Method and system for monitoring and controlling production quality of water-soluble fertilizer

Abstract

The invention relates to the cross technical field of agricultural chemistry and intelligent process control, and discloses a production quality monitoring and control method and system for a water-soluble fertilizer. The method comprises the following steps: acquiring a multi-source voiceprint signal through a reaction kettle outer wall acoustic acquisition array; preprocessing and constructing an acoustic feature vector space; inputting a deep learning state mapping model to output a dissolved state level; kinetic parameters are calculated in combination with a rheological model, and a dissolution end point is cooperatively judged; therefore, the stirring speed and the thermal compensation parameter are dynamically adjusted. The system comprises an acoustic acquisition module, a signal processing module, a state evaluation module, a central control module and a data storage module. According to the invention, through non-intrusive acoustic perception and model fusion, accurate, reliable and closed-loop quality monitoring and control are realized.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of agricultural chemistry and intelligent process control, specifically relating to a method and system for monitoring and controlling the production quality of water-soluble fertilizers. Background Technology

[0002] As modern agriculture develops towards precision and efficiency, water-soluble fertilizers, as a new type of fertilizer that can dissolve quickly in water and is easily absorbed by crops, play a crucial role in improving fertilizer utilization and ensuring the quality of agricultural products. Their large-scale industrial production is highly dependent on automated control technology.

[0003] Monitoring the dissolution state and automatically controlling the quality of water-soluble fertilizers during production is a crucial step in ensuring uniform fertilizer composition and accurate formulation. This technology provides data support for optimizing and adjusting production process parameters by real-time sensing of the physical state of materials inside the reactor, thus achieving closed-loop quality management throughout the entire production process.

[0004] Existing water-soluble fertilizer production monitoring technologies suffer from the following problems under complex operating conditions: Due to the generally high salinity and strong corrosiveness of production systems, traditional contact sensors such as conductivity meters and viscometers are susceptible to strong chemical corrosion or surface scaling. This leads to sluggish response or signal distortion of sensor probes due to crystallization, causing delays in control decisions. Traditional detection methods struggle to capture the fluid dynamics changes during solid-liquid interactions, exhibiting real-time deficiencies and blind spots in determining the dissolution endpoint, and failing to accurately pinpoint the exact moment of complete reaction. These problems not only waste production resources and increase equipment maintenance costs but also restrict the stability of water-soluble fertilizer product quality. Therefore, developing a production quality monitoring technology that can circumvent harsh environmental interference and achieve accurate status perception has become a pressing issue for the industry. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for monitoring and controlling the production quality of water-soluble fertilizers, solving the problems mentioned in the background art. By deeply integrating non-invasive acoustic sensing technology with rheological models, this invention aims to eliminate the data distortion and equipment wear problems of traditional contact sensors in high-salt and highly corrosive environments, while also solving the sensing blind spot in determining the dissolution endpoint, achieving precise monitoring and closed-loop optimization of the entire production process.

[0006] To achieve the above objectives, the present invention provides a method for monitoring and controlling the production quality of water-soluble fertilizers, comprising the following steps:

[0007] S1. By deploying an acoustic acquisition array on the outer wall of the reactor, multi-source acoustic signature signals during the production process of water-soluble fertilizer are acquired in real time. The acoustic acquisition array includes multiple high-sensitivity piezoelectric acoustic sensors distributed at different levels on the outer wall of the reactor. Each sensor is fixed by a magnetic adsorption device and its contact surface is filled with acoustic coupling agent. It captures broadband vibration signals that reflect the evolution of the macroscopic flow state of the fluid, pulse acoustic signals that reflect the degree of solid material residue, and cavitation acoustic signals that reflect the microscopic dynamics of dissolution.

[0008] S2. The multi-source acoustic signature feature signal is preprocessed and reconstructed to construct an acoustic feature vector space characterizing the dissolution state of the material. The preprocessing is carried out by establishing a background noise benchmark model and calculating the cross-correlation between the measured signal and the background model to remove background mechanical noise. The one-dimensional time series acoustic signal is converted into a two-dimensional time-frequency feature spectrum using short-time Fourier transform. At the same time, the converted spectrum is normalized by mean removal and variance scaling.

[0009] S3. Input the acoustic feature vector space into the pre-constructed deep learning state mapping model and output the current material dissolution state level; the deep learning state mapping model extracts local texture features and global contour features in the acoustic feature map through a multi-scale convolutional network structure, and integrates an attention mechanism to learn the core acoustic frequency points corresponding to different dissolution stages.

[0010] S4. Combine the rheological model to calculate the dynamic parameters of the fluid in the reactor, and determine the dissolution endpoint in conjunction with the dynamic parameters and the material dissolution state level; establish a rheological coupling equation based on stirring power, fluid density and sound velocity attenuation rate, calculate the equivalent viscosity and flow index, and estimate the real-time volume fraction of undissolved solids in the fluid by analyzing the attenuation coefficient of sound wave energy.

[0011] S5. Based on the determination result of the dissolution endpoint, generate production process control instructions to adjust the stirring speed and thermal compensation parameters of the reactor in real time.

[0012] Preferably, the sensor distribution of the acoustic acquisition array is as follows:

[0013] Multiple sensor groups are set at the bottom, middle liquid level line and top stirring layer of the reactor. The sensor groups are arranged at equal intervals in the circumferential direction. The sound field changes in different spatial regions inside the reactor are obtained through a synchronous triggering mechanism.

[0014] The broadband vibration signal reflects the change in damping coefficient as the fluid transitions from a rarefied to a viscous state through frequency drift and energy intensity variation.

[0015] The pulsed acoustic signal determines the particle size distribution and number density of solid particles in the fluid based on the transient energy spikes generated by solid particles impacting the vessel wall.

[0016] The cavitation acoustic signal is based on the high-frequency signal generated by bubble collapse to monitor the intensity of chemical reactions and local phase transitions during the dissolution process.

[0017] Preferably, the feature reconstruction process includes: applying a window function to the digital sequence after background noise removal to reduce spectral leakage caused by the truncation effect;

[0018] By introducing cepstral analysis logic, the Mel frequency cepstral coefficients of the acoustic signal are extracted, separating the mechanical excitation source characteristics and fluid system response characteristics in the acoustic signal, and capturing acoustic wave fluctuations that are sensitive to the dissolution state.

[0019] The acoustic feature vector space also incorporates an environmental compensation dimension. By acquiring the real-time temperature of the reactor wall, the temperature parameter is input into the model as a continuous variable along with the acoustic features. The compensation layer in the deep learning network is used to correct the influence of environmental temperature on the sound wave propagation speed and attenuation rate.

[0020] Preferably, step S3 specifically includes the following steps:

[0021] S31. The acoustic feature vector space is decomposed in multiple dimensions to obtain the time-domain energy distribution characteristics that reflect the overall activity of the dissolution process, the frequency-domain power spectral density characteristics used to locate the characteristic response of frequency points, and the time-frequency joint distribution characteristics that reflect the evolution of signal energy over time.

[0022] S32. Calculate the correlation between each dimension of features and the material's dissolution state using a feature weighting algorithm, and extract a subset of key features. The feature weighting algorithm calculates the correlation coefficient between each acoustic feature component and the target value of the dissolution concentration, marks candidate features with a correlation higher than a preset threshold, and executes principal component analysis logic to compress high-dimensional interrelated features into multiple independent principal components.

[0023] S33. Based on a deep neural network, a nonlinear mapping is performed on a subset of key features to output the probability distribution of the continuous state of the material from initial solid accumulation to complete dissolution.

[0024] Preferably, the deep neural network uses a combination of cross-entropy loss function and mean square error loss function during the training phase. By dynamically adjusting the weight coefficients, it identifies the stage of the material in the early and middle stages of dissolution and provides quantitative kinetic parameters at the end point. The attention mechanism learns the correlation between feature channels, automatically identifies the core acoustic frequency points of different production stages, and triggers early warnings for abnormal crystallization or material agglomeration.

[0025] Preferably, the rheological model calculation process includes:

[0026] Obtain the current, voltage, and speed parameters of the stirring motor, and calculate the actual stirring power;

[0027] A preset excitation source drives a designated acoustic sensor to emit a standard pulse signal. The signal time difference between the transmitter and the receiver of other sensor groups is obtained through a synchronous triggering mechanism. Combined with the pre-calibrated sound wave propagation path length, the propagation speed of the sound wave in the fluid is calculated.

[0028] By analyzing the scattering characteristics and energy attenuation coefficient of sound waves in heterogeneous fluids, an acoustic feedback distribution function is established.

[0029] The calculation of the equivalent viscosity takes into account the characteristics of non-Newtonian fluids. By analyzing the phase difference of acoustic waves received by sensors at different vertical heights, the local flow velocity of the fluid at different levels is calculated. The shear rate distribution field inside the fluid is constructed by combining the geometry of the stirring impeller. The rheological coefficient in the rheological model is corrected in real time using the wall shear stress distribution fed back by the acoustic signal.

[0030] Preferably, the synergistic determination of the dissolution endpoint includes:

[0031] Set preset thresholds for the material dissolution state level and stability indices for kinetic parameters;

[0032] When the current material dissolution state level reaches a preset threshold and the calculated equivalent viscosity fluctuation rate remains below a predetermined change step within a preset time period, it is determined to be the dissolution endpoint.

[0033] The judgment process also incorporates a confidence assessment based on logistic regression. By combining the classification probability distribution of the deep learning model, the stationarity index of the dynamic parameters, and the time statistical characteristics of similar batches in history, a percentage value reflecting the accuracy of the judgment is calculated. When this percentage value is lower than a preset threshold, the observation time is extended.

[0034] Preferably, the generation of production process control instructions includes:

[0035] In the initial stage of dissolution, the frequency of the inverter of the stirring motor is increased according to the energy intensity of the pulse acoustic signal, thereby increasing the stirring speed to enhance solid-liquid convection mass transfer.

[0036] During the middle stage of dissolution, the thermal compensation parameters are adjusted according to the change in the sound velocity decay rate. The temperature inside the vessel is maintained within the preset dissolution temperature range by controlling the opening of the steam valve or adjusting the electric heating power.

[0037] At the end of the dissolution process, reduce the output frequency of the stirring motor to reduce energy consumption and eliminate bubble interference;

[0038] The production process control instructions are generated using a predictive control algorithm. Based on the rate of change of the current solubility level, the material state trend in the future time period is predicted, and the thermal compensation instructions are fine-tuned in advance to suppress local overheating and agglomeration.

[0039] This invention provides a production quality monitoring and control system for water-soluble fertilizers. The system utilizes the aforementioned method for production quality monitoring and control of water-soluble fertilizers, comprising:

[0040] The acoustic acquisition module is equipped with a non-invasive acoustic sensor group installed on the outer wall of the reactor to collect full-frequency acoustic signals during the production process; the non-invasive acoustic sensor group is fixed by magnetic adsorption or coupling agent to realize the transmission of acoustic energy through the reactor wall;

[0041] The signal processing module, connected to the acoustic acquisition module, is used to realize signal noise reduction, transformation and feature extraction; the signal processing module is equipped with a high-speed digital signal processor, which performs synchronous analog-to-digital conversion on multiple acoustic signals at a predetermined sampling frequency, and uses a ping-pong buffer mechanism to realize the temporal overlap of data acquisition and mathematical processing;

[0042] The status assessment module integrates a deep learning engine to assess the material dissolution progress and fluid dynamics state based on feature signals; the status assessment module is also equipped with an anomaly diagnosis submodule, which identifies equipment fault status by comparing with a standard voiceprint template.

[0043] The central control module is connected to the status assessment module and the actuator of the production equipment, and is used to output frequency conversion control signals and thermal regulation signals according to the assessment results.

[0044] The data storage module is used to record voiceprint data, characteristic parameters, and final quality judgment results throughout the entire production process, and to build a quality traceability database.

[0045] Preferably, the system further includes a human-computer interaction module for real-time display of the dissolution progress curve, acoustic signature spectrum, and dynamic parameter change trends;

[0046] The abnormality diagnosis submodule periodically executes an acoustic transparency self-test procedure. During the reactor evacuation phase, it emits a standard pulse through an excitation source and analyzes the echo characteristics to assess the degree of fouling on the reactor wall and automatically updates the input correction coefficients of the deep learning model.

[0047] The central control module has a reserved application programming interface for transmitting real-time data of the dissolving endpoint to the production management system to achieve dynamic optimization of production scheduling.

[0048] The acoustic acquisition module utilizes beamforming technology to introduce time delays into the signals of each channel of the sensor array, thereby enabling the focusing of the internal region of the reactor at the algorithm level, locating and eliminating localized areas of lag in dissolution.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. This invention overcomes the survival challenges of traditional contact sensors in the high-salt, highly corrosive production environment of water-soluble fertilizers by introducing a non-invasive acoustic sensing solution. By deploying the acoustic sensor array on the outer wall of the reactor, the system is no longer affected by chemical corrosion, electrochemical corrosion, or surface scaling and crystallization, thus solving the control distortion problem caused by sensor data drift or damage. This not only extends the hardware lifespan of the monitoring system and reduces maintenance frequency and costs, but also ensures the continuity and accuracy of production data.

[0051] 2. This invention overcomes the limitations of traditional single-indicator judgment of the dissolution endpoint by synergistically combining a deep learning state mapping model and a rheological model. The deep learning model can automatically extract subtle features highly correlated with the material state from complex acoustic signals, while the rheological model provides quantitative support for fluid dynamics from a physical perspective. The combination of the two eliminates the perception blind spots in the dissolution process. This allows the production process to accurately capture the critical point at which solid particles completely disappear, avoiding both uneven product composition due to insufficient stirring time and energy waste caused by excessive stirring, thus improving production efficiency and product quality stability.

[0052] 3. By collecting multi-source acoustic signature signals, this invention achieves comprehensive perception of the microscopic dynamic processes inside the reactor. Acoustic information from different dimensions, such as stirring interaction, collision sounds, and cavitation sounds, corresponds to macroscopic flow, microscopic mixing, and phase change processes, respectively. This multi-dimensional information fusion enables the system to perform targeted optimization at each stage of the entire production cycle, realizing a shift from experience-driven to data-driven production modes. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0054] Figure 2 This is a schematic diagram of the core principle framework of the present invention, which is based on the collaborative determination of the dissolution endpoint using acoustic features and rheological models.

[0055] Figure 3 This is a logical flowchart of the preprocessing of multi-source voiceprint feature signals and the construction of acoustic feature vector space in this invention;

[0056] Figure 4This is a logical framework diagram of the material dissolution state assessment based on the deep learning state mapping model in this invention;

[0057] Figure 5 This is a schematic diagram of the closed-loop control and real-time parameter adjustment of the production process based on the determination of the dissolution endpoint in this invention. Detailed Implementation

[0058] Example 1: Reference Figures 1 to 5 This embodiment describes in detail a method for monitoring and controlling the production quality of water-soluble fertilizers, aiming to solve the failure problem of traditional contact sensors in high-salt and highly corrosive environments, and to achieve precise closed-loop control of the dissolution endpoint.

[0059] In the industrial production process of water-soluble fertilizers, the evolution of materials inside the reactor is a complex physicochemical process.

[0060] Step S1 involves acquiring multi-source acoustic signature signals in real time during the water-soluble fertilizer production process using an acoustic acquisition array deployed on the outer wall of the reactor. In this embodiment, the acoustic acquisition array consists of 12 high-sensitivity piezoelectric acoustic sensors distributed at different levels on the outer wall of the reactor. These sensors are fixed to the outer wall of the reactor (made of carbon steel or stainless steel) using a high-strength magnetic adsorption device, and a special ultrasonic coupling agent is applied to the contact surface to ensure that acoustic energy can pass through the reactor wall from the internal medium and be transmitted to the sensing element with low loss.

[0061] In terms of spatial layout, the acoustic sensor array employs a strategy of alternating vertical and circumferential arrangement. Four sensor groups are positioned at the bottom, middle liquid level line, and top stirring layer of the reactor, with each group arranged at 90-degree intervals along the circumference. This omnidirectional sensing layout can capture changes in the sound field in different spatial regions within the reactor. The multi-source acoustic signature signals are not merely noise but also contain rich information reflecting the production status.

[0062] The broadband vibration signal generated by the interaction between the stirring device and the fluid has its frequency components mainly concentrated in the low frequency band. By acquiring the frequency drift and energy intensity changes of this broadband vibration signal, the evolution of the macroscopic flow state of the fluid can be reflected in real time. For example, when the fluid changes from a rarefied state to a viscous state, the damping coefficient of the sound wave in the medium will change.

[0063] The pulsed acoustic signals generated by the collision of solid particles with the inner wall of the reactor are a key basis for determining solid residue. In the initial stage of dissolution, a large number of undissolved fertilizer particles impact the reactor wall under the action of the agitator, generating high-energy transient spikes. The trigger frequency and peak pressure of these pulse signals are directly related to the particle size distribution and number density of solid particles in the current fluid. As the dissolution process progresses, the particles gradually shrink until they disappear, and the intensity and frequency of the pulse signals show a decreasing trend.

[0064] The cavitation acoustic signals generated by the collapse of tiny bubbles reflect the microscopic dynamics of the dissolution process. In particular, during dissolution processes accompanied by exothermic or endothermic reactions, the collapse of bubbles caused by local phase transitions emits unique high-frequency signals, which provides a non-contact sensing method for monitoring the intensity of chemical reactions.

[0065] Step S2 involves preprocessing and reconstructing the multi-source acoustic signature signal. Due to the presence of significant mechanical motor noise, pump vibration, and electromagnetic interference at the production site, the original acquired electrical signal contains substantial background noise. This embodiment employs an adaptive filtering algorithm. A baseline model of the background noise is established during the idle phase before production begins, and the cross-correlation between the current signal and the background model is calculated in real-time during actual acquisition. This dynamically removes periodic mechanical noise unrelated to production logic from the composite signal.

[0066] After noise reduction, the one-dimensional time-series acoustic signal is transformed into a two-dimensional time-frequency feature map using short-time Fourier transform (SFT) technology. This process involves selecting a specific time window, performing continuous sliding sampling on the time axis, and converting the signal within each window into the frequency domain to obtain the evolution trajectory of the signal energy with both time and frequency dimensions.

[0067] To eliminate the impact of sensitivity differences between different sensors and gain fluctuations in the acquisition circuit, this embodiment normalizes the converted spectrum. By calculating the statistical average of the full-range data and subtracting it from the original data, and then dividing the processed result by the standard deviation, all data features are mapped to a numerical space that conforms to a standard normal distribution. This provides a high-quality input benchmark for subsequent deep learning models and avoids the gradient explosion or vanishing problem caused by differences in data magnitude.

[0068] Step S3: Input the acoustic feature vector space into the pre-constructed deep learning state mapping model and output the current material dissolution state level.

[0069] Step S31 involves decomposing the acoustic feature vector space into multiple dimensions. The system automatically identifies and separates the time-domain energy distribution, frequency-domain power spectral density, and time-frequency joint distribution features. The time-domain energy distribution mainly reflects the overall activity of the dissolution process, while the frequency-domain power spectral density is used to locate the characteristic response at a frequency point. For example, the energy peak at a specific frequency often corresponds to the collision frequency of particles of a specific size.

[0070] Step S32 involves calculating the correlation between each dimension of features and the material's dissolution state using a feature weighting algorithm. This embodiment employs an information gain evaluation mechanism, prioritizing hundreds of original features by calculating the contribution of different feature components in the sample classification process. Only those key features exhibiting high sensitivity to changes in material concentration and minimal fluctuations due to environmental disturbances are selected into the key feature subset. For example, in a certain formulation, the energy attenuation rate in the 15 kHz to 20 kHz frequency band has a strong linear correlation with the mass fraction of solid particles; this feature is then assigned a higher weight.

[0071] Step S33 involves nonlinear mapping of key feature subsets based on a deep neural network. This deep neural network employs a multi-scale convolutional network structure, using convolutional kernels of different sizes to extract features from the time-frequency feature map. Smaller convolutional kernels focus on extracting local textures in the map, such as fine stripes generated by instantaneous pulses; while larger convolutional kernels are used to capture the global contour of the map, reflecting the macroscopic evolution of the dissolution trend. An attention mechanism is also integrated, which can automatically learn which frequency windows should be focused on at different stages of dissolution. The output reflects the probability distribution of the continuous state of the material from initial solid accumulation to final complete dissolution, and is discretized into dissolution state levels between 0 and 100. The dissolution state level is linearly negatively correlated with the mass fraction of solid residue in the reactor, with level 0 corresponding to a solid residue mass fraction of 100% (initial accumulation state), level 100 corresponding to a solid residue mass fraction of less than 0.1% (completely dissolved state), and intermediate levels corresponding to a linear decreasing relationship of solid residue mass fraction.

[0072] Step S4 involves calculating the dynamic parameters of the fluid inside the reactor using a rheological model. This step relies not only on qualitative judgment based on acoustic characteristics but also on quantitative calculations of the physical essence. The system establishes a set of rheological coupling equations based on stirring power, fluid density, and sound velocity attenuation rate. By acquiring the current, voltage, and speed parameters of the stirring motor in real time, the current actual stirring power is calculated.

[0073] The method of calculating the propagation speed of sound waves in fluid using the transmission and reception time difference between sensor groups is as follows: The system drives a piezoelectric acoustic sensor at a designated location as the transmitter through a preset active excitation source to emit a standard pulse sound wave signal of a preset frequency into the fluid inside the vessel. Other acoustic sensors at the same or different levels act as receivers. The transmitter and receiver are synchronized using a hardware synchronization trigger mechanism, with the rising edge of the transmitter's excitation signal as the starting point of the sound wave transmission time. The receiver performs cross-correlation calculations on the acquired signals, using the peak point of the cross-correlation function as the arrival time of the sound wave, and calculates the propagation time difference of the sound wave between the transmitter and receiver. Combined with the pre-calibrated direct propagation path length of the fluid between the transmitter and receiver, the propagation speed of sound waves is calculated using... The real-time propagation speed of sound waves in the fluid inside the vessel is calculated. During the calculation, the propagation time delay of the vessel wall is pre-calibrated to eliminate interference from the vessel wall on the sound wave propagation time. The signal sampling frequency is no less than 1 MHz to ensure that the time delay estimation accuracy is no less than 1 microsecond. When the fluid contains a large number of solid particles or microbubbles, sound waves will undergo scattering and energy attenuation. By analyzing the attenuation coefficient of the sound wave energy and combining it with preset fluid density parameters, the real-time volume fraction of undissolved solids in the fluid can be estimated.

[0074] Based on this, the system further calculates the current equivalent viscosity and flow index. Considering that water-soluble fertilizers typically exhibit non-Newtonian fluid characteristics, their viscosity is not a constant value but changes with the shear rate. In this embodiment, the rheological coefficients in the rheological model are corrected in real time using the wall shear stress distribution fed back by acoustic signals, obtaining high-precision real-time viscosity data.

[0075] The process of determining the dissolution endpoint employs collaborative logic. The system sets a preset threshold for the dissolution state level (e.g., reaching level 98 or higher) and a stability index for kinetic parameters. When the current material dissolution state level reaches the preset threshold, and the calculated equivalent viscosity fluctuation rate remains below the preset minimum change step size for a preset 60-second time period, the system determines that the material has achieved complete dissolution in a microscopic sense, thus identifying the dissolution endpoint. This dual-index determination mechanism avoids misjudgments caused by single sensor drift or localized uneven mixing.

[0076] Step S5: Based on the dissolution endpoint determination result, generate production process control instructions. In the initial stage of dissolution, when a strong pulse signal generated by solid particle collisions is detected, the central control module outputs a frequency increase command, increasing the inverter output of the stirring motor and raising the stirring speed to enhance convective mass transfer between solid and liquid, preventing fertilizer particles from settling and clumping. In the middle stage of dissolution, based on changes in the sound velocity decay rate, when the dissolution rate slows down, the system automatically adjusts the thermal compensation parameters, increasing the steam valve opening or increasing the electric heating power to maintain the reactor temperature within the optimal dissolution temperature range. At the dissolution endpoint, the system automatically reduces the stirring speed to reduce energy consumption and eliminate bubble interference, achieving final quality confirmation through rheological stability.

[0077] Example 2: This example focuses on a water-soluble fertilizer production quality monitoring and control system for implementing the above-described monitoring and control methods. The system employs a modular hardware architecture and highly reliable software logic.

[0078] The acoustic acquisition module is equipped with a non-invasive acoustic sensor array mounted on the outer wall of the reactor. Each sensor utilizes an industrial-grade piezoelectric ceramic element with a high signal-to-noise ratio, covering a frequency range from low-frequency vibrations of a few hertz to high-frequency ultrasound of hundreds of kilohertz. The sensor housing is encapsulated in a corrosion-resistant special alloy, and the weak charge signal is transmitted to the signal preamplifier via a shielded cable. The non-invasive installation of the sensor array eliminates the need for drilling or welding in the reactor during system deployment, leaving the original pressure vessel structure completely unaltered. This facilitates upgrades for existing production lines.

[0079] The signal processing module is connected to the acoustic acquisition module via a high-speed industrial bus. Internally, this module is equipped with a dedicated high-speed digital signal processor capable of performing billions of operations per second. The module can perform synchronous analog-to-digital conversion on multiple acoustic signals at sampling frequencies up to 2 MHz. In the digital domain, the processor executes complex real-time noise reduction, feature extraction, and short-time Fourier transform logic. To ensure real-time data processing, the module employs a ping-pong buffering mechanism, ensuring complete timing overlap between data acquisition and mathematical processing, eliminating the impact of processing delays on control accuracy.

[0080] The state assessment module, acting as the system's intelligent brain, integrates a high-performance deep learning engine. This engine runs on top of the aforementioned deep neural network model and can receive pre-processed acoustic feature vectors in real time. The assessment module not only outputs the dissolution state level but also includes an anomaly diagnosis submodule. This submodule stores a database of acoustic signature templates from the standard production process.

[0081] By comparing measured acoustic signatures with standard templates in real time, the system can identify abnormal production conditions, such as abnormal vibration frequencies caused by wear on the agitator, cavitation noises caused by malfunctioning circulating pumps, and even changes in acoustic impedance caused by scaling on the reactor walls. Once an anomaly is detected, the system will immediately trigger an early warning signal and display the fault type and suggested maintenance location on the human-machine interface.

[0082] The central control module serves as the command issuing center, connecting to the status assessment module and the actuators of the production equipment, such as frequency converters, proportional control valves, and heating controllers. The central control module communicates with the reactor's control system via the industrial Ethernet protocol, achieving millisecond-level command response. The logic includes a multi-parameter coupled control matrix, dynamically adjusting the weight ratio of stirring speed and thermal compensation according to different stages of the dissolution process, ensuring the dissolution process is completed in the shortest time with the lowest energy consumption.

[0083] The system also includes a data storage module to record raw acoustic signature data, characteristic parameters, and final quality assessment results throughout the entire production process. This data is not only used for current production monitoring but also supports subsequent process optimization by building a quality traceability database. Through long-term analysis of historical data, technicians can discover the impact of different seasons and batches of raw materials on dissolution efficiency, thereby continuously fine-tuning production parameters and improving overall product consistency.

[0084] The system's human-machine interface module utilizes an industrial-grade touchscreen display, capable of dynamically displaying the dissolution progress curve, acoustic signature spectrum, and trends in key kinetic parameters in real time. Operators can understand the microscopic evolution of materials within the reactor through an intuitive graphical interface, replacing the previous method of relying solely on observation windows or experience.

[0085] Example 3: This example further illustrates how, in the method and system of the present invention, monitoring accuracy can be improved by optimizing logic details for specific production environments and material characteristics.

[0086] In the production of water-soluble fertilizers, high-viscosity non-Newtonian fluids are often encountered. For such materials, this embodiment introduces a more detailed shear stress correction logic in the rheological model calculation of step S4. The system calculates the local flow velocity of the fluid at different levels by analyzing the phase difference of the acoustic waves received by sensors at different vertical heights. Using this local velocity gradient, combined with the geometric parameters of the stirring impeller, a shear rate distribution field inside the fluid is constructed.

[0087] The calculation of equivalent viscosity is no longer a single average value, but a distribution function based on acoustic feedback. This deep integration of physical models allows the system to maintain accurate capture of the hydrodynamic state even at the end of the dissolution process when the material viscosity changes drastically.

[0088] At the feature extraction algorithm level, this embodiment details the specific execution process of the feature weighting algorithm in step S32. The system first calculates the Pearson correlation coefficient between each acoustic feature component and the target concentration value. This Pearson correlation coefficient reflects the synchronicity of the feature changes with concentration. For features with absolute correlation coefficient values ​​higher than a preset threshold (e.g., 0.85), the system marks them as candidate key features. Subsequently, to eliminate redundant information between features, the system executes principal component analysis logic, compressing high-dimensional, interrelated features into several independent principal components. These principal components collectively represent more than 90% of the useful information in the voiceprint signal. This processing method reduces the input dimensionality of the deep learning model, improving the model's inference speed while maintaining accuracy, making it possible to run complex neural networks on low-power embedded hardware.

[0089] In the training phase of the deep learning model, to simultaneously ensure the accuracy of qualitative classification and the precision of quantitative prediction, this embodiment employs a composite loss function. This composite loss function combines a cross-entropy term for state level classification and a mean squared error term for predicting kinetic parameters such as viscosity and volume fraction. During training, the weights of these two terms are dynamically adjusted, enabling the model to accurately identify the material's stage in the early and middle stages of dissolution, and to provide quantitative parameters at the endpoint. This training strategy ensures high reliability of the model throughout the entire lifecycle.

[0090] The specific weighting strategy of the composite loss function is implemented through a combined loss function formula, as follows:

[0091]

[0092] in, This represents the total loss value of the deep neural network. The cross-entropy loss weight coefficients are... For cross-entropy loss, The mean squared error loss weighting coefficient is used. This represents the mean squared error loss. The weighting coefficient α is dynamically switched according to the dissolution state level. When the level is less than or equal to 60 (early to mid-dissolution), α equals 0.7, and when the level is greater than 60 (late dissolution), α = 0.3.

[0093] In this embodiment, the omnidirectional spatial sound field perception also utilizes beamforming technology. By introducing a slight time delay into the signals of each channel in the sensor array, the system can achieve focusing on the internal region of the reactor at the algorithm level. If a corner of the reactor experiences dissolution lag due to a dead zone in the stirring, the particle collision sound signal generated in that area will be amplified by spatial filtering logic. Based on this spatial positioning information, the central control module can trigger a local jet disturbance device or adjust the forward and reverse rotation logic of the stirring paddle, thereby eliminating local dead zones and improving mixing uniformity.

[0094] To address the issue of strong interference in the production environment, this embodiment employs a sliding window dynamic mean removal method for normalization. Instead of uniformly normalizing the entire production batch, the system uses the signal average over the past 10 seconds as a reference. This dynamic calibration mechanism can offset slow baseline shifts caused by environmental temperature drift or long-term equipment operation, ensuring that the input to the deep learning model remains within the most sensitive linear region.

[0095] Example 4: This example focuses on describing the application scenario of the present invention in a real large-scale fertilizer production base, and how to achieve clustered production quality control through multi-system collaboration.

[0096] In large-scale production sites, multiple reactors typically operate simultaneously. The system of this invention supports clustered deployment, with the acoustic acquisition modules of each reactor connected to a unified central processing server via an industrial ring network. In this architecture, deep learning models can continuously learn online on the server side. The system automatically collects acoustic signature samples from batches of products deemed high-quality, fine-tunes the model during production breaks, and then redistributes the optimized model parameters to each front-end processing unit. This self-evolving mechanism allows the monitoring system's accuracy to continuously improve with accumulated production experience.

[0097] In terms of quality traceability, the data storage module not only records whether the final dissolution endpoint has been reached, but also completely preserves the acoustic fingerprint of the entire dissolution process. If quality fluctuations are found in a batch of products in downstream processes, technicians can retrieve the original acoustic fingerprint spectrum of that batch during production. By comparison, it can be found that, for example, abnormal fluctuations in the sound wave energy distribution occurred at the 30-minute mark of dissolution, which may correspond to the abnormal introduction of a certain impurity in the raw materials. This in-depth quality traceability capability provides an unprecedented data dimension for the diagnosis and improvement of chemical production processes.

[0098] In the process control command generation logic, this embodiment also introduces a predictive control algorithm. The system not only adjusts based on the current dissolution level, but also predicts the material's state trend within the next 3 minutes based on the rate of change of the current dissolution level. If the prediction indicates that an excessively rapid dissolution rate may lead to localized overheating and agglomeration, the system will preemptively fine-tune the thermal compensation command. This proactive control logic makes the production process smoother and reduces the fluctuation range of process parameters.

[0099] The system integrates full lifecycle equipment health monitoring functions. Through an attention mechanism, the deep learning model is endowed with the ability to identify atypical sound patterns. For example, when the bearing of the stirring shaft experiences minor wear, it will generate discontinuous frictional sound signals in a specific high-frequency range. The attention mechanism captures these weak features that are ignored by traditional sensors and issues pre-maintenance reminders to operators through the human-machine interface module before the failure leads to downtime. This not only ensures the production quality of water-soluble fertilizers but also improves the operational safety and equipment utilization rate of the entire production line.

[0100] In practical industrial implementation, the system's preset time period and stability indicators can be dynamically adjusted for different formulations of water-soluble fertilizers, such as high-phosphorus, high-potassium, or balanced formulations containing multiple trace elements. The system automatically retrieves the corresponding control strategy template from the database based on the formulation code entered by the operator. For formulations with slower dissolution, the stability assessment time will be appropriately increased, and the volatility threshold will be tightened.

[0101] In summary, this invention solves the survival problem of traditional contact sensors in harsh chemical environments through non-invasive acoustic sensing. By integrating deep learning with rheological physical models, it achieves a leap from experience-based production to digital and intelligent production. It not only improves the consistency of water-soluble fertilizer product quality but also achieves energy conservation and emission reduction by optimizing process parameters, demonstrating high industrial application value and economic benefits.

[0102] Example 5: In this example, the construction logic of the acoustic feature vector space and its adaptive processing for complex working conditions will be further refined.

[0103] The feature reconstruction described in step S2 essentially involves constructing a high-dimensional mathematical space that comprehensively and objectively reflects the physicochemical state of materials. In practice, the raw waveform acquired by the acoustic acquisition array first passes through a gain-controlled preamplifier, boosting the microvolt-level electrical signal to a volt level suitable for analog-to-digital conversion. The sampled digital sequence is then processed using a Hanning window function before undergoing a short-time Fourier transform. The Hanning window is used to reduce spectral leakage caused by truncation effects, resulting in a smoother energy distribution in the frequency domain that better reflects real physical phenomena.

[0104] In the feature reconstruction process, this invention introduces cepstral analysis logic. Cepstral analysis can separate the excitation source features in the acoustic signal, such as the mechanical impact of the agitator, from the system response features, such as the resonant characteristics of the fluid inside the vessel. By extracting the Mel-frequency cepstral coefficients of the acoustic signal, the system can simulate the sensitivity of human hearing to different frequency bands, capturing those subtle acoustic wave fluctuations that are invisible to the human eye but sensitive to the state of dissolution. These coefficients, together with the aforementioned time-domain and frequency-domain indices, constitute a raw feature vector with up to several hundred dimensions.

[0105] To enhance the model's robustness to changes in ambient temperature, this embodiment incorporates an environmental compensation dimension into the feature vector space. Although the sensor does not contact the material, the temperature of the vessel wall affects the propagation speed and attenuation rate of sound waves in the metallic medium. The system acquires the real-time temperature using a platinum resistance temperature sensor mounted on the vessel wall and inputs this real-time temperature parameter as a continuous variable, along with the acoustic features, into the model. The compensation layer in the deep learning network automatically learns the nonlinear coupling relationship between temperature and acoustic features, enabling it to output accurate solubility levels under varying ambient temperatures.

[0106] In the dynamic parameter calculation of step S4, this embodiment details the relationship between acoustic energy attenuation and fluid microstructure for the calculation of equivalent viscosity. As fertilizer particles gradually shrink in the solvent, a multiphase suspension system is formed. When sound waves pass through this system, energy loss occurs due to viscous dissipation at the solid-liquid interface and thermal conduction dissipation. This system measures the amplitude attenuation ratio of sound waves along different propagation paths and, combined with the known sound wave emission frequency, inversely calculates the viscous damping coefficient within the fluid. This calculation method, based on wave physics, completely avoids the risks of mechanical measurement methods such as rotational viscometers, thus preventing the mechanical probe from being jammed by crystallization or corroded by acids and alkalis.

[0107] In the collaborative determination of the dissolution endpoint, this embodiment also introduces a confidence assessment based on logistic regression. The system not only provides a switching signal indicating dissolution completion but also a percentage value reflecting the accuracy of the determination. This percentage value combines the classification probability distribution of the deep learning model, the stationarity index of rheological parameters, and the time statistical characteristics of historical batches of the same type. If the confidence level is lower than a preset threshold (e.g., 95%), the system will request an extension of the observation time and prompt manual sampling for verification. This multi-level logical determination improves the fault tolerance of the automated production line.

[0108] When generating process control instructions, the central control module employs a strategy called linear quadratic regulation, but during execution, it is entirely converted into a purely textual logical description. The system first determines the current production target state, i.e., the ideal dissolution curve trajectory, and then calculates the deviation between the current actual state and the target state. If the deviation is mainly reflected in a slow dissolution rate, the system calculates the amount of heat that needs to be compensated and converts it into a pulse width modulation signal for the steam electronically controlled valve. If the deviation is reflected in uneven material mixing, determined by differences in the spatial sound field, the system calculates the increment of the stirring speed and sends it to the frequency converter. The entire control process is a dynamic balancing process, aiming to obtain the most stable process output with minimal control action costs.

[0109] This invention also considers the impact of internal scaling on monitoring. With increasing production cycles, a layer of hardened scale may adhere to the inner wall of the reactor. This scale increases the acoustic impedance when sound waves pass through the reactor wall. To address this issue, the system's anomaly diagnosis submodule periodically performs an acoustic transparency self-check. During the reactor evacuation phase, a standard pulse is emitted using a preset excitation source. By analyzing the received echo characteristics, the current degree of scaling on the reactor wall is assessed, and the input correction coefficients of the deep learning model are automatically updated. This system, with its self-sensing and self-compensation capabilities, can truly adapt to the continuous production needs of a factory over many years.

[0110] Regarding data storage and quality traceability, this embodiment further illustrates the structured design of the database. Each production batch is assigned a unique global identifier, and its associated data items include: raw material batch number, ambient temperature and humidity, full-process acoustic signature energy evolution matrix, kinetic parameter snapshots at each key time point, and the final physicochemical test certificate number. Through cluster analysis of massive historical data, the system can automatically identify potential factors causing product quality fluctuations. For example, the system may discover that when the ambient humidity exceeds a certain critical point, the pre-treatment acoustic signature of the raw materials changes, thereby affecting the subsequent dissolution rate. This process knowledge obtained through big data mining is incomparable to the traditional reliance on the accumulated experience of master craftsmen.

[0111] This invention also features specially designed power failure protection and data recovery logic. In the event of an unexpected power outage during production, the large capacitor built into the signal processing module will maintain system operation for a sufficient period, storing a snapshot of the current production state in non-volatile memory. Once power is restored, the system can quickly locate the current dissolution stage based on acoustic characteristics, reconnecting with previous control commands and preventing the entire batch of material from being scrapped due to downtime.

[0112] In summary, this embodiment demonstrates the depth and breadth of this invention as a mature industrial solution from the microscopic dimension of signal processing, the macroscopic dimension of control logic, and the dimension of ensuring industrial operation.

[0113] Example 6: This example further explores the flexible production capabilities of the present invention when facing complex formula changes, and how the system achieves automatic parameter calibration through an adaptive algorithm.

[0114] In the production of water-soluble fertilizers, it is often necessary to alternately produce fertilizers with different components in a reactor. These different components exhibit vastly different physicochemical properties. For example, some fertilizers containing a high proportion of humic acid exhibit strong thixotropy in their fluid composition; while some high-phosphorus fertilizers experience intense exothermic chemical reactions during dissolution. To address this need for flexible, multi-variety production, this system incorporates multiple expert weight matrices in its state assessment module.

[0115] When the production plan is sent to the central control module, the system automatically loads the corresponding acoustic feature weighting logic based on the formula identifier. For materials with strong thixotropy, the system increases the monitoring weight of low-frequency components in the interaction signals of the stirring device, as these signals best reflect shear thinning. For materials with intense exothermic reactions, the system prioritizes analyzing the frequency distribution of cavitation acoustic signals to estimate the reaction rate. This intelligent adjustment capability, tailored to each material, is the advantage that allows this system to be widely applied to the production of various complex liquid fertilizers.

[0116] In the initial stages of system deployment, benchmark calibration is typically required. This embodiment describes a self-calibration method based on standard samples. The operator adds a standard material with known physicochemical properties to the reactor and starts the system into calibration mode. The system automatically traverses different stirring speeds and temperature ranges, recording the multi-source acoustic signature features at each state. Through this process, the deep learning model can quickly establish an initial mapping relationship for the specific hardware environment, such as a reactor wall of a specific thickness or a specific type of agitator. This self-calibration logic shortens the on-site debugging cycle of the system.

[0117] To further enhance the system's anti-interference performance, spatial diversity reception technology was introduced into the signal processing module. Due to the multipath effect of sound waves propagating within the metal reactor wall, signals from the same internal sound source will reach the sensor array via different paths. By calculating the time delay and phase shift of each channel signal and utilizing beamforming technology, the system can artificially suppress noise from the external environment, such as the noise from exhaust fans on the factory roof, while enhancing the signal gain from the internal region. This "physical sound insulation" achieved at the electrical signal level ensures that even in noisy heavy industrial workshops, the system can still capture the weak pulses generated by tiny particles impacting the reactor wall.

[0118] This embodiment also defines in detail the thermal compensation parameter adjustment logic in the production process control instructions. The system not only adjusts the power of steam or electric heating but also controls the inflow rate of cooling circulating water based on changes in the sound velocity decay rate. When the dissolution reaction is too vigorous, causing the temperature to rise too rapidly, the system generates a cooling control signal to achieve negative adjustment of the thermal parameters, precisely controlling the temperature rise curve within the reactor. This bidirectional thermal closed-loop control is crucial for ensuring the activity of the heat-sensitive water-soluble fertilizer components.

[0119] In terms of human-computer interaction, this system not only provides real-time chart displays but also integrates a natural language processing-based early warning system. When the system detects a potential malfunction, such as the impeller possibly becoming entangled, the interface no longer simply displays an error code. Instead, it uses semantic descriptions of voiceprint characteristics to inform the operator: "Asymmetrical fluctuations in the stirring cycle signal have been detected, and low-frequency energy has abnormally increased. This indicates that there may be material clumping on the impeller blades. It is recommended to reduce the speed and check and clean them." This user-friendly interactive design reduces the professional knowledge required of the operators.

[0120] The central control module has a reserved standard application programming interface (API), allowing the company's existing production management system to directly read the dissolution status data. By integrating the real-time data of the dissolution endpoint into the company's resource planning system, dynamic optimization of production scheduling can be achieved. For example, when the first reactor reaches the dissolution endpoint ahead of schedule, the system will automatically notify the downstream filling line to prepare for receiving the material.

[0121] In summary, as described in the embodiments above, this invention is not only a collection of technical means, but also a complete, practical industrial solution. By organically combining non-invasive sensing, deep learning state assessment, and rheological physical modeling, it overcomes the monitoring challenges in the production process of water-soluble fertilizers, making a substantial technical contribution to improving the automation level and product quality of the fertilizer industry.

[0122] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for monitoring and controlling the production quality of water-soluble fertilizers, characterized in that, Includes the following steps: S1. By deploying an acoustic acquisition array on the outer wall of the reactor, multi-source acoustic signature signals during the production process of water-soluble fertilizer are acquired in real time. The acoustic acquisition array includes multiple high-sensitivity piezoelectric acoustic sensors distributed at different levels on the outer wall of the reactor. Each sensor is fixed by a magnetic adsorption device and its contact surface is filled with acoustic coupling agent. It captures broadband vibration signals that reflect the evolution of the macroscopic flow state of the fluid, pulse acoustic signals that reflect the degree of solid material residue, and cavitation acoustic signals that reflect the microscopic dynamics of dissolution. S2. Preprocess and reconstruct the multi-source acoustic signature signal to construct an acoustic feature vector space characterizing the dissolution state of the material; The preprocessing process removes background mechanical noise by establishing a background noise benchmark model and calculating the cross-correlation between the measured signal and the background model. It also uses short-time Fourier transform to convert the one-dimensional time series acoustic signal into a two-dimensional time-frequency feature spectrum, and performs mean removal and variance scaling normalization processing on the converted spectrum. S3. Input the acoustic feature vector space into the pre-constructed deep learning state mapping model and output the current material dissolution state level; the deep learning state mapping model extracts local texture features and global contour features in the acoustic feature map through a multi-scale convolutional network structure, and integrates an attention mechanism to learn the core acoustic frequency points corresponding to different dissolution stages. S4. Combine the rheological model to calculate the dynamic parameters of the fluid in the reactor, and determine the dissolution endpoint in conjunction with the dynamic parameters and the material dissolution state level; establish a rheological coupling equation based on stirring power, fluid density and sound velocity attenuation rate, calculate the equivalent viscosity and flow index, and estimate the real-time volume fraction of undissolved solids in the fluid by analyzing the attenuation coefficient of sound wave energy. S5. Based on the determination result of the dissolution endpoint, generate production process control instructions to adjust the stirring speed and thermal compensation parameters of the reactor in real time.

2. The method for monitoring and controlling the production quality of water-soluble fertilizer according to claim 1, characterized in that, The sensor distribution of the acoustic acquisition array is as follows: Multiple sensor groups are set at the bottom, middle liquid level line and top stirring layer of the reactor. The sensor groups are arranged at equal intervals in the circumferential direction. The sound field changes in different spatial regions inside the reactor are obtained through a synchronous triggering mechanism. The broadband vibration signal reflects the change in damping coefficient as the fluid transitions from a rarefied to a viscous state through frequency drift and energy intensity variation. The pulsed acoustic signal determines the particle size distribution and number density of solid particles in the fluid based on the transient energy spikes generated by solid particles impacting the vessel wall. The cavitation acoustic signal is based on the high-frequency signal generated by bubble collapse to monitor the intensity of chemical reactions and local phase transitions during the dissolution process.

3. The method for monitoring and controlling the production quality of water-soluble fertilizer according to claim 2, characterized in that, The feature reconstruction process includes: applying a window function to the digital sequence after background noise removal to reduce spectral leakage caused by truncation effect; By introducing cepstral analysis logic, the Mel frequency cepstral coefficients of the acoustic signal are extracted, separating the mechanical excitation source characteristics and fluid system response characteristics in the acoustic signal, and capturing acoustic wave fluctuations that are sensitive to the dissolution state. The acoustic feature vector space also incorporates an environmental compensation dimension. By acquiring the real-time temperature of the reactor wall, the temperature parameter is input into the model as a continuous variable along with the acoustic features. The compensation layer in the deep learning network is used to correct the influence of environmental temperature on the sound wave propagation speed and attenuation rate.

4. The method for monitoring and controlling the production quality of water-soluble fertilizer according to claim 3, characterized in that, Step S3 specifically includes the following steps: S31. The acoustic feature vector space is decomposed in multiple dimensions to obtain the time-domain energy distribution characteristics that reflect the overall activity of the dissolution process, the frequency-domain power spectral density characteristics used to locate the characteristic response of frequency points, and the time-frequency joint distribution characteristics that reflect the evolution of signal energy over time. S32. Calculate the correlation between each dimension of features and the material's dissolution state using a feature weighting algorithm, and extract a subset of key features. The feature weighting algorithm calculates the correlation coefficient between each acoustic feature component and the target value of the dissolution concentration, marks candidate features with a correlation higher than a preset threshold, and executes principal component analysis logic to compress high-dimensional interrelated features into multiple independent principal components. S33. Based on a deep neural network, a nonlinear mapping is performed on a subset of key features to output the probability distribution of the continuous state of the material from initial solid accumulation to complete dissolution.

5. The method for monitoring and controlling the production quality of water-soluble fertilizer according to claim 4, characterized in that, The deep neural network employs a combination of cross-entropy loss function and mean squared error loss function during the training phase. By dynamically adjusting the weight coefficients, it identifies the stage of the material in the early and middle stages of dissolution and provides quantitative kinetic parameters at the end point. The attention mechanism automatically identifies the core acoustic frequency points of different production stages by learning the correlation between feature channels and triggers early warnings for abnormal crystallization or material agglomeration.

6. The method for monitoring and controlling the production quality of water-soluble fertilizer according to claim 5, characterized in that, The calculation process of the rheological model includes: Obtain the current, voltage, and speed parameters of the stirring motor, and calculate the actual stirring power; A preset excitation source drives a designated acoustic sensor to emit a standard pulse signal. The signal time difference between the transmitter and the receiver of other sensor groups is obtained through a synchronous triggering mechanism. Combined with the pre-calibrated sound wave propagation path length, the propagation speed of the sound wave in the fluid is calculated. By analyzing the scattering characteristics and energy attenuation coefficient of sound waves in heterogeneous fluids, an acoustic feedback distribution function is established. The calculation of the equivalent viscosity takes into account the characteristics of non-Newtonian fluids. By analyzing the phase difference of acoustic waves received by sensors at different vertical heights, the local flow velocity of the fluid at different levels is calculated. The shear rate distribution field inside the fluid is constructed by combining the geometry of the stirring impeller. The rheological coefficient in the rheological model is corrected in real time using the wall shear stress distribution fed back by the acoustic signal.

7. The method for monitoring and controlling the production quality of water-soluble fertilizer according to claim 6, characterized in that, The collaborative determination of the dissolution endpoint includes: Set preset thresholds for the material dissolution state level and stability indices for kinetic parameters; When the current material dissolution state level reaches a preset threshold and the calculated equivalent viscosity fluctuation rate remains below a predetermined change step within a preset time period, it is determined to be the dissolution endpoint. The judgment process also incorporates a confidence assessment based on logistic regression. By combining the classification probability distribution of the deep learning model, the stationarity index of the dynamic parameters, and the time statistical characteristics of similar batches in history, a percentage value reflecting the accuracy of the judgment is calculated. When this percentage value is lower than a preset threshold, the observation time is extended.

8. The method for monitoring and controlling the production quality of water-soluble fertilizer according to claim 7, characterized in that, The generated production process control instructions include: In the initial stage of dissolution, the frequency of the inverter of the stirring motor is increased according to the energy intensity of the pulse acoustic signal, thereby increasing the stirring speed to enhance solid-liquid convection mass transfer. During the middle stage of dissolution, the thermal compensation parameters are adjusted according to the change in the sound velocity decay rate. The temperature inside the vessel is maintained within the preset dissolution temperature range by controlling the opening of the steam valve or adjusting the electric heating power. At the end of the dissolution process, reduce the output frequency of the stirring motor to reduce energy consumption and eliminate bubble interference; The production process control instructions are generated using a predictive control algorithm. Based on the rate of change of the current solubility level, the material state trend in the future time period is predicted, and the thermal compensation instructions are fine-tuned in advance to suppress local overheating and agglomeration.

9. A production quality monitoring and control system for water-soluble fertilizers, comprising using the production quality monitoring and control method for water-soluble fertilizers as described in any one of claims 1 to 8, characterized in that, include: The acoustic acquisition module is equipped with a non-invasive acoustic sensor group installed on the outer wall of the reactor to collect full-frequency acoustic signals during the production process; The non-invasive acoustic sensor array is fixed by magnetic adsorption or coupling agent, enabling the transmission of acoustic energy through the vessel wall; The signal processing module, connected to the acoustic acquisition module, is used to realize signal noise reduction, transformation and feature extraction; the signal processing module is equipped with a high-speed digital signal processor, which performs synchronous analog-to-digital conversion on multiple acoustic signals at a predetermined sampling frequency, and uses a ping-pong buffer mechanism to realize the temporal overlap of data acquisition and mathematical processing; The status assessment module integrates a deep learning engine to assess the material dissolution progress and fluid dynamics state based on feature signals; the status assessment module is also equipped with an anomaly diagnosis submodule, which identifies equipment fault status by comparing with a standard voiceprint template. The central control module is connected to the status assessment module and the actuator of the production equipment, and is used to output frequency conversion control signals and thermal regulation signals according to the assessment results. The data storage module is used to record voiceprint data, characteristic parameters, and final quality judgment results throughout the entire production process, and to build a quality traceability database.

10. A production quality monitoring and control system for water-soluble fertilizer according to claim 9, characterized in that, The system also includes a human-computer interaction module, which is used to display the dissolution progress curve, acoustic signature spectrum, and dynamic parameter change trends in real time. The abnormality diagnosis submodule periodically executes an acoustic transparency self-test procedure. During the reactor evacuation phase, it emits a standard pulse through an excitation source and analyzes the echo characteristics to assess the degree of fouling on the reactor wall and automatically updates the input correction coefficients of the deep learning model. The central control module has a reserved application programming interface for transmitting real-time data of the dissolving endpoint to the production management system to achieve dynamic optimization of production scheduling. The acoustic acquisition module utilizes beamforming technology to introduce time delays into the signals of each channel of the sensor array, thereby enabling the focusing of the internal region of the reactor at the algorithm level, locating and eliminating localized areas of lag in dissolution.

Images