Decontamination liquid concentration real-time regulation and control system

Abstract

The invention relates to the technical field of automatic control, and discloses a decontamination liquid concentration real-time regulation and control system which comprises a concentration acquisition module, a parameter analysis module, a regulation and control strategy generation module, an execution control module and an output feedback module. The concentration acquisition module acquires real-time concentration sampling data in the pipeline and preprocesses the real-time concentration sampling data; the parameter analysis module performs hierarchical decomposition and feature extraction on the data to obtain a concentration feature vector set; a regulation and control strategy generation module performs correlation optimization on the feature vectors to generate an optimized regulation and control vector set; the execution control module globally reconstructs the optimization regulation and control vector to obtain an execution control feature map; and the output feedback module generates a concentration regulation and control output signal accordingly. The system further comprises a physical regulation and control module, and regulation sequence conversion and composite regulation protocol generation are achieved through a hardware controller and a communication interface. According to the system, the real-time performance and the accuracy of concentration regulation and control are improved, and the stability and the reliability of the decontamination effect are ensured.

Description

Technical Field

[0001] The present invention relates to the field of automation control technology, and in particular to a real-time control system for the concentration of a cleaning solution. Background Art

[0002] In areas such as public health and safety, medical epidemic prevention, and industrial cleaning, the effective use of disinfectants plays a key role in killing pathogens and removing pollutants. Precise control of disinfectant concentration directly impacts disinfection effectiveness and safety. Traditional methods for controlling disinfectant concentration have significant limitations: early systems often employed single-point sampling and open-loop control, which were unable to capture the spatial distribution and dynamic trends of disinfectant concentration within the pipeline in real time. When faced with complex operating conditions such as flow fluctuations, temperature changes, or multi-channel mixing, concentration control exhibited significant lag, often leading to overshoot or undershoot, resulting in reduced disinfection efficiency or excessive chemical residues.

[0003] Existing control systems have weak data processing capabilities and lack effective hierarchical decomposition and feature extraction mechanisms for sampled data, making it difficult to accurately identify characteristic patterns of concentration changes from raw data heavily tainted by noise. For example, some systems process data using only simple mean filtering, ignoring the sudden changes and periodic patterns of concentration signals at different time scales. This results in control strategies that are unable to adapt to sudden pollution events or the demands of sustained operations.

[0004] Traditional control strategy generation methods mostly rely on preset thresholds or empirical rules, and lack the ability to analyze the temporal correlation of historical data and dynamically optimize it. When the use scenario of the disinfectant changes (such as different pollutant types, different disinfection objects), fixed control parameters are difficult to achieve adaptive adjustment, requiring frequent manual intervention, which increases the complexity of operation and the risk of human error. At the same time, the execution control link lacks a global reconstruction mechanism for multi-source control signals, and the coordination between various control instructions is poor, which can easily cause frequent start-stop or dosage fluctuations of the filling equipment, affecting the stability of the system and the service life of the equipment.

[0005] Existing systems generally lack closed-loop feedback and dynamic correction capabilities for the control process, making it impossible to adjust control strategies in real time based on actual concentration changes. This is especially true during long-term continuous operation, where accumulated errors can cause concentrations to deviate from target values, seriously impacting the sustainability of disinfection effectiveness. With the increasing prevalence of intelligent and unmanned disinfection scenarios, the shortcomings of traditional control technologies in terms of real-time performance, accuracy, and adaptability have become increasingly prominent. There is an urgent need for a new control system that can achieve real-time concentration data collection, in-depth feature analysis, intelligent strategy generation, and precise control execution. Summary of the Invention

[0006] The purpose of the present invention is to provide a real-time control system for the concentration of disinfectant to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a real-time control system for the concentration of a cleaning solution, the system comprising:

[0008] Concentration acquisition module, used to obtain real-time concentration sampling data in the disinfectant circulation pipeline;

[0009] A parameter analysis module, configured to perform hierarchical decomposition and feature extraction on the real-time concentration sampling data to obtain a set of concentration feature vectors;

[0010] a control strategy generation module, configured to perform association optimization on the set of concentration feature vectors to obtain a set of optimized control vectors;

[0011] an execution control module, configured to globally reconstruct the set of optimized control vectors to obtain an execution control feature graph;

[0012] An output feedback module is used to generate a concentration control output signal according to the execution control characteristic diagram.

[0013] Preferably, the concentration acquisition module includes: a sensor array unit, used to divide the real-time concentration data in the disinfectant circulation pipeline according to a preset sampling period to obtain a set of concentration sampling units; a data preprocessing unit, used to input the data of each unit in the set of concentration sampling units into a processor based on dynamic filtering to obtain a set of concentration feature vectors.

[0014] Preferably, the control strategy generation module includes: a feature vector reorganization unit, which reorganizes the concentration feature vector along the time dimension to obtain a set of reorganized control vectors; a feature correlation calculation unit, which is used to calculate the correlation between adjacent vectors in the set of reorganized control vectors to generate a control correlation matrix; a weight correction unit, which is used to adjust the weight of the control correlation matrix according to the temporal relationship between adjacent vectors in the set of reorganized control vectors to obtain a constraint correlation matrix; and a strategy optimization unit, which is used to convolute and fuse the constraint correlation matrix with the set of reorganized control vectors to obtain the optimized control vector.

[0015] Preferably, the feature correlation calculation unit includes: a state space mapping subunit, used to map each feature vector in the set of recombinant control vectors to the state density space to obtain a set of mapped recombinant control vectors; a correlation analysis subunit, used to calculate the cosine correlation coefficient between adjacent mapping vectors in the set of mapped recombinant control vectors to generate the control correlation matrix composed of multiple control correlation values.

[0016] Preferably, the strategy optimization unit is specifically implemented as follows: performing deep convolution processing on the constraint correlation matrix to obtain a constraint correlation feature matrix; inputting the set of reorganized control vectors and the constraint correlation feature matrix into a gated recurrent network to obtain a set of context control vectors; and stacking the set of context control vectors to obtain the optimized control vector.

[0017] Preferably, the execution control module includes: an optimized feature dimensionality reduction unit, used to perform channel dimension mean pooling processing on each vector in the set of optimized control vectors to obtain a set of reduced-dimensionality control vectors; a feature dynamic distribution degree calculation unit, used to calculate the distribution degree value of each feature vector in the set of reduced-dimensionality control vectors to generate a control distribution degree set; a control reference center determination unit, used to select the reduced-dimensionality control vector corresponding to the maximum distribution degree value in the control distribution degree set as the initial control center vector; a feature dynamic weight allocation unit, used to calculate the dynamic control weight of each feature vector in the set of reduced-dimensionality control vectors and the initial control center vector based on the logical distance between each feature vector and the distribution degree value of each feature vector to generate a dynamic control weight set; a global feature reconstruction unit, used to use the dynamic control weight set to perform weighted fusion on the set of reduced-dimensionality control vectors to generate the execution control feature graph.

[0018] Preferably, the characteristic dynamic distribution degree calculation unit is specifically implemented as follows: calculating the median vector and standard deviation vector of the dimensionality reduction control vector; performing element-by-element difference calculation between the dimensionality reduction control vector and the median vector, and performing a cubic operation on the difference result to obtain a control difference vector; calculating the overall median value of the control difference vector; performing a product operation on the median value and the cube value of the standard deviation vector, and inputting the result into a normalization function to obtain the distribution degree value.

[0019] Preferably, the characteristic dynamic weight allocation unit is specifically implemented as follows: multiplying the distribution value of the dimensionality reduction control vector and the distribution value of the initial control center vector by a first correction coefficient to obtain a first dynamic control factor; multiplying the Chebyshev distance between the dimensionality reduction control vector and the initial control center vector by a second correction coefficient to obtain a second dynamic control factor; and performing a product operation on the first dynamic control factor and the second dynamic control factor to obtain the dynamic control weight.

[0020] Preferably, the output feedback module is specifically implemented as follows: inputting the execution control characteristic diagram into an output signal generator based on a decision tree to obtain the concentration control output signal, and the concentration control output signal is used to indicate the adjustment instruction of the disinfectant concentration.

[0021] Preferably, the system also includes a physical control module, which is composed of a hardware controller and a communication interface, wherein: the hardware controller has an embedded dedicated processing chip for converting the execution control characteristic diagram into an optimized adjustment sequence; the communication interface is arranged in the disinfectant filling channel, for synchronously collecting flow pressure data, and generating a composite adjustment protocol after multi-dimensional cross-integration with the optimized adjustment sequence; the composite adjustment protocol is transmitted to the execution terminal through a wired channel to complete the concentration adjustment action.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] The real-time control system for cleaning solution concentration provided by this invention achieves high-precision, real-time control of cleaning solution concentration through the collaborative operation of multiple modules, offering significant technical advantages and practical value. The concentration acquisition module divides the sensor array into a set of concentration sampling units according to a preset sampling period and pre-processes the data using a dynamic filtering processor. This effectively improves the signal-to-noise ratio of the raw concentration data, ensuring that the collected data truly reflects the temporal and spatial distribution characteristics of the concentration within the pipeline, providing a reliable data foundation for subsequent control.

[0024] The parameter analysis module performs hierarchical decomposition and feature extraction on real-time concentration sampling data, and can separate feature vectors of different time scales and frequency components from complex concentration signals to form a set of concentration feature vectors. This deep feature extraction method can accurately capture the mutation points, trend items and periodic laws of concentration changes. Compared with traditional simple filtering processing, it significantly improves the system's perception of dynamic changes in concentration.

[0025] The control strategy generation module reorganizes concentration feature vectors along the time dimension through a series of processes, including feature vector reorganization, correlation calculation, weight correction, and strategy optimization. This generates a set of reorganized control vectors. It then calculates the correlations between adjacent vectors to generate a control correlation matrix, adjusts weights based on temporal relationships, and ultimately obtains the optimized control vector through convolutional fusion. This process fully considers the temporal correlation of concentration data and utilizes technologies such as gated recurrent networks to achieve dynamic optimization of the control strategy, enabling the strategy to adapt to concentration changes under different operating conditions. Compared to traditional control methods based on preset thresholds, this improves the strategy's flexibility and adaptability.

[0026] The execution control module performs dimensionality reduction, distribution calculation, benchmark center determination, and dynamic weight assignment on the optimized control vector, ultimately reconstructing global features to generate an execution control feature map. The dynamic distribution calculation and dynamic weight assignment mechanism fully consider the importance differences and spatial distribution relationships of each feature vector, ensuring that control instructions accurately reflect the actual needs of concentration control, avoiding control deviations caused by traditional averaging, and improving the accuracy and effectiveness of execution instructions.

[0027] The output feedback module feeds the execution control characteristic diagram into a decision-tree-based output signal generator to generate a concentration control output signal. Combined with the physical control module's hardware controller and communication interface, this module achieves precise conversion from digital signals to physical control actions. The hardware controller, embedded with a dedicated processing chip, converts the execution control characteristic diagram into an optimized control sequence. The communication interface simultaneously collects flow and pressure data and integrates it with the control sequence to generate a composite control protocol, ensuring that control commands respond in real time to flow and pressure changes within the pipeline. This enables closed-loop feedback and dynamic correction of concentration control, effectively avoiding cumulative errors and ensuring concentration stability during long-term operation.

[0028] Through intelligent processing of the entire process from data collection, feature analysis, strategy generation, execution control to feedback correction, the system achieves real-time, precise and adaptive regulation of the concentration of disinfectant, greatly improving the efficiency and safety of disinfection operations and reducing the need for manual intervention. It can be widely used in various scenarios that require precise control of the concentration of disinfectant, and has significant economic and social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a working principle diagram of the real-time control system for the concentration of the disinfectant according to the present invention;

[0030] Figure 2 Detailed design diagram for the concentration acquisition module;

[0031] Figure 3 Detailed design diagram for the feature correlation calculation unit;

[0032] Figure 4 This is a detailed design drawing of the physical control module. DETAILED DESCRIPTION

[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0034] See also Figures 1-4 The present invention relates to a real-time control system for the concentration of disinfectant, and the specific implementation methods are as follows:

[0035] The concentration acquisition module acquires real-time concentration sampling data within the disinfectant flow pipeline. This module uses a sensor array deployed within the pipeline to collect real-time concentration data within the pipeline according to a preset sampling period, forming a collection of concentration sampling units.

[0036] The parameter analysis module performs hierarchical decomposition and feature extraction on real-time concentration sampling data to obtain a set of concentration feature vectors. A dynamic filter processor processes the data from each unit in the concentration sampling unit set, removing noise and extracting key features, converting them into concentration feature vectors that can be used for analysis.

[0037] The control strategy generation module performs correlation optimization on the set of concentration feature vectors to obtain a set of optimized control vectors. The optimized control vectors are generated by reorganizing the feature vectors in the time dimension, calculating correlations, adjusting weights, and performing convolution fusion operations.

[0038] The execution control module globally reconstructs the set of optimized control vectors to produce an execution control feature graph. Through dimensionality reduction, distribution calculation, benchmark center determination, weight assignment, and weighted fusion, the optimized control vectors are reconstructed into a feature graph that can guide execution.

[0039] The output feedback module generates a concentration control output signal based on the execution control feature map. The feature map is input into the decision tree-based output signal generator to obtain an output signal indicating the disinfectant concentration adjustment instruction.

[0040] Example 1: In this example, the concentration acquisition module includes a sensor array unit and a data preprocessing unit, and its specific implementation is as follows:

[0041] The sensor array unit is deployed inside the disinfectant circulation pipeline. The array consists of multiple different types of concentration sensors, such as electrochemical sensors, optical sensors, etc. These sensors are installed in different positions of the pipeline according to a specific layout, such as around the cross-section of the pipeline, at different axial heights, etc., to ensure that the concentration data of the disinfectant in the pipeline can be fully collected. When the sensor array unit is working, the concentration of the disinfectant in the pipeline is collected in real time according to the preset sampling period. The preset sampling period can be set according to the actual application scenario and needs. For example, in scenarios that are more sensitive to concentration changes, the sampling period can be set to 50 milliseconds to achieve high-frequency data collection; in scenarios where the concentration changes relatively slowly, the sampling period can be appropriately extended to 100 milliseconds or longer.

[0042] During each sampling cycle, each sensor in the sensor array operates simultaneously, acquiring the detergent concentration data at its location, thus forming a concentration sampling unit. Over time, the concentration sampling units acquired over multiple sampling cycles are accumulated to form a collection of concentration sampling units. Each concentration sampling unit contains the concentration data for different points within the pipeline during that sampling cycle. For example, for a pipeline with four sensors installed across a cross-section, each sampling unit might contain concentration values ​​for four different points, along with information such as the sampling timestamp. This data comprehensively reflects the spatial distribution of the detergent concentration within the pipeline at that moment.

[0043] The data preprocessing unit is electrically connected to the sensor array unit and is responsible for processing the data from each unit in the collection of concentration sampling units. This unit integrates a processor based on dynamic filtering, whose core function is to filter the input concentration data to remove noise and extract useful feature information. A characteristic of dynamic filtering is that its filter parameters can be adaptively adjusted based on the real-time input data characteristics, thereby better adapting to the various possible changes in the disinfectant concentration.

[0044] After the collection of concentration sampling units is input into the data preprocessing unit, the processor first performs preliminary denoising on the data at each point in each concentration sampling unit. For the concentration data at each point, the processor analyzes its changing trend in the time series to determine whether there are any outliers. For example, when the concentration value of a certain point suddenly jumps significantly within adjacent sampling cycles and is significantly inconsistent with the data change trend of other points, the processor will identify it as possible noise data and use appropriate filtering algorithms, such as median filtering, mean filtering, etc., to eliminate or correct the outlier.

[0045] In addition to removing noise, the processor also performs feature extraction on the data to obtain a set of concentration feature vectors. During feature extraction, the processor analyzes the changing characteristics of the concentration data in both the temporal and spatial dimensions. In the temporal dimension, the processor focuses on the concentration trend over time, such as the rate of increase, rate of decrease, and steady-state state. In the spatial dimension, the processor focuses on the distribution differences in concentration at different points in the pipeline, including concentration gradients. By extracting and quantifying these features, each concentration sampling unit is converted into a multidimensional concentration feature vector.

[0046] For example, for a sampling unit containing concentration data from four points, after denoising and feature extraction, the data from each point may yield characteristic parameters such as the current concentration value at that point, the difference from the previous sampling period, and the rate of change. These parameters together constitute a multidimensional feature vector. Arranging the feature vectors from multiple sampling periods in chronological order forms a collection of concentration feature vectors.

[0047] The data preprocessing unit also features data synchronization and calibration. Because individual sensors in the sensor array may have varying response times or performance variations, the data preprocessing unit synchronizes the data collected by different sensors to ensure temporal consistency across all points within the same sampling cycle. Furthermore, the unit regularly calibrates sensor data, adjusting sensor output data by comparing it with standard concentration samples to improve data accuracy and reliability.

[0048] In hardware implementation, the sensors in the sensor array unit are connected to the data preprocessing unit via waterproof, corrosion-resistant cables to accommodate the potentially corrosive environment of disinfectant solutions. The data preprocessing unit can utilize a high-performance embedded processor, such as an ARM series processor, in conjunction with appropriate analog and digital signal processing circuits to enable real-time collection and processing of concentration data. The dynamic filtering and feature extraction algorithms running within the processor can be written in C or assembly language to ensure processing speed and real-time performance.

[0049] Example 2: In this example, the control strategy generation module includes a feature vector recombination unit, a feature correlation calculation unit, a weight correction unit, and a strategy optimization unit. The specific implementation method is as follows:

[0050] The core function of the feature vector recombination unit is to reorganize the concentration feature vectors along the time dimension. The unit receives a set of concentration feature vectors from the parameter analysis module. These vectors are generated in the order of the sampling period, and each vector represents the characteristics of the disinfectant concentration at the corresponding moment. The recombination unit arranges these vectors in chronological order to form a time series. For example, if the sampling period is 50 milliseconds and 100 periods are collected continuously, 100 concentration feature vectors will be obtained. The recombination unit will arrange them in chronological order from the 1st to the 100th to form a set of reorganized control vectors. This recombination method enables the vector sequence to reflect the evolution of the concentration characteristics over time, laying the foundation for the subsequent analysis of the correlation relationship in the time dimension. In hardware implementation, the unit can use a cache to store the vector sequence and use a timing control circuit to ensure that the vectors are arranged in the correct order.

[0051] The feature correlation calculation unit includes a state-space mapping subunit and a correlation analysis subunit. The state-space mapping subunit maps each eigenvector in the reorganized control vector set to the state density space. The state density space is a multidimensional abstract space whose dimensions are determined by the dimensions of the eigenvector. For example, when the eigenvector is 10-dimensional, the state density space is a 10-dimensional space. During the mapping process, the subunit normalizes the data of each dimension of the eigenvector to eliminate the influence of different dimensional dimensions, and then maps its coordinates to the state density space. For example, if the dimensional values ​​of a certain eigenvector are normalized to [0.2, 0.5, 0.3, ...], then there is a corresponding coordinate point in the state density space. Through this mapping, the distribution characteristics of the eigenvector can be intuitively reflected in the state density space, facilitating subsequent correlation analysis.

[0052] The correlation analysis subunit calculates the cosine correlation coefficient between adjacent vectors based on the set of mapped vectors. For two adjacent mapping vectors A and B in the state density space, the cosine correlation coefficient measures their similarity by calculating the cosine value of the angle between the two vectors. In the specific calculation, first obtain the numerical values ​​of each dimension of the two vectors, calculate their dot product, and then divide it by the product of the moduli of the two vectors. The result is the cosine correlation coefficient, which ranges from [-1,1]. The closer the coefficient is to 1, the more similar the two vectors are, that is, the stronger the correlation between the concentration characteristics at adjacent time points; the closer it is to -1, the weaker the correlation. The subunit will calculate the cosine correlation coefficient of each pair of adjacent mapping vectors in turn, and generate a regulatory correlation matrix consisting of multiple regulatory correlation values. For example, 100 mapping vectors will generate a 99×1 correlation matrix, and each element in the matrix corresponds to the correlation value of adjacent vectors.

[0053] The weight correction unit adjusts the weights of the control correlation matrix based on the temporal relationship between adjacent vectors in the reorganized control vector set. The key to temporal relationships lies in the impact of time intervals on correlation. It is generally believed that vectors closer to the current moment have greater reference value for subsequent control. This unit sets a time decay factor, for example, based on sampling periods, with the weight decaying proportionally with each additional period. For example, assuming a decay factor of 0.95, and the correlation between the nth and n+1th vectors is processed, if the time interval between them is 1 period, the weight is 1; if the time interval is 2 periods, the weight is 0.95 × 1; if the time interval is 3 periods, the weight is 0.95 × 0.95, and so on. The constraint correlation matrix is ​​obtained by multiplying each element in the control correlation matrix by the corresponding temporal weight. This adjustment gives the correlation of recent data a greater weight in the matrix, better meeting the timeliness requirements of real-time control.

[0054] The policy optimization unit first performs deep convolution on the constraint correlation matrix. This convolution uses multiple convolution kernels of different sizes, such as 3×1 and 5×1 one-dimensional kernels, to extract features from the correlation values ​​in the matrix. Convolution can capture local patterns and trends in the correlation sequence, such as areas of continuous high correlation or sudden changes. Through multiple layers of convolution, a constraint correlation feature matrix is ​​generated, which contains the high-order features of the correlation data in the time dimension.

[0055] Next, the policy optimization unit inputs the reorganized control vector set and the constraint association feature matrix into the gated recurrent network (GRU). GRU is a recursive neural network that can process time series data and capture long-term dependencies. The input layer of the network receives the fusion data of the reorganized control vector and the association feature matrix, and the hidden layer selectively retains historical information and current input information through the gating mechanism (update gate and reset gate). For example, the update gate determines how much of the state at the previous moment is retained to the current moment, and the reset gate determines how much of the state at the previous moment is ignored. Through this mechanism, GRU can learn the contextual relationship in the concentration feature vector sequence. For example, when the correlation of concentration features is high within a certain period of time, the network will strengthen the memory of the features in this segment, thereby generating a set of contextual control vectors that is more in line with the time series law.

[0056] Finally, the policy optimization unit stacks the context control vectors. Stacking involves concatenating multiple context control vectors by dimension. For example, 10 context vectors (each with 20 dimensions) are stacked into a 200-dimensional vector to generate an optimized control vector. This vector integrates feature correlation information and context dependencies across time, providing more comprehensive and optimized control strategy guidance for the subsequent execution control module.

[0057] In terms of hardware implementation, the control strategy generation module can use an FPGA (field-programmable gate array) or GPU (graphics processing unit) as its core computing unit. FPGAs, with their strong parallel computing capabilities and low latency, are suitable for processing feature vector recombination and convolution operations. The GPU's massively parallel computing core is well-suited for state-space mapping and matrix operations in GRU. Internally, each unit in the module is connected via a high-speed data bus to ensure real-time data transmission. At the software level, algorithms are developed using parallel computing frameworks such as CUDA or OpenCL to improve computational efficiency.

[0058] Example 3: In this embodiment, the execution control module includes an optimization feature dimension reduction unit, a feature dynamic distribution degree calculation unit, a control reference center determination unit, a feature dynamic weight allocation unit and a global feature reconstruction unit. The specific implementation method is as follows:

[0059] The core function of the optimized feature dimensionality reduction unit is to perform mean pooling processing on the channel dimension of each vector in the optimized control vector set. The optimized control vector set is output by the control strategy generation module, and each vector may contain feature data of multiple channels, such as 128 dimensions or 256 dimensions. During mean pooling processing, the dimensionality reduction unit calculates the mean of all data under each channel dimension of each vector, thereby compressing the high-dimensional vector into a low-dimensional vector. For example, a 256-dimensional vector can be reduced to 64 dimensions or 32 dimensions after mean pooling, forming a set of reduced-dimensional control vectors. This dimensionality reduction operation reduces the data dimension while retaining the main feature information, thereby reducing the complexity of subsequent calculations. In hardware implementation, the unit can simultaneously process the mean calculation of multiple channels through parallel computing circuits, and use cache to store intermediate results to ensure the real-time performance of the dimensionality reduction process.

[0060] The dynamic distribution calculation unit is used to calculate the distribution value of each feature vector in the dimensionality reduction control vector set. The specific steps are as follows: Calculate the median vector and standard deviation vector of the dimensionality reduction control vector. Each dimension of the median vector is the median value of the corresponding dimension of all vectors in the dimensionality reduction control vector set. For example, for 100 32-dimensional dimensionality reduction vectors, calculate the median of 100 values ​​in each dimension to form a 32-dimensional median vector. Each dimension of the standard deviation vector is the standard deviation of the corresponding dimension data, reflecting the degree of dispersion of the dimension data.

[0061] Each dimension-reduced control vector is subtracted from the median vector element-by-element to obtain the difference along each dimension. The resulting difference is then cubed to produce the control difference vector. The purpose of the cubing operation is to amplify large differences, making abnormalities in the data distribution more prominent. For example, if the difference in a dimension is 2, it will be 8 after cubing; if the difference is -3, it will be -27 after cubing, thus more clearly indicating the degree to which the data deviates from the median.

[0062] Calculate the overall median of the control difference vector. This calculation takes the median of the cubed differences across all dimensions of the control difference vector. This reduces the influence of extreme values ​​and improves the robustness of the distribution calculation. For example, if the control difference vector has 32 cubed differences, sort these 32 values ​​and take the median as the overall median.

[0063] Multiply the overall median by the cube of the standard deviation vector and input the result into a normalization function to obtain the distribution value. Normalization functions typically use linear normalization to map the product to the interval [0, 1], making the distribution values ​​of different vectors comparable. For example, if the product is 50, it might be 0.75 after normalization, indicating that the vector has a high distribution value and a large degree of data dispersion.

[0064] The control reference center determination unit selects the reduced-dimensionality control vector corresponding to the maximum distribution value from the control distribution set as the initial control center vector. A larger distribution value indicates a more dispersed data distribution, often containing richer feature information and better representing the typical characteristics of the current control scenario. For example, if there are 100 distribution values ​​in the control distribution set, with the largest distribution value being 0.92, the corresponding reduced-dimensionality vector is the initial control center vector, which will serve as the benchmark for subsequent weight allocation.

[0065] The characteristic dynamic weight allocation unit calculates the dynamic control weight based on the logical distance between the dimensionality reduction control vector and the initial control center vector and the distribution value of each vector. The specific process is as follows: First, the product of the distribution value of the dimensionality reduction control vector and the distribution value of the initial control center vector is calculated, and multiplied by the first correction coefficient to obtain the first dynamic control factor. The first correction coefficient is a parameter pre-set according to the system control accuracy requirements, such as 0.8 or 1.2, which is used to adjust the degree of influence of the distribution on the weight. For example, if the distribution degree of a dimensionality reduction vector is 0.8, the distribution degree of the initial center vector is 0.92, and the first correction coefficient is 0.8, then the first dynamic control factor is 0.8×0.92×0.8=0.5888.

[0066] The Chebyshev distance between the reduced-dimensional control vector and the initial control center vector is calculated and multiplied by the second correction coefficient to obtain the second dynamic control factor. The Chebyshev distance is the maximum absolute value of the difference between the two vectors in each dimension, which can reflect the maximum difference between the two vectors in each dimension. The second correction coefficient is also a preset parameter, such as 0.5 or 0.6, which is used to adjust the impact of the distance on the weight. For example, if the Chebyshev distance is 0.3 and the second correction coefficient is 0.5, then the second dynamic control factor is 0.3 × 0.5 = 0.15.

[0067] The first dynamic control factor is multiplied by the second dynamic control factor to obtain the dynamic control weight. This weight takes into account the vector's distribution and distance from the reference center. Vectors with higher distribution and closer distances receive a larger weight, while vectors with lower distribution and closer distances receive a smaller weight. For example, in the above example, the dynamic control weight is 0.5888 × 0.15 = 0.08832, which will be used in subsequent weighted fusion.

[0068] The global feature reconstruction unit uses a set of dynamic control weights to perform weighted fusion on the set of reduced-dimensionality control vectors to generate an execution control feature map. During weighted fusion, each reduced-dimensionality control vector is multiplied by its corresponding dynamic control weight, and all weighted vectors are then summed to produce the final execution control feature map. For example, 100 reduced-dimensionality vectors are multiplied by their respective weights and added together to form a composite vector, which is the execution control feature map. The value of each dimension reflects the importance of the corresponding feature in global control.

[0069] In terms of hardware implementation, the execution control module can utilize either an ASIC (application-specific integrated circuit) or a high-performance CPU as its processing core. ASICs are customized for specific operations such as dimensionality reduction and distribution calculation, offering advantages such as low power consumption and high speed. The CPU, through multi-threaded programming, enables parallel computing, making it suitable for handling complex weight distribution and fusion operations. High-speed cache and bus architecture within the module enable rapid data transmission, ensuring real-time control requirements.

[0070] Example 4: In this example, the specific implementation of the output feedback module and the composition and workflow of the physical control module are as follows:

[0071] The core function of the output feedback module is to convert the execution control feature map into a concentration control output signal. This module integrates an output signal generator based on a decision tree. The execution control feature map, output by the execution control module, is a multidimensional feature vector containing all the characteristic information required for the current disinfectant concentration control. The output signal generator incorporates a pre-built decision tree model, trained using historical concentration control data, that analyzes the input feature map and generates corresponding control instructions.

[0072] After executing the control feature map input and output feedback module, the signal generator first preprocesses the feature map, including normalization, mapping the feature values ​​of each dimension to a unified numerical range to avoid the impact of different dimensions on the decision results. After the preprocessing is completed, the feature map data enters the root node of the decision tree model, and the decision tree traverses downward layer by layer according to the preset feature judgment conditions. For example, if a dimension in the feature map represents the difference between the current concentration and the target concentration, the decision tree will first determine whether the difference exceeds the preset threshold: if the difference is greater than the positive threshold, it means that the current concentration is low and an instruction to increase the filling amount needs to be generated; if the difference is less than the negative threshold, it means that the concentration is high and an instruction to reduce the filling amount needs to be generated; if the difference is within the threshold range, an instruction to maintain the current filling amount may be generated.

[0073] Each internal node of the decision tree corresponds to a feature judgment condition, and each leaf node corresponds to a specific control instruction. Through this hierarchical judgment logic, the output signal generator can comprehensively judge the multi-dimensional features in the feature graph and generate an accurate concentration control output signal. This signal is output in the form of an electrical signal, such as a TTL level signal or a 4-20mA current signal. The signal contains information such as the control direction (increase, decrease, maintain) and the control amplitude, which is used to indicate the adjustment instruction of the disinfectant concentration.

[0074] The physical control module consists of a hardware controller and a communication interface, which work together to realize the physical execution of control instructions. The hardware controller is embedded with a dedicated processing chip, which usually adopts a high-performance microprocessor or digital signal processor (DSP) with high-speed data processing and real-time control capabilities. The main function of the hardware controller is to convert the execution control characteristic diagram into an optimized adjustment sequence. After receiving the execution control characteristic diagram, the dedicated processing chip first parses the characteristic diagram and extracts the control characteristics therein, such as the degree of concentration deviation, the rate of change of deviation, the pipeline flow status and other information. Based on this information, the chip generates a series of orderly adjustment steps according to the preset control algorithm, that is, the optimized adjustment sequence. This sequence contains specific adjustment parameters, such as the start and stop time of the filling pump, the speed adjustment amount, the valve opening change, etc. Each step corresponds to a specific physical execution action to ensure the accuracy and orderliness of the control process.

[0075] The communication interface is set near the disinfectant filling channel and adopts a waterproof and dustproof industrial-grade interface design to ensure stable operation in complex environments. The main function of the communication interface is to synchronously collect flow and pressure data and perform multi-dimensional cross-integration with the optimization adjustment sequence. Flow sensors and pressure sensors are installed in the filling channel to monitor the liquid flow and pressure values ​​in the channel in real time. The communication interface is connected to these sensors through a dedicated cable to collect flow and pressure data at a fixed sampling frequency (such as 100Hz). The data obtained includes instantaneous flow value, cumulative flow value, real-time pressure value, etc.

[0076] After acquiring the flow and pressure data, the communication interface integrates this data with the optimized adjustment sequence in multiple dimensions. During the integration process, the flow and pressure data are first time-synchronized with each step in the adjustment sequence to ensure the correspondence between the data and the adjustment action. The impact of the flow and pressure data on the adjustment effect is then analyzed. For example, when the channel pressure is high, the same change in valve opening may result in a smaller change in the actual flow rate. In this case, the parameters in the adjustment sequence need to be corrected based on the pressure data. Integration also includes incorporating flow and pressure data as feedback information into the adjustment sequence to form closed-loop control logic. For example, the speed of the filling pump is adjusted based on real-time flow data to compensate for the impact of changes in pipeline resistance.

[0077] After multi-dimensional cross-integration, the communication interface generates a composite regulation protocol. This protocol contains all the execution steps of the optimized regulation sequence, the corresponding flow and pressure constraints, and the feedback correction mechanism, and is encapsulated in a specific communication protocol format, such as the Modbus protocol or the CANopen protocol, to ensure the reliability and compatibility of data transmission. The composite regulation protocol is transmitted to the execution terminal through a wired channel (such as a shielded twisted pair or optical fiber). The execution terminal is usually the control unit of the decontamination liquid filling equipment, such as a PLC (programmable logic controller). After receiving the protocol, the execution terminal parses the instructions and parameters therein and controls the corresponding actuators (such as filling pumps, solenoid valves, etc.) to complete the concentration adjustment action, such as increasing the filling amount of the decontamination liquid to increase the concentration, or reducing the filling amount to reduce the concentration.

[0078] In terms of hardware implementation, the output feedback module utilizes an embedded circuit board design, integrating the decision tree model's computing chip and signal output driver circuitry to ensure real-time signal generation and output. The physical control module's hardware controller utilizes an industrial-grade processing chip with anti-interference capabilities and wide-temperature operating characteristics. The communication interface utilizes a high-speed communication chip, supports multiple industrial communication protocols, and is equipped with signal isolation circuitry to prevent external interference from affecting data transmission. At the software level, the output feedback module's decision tree model is trained and generated using a machine learning algorithm. The physical control module's control algorithm and communication protocol parser are written in C or C++ to ensure system stability and maintainability.

[0079] Example 5: In this embodiment, the specific implementation of the feature dynamic distribution degree calculation unit and the feature dynamic weight allocation unit, as well as the hardware details of the physical control module are as follows:

[0080] When calculating the median vector of the dimensionality reduction control vector, the characteristic dynamic distribution degree calculation unit needs to perform median calculation on the corresponding dimensions of all vectors in the dimensionality reduction control vector set. For example, the dimensionality reduction control vector is 32-dimensional, and the set contains 100 vectors. Then, for each dimension (such as the 1st dimension, the 2nd dimension... the 32nd dimension), extract the values ​​of 100 vectors in that dimension respectively, sort these values ​​from small to large, and take the middle value to form a 32-dimensional median vector. This calculation method can avoid the influence of extreme values ​​on the median, and ensure that the median vector can represent the central trend of the vector set. The calculation of the standard deviation vector is to calculate the standard deviation of 100 values ​​for each dimension to reflect the degree of discreteness of the data in that dimension. For example, the larger the standard deviation of the values ​​of a certain dimension, the more dispersed the data distribution of that dimension.

[0081] After obtaining the median vector and standard deviation vector, each reduced-dimensionality control vector is subtracted from the median vector element by element. For example, for a 32-dimensional reduced-dimensionality vector, the difference between its i-th dimension and the i-th dimension of the median vector is the deviation value for that dimension. The deviation values ​​for all dimensions form the difference vector. Each element of the difference vector is then cubed. For example, a deviation value of 2 is cubed to 8, while a deviation value of -1.5 is cubed to -3.375. The purpose of the cubing operation is to amplify large deviation values, making abnormal features of the data distribution more prominent and facilitating subsequent analysis of the data's dispersion.

[0082] To calculate the overall median of the manipulated difference vector, the values ​​for all 32 dimensions of the manipulated difference vector (i.e., the cubed difference vector) must be collected to form a set of 32 data points. These data points are then sorted and the median value is taken (if the number of data points is even, the average of the two middle values ​​is taken). This overall median calculation method can prevent the impact of extreme deviations in a single dimension on the results. For example, if the deviation value of one dimension is extremely large, but the deviation values ​​of other dimensions are relatively small, the overall median can still reflect the deviation level of most dimensions, ensuring the robustness of the distribution calculation.

[0083] When multiplying the overall median value by the cube of the standard deviation vector, the cube of each dimension of the standard deviation vector must first be calculated. The overall median value is then multiplied by the cube of each dimension. The resulting product is then input into a normalization function. Normalization functions typically employ linear transformations, such as mapping the product to the interval [0, 1]. The formula is: Normalized value = (product result - minimum value) / (maximum value - minimum value), where the minimum and maximum values ​​are the extreme values ​​of the product accumulated from historical calculations. This normalization process makes the distribution values ​​of different dimensionality reduction control vectors comparable, facilitating the subsequent selection of the initial control center vector.

[0084] In the characteristic dynamic weight allocation unit, the first correction coefficient and the second correction coefficient are system preset parameters, which can be adjusted according to the needs of the actual control scenario. For example, in scenarios that are sensitive to concentration changes, the first correction coefficient can be set to 1.0 to make the distribution value have a more direct impact on the weight; in scenarios where distance factors need to be balanced, the second correction coefficient can be set to 0.6 to enhance the regulatory effect of the Chebyshev distance on the weight. The calculation of the first dynamic control factor requires first obtaining the distribution value of the dimensionality reduction control vector and the distribution value of the initial control center vector, multiplying the two together, and then multiplying by the first correction coefficient. For example, the distribution degree of a dimensionality reduction vector is 0.75, the distribution degree of the initial center vector is 0.85, and the first correction coefficient is 0.9. Then the first dynamic control factor is 0.75×0.85×0.9=0.57375.

[0085] The Chebyshev distance is calculated by taking the maximum absolute value of the difference between the reduced-dimensional control vector and the initial control center vector in each dimension. For example, if the absolute values ​​of the differences between the two vectors in 32 dimensions are 0.1, 0.3, 0.2, …, and 0.5, respectively, with the maximum value being 0.5, then the Chebyshev distance is 0.5. The second dynamic control factor is obtained by multiplying the Chebyshev distance by the second correction coefficient. For example, when the second correction coefficient is 0.5, the second dynamic control factor is 0.5 × 0.5 = 0.25. The final dynamic control weight is the product of the two factors, namely 0.57375 × 0.25 = 0.1434375. This weight is used in subsequent weighted fusion to give vectors with high distribution and close distance to the reference center a greater proportion in the reconstruction.

[0086] The dedicated processing chip embedded in the hardware controller of the physical control module usually adopts the ARM Cortex-A series processor or FPGA chip. Taking the ARM processor as an example, it has an internal integrated high-speed cache and multi-core processing unit, which can quickly parse and execute the control feature map and generate an optimized adjustment sequence. The chip is connected to the storage unit through a bus interface. The storage unit pre-stores the control algorithm and the adjustment parameter table, such as the filling amount adjustment mapping table corresponding to different concentration deviations. When the processing chip executes the control algorithm, it first extracts the concentration deviation, deviation change rate and other features in the feature map, and then generates specific adjustment steps according to the adjustment parameter table. For example, the first step is to control the filling pump speed to increase by 10%, and the second step is to adjust the valve opening after a delay of 500 milliseconds, etc., forming an orderly optimized adjustment sequence.

[0087] The communication interface uses RS-485 or Ethernet interface, equipped with ESD electrostatic protection and surge protection circuits to ensure stable data transmission in industrial environments. The interface is connected to the flow sensor (such as electromagnetic flowmeter) and pressure sensor (such as diffused silicon pressure transmitter) of the filling channel through a dedicated cable, and collects flow and pressure data with a sampling period of 200ms. The collected data includes instantaneous flow value (unit: L / min), cumulative flow value (unit: m 3 ), pipeline pressure value (unit: MPa), etc. These data are converted into digital signals by the ADC conversion circuit of the communication interface and temporarily stored in the buffer.

[0088] During multi-dimensional cross-integration, the communication interface aligns the flow and pressure data with the optimized adjustment sequence according to the timestamp. For example, the nth step in the adjustment sequence corresponds to time t. The communication interface will search for the flow and pressure data before and after time t and analyze the pipeline working conditions when the action is executed. If the pressure value is found to be abnormally high during the integration process (such as exceeding the preset threshold of 1.5MPa), the adjustment sequence correction mechanism will be automatically triggered to reduce the valve opening adjustment amount of this step to avoid damage to the pipeline due to excessive pressure. The integrated composite adjustment protocol contains information such as adjustment steps, flow and pressure constraints, and abnormal correction rules. It is encapsulated in binary format and is transmitted to the execution terminal through a wired channel. A CRC check is first performed to ensure the accuracy of data transmission.

[0089] After the execution terminal receives the composite control protocol, the PLC parses the instructions within the protocol, controlling the VFD of the filling pump to adjust the speed or driving the electric valve to adjust the opening. For example, if the protocol instructions require a 20% increase in the filling volume, the PLC calculates the VFD's target frequency based on the current flow data, sends a 4-20mA current signal to the VFD, and simultaneously monitors the valve opening feedback signal to ensure accurate execution of the control action. The hardware controller and communication interface of the physical control module are enclosed in a metal casing that meets the IP65 protection rating and can adapt to harsh environments such as moisture and dust. The internal circuit board is treated with a three-conformal coating to enhance corrosion resistance.

[0090] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0091] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A real-time control system for the concentration of disinfectant, characterized in that: include: Concentration acquisition module, used to obtain real-time concentration sampling data in the disinfectant circulation pipeline; A parameter analysis module, configured to perform hierarchical decomposition and feature extraction on the real-time concentration sampling data to obtain a set of concentration feature vectors; a control strategy generation module, configured to perform association optimization on the set of concentration feature vectors to obtain a set of optimized control vectors; an execution control module, configured to globally reconstruct the set of optimized control vectors to obtain an execution control feature graph; An output feedback module is used to generate a concentration control output signal according to the execution control characteristic diagram.

2. A real-time control system for cleaning solution concentration according to claim 1, characterized in that: The concentration acquisition module includes: a sensor array unit, which is used to divide the real-time concentration data in the disinfectant circulation pipeline according to a preset sampling period to obtain a set of concentration sampling units; and a data preprocessing unit, which is used to input the data of each unit in the set of concentration sampling units into a processor based on dynamic filtering to obtain a set of concentration feature vectors.

3. A real-time control system for cleaning solution concentration according to claim 2, characterized in that: The control strategy generation module includes: a feature vector reorganization unit, which reorganizes the concentration feature vector along the time dimension to obtain a set of reorganized control vectors; a feature correlation calculation unit, which is used to calculate the correlation between adjacent vectors in the set of reorganized control vectors to generate a control correlation matrix; a weight correction unit, which is used to adjust the weight of the control correlation matrix according to the temporal relationship between adjacent vectors in the set of reorganized control vectors to obtain a constraint correlation matrix; and a strategy optimization unit, which is used to convolute and fuse the constraint correlation matrix with the set of reorganized control vectors to obtain the optimized control vector.

4. A real-time control system for cleaning solution concentration according to claim 3, characterized in that: The feature correlation calculation unit includes: a state space mapping subunit, which is used to map each feature vector in the set of reorganized control vectors to the state density space to obtain a set of mapped reorganized control vectors; a correlation analysis subunit, which is used to calculate the cosine correlation coefficient between adjacent mapping vectors in the set of mapped reorganized control vectors to generate the control correlation matrix composed of multiple control correlation values.

5. A real-time control system for cleaning solution concentration according to claim 4, characterized in that: The strategy optimization unit is specifically implemented as follows: performing deep convolution processing on the constraint correlation matrix to obtain a constraint correlation feature matrix; inputting the set of reorganized control vectors and the constraint correlation feature matrix into a gated recurrent network to obtain a set of context control vectors; The set of context control vectors is stacked to obtain the optimized control vector.

6. A real-time control system for cleaning solution concentration according to claim 5, characterized in that: The execution control module includes: an optimized feature dimensionality reduction unit, which is used to perform channel dimension mean pooling processing on each vector in the set of optimized control vectors to obtain a set of reduced-dimensionality control vectors; a feature dynamic distribution degree calculation unit, which is used to calculate the distribution degree value of each feature vector in the set of reduced-dimensionality control vectors to generate a control distribution degree set; a control reference center determination unit, which is used to select the reduced-dimensionality control vector corresponding to the maximum distribution degree value in the control distribution degree set as the initial control center vector; a feature dynamic weight allocation unit, which is used to calculate the dynamic control weight of each feature vector in the set of reduced-dimensionality control vectors and the initial control center vector based on the logical distance between each feature vector and the distribution degree value of each feature vector to generate a dynamic control weight set; a global feature reconstruction unit, which is used to use the dynamic control weight set to perform weighted fusion on the set of reduced-dimensionality control vectors to generate the execution control feature map.

7. A real-time control system for cleaning solution concentration according to claim 6, characterized in that: The characteristic dynamic distribution degree calculation unit is specifically implemented as follows: calculating the median vector and the standard deviation vector of the dimension reduction control vector; performing element-by-element difference calculation between the dimension reduction control vector and the median vector, and performing a cubic operation on the difference result to obtain a control difference vector; Calculate the overall median value of the control difference vector; multiply the median value by the cube of the standard deviation vector, and input the result into a normalization function to obtain the distribution value.

8. A real-time control system for cleaning solution concentration according to claim 7, characterized in that: The characteristic dynamic weight allocation unit is specifically implemented as follows: multiplying the distribution value of the reduced-dimensionality control vector and the distribution value of the initial control center vector by a first correction coefficient to obtain a first dynamic control factor; multiplying the Chebyshev distance between the reduced-dimensionality control vector and the initial control center vector by a second correction coefficient to obtain a second dynamic control factor; and performing a product operation on the first dynamic control factor and the second dynamic control factor to obtain the dynamic control weight.

9. A real-time control system for cleaning solution concentration according to claim 8, characterized in that: The output feedback module is specifically implemented as follows: inputting the execution control characteristic diagram into an output signal generator based on a decision tree to obtain the concentration control output signal, and the concentration control output signal is used to indicate the adjustment instruction of the disinfectant concentration.

10. A real-time control system for cleaning solution concentration according to claim 9, characterized in that: It also includes a physical control module, which is composed of a hardware controller and a communication interface, wherein: the hardware controller has an embedded dedicated processing chip for converting the execution control characteristic diagram into an optimized adjustment sequence; the communication interface is set in the disinfectant filling channel, for synchronously collecting flow pressure data, and generating a composite adjustment protocol after multi-dimensional cross-integration with the optimized adjustment sequence; the composite adjustment protocol is transmitted to the execution terminal through a wired channel to complete the concentration adjustment action.

Images