flowchart illustrating the intelligent monitoring method for sulfur dioxide flue gas during the waste lead-acid battery treatment process according to an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of the data flow of the intelligent monitoring method for sulfur dioxide flue gas in the waste lead-acid battery treatment process according to an embodiment of this application.

[0011] Figure 3 This is a schematic flowchart of step S1 in the intelligent monitoring method for sulfur dioxide flue gas during the waste lead-acid battery treatment process according to an embodiment of this application.

[0012] Figure 4 This is a schematic flowchart of step S2 in the intelligent monitoring method for sulfur dioxide flue gas during the waste lead-acid battery treatment process according to an embodiment of this application.

[0013] Figure 5 This is a schematic flowchart of step S3 in the intelligent monitoring method for sulfur dioxide flue gas during the waste lead-acid battery treatment process according to an embodiment of this application.

[0014] Figure 6 This is a schematic flowchart of step S4 in the intelligent monitoring method for sulfur dioxide flue gas during the waste lead-acid battery treatment process according to an embodiment of this application.

[0015] Figure 7 This is a schematic block diagram of an intelligent monitoring system for sulfur dioxide flue gas during the waste lead-acid battery treatment process, as described in an embodiment of this application. Detailed Implementation

[0016] The embodiments of this application will now be described in more detail with reference to the accompanying drawings. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0017] like Figure 1 and Figure 2 As shown, this application provides an intelligent monitoring method for sulfur dioxide flue gas during the treatment of waste lead-acid batteries, comprising: S1. A time-series monitoring dataset for sulfur dioxide flue gas is obtained by clock synchronization and fusion of the collected multi-channel sensor data using a unified time reference. This multi-channel sensor data includes the sulfur dioxide concentration signal at the absorber inlet, the temperature signal inside the absorber, the flow rate signal of the smelting flue gas, the flow rate signal of the lean ion liquid, and the temperature signal at the bottom of the regeneration tower. It should be understood that in actual industrial deployment environments, the sensors are distributed in different spatial locations, such as the absorber inlet pipe, inside the absorber body, the main pipeline of the smelting flue gas, the lean ion liquid circulation pipeline, and the bottom of the regeneration tower. The communication link length, communication protocol type, and number of intermediate forwarding nodes between each sensor and the host computer gateway vary, resulting in different transmission delays experienced by each channel's data packets from physical acquisition to reception by the host computer. If this transmission delay difference is not compensated for, and the host computer's reception timestamp for each channel is directly regarded as the actual physical sampling time, then the multi-channel data read under the same time index actually corresponds to different physical times. This leads to the statistical features extracted subsequently within the same sliding window failing to accurately reflect the system's operating status at the same time segment.

[0018] In one embodiment of this application, such as Figure 3As shown, step S1 includes: S11, based on the pre-calibrated network communication transmission delay and the inherent hardware delay of analog-to-digital conversion, performing sensing delay calculation and timestamp compensation on the sampling points of each channel in the multi-channel sensing data to obtain a sensing data stream with compensated timestamps; S12, performing global time base mapping and resampling on the sensing data stream with compensated timestamps to obtain a synchronization-aligned data matrix; S13, performing horizontal feature splicing and structured encapsulation and recombination on the resampled values ​​at each time moment in the synchronization-aligned data matrix to obtain a sulfur dioxide flue gas time-series monitoring dataset.

[0019] First, the time information carried by each discrete sampling point in the multi-channel sensor data is corrected to restore the system arrival time of the data packets recorded by the host computer gateway to the actual sampling time when the sensor completes the analog-to-digital conversion at the physical end. In the industrial communication environment of the waste lead-acid battery processing site, the analog signals collected by each sensor are converted into digital signals by the analog-to-digital converter and then transmitted to the host computer gateway through communication media such as RS485 bus or industrial Ethernet. The delay introduced by this transmission process mainly consists of two parts: one is the network communication transmission delay, that is, the time taken for the digitized data packet to be sent from the sensor end to the host computer gateway to receive the complete data frame; the other is the inherent hardware delay of the analog-to-digital conversion, that is, the inherent processing time required for the sensor to convert the physical signal into a digital quantity.

[0020] For the m-th sampling point of the i-th sensing channel, the calculation process for its actual physical sampling timestamp is as follows: subtract the sum of the network communication transmission delay and the inherent hardware delay of the analog-to-digital conversion corresponding to that channel from the system arrival time of the data packet recorded by the host computer gateway. The difference obtained is the actual time when the sampling point is collected at the physical end. This calculation relationship is expressed as: in, This represents the actual physical sampling timestamp of the m-th sampling point of the i-th sensing channel after compensation; This represents the system arrival time recorded by the host computer gateway when the m-th data packet of the i-th sensor channel is received; This represents the pre-calibrated network communication transmission delay of the i-th sensing channel; This represents the inherent hardware delay constant of the analog-to-digital converter inside the i-th sensing channel.

[0021] The calculated actual physical sampling timestamp is key-bound with the original sampling value, that is, a correspondence between the real time and the measured value is established for each discrete sampling point, thereby generating a sensor data stream with a compensated timestamp.

[0022] In the first embodiment, the network communication transmission delay is considered a static constant obtained through a one-time calibration during the initial system debugging phase, and is not updated or corrected during the subsequent entire lifecycle operation. Specifically, during the debugging phase after system deployment, test signals with known timestamps are sent to each sensor channel, and the time difference between the signal's transmission from the sensor and its reception by the host computer gateway is recorded. This process is repeated multiple times, and the average value is taken as the calibrated transmission delay value for that channel. This method is simple to implement, has low computational overhead, and is suitable for scenarios with relatively stable communication environments. In this embodiment, the timestamp backtracking operation performed on each sampling point is the system arrival time of the data packet minus the sum of the pre-calibrated network communication transmission delay and the inherent hardware delay of the analog-to-digital conversion. For example, if the network communication transmission delay obtained by calibration for a certain channel is 15 milliseconds and the inherent hardware delay of the analog-to-digital conversion is 3 milliseconds, then a fixed compensation of 18 milliseconds is subtracted from any sampling point of that channel.

[0023] In the second embodiment, considering the actual industrial site of pyrometallurgical smelting of waste lead-acid batteries, the high-power electric arc equipment of the recycled lead smelting furnace operates on a periodic duty cycle, with ignition, heat preservation, and shutdown cycles. The strong electromagnetic pulse generated during the arc ignition stage is coupled to the communication bus within the workshop via spatial radiation, causing signal distortion and retransmission of data frames during transmission. This results in a quasi-periodic fluctuation characteristic in network communication transmission delay, coupled with the furnace's duty cycle. Static calibration values ​​cannot track this dynamic fluctuation, introducing residual time errors that vary with time at each sampling point. To address this problem, the second embodiment employs a dynamic adaptive delay compensation method.

[0024] In the second embodiment of this application, based on the pre-calibrated network communication transmission delay and the inherent hardware delay of analog-to-digital conversion, the sensing delay of each sampling point in the multi-channel sensing data is calculated and timestamp compensation is performed to obtain a sensing data stream with compensated timestamps. This includes: extracting and smoothing the communication delay drift trend of the multi-channel sensing data to obtain a time-varying delay drift trend sequence; performing residual deviation anchoring calibration and superposition correction on the time-varying delay drift trend sequence to obtain a dynamic delay compensation sequence after anchoring calibration; and based on the adaptive compensation value at the corresponding time in the dynamic delay compensation sequence after anchoring calibration and the inherent hardware delay of analog-to-digital conversion, backtracking and binding the system arrival time of each sampling point in the multi-channel sensing data with timestamp key values ​​to obtain a sensing data stream with compensated timestamps.

[0025] First, communication delay drift trend extraction and smoothing modeling are performed. Since each sensor periodically collects data with a known fixed nominal sampling period, in an ideal, interference-free communication environment, the time interval between adjacent data packets received by the host computer gateway should be strictly equal to the nominal sampling period. Once the communication link is affected by electromagnetic interference, the actual arrival interval will deviate from the nominal period value. Based on this, the actual arrival time interval of adjacent data packets at the host computer gateway is extracted channel by channel from the multi-channel sensor data. This actual arrival interval is then subtracted from the known nominal sampling period to obtain the instantaneous arrival jitter at each sampling point. The formula for calculating this jitter is: in, This represents the instantaneous arrival jitter calculated by the i-th sensing channel at the m-th sampling point; and These represent the system arrival times recorded at the host computer gateway for the (m+1)th and mth data packets of the i-th channel, respectively. Represents the known nominal sampling period constant of the i-th sensing channel.

[0026] Since instantaneous jitter not only includes slow-varying drift components related to the smelting furnace's operating mode, but also high-frequency random disturbances caused by random collisions on the communication bus and environmental electrical noise, directly using instantaneous jitter for delay compensation will introduce unnecessary high-frequency oscillations. To extract the slow-varying envelope reflecting the systematic trend of communication delay, an exponentially weighted moving average filter is applied to the instantaneous jitter sequence. The calculation formula is as follows: in, This represents the delay drift trend value obtained by smoothing and filtering at the m-th sampling point of the i-th channel, and constitutes an element of the time-varying delay drift trend sequence; This represents the drift trend value at the previous sampling point; This represents the forgetting factor in the exponentially weighted moving average. The closer the value of is to 1, the smoother the trend tracking and the stronger the suppression effect on high-frequency random disturbances. In practical applications, The value can be determined based on the ratio of the furnace's working cycle to the sampling frequency. For example, when the complete cycle of the furnace ignition-holding-stop is approximately 300 seconds and the sensor sampling period is 1 second, The value can be between 0.95 and 0.98 to ensure that the filter only allows the low-frequency drift components that match the furnace cycle to pass through.

[0027] Then, residual bias anchoring calibration and superposition correction are performed. A drift trend model established solely based on the statistical laws of data packet arrival jitter is an indirect observation and open-loop estimation of communication delay changes, which carries the risk of the trend estimate gradually deviating from the true value due to improper model parameter selection. To address this issue, this embodiment utilizes the naturally existing deterministic physical causal relationship during the absorption of lead smelting flue gas from waste lead-acid batteries into ionic liquid as an anchoring benchmark. Specifically, when the sulfur dioxide concentration at the inlet of the absorption tower undergoes a step change, since the reaction between sulfur dioxide and the ionic liquid is exothermic, the internal temperature of the absorption tower will change after an inherent response time lag uniquely determined by the reaction kinetic constant and heat transfer characteristics. This physical causal time lag is entirely determined by the laws of chemical thermodynamics and heat transfer, and is independent of the communication link state of the sensor, serving as an absolute anchoring benchmark in the time dimension.

[0028] Based on this physical prior, the sulfur dioxide concentration channel and the internal temperature channel of the absorption tower in the multi-channel sensor data are first compensated for using a time-varying delay drift trend sequence to eliminate the modeled systematic delay offset. Then, within a set sliding analysis window, cross-correlation calculations are performed on the pre-compensated concentration and temperature channels. The measured time lag between the two channels is located by searching for the peak position of the cross-correlation function. If the drift trend model is completely accurate, this measured time lag should be exactly equal to the predicted inherent physical time lag of the thermochemical causal response; otherwise, the difference is the residual calibration bias that the current drift model has not yet captured. The formula for calculating this residual bias is: in, This represents the measured time delay between channels corresponding to the peak position of the cross-correlation function after performing cross-correlation calculation on the sulfur dioxide concentration channel signal after preliminary compensation and the temperature channel signal inside the absorption tower within the m-th sliding analysis window. This represents the inherent physical time delay constant of the concentration-temperature causal response, which is predetermined by the thermodynamic constant of the sulfur dioxide absorption reaction, and serves as the intrinsic time anchor point. This represents the drift calibration residual deviation calculated for the i-th channel in the m-th analysis window.

[0029] The calculated residual bias is fed back and superimposed onto the original drift trend value and the static calibration reference to form an adaptive dynamic delay compensation value after physical causality anchoring calibration, generating a dynamic delay compensation sequence after anchoring calibration. The formula for generating this compensation value is: in, This represents the adaptive dynamic delay compensation value output after physical anchoring calibration. This represents the original statically calibrated reference value for network communication transmission delay. This represents the drift trend estimate derived from the corresponding position in the time-varying delay drift trend sequence; This represents the residual bias correction calculated using the physical causal anchoring method. This compensation value is composed of three progressively stacked levels: the static baseline provides a coarse initial estimate, the drift trend provides dynamic tracking correction, and the residual bias provides precise calibration of the physical anchoring.

[0030] In practical applications, The value can be determined by artificially applying a concentration step disturbance to the inlet of the absorption tower and simultaneously recording the temperature response delay during the system commissioning phase. For example, in a certain ionic liquid absorption tower, this value was calibrated to be 8 to 12 seconds.

[0031] Finally, adaptive dynamic delay superposition compensation and timestamp binding are performed. For each discrete sampling point in the multi-channel sensor data, the system indexes the corresponding adaptive dynamic delay compensation value from the anchored calibration dynamic delay compensation sequence. This value, along with the inherent hardware delay of the analog-to-digital conversion, constitutes the total time compensation for the current sampling point. A backtracking subtraction operation is performed on the system arrival time of the data packet to reconstruct the high-precision actual physical sampling time, and the reconstructed timestamp is key-bound to the original sampling value. The calculation formula is as follows: in, This represents the actual physical sampling timestamp restored from the m-th sampling point of the i-th channel after adaptive dynamic compensation; This represents the system arrival time recorded at the host computer gateway for the m-th data packet in the i-th channel; This represents the adaptive dynamic compensation value at the corresponding time obtained from the dynamic delay compensation sequence after anchoring calibration; This represents the inherent hardware delay constant of the analog-to-digital conversion within the i-th channel instrument.

[0032] After processing by the second embodiment, the timestamps carried by each sampling point in the output sensor data stream with compensated timestamps are integrated with drift trend tracking and physical causal anchoring calibration, so that the data of each channel can still maintain stable time alignment accuracy under the condition of periodic fluctuation of electromagnetic interference in smelting furnace.

[0033] Then, the sampling frequency differences and phase shifts that still exist in each channel after timestamp compensation are eliminated, so that the data of all channels are mapped to the same set of uniformly distributed time scales. Since the inherent sampling frequencies of each sensor may not be completely consistent (for example, the sampling period of a temperature sensor is 1 second while that of a flow sensor is 0.5 seconds), and even if the nominal sampling period is the same, there are random deviations in the actual sampling start phase of each channel at the physical end. Therefore, although the data streams of each channel after timestamp compensation carry accurate time stamps, their time scales are still not uniform, and it is not possible to directly perform cross-sectional comparisons time by time.

[0034] To address this issue, the system constructs a globally synchronized time axis with a fixed step size, which uses a preset uniform sampling period. A set of equally spaced global synchronization target time sequences is generated, with a step interval as the reference. For any global synchronization target time on this time axis... The algorithm iterates through the sensor data stream with compensated timestamps in the specified channel, finding two adjacent sampling nodes whose timestamps are immediately before and after the target time. Using the values ​​of these two adjacent nodes and their respective timestamps, it performs first-order linear interpolation to accurately estimate the corresponding value at the target time. The formula for this resampling interpolation is: in, This indicates that the i-th sensing channel is at the global synchronization target time. The reconstructed sample values ​​obtained from the calculation; This represents the k-th unified target time node on the global synchronization timeline; and These represent the sensor data streams with compensated timestamps, respectively, at the target time on the timeline. The m-th original sample value preceding it and the (m+1)-th original sample value following it; and These represent the actual physical sampling timestamps associated with the two sampling values ​​mentioned above.

[0035] In practical applications, the step size of the global synchronization time axis It can be set according to the highest sampling frequency in each channel. For example, when the shortest sampling period among all channels is 0.5 seconds, It can be set to 0.5 seconds or 1 second.

[0036] By performing the above resampling operation on each channel, the asynchronous sampling points of each channel are uniformly mapped to the same time reference plane, generating a synchronization-aligned data matrix. The number of rows in this matrix is ​​equal to the total number of global synchronization target times, the number of columns is equal to the number of sensor channels, and the values ​​of each column in the same row correspond to the reconstructed values ​​of each monitoring parameter at the same physical time.

[0037] Finally, the resampled values ​​of each time point and channel in the synchronized data matrix are structured to form a standardized dataset that meets the input data format requirements of subsequent time-series deep learning models and anomaly evaluation models. Specifically, using the global synchronization time axis as the row index and five channel attributes—sulfur dioxide concentration at the absorber inlet, absorber internal temperature, smelting flue gas flow rate, lean ion liquid flow rate, and regeneration tower bottom temperature—as column fields, the resampled values ​​of all channels at the same time reference are horizontally concatenated to construct a standard two-dimensional matrix structure. The generation of this matrix can be expressed as: in, This is the complete two-dimensional data matrix corresponding to the generated sulfur dioxide flue gas time-series monitoring dataset; column indices 1 to 5 correspond to the sulfur dioxide concentration at the inlet of the absorption tower, the internal temperature of the absorption tower, the flow rate of smelting flue gas, the flow rate of lean ion liquid, and the bottom temperature of the regeneration tower, respectively; K represents the total number of globally synchronized target time nodes within a processing batch, i.e., the total number of rows in the matrix.

[0038] S2 preprocesses and extracts window features from the sulfur dioxide flue gas time-series monitoring dataset to obtain a multi-dimensional feature matrix for flue gas desulfurization. Based on the combination of concentration and temperature ranges, each window is labeled with its operating condition to obtain a desulfurization condition label sequence. It is understandable that in the ionic liquid desulfurization process for lead smelting from waste lead-acid batteries, the operating state of the absorber tower varies significantly depending on the feed conditions. When the sulfur dioxide concentration in the smelting flue gas is high and the temperature inside the absorber tower rises due to the exothermic reaction, the desulfurization system operates under high load. At this time, the absorption capacity of the ionic liquid tends to be saturated, requiring an increase in the lean liquid circulation flow rate or an increase in the heating power of the regeneration tower to maintain desulfurization efficiency. When the sulfur dioxide concentration is low, the system operates under low load, and the circulation pump frequency can be appropriately reduced to save energy. If the subsequent prediction model does not understand the current operating state of the system, its prediction results will lack specificity, and the formulation of control strategies will be unable to distinguish the differentiated needs under different operating scenarios. Therefore, labeling the operating conditions of each time window during feature extraction, so that the operating condition information can be used as contextual input for subsequent inference and prediction, is a necessary prerequisite for realizing adaptive intelligent monitoring of operating conditions.

[0039] In one embodiment of this application, such as Figure 4As shown, step S2 includes: S21, performing data normalization and sliding window time-series slicing on the sulfur dioxide flue gas time-series monitoring dataset to obtain a normalized window segment tensor; S22, performing window statistical domain feature calculation and dimensionality reduction aggregation on the normalized window segment tensor to obtain a multi-dimensional feature matrix for flue gas desulfurization; S23, based on the preset multi-dimensional combination judgment conditions of sulfur dioxide concentration range and absorption tower temperature range, performing a step-by-step logical comparison and operating condition category labeling on the mean features of each window in the multi-dimensional feature matrix for flue gas desulfurization to obtain a desulfurization operating condition label sequence.

[0040] First, data normalization is used to eliminate the differences in numerical magnitude caused by different physical dimensions among the channels in the sulfur dioxide flue gas time series monitoring dataset. Then, the normalized long-term series is segmented into fixed-length sliding windows to form a three-dimensional tensor structure suitable for subsequent window-by-window statistical feature extraction.

[0041] In the desulfurization section of the waste lead-acid battery processing, there are significant differences in the physical dimensions and numerical ranges of the various monitoring channels. For example, the typical range of sulfur dioxide concentration at the inlet of the absorption tower is 0 to tens of thousands of milligrams per cubic meter; the typical range of the internal temperature of the absorption tower is 30 to 80 degrees Celsius; the value of the smelting flue gas flow rate can reach thousands to tens of thousands of cubic meters per hour; the flow rate of the lean ion liquid is tens to hundreds of cubic meters per hour; and the temperature at the bottom of the regeneration tower is 80 to 160 degrees Celsius. If these values ​​are not normalized and are directly used for feature calculations, channels with larger numerical magnitudes will dominate the statistical features, masking the effective information carried by channels with smaller magnitudes.

[0042] Min-max normalization was performed on the channel data of each column in the sulfur dioxide flue gas time-series monitoring dataset, mapping each channel data to a dimensionless interval of 0 to 1. For the original sampled value of the c-th channel at time t, the normalization calculation formula is: in, This represents the original sampled value of the c-th channel in the sulfur dioxide flue gas time-series monitoring dataset at time t; This represents the maximum extreme value observed in the global historical dataset for the c-th sensor channel; This represents the minimum extreme value observed by the c-th sensor channel in the global historical dataset; This represents the dimensionless data points after min-max normalization.

[0043] In practical applications, and This can be determined based on the extreme operating condition records of each channel in the system's historical operating data. For example, for the sulfur dioxide concentration channel at the inlet of the absorption tower, 0 mg / m³ is acceptable. The historical peak concentration of the smelting process under maximum feed load (e.g., 50,000 mg / m³) can be used; for the temperature channels inside the absorption tower, Ambient temperature (e.g., 20 degrees Celsius) is acceptable. The highest operating temperature allowed by the tower design can be taken (e.g., 90 degrees Celsius).

[0044] After normalization, a sliding time window of a preset length is introduced on the global synchronization time axis. Moving along the time axis with a fixed sliding step, the normalized data of the long sequence is truncated and sliced ​​into multiple local time segments. The slice range of the j-th time window is defined as: in, This represents the j-th multi-channel local time window segment. This represents the fixed step size value by which the sliding time window slides along the stepping direction on the time axis; This indicates the length of a fixed data point that the sliding time window captures on the time axis; This indicates the total number of channels, which is 5 in this implementation.

[0045] In practical applications, The value of should be determined based on the dynamic response time constant of the desulfurization process. Considering that the reaction time of ionic liquid absorption of sulfur dioxide and the cycle of lean liquid circulation are usually on the order of tens of seconds to several minutes, the window length can be set to 60 to 300 sampling points (60 to 300 seconds when the sampling cycle is 1 second). The value of determines the degree of overlap between adjacent windows, and is usually set to one-half to three-quarters of the window length. For example, when = 120, It can be set to 60 or 30 to control the computational load while ensuring time resolution.

[0046] Through the above operations, the two-dimensional normalized time series data is converted into a three-dimensional tensor structure containing the number of windows × window length × number of channels, i.e., the normalized window segmented tensor.

[0047] Then, statistical domain features that characterize the system's operating state within each local time window in the normalized window segment tensor are extracted, and the feature vectors of each window are stacked in rows to achieve dimensionality reduction and aggregation from a three-dimensional tensor to a two-dimensional feature matrix.

[0048] Iterate through each local time window in the normalized window segmented tensor, and calculate the statistical domain characteristics for each sensing channel within the window. Specifically, this includes extracting the mean within the window to reflect the basic operating level of each parameter during that time period, and extracting the standard deviation to reflect the degree of fluctuation of each parameter during that time period.

[0049] The formula for calculating the mean characteristic of the c-th channel within the j-th time window is as follows: in, This represents the k-th normalized time series data point in the c-th channel within the j-th time window of the normalized window segment tensor. This represents the window mean feature of the extracted signal from the c-th channel within the j-th window; Indicates the length of the local time window.

[0050] The standard deviation feature of the c-th channel within the j-th time window is calculated using the following formula: in, The window standard deviation feature of the extracted c-th channel signal within the j-th window is represented by the unbiased estimation form.

[0051] The mean and standard deviation of all channels within the same window are concatenated sequentially into a one-dimensional feature vector. For the case where the mean and standard deviation are extracted for each of the five channels, the feature vector for each window has a dimension of 10. The feature vectors for all windows are then stacked row-wise in chronological order to form a two-dimensional multi-dimensional feature matrix for flue gas desulfurization. The number of rows in this matrix equals the total number of sliding windows, and the number of columns equals the total dimension of the concatenated statistical features for each channel.

[0052] In practical applications, in addition to the mean and standard deviation, features such as the rate of change within the window (i.e., the difference between the window's end value and its beginning value divided by the time span corresponding to the window length) and the Pearson correlation coefficient between channels can be further introduced as needed to enrich the information representation capability of the feature matrix. Whether to include these additional features can be determined based on the validation results of the model's performance in the specific application scenario.

[0053] Finally, based on the actual operating characteristics of the lead smelting ionic liquid desulfurization process for waste lead-acid batteries, and using the preset combination of concentration and temperature ranges as judgment conditions, the desulfurization operation condition categories of each time window in the multidimensional feature matrix of flue gas desulfurization are automatically classified and labeled, generating a desulfurization operation condition label sequence.

[0054] In the process of removing sulfur dioxide from flue gas from recycled lead smelting using ionic liquid circulating absorption, the concentration of sulfur dioxide in the flue gas directly determines the pollution treatment load level of the desulfurization system, while the change in the internal temperature of the absorption tower reflects the degree of absorption reaction. When the sulfur dioxide concentration in the inlet flue gas is high, the heat released by the absorption reaction between the ionic liquid and sulfur dioxide increases, leading to an increase in the internal temperature of the absorption tower; conversely, when the concentration is low, the heat release from the reaction decreases, and the temperature inside the tower tends to stabilize. Therefore, the combined state of these two parameters—sulfur dioxide concentration and absorption tower temperature—can effectively characterize the current operating condition of the desulfurization system.

[0055] From the multi-dimensional feature matrix of flue gas desulfurization, the window mean features characterizing the sulfur dioxide concentration at the absorber inlet and the window mean features characterizing the internal temperature of the absorber are extracted separately. Based on preset high and low load concentration thresholds and temperature critical thresholds, multi-condition logic branch judgment is performed for each time window. The judgment formula is as follows: in, This represents the discrete state label output after determining the working condition for the j-th time window; This represents the mean value of the j-th window corresponding to the sulfur dioxide concentration at the inlet of the absorption tower in the multidimensional feature matrix of flue gas desulfurization (this value is a normalized dimensionless value). This represents the mean value of the j-th window corresponding to the internal temperature of the absorption tower in the multidimensional feature matrix of flue gas desulfurization (both are normalized dimensionless values). This indicates a pre-set threshold for judging high-load concentrations used to define sulfur dioxide concentration overload; This refers to the pre-set high-load temperature threshold used to define the risk of high temperatures inside the absorption tower. This indicates a pre-set low concentration threshold used to define low-load operating conditions.

[0056] A label value of 2 indicates that the desulfurization system is under high load, meaning that the inlet sulfur dioxide concentration is high and the absorption tower temperature is rising simultaneously, and the system is facing significant desulfurization pressure. A label value of 0 indicates that the system is under low load, the inlet sulfur dioxide concentration is at a low level, and the system is under low operating pressure. A label value of 1 indicates that the system is under normal operating conditions, and all parameters are in the middle range of the design operating range.

[0057] In practical applications, the above thresholds should be determined based on the design parameters and operational experience of the specific desulfurization process. For example, in the ion-liquid desulfurization system of a waste lead-acid battery processing plant, the designed sulfur dioxide concentration in the smelting flue gas entering the absorption tower is approximately 4%, the initial lean liquid flow rate is controlled at around 60% and gradually increases with the load, and the absorption temperature is controlled within the range of 30 to 40 degrees Celsius. Based on these process parameters, It can be set to a value corresponding to 80% or more of the design concentration within the normalized interval (e.g., 0.7). It can be set to a value corresponding to the upper limit of the normal absorption temperature within the normalized range (e.g., 0.65). It can be set to a value corresponding to below 30% of the design concentration within the normalized interval (e.g., 0.25). These thresholds can also be calibrated by statistical analysis of historical system operating data and combined with the experience of maintenance personnel in judging the operating conditions.

[0058] The classification values ​​of all windows are assembled into a one-dimensional vector in chronological order, thus generating a desulfurization operating condition label sequence. The length of this sequence is the same as the number of rows in the multi-dimensional feature matrix of flue gas desulfurization, and the two correspond one-to-one in the index dimension, so that the statistical feature vector of each window has its corresponding operating condition category label.

[0059] S3 involves splicing the multidimensional feature matrix of flue gas desulfurization with the desulfurization operating condition label sequence and performing time-series inference prediction to obtain a sulfur dioxide concentration prediction sequence. Simultaneously, anomaly assessment is performed on each window to obtain the desulfurization equipment health vector. It is understandable that in the ion-liquid desulfurization process for lead smelting from waste lead-acid batteries, the operation of the absorption tower is affected by a combination of factors. The smelting flue gas flow rate fluctuates with the batches of furnace feed, the absorption capacity of the lean ion-liquid gradually decreases with the number of cycles, and the regeneration efficiency of the regeneration tower is constrained by the accuracy of the bottom temperature control. The interaction of these factors means that the future evolution of sulfur dioxide concentration cannot be predicted by simple linear extrapolation, but requires a sequence modeling method capable of capturing long-term time dependencies. Furthermore, when the desulfurization equipment experiences abnormal conditions (such as absorption tower packing blockage, decreased lean liquid circulation pump efficiency, or scaling of the regeneration tower heat exchanger), its performance pattern in the multidimensional feature space will deviate from the normal operating data distribution. However, this deviation often does not directly lead to excessive emission concentrations in the early stages, and traditional single concentration threshold alarm mechanisms cannot capture such potential risks. Therefore, by executing the two reasoning paths of concentration trend prediction and equipment anomaly detection in parallel, a comprehensive assessment of system risk can be conducted from two complementary dimensions: emission trends and equipment status.

[0060] In one embodiment of this application, such as Figure 5As shown, step S3 includes: S31, converting each discrete state label in the desulfurization operating condition label sequence into an equal-length sparse numerical vector through one-hot encoding to obtain the encoded operating condition numerical vector, and concatenating and fusing the column dimensions of the flue gas desulfurization multidimensional feature matrix and the encoded operating condition numerical vector to obtain the fusion inference input tensor; S32, using a pre-trained long short-term memory network model to perform temporal dynamic expansion and forward inference prediction on the fusion inference input tensor to obtain the sulfur dioxide concentration prediction sequence; S33, using a pre-trained isolated forest algorithm model to perform anomaly inference and reverse health mapping on the fusion inference input tensor to obtain the desulfurization equipment health vector.

[0061] First, the continuous numerical multidimensional feature matrix of flue gas desulfurization and the discrete integer desulfurization operating condition label sequence are uniformly numerically represented and dimensionally concatenated to construct a unified input tensor that can be directly read by subsequent prediction models and anomaly detection models.

[0062] The elements in the desulfurization operating condition label sequence are discrete integer values ​​(e.g., 0, 1, and 2 represent low-load, normal, and high-load operating conditions, respectively). If these integer values ​​are directly concatenated with the continuous statistical features in the multi-dimensional feature matrix of flue gas desulfurization, the model may incorrectly interpret the magnitude relationship of the label values ​​as a natural order and proportional relationship between operating condition categories, introducing implicit assumptions that do not conform to physical meaning. To eliminate this problem, a one-hot encoding method is first used to transform each discrete label into a sparse vector of equal length, so that each operating condition category is expressed in a mutually orthogonal manner in the numerical space, and there is no distance bias between the categories.

[0063] For the discrete operating condition status label of the j-th time window in the desulfurization operating condition label sequence The one-hot encoding conversion process is as follows: generate a value with a length equal to the total number of operating condition categories. Given a vector of all zeros, find the index of that vector that is equal to... The element is set to 1, and the rest remain 0. This encoding operation is expressed as: in, This represents the discrete operating condition status label for the j-th time window in the desulfurization operating condition label sequence; This represents the total number of predefined operating condition categories for the desulfurization system; This indicates that the length of the output after one-hot encoding is... The sparse vector. In this embodiment, the operating condition categories include three types: low-load operating condition, normal operating condition, and high-load operating condition. Therefore = 3, the encoded vector dimension is 3. For example, when = 2 (high load condition), = [0, 0, 1]; when = 0 (low load condition), = [1, 0, 0].

[0064] After encoding, the operating condition vector processed by single-thermal encoding is concatenated and spliced ​​with the one-dimensional window statistical feature vector of the corresponding row in the multi-dimensional feature matrix of flue gas desulfurization along the feature column dimension. This splicing process is expressed as follows: in, This represents the one-dimensional window statistical eigenvector in the j-th row of the multi-dimensional feature matrix for flue gas desulfurization; This represents the vector concatenation operator; This represents the feature vector corresponding to the j-th sample row of the fused inference input tensor generated after splicing and fusion.

[0065] In this embodiment, each row of the flue gas desulfurization multidimensional feature matrix contains 10 statistical features, including the mean and standard deviation of each of the 5 channels. The one-heat encoding vector has a length of 3, so the feature dimension of each row of the concatenated fusion inference input tensor is 13. The feature vectors concatenated from all time windows are stacked row by row in chronological order to form the complete fusion inference input tensor, which has the same number of rows as the total number of time windows and 13 columns.

[0066] Then, the long short-term memory network model is used to extract the long-term dependence evolution of sulfur dioxide concentration in the time dimension from the fusion inference input tensor, and the sulfur dioxide concentration in the future preset time period is continuously predicted based on the learned time series pattern.

[0067] Long Short-Term Memory (LSTM) networks are recurrent neural network structures specifically designed for processing sequential data. Through three sets of gating mechanisms—forget gate, input gate, and output gate—they selectively forget old information, write new information, and read the current output between time steps, effectively mitigating the gradient vanishing problem in long-sequence processing and preserving the long-term memory of key historical events. In the desulfurization section of the waste lead-acid battery processing, the future evolution of sulfur dioxide concentration is influenced by various long-period factors, such as the furnace feeding rhythm, the decay trend of ion liquid absorption efficiency, and the regeneration tower cycle. The effects of these factors may not become apparent until dozens or even hundreds of time windows later. Therefore, a sequence model capable of modeling long-distance temporal dependencies is needed.

[0068] The fusion inference input tensor is fed into a pre-trained Long Short-Term Memory (LSTM) network model step by step. For the t-th time step, the model receives the feature vector of the current time step. The hidden state vector and cell state vector passed from the previous time step are used to calculate and update the hidden state and cell state at the current time step through a gating mechanism. This state update process is expressed as follows: in, This represents the feature vector input to the model from the fusion inference input tensor at the t-th time series step. This represents the hidden state vector passed from the Long Short-Term Memory network at the previous time step t-1; This represents the cell state vector passed from the Long Short-Term Memory network at the previous time step t-1; and These represent the hidden state vector and the cell state vector, respectively, calculated and updated by the gating mechanism at the current time step.

[0069] After the model completes the dynamic unfolding of the input sequence at all time steps, the hidden state vector output from the last time step is decoded through a fully connected mapping layer. This decodes the abstract representation in the high-dimensional hidden state space into a predicted sulfur dioxide concentration with clear physical meaning. This fully connected decoding and prediction process is expressed as follows: in, This represents the mapping weight matrix of the fully connected network layer at the output layer; This represents the bias constant vector of the fully connected network layer at the output level. This indicates the predicted future concentration of sulfur dioxide generated at the current time step.

[0070] The system reassembles the concentration values ​​generated at each time step in chronological order to form a sulfur dioxide concentration prediction sequence characterizing future emission trends. The length of this prediction sequence is determined by the number of prediction steps set in the model, and in practical applications, it can be determined based on the required control response time. For example, if the typical response time for a desulfurization system from receiving a control command to the effective adjustment of process parameters is 3 to 5 minutes, the number of prediction steps can be set to the corresponding 3 to 5 window steps to ensure that the system has sufficient time to execute control actions after the warning is issued.

[0071] Regarding the pre-training process of this Long Short-Term Memory (LSTM) network model, during the offline training phase before system deployment, fused feature sequences from the historical operating data of the desulfurization system are used as model input, with subsequent values ​​of the corresponding actual sulfur dioxide concentrations serving as supervision labels. The model parameters are optimized by minimizing the mean squared error between the predicted and actual values. The selection of the hidden layer dimension can be determined based on the dimension of the input features and the temporal complexity. For example, when the input feature dimension is 13 and the prediction steps are 5, the hidden layer dimension can be set between 64 and 128.

[0072] Finally, from the perspective of equipment operation status monitoring, the anomaly assessment of the feature vectors of each time window in the fusion inference input tensor is performed to quantify the degree to which the desulfurization equipment operation mode deviates from the normal distribution in each time period, and the anomaly score is converted into an intuitive health score output.

[0073] The Isolation Forest algorithm is an unsupervised anomaly detection method based on the idea of ​​random partitioning. Its core principle is that in a multidimensional feature space, anomalous sample points, due to their deviation from the main data distribution, are more easily isolated early during random partitioning, resulting in a shorter average search path length in the randomly constructed binary decision tree. Normal sample points, located in densely populated areas, require more random partitioning to be isolated, thus exhibiting a longer average search path length. This algorithm does not require any parameterized assumptions about the distribution of normal data and is suitable for complex industrial scenarios with multi-parameter coupling and diverse operating modes, such as desulfurization systems.

[0074] The feature vectors of each time window in the fusion inference input tensor are input into a pre-trained isolated forest model. This model consists of multiple randomly constructed isolated trees. For each input feature vector, the model performs a path walk from the root node to a leaf node in each isolated tree, recording the number of edges traversed to reach the leaf node as the path length for that tree. The average path length of all trees is then taken as the average search path length for that sample. Based on the ratio of this average path length to the expected path constant, an anomaly score is calculated for each time window. The formula for calculating this anomaly score is: in, Represents the current input feature vector The anomaly score output by the isolated forest model ranges from 0 to 1. Indicates input sample The expected value of the average search path length obtained by walk calculation in all random decision trees of the isolated forest model; Indicates based on a given subsample space size The derived average path length constant of the binary search tree is used as a normalization reference.

[0075] The closer the anomaly score is to 1, the shorter the path required for the sample to be isolated in the random tree, which means that the probability of the device's operating status deviating from the normal mode is greater for that time window; the closer the anomaly score is to 0.5, the more obvious the difference between the sample and the normal data subject; the lower the anomaly score is to 0.5, the more the sample is in the normal area with dense data.

[0076] To ensure that the joint rating operation in subsequent steps can be performed directly using the positive logic of "the higher the health, the better," anomaly scores are reverse-mapped and converted into health scores. The formula for this reverse mapping is: in, This represents the device health assessment score at time step t, calculated after inversion mapping. The closer the value is to 1, the more normal the device's operating status; the closer it is to 0, the more severe the abnormality.

[0077] The health scores of all time windows are arranged and combined in chronological order to generate a desulfurization equipment health vector. The length of this vector is equal to the total number of time windows, and is consistent with the number of rows in the multi-dimensional feature matrix of flue gas desulfurization and the length of the desulfurization operating condition label sequence.

[0078] Regarding the pre-training process of this isolated forest model, during the offline training phase before system deployment, historical fused feature data accumulated by the desulfurization system under normal operating conditions is used as the training set. The model learns from the distribution characteristics of the normal data to establish a baseline reference for the normal operating mode. The number of isolated trees in the model can be set to 100 to 200, and the subsample size of each tree... The number of sample points can be set to 256. The selection of these hyperparameters can be determined by tuning the model to evaluate its detection rate of known anomalous samples on the validation set.

[0079] In practical applications, when packing blockage in desulfurization equipment leads to a decrease in gas-liquid contact efficiency, the mean temperature in the multidimensional feature matrix may rise abnormally (the heat of reaction cannot be effectively removed), and the standard deviation of concentration may increase (uneven absorption leads to increased fluctuations in outlet concentration). These multidimensional collaborative shifts will significantly shorten the path length of the corresponding time window in the isolated forest model, thereby increasing the anomaly score and decreasing the health score, thus enabling early warning of potential equipment failures.

[0080] S4, the sulfur dioxide concentration prediction sequence and the desulfurization equipment health vector are jointly classified to obtain a graded early warning signal. This is then combined with the desulfurization operating condition label sequence query decision rule base to map the control strategy and obtain the desulfurization process control instruction set. It should be understood that in the ionic liquid desulfurization process for recycled lead smelting from waste lead-acid batteries, the risk of exceeding emission standards may originate from two different mechanisms: firstly, changes in the feed conditions of the upstream smelting process may lead to a sudden increase in the inlet sulfur dioxide load, causing the outlet concentration to exceed the limit if the desulfurization system does not respond in time; secondly, the desulfurization equipment itself may experience performance degradation (such as decreased ionic liquid absorption efficiency, scaling in the regeneration tower heat exchanger, and reduced efficiency of the lean liquid circulation pump), meaning that even if the inlet load does not change significantly, the outlet emission concentration may gradually increase. If early warning judgment is based solely on the concentration prediction results, risk signals may be missed when the equipment deteriorates early but the concentration has not yet exceeded the limit; if judgment is based solely on equipment health, false alarms may occur when the load suddenly increases but the equipment itself is in normal condition. Therefore, by jointly assessing the risks from both dimensions and taking the most severe risk level as the overall warning level of the system, we can ensure that no risk is overlooked, regardless of which dimension it originates from.

[0081] Furthermore, the control strategies adopted for the same warning level vary significantly under different operating conditions. For example, when a level 2 warning is triggered under high load conditions, the lean liquor circulation flow rate should be increased first to improve absorption capacity; while when the same level warning is triggered under normal operating conditions, it may be more appropriate to increase the heating power of the regeneration tower to improve the regeneration quality of the lean liquor. Therefore, the formulation of control strategies cannot rely solely on the single dimension of the warning level, but also needs to be differentiated based on the current operating condition of the system.

[0082] In one embodiment of this application, such as Figure 6 As shown, step S4 includes: S41, comparing the concentration values ​​at each prediction time in the sulfur dioxide concentration prediction sequence with emission multi-level thresholds and quantifying the concentration warning to obtain a concentration warning judgment vector; S42, comparing each score value in the desulfurization equipment health vector with the equipment health threshold to obtain the equipment warning level, and fusing the concentration warning judgment vector and the equipment warning level at the same time step index to obtain a graded warning signal; S43, constructing a two-dimensional retrieval combination key with the warning level in the graded warning signal and the operating condition category in the desulfurization operating condition label sequence, and performing table lookup addressing and control strategy mapping on the preset decision rule base matrix to obtain the desulfurization process control instruction set.

[0083] First, the sulfur dioxide concentration values ​​at each prediction time in the sulfur dioxide concentration prediction sequence are compared step by step with the preset multi-level emission threshold system to determine the risk range at each time in terms of emission standards, and then quantify it into the corresponding integer warning level value.

[0084] In the lead smelting process of recycling waste lead-acid batteries, flue gas emissions must meet the sulfur dioxide emission concentration limits stipulated by national environmental protection standards. To achieve graded prediction and gradient response to the risk of exceeding emission standards, the system establishes a discrete multi-level threshold hierarchy based on environmental emission standards. This threshold hierarchy divides the continuous concentration space from the safe zone to the severely exceeding zone into several discrete intervals, each interval corresponding to a different risk level.

[0085] For each predicted time point in the sulfur dioxide concentration prediction sequence, the concentration value is compared step-by-step with a multi-level threshold hierarchy using a hierarchical logical comparison. Based on its corresponding interval range, the value is quantified into a corresponding risk penalty level. The calculation formula for this quantification process is as follows: in, This represents the sulfur dioxide concentration warning level value output after step quantization at the current time step; This represents the predicted value of sulfur dioxide concentration at the current time step, derived from the sulfur dioxide concentration prediction sequence. , , These represent the preset warning and monitoring threshold constants for sulfur dioxide emission concentrations at levels one, two, and three, respectively.

[0086] A level value of 0 indicates that the predicted concentration is within a safe range and the system is operating normally without intervention; a level value of 1 indicates that the predicted concentration is close to the emission standard, posing a primary risk that requires attention; a level value of 2 indicates that the predicted concentration is approaching the emission standard, posing a medium risk that requires the initiation of control measures; and a level value of 3 indicates that the predicted concentration has reached or exceeded the emission standard limit, posing a serious risk of exceeding the standard and requiring immediate emergency control measures.

[0087] In practical applications, the above thresholds should be determined in conjunction with specific emission standards and safety margins. For example, if the applicable sulfur dioxide emission concentration limit for this waste lead-acid battery processing enterprise is 400 milligrams per cubic meter, then... It can be set to 80% of the emission standard, i.e., 320 milligrams per cubic meter. It can be set to 90% of the emission standard, i.e., 360 milligrams per cubic meter. This can be set to 100% of the emission standard, i.e., 400 mg / m³. Since the values ​​in the sulfur dioxide concentration prediction sequence may be normalized dimensionless values, the above threshold also needs to be converted to the corresponding values ​​in the normalization domain. For example, if the maximum concentration used for normalization is 50,000 mg / m³, then... The corresponding normalized value is 0.0064. Corresponding to 0.0072, This corresponds to 0.008. In actual systems, the predicted output can also be inversely normalized to a physical concentration value before being compared with the physical threshold.

[0088] The quantified risk levels of all predicted time steps are assembled sequentially to generate a concentration warning determination vector. The length of this vector is equal to the length of the predicted sequence.

[0089] Then, the continuous score values ​​in the desulfurization equipment health vector are converted into discrete equipment warning levels, and integrated with the concentration warning judgment vector under the same time step index to generate a unified graded warning signal.

[0090] The elements in the desulfurization equipment health vector range from 0 to 1 as continuous values. These values ​​need to be compared inversely with a preset equipment health threshold (i.e., the lower the score, the higher the risk) to map discrete warning levels for each equipment dimension. The calculation formula for this mapping process is as follows: in, This indicates the numerical value of the equipment anomaly risk level derived from the equipment health score at the current time step; This represents the equipment health score calculated at the current time step, which is derived from the desulfurization equipment health vector. This represents a predefined safety baseline health threshold constant; a value below this indicates that the equipment faces a serious risk of failure. This represents a predefined warning threshold constant for device performance degradation. A value below this threshold indicates that the device's health condition has begun to deteriorate and requires attention.

[0091] In practical applications, It can be set to 0.6, which means that when the device health score is below 0.6, the device performance is considered to have deteriorated significantly. A threshold of 0.3 can be set, indicating that the equipment is considered to be in a severely abnormal state when the health score is below 0.3. The specific values ​​of these thresholds can be determined by statistical backtracking analysis of the equipment health scores before historical failure events.

[0092] After mapping the warning levels at the equipment level, under the same time step index, the equipment warning level and the concentration warning level in the concentration warning judgment vector are fused using an element-wise maximum fusion operation. This fusion logic follows the worst-case principle, meaning that regardless of whether the risk of exceeding emission limits or the risk of equipment anomaly, the most severe risk level between the two dimensions is used as the final system warning level for that time step. The fusion formula is: in, This represents the final system-wide comprehensive early warning level value established after merging the largest values, and constitutes the discrete elements of the graded early warning signal; This represents a mathematical operation function that takes the maximum value between two options.

[0093] This fusion mechanism ensures that when a high-risk signal appears in any dimension, the overall system warning level will be raised to the corresponding high level, preventing temporary safety in one dimension from masking risk signals already appearing in another. For example, when the predicted concentration is still within the safe range (concentration warning level 0) but the equipment health score has dropped to 0.25 (equipment warning level 3), the overall system warning level will be set to level 3, indicating that immediate measures must be taken to address the equipment anomaly and prevent further deterioration leading to uncontrolled emissions.

[0094] The comprehensive early warning levels of all time steps are assembled sequentially to generate a tiered early warning signal. In subsequent processing, the highest level value or the latest time value in the tiered early warning signal is taken as the effective early warning level for the current decision-making cycle.

[0095] Finally, the current warning level and the system's operating condition are used together to form the search conditions. The system then performs a lookup in the pre-set decision rule base to map out specific process parameter adjustment strategies that match the current risk situation and operating condition, generating a set of desulfurization process control instructions that can be parsed by the actuators.

[0096] The decision rule base is a static mapping table pre-compiled by experts in the desulfurization process field based on operational experience and process characteristics before system deployment. In the ion-liquid desulfurization process for lead smelting in the recycling of waste lead-acid batteries, the control of sulfur dioxide concentration in flue gas is mainly achieved by adjusting the circulation flow rate of the lean ion-liquid to change the absorption capacity, adjusting the heating power of the regeneration tower to improve the regeneration quality of the lean liquid, and controlling the distribution of flue gas flow into the desulfurization system by adjusting the opening of the flue gas regulating valve. The required control intensity varies under different warning levels, and the priority and adjustment direction of each control parameter also differ under different operating conditions. This prior knowledge in the field is systematically encoded and stored in the decision rule base matrix.

[0097] The system concatenates the warning level values ​​from the graded early warning signals with the current operating status values ​​from the desulfurization operating condition label sequence to construct a two-dimensional retrieval key. This key is then used to perform a lookup operation in the decision rule base matrix, uniquely mapping an action vector containing the target adjustment amounts of each physical control parameter. This lookup mapping process is expressed as follows: in, This represents the multidimensional physical control increment vector generated after addressing and mapping, which is the core content of the desulfurization process control instruction set; This indicates the specific increase or decrease in the flow rate of the lean ionized liquid required by the instruction; This indicates the specific increase or decrease in the heating power of the regeneration tower required by the instruction. This indicates the specific compensation amount required by the instruction to adjust the opening of the flue gas regulating valve; This represents the static decision rule base retrieval mapping function built into the storage medium; This indicates the overall warning level value used as the first search key; The label representing the current desulfurization operating condition status is derived from the desulfurization operating condition label sequence and serves as the second search key.

[0098] In practical applications, the specific content of the decision rule base can be compiled with reference to the following example. When the warning level is 1 (basic risk) and the operating condition category is 1 (normal operating condition), the lean liquor flow rate increases by 5%, the regeneration tower heating power increases by 2%, and the flue gas regulating valve opening remains unchanged; when the warning level is 2 (medium risk) and the operating condition category is 2 (high load condition), the lean liquor flow rate increases by 15%, the regeneration tower heating power increases by 8%, and the flue gas regulating valve opening decreases by 5% to reduce the flue gas load entering the desulfurization system; when the warning level is 3 (severe risk) and the operating condition category is 2 (high load condition), the lean liquor flow rate increases by 25%, the regeneration tower heating power increases by 15%, and the flue gas regulating valve opening decreases by 10%; when the warning level is 0 (safe state), regardless of the operating condition category, the adjustment amount of each parameter is 0, and the system maintains its current operating state.

[0099] In the ion-liquid desulfurization process for waste lead-acid batteries, the adjustment of the lean liquid flow rate directly affects the liquid-to-gas ratio in the gas-liquid phase contact within the absorber. The lean liquid flow rate is initially controlled at around 60% and gradually increases with the rise in the temperature at the bottom of the regeneration tower and the increase in load. Therefore, the adjustment range of the lean liquid flow rate in the rule library must consider the current flow rate position and the allowable range of pump frequency. The adjustment of the regeneration tower heating power affects the desorption efficiency of sulfur dioxide in the rich liquid. After passing through the packing in the regeneration tower, the rich liquid enters the reboiler for further heating and regeneration. The increase or decrease in heating power must be set within the allowable heat load range of the equipment. The adjustment of the flue gas regulating valve opening affects the distribution ratio of flue gas entering the desulfurization system. Its adjustment range must ensure that the sulfur dioxide content in the flue gas of the dry absorption section is maintained within the design range of 4% to 6%.

[0100] The target adjustment vector obtained from the lookup table is structured and packaged to generate a desulfurization process control instruction set. This instruction set not only contains the adjustment range values ​​of each parameter, but also carries the corresponding parameter identifiers and target actuator address information, so that the downstream step S5 can parse and execute it step by step.

[0101] S5 determines the control execution rate based on the graded early warning signals and issues the desulfurization process control command set in stages. It should be understood that in the ion-liquid desulfurization process for lead smelting from waste lead-acid batteries, the execution of control commands involves the actual actions of physical actuators such as the frequency of the lean liquor circulation pump, the power of the regeneration tower heater, and the opening degree of the flue gas electric regulating valve. The changes in process parameters driven by these actuators have inertia and coupling effects. If the overall control increment generated in step S4 is issued to the actuators all at once, it may cause severe disturbances to the process system. Specifically, a sudden increase in the lean liquor flow rate will cause a sharp fluctuation in the liquid level in the absorber, potentially leading to flooding or liquid sealing; a sudden increase in the heating power of the regeneration tower will cause a rapid rise in the tower bottom temperature, potentially leading to localized overheating and decomposition of the ion-liquid; and a sudden and significant change in the opening degree of the flue gas regulating valve will cause a sudden change in the pipeline system pressure, affecting the normal operation of the upstream smelting furnace.

[0102] On the other hand, when the system faces a serious risk of exceeding standards (warning level 3), if process parameters are adjusted too slowly, sulfur dioxide emissions may exceed the standard limit before the adjustment is completed, resulting in actual emission violations. Therefore, the rate of control execution must be commensurate with the urgency indicated by the warning level: the higher the warning level, the faster the execution rate, to ensure that the control action is completed within the limited response time window; when the warning level is low, a slower, gradual adjustment should be adopted to avoid unnecessary disturbances to the steady-state operation of the process system.

[0103] In one embodiment of this application, step S5 includes: S51, determining the control execution time step as the control execution rate based on the warning level in the graded warning signal; S52, equally discretizing and timing-packing the overall control increment in the desulfurization process control instruction set to obtain a step-by-step execution instruction sequence; S53, issuing the step-by-step execution instruction sequence.

[0104] First, the discrete warning level values ​​carried in the graded warning signals are converted into control execution time step parameters with continuous time dimensions, and a quantitative mapping relationship between the urgency of the warning and the execution rhythm is established.

[0105] The system extracts the warning level values ​​from the tiered early warning signals and converts these discrete level values ​​into corresponding control execution time steps using an exponentially decaying mapping function. The characteristic of this mapping function is that as the warning level value increases, the output time step decreases exponentially, resulting in a sharp increase in execution rate under high-risk conditions and a relatively relaxed execution pace under low-risk conditions. The calculation formula is as follows: in, This indicates the time step for implementing control measures, dynamically calculated based on the current early warning situation. This represents the baseline execution time step constant set by the system under normal, risk-free operating conditions. This represents the preset warning level index decay sensitivity coefficient, used to adjust the sensitivity of the execution rate to the risk level; Represents the base of the natural logarithm; This indicates the current overall warning level value carried in the graded warning signal.

[0106] In practical applications, The value of should be determined based on the maximum allowable transition time for adjusting the process parameters of the desulfurization system. Considering that the typical response time for a significant change in absorption efficiency from flow rate adjustment in the lean liquor circulation of the ion-liquid desulfurization process is between 30 and 120 seconds, It can be set from 10 seconds to 30 seconds. For example, take... = 20 seconds means that a step size is issued every 20 seconds in a risk-free state.

[0107] The value of determines the steepness of the increase in execution rate as the warning level rises. The larger the value, the more drastic the shortening of the time step under the high warning level. In practical applications, The value can be between 0.5 and 1.5. For example, take... When = 1.0, the warning level is 0 (safe state). Seconds; when the warning level is 1 (basic risk). Seconds; when the warning level is 2 (medium risk). Seconds; when the warning level is 3 (severe risk). Seconds. This shows that under severe risk conditions, the system's command issuance interval is shortened to approximately 1 second, achieving rapid response to emergency situations.

[0108] Finally, based on the control execution time step determined in the previous sub-step and the system's preset target response total time window, the total physical control increment contained in the desulfurization process control instruction set is calculated to determine the required number of steps. The total increment is then equally divided into a corresponding number of small step increments and packaged in chronological order to form a step-by-step execution instruction sequence.

[0109] First, based on the time step of the regulation implementation The total response time window length expected to complete this control action as preset by the system. Calculate how many steps to divide the total increment into. The formula for calculating the number of steps is: in, This represents the total number of steps calculated. This indicates the total response time window length that the system is expected to complete this control action. This indicates the time step for the control implementation determined in the previous sub-step; This represents the floor function, used to ensure that the number of splitting steps is a valid positive integer.

[0110] In practical applications, The value should be determined based on the reasonable transition time required for the desulfurization system to fully stabilize at the new setpoints from startup and adjustment. Considering the typical time constant for the lean liquor in the ion-liquid desulfurization process from flow pump frequency adjustment to the re-establishment of gas-liquid balance in the absorber, It can be set from 60 seconds to 180 seconds. For example, take... Seconds. When the warning level is 1, seconds, then Step; when the warning level is 3, seconds, then step.

[0111] Subsequently, the overall physical control increment vector contained in the desulfurization process control instruction set is divided by the calculated number of steps to obtain the single-step control amount for each execution step. The formula for calculating this step amount is: in, This represents the action step vector issued within a single execution time step after discretization and segmentation. This represents the overall physical control increment vector contained in the desulfurization process control instruction set.

[0112] In the example above, if the overall control increment is a 15% increase in lean liquor flow rate, an 8% increase in regeneration tower heating power, and a 5% decrease in flue gas regulating valve opening, and the warning level is 1 (17 steps), then each step increases the lean liquor flow rate by approximately 0.88%, the regeneration tower heating power by approximately 0.47%, and the flue gas regulating valve opening by approximately 0.29%. When the warning level is 3 (120 steps), although the single-step control increment is smaller (the lean liquor flow rate increases by approximately 0.125% per step), the overall completion time remains 120 seconds because the step interval is only 1 second, and the overall response speed is the same as under the low warning level. It maintains consistency while making the execution process smoother and more refined.

[0113] Each step quantity is packaged and grouped in chronological order, with each step quantity accompanied by its execution time information (i.e., the nth step quantity should be executed after the start time). (issuing instructions at specific times), thus forming a step-by-step execution instruction sequence with strict timing constraints.

[0114] Finally, the generated step-by-step execution instruction sequence is sent to each physical actuator of the desulfurization system according to the timing constraints marked in its internal label, so that the process parameters transition from the current operating value to the control target value in a uniform and gradual manner.

[0115] The system sequentially sends the corresponding step quantity commands to each actuator in the flue gas treatment system according to the execution time marked in the step-by-step execution command sequence. In the ion-liquid desulfurization process for lead smelting in the recycling of waste lead-acid batteries, these actuators specifically include: a lean liquid flow regulating valve, used to control the circulating flow rate of the lean ion-liquid entering the absorption tower; a regeneration tower heater, used to control the heating power of the rich liquid regeneration process to affect the quality of lean liquid regeneration; and a flue gas electric regulating valve, used to control the distribution ratio of smelting flue gas flow entering the desulfurization system. After receiving the step quantity command, each actuator adds the current actual control quantity to the step increment, performs a physical action to move the process parameters to the new intermediate value, and then waits for the arrival of the next step quantity command.

[0116] In the actual operation of the ion-liquid desulfurization process, the lean liquor flow rate is adjusted by changing the frequency of the frequency converter of the lean liquor circulation pump. The initial lean liquor flow rate is controlled at around 60%, and gradually increases as the temperature at the bottom of the regeneration tower rises and the load increases. The step-by-step distribution mechanism allows the pump frequency to gradually increase to the target value in a small and uniform manner, avoiding mechanical shock to the pump body and water hammer effect in the pipeline caused by sudden acceleration. The heating power of the regeneration tower is adjusted by controlling the output power of the electric heating element of the reboiler. After passing through the packing in the regeneration tower, the rich liquor enters the reboiler for further heating and regeneration into lean liquor. The gradual increase in heating power ensures a stable rise in the bottom temperature of the tower without local overheating. The adjustment of the flue gas regulating valve opening affects the distribution ratio of flue gas entering the dry absorption section and the desulfurization section. The flue gas flow rate is controlled to maintain the sulfur dioxide content in the flue gas entering the dry absorption section at 4% to 6%. Gradual adjustment ensures that this distribution ratio does not change drastically in an instant.

[0117] Through the aforementioned step-by-step distribution mechanism, each actuator performs only a small step within each time interval, allowing the process parameters to transition smoothly and gradually from their current value to the target value throughout the entire control process. When the warning level is high, the step interval is short, and each step is distributed intensively within a short period of time, enabling the process parameters to quickly converge to the target value to cope with emergency situations. When the warning level is low, the step interval is long, and the process parameters transition slowly to the target value, resulting in a lower degree of disturbance to the steady-state operation of the system.

[0118] In one embodiment of this application, the method further includes: continuously reading the latest actual response parameters in the sulfur dioxide flue gas time-series monitoring data after issuing the initial step amount of the step-by-step execution instruction sequence; performing vector subtraction and column stacking on the actual response state vector and the expected target state vector to obtain the control deviation compensation matrix; performing weighted summation and compensation increment derivation on the control deviation compensation matrix to obtain the compensation increment, and superimposing the compensation increment with the current step amount in the step-by-step execution instruction sequence to obtain the control signal set of the desulfurization actuator.

[0119] First, after the initial step size of the step-by-step execution command sequence has been issued to the actuator and generated physical action, the latest actual measurement data is continuously obtained from the sulfur dioxide flue gas time-series monitoring dataset as a real-time feedback signal source for evaluating the control effect.

[0120] After the first step of the step-by-step execution command sequence is issued to the lean liquid flow regulating valve, the regeneration tower heater, and the flue gas electric regulating valve, each actuator begins to perform physical actions, and the process parameters change accordingly. The system continuously reads the latest actual monitoring data from the sulfur dioxide flue gas time-series monitoring dataset using the same sampling period as the global synchronization time axis in step S1. Since the sulfur dioxide flue gas time-series monitoring dataset is continuously updated in step S1 through continuous acquisition and clock synchronization, measurement data reflecting the latest state of the system after the actuator actions can be obtained at any time during the execution of this step.

[0121] Specifically, the system reads the measured values ​​of sulfur dioxide concentration at the absorber inlet, internal temperature of the absorber, smelting flue gas flow rate, lean ion liquid flow rate, and bottom temperature of the regeneration tower from the latest row of the sulfur dioxide flue gas time-series monitoring dataset. It then assembles these multi-channel measured values ​​into the actual response state vector for the current moment. The dimension of this vector is equal to the number of monitoring channels, which is 5 dimensions in this implementation.

[0122] Then, the actual response state vector read at each time point is compared with the target state vector that the system expects each monitored parameter to reach at that time point, the deviation between the two is calculated, and the deviation vectors at multiple times are stacked in the time dimension to form a control deviation compensation matrix.

[0123] The expected target state vector of the system This represents the theoretically ideal value that each process parameter should achieve under the current cumulative step increment. This target value can be determined by superimposing the baseline values ​​of each parameter at the start of the control process with the cumulative step increments. For example, if the measured lean liquid flow rate before control is initiated is 60%, and the step-by-step execution command sequence has issued 3 steps with a lean liquid flow rate increment of 0.88% per step, then the target value for the lean liquid flow rate at the current moment is... .

[0124] The actual response state vector is subtracted from the expected target state vector to obtain the actual control error of each monitoring parameter channel at the current moment. The formula for calculating this deviation is: in, This represents the actual control error vector of each channel calculated at time t; This represents the ideal target state vector that the system expects each monitored parameter to reach at time t. This represents the latest actual physical state vector at time t extracted from the sulfur dioxide flue gas time-series monitoring dataset.

[0125] The sign of each element in the error vector carries directional information: a positive value indicates that the actual parameter has not yet reached the target value (insufficient response), and a negative value indicates that the actual parameter has exceeded the target value (overshoot).

[0126] The error vectors from each acquisition point from the start of the control process to the current time are stacked column-wise to form a control deviation compensation matrix. The number of rows in this matrix equals the number of monitoring parameter channels (i.e., 5), and the number of columns equals the number of time steps in which deviation acquisition has been completed. As the control process progresses, the number of columns continuously increases. The construction of the control deviation compensation matrix enables the system to not only grasp the instantaneous deviation at the current moment but also retain the historical trajectory information of the deviation's evolution over time, providing a global view in the time dimension for subsequent dynamic compensation simulations.

[0127] Finally, the time evolution law of the error vector is extracted from the control deviation compensation matrix, and the dynamic compensation control increment required at the current time is derived using the proportional-integral-derivative control algorithm. This increment is then superimposed and corrected with the basic step instruction at the current moment in the step-by-step execution instruction sequence. After hardware safety limiting verification, the final control signal set of the desulfurization actuator is generated.

[0128] The system extracts the current error vector and its historical evolution sequence from the control deviation compensation matrix and inputs it into a preset proportional-integral-derivative closed-loop control algorithm module for compensation increment derivation. The proportional term generates an instantaneous correction proportional to the current instantaneous error; the integral term sums the cumulative error from the start of control to the current moment, eliminating the steady-state deviation; and the derivative term generates a predictive correction based on the rate of change of error between two adjacent moments, suppressing the continued increase in error. The weighted sum of these three terms yields the required dynamic compensation control increment. The calculation formula is as follows: in, This represents the dynamic compensation control increment vector that needs to be superimposed at time t, calculated using the proportional-integral-differential algorithm. This represents the proportional gain coefficient, which determines the instantaneous response magnitude of the compensation amount to the current error. This represents the integral gain coefficient, which determines the effectiveness of the compensation in eliminating accumulated errors. The differential gain coefficient determines the strength of the compensation amount in suppressing the rate of change of error. This represents the vector of the cumulative historical control error integrals from the start of the control process to the current time t. It represents the difference between the error at the current time and the error at the previous time, and represents the rate of change of the error.

[0129] In practical applications, , , The values ​​of the three coefficients need to be adjusted according to the dynamic response characteristics of each execution loop in the desulfurization system. For the lean liquor flow control loop, since the flow regulating valve responds quickly but has a certain pipeline inertia, The possible values ​​are 0.3 to 0.8. The possible values ​​are 0.01 to 0.05. The value can be between 0.05 and 0.2. For the heating power control loop of the regeneration tower, due to the large thermal inertia and pure time lag in the temperature response, A smaller value (e.g., 0.1 to 0.3) should be used to avoid overshoot. A value of 0.005 to 0.02 can be used to slowly eliminate steady-state error. A value of 0.1 to 0.5 can be used to suppress temperature overshoot. For the flue gas regulating valve opening control loop, The possible values ​​are 0.2 to 0.5. The possible values ​​are 0.01 to 0.03. The values ​​can range from 0.05 to 0.15. These coefficients can be determined by combining step response tests during the system commissioning phase with engineering tuning methods (such as the critical proportional method or the attenuation curve method).

[0130] The dynamically compensated control increment obtained from the derivation is superimposed and synthesized with the base step size at the current moment in the step-by-step execution instruction sequence and the actual control reference size of the actuator at the previous moment. Then, it undergoes saturation verification using upper and lower limit hardware safety limiting functions to prevent the calculated control quantity from exceeding the allowable physical operation extreme range of each actuator. The formula for generating the final control signal is: in, This represents the absolute physical control vector that is finally encapsulated after superposition correction and saturation verification, which is the core data of the desulfurization actuator control signal set; This represents the current actual physical control reference quantity of the actuator at the previous time t-1; This indicates the basic action step size issued at this moment in the step-by-step execution instruction sequence; This represents the dynamic compensation increment derived using the proportional-integral-differential algorithm; This represents the upper and lower limit amplitude functions, used to prevent the calculated control quantity from exceeding the safe physical limits allowed by the hardware; This represents the minimum operational bound vector allowed by each underlying physical actuator; This represents the maximum allowed operational limit vector for each underlying physical actuator.

[0131] In practical applications, the safe operating range of each actuator can be set as follows. For lean liquid flow regulating valves, The corresponding minimum allowable flow rate is 30% of the design flow rate (too low a flow rate will cause the liquid film inside the absorber to break and the gas to short-circuit). The corresponding maximum allowable flow rate is 100% of the design flow rate (exceeding this will cause pump overload). Regarding the heating power of the regeneration tower, The corresponding minimum heating power is the minimum heat load required to maintain the non-condensable liquid in the regeneration tower. The corresponding maximum heating power is the rated power of the reboiler's electric heating element. For the flue gas regulating valve, The corresponding minimum opening is the minimum required to ensure that negative pressure backflow does not occur within the pipeline. The corresponding maximum opening is the fully open position of the valve.

[0132] After amplitude limiting and verification, the absolute control quantity is encapsulated according to the underlying communication protocol and output as a set of control signals for desulfurization actuators that directly drive the actions of field hardware. This set of control signals includes control signals for lean liquor flow valves, regeneration tower heating power, and flue gas regulating valve opening. This set of control signals is sent to the local controllers of each actuator via the industrial fieldbus to drive the physical equipment to perform actions.

[0133] The aforementioned closed-loop feedback compensation process is repeated at each step of the step-by-step instruction sequence. That is, before issuing each new step, the latest actual response data is read, the control deviation compensation matrix is ​​updated, the compensation increment is recalculated, and the final control signal is generated by superimposing these data. As the control process progresses, if the errors of each channel continue to converge to near zero, the outputs of the integral and proportional terms will approach zero, the compensation increment will no longer generate additional corrections, and the system will enter a steady-state maintenance phase. If the error of a certain channel fails to converge or shows a diverging trend after multiple step cycles, the continuous accumulation of the integral term and the rate of change response of the derivative term will automatically increase the compensation intensity, driving the system to converge towards the target value.

[0134] With the introduction of the aforementioned closed-loop feedback compensation mechanism, the step-by-step distribution process is upgraded from a purely open-loop feedforward execution to a closed-loop feedforward-feedback composite control. The basic step size provides feedforward guidance for the control direction and basic intensity, while the dynamic compensation increment provides feedback correction based on real-time deviations. The combination of these two ensures that, even under process disturbances and non-ideal actuator responses, each process parameter can still accurately converge to the expected target value, achieving precise closed-loop control of sulfur dioxide emissions during the waste lead-acid battery processing.

[0135] In summary, the intelligent monitoring method for sulfur dioxide flue gas in the waste lead-acid battery treatment process of this application has been clarified. Firstly, it synchronizes and merges the clocks of each sensor channel using a unified time reference, ensuring that subsequent analysis is based on high-quality data with strict time alignment. Secondly, it introduces a sliding window mechanism to extract multi-dimensional statistical features and uses a combination of concentration and temperature ranges to identify and label the operating conditions of each window, enabling the system to have real-time perception capabilities of the desulfurization operation status. Thirdly, it fuses and encodes the operating condition labels and feature matrices, then feeds them into a time-series prediction model and anomaly detection model for parallel inference, simultaneously obtaining quantitative assessment results of future concentration evolution trends and equipment health status. Finally, it jointly classifies the concentration prediction and equipment health status to generate graded early warnings, and maps the corresponding process control strategies in the decision rule base based on the current operating condition category. The execution rhythm is dynamically determined according to the urgency of the early warning, and the strategies are implemented in stages, thereby achieving closed-loop intelligent monitoring driven by prediction, adaptive to operating conditions, and precisely matched with control intensity.

[0136] Figure 7 This is a schematic block diagram of the intelligent monitoring system for sulfur dioxide flue gas during the waste lead-acid battery treatment process described in this application. Figure 7As shown, the intelligent monitoring system 700 for sulfur dioxide flue gas in the waste lead-acid battery treatment process according to an embodiment of this application includes: a data synchronization module 710, used to synchronize and fuse multi-channel sensor data collected through a unified time base to obtain a sulfur dioxide flue gas time-series monitoring dataset; a feature annotation module 720, used to preprocess and extract window features from the sulfur dioxide flue gas time-series monitoring dataset to obtain a multi-dimensional feature matrix for flue gas desulfurization, and to annotate each window with operating conditions according to the combination of concentration range and temperature range to obtain a desulfurization operating condition label sequence; and a prediction and diagnosis module 730, used to... The multidimensional feature matrix of flue gas desulfurization is spliced ​​with the desulfurization operating condition label sequence and time-series reasoning prediction to obtain the sulfur dioxide concentration prediction sequence. At the same time, the anomaly degree of each window is evaluated to obtain the health vector of the desulfurization equipment. The decision mapping module 740 is used to jointly classify the sulfur dioxide concentration prediction sequence and the health vector of the desulfurization equipment to obtain the graded early warning signal. It also combines the desulfurization operating condition label sequence to query the decision rule base to perform control strategy mapping to obtain the desulfurization process control instruction set. The instruction issuance module 750 is used to determine the control execution rate according to the graded early warning signal and issue the desulfurization process control instruction set in steps.

[0137] It should be noted that the intelligent monitoring system for sulfur dioxide flue gas in the waste lead-acid battery processing process of this application embodiment is similar in principle to the aforementioned intelligent monitoring method for sulfur dioxide flue gas in the waste lead-acid battery processing process. Therefore, the implementation process, implementation principle, and beneficial effects of the intelligent monitoring system for sulfur dioxide flue gas in the waste lead-acid battery processing process can be found in the description of the implementation process, implementation principle, and beneficial effects of the aforementioned method, and will not be repeated.

[0138] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for intelligent monitoring of sulfur dioxide flue gas during the treatment of waste lead-acid batteries, characterized in that, include: S1, clock synchronization and fusion splicing of the collected multi-channel sensor data are performed using a unified time reference to obtain a sulfur dioxide flue gas time-series monitoring dataset; S2, preprocess and extract window features from the sulfur dioxide flue gas time series monitoring dataset to obtain the flue gas desulfurization multidimensional feature matrix, and label each window with operating conditions according to the combination of concentration range and temperature range to obtain the desulfurization operating condition label sequence. S3 involves concatenating the multidimensional feature matrix of flue gas desulfurization with the desulfurization operating condition label sequence and performing temporal inference prediction to obtain the sulfur dioxide concentration prediction sequence. Simultaneously, anomaly assessment is performed on each window to obtain the desulfurization equipment health vector. This includes: converting each discrete state label in the desulfurization operating condition label sequence into an equal-length sparse numerical vector using one-hot encoding to obtain the encoded operating condition numerical vector; concatenating and fusing the multidimensional feature matrix of flue gas desulfurization with the encoded operating condition numerical vector using column-dimensional concatenation to obtain the fused inference input tensor; using a pre-trained long short-term memory network model to perform temporal dynamic expansion and forward inference prediction on the fused inference input tensor to obtain the sulfur dioxide concentration prediction sequence; and using a pre-trained isolated forest algorithm model to perform anomaly inference and reverse health mapping on the fused inference input tensor to obtain the desulfurization equipment health vector. S4. The sulfur dioxide concentration prediction sequence and the desulfurization equipment health vector are jointly classified to obtain a graded early warning signal. The desulfurization operating condition label sequence is then combined with the decision rule base for querying to obtain a control strategy mapping to obtain a desulfurization process control instruction set. This includes: comparing the concentration values ​​at each prediction time in the sulfur dioxide concentration prediction sequence with emission multi-level thresholds and quantifying the concentration early warning to obtain a concentration early warning judgment vector; comparing each score value in the desulfurization equipment health vector with the equipment health threshold to obtain the equipment early warning level; and fusing the concentration early warning judgment vector and the equipment early warning level at the same time step index by taking the maximum value element-wise to obtain a graded early warning signal; constructing a two-dimensional retrieval key using the early warning level in the graded early warning signal and the operating condition category in the desulfurization operating condition label sequence, and performing table lookup and control strategy mapping on the preset decision rule base matrix to obtain a desulfurization process control instruction set. S5 determines the control execution rate based on the graded early warning signals and issues the desulfurization process control instruction set in stages.

2. The intelligent monitoring method for sulfur dioxide flue gas in the waste lead-acid battery treatment process according to claim 1, characterized in that, The multi-channel sensor data includes the sulfur dioxide concentration signal at the inlet of the absorption tower, the internal temperature signal of the absorption tower, the flow rate signal of the smelting flue gas, the flow rate signal of the lean ion liquid, and the temperature signal at the bottom of the regeneration tower.

3. The intelligent monitoring method for sulfur dioxide flue gas in the waste lead-acid battery treatment process according to claim 2, characterized in that, Step S1 includes: Based on the pre-calibrated network communication transmission delay and the inherent hardware delay of analog-to-digital conversion, the sensing delay of each channel sampling point in the multi-channel sensing data is calculated and timestamp compensation is performed to obtain a sensing data stream with compensated timestamps. Global time base mapping and resampling are performed on the sensor data stream with compensated timestamps to obtain a synchronized data matrix; The resampled values ​​at each time step in the synchronized data matrix are horizontally feature-stitched and structured encapsulated and recombined to obtain the sulfur dioxide flue gas time-series monitoring dataset.

4. The intelligent monitoring method for sulfur dioxide flue gas in the waste lead-acid battery treatment process according to claim 1, characterized in that, Step S2 includes: Data normalization and sliding window time-series slicing were performed on the sulfur dioxide flue gas time-series monitoring dataset to obtain the normalized window segment tensor; Window statistical domain features are calculated and dimension reduction aggregated on the normalized window segment tensor to obtain the multidimensional feature matrix of flue gas desulfurization. Based on the multi-dimensional combination judgment conditions of the preset sulfur dioxide concentration range and the absorption tower temperature range, the mean features of each window in the flue gas desulfurization multi-dimensional feature matrix are compared in a step-by-step logical manner and labeled with the operating condition category to obtain the desulfurization operating condition label sequence.

5. The intelligent monitoring method for sulfur dioxide flue gas in the waste lead-acid battery treatment process according to claim 1, characterized in that, Step S5 includes: Based on the warning level in the graded warning signal, the time step of the control execution is determined as the control execution rate; The overall control increment in the desulfurization process control instruction set is equally discretized and time-sequentially packaged to obtain a step-by-step execution instruction sequence; Issue a sequence of instructions for step-by-step execution.

6. The intelligent monitoring method for sulfur dioxide flue gas in the waste lead-acid battery treatment process according to claim 5, characterized in that, Also includes: After issuing the initial step size of the step-by-step execution instruction sequence, continuously read the latest actual response parameters in the sulfur dioxide flue gas time-series monitoring dataset; The actual response state vector and the expected target state vector are subtracted and stacked in columns to obtain the control deviation compensation matrix. The compensation increment is obtained by weighted summation and compensation increment derivation of the control deviation compensation matrix, and the compensation increment is superimposed with the current step amount in the step-by-step execution instruction sequence to obtain the control signal set of the desulfurization actuator.

7. The intelligent monitoring method for sulfur dioxide flue gas in the waste lead-acid battery treatment process according to claim 3, characterized in that, Based on pre-calibrated network communication transmission delay and inherent hardware delay of analog-to-digital conversion, sensing delay calculation and timestamp compensation are performed on the sampling points of each channel in the multi-channel sensing data to obtain a sensing data stream with compensated timestamps, including: Communication delay drift trend extraction and smoothing modeling are performed on multi-channel sensor data to obtain a time-varying delay drift trend sequence; Residual bias anchoring calibration and superposition correction are performed on the time-varying delay drift trend sequence to obtain the dynamic delay compensation sequence after anchoring calibration; Based on the adaptive compensation value at the corresponding moment in the dynamic delay compensation sequence after anchoring calibration and the inherent hardware delay of analog-to-digital conversion, the system arrival time of each sampling point in the multi-channel sensor data is traced back and bound to the timestamp key value to obtain a sensor data stream with compensation timestamp.

8. An intelligent monitoring system for sulfur dioxide flue gas during the treatment of waste lead-acid batteries, characterized in that, include: The data synchronization module is used to synchronize and fuse multi-channel sensor data collected by a unified time reference to obtain a sulfur dioxide flue gas time-series monitoring dataset. The feature annotation module is used to preprocess and extract window features from the sulfur dioxide flue gas time series monitoring dataset to obtain a multi-dimensional feature matrix for flue gas desulfurization, and to annotate each window with operating conditions based on the combination of concentration range and temperature range to obtain a desulfurization operating condition label sequence. The prediction and diagnosis module is used to concatenate and perform temporal inference prediction on the multidimensional feature matrix of flue gas desulfurization and the desulfurization operating condition label sequence to obtain the sulfur dioxide concentration prediction sequence. At the same time, it performs anomaly assessment on each window to obtain the desulfurization equipment health vector. This includes: converting each discrete state label in the desulfurization operating condition label sequence into an equal-length sparse numerical vector through one-hot encoding to obtain the encoded operating condition numerical vector; concatenating and fusing the multidimensional feature matrix of flue gas desulfurization and the encoded operating condition numerical vector through column-dimensional concatenation to obtain the fused inference input tensor; using a pre-trained long short-term memory network model to perform temporal dynamic expansion and forward inference prediction on the fused inference input tensor to obtain the sulfur dioxide concentration prediction sequence; and using a pre-trained isolated forest algorithm model to perform anomaly inference and reverse health mapping on the fused inference input tensor to obtain the desulfurization equipment health vector. The decision mapping module is used to jointly classify the sulfur dioxide concentration prediction sequence and the desulfurization equipment health vector to obtain a graded early warning signal. It also combines the desulfurization operating condition label sequence with a decision rule base to perform control strategy mapping to obtain a desulfurization process control instruction set. This includes: comparing the concentration values ​​at each prediction time in the sulfur dioxide concentration prediction sequence with emission multi-level thresholds and quantifying the concentration early warning to obtain a concentration early warning judgment vector; comparing each score value in the desulfurization equipment health vector with the equipment health threshold to obtain the equipment early warning level; fusing the concentration early warning judgment vector and the equipment early warning level at the same time step index using element-wise maximum fusion to obtain a graded early warning signal; constructing a two-dimensional retrieval key using the early warning level in the graded early warning signal and the operating condition category in the desulfurization operating condition label sequence; and performing table lookup and control strategy mapping on a pre-set decision rule base matrix to obtain a desulfurization process control instruction set. The instruction issuance module is used to determine the control execution rate based on the graded early warning signals and to issue the desulfurization process control instruction set in stages.

Images