A method for optimizing a process of selective lithium extraction and nickel-cobalt-manganese co-extraction from ternary lithium black powder

Abstract

The present application relates to a kind of ternary lithium black powder selective lithium extraction and nickel cobalt manganese co-extraction process optimization method.For the strong dynamic characteristics of process variables, it is difficult to accurately identify and control the stability of sudden abnormalities, etc. Through graph convolution network, non-parametric process state semantic fingerprint is generated, and combined with perturbation response, selective critical zone identification and dynamic stability determination are carried out. The system can intelligently switch the pre-calibration control parameters when the state crosses the cluster, and use the time sequence knowledge graph to realize the cause and effect tracing of process abnormalities. This method improves the data-driven intelligent perception and steady-state control capability of hydrometallurgical process, and helps to timely warn and optimize the process under abnormal conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for hydrometallurgical processes, and in particular to an optimization method for selective lithium extraction from ternary lithium black powder and nickel-cobalt-manganese co-extraction process. Background Technology

[0002] With the widespread application of ternary lithium black powder in battery recycling and resource reuse, the process stability control of its hydrometallurgical lithium extraction and nickel-cobalt-manganese co-extraction processes has gradually become a key focus of the industry. Currently, mainstream control strategies in the industry mostly rely on setpoint or range control methods based on traditional process variables (such as pH, temperature, flow rate, etc.), manually or semi-automatically adjusting the process by setting thresholds and rules. These methods typically use preset upper and lower limits, empirical judgment logic, etc., as the basis for process alarms and control switching. In practice, they mostly use conventional PID parameter groups, rule bases, and simplified statistical analysis based on historical process data.

[0003] Existing technologies typically employ the following methods for process stability control: First, they rely heavily on the accumulated experience of operators, with the judgment of critical conditions and abnormal events primarily achieved through experience-based threshold setting and manual observation. Second, in terms of parameter setting, they heavily depend on the absolute value or range of a single variable, lacking the ability to characterize the dynamic correlation between operating conditions and complex transition mechanisms. Third, when dealing with sudden process changes caused by factors such as fluctuations in raw material composition, changes in interfacial mass transfer kinetics, and equipment deterioration, the adjustment response is lagging, easily leading to problems such as system oscillations and product quality fluctuations.

[0004] Currently, the industry has explored the introduction of data-driven modeling methods and intelligent control algorithms, such as using multiple regression, support vector machines, or shallow neural networks to predict and regulate some process parameters. However, most of these methods are limited to modeling shallow relationships between observable variables and heavily rely on massive labeled samples, lacking in-depth semantic characterization of complex physicochemical phenomena such as multi-metal competitive complexation, interface relaxation, and nonlinear transitions in hydrometallurgy. Furthermore, with the diversification of raw materials for lithium battery material recycling, significant batch differences in feed, and the combined effects of instrument drift and extractant aging caused by long-term operation, fixed threshold and static parameter group control exhibit significant inadequacy in actual production: the control bandwidth cannot be dynamically adjusted, making it difficult to identify critical operating conditions in a timely manner, leading to frequent stability risk events.

[0005] In the synergistic process of lithium extraction and co-extraction from ternary lithium black powder, the process state is affected by multiple factors, including the dynamic concentration of various metal ions, pH jumps, and interfacial rheology, leading to frequent transient changes and critical boundary phenomena. Traditional manually set operating condition segmentation, threshold determination, or historical experience-based PID parameter scheduling modes suffer from the following prominent problems when facing strong nonlinearity, multi-condition interaction, and variable drift: The lack of sophisticated identification at the process semantic level and dynamic characterization methods under multiple operating conditions makes it difficult to integrate underlying physicochemical mechanisms such as interface mass transfer lag and complex metal ion complexation with process control strategies.

[0006] The control parameters are set in a highly static manner, making it impossible to automatically adjust according to real-time operating conditions. This results in the system being either too loose or too tight when it deviates from typical operating conditions, making it difficult to balance production efficiency and process stability.

[0007] Relying too much on historical experience and static alarms based on a single variable makes it impossible to intelligently identify abnormal operating conditions and critical stability zones based on multi-source data. The system has poor self-adaptability and is sensitive to disturbances such as sudden changes in raw materials, aging of extractants, and instrument hysteresis, which can easily lead to abnormal fluctuations in product indicators.

[0008] The lack of dynamic threshold self-adjustment and migration buffer mechanism can lead to lag or excessively rapid parameter switching after sudden changes in process status, which can easily cause short-term oscillations, instability or even production failures in the system.

[0009] Traditional data modeling and shallow feature input alone are insufficient to extract highly interpretable process state fingerprints, which cannot support subsequent instability tracing and causal reasoning.

[0010] In summary, the existing stability control technology for lithium extraction-co-extraction processes in the hydrometallurgical process of ternary lithium black powder urgently needs innovative breakthroughs. The following objectives are urgently required: Under dynamic, multi-condition, and strongly physicochemical coupling conditions, fully explore the implicit semantic features of the process to break free from reliance on manual experience and static thresholds, and realize a stability identification and control system that can adaptively adjust to deviations from operating conditions; establish a novel intelligent process control scheme based on process semantic fingerprints, incorporating process perturbation identification and threshold drift strategies, thereby improving the system's adaptive capabilities and stability control efficiency under extreme conditions, providing theoretical support and practical assurance for intelligent recycling and the safe and efficient operation of metallurgical processes. Summary of the Invention

[0011] This application provides an optimized method for selective lithium extraction from ternary lithium black powder and nickel-cobalt-manganese co-extraction process, aiming to solve one of the problems or issues of the prior art mentioned in the background.

[0012] This application provides an optimized method for selective lithium extraction from ternary lithium black powder and nickel-cobalt-manganese co-extraction process, specifically including: S1: Deploy sensing devices in the boundary area between the lithium extraction section and the co-extraction section to simultaneously collect raw indicators reflecting changes in the dynamics of the two-phase interface and the liquid phase chemical environment, and generate a raw signal set of the ternary lithium black powder wet metallurgical process. S2: Based on the original signal set of the ternary lithium black powder wet metallurgy process, the energy entropy spectrum distribution is extracted as a frequency domain feature, and the frequency domain feature is fused with real-time flow, temperature and redox potential data to construct a six-dimensional working condition input vector. S3: Based on multiple preset semantic cluster centers, the six-dimensional working condition input vector is input into a graph convolutional network to calculate the distance weighted value of the current working condition in each semantic cluster center, and generate a non-parametric process state semantic fingerprint. S4: If the non-parametric process state semantic fingerprint undergoes cross-cluster migration, a migration buffer window is activated in the overlapping area of ​​the old and new semantic clusters. While maintaining the original control strategy, the pre-calibrated PID parameter group corresponding to the new semantic cluster is loaded in parallel to generate a disturbance response quality factor that includes the phase margin improvement rate and the gain settling time reduction rate. S5: If the phase margin improvement rate and the gain stabilization time reduction rate meet the preset convergence conditions, the control parameters will be updated to the pre-calibrated PID parameter group corresponding to the new semantic cluster; otherwise, the original control strategy will be maintained and the migration buffer window duration will be extended, and the final closed-loop stability control command will be output.

[0013] This application provides an optimized process for selective lithium extraction from ternary lithium black powder and nickel-cobalt-manganese co-extraction, which has the following beneficial effects: (1) By deploying a micro-rheological sensor array and a micro-differential pH probe group in the junction area between the lithium extraction section and the co-extraction section, key dynamic features such as the liquid phase viscosity change response time, the two-phase interface tension relaxation period and the pH jump front propagation speed are obtained simultaneously. Combined with wavelet packet decomposition to extract its energy entropy spectrum distribution, the problem of lag and high misjudgment rate caused by the traditional online monitoring method relying only on a single parameter threshold alarm is effectively overcome. Furthermore, the above features are integrated with flow rate, temperature and redox potential to construct a six-dimensional input vector, which is mapped to a predefined process state semantic space through a lightweight graph convolutional network to generate a "semantic fingerprint" with clear physical interpretation. This achieves a highly sensitive and interpretable characterization of multi-metal competitive complexation and interfacial mass transfer behavior in complex hydrometallurgical processes, significantly improving the accuracy and response sensitivity of working condition identification, especially showing stronger robustness when facing non-steady-state disturbances such as raw material composition fluctuations or pulse feeding.

[0014] (2) The introduction of the "perturbation excitation-response identification" mechanism and the adaptive stability threshold dynamic modulation strategy fundamentally breaks through the technical limitations of traditional control systems that rely on fixed boundary conditions for stability determination. When the system detects that the semantic fingerprint is continuously located in the edge region of a semantic cluster, it actively applies a short-term micro square wave perturbation and analyzes the phase shift and gain attenuation characteristics of the nickel-cobalt-manganese concentration ratio at the extraction phase outlet. It accurately identifies whether the current condition is in the critical region of co-extraction selectivity and adjusts the parameter tolerance range under this condition in real time. Based on the historical optimal operating range, the control slackness is linearly scaled according to the distance of the semantic fingerprint from the cluster center. The intelligent adaptation logic of "flexible control of typical operating mode and conservative response of abnormal operating mode" is realized, which greatly improves the long-term operating stability of the system in complex and variable operating environments, effectively avoids the risk of chain instability caused by slight drift of local parameters, and at the same time ensures the synergistic optimization of valuable metal recovery rate and separation selectivity.

[0015] (3) The design of the “migration buffer window” mechanism and the event-flow-based temporal knowledge graph recording architecture enables the system to have the ability to smoothly transition across operating conditions and deep causal tracing function. When the semantic fingerprint migrates across clusters, the control parameters are not switched immediately. Instead, the original strategy is maintained for two minutes in the overlapping area of ​​the old and new semantic clusters. During this period, the target cluster pre-calibrated PID parameter group is loaded in parallel. The system decides whether to switch based on the disturbance response quality factors (such as the phase margin improvement rate and the gain stabilization time reduction rate) during the buffer period. This significantly reduces the transient impact and control oscillation risk caused by mode jump. At the same time, all semantic coordinate evolution trajectories, disturbance response data and threshold drift paths are written into the temporal knowledge graph in the form of structured event streams. This supports the subsequent inversion of the root cause of stability imbalance through the causal reasoning engine. It distinguishes whether the cause is due to different physical sources such as raw material batch mutation, extractant activity decay or sensor response lag. This provides a reliable basis for process optimization and equipment maintenance. The system constructs a closed-loop adaptive control system from perception, cognition, decision-making to backtracking. Without requiring a large number of labeled samples or external control paradigm shifts, it closely follows the core chemical nature of ternary lithium black powder wet recycling, such as acid-base sensitive transitions, multi-metal complexation competition, and mass transfer hysteresis, achieving a unified approach to intelligent, lightweight, and interpretable process stability control.

[0016] In summary, this solution comprehensively improves the dynamic response capability, operational robustness, and fault attribution accuracy of the co-extraction stage in hydrometallurgical processes by constructing a lightweight graph neural network model oriented towards process semantic understanding, introducing active excitation identification and adaptive threshold adjustment mechanisms, designing cross-condition buffer switching logic, and a causal traceability knowledge graph architecture. It not only effectively solves the problems of poor adaptability and rigid control of traditional methods when facing atypical operating conditions, but also achieves essential-level modeling of complex chemical processes without relying on large-scale labeled data and cross-domain control templates, providing a brand-new technical path for high-precision and high-reliability separation control in lithium battery recycling systems. Attached Figure Description

[0017] Figure 1 This is a main flow chart of an optimized process for selective lithium extraction from ternary lithium black powder and nickel-cobalt-manganese co-extraction. Figure 2 This is a sub-flowchart of an optimized process for selective lithium extraction from ternary lithium black powder and nickel-cobalt-manganese co-extraction. Figure 3 This is another sub-flowchart of an optimized process for selective lithium extraction from ternary lithium black powder and co-extraction of nickel, cobalt, and manganese. Figure 4 This is an application environment diagram of an optimized process for selective lithium extraction and nickel-cobalt-manganese co-extraction of ternary lithium black powder in one embodiment. Figure 5 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0020] like Figure 1 As shown, this application provides an optimized method for selective lithium extraction from ternary lithium black powder and nickel-cobalt-manganese co-extraction process, specifically including steps S1 to S5.

[0021] S1: Deploy sensing devices at the junction of the lithium extraction section and the co-extraction section to simultaneously collect raw indicators reflecting changes in the dynamics of the two-phase interface and the liquid phase chemical environment, and generate a raw signal set of the ternary lithium black powder wet metallurgical process.

[0022] The sensing devices may include, but are not limited to, micro-rheological sensor arrays and micro-differential pH probe groups; the original indicators reflecting the dynamics of the two-phase interface and changes in the liquid phase chemical environment may include, but are not limited to, the liquid phase viscosity change response time, the two-phase interface tension relaxation decay period, and the pH jump front propagation speed; the original signal set of the ternary lithium black powder wet metallurgical process may include, but is not limited to, time-series waveform characteristics.

[0023] S2: Based on the original signal set of the ternary lithium black powder wet metallurgical process, the energy entropy spectrum distribution is extracted as a frequency domain feature, and the frequency domain feature is fused with real-time flow, temperature, and redox potential data to construct a six-dimensional operating condition input vector.

[0024] Specifically, wavelet packet decomposition is performed on the original signal set of the ternary lithium black powder wet metallurgy process to extract the energy entropy spectrum distribution as a frequency domain feature. The frequency domain feature is then fused with real-time flow rate, temperature, and redox potential data to construct a six-dimensional operating condition input vector.

[0025] S3: Based on multiple preset semantic cluster centers, the six-dimensional working condition input vector is input into a graph convolutional network to calculate the distance weighted value of the current working condition in each semantic cluster center, and generate a non-parametric process state semantic fingerprint.

[0026] Specifically, the semantic cluster centers include twelve semantic cluster centers defined by unsupervised clustering of typical operating conditions such as high lithium and low nickel, low acid and high manganese, and pulse feeding in the offline stage; based on the above twelve semantic cluster centers, the six-dimensional operating condition input vector is input into a lightweight graph convolutional network to calculate the distance weighted value of the current operating condition in each semantic cluster center, and generate a non-parametric process state semantic fingerprint representing the current operating state.

[0027] S4: If the non-parametric process state semantic fingerprint undergoes cross-cluster migration, a migration buffer window is activated in the overlapping area of ​​the old and new semantic clusters. While maintaining the original control strategy, the pre-calibrated PID parameter group corresponding to the new semantic cluster is loaded in parallel to generate a disturbance response quality factor that includes the phase margin improvement rate and the gain settling time reduction rate.

[0028] Specifically, it detects whether the non-parametric process state semantic fingerprint has undergone cross-cluster migration. If a migration event is detected, a migration buffer window is started in the overlapping area of ​​the old and new semantic clusters. While maintaining the original control strategy, the pre-calibrated PID parameter group corresponding to the new semantic cluster is loaded in parallel to generate a disturbance response quality factor that includes the phase margin improvement rate and the gain settling time reduction rate.

[0029] S5: If the phase margin improvement rate and the gain stabilization time reduction rate meet the preset convergence conditions, the control parameters will be updated to the pre-calibrated PID parameter group corresponding to the new semantic cluster; otherwise, the original control strategy will be maintained and the migration buffer window duration will be extended, and the final closed-loop stability control command will be output.

[0030] Specifically, the control strategy switching decision is executed based on the numerical performance of the disturbance response quality factor. If the phase margin improvement rate and the gain stabilization time reduction rate meet the preset convergence conditions, the control parameters are updated to the pre-calibrated PID parameter group corresponding to the new semantic cluster. Otherwise, the original control strategy is maintained and the migration buffer window duration is extended, and the final closed-loop stability control command is output.

[0031] Step S1: Deploy a micro-rheological sensor array and a micro-differential pH probe group at the interface between the lithium extraction section and the co-extraction section to simultaneously collect three raw indicators: the liquid phase viscosity change response time, the interphase interface tension relaxation decay period, and the pH jump front propagation speed, generating a raw signal set of the ternary lithium black powder hydrometallurgical process containing time-series waveform characteristics. Specifically, this includes: S1.1: Based on the hydrodynamic characteristics of the boundary region between the lithium extraction section and the co-extraction section in the ternary lithium black powder wet metallurgy process, the installation topology of the micro rheological sensing array and the micro differential pH probe group is determined in order to obtain a set of key monitoring points that can characterize the mass transfer efficiency of the two-phase mixing.

[0032] In the hydrometallurgical process of ternary lithium black powder, the mixing and mass transfer behavior of the liquid and organic phases at the interface between the lithium extraction and co-extraction sections constitutes a crucial scenario for initial signal acquisition. It is necessary to determine the installation topology of the micro-rheological sensor array and the micro-differential pH probe group based on the fluid dynamics characteristics of this region to obtain a set of key monitoring points characterizing the two-phase mixing and mass transfer efficiency. Based on the liquid flow path between the reactor, mixing tank, and extraction tower, the velocity gradient distribution of the local flow field is calculated. The three-dimensional velocity vector field is solved through fluid simulation, and the coordinate data of the high-shear zone where the velocity gradient exceeds a set threshold are extracted. The high-shear zone coordinate data and the interface fluctuation zone location data are spatially superimposed and analyzed to screen the intersection points where both high shear and interface instability exist simultaneously as candidate monitoring points. The hydraulic radius formula is then used... With the cross-sectional area of ​​the flow channel Wetted perimeter data is used to calculate local fluid resistance to assess pressure stability at candidate points and eliminate areas with excessive pressure fluctuations. The pH front propagation velocity is quantitatively measured using an acid-base indicator diffusion experiment. This velocity data is then matched with the locations corresponding to critical values ​​of liquid phase viscosity abrupt changes, narrowing the monitoring point range to nodes with the most representative mass transfer characteristics. Through the above chain-like calculation and matching process, the flow field simulation results are integrated with the characteristic parameters obtained from physical experiments to output the installation topology matrix of the micro-rheological sensor array and the micro-differential pH probe group, achieving precise positioning of key monitoring points.

[0033] For example, in a ternary lithium black powder hydrometallurgical production line, fluid simulation was performed on the interface between the lithium extraction section and the co-extraction section. The inlet velocity was set to 0.45 m / s, the liquid phase density to 1050 kg / m³, and the organic phase density to 880 kg / m³. The calculated velocity gradient in the high-shear zone was 120 s. -1 Locations with interface fluctuation frequencies exceeding 8Hz were extracted and superimposed with the coordinates of the high-shear zone to obtain six candidate monitoring points. Using the hydraulic radius formula (where A = 0.003 m² and wetted perimeter = 0.072 m), local fluid resistance was calculated, and two points with pressure fluctuations exceeding 15 kPa were eliminated, leaving four candidate monitoring points. The pH front advance velocity was measured to be 0.0065 m / s using acid-base diffusion experiments, and matched with the location corresponding to the critical value of liquid phase viscosity abrupt change, ultimately determining three key monitoring points. A micro-rheological sensor array was deployed at these three points, and the micro-differential pH probe group was installed with a 5 mm spacing between two points. The collected liquid phase viscosity abrupt change and pH front data significantly improved the signal-to-noise ratio in subsequent steps, enhancing the accuracy of process state determination.

[0034] S1.2: Using a micro-rheological sensor array deployed on the set of key monitoring points, high-frequency shear stress response scanning processing is performed on the liquid medium flowing through the interface region to obtain time series data of liquid phase viscosity change response reflecting the change in the degree of emulsification between the organic phase and the aqueous phase.

[0035] For a micro-rheological sensor array deployed on a set of key monitoring points, the high-frequency sampling units of each channel in the array are invoked to execute a shear stress scanning command set on the liquid medium flowing through the interface region. The scanning frequency and amplitude are set according to the results of the previous fluid dynamics analysis to ensure that the sampling period matches the time scale of local fluid vortex changes.

[0036] The high-frequency shear stress signal is subjected to gain equalization filtering to eliminate the amplitude drift introduced by the equipment characteristics, ensuring that the signal amplitude of the input for subsequent viscosity calculation is stable and free from low-frequency baseline drift.

[0037] Based on the relationship between shear rate and shear stress, the instantaneous viscosity is calculated using the standard rheological formula: in Instantaneous viscosity, For shear stress, denoted as shear rate.

[0038] By detecting abrupt changes in viscosity values ​​over time, and by setting a threshold for abrupt change detection and a sliding window length, the starting time index of the abrupt change is extracted, and a data segment for the abrupt change response interval is established.

[0039] The response delay time difference, i.e. the time difference between the viscosity mutation point and the trigger shear stress change command, is calculated within the mutation response interval to quantify the interfacial dynamic hysteresis characteristics corresponding to the change in the degree of emulsification between the organic phase and the aqueous phase.

[0040] The aforementioned time delay difference is continuously recorded in the order of sampling timestamps to form liquid phase viscosity change response time series data, and metadata such as monitoring points, scanning frequency, and threshold settings are added as process data integrity identifiers.

[0041] By using high-frequency shear stress scanning and time delay difference calculation, the key monitoring signals obtained in the previous step are transformed into liquid phase viscosity change response time series data that can characterize the changes in two-phase emulsification and mass transfer efficiency, thereby achieving accurate identification and quantification of the rheological state of the interface region.

[0042] For example, a four-channel micro-rheological sensor array is deployed at the boundary between the lithium extraction section and the co-extraction section. The channel spacing is set to 5 cm, the scanning frequency is configured to 200 Hz, and the shear stress test amplitude is 1.5 Pascals. The acquired signal is processed by a high-pass filter with a filter cutoff frequency set to 2 Hz to remove low-frequency drift. When calculating the instantaneous viscosity, the shear rate is measured by the rotor speed, which is 125 s in this embodiment. -1 Substitute into the formula The instantaneous viscosity value was obtained as 0.012 Pa·s. The mutation detection threshold was set to be 0.005 Pa·s higher than the baseline viscosity value, the sliding window width was 25 sampling points, and the mutation start time index was detected at sampling point 512, corresponding to a delay time difference of 0.125 seconds. The delay time differences collected at different monitoring points were 0.118 seconds, 0.130 seconds, 0.125 seconds, and 0.122 seconds, respectively. The time series was recorded and combined with metadata such as monitoring points and scanning frequency to form a complete liquid phase viscosity mutation response time series. This data was used for subsequent synchronous analysis with the pH jump front velocity and interfacial tension changes, significantly improving the quantitative accuracy of rheological state changes in the interface region.

[0043] S1.3: Based on the same spatiotemporal window corresponding to the liquid phase viscosity change response time series data, a dual-point potential difference dynamic tracking process is performed through a micro-differential pH probe group to extract the pH jump front propagation speed waveform data that characterizes the advancement rate of the acid-base neutralization reaction front.

[0044] Based on the same spatiotemporal window of the liquid phase viscosity change response time series, a micro-differential pH probe group deployed in the interface region is invoked and the scale of the two-point detection spacing matched with the liquid phase flow rate is set to ensure that the acid-base mass transfer gradient between the detection nodes can be completely captured.

[0045] Differential amplification and zero-drift compensation are performed on the potential signal of the dual-point detection node. The transient potential difference is precisely amplified using a high-impedance amplification module, and a time-continuous potential difference sampling sequence is generated through a digital-to-analog conversion unit.

[0046] The first derivative edge detection algorithm is applied to the potential difference sampling sequence to identify the starting position of the pH value jump and calculate the integral length of the front, which is then converted into the front propagation distance.

[0047] Based on the known physical properties of the liquid medium and the actual measured propagation distance of the leading edge, combined with the probe spacing and sampling time interval, the propagation distance is divided by the time difference of the leading edge arrival to obtain the leading edge propagation rate using the velocity calculation formula, which is expressed as: in, Indicates the propulsion speed of the pH jump front. Indicates the probe spacing. This indicates the time difference between the arrival of the leading edge at the two probes.

[0048] The propagation rate of the leading edge is sorted by sampling time and mapped to the time axis to construct waveform data of the propagation velocity of the pH jump leading edge. Noise is filtered and baseline correction is performed on the waveform to form data input that can be directly used for subsequent interfacial tension measurement.

[0049] By using the above processing method, the pH potential difference within the window corresponding to the liquid phase viscosity change response time series is converted into the pH jump front propagation velocity waveform data, which characterizes the mass transfer dynamics of the acid-base neutralization reaction, thereby achieving a high-precision quantitative characterization of the acid-base reaction rate characteristics.

[0050] For example, the two-point spacing of the micro-differential pH probe array was set to 5 mm to detect a medium with a liquid flow rate of 12 mm / s. The differential amplifier gain was set to 1000 to optimize the signal-to-noise ratio of the potential difference. In the acquired potential difference sequence, the starting time of the leading edge was found to be 1.25 s and the ending time was 1.67 s using first-order derivative edge detection. The time difference between the two probes receiving the leading edge... , probe spacing Substitute into the velocity calculation formula The obtained leading-edge propulsion velocity was approximately 11.9 mm / s. After noise filtering, this velocity waveform exhibited a smooth change, which significantly improved the accuracy of identifying acid-base mass transfer dynamics in subsequent interfacial tension relaxation decay period calculations, and maintained stable output in multiple batches of tests.

[0051] S1.4: Combining the liquid phase viscosity change response time series data and the pH jump front propagation velocity waveform data, the optical interfacial tension detection unit is simultaneously invoked to perform transient relaxation decay period calculation on the two-phase contact surface, so as to generate a two-phase interfacial tension relaxation decay period dataset containing interfacial energy change characteristics.

[0052] Input binding is performed on the liquid phase viscosity change response time series data and the pH jump front propagation velocity waveform data to establish multi-source data joint retrieval conditions for corresponding spatiotemporal windows, so as to ensure that the triggering time of the interfacial tension detection unit is consistent with the time reference of the aforementioned two data acquisitions.

[0053] The optical interface tension detection unit is invoked to apply a standardized micro-perturbation light signal to the two-phase contact interface in the boundary region. The interferometric diffraction imaging module captures the mapping sequence of interface deformation over time and generates high frame rate image data of the interface relaxation process.

[0054] An edge contour extraction algorithm is used to perform frame-by-frame interface boundary recognition on high frame rate image data. The transient interface tension value is calculated based on the interface curvature change rate to form a tension time series curve covering the entire time domain of the relaxation process.

[0055] The tension time-series curve is fitted with an exponential decay model, the relaxation coefficient is solved using the least squares method, and the relaxation decay period is calculated using the following formula: in, The relaxation decay coefficient is obtained by fitting.

[0056] By combining the relaxation decay period with the initial and steady-state values ​​of the interfacial tension curve, a multidimensional feature vector containing the characteristics of interfacial energy change is constructed and written into the two-phase interfacial tension relaxation decay period dataset.

[0057] By using interfacial tension optical detection and relaxation period calculation, the liquid phase viscosity and pH jump data from the previous step are transformed into a standardized dataset that characterizes the changes in interfacial mass transfer kinetics, enabling early quantification of potential imbalances in the stability of two-phase mixing.

[0058] For example, on a ternary lithium black powder hydrometallurgical production line with an annual processing capacity of 500 tons, the sampling frequency of the liquid phase viscosity change response time series was set to 200Hz, the sampling frequency of the pH jump propagation velocity was set to 100Hz, and the interfacial tension optical detection unit used an interferometric laser source with a wavelength of 532nm, with the interferometric imaging frame rate set to 1000fps. When fitting an exponential decay model, the relaxation decay coefficient was obtained from the interfacial tension curve. The relaxation decay period is then... The data is calculated in seconds, specifically 4 seconds. This periodic data, combined with the initial tension of 35 mN / m and the steady-state tension of 28 mN / m, constructs a three-dimensional interfacial energy change feature vector [4, 35, 28], serving as a standard record for the two-phase interfacial tension relaxation decay period dataset. Under 20 batches of different raw material conditions, the relaxation decay period variation range of this feature vector was effectively monitored, significantly improving the system's early warning capability for interfacial stability anomalies.

[0059] S1.5: Perform multi-channel timestamp alignment and fusion encapsulation processing on the liquid phase viscosity change response time series data, pH jump front propagation velocity waveform data, and two-phase interfacial tension relaxation decay cycle dataset to generate a raw signal set of the ternary lithium black powder wet metallurgical process containing complete time-series waveform features.

[0060] Step S2: The original signal set from the ternary lithium black powder wet metallurgical process is subjected to wavelet packet decomposition to extract the energy entropy spectrum distribution as a frequency domain feature. This frequency domain feature is then fused with real-time flow rate, temperature, and redox potential data to construct a six-dimensional operating condition input vector. Specifically, this includes: S2.1: Wavelet packet decomposition is performed on the liquid phase viscosity change response time-series waveform, the two-phase interfacial tension relaxation decay period time-series waveform, and the pH jump front propagation velocity time-series waveform in the original signal set of the ternary lithium black powder wet metallurgy process. Orthogonal wavelet basis functions are used to map each time-series waveform to a multi-resolution frequency band space to obtain a multi-level wavelet packet decomposition coefficient sequence containing high-frequency transient details and low-frequency trend components.

[0061] Targeted frequency domain decomposition was performed on the time-series waveforms of liquid phase viscosity abrupt change response, interphase tension relaxation decay period, and pH jump front propagation velocity from the original signal set of the ternary lithium black powder hydrometallurgical process. The sampling frequency and signal length parameters of each time-series waveform were clarified, and a corresponding signal matrix representation was established. Based on this matrix, the orthogonal wavelet basis function type, decomposition level, and frequency band division rules were set during the wavelet packet decomposition initialization stage to ensure that the multi-resolution analysis results covered high-frequency transient details and low-frequency trend components. For each type of time-series waveform, the wavelet packet decomposition algorithm was called to perform complete decomposition node by node, extracting the coefficient sequences of each level and frequency band, and maintaining a one-to-one correspondence between the coefficient sequences and the original signal time sequence according to the time index. For high-frequency transient coefficients, the coefficient peaks at abrupt change locations were extracted using a local peak detection algorithm; for low-frequency trend coefficients, background drift was eliminated through polynomial fitting to obtain stable trend components. The coefficient sequences of different frequency bands were combined in a hierarchical manner to form a multi-level wavelet packet decomposition coefficient matrix, and sorted according to the frequency band energy distribution for subsequent energy entropy spectrum calculation. By using orthogonal wavelet packet decomposition, the time-series waveform features of the previous step are transformed into a multi-level wavelet packet decomposition coefficient sequence containing high-frequency transients and low-frequency trends, thus achieving accurate characterization of liquid phase rheological signals, interfacial tension changes, and acid-base front propagation characteristics in a multi-resolution frequency band space.

[0062] For example, in the hydrometallurgical process of ternary lithium black powder, at the interface between lithium extraction and co-extraction, the sampling time series of the liquid phase viscosity change response is 1024 points long with a sampling frequency of 256 Hz; the interfacial tension relaxation period sequence is 2048 points long with a sampling frequency of 128 Hz; and the pH jump front propagation velocity sequence is 512 points long with a sampling frequency of 64 Hz. Daubechies-4 is selected as the orthogonal wavelet basis function, the decomposition level is set to 5 levels, and the frequency band division rule adopts a binary tree structure. When performing wavelet packet decomposition on the liquid phase viscosity change response signal, a sequence of 32 frequency band coefficients in the 5th level is obtained. Significant coefficient fluctuations are found in the 24th frequency band through peak detection, with a peak coefficient of 3.57. The low-frequency trend component of the interfacial tension relaxation period signal is fitted with a third-order polynomial, and the standard deviation of the fitting residual is 0.012. Energy calculation is performed on the high-frequency component coefficients of the pH jump front propagation velocity signal, and the energy value of the 5th frequency band in the 3rd level is 15.8. The decomposition coefficients corresponding to each waveform are sorted according to the frequency band energy to form a multi-level coefficient matrix, where the number of columns in the matrix corresponds to the frequency band position and the number of rows corresponds to the waveform type. The matrix is ​​then passed as input to the subsequent energy entropy spectrum distribution calculation module. The verification results show that the decomposition coefficient matrix has a significantly improved analytical capability in capturing transient fluctuations and acid-base mass transfer hysteresis characteristics caused by multi-metal competitive complexation, supporting subsequent high-precision working condition determination.

[0063] S2.2: Calculate the energy proportion value of each frequency band node based on the multi-level wavelet packet decomposition coefficient sequence, and perform Shannon entropy operation on the energy proportion value to quantify the disorder of the frequency band energy distribution, thereby generating an energy entropy spectrum distribution feature vector that can characterize the liquid phase rheological properties and the complexity of acid-base mass transfer kinetics.

[0064] S2.3: Acquire the instantaneous values ​​of feed flow rate, reactor temperature, and redox potential as static operating parameters, and perform timestamp alignment processing between the static operating parameters and the energy entropy spectrum distribution feature vector to eliminate phase deviation of multi-source heterogeneous data at the sampling time and generate a time-synchronized multi-source data set.

[0065] S2.4: Perform normalization scaling on the time-synchronized multi-source data set, and use the range normalization algorithm to uniformly map the energy entropy spectrum distribution feature vector and the numerical range of each static working condition parameter to the dimensionless interval of zero to one, so as to eliminate the influence of different physical dimensions on the convergence speed of the subsequent neural network model and generate a standardized multidimensional data matrix.

[0066] Normalization and scaling are performed on a time-synchronized multi-source dataset. The input includes a timestamp-aligned energy entropy spectrum distribution feature vector and three static operating parameters: feed flow rate, reactor temperature, and redox potential. For each physical quantity component, the range normalization algorithm is called to obtain its minimum and maximum values, which are then used to construct a normalization coefficient matrix. Using this normalization coefficient matrix, numerical mapping is performed on each component, proportionally transforming the original values ​​to a dimensionless range of zero to one, thereby eliminating numerical scale differences introduced by different physical dimensions. For the multi-frequency band components in the energy entropy spectrum distribution feature, the above mapping operation is performed separately, and the mapped results are stored in the frequency domain feature normalization cache unit. For the static operating parameters, the same range mapping process is directly applied, and the results are stored in the static parameter normalization cache unit. The data matrix assembly module is called to arrange the components of the frequency domain feature normalization cache and the static parameter normalization cache in a preset column order, generating a normalized multidimensional data matrix, ensuring that each column in the matrix is ​​consistent with the input structure of the subsequent lightweight graph convolutional network. By using the range standardization algorithm, the multi-source data from the previous step is transformed into a standardized multidimensional data matrix with a unified value range and dimensionless dimensions. This effectively eliminates the influence of different physical quantities on the convergence speed of the neural network and maintains the proportional relationship between features.

[0067] For example, in a monitoring of a ternary lithium black powder hydrometallurgical process, the energy entropy spectrum distribution feature vector contains five frequency band components, with values ​​ranging from 0.15 to 0.67, 0.21 to 0.74, 0.18 to 0.62, 0.25 to 0.79, and 0.19 to 0.66, respectively. The instantaneous feed flow rate ranges from 3.25 to 7.80 L / min, the instantaneous reactor temperature ranges from 58.2 to 83.5 ℃, and the instantaneous redox potential ranges from 215 to 348 mV. The normalized mapping of each component is achieved according to the following formula: in, This represents the original value of a certain component. and These represent the minimum and maximum values ​​of the component, respectively. The five frequency band components are normalized sequentially. For example, the first frequency band component, 0.35, is normalized to 0.416 after mapping, and the feed flow rate, 5.60 L / min, is normalized to 0.456 after mapping. All normalization results are arranged in column order with frequency domain components first and static operating parameters last, generating a standardized multidimensional data matrix of size 8×1. In the pre-training test of the lightweight graph convolutional network, this matrix input significantly reduced the total number of iterations required for model convergence, indicating that normalization effectively improves the convergence speed, maintains the proportional information of different features, and enhances the discriminative stability of the operating condition semantic fingerprint.

[0068] S2.5: Based on the standardized multidimensional data matrix, perform vector concatenation operation according to the preset frequency domain feature priority order and process parameter association order, combine the energy entropy spectrum distribution component with the flow rate, temperature, and redox potential components into a fixed-length array structure, and finally generate a six-dimensional operating condition input vector for inputting the lightweight graph convolutional network.

[0069] like Figure 2 As shown, step S3 involves defining twelve semantic cluster centers based on unsupervised clustering of typical operating conditions such as high-lithium low-nickel, low-acid high-manganese, and pulse feeding during the offline phase. The six-dimensional operating condition input vector is then input into a lightweight graph convolutional network to calculate the distance weighted value of the current operating condition at each semantic cluster center, generating a non-parametric process state semantic fingerprint representing the current operating state. Specifically, this includes: S3.1: Obtain the coordinate set of semantic cluster centers for twelve typical operating conditions such as high lithium and low nickel, low acid and high manganese, and pulse feeding generated by unsupervised clustering in the offline stage, and construct a static process state topology graph structure containing node features and edge weights based on the coordinate set of the semantic cluster centers, so as to establish a benchmark reference framework for the generation of non-parametric process state semantic fingerprints.

[0070] Based on the coordinate set of semantic cluster centers for twelve typical operating conditions, including high-lithium low-nickel, low-acid high-manganese, and pulse feeding, generated through unsupervised clustering in the offline stage, this set is loaded into the process state initialization module as input to form a callable cluster center coordinate dataset. Node attribute parsing is performed on the cluster center coordinate dataset to extract the operating condition category label, feature vector elements, and vector dimension information corresponding to each cluster center, establishing a node feature index table for subsequent topology construction. Coordinate space mapping is performed using the node feature index table and the cluster center coordinate dataset to construct a similarity measurement matrix between nodes. The element values ​​of this matrix are obtained by combining cosine similarity and Euclidean distance, where the Euclidean distance is calculated using the formula: in, and These are the feature coordinate vectors of the two cluster centers. Edge weight parameters are set based on this similarity metric matrix, with weight values ​​decreasing as the similarity metric increases to strengthen the connection between cluster centers with small differences. All node features and edge weight data are combined according to the graph structure definition to generate a static process state topology graph structure, and a unique identifier is assigned to it so that the subsequent lightweight graph convolutional network can achieve stable neighborhood information aggregation based on this structure. Topology consistency verification ensures that the connectivity and weight distribution of edges between nodes meet the reference requirements for process state semantic fingerprint calculation. When the six-dimensional process condition input vector from the previous step is embedded into this static process state topology graph, a complete spatial reference framework is maintained, achieving seamless connection between standardized input and semantic space mapping.

[0071] For example, four sets of samples were collected offline for each of three operating conditions: high lithium / low nickel, low acid / high manganese, and pulse feeding. K-means clustering was used to obtain twelve cluster centers. Each cluster center has a six-dimensional feature vector, corresponding to a four-dimensional energy entropy spectrum distribution and two-dimensional flow and temperature components. When performing node attribute parsing, the cluster center coordinates were uniformly stored as floating-point numbers, retaining three decimal places of precision. The cosine similarity threshold was set to 0.85, and the Euclidean distance was calculated according to the above formula. When the maximum distance did not exceed 1.2, an edge connection was established between nodes, and the edge weight was assigned as the reciprocal of the similarity value. For example, the eigenvectors of the high-lithium, low-nickel cluster center C1 and the low-acid, high-manganese cluster center C5 are (0.325, 0.188, 0.412, 0.275, 0.562, 0.498) and (0.291, 0.175, 0.398, 0.284, 0.570, 0.501), respectively. Calculate the Euclidean distance: The calculated result is 0.0503, which is below the threshold of 1.2. An edge connection is established with a weight of 1 / (0.0503+ε), where ε is a minimum value to prevent division by zero. The final topology graph structure significantly improves neighborhood aggregation performance when loaded with a lightweight graph convolutional network, ensuring stable feature propagation paths and full domain coverage during dynamic query node input, thus improving the accuracy and interpretability of semantic fingerprint generation.

[0072] S3.2: The six-dimensional operating condition input vector generated by the previous steps is embedded into the static process state topology graph structure as a dynamic query node. A lightweight graph convolutional network is used to perform neighborhood information aggregation and nonlinear feature transformation processing to output a high-dimensional latent feature representation vector of the current real-time operating condition in the implicit semantic space.

[0073] The six-dimensional operating condition input vector generated by the fusion of the previous steps is embedded as a dynamic query node into the static process state topology graph structure, and a node index mapping relationship is established to ensure that each operating condition input vector is uniquely bound to its corresponding topology graph node.

[0074] The first layer convolution operator of the lightweight graph convolutional network is called in the topology graph to perform feature aggregation processing on the direct neighborhood of the query node. The feature vectors of adjacent nodes are weighted and summed according to the preset edge weight coefficients, and the normalization matrix is ​​used to eliminate the bias of multi-degree nodes on the aggregation result.

[0075] The intermediate feature vector obtained by neighborhood aggregation is input into the nonlinear activation function layer, and the ReLU function is used to realize the sparsification mapping of the feature space, thereby enhancing the distinguishability between features of different working conditions.

[0076] In subsequent layers of the graph convolutional network, second-order neighborhood information aggregation is performed. Multi-resolution fusion is then performed by combining the output of the first layer with the second-order neighborhood features to generate an intermediate representation vector that can reflect the coupling relationship between global and local topological characteristics.

[0077] The multi-level feature representations output from all convolutional layers are merged through a fully connected mapping layer, and batch normalization is used to eliminate scale differences between feature components, resulting in the final high-dimensional latent feature representation vector.

[0078] Through the above embedding and convolution processing methods, the standardized six-dimensional working condition input from the previous step is transformed into a high-dimensional latent feature representation that carries the global topological pattern and local working condition differences, thereby achieving accurate positioning of the process state in the implicit semantic space.

[0079] For example, consider a static process state topology graph containing 12 semantic cluster nodes, where the node feature dimension is 6, and the edge weights are constructed based on the reciprocal of the Euclidean distance, ranging from 0.05 to 0.5. Input a six-dimensional process state vector. As a query node embedded in the topological graph, feature aggregation is performed on the first-order neighborhood. Assuming the neighborhood contains three nodes, their feature vectors are as follows: , , The edge weights are 0.3, 0.25, and 0.2, respectively. The aggregation operation uses... Formula, where This is the weight matrix. For ReLU function, For edge weights, and These are the feature vectors of the query node and the neighboring nodes, respectively. After two layers of convolution and nonlinear mapping, the dimension of the high-dimensional latent feature vector is expanded to 64 dimensions, and the variance of the feature components is reduced to below 0.02 after batch normalization, which significantly improves the stability and accuracy of the working condition classification in the latent semantic space.

[0080] S3.3: Based on the high-dimensional latent feature representation vector and the coordinate set of the pre-constructed twelve semantic cluster centers, calculate the Euclidean distance metric between the current real-time operating condition and each semantic cluster center to generate the original distance distribution sequence characterizing the degree of deviation of the current operating condition from each typical pattern.

[0081] Based on the high-dimensional latent feature representation vector output by the lightweight graph convolutional network and the coordinate set of twelve semantic cluster centers pre-constructed in the offline stage, the Euclidean distance metric algorithm module is called to perform component difference calculation with the coordinates of each cluster center as static reference input and the current high-dimensional latent feature representation vector.

[0082] The component difference results are squared to eliminate directionality, and summation is applied to the squared values ​​of each component to aggregate the overall squared distance values, forming the sum of squared distances corresponding to each semantic cluster.

[0083] The square root function is called using the sum of squares to calculate the distance scalar and establish an initial sequence of Euclidean distances between the current working condition and the centers of all twelve semantic clusters.

[0084] The original distance distribution sequence is constructed using the above sequence to ensure that each distance data point directly maps the degree to which the current operating condition deviates from each typical pattern in numerical form.

[0085] To avoid metric distortion caused by differences in feature dimensions, consistency calibration is performed on the component structure in the high-dimensional latent feature representation vector and the coordinates of each cluster center to ensure the comparability and accuracy of Euclidean distance calculation.

[0086] Using the Euclidean distance formula: in, These are the components of the current high-dimensional latent feature representation vector. These are the corresponding components of the cluster center coordinates. Number the clusters. As a component index, the original distance distribution sequence is generated through cluster-by-cluster and component-by-component operations. This represents the total number of dimensions of the vector.

[0087] Through the above processing method, the high-dimensional latent feature representation vector of the previous step is transformed into the original distance distribution sequence that fully quantifies the deviation of the current working condition from each typical mode, thereby achieving accurate spatial positioning of the working condition location.

[0088] For example, in the lithium extraction-co-extraction hydrometallurgical process of ternary lithium black powder, the high-dimensional latent feature representation vector is set to a 16-dimensional structure, with each component taking values ​​from 0 to 1; the coordinates of the center of the twelve semantic clusters are also a 16-dimensional structure, with the value range consistent with that of the high-dimensional latent feature vector. For a given sampling period, the input high-dimensional latent feature representation vector is {0.12, 0.35, 0.87, 0.65, 0.44, 0.31, 0.90, 0.52, 0.23, 0.77, 0.64, 0.41, 0.58, 0.19, 0.82, 0.56}, and the coordinates of a cluster center are {0.15, 0.34, 0.88, 0.66, 0.45, 0.30, 0.91, 0.50, 0.20, 0.75, 0.66, 0.40, 0.60, 0.18, 0.80, 0.55}. Using the Euclidean distance formula, the sum of squared distances for this cluster is 0.0045, and the square root of the distance scalar is approximately 0.06708. Similarly, the distance scalars of other cluster centers are calculated to form the original distance distribution sequence, such as {0.06708, 0.54215, 0.11233, ..., 0.25498}. This sequence is output to the subsequent normalized weighted fusion module. In this embodiment, the average difference between the output original distance distribution sequence and the manually calculated sequence is less than 0.0001, indicating that this step can significantly improve the spatial positioning accuracy of the semantic fingerprint generation process and provide highly reliable basic data for edge condition determination and perturbation excitation triggering.

[0089] S3.4: Perform normalized weighted fusion processing on the original distance distribution sequence, and allocate the contribution weight of each semantic cluster according to the inverse distance relationship to calculate the comprehensive position coordinates of the current working condition relative to all typical working condition modes, thereby eliminating the local uncertainty of single cluster classification.

[0090] S3.5: Generate a unique non-parametric process state semantic fingerprint based on the comprehensive position coordinates. This fingerprint quantifies the current operating state of the lithium extraction and co-extraction process of ternary lithium black powder in the form of a multi-dimensional vector, and serves as the direct input basis for subsequent judgment of the edge region of the operating condition and triggering perturbation excitation.

[0091] The weighted fusion-derived comprehensive location coordinates are used as input data. A unique identifier generation operation is performed in the feature space. A multidimensional vector encoding scheme is used to convert the normalized Euclidean distance of each semantic cluster from the current working condition into vector components.

[0092] The coordinate axes of the comprehensive position coordinates are rearranged so that each component corresponds to specific working condition features in a preset dimensional order, ensuring that the vector structure remains consistent in subsequent edge determination.

[0093] For each normalized Euclidean distance, a stability sensitivity mapping is performed, adjusting the distance values ​​to weights for stability control through a nonlinear scaling function, forming component values ​​that enhance semantic differences.

[0094] A multidimensional fingerprint vector is constructed using the component value sequence, encapsulated in a fixed-length array structure, and a unique time sequence marker is appended to the head of the vector to locate the current running state in the time series.

[0095] A consistency check is performed on the encapsulated multidimensional fingerprint vector to ensure that the numerical range of all components meets the dimensionless constraint from zero to one, thus preventing abnormal parameters from occurring when the perturbation excitation mechanism is triggered.

[0096] By generating unique identifiers and using structured encapsulation, the comprehensive location coordinates are transformed into multidimensional non-parametric process state semantic fingerprints, enabling quantitative characterization of the current operating status of lithium extraction and co-extraction processes of ternary lithium black powder, and providing direct input basis for subsequent judgment of edge regions of operating conditions and triggering of perturbation excitation.

[0097] For example, under the operating conditions of a hydrometallurgical recovery line, the six-dimensional input vector is {0.45, 0.62, 0.51, 0.37, 0.48, 0.59}. The normalized Euclidean distance sequence corresponding to the twelve semantic cluster centers is then weighted and fused to obtain the comprehensive position coordinates {0.087, 0.094, 0.092, 0.081, 0.096, 0.089}. These position values ​​are then mapped using a stability sensitivity mapping function. Calculation, where To normalize the Euclidean distance, the mapped component values ​​are {0.519, 0.523, 0.522, 0.520, 0.524, 0.521}. This sequence of component values ​​is encapsulated into a fingerprint vector of length 12 in a preset order, with the current sampling timestamp "2024-05-16T10:24:00Z" appended, resulting in an output vector structure {0.519, 0.523, 0.522, 0.520, 0.524, 0.521, ..., timestamp}. Under different batches of raw materials, such as high-lithium, low-nickel conditions, the mapped component values ​​are generally higher. The fingerprint vector triggers an edge detection mechanism in the stability determination module, significantly improving the accuracy of critical condition identification and maintaining stability control in actual verification.

[0098] Step S4: Detect whether cross-cluster migration occurs in the non-parametric process state semantic fingerprint. If a migration event is detected, a migration buffer window is activated in the overlapping area of ​​the old and new semantic clusters. While maintaining the original control strategy, the pre-calibrated PID parameter set corresponding to the new semantic cluster is loaded in parallel to generate a disturbance response quality factor that includes the phase margin improvement rate and the gain settling time reduction rate. Specifically, this includes: S4.1: Perform continuous temporal trajectory monitoring on the non-parametric process state semantic fingerprint. Based on the rate of change of the Euclidean distance between the fingerprint coordinates at the current sampling time and the historical trajectory points, determine whether the non-parametric process state semantic fingerprint crosses a preset semantic cluster boundary threshold. If a cross-cluster migration event is determined to have occurred, a migration trigger signal is generated.

[0099] High-precision monitoring and processing are performed on the continuous temporal trajectory of the non-parametric process state semantic fingerprint. The fingerprint coordinate acquisition module outputs the multi-dimensional coordinate vector of the current sampling moment and writes it into the trajectory cache sequence. Based on the historical trajectory point data in the trajectory cache sequence, the spatial distance difference between the current sampling point and each historical trajectory point is calculated using an Euclidean distance metric, and a distance change rate sequence is formed to characterize the temporal drift speed of the fingerprint coordinates. A preset semantic cluster boundary threshold library is called, and the distance change rate sequence is synchronously compared with the boundary threshold of the old semantic cluster center position to generate a Boolean matrix for the current point to cross the boundary. Logical scanning is performed on the Boolean matrix to identify sampling segments that continuously meet the boundary threshold crossing conditions, and the migration judgment module is called to solidify the judgment results. If the crossing condition meets the preset continuous sampling judgment limit, a migration trigger signal is generated and written into the migration event queue as the start condition for the subsequent migration buffer window activation. Through continuous trajectory monitoring and boundary discrimination processing, the dynamic stability judgment threshold set and fingerprint coordinate temporal data generated in the previous step are transformed into cross-cluster migration event signals, realizing accurate capture of the transition of the operating mode.

[0100] For example, in the real-time operation of the ternary lithium black powder hydrometallurgical process at the lithium extraction-co-extraction interface, non-parametric process state semantic fingerprint six-dimensional coordinate data is collected, with a sampling period of 2 seconds and a trajectory buffer sequence length of 300 points. In the migration determination, the Euclidean distance formula is used. in, For each component of the fingerprint vector, This represents the component value at the current sampling time. This represents the component at the previous sampling time. The dimension of the vector is denoted as 0.85. The semantic cluster boundary threshold is set to 0.85, and the consecutive crossing condition is defined as three sampling periods. During operation, when the calculated distance change rate sequence exceeds 0.85 three times consecutively, the migration determination module confirms the cross-cluster event and outputs a migration trigger signal. This signal triggers the migration buffer window timer in S4.2. Verification results show that under scenarios of fluctuations in different raw material batches and degradation of extractant performance, this processing method can significantly improve the accuracy of pattern migration detection and ensure timely adjustment of the stability control strategy.

[0101] S4.2: Based on the migration trigger signal, activate the migration buffer window timer in the overlapping area of ​​the feature spaces of the old semantic cluster and the new semantic cluster, lock the currently executing original control strategy parameter group to maintain the instantaneous flow ratio of the feed pump and the extractant, and at the same time retrieve and load the target pre-calibrated PID parameter group that matches the center of the new semantic cluster from the pre-calibration parameter library and enter the parallel standby state.

[0102] The old semantic cluster center refers to the center coordinate of the semantic cluster to which the non-parametric process state semantic fingerprint stably belongs in the process state semantic feature space before the current time. It represents the optimized typical operating mode adapted to the process operation before the system triggers the migration. The original control strategy parameter set associated with it is a set of controller parameters that have been verified to be applicable to the operating conditions represented by the old cluster. The new semantic cluster center refers to the center coordinates of another semantic cluster that is about to be entered in the semantic feature space of the process state, which is predicted based on the rate of change of the Euclidean distance of the monitored fingerprint trajectory and the threshold of crossing the semantic cluster boundary. It represents another typical operating mode that the system expects the process state to switch to.

[0103] The overlapping region of the feature spaces of the old and new semantic clusters refers to the spatial domain formed in the semantic feature space of the process state, with the centers of the old and new semantic clusters as focal points, and jointly covered by the effective influence range (or probability distribution) of the two clusters. This spatial domain represents the fuzzy transition and undetermined state between two typical operating conditions. The semantic fingerprint of the process state located in this region has an uncertain affiliation (belonging to the old or new cluster) or is in the process of dynamic transformation. Setting the activation condition of the migration buffer window timer here is to initiate a safe and controllable parameter switching verification process during the critical fuzzy period when the process state has clearly left the old cluster but has not yet been fully established in the new cluster, thereby achieving a smooth and uninterrupted control strategy migration.

[0104] S4.3: Using the target pre-calibrated PID parameter set, apply a small-amplitude tentative adjustment to the system within the migration buffer window period, and simultaneously acquire the transient response waveform of the nickel-cobalt-manganese concentration ratio at the extract phase outlet. Perform frequency domain phase calculation on the transient response waveform to obtain the phase margin improvement rate value, and perform decay time fitting on the envelope of the transient response waveform to obtain the gain stabilization time reduction rate value.

[0105] During the migration buffer window, the target pre-calibrated PID parameter set is loaded into the parallel standby port of the controller, and a small, tentative adjustment is made to the feed pump and extractant flow ratio, with an amplitude not exceeding one percent of the baseline set value, to ensure that the system disturbance response generates sufficient observable dynamic change signals within a controllable range.

[0106] Transient time-series data of nickel, cobalt, and manganese concentration ratios were collected at the extraction phase outlet using an online mass spectrometer. A complete input-output response mapping was established by combining the sampling time with the time synchronization marker of the adjustment signal.

[0107] A Fast Fourier Transform (FFT) operation is performed on the transient response waveform to transform it from the time domain to the frequency domain and extract the amplitude and phase spectrum. The phase change amplitude from zero frequency to the cutoff frequency is calculated based on the frequency domain amplitude and phase spectrum, compared to the reference phase difference under open-loop conditions, to obtain the phase margin improvement rate. The formula is as follows: in, To apply the phase margin of the system after trial adjustment, This represents the phase margin under the baseline conditions.

[0108] An exponential decay curve is fitted to the envelope of the transient response waveform, and the decay constant is solved using the least squares method. The difference between the reference settling time and the current decay time is calculated to obtain the gain settling time reduction rate, as shown in the following formula: in, The gain settling time under the reference condition. This is the gain settling time for the current response.

[0109] By using the frequency domain phase calculation and time domain attenuation fitting described above, the transient dynamic characteristics of the trial adjustment phase are quantified into the phase margin improvement rate and the gain stabilization time reduction rate, thereby enabling rapid performance evaluation of the adaptability of the new PID parameter set within the migration buffer window.

[0110] For example, during the cross-cluster migration buffer period of the ternary lithium black powder wet metallurgical co-extraction process, the feed pump reference flow rate was set to 3.5 L / min, the trial adjustment amplitude was set to 0.035 L / min, and the duration was 200 ms; the mass spectrometer sampling frequency was set to 100 Hz and synchronized with the control signal generation unit to obtain the transient waveform of the nickel-cobalt-manganese concentration ratio; after performing FFT calculation, the phase spectrum was intercepted up to the cutoff frequency of 2 Hz, and the measured values ​​were obtained. , Substituting this into the phase margin boost formula yields a boost value of 3°; the baseline settling time is obtained from the envelope fitting. Current decay time Substituting the values ​​into the formula for the reduction rate of gain settling time, we obtain a reduction rate of 0.08. This result indicates that the new PID parameter set significantly improves the phase margin and shortens the settling time during the buffer period, which helps to facilitate a smooth transition in the cross-cluster migration process. The final output phase margin improvement rate and gain settling time reduction rate are stored in a temporary evaluation register for subsequent control strategy switching decisions.

[0111] S4.4: The phase margin improvement rate and the gain settling time reduction rate are vector-concatenated to construct a disturbance response quality factor characterizing the dynamic performance of the current transient process, and the disturbance response quality factor is written into a temporary evaluation register for subsequent control strategy switching decision module to call.

[0112] After receiving the phase margin improvement rate generated by frequency domain phase calculation within the migration buffer window and the gain stabilization time reduction rate generated by envelope decay time fitting, the two indicators are treated as independent scalar inputs and parallel normalization is performed. The range normalization algorithm is used to perform a dimensionless mapping from zero to one on their respective numerical ranges to ensure that the numerical scales of different physical quantities are consistent during subsequent vector concatenation.

[0113] After completing the normalization mapping of the two scalar outputs, the phase margin boost rate normalized value and the gain settling time reduction rate normalized value are concatenated in a fixed sequence by the multidimensional vector constructor to form a two-dimensional dynamic performance feature vector. The first component of this vector corresponds to the phase margin boost rate, and the second component corresponds to the gain settling time reduction rate.

[0114] During the vector splicing process, the consistency verification module is called to check the structural integrity, numerical validity, and component order of the spliced ​​data. It also automatically triggers re-acquisition of signals for any missing or out-of-bounds anomalies to ensure the reliability of the disturbance response quality factor construction.

[0115] The verified two-dimensional dynamic performance feature vector is assigned a unique disturbance response quality factor identifier and written to a temporary evaluation register. The register structure adopts a multi-slot caching mechanism sorted by timestamp to support the control strategy switching decision module to obtain the latest and valid quality factor data in subsequent calls.

[0116] Through the above multi-stage processing method, the exploratory adjustment response data within the buffer window period is transformed into a disturbance response quality factor that can be directly quantified to evaluate the transition performance, thereby achieving standardized recording and real-time storage of the dynamic performance of cross-cluster migration.

[0117] For example, within a single cross-cluster migration buffer window, the frequency domain phase calculation yields a phase margin improvement rate of 18.5, and the envelope decay time fitting yields a gain settling time reduction rate of 12.7. Using the range normalization algorithm, assuming the historical data phase margin improvement rate ranges from 10.0 to 20.0 and the gain settling time reduction rate ranges from 8.0 to 15.0, the normalization calculation formula is as follows: ,in This is the current indicator value. This is the historical minimum. This is the historical maximum value. The normalized value for calculating the phase margin boost rate is... The normalized value of the gain settling time reduction rate is The two normalized values ​​are concatenated in order to form the disturbance response quality factor vector. The identifier RFQ-20240615-001 is written to a temporary evaluation register. In subsequent calls, the control strategy switching decision module directly reads this quality factor and, in conjunction with preset convergence conditions, completes the adaptation determination of the new control parameter group. Execution results show that, in this scenario, the first component of the quality factor vector is high, the second component is moderate, and the overall convergence index is significantly improved, satisfying the parameter update allowance conditions and achieving a smooth transition and dynamic performance guarantee during cross-cluster migration.

[0118] Step S5: Based on the numerical performance of the disturbance response quality factor, a control strategy switching decision is executed. If the phase margin improvement rate and the gain settling time reduction rate meet the preset convergence conditions, the control parameters are updated to the pre-calibrated PID parameter group corresponding to the new semantic cluster; otherwise, the original control strategy is maintained and the migration buffer window duration is extended, and the final closed-loop stability control command is output. Specifically, this includes: S5.1: Obtain the disturbance response quality factor data collected within the migration buffer window, and normalize the phase margin improvement rate and gain settling time reduction rate in the disturbance response quality factor to generate a standardized dynamic performance evaluation vector as the basic input for subsequent threshold comparison.

[0119] S5.2: Based on the standardized dynamic performance evaluation vector, the comprehensive convergence index under the current working condition is calculated using a preset multidimensional convergence discrimination algorithm to generate a convergence judgment index that characterizes the degree of adaptation of the pre-calibrated PID parameter group of the new semantic cluster.

[0120] Based on the standardized dynamic performance evaluation vector obtained through normalization within the migration buffer window, a preset multidimensional convergence discrimination algorithm input module is invoked to map the vector content to the parameter positions of the discrimination formula. The evaluation vector is then multiplied by its components and weight coefficients, and the results are accumulated. A preset normalization coefficient is used to scale the accumulated value to eliminate dimensional differences between different performance factors. Nonlinear smoothing is applied to the adjusted weighted composite value, and a quadratic curve fitting algorithm is used to weaken the impact of extreme high or low values ​​on the overall convergence metric. An exponential decay factor correction is applied to the smoothed result, and stability trend information from the operating condition time series is introduced to form a time-sensitive comprehensive convergence metric. Based on this metric, the comprehensive convergence index is calculated, and the calculation process can be defined by the following formula: in, Let be the weighting coefficient of the i-th performance factor. This represents the normalized value of the i-th performance factor. The number of performance factors is denoted by . The calculated comprehensive convergence index is used as a convergence criterion and transmitted to the control strategy switching decision module to achieve a quantitative assessment of the adaptation degree of the pre-calibrated PID parameter group for the new semantic cluster. By combining a multi-dimensional convergence discrimination method with weight and trend correction, the standardized dynamic performance evaluation vector from the previous step is transformed into a comprehensive convergence index, thereby achieving a stable assessment of the adaptation performance of the control strategy for the new semantic cluster.

[0121] For example, during the cross-cluster migration buffer window of the ternary lithium black powder hydrometallurgical process, the normalized phase margin improvement rate was 0.42, the gain stabilization time reduction rate was 0.36, and the number of performance factors was 2. The preset weighting coefficients were 0.6 for the phase margin improvement rate and 0.4 for the gain stabilization time reduction rate. Multiplying each component by the weighting coefficient yielded 0.252 and 0.144, respectively, and summing them resulted in 0.396. The sum was adjusted using a normalization coefficient of 0.95, yielding 0.3762. A quadratic curve fitting smoothing process was performed on the result, outputting a smoothed value of 0.375. An exponential decay factor of 0.98 was introduced for time-series trend correction, ultimately yielding a comprehensive convergence index of 0.3675, which was used as the convergence criterion and passed to the switching decision module. In the test scenario, after comparing this index with the preset convergence threshold, it significantly improved the accuracy of the new semantic cluster PID parameter group adaptation determination and maintained stable system dynamic performance during cross-cluster migration.

[0122] S5.3: Perform a logical comparison operation between the convergence determination index and the preset convergence condition threshold built into the system to generate a control strategy switching decision signal that includes a parameter update permission flag or a strategy maintenance flag.

[0123] Based on the convergence judgment index data generated in the previous steps, it is used as the core input signal for the current control strategy switching decision, and the system's built-in preset convergence condition threshold vector is loaded to ensure the consistency of the comparison benchmark.

[0124] Logical comparison operations are performed on the convergence judgment index and the threshold vector corresponding to the elements. A Boolean condition criterion that includes both greater than or equal to the judgment and parameter group adaptation priority is adopted to comprehensively evaluate the adaptability of the new semantic cluster pre-calibrated PID parameter group under the current working conditions.

[0125] During the logical comparison process, component-level judgment is performed on the convergence judgment vector containing multiple performance indicators. First, the phase margin improvement rate component is compared with a threshold to determine whether it meets the minimum requirements for system stability improvement. Then, the gain settling time reduction rate component is compared with a threshold to check whether it is within the allowable dynamic response optimization range.

[0126] When both performance components meet the corresponding threshold conditions, the synthesis parameters update the allow flag; if either component is below the threshold, the synthesis strategy maintains the flag.

[0127] The above synthesis results are packaged with convergence criteria into a control strategy switching decision signal, which includes a flag field, a decision timestamp, and an operation permission identifier for subsequent execution.

[0128] By using logical comparison operations and flag generation processing, the convergence judgment result of the previous step is transformed into a control strategy switching decision signal that can directly drive the actuator, enabling the system to quickly and accurately update control parameters in cross-cluster migration scenarios.

[0129] S5.4: Based on the parameter update permission flag in the control strategy switching decision signal, perform a hot replacement operation of control parameters to seamlessly update the original control strategy parameters in the currently running controller to the pre-calibrated PID parameter group corresponding to the new semantic cluster, so as to generate a real-time closed-loop control parameter set with adaptability to new operating conditions.

[0130] The parameter update permission flag in the control strategy switching decision signal is used as the trigger input condition for performing the hot replacement operation.

[0131] The original control strategy parameter set in the memory of the currently running controller is called and the pre-calibrated PID parameter set corresponding to the new semantic cluster is loaded. A dual-parameter set mapping table is constructed to match the parameter position index of each control link.

[0132] By comparing the parameter values ​​in the dual-parameter mapping table, numerical differences and sign differences are identified, and a partition-weighted update strategy is applied to the differences, prioritizing the update of key parameters with large changes to suppress transient oscillations.

[0133] Closed-loop error correction is performed on the updated parameters using real-time acquired operating condition feedback signals, employing a gain adjustment formula. in The current control gain, For the corrected gain, For closed-loop response error, For error sensitivity coefficient, The sampling period is To ensure the smoothing coefficient, parameters are made stable rapidly during the heat exchange process.

[0134] After the entire parameter set is updated and the closed-loop error is controlled within the allowable range, the updated PID parameter set is written into the controller's execution register to form a real-time closed-loop control parameter set that is adaptable to new operating conditions.

[0135] By using partitioned weighted updates and closed-loop error correction, the decision signal from the previous step is transformed into real-time control parameters adapted to the new operating conditions, thereby achieving smooth parameter switching and stability maintenance in cross-cluster migration scenarios.

[0136] For example, in the cross-cluster migration scenario during the lithium extraction and co-extraction process of ternary lithium black powder, the parameter update permission flag of the control strategy switching decision signal triggers the execution of the hot replacement module. The original PID parameter set of the currently running controller is P=1.8, I=0.12, D=0.05, and the new semantic cluster pre-calibration PID parameter set is P=2.0, I=0.15, D=0.04. The dual-parameter set mapping table shows that P and I have significant differences, while D has a smaller difference. The system identifies P and I as high-impact factors, updates them first, and collects the disturbance response waveform to calculate the closed-loop error e=0.08, with a sampling period T=1.0 second, an error sensitivity coefficient α=0.5, and a smoothing coefficient β=0.1, using the formula... The calculated gain of the updated P channel is approximately 1.8 + 0.03636 ≈ 1.83636. The PID parameter set, after closed-loop error correction, takes effect in the execution register. The interfacial tension fluctuations within the operating condition are significantly reduced, and the nickel-cobalt-manganese concentration ratio stabilization time is shortened, indicating that the parameter set after thermal replacement has good adaptability to the new operating conditions.

[0137] S5.5: Based on the real-time closed-loop control parameter set or the delayed waiting instruction generated when the control strategy switching decision signal is the strategy maintenance flag, drive the actuator to output the final closed-loop stability control instruction to complete the stability closed-loop control of the ternary lithium black powder wet metallurgy process in the cross-cluster migration scenario.

[0138] Secondly, a process optimization system for selective lithium extraction and nickel-cobalt-manganese co-extraction of ternary lithium black powder is provided, comprising: a module for acquiring the original signal set of the ternary lithium black powder hydrometallurgical process, a module for acquiring the six-dimensional operating condition input vector, a module for acquiring the non-parametric process state semantic fingerprint, a module for acquiring the disturbance response quality factor, and a module for acquiring the closed-loop stability control command, wherein: The ternary lithium black powder hydrometallurgical process raw signal set acquisition module is used to deploy sensing devices in the interface area between the lithium extraction section and the co-extraction section, synchronously collect raw indicators reflecting the changes in the dynamics of the two-phase interface and the liquid phase chemical environment, and generate the raw signal set of the ternary lithium black powder hydrometallurgical process. Six-dimensional operating condition input vector acquisition module: It is used to extract the energy entropy spectrum distribution as a frequency domain feature based on the original signal set of the ternary lithium black powder wet metallurgy process, and to fuse the frequency domain feature with real-time flow, temperature and redox potential data to construct a six-dimensional operating condition input vector. Non-parametric process state semantic fingerprint acquisition module: Based on multiple preset semantic cluster centers, the six-dimensional working condition input vector is input into a graph convolutional network, the distance weighting value of the current working condition in each semantic cluster center is calculated, and a non-parametric process state semantic fingerprint is generated. The disturbance response quality factor acquisition module is used to start a migration buffer window in the overlapping area of ​​the old and new semantic clusters if the non-parametric process state semantic fingerprint migrates across clusters. While maintaining the original control strategy, it loads the pre-calibrated PID parameter group corresponding to the new semantic cluster in parallel to generate a disturbance response quality factor that includes the phase margin improvement rate and the gain settling time reduction rate. The closed-loop stability control command acquisition module is used to update the control parameters to the pre-calibrated PID parameter group corresponding to the new semantic cluster if the phase margin improvement rate and the gain stabilization time reduction rate meet the preset convergence conditions; otherwise, it maintains the original control strategy and extends the migration buffer window duration, and outputs the final closed-loop stability control command.

[0139] This application provides a central air conditioning source-load coordinated control method, which can be applied to, for example... Figure 4 In the application environment shown, terminal 102 communicates with server 104 via a network. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices, and server 104 can be a standalone server or a server cluster consisting of multiple servers.

[0140] Thirdly, such as Figure 5 The diagram shows a computer device including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described in the above method embodiments.

[0141] Fourthly, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of any of the methods described in the above method embodiments.

[0142] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0143] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0144] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0145] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An optimized method for selective lithium extraction from ternary lithium black powder and nickel-cobalt-manganese co-extraction process, characterized in that, include: S1: Deploy sensing devices in the boundary area between the lithium extraction section and the co-extraction section to simultaneously collect raw indicators reflecting changes in the dynamics of the two-phase interface and the liquid phase chemical environment, and generate a raw signal set of the ternary lithium black powder wet metallurgical process. S2: Based on the original signal set of the ternary lithium black powder wet metallurgy process, the energy entropy spectrum distribution is extracted as a frequency domain feature, and the frequency domain feature is fused with real-time flow, temperature and redox potential data to construct a six-dimensional working condition input vector. S3: Based on multiple preset semantic cluster centers, the six-dimensional working condition input vector is input into a graph convolutional network to calculate the distance weighted value of the current working condition in each semantic cluster center, and generate a non-parametric process state semantic fingerprint. S4: If the non-parametric process state semantic fingerprint undergoes cross-cluster migration, a migration buffer window is activated in the overlapping area of ​​the old and new semantic clusters. While maintaining the original control strategy, the pre-calibrated PID parameter group corresponding to the new semantic cluster is loaded in parallel to generate a disturbance response quality factor that includes the phase margin improvement rate and the gain settling time reduction rate. S5: If the phase margin improvement rate and the gain stabilization time reduction rate meet the preset convergence conditions, the control parameters will be updated to the pre-calibrated PID parameter group corresponding to the new semantic cluster; otherwise, the original control strategy will be maintained and the migration buffer window duration will be extended, and the final closed-loop stability control command will be output.

2. The optimized method for selective lithium extraction and nickel-cobalt-manganese co-extraction process of ternary lithium black powder according to claim 1, characterized in that, The original indicators include: liquid phase viscosity change response time, interphase tension relaxation decay period, and pH jump front propagation speed.

3. The optimized method for selective lithium extraction and nickel-cobalt-manganese co-extraction process of ternary lithium black powder according to claim 2, characterized in that, S2 specifically includes: Wavelet packet decomposition was performed on the time series waveforms of liquid phase viscosity change response time, interphase tension relaxation decay period, and pH jump front propagation velocity, respectively. Orthogonal wavelet basis functions were used to map each time series waveform to a multi-resolution frequency band space to obtain a multi-level wavelet packet decomposition coefficient sequence. The energy proportion of each frequency band node is calculated based on the multi-level wavelet packet decomposition coefficient sequence, and Shannon entropy operation is performed on the energy proportion to generate an energy entropy spectrum distribution feature vector. The instantaneous values ​​of feed flow rate, reactor temperature, and redox potential are acquired in real time as static operating parameters. The static operating parameters are then time-stamped with the energy entropy spectrum distribution feature vector to generate a multi-source data set. Normalization scaling is performed on the multi-source data set to uniformly map the energy entropy spectrum distribution feature vector and the numerical range of each static operating condition parameter to a dimensionless interval from zero to one, thereby generating a standardized multidimensional data matrix. Based on the standardized multidimensional data matrix, a vector concatenation operation is performed according to the preset frequency domain feature priority order and process parameter association order to generate the six-dimensional operating condition input vector.

4. The optimized method for selective lithium extraction and nickel-cobalt-manganese co-extraction process of ternary lithium black powder according to claim 1, characterized in that, The semantic cluster center is a semantic cluster center defined by unsupervised clustering of typical operating conditions based on the offline stage. The typical operating conditions include: high lithium and low nickel, low acid and high manganese, and pulse feeding.

5. The optimized method for selective lithium extraction and nickel-cobalt-manganese co-extraction process of ternary lithium black powder according to claim 4, characterized in that, S3 specifically includes: Obtain the coordinate set of the semantic cluster centers, and construct a static process state topology graph structure containing node features and edge weights based on the coordinate set of the semantic cluster centers; The six-dimensional operating condition input vector is embedded into the static process state topology graph structure, and a graph convolutional network is used to perform neighborhood information aggregation and nonlinear feature transformation processing to output a high-dimensional latent feature representation vector of the current real-time operating condition in the implicit semantic space. Based on the high-dimensional latent feature representation vector and the coordinate set of the semantic cluster center, a multi-dimensional space Euclidean distance metric algorithm is executed to generate a non-parametric process state semantic fingerprint representing the degree to which the current operating state deviates from the typical operating condition mode, and the Euclidean distance scalar from the semantic cluster center. Normalized weighted fusion processing is performed on the Euclidean distance scalar, and the contribution weights of each semantic cluster are allocated according to the inverse distance relationship to calculate the comprehensive position coordinates of the current working condition relative to the typical working condition mode. A uniquely identifiable non-parametric process state semantic fingerprint is generated based on the comprehensive location coordinates.

6. The optimized method for selective lithium extraction and nickel-cobalt-manganese co-extraction process of ternary lithium black powder according to claim 1, characterized in that, S4 specifically includes: Continuous temporal trajectory monitoring is performed on the non-parametric process state semantic fingerprint. Based on the rate of change of the Euclidean distance between the fingerprint coordinates at the current sampling time and the historical trajectory points, if the non-parametric process state semantic fingerprint crosses a preset semantic cluster boundary threshold, a migration trigger signal is generated. Based on the migration trigger signal, the migration buffer window timer is activated in the feature space overlap region of the old semantic cluster center and the new semantic cluster center, the original control strategy parameter group currently being executed is locked, and the target pre-calibrated PID parameter group matching the new semantic cluster center is retrieved from the pre-calibration parameter library and loaded into the parallel standby state. The target pre-calibrated PID parameter set is used to apply a small-amplitude tentative adjustment to the system during the migration buffer window period, and the transient response waveform of the nickel-cobalt-manganese concentration ratio at the extract phase outlet is acquired simultaneously; The transient response waveform is subjected to frequency domain phase calculation to obtain the phase margin improvement rate, and the envelope of the transient response waveform is fitted with decay time to obtain the gain settling time reduction rate. The phase margin improvement rate and the gain settling time reduction rate are vector-concatenated to construct the disturbance response quality factor that characterizes the dynamic performance of the current transient process.

7. The optimized method for selective lithium extraction and nickel-cobalt-manganese co-extraction process of ternary lithium black powder according to claim 1, characterized in that, The sensing device includes a micro-rheological sensor array and a micro-differential pH probe array.

8. An optimized system for selective lithium extraction and nickel-cobalt-manganese co-extraction processes from ternary lithium black powder, characterized in that, include: The ternary lithium black powder hydrometallurgical process raw signal set acquisition module is used to deploy sensing devices in the interface area between the lithium extraction section and the co-extraction section, synchronously collect raw indicators reflecting the changes in the dynamics of the two-phase interface and the liquid phase chemical environment, and generate the raw signal set of the ternary lithium black powder hydrometallurgical process. Six-dimensional operating condition input vector acquisition module: It is used to extract the energy entropy spectrum distribution as a frequency domain feature based on the original signal set of the ternary lithium black powder wet metallurgy process, and to fuse the frequency domain feature with real-time flow, temperature and redox potential data to construct a six-dimensional operating condition input vector. Non-parametric process state semantic fingerprint acquisition module: Based on multiple preset semantic cluster centers, the six-dimensional working condition input vector is input into a graph convolutional network, the distance weighting value of the current working condition in each semantic cluster center is calculated, and a non-parametric process state semantic fingerprint is generated. The disturbance response quality factor acquisition module is used to start a migration buffer window in the overlapping area of ​​the old and new semantic clusters if the non-parametric process state semantic fingerprint migrates across clusters. While maintaining the original control strategy, it loads the pre-calibrated PID parameter group corresponding to the new semantic cluster in parallel to generate a disturbance response quality factor that includes the phase margin improvement rate and the gain settling time reduction rate. The closed-loop stability control command acquisition module is used to update the control parameters to the pre-calibrated PID parameter group corresponding to the new semantic cluster if the phase margin improvement rate and the gain stabilization time reduction rate meet the preset convergence conditions; otherwise, it maintains the original control strategy and extends the migration buffer window duration, and outputs the final closed-loop stability control command.

9. A computer device, characterized in that, include: A processor, a memory, and a computer program stored in the memory and executable on the processor; when executed by the processor, the computer program implements the steps of the optimized method for selective lithium extraction and nickel-cobalt-manganese co-extraction process of ternary lithium black powder as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for optimizing the selective lithium extraction and nickel-cobalt-manganese co-extraction process of ternary lithium black powder as described in any one of claims 1 to 7.

Images