Electrolyte acid-zinc specific concentration regulation and control method based on deep learning

Abstract

The invention discloses an electrolyte acid-zinc ratio concentration regulation and control method based on deep learning, and the method comprises the following steps: S1, collecting the operation data of an electrolyte system, and constructing an input data set; s2, inputting the input data set into the improved SSDNet model, and calculating a residual value; s3, identifying a variable type and a concentration response interval corresponding to the residual value; s4, constructing a dynamic weight sensitive input vector, and generating a differential weighted residual value by responding to a dynamic reconstruction function and weighting processing; s5, generating a state feature vector based on the state change rate, the conductivity change trend and the regulation and control frequency; adjusting and responding to parameter configuration of the dynamic reconstruction function, and updating a residual weighting strategy; s6, executing joint training operation; and S7, if the predicted concentration value is not within the set target range, outputting an acid liquid or zinc liquid injection instruction, and completing closed-loop regulation and control of the acid-zinc ratio concentration. According to the invention, high-precision prediction and flexible regulation and control of the acid-zinc specific concentration are realized, and the real-time performance and stability of concentration control are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent electrolyte concentration control technology, and in particular to a method for controlling the zinc-to-acid ratio concentration of electrolyte based on deep learning. Background Technology

[0002] With the continuous expansion of electrolytic metallurgy in fields such as non-ferrous metal purification and resource recycling, the precise control of the acid-zinc ratio concentration in the electrolyte has gradually become a key aspect in improving production capacity and product consistency. Currently, the injection of acid and zinc solutions into the electrolyte system mostly relies on empirically based mixing schemes or simple proportional control methods. These methods have weak responsiveness in the face of complex operating conditions or disturbed environments, and the control strategies lack sufficient adaptive capability and feedback accuracy.

[0003] Existing methods primarily rely on the deviation between a preset target range for the zinc acid-to-acid ratio and the measured value for manual or semi-automatic adjustment. However, in situations involving multivariate coupling, significant nonlinearity in concentration response, and frequent changes in conductivity and temperature, traditional methods often fail to accurately capture the source of residual changes, leading to problems such as misjudgment of variable attributes, lag in injection adjustment, or over-response. Furthermore, concentration prediction models are generally based on static modeling or shallow neural network structures, lacking the ability to dynamically model historical injection records, changes in operating status, and residual feedback, thus failing to establish a closed-loop adaptive control mechanism of prediction-response-correction.

[0004] Meanwhile, existing methods lack the ability to identify phases when facing rapid system state transitions, such as from the startup phase to the stable operation phase. This makes it difficult to adjust control strategies in a timely manner, which can easily lead to increased fluctuations and frequent interventions after the zinc acid-to-acid ratio concentration deviates from the target range.

[0005] Therefore, how to provide a method for regulating the zinc-to-acid ratio in electrolytes based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based method for regulating the zinc-to-acid ratio concentration in electrolytes. This invention details the algorithm structure and execution flow for intelligent prediction and closed-loop regulation of the zinc-to-acid ratio concentration using an improved SSDNet model. This invention constructs an input dataset by collecting operational data from the electrolyte system. It establishes a multi-path fusion and residual self-feedback mechanism using a drift coding module, a fusion construction module, a self-adjustment module, and a recursive feedback module to perform dynamic prediction of the zinc-to-acid ratio concentration and residual response modeling. Differential weighted residual values ​​are generated through a dynamic response reconstruction function, enabling partitioned weighted regulation of acid and zinc variables. The residual weight distribution strategy is automatically adjusted according to the system's operational phase. The model is then trained end-to-end using a joint loss function to jointly optimize the prediction path and residual feedback path. Finally, intelligent prediction and automatic electrolyte injection control of the zinc-to-acid ratio concentration are achieved in a real electrolyte system, forming a closed-loop adaptive regulation process. This invention possesses advantages such as comprehensive data acquisition, a reasonable model structure, high prediction accuracy, strong response sensitivity, and good control stability.

[0007] A method for regulating the zinc-to-acid ratio of an electrolyte based on deep learning, according to an embodiment of the present invention, includes the following steps: S1. Collect the operating data of the electrolyte system and construct the input dataset; the operating data includes acid concentration, zinc concentration, conductivity, current density, temperature and historical electrolyte injection records; S2. Input the input dataset into the improved SSDNet model, perform the zinc acid ratio concentration prediction operation, output the predicted concentration value at the target time, and calculate the residual value between the predicted concentration value and the measured concentration value. S3. Identify the variable type and concentration response range corresponding to the residual value, classify the variable type into acid variable and zinc variable, and classify the concentration response range into high response range, medium response range and low response range. S4. Construct a dynamic weighted input vector and generate residual weighted values ​​by responding to a dynamic reconstruction function; perform weighted processing on the residual values ​​according to the residual weighted values ​​to generate differentiated weighted residual values; the dynamic weighted input vector includes variable type, concentration response interval, residual change rate, historical residual sequence and operating characteristics; S5. Based on the state change rate, conductivity change trend and control frequency in the previous sampling period, generate a state feature vector and determine the operating stage of the electrolyte system; adjust the parameter configuration of the response dynamic reconstruction function according to the operating stage and update the residual weighting strategy; the operating stage includes the startup stage, the stable stage, the rapid adjustment stage and the decay stage. S6. Feed the differentiated weighted residual value as a loss signal to the improved SSDNet model, perform joint training operation, and realize the collaborative optimization of zinc acid ratio concentration prediction and response dynamic reconstruction function. S7. Deploy the trained improved SSDNet model to the electrolyte system; if the predicted concentration value is not within the set target range, output an acid or zinc injection command to complete the closed-loop control of the acid-zinc ratio concentration.

[0008] Preferably, S1 specifically comprises: Operational data is collected using an acid concentration sensor, a zinc concentration sensor, a conductivity detection device, a current density detection device, and a temperature detection device. The operational data includes acid concentration, zinc concentration, conductivity, current density, and temperature. The data on the injection volume of acid and zinc solution recorded in the injection control system is also read. The running data is time aligned to unify all data sequences to the same sampling period and time index; the time-aligned data is then filled with missing data by linear interpolation to fill in abnormal discontinuities; and the filled data is then normalized by linearly scaling each data sequence according to a preset range. The acid concentration sequence, zinc concentration sequence, conductivity sequence, current density sequence, temperature sequence, and historical injection record sequence are concatenated to form an input dataset in continuous time series format.

[0009] Preferably, the improved SSDNet model includes a drift coding module, a fusion construction module, a self-adjusting module, and a recursive feedback module, specifically: The drift coding module performs feature extraction operations on acid concentration, zinc concentration, conductivity, current density, temperature and historical injection records, constructs a time drift matrix, models the time drift of the feature weights of various variables, and outputs a time-series drift feature vector. The fusion construction module includes a trend path and a periodic path. The trend path extracts long-term trend features based on a state-space modeling structure, while the periodic path extracts periodic perturbation features based on a multi-scale convolutional structure. The fusion construction module constructs an interaction matrix based on the output information of the trend path and the periodic path, performs an interactive fusion operation on the trend features and the periodic features, and generates a fused feature vector. The self-adjustment module receives the features of the operation phase, constructs a gated adjustment structure, generates an adjustment weight matrix based on the operation phase, adjusts the output amplitude and path channel ratio of the fusion construction module, and outputs the phase adjustment feature vector. The recursive feedback module constructs a recursive residual signal based on the differential weighted residual value, combines the historical residual signal with the current residual signal to generate a feedback vector, inputs the feedback vector to the drift coding module and the fusion construction module, and participates in the parameter update operation during the training process as a loss feedback path, thus completing the collaborative optimization of the concentration prediction path and the residual feedback path.

[0010] Preferably, S2 specifically includes: The input dataset is fed into the drift coding module to generate a time-drift feature vector. The time-series drift feature vector is input into the fusion construction module, where trend features are extracted from the trend path and periodic features are extracted from the periodic path. Based on the interaction matrix in the fusion construction module, the fusion operation of trend features and periodic features is performed to generate a fused feature vector. The fused feature vector is input into the self-adjustment module, which generates an adjustment weight matrix according to the running stage, performs channel weighting on the fused feature vector, and generates a stage-adjusted feature vector. Perform a fully connected computation on the stage-adjusted feature vector to generate the predicted concentration value at the target time. The predicted concentration values ​​are matched one-to-one with the collected measured concentration values. Time alignment is performed according to the set sampling index matching rules. The difference at each sampling point is calculated to generate the corresponding residual sequence.

[0011] Preferably, S3 specifically includes: Based on the index position information of each residual value in the residual sequence, the input variable type corresponding to the residual value is determined; if the residual value comes from the acid concentration data field, it is marked as an acid variable; if the residual value comes from the zinc concentration data field, it is marked as a zinc variable; the variable type is classified according to the source of the variable. For each residual value corresponding to a time index, calculate the rate of change of conductivity between adjacent sampling points before and after the current time point, and extract the conductivity change trend index; retrieve multiple residual values ​​adjacent to the current time index in the historical residual sequence, and calculate the local fluctuation amplitude index; combine the conductivity change trend index and the residual fluctuation amplitude index with weights to construct the concentration response index value. The concentration response index value is compared with the set response interval division threshold. If the concentration response index value is greater than the high response threshold, it is marked as a high response zone; if the concentration response index value is between the high response threshold and the low response threshold, it is marked as a medium response zone; if the concentration response index value is less than the low response threshold, it is marked as a low response zone. The concentration response interval is divided according to the concentration response index value. The variable type label corresponding to each residual value is merged with the concentration response interval label to generate a joint label result.

[0012] Preferably, S4 specifically comprises: The variable type, concentration response range, residual change rate, historical residual sequence, and operational features are concatenated to construct a dynamic weighted input vector; the residual change rate is the difference result between adjacent residual values, the historical residual sequence is the set of residual values ​​of multiple sampling points before the current residual value time index, and the operational features include conductivity change trend, current density fluctuation amplitude, and temperature offset. The dynamic weighted input vector is input to the response dynamic reconstruction function to generate residual weighted values. The response dynamic reconstruction function includes an input mapping structure, a feature association structure, and a weight generation structure. The input mapping structure performs a normalization transformation on the dynamic weighted input vector to generate a standardized feature tensor. The feature association structure calculates the combination relationship between variable type, response interval, and running features based on the standardized feature tensor and outputs the associated feature results. The weight generation structure performs a nonlinear function operation on the associated feature results and outputs the residual weighted values ​​corresponding to the residual values. The residuals are weighted according to their weighted values. Weighted multiplication is performed on the residuals corresponding to acid and zinc variables respectively. When the variable type is acid and the concentration response range is in the high response range, the residual weighting value adopts a high weighting factor. When the variable type is acid and the concentration response range is in the low response range, the residual weighting value adopts a low weighting factor. When the variable type is zinc and the concentration response range is in the medium response range, the residual weighting value adopts a medium weighting factor. All weighted residuals are then combined to form a differentiated weighted residual sequence.

[0013] Preferably, the step of generating a state feature vector based on the rate of change of state, the trend of change of conductivity, and the control frequency in the previous sampling period, and determining the operating stage of the electrolyte system, specifically involves: Based on the acid concentration, zinc concentration, conductivity, current density, and temperature data recorded in the previous sampling period, the rate of state change is calculated. The rate of state change is obtained by the numerical difference results of adjacent sampling points. The trend of conductivity change in multiple consecutive sampling points is calculated as the trend slope value. The control frequency is calculated according to the ratio of the number of injection command triggers to the sampling period. The rate of state change, the conductivity change trend, and the control frequency are concatenated to construct a state feature vector. The state feature vector is input into the stage determination structure. The stage determination structure includes a feature extraction unit, a threshold mapping unit, and a stage output unit. The feature extraction unit performs standardized calculations on the state feature vector to generate a feature index sequence. The threshold mapping unit calculates the stage matching degree based on the feature index sequence. The stage output unit determines the operating stage label based on the stage matching degree. When both the state change rate and the control frequency are higher than the stage change threshold, it is determined to be a rapid adjustment stage. When the state change rate is stable and the conductivity change trend is within the threshold range, it is determined to be a stable stage. When the control frequency is low and the conductivity change trend is decreasing, it is determined to be a decay stage. When both the state change rate and the conductivity change trend are increasing, it is determined to be a startup stage.

[0014] Preferably, the step of adjusting the parameter configuration of the dynamic reconstruction function based on the operational phase and updating the residual weighting strategy specifically involves: Select the corresponding parameter configuration set according to the running phase label, and adjust the weight generation structure parameters in the response dynamic reconstruction function; when the running phase is the fast adjustment phase, increase the high weight factor in the weight generation structure; when the running phase is the stable phase, maintain the equilibrium factor in the weight generation structure; when the running phase is the decay phase, decrease the residual response sensitivity coefficient in the weight generation structure; when the running phase is the startup phase, enable the initial bias coefficient in the weight generation structure. After updating the residual weighting strategy, the residual weight values ​​are recalculated, and the weight distribution of the weighted residual sequence is adjusted to complete the process of identifying and dynamically adjusting residual weights during the operation phase.

[0015] Preferably, S6 specifically includes: The differentially weighted residual values ​​are input into the recursive feedback module to construct the loss feedback vector. The recursive feedback module includes a residual embedding structure, a temporal memory structure, and a residual fusion structure. The residual embedding structure performs vector encoding on the differentially weighted residual values ​​to generate a residual embedding representation. The temporal memory structure combines the residual embedding representation with historical residual signals to extract temporal recursive features. The residual fusion structure performs fusion operations on the temporal recursive features and the output features of the drift encoding module to generate a joint feedback tensor. The joint feedback tensor is used as the training signal input to the drift coding module, fusion construction module, and self-adjustment module to construct a joint loss function during end-to-end training. The joint loss function includes a prediction error term and a weighted residual penalty term. The prediction error term is calculated based on the mean square error between the predicted concentration value and the measured concentration value, and the weighted residual penalty term is calculated based on the degree of interval offset of the differentiated weighted residual values. By performing backpropagation training on the model parameters using a joint loss function, the time drift parameters in the drift coding module, the path weight parameters in the fusion construction module, the adjustment matrix parameters in the self-adjustment module, and the weight generation parameters in the response dynamic reconstruction function are updated respectively. This achieves the coordinated updating of the concentration prediction path and the residual control path, completing the joint training process for the acid-zinc ratio concentration prediction task and the residual response modeling task.

[0016] Preferably, S7 specifically includes: The trained improved SSDNet model is deployed to the data control unit of the electrolyte system. The newly constructed input dataset is input in each sampling period, the zinc acid ratio concentration prediction operation is performed, and the target predicted concentration value is output. The target predicted concentration value is compared with the preset target range of acid-zinc ratio concentration to determine whether the target predicted concentration value is within the target range boundary. If the target predicted concentration value is less than the lower limit, an acid injection command is generated; if the target predicted concentration value is greater than the upper limit, a zinc injection command is generated. The injection command includes the target injection channel identifier, the injecting agent type identifier, the control flow parameters, and the injection duration. Based on the injection command, the liquid injection control system is driven to complete the acid or zinc injection process and record the injection volume data to the liquid injection control recording module. After the injection is completed, acid and zinc concentration monitoring data are collected to construct an updated input dataset. In the next sampling cycle, the improved SSDNet model is called to perform concentration prediction and deviation judgment operations to complete the closed-loop control of acid-zinc ratio concentration.

[0017] The beneficial effects of this invention are: This invention addresses the problems of large prediction errors, sluggish control response, and poor adaptability during operation in electrolyte systems by constructing an improved SSDNet model that integrates time drift modeling, trend-period interaction paths, self-regulation mechanisms, and recursive feedback structures for acid-zinc ratio concentration regulation. It constructs a standard input dataset by collecting acid concentration, zinc concentration, conductivity, current density, temperature, and historical injection records. A time-series drift feature vector is generated based on a drift encoding module, and interactive fusion is performed using trend and periodic paths from the fusion construction module. Furthermore, a self-regulation module driven by the operation phase is introduced to generate phase-specific regulation feature vectors, achieving high-precision prediction of the acid-zinc ratio concentration. Regarding residual modeling, this invention… A dynamic weighted input vector is constructed by introducing a response dynamic reconstruction function. This vector is then combined with variable type, concentration response range, and operational characteristics to perform residual weighting, generating differentiated weighted residual values. The weight generation strategy is dynamically adjusted according to the system's operational stage to achieve adaptive compensation for concentration prediction errors. During model training, a recursive feedback module integrates current and historical residual signals, constructs a joint loss function, and performs end-to-end collaborative training to ensure joint optimization of the prediction path and the residual feedback path. Finally, the trained improved SSDNet model is deployed to the electrolyte system. Closed-loop control logic outputs acid or zinc injection commands to precisely regulate the acid-zinc ratio concentration, improving concentration stability and system control response speed. This invention achieves the fusion optimization of concentration prediction and feedback control, possessing advantages such as high prediction accuracy, fast response speed, strong adaptability, and high control stability. It effectively solves the key problems of large control offsets and adjustment lags in traditional methods under dynamic operating conditions. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is a flowchart of a deep learning-based method for regulating the zinc-to-acid ratio in an electrolyte, as proposed in this invention. Figure 2 This is a schematic diagram of a deep learning-based method for regulating the zinc-to-acid ratio in an electrolyte, as proposed in this invention. Figure 3 This is a data flow diagram of a deep learning-based method for regulating the zinc-to-acid ratio in an electrolyte, as proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1-3 A deep learning-based method for regulating the zinc-to-acid ratio in an electrolyte comprises the following steps: S1. Collect the operating data of the electrolyte system and construct the input dataset; the operating data includes acid concentration, zinc concentration, conductivity, current density, temperature and historical electrolyte injection records; S2. Input the input dataset into the improved SSDNet model, perform the zinc acid ratio concentration prediction operation, output the predicted concentration value at the target time, and calculate the residual value between the predicted concentration value and the measured concentration value. S3. Identify the variable type and concentration response range corresponding to the residual value, classify the variable type into acid variable and zinc variable, and classify the concentration response range into high response range, medium response range and low response range. S4. Construct a dynamic weighted input vector and generate residual weighted values ​​by responding to a dynamic reconstruction function; perform weighted processing on the residual values ​​according to the residual weighted values ​​to generate differentiated weighted residual values; the dynamic weighted input vector includes variable type, concentration response interval, residual change rate, historical residual sequence and operating characteristics; S5. Based on the state change rate, conductivity change trend and control frequency in the previous sampling period, generate a state feature vector and determine the operating stage of the electrolyte system; adjust the parameter configuration of the response dynamic reconstruction function according to the operating stage and update the residual weighting strategy; the operating stage includes the startup stage, the stable stage, the rapid adjustment stage and the decay stage. S6. Feed the differentiated weighted residual value as a loss signal to the improved SSDNet model, perform joint training operation, and realize the collaborative optimization of zinc acid ratio concentration prediction and response dynamic reconstruction function. S7. Deploy the trained improved SSDNet model to the electrolyte system; if the predicted concentration value is not within the set target range, output an acid or zinc injection command to complete the closed-loop control of the acid-zinc ratio concentration.

[0022] This implementation provides a deep learning-based method for regulating the acid-zinc ratio concentration of an electrolyte. By collecting data on the acid concentration, zinc concentration, conductivity, current density, temperature, and historical injection records of the electrolyte system, a comprehensive input dataset characterizing the system state is constructed, which improves the model's accuracy in perceiving system dynamics. The input dataset is then fed into an improved SSDNet model, which outputs predicted concentration values ​​and calculates the residual between the predicted and measured concentration values, enabling real-time identification of concentration deviations. Furthermore, by identifying the variable type and concentration response interval corresponding to the residual values, and classifying the variable type into acid and zinc variables, and the concentration response interval into high, medium, and low response regions, the ability to characterize the influence weights of variables and concentration sensitivity is enhanced. Further, a dynamic weighted input vector containing variable type, response interval, residual change rate, historical residual sequence, and operational characteristics is constructed, and a dynamic response reconstruction function is called to generate residuals. Weighting the residuals with differentiated weighting enhances the model's adaptability to complex and fluctuating conditions. Simultaneously, a state feature vector is constructed based on the rate of change of state, conductivity trends, and control frequency to determine the operating stage of the electrolyte system. This allows for adjustments to the parameters of the dynamic reconstruction function and updates to the residual weighting strategy, enabling the residual processing strategy to dynamically evolve with the system state and enhancing the flexibility and stability of the control logic. Furthermore, the differentiated weighted residuals are fed back as a loss signal to the improved SSDNet model for joint training of concentration prediction and response modeling. This achieves synergistic optimization of the residual mechanism and the main prediction path, improving the overall prediction accuracy and convergence stability of the model. Finally, the trained model is deployed to the electrolyte system, automatically outputting acid or zinc injection commands when the predicted concentration deviates from the set target range, forming a closed-loop control mechanism. This improves the dynamic control efficiency of the acid-zinc ratio and the concentration stability during operation.

[0023] In this embodiment, S1 specifically refers to: Operational data is collected using an acid concentration sensor, a zinc concentration sensor, a conductivity detection device, a current density detection device, and a temperature detection device, generating acid concentration sequences, zinc concentration sequences, conductivity sequences, current density sequences, and temperature sequences, respectively. The acid and zinc injection volumes recorded in the injection control system are read to construct an injection record sequence. The above data together constitute the original operational data set within the sampling period. Time alignment processing is performed on the original set of operating data to construct a unified time index set, and interpolation mapping is performed on different types of data sequences according to the unified time index so that all data sequences have complete values ​​at the same sampling time point; the sampling period length and time index interval are set according to the system sampling frequency, and the set values ​​are obtained by fitting the data fluctuation range under steady state of the electrolysis system based on an empirical model to meet the dynamic response characteristics of concentration change; Missing data is filled in the time-aligned data sequence to identify null values ​​or non-numeric outliers. Linear interpolation is then performed using the nearest valid data points to fill in the gaps. The interpolation function is a two-point-one-line model to ensure data continuity and trend consistency. Normalization is performed on all the completed data sequences. A normalization interval is constructed based on the historical maximum and minimum values. Each value is linearly mapped to the set range interval. The normalization function adopts the Min-Max scaling function, and the normalization range is set to [0, 1]. This operation makes variables under different units comparable and improves the model's sensitivity to different variables. The acid concentration sequence, zinc concentration sequence, conductivity sequence, current density sequence, temperature sequence, and injection record sequence are concatenated with the injection record sequence according to the time index to construct a multivariate, multi-time series input data tensor. Each time step in the input data tensor contains the normalized values ​​of six variables, forming an input dataset in continuous time series format.

[0024] In this embodiment, the improved SSDNet model includes a drift coding module, a fusion construction module, a self-adjusting module, and a recursive feedback module, specifically: The drift encoding module receives acid concentration sequences, zinc concentration sequences, conductivity sequences, current density sequences, temperature sequences, and injection record sequences, and constructs a multivariate input tensor. For each variable sequence, it performs time-step feature extraction, calling a one-dimensional convolutional structure to extract local change patterns and introducing a positional encoding mechanism to embed time index information, constructing a positional feature vector. A drift-aware matrix is ​​constructed in the time dimension. Based on the change trend and historical offset measure of each variable at different time steps, a drift factor matrix is ​​calculated. The drift factor is obtained by fitting the local change rate within the historical window to the global drift mean. The convolutional features and the drift factor matrix are fused to generate a temporal drift feature vector, which serves as the input for subsequent modules. The fusion construction module includes a trend path and a periodic path. The trend path constructs a dynamic trend estimator based on a state-space modeling structure. The dynamic trend estimator consists of a gated recursive unit and a moving average channel, which recursively calculates the temporal drift feature vector and outputs a long-term trend vector. The periodic path adopts a multi-scale convolutional structure to extract local features from the input sequence at different convolutional kernel scales and outputs periodic perturbation features. An interaction matrix is ​​constructed, and the trend vector and periodic perturbation features are interactively fused through a dot product attention mechanism. The fusion result is input into a fully connected structure to generate a fused feature vector. The self-adjustment module receives the operation stage label information generated by the external state determination path, constructs a gating adjustment structure, which includes a channel activation submodule and an amplitude control submodule. It selects the activation channel combination and amplitude scaling factor according to different operation stages. Based on the preset adjustment weight matrix mapped by the operation stage label, it performs channel dimension weight adjustment operation on the fused feature vector and outputs the stage adjustment feature vector. This feature vector dynamically reflects the response structure changes of the system under different operation stages. The recursive feedback module receives differentiated weighted residual values ​​and constructs a recursive residual signal sequence. A recurrent neural network structure is used to fuse the current residual and historical residual sequences to extract residual evolution features. The residual evolution features are concatenated with the original features to construct a feedback vector. The feedback vector is input into the intermediate layer structure between the drift coding module and the fusion construction module as an auxiliary training signal to influence weight updates. At the same time, the feedback vector participates in the construction of the joint loss function, which includes a prediction error loss term and a residual collaborative loss term. The joint optimization objective is generated through a residual weighting mechanism.

[0025] This implementation method improves the model's ability to structurally model dynamic changes in concentration and optimize long-term error feedback by constructing a time-drift sensing structure, a trend-period dual-channel fusion structure, a stage adaptive adjustment structure, and a recursive feedback training structure in the improved SSDNet model. This effectively enhances the timeliness, accuracy, and stability of concentration prediction.

[0026] In this embodiment, S2 specifically refers to: The input dataset is constructed into a unified input tensor in chronological order. The input tensor contains a normalized sequence of acid concentration, zinc concentration, conductivity, current density, temperature, and historical injection records. The input tensor is then fed into a drift encoding module, where one-dimensional convolution operations are performed in the channel dimension to extract local feature information. A drift perception matrix is ​​constructed in the time dimension, which is generated based on the average volatility and trend offset measure within a sliding time window. The convolutional features and the drift perception matrix are fused to form a temporal drift feature vector, which preserves the structural change trends of each variable across multiple time steps. The temporal drift feature vector is input into the fusion construction module. In the trend path, a gated recursive network is used to perform time state modeling on the long sequence and output trend features. In the periodic path, a multi-scale convolutional structure is introduced to extract periodic perturbation features with different convolutional kernel sizes. The trend features and periodic perturbation features are input into the interaction matrix structure. The interaction matrix calculates the correlation weight between channels through feature dot product, performs fusion calculation, and outputs a fused feature vector. The fused feature vector contains both long-period variation and short-period perturbation information. The fused feature vector is input into the self-adjustment module. The self-adjustment module receives the operation stage information, retrieves the adjustment weight configuration according to the stage label, and constructs the channel adjustment weight matrix and amplitude adjustment coefficient. The adjustment weight matrix is ​​applied to the channel dimension of the fused feature vector to perform weight scaling and path proportion adjustment on each channel, and outputs the stage adjustment feature vector. The stage adjustment feature vector reflects the change in response intensity of the fused feature under different operating states. The stage-adjusted feature vector is input into a fully connected computational structure, where feature compression and dimension mapping operations are performed. The predicted concentration value at the target time is output, and the predicted concentration value remains consistent with the input sampling index in the time dimension. The predicted concentration value is index-aligned with the collected measured concentration value. A residual index table is constructed based on the timestamp matching principle. The difference at each sampling point is calculated according to the residual index table, and the residual sequence is output.

[0027] In this embodiment, S3 specifically refers to: Based on the index position information of each residual value in the residual sequence, the input field identifier at the corresponding time index is extracted from the input dataset; the variable type of the residual value source is determined according to the field identifier; if the residual value comes from the acid concentration field, it is marked as an acid variable; if the residual value comes from the zinc concentration field, it is marked as a zinc variable; all residual values ​​are classified according to variable type, and a variable type label sequence is constructed. For each residual value corresponding to a time index, data from the same index and one sampling point before and after it in the conductivity sequence are called to calculate the conductivity change rate between adjacent time points, which is recorded as the conductivity change trend index. Residual values ​​from multiple sampling points before and after the time index in the residual sequence are called to construct a residual sliding window, and the standard deviation and maximum / minimum difference indices within the residual sliding window are calculated to extract the residual fluctuation amplitude. The conductivity change trend index and the residual fluctuation amplitude index are weighted and superimposed to construct the concentration response index value. The weighting coefficients are obtained by fitting the system's empirical model to the sample data. The fitting process uses the minimum mean square error principle to estimate the weighting ratio at different stages, generating a concentration response sensitivity combination function. The concentration response index value is compared with a set threshold for interval judgment. If the index value is greater than the high response threshold, it is marked as a high response zone; if the index value is between the high response threshold and the low response threshold, it is marked as a medium response zone; if the index value is less than the low response threshold, it is marked as a low response zone. A concentration response interval label sequence is constructed and concatenated with the variable type label sequence. In the time index dimension, a joint label sequence is constructed, where the label at each index consists of the variable type and the concentration response range. The joint label sequence serves as one of the structured inputs in the dynamic weighting mechanism and is used for weight configuration and sensitivity response modeling in the subsequent differential weighting process.

[0028] This implementation improves the accuracy and context-awareness of residual structure identification by constructing a joint labeling mechanism based on variable sources and response fluctuation characteristics. It provides quantifiable support for the construction of dynamic weighted input vectors and the parameter update of response functions, which helps to enhance the diversity and adaptability of the weighting mechanism.

[0029] In this embodiment, S4 specifically refers to: The variable type labels, concentration response interval labels, residual change rate, historical residual sequence, and operational features are encoded and standardized representation structures are constructed respectively. Specifically, the variable type labels are converted into binary feature vectors using one-hot encoding; the concentration response interval labels are embedded with multi-class labels based on three-class classifications; the residual change rate is obtained by calculating the difference between the current residual value and the residual value at the previous time point; the historical residual sequence consists of multiple residual values ​​prior to the current time index, forming a local residual window in chronological order; the operational features include conductivity change trend, current density fluctuation amplitude, and temperature offset, constructed using sliding standard deviation calculation, range calculation, and mean offset calculation methods, respectively; and the standardized feature vectors are concatenated to form a multi-dimensional feature tensor, which serves as the dynamic weighted input vector. The dynamic weight-sensitive input vector is input into the response dynamic reconstruction function, which includes an input mapping structure, a feature association structure, and a weight generation structure. The input mapping structure uses a linear normalization mechanism to stretch all numerical features to the range of [0, 1], generating a standardized feature tensor. The feature association structure uses a multilayer perceptron to construct a nonlinear combination mapping path, performing joint modeling operations on the cross features between variable types and concentration response ranges, and the degree of association between running features, outputting the associated feature results. The associated feature results are then extracted using a gated multiplication structure to extract the path combination most sensitive to the residual weights. The weight generation structure uses a two-layer activation network to perform nonlinear mapping, ultimately generating residual weights. The residual weights are decimals between [0, 1], representing the relative importance of the corresponding residual value in the overall weighted structure. Element-wise weighting is performed on the residual values ​​based on their weighted values. Each residual value is multiplied by its corresponding weighted value to generate a weighted residual value. Weighting factor boundary conditions are set according to the variable type and concentration response range. When the variable type is acid and the concentration response range is in the high response zone, the weighting guide function adopts an aggressive weighting strategy, outputting a weighting factor greater than 0.8. When the variable type is acid and the concentration response range is in the low response zone, a conservative weighting strategy is adopted, with a weighting factor less than 0.3. When the variable type is zinc and the concentration response range is in the medium response zone, a neutral weighting strategy is adopted, with a weighting factor approximately 0.5. The coefficients of these strategies are obtained through interval regression fitting using large-sample residual sensitivity analysis data. The fitting model uses minimizing the residual fluctuation loss function as the optimization objective. All weighted residual values ​​are reorganized according to time index order to form a differentiated weighted residual sequence, which serves as one of the input paths for the subsequent loss feedback structure.

[0030] This implementation constructs a dynamic weighted input vector that includes residual source features and operating state features, and introduces a residual weighting mechanism through a response dynamic reconstruction function. This achieves refined control of residual weights and enhanced feature perception, which helps improve the sensitivity of identifying abnormal biases and the adaptability of the model in the subsequent training stage.

[0031] In this embodiment, the step of generating a state feature vector based on the rate of change of state, the trend of change of conductivity, and the control frequency in the previous sampling period, and determining the operating stage of the electrolyte system, specifically involves: Based on the acid concentration, zinc concentration, conductivity, current density and temperature data recorded in the previous sampling period, the continuous values ​​of each variable in the sampling period are retrieved, the difference results between adjacent sampling points are calculated, and the state change rate vector is extracted; the difference values ​​of each variable in the state change rate vector are used to construct a state change index through the mean and range reduction method. The conductivity variable is sampled at multiple consecutive sampling points, and the slope of the linear trend of the time series is calculated based on the least squares fitting method to construct a conductivity change trend index. The system output injection control commands within the same sampling period are accumulated, and the total number of times the two types of commands, acid injection and zinc injection, are triggered is counted. The control frequency value is calculated, which is defined as the number of times the injection behavior is triggered per unit time. The control frequency is mapped to the [0, 1] interval through a normalization formula. The state change index, conductivity change trend index and control frequency index are concatenated to construct a state feature vector. The state feature vector corresponds to three sub-indicators in the dimension and forms a two-dimensional tensor structure in the sample dimension, which is then input into the stage judgment structure. The stage determination structure includes a feature extraction unit, a threshold mapping unit, and a stage output unit. The feature extraction unit performs z-score standardization on the input state feature vector to generate a standardized feature index sequence. The threshold mapping unit calls a preset stage determination function to calculate the matching degree vector for each stage by performing rule matching on the standardized feature index sequence. The stage output unit outputs the running stage label according to the position corresponding to the maximum value in the matching degree vector. When the state change index is greater than the set upper threshold and the control frequency index is higher than the set frequency threshold, the output label is "rapid adjustment stage"; when the state change index is within the set intermediate fluctuation range and the conductivity trend index is within the preset stable range, the output label is "stable stage"; when the control frequency index is at a low level and the conductivity trend index has a negative slope value, the output label is "decline stage"; when both the state change index and the conductivity trend index show a positive upward trend and the control frequency begins to rise, the output label is "start-up stage". The stage decision function is a multi-class decision mapping function. The structure is fitted by a KNN classifier. The training samples are derived from multiple rounds of system state records and engineering annotation stage results. The optimal decision boundary is obtained by leave-one-out cross-validation.

[0032] This implementation constructs a state feature vector based on the fluctuation characteristics of electrolyte operation and the frequency of control behavior, and introduces a stage judgment structure and a multi-condition combined threshold mapping method. This can accurately identify the operating stage of the system at different times, providing a highly reliable basis for subsequent weighted residual processing and dynamic adjustment of control strategies.

[0033] In this embodiment, the step of adjusting the parameter configuration of the response dynamic reconstruction function according to the running phase and updating the residual weighting strategy specifically involves: Based on the running stage label, the parameter configuration set corresponding to the stage is called from the preset parameter configuration library; the parameter configuration set includes high weight factor, balance factor, residual response sensitivity coefficient and initial bias coefficient; the weight generation structure dynamically sets the internal calculation path and activation function output amplitude according to each parameter in the parameter configuration set; When the running phase label is the rapid adjustment phase, the high weight factor is called from the configuration library and assigned to the weight amplification parameter in the weight generation structure, so that the weighting strategy pays more attention to the residuals in the high response zone and enhances the response intensity of the control action. When the running phase label is stable, the equalization factor is called from the configuration library to adjust the response weights of each residual source in the weight generation structure to maintain a relatively balanced state, control the distribution of the predicted residuals to prevent excessive shift, and maintain the stability of the residual weights. When the running phase is labeled as the decay phase, the residual response sensitivity coefficient is called from the configuration library and applied to the residual amplification path in the weight generation structure to reduce the model's response to small disturbances and keep the weighting process in a convergent state in response to changes in system stability. When the running phase label is the startup phase, the initial bias coefficient is called from the configuration library. This coefficient is input into the starting bias channel of the weight generation structure. In the initial stage with few samples and large fluctuations, it provides the initial guiding bias for residual weighting, enhancing the stability and response guidance capability of the model in the early stage. The above parameter configuration set is constructed from multi-stage system operation data and expert experience annotations. The parameters are adjusted and optimized using a multi-dimensional minimum error fitting method based on historical system response data. The fitting objective function is the mean square error function between the expected distribution of residual weights and the actual weighted distribution of residuals. After completing the matching of the runtime stage labels and parameter configuration sets, the weight generation structure is called to recalculate the residual weight value corresponding to each residual value and construct a new residual weight vector; the weight vector is applied to the original residual sequence, and point-by-point multiplication is performed to generate the updated weighted residual sequence.

[0034] This implementation method constructs a state-aware weighted strategy update mechanism by dynamically switching weights based on operational stage labels to generate parameter configuration paths. This enables a structural transition of the weighting function from static response to stage-adaptive adjustment, significantly improving the accuracy and robustness of concentration prediction residual feedback and providing a sensitive and controllable dynamic weighting basis for concentration regulation tasks.

[0035] In this embodiment, S6 specifically refers to: The differentiated weighted residual values ​​are input into the recursive feedback module to construct the loss feedback vector. The recursive feedback module includes a residual embedding structure, a time memory structure, and a residual fusion structure. The residual embedding structure uses a fully connected mapping method to perform vector encoding on the differentiated weighted residual values ​​to generate a residual embedding representation, which retains the time variation characteristics of the original residual information. The temporal memory structure is constructed using gated recursive units (GRUs). It takes the residual embedding representation and the historical residual signal sequence as input and calculates the temporal recursive features. The historical residual signal sequence is dynamically updated through a sliding time window mechanism. The memory state of the GRU constructs a recursive path through the hidden state vector, extracts the temporal dependence characteristics of residual changes, and outputs the temporal recursive features. The residual fusion structure fuses the temporal recursive features and the output feature vector of the drift coding module, and uses a residual connection mechanism and linear fusion transformation to generate a joint feedback tensor. The joint feedback tensor expresses the importance control signal of the residual change trend on the coding path features. The joint feedback tensor is used as the training signal input to the drift coding module, fusion construction module, and self-adjustment module. During training, a joint loss function is constructed, which consists of a prediction error term and a weighted residual penalty term. The prediction error term is obtained by fitting the mean square error formula between the predicted concentration value and the measured concentration value. The weighted residual penalty term is constructed based on the offset magnitude between the differential weighted residual value and the response interval boundary, and is obtained by fitting a dynamic interval penalty function. The joint loss function is defined as the total loss function, and the scaling factor is determined by parameter tuning experiments or dynamic fitting using a Bayesian optimization algorithm. The joint loss function is input into the optimizer to perform backpropagation training; the adaptive gradient optimizer Adam is used to update the parameters of each substructure of the model; the time drift parameter in the drift encoding module is obtained by fitting a time-weighted difference function; the path weight parameter in the fusion construction module is adjusted according to the importance score of the fusion feature; the adjustment matrix parameter in the self-adjustment module is dynamically reconstructed by the running stage label; the weight generation parameter in the response dynamic reconstruction function is adjusted by the weighted residual minimization direction. After updating the parameters, the concentration prediction path and the residual control path are optimized in a coordinated manner. The prediction path improves the accuracy and stability of concentration prediction, while the residual path improves the ability to perceive differences in variable type and response state.

[0036] This implementation constructs a multi-temporal residual learning mechanism through a recursive feedback module, establishes a highly coupled feedback chain between concentration prediction and residual evaluation, and drives multi-structure collaborative training of the model through a joint loss function, thereby improving the concentration control accuracy, response sensitivity and model convergence ability of the electrolyte system.

[0037] In this embodiment, S7 specifically refers to: The trained improved SSDNet model is deployed to the data control unit of the electrolyte system. At each sampling period, the newly constructed input dataset is input, and the improved SSDNet model is called to sequentially perform drift encoding, fusion construction, self-regulation and recursive feedback to generate the predicted concentration value at the target time. The predicted concentration value is the predicted result of the ratio of acid concentration to zinc concentration in the current period. The target predicted concentration value is compared with the preset target range of acid-zinc ratio concentration. The boundary comparison function is called to calculate the difference between the predicted value and the upper and lower limits. The judgment label is output according to the comparison result. If the predicted concentration value is less than the lower limit of the target range, an acid injection instruction is generated. If the predicted concentration value is greater than the upper limit of the target range, a zinc injection instruction is generated. If the predicted concentration value is within the range, the record is kept and the injection operation is not performed. Construct an injection command format, which includes: injection channel identifier field, injection agent type field, target injection volume parameter, control flow rate parameter and injection duration field; send the injection command to the injection control system, which parses the command content and drives the execution unit to start the target channel pump valve, and controls the corresponding injection type and injection volume parameter to complete the injection action; Collect acid and zinc concentration monitoring values ​​after the injection action is completed, record the corresponding injection time, injection parameters and concentration response values, and construct an updated input dataset containing a new historical injection record sequence. In the next sampling cycle, the improved SSDNet model is called, the updated input dataset is input, the drift coding, fusion construction, self-adjustment and recursive feedback process is re-executed, a new predicted concentration value is generated and the target deviation is judged, and the control command for whether to inject liquid is output again, forming a closed-loop control process with continuous cycle. The control rhythm vector is formed by the control frequency, injection flow rate and injection duration fields in the injection command generation process. The control rhythm vector is input to the control rhythm monitoring module and dynamically compared with the historical injection frequency standard value to extract the control deviation index. Based on the control deviation index, the residual response threshold and weighting function sensitivity of the response dynamic reconstruction function are dynamically adjusted to improve the ability to respond quickly to concentration deviation trends. An evaluation index for the regulation effect was constructed, including three dimensions: the regression time of the target value after injection, the stability of concentration fluctuation, and the accuracy of target recovery. The evaluation index was obtained by fitting the response effect data of the most recent N sampling periods using an exponential moving average function, and a regulation reliability label was further generated.

[0038] This implementation method constructs an automatic closed-loop control path between concentration prediction and injection control. When the concentration prediction result deviates from the target range, it quickly generates an injection control command to drive the injection system to complete the adjustment action. By combining the dynamic adjustment model residual processing mechanism and control strategy configuration with the control frequency and effect feedback, it improves the stability of the zinc acid ratio concentration, control response efficiency and concentration control accuracy of the electrolyte system under multivariable dynamic environment.

[0039] Example 1: To verify the feasibility of this invention in practice, it was applied to the automatic electrolyte control system of a large-scale zinc electrolysis production enterprise. This enterprise has a continuous zinc smelting production line, and the control precision required for the zinc-acid ratio concentration of the electrolyte during production is extremely high. Excessive or insufficient concentration can lead to problems such as electrode passivation, uneven crystallization, and increased energy consumption. Previous methods relying on traditional manual sampling and experience-based electrolyte injection strategies suffer from response lag, unreasonable control frequency, and large fluctuations in electrolyte injection decisions.

[0040] In this embodiment, the improved SSDNet model is deployed in the intelligent central control system of the production line. It collects and inputs data in real time, including acid concentration, zinc concentration, conductivity, current density, temperature, and injection records. In each sampling cycle, it predicts the acid-zinc ratio concentration, determines deviations, and generates injection commands. The experiment lasted three months, and the operational data before and after applying this invention were compared, resulting in the data comparison results shown in the table below.

[0041] The following is a comparison of key indicators before and three months after the implementation of this invention. The data is statistically summarized monthly as follows: Table 1 Comparison of Zinc Ratio Concentration Control Accuracy Before and After Implementation

[0042] As shown in Table 1, before applying this invention, the zinc acid ratio concentration deviation control was at a high level, with an average monthly error as high as 6.87% and a maximum deviation exceeding 10%, indicating significant control instability. In the second month, the system of this invention went online, automatically identifying response intervals and dynamically weighting residuals. The control strategy responded promptly to changes in state, reducing the average error to 3.42%. In the third month, the system operated stably. Through continuous optimization of the concentration prediction path and residual feedback path using the improved SSDNet model, the error further decreased to 1.16%, and the maximum deviation was controlled within 2.3%, demonstrating that this invention significantly improved control accuracy and system stability.

[0043] Meanwhile, regarding the responsiveness and resource consumption of the injection control system, the injection control parameters recorded in this embodiment are as follows: Table 2 Comparison of Injection Control Response Performance Before and After Implementation

[0044] As shown in Table 2, the number of injection cycles was significantly reduced after adopting the technology of this invention. This indicates that the improved prediction accuracy of the system reduced unnecessary adjustments and avoided system interference caused by frequent control. In particular, the over-injection ratio decreased dramatically from 19.4% before implementation to 4.3%, demonstrating that the dynamic reconstruction function and residual weighting strategy effectively compressed the injection error space. Regarding response delay, it decreased from an average of 97 seconds to 18 seconds, improving the control response capability by more than four times and significantly improving delay control. Furthermore, the number of alarms triggered by abnormal zinc acid ratios on the production line decreased from 13 to 1, effectively ensuring the continuous and stable operation of the production line.

[0045] The analysis results of the two tables above show that this invention constructs a concentration control strategy based on the fusion of deep learning and residual dynamic regulation. It comprehensively optimizes traditional control methods to address issues such as slow response, low prediction accuracy, and rigid adjustment strategies. By introducing drift coding modeling, trend-period fusion paths, self-regulating gating mechanisms, and residual recursive feedback modules, combined with dynamic identification and weighted strategy update mechanisms during the operational phase, the model maintains good adaptability and control capabilities across different operational phases. Therefore, this invention not only outperforms traditional methods in terms of accuracy, stability, and response speed, but also significantly reduces system energy consumption and interference frequency, demonstrating good engineering application value and promising prospects for widespread adoption.

[0046] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for regulating the zinc-to-acid ratio concentration in an electrolyte based on deep learning, characterized in that, Includes the following steps: S1. Collect the operating data of the electrolyte system and construct the input dataset; the operating data includes acid concentration, zinc concentration, conductivity, current density, temperature and historical electrolyte injection records; S2. Input the input dataset into the improved SSDNet model, perform the zinc acid ratio concentration prediction operation, output the predicted concentration value at the target time, and calculate the residual value between the predicted concentration value and the measured concentration value. S3. Identify the variable type and concentration response range corresponding to the residual value, classify the variable type into acid variable and zinc variable, and classify the concentration response range into high response range, medium response range and low response range. S4. Construct a dynamic weighted input vector and generate residual weighted values ​​by responding to a dynamic reconstruction function; perform weighted processing on the residual values ​​according to the residual weighted values ​​to generate differentiated weighted residual values; the dynamic weighted input vector includes variable type, concentration response interval, residual change rate, historical residual sequence and operating characteristics; S5. Based on the state change rate, conductivity change trend and control frequency in the previous sampling period, generate a state feature vector and determine the operating stage of the electrolyte system; adjust the parameter configuration of the response dynamic reconstruction function according to the operating stage and update the residual weighting strategy; the operating stage includes the startup stage, the stable stage, the rapid adjustment stage and the decay stage. S6. Feed the differentiated weighted residual value as a loss signal to the improved SSDNet model, perform joint training operation, and realize the collaborative optimization of zinc acid ratio concentration prediction and response dynamic reconstruction function. S7. Deploy the trained improved SSDNet model to the electrolyte system; If the predicted concentration value is not within the set target range, an acid or zinc solution injection command will be output to complete the closed-loop control of the acid-zinc ratio concentration.

2. The method for controlling the zinc-to-acid ratio of an electrolyte based on deep learning according to claim 1, characterized in that, Specifically, S1 is: Operational data is collected using an acid concentration sensor, a zinc concentration sensor, a conductivity detection device, a current density detection device, and a temperature detection device. The operational data includes acid concentration, zinc concentration, conductivity, current density, and temperature. The data on the injection volume of acid and zinc solution recorded in the injection control system is also read. The running data is time aligned to unify all data sequences to the same sampling period and time index; the time-aligned data is then filled with missing data by linear interpolation to fill in abnormal discontinuities; and the filled data is then normalized by linearly scaling each data sequence according to a preset range. The acid concentration sequence, zinc concentration sequence, conductivity sequence, current density sequence, temperature sequence, and historical injection record sequence are concatenated to form an input dataset in continuous time series format.

3. The method for controlling the zinc-to-acid ratio of an electrolyte based on deep learning according to claim 2, characterized in that, The improved SSDNet model includes a drift coding module, a fusion construction module, a self-adjusting module, and a recursive feedback module, specifically: The drift coding module performs feature extraction operations on acid concentration, zinc concentration, conductivity, current density, temperature and historical injection records, constructs a time drift matrix, models the time drift of the feature weights of various variables, and outputs a time-series drift feature vector. The integrated construction module includes trend path and periodic path. The trend path extracts long-term trend features based on the state space modeling structure, and the periodic path extracts periodic perturbation features based on the multi-scale convolution structure. The fusion construction module constructs an interaction matrix based on the output information of the trend path and the cycle path, performs an interactive fusion operation on the trend features and cycle features, and generates a fusion feature vector. The self-adjustment module receives the features of the operation phase, constructs a gated adjustment structure, generates an adjustment weight matrix based on the operation phase, adjusts the output amplitude and path channel ratio of the fusion construction module, and outputs the phase adjustment feature vector. The recursive feedback module constructs a recursive residual signal based on the differential weighted residual value, combines the historical residual signal with the current residual signal to generate a feedback vector, inputs the feedback vector to the drift coding module and the fusion construction module, and participates in the parameter update operation during the training process as a loss feedback path, thus completing the collaborative optimization of the concentration prediction path and the residual feedback path.

4. The method for controlling the zinc-to-acid ratio of an electrolyte based on deep learning according to claim 3, characterized in that, Specifically, S2 is: The input dataset is fed into the drift coding module to generate a time-drift feature vector. The time-series drift feature vector is input into the fusion construction module, where trend features are extracted from the trend path and periodic features are extracted from the periodic path. Based on the interaction matrix in the fusion construction module, the fusion operation of trend features and periodic features is performed to generate a fused feature vector. The fused feature vector is input into the self-adjustment module, which generates an adjustment weight matrix according to the running stage, performs channel weighting on the fused feature vector, and generates a stage-adjusted feature vector. Perform a fully connected computation on the stage-adjusted feature vector to generate the predicted concentration value at the target time. The predicted concentration values ​​are matched one-to-one with the collected measured concentration values. Time alignment is performed according to the set sampling index matching rules. The difference at each sampling point is calculated to generate the corresponding residual sequence.

5. The method for controlling the zinc-to-acid ratio of an electrolyte based on deep learning according to claim 4, characterized in that, Specifically, S3 is: Based on the index position information of each residual value in the residual sequence, the type of input variable corresponding to the residual value is determined; if the residual value comes from the acid concentration data field, it is marked as an acid variable. If the residual value originates from the zinc concentration data field, it is marked as a zinc variable; the variable type is classified according to the source of the variable. For each residual value corresponding to a time index, calculate the rate of change of conductivity between adjacent sampling points before and after the current time point, and extract the conductivity change trend index; retrieve multiple residual values ​​adjacent to the current time index in the historical residual sequence, and calculate the local fluctuation amplitude index; combine the conductivity change trend index and the residual fluctuation amplitude index with weights to construct the concentration response index value. The concentration response index value is compared with the set response interval division threshold. If the concentration response index value is greater than the high response threshold, it is marked as a high response zone; if the concentration response index value is between the high response threshold and the low response threshold, it is marked as a medium response zone; if the concentration response index value is less than the low response threshold, it is marked as a low response zone. The concentration response interval is divided according to the concentration response index value; the variable type label corresponding to each residual value is merged with the concentration response interval label to generate a joint label result.

6. The method for controlling the zinc-to-acid ratio of an electrolyte based on deep learning according to claim 5, characterized in that, Specifically, S4 is: The variable type, concentration response range, residual change rate, historical residual sequence, and operational features are concatenated to construct a dynamic weighted input vector; the residual change rate is the difference result between adjacent residual values, the historical residual sequence is the set of residual values ​​of multiple sampling points before the current residual value time index, and the operational features include conductivity change trend, current density fluctuation amplitude, and temperature offset. The dynamic weighted input vector is fed into the response dynamic reconstruction function to generate residual weighted values; The response dynamic reconstruction function includes an input mapping structure, a feature association structure, and a weight generation structure. The input mapping structure performs a normalization transformation on the dynamic weight-sensitive input vector to generate a standardized feature tensor. feature The correlation structure calculates the combination relationship between variable types, response intervals and operational features based on the standardized feature tensor, and outputs the correlation feature results. The weight generation structure performs a non-linear function operation on the associated feature results and outputs the residual weighted value corresponding to the residual value. The residuals are weighted according to their weighted values. Weighted multiplication is performed on the residuals corresponding to acid and zinc variables respectively. When the variable type is acid and the concentration response range is in the high response range, the residual weighting value adopts a high weighting factor. When the variable type is acid and the concentration response range is in the low response range, the residual weighting value adopts a low weighting factor. When the variable type is zinc and the concentration response range is in the medium response range, the residual weighting value adopts a medium weighting factor. All weighted residuals are then combined to form a differentiated weighted residual sequence.

7. The method for controlling the zinc-to-acid ratio of an electrolyte based on deep learning according to claim 6, characterized in that, Based on the rate of change of state, the trend of change of conductivity, and the control frequency in the previous sampling period, a state feature vector is generated, and the operating stage of the electrolyte system is determined, specifically as follows: The rate of change of state is calculated based on the acid concentration, zinc concentration, conductivity, current density and temperature data recorded in the previous sampling period. The rate of change of state is obtained by the numerical difference results of adjacent sampling points. The trend of conductivity change in multiple consecutive sampling points is calculated as the trend slope value. The control frequency is calculated according to the ratio of the number of injection command triggers to the sampling period. The state feature vector is constructed by concatenating the rate of change of state and the trend of change of conductivity with the control frequency. The state feature vector is input into the stage decision structure; the stage decision structure includes a feature extraction unit, a threshold mapping unit, and a stage output unit; features The extraction unit performs standardized calculations on the state feature vectors to generate a feature index sequence; the threshold mapping unit calculates the stage matching degree based on the feature index sequence; the stage output unit determines the operating stage label based on the stage matching degree; when both the state change rate and the control frequency are higher than the stage change threshold, it is determined to be a rapid adjustment stage; when the state change rate is stable and the conductivity change trend is within the threshold range, it is determined to be a stable stage; when the control frequency is low and the conductivity change trend is decreasing, it is determined to be a decay stage; when both the state change rate and the conductivity change trend are in an upward state, it is determined to be a startup stage.

8. The method for controlling the zinc-to-acid ratio of an electrolyte based on deep learning according to claim 7, characterized in that, The adjustment of the parameter configuration of the dynamic reconstruction function based on the operational phase and the updating of the residual weighting strategy are specifically as follows: Select the corresponding parameter configuration set based on the running phase label, and adjust the weight generation structure parameters in the response dynamic reconstruction function; when the running phase is the fast adjustment phase, increase the high weight factor in the weight generation structure. When the operation phase is a stable phase, maintain the balance factor in the weight generation structure; When the running phase is the decay phase, reduce the residual response sensitivity coefficient in the weight generation structure; When the running phase is the startup phase, the initial bias coefficients in the weight generation structure are enabled. After updating the residual weighting strategy, the residual weight values ​​are recalculated, and the weight distribution of the weighted residual sequence is adjusted to complete the process of identifying and dynamically adjusting residual weights during the operation phase.

9. The method for controlling the zinc-to-acid ratio of an electrolyte based on deep learning according to claim 8, characterized in that, Specifically, S6 is: The differentially weighted residual values ​​are input into the recursive feedback module to construct the loss feedback vector. The recursive feedback module includes a residual embedding structure, a temporal memory structure, and a residual fusion structure. The residual embedding structure performs vector encoding on the differentially weighted residual values ​​to generate a residual embedding representation. The temporal memory structure combines the residual embedding representation with historical residual signals to extract temporal recursive features. The residual fusion structure performs fusion operations on the temporal recursive features and the output features of the drift encoding module to generate a joint feedback tensor. The joint feedback tensor is used as the training signal input to the drift coding module, fusion construction module, and self-adjustment module to construct a joint loss function during end-to-end training. The joint loss function includes a prediction error term and a weighted residual penalty term. The prediction error term is calculated based on the mean square error between the predicted concentration value and the measured concentration value, and the weighted residual penalty term is calculated based on the degree of interval offset of the differentiated weighted residual values. By performing backpropagation training on the model parameters using a joint loss function, the time drift parameters in the drift coding module, the path weight parameters in the fusion construction module, the adjustment matrix parameters in the self-adjustment module, and the weight generation parameters in the response dynamic reconstruction function are updated respectively. This achieves the coordinated updating of the concentration prediction path and the residual control path, completing the joint training process for the acid-zinc ratio concentration prediction task and the residual response modeling task.

10. The method for controlling the zinc-to-acid ratio of an electrolyte based on deep learning according to claim 9, characterized in that, Specifically, S7 is: The trained improved SSDNet model is deployed to the data control unit of the electrolyte system. The newly constructed input dataset is input in each sampling period, the zinc acid ratio concentration prediction operation is performed, and the target predicted concentration value is output. The target predicted concentration value is compared with the preset target range of acid-zinc ratio concentration to determine whether the target predicted concentration value is within the boundary of the target range; If the target predicted concentration value is less than the lower limit, an acid injection command is generated; If the target predicted concentration value is greater than the upper limit value, a zinc liquid injection command is generated; the injection command includes the target injection channel identifier, the injectant type identifier, the control flow parameters, and the injection duration; The system drives the injection control system to complete the injection process of acid or zinc solution according to the injection command, and records the injection volume data to the injection control recording module. After the injection is completed, acid and zinc concentration monitoring data are collected to construct an updated input dataset. In the next sampling cycle, the improved SSDNet model is called to perform concentration prediction and deviation judgment operations to complete the closed-loop control of acid-zinc ratio concentration.

Images