Intelligent spraying and crust control system based on advanced control

Abstract

This invention belongs to the field of advanced control technology, specifically relating to an intelligent spray-based crispy skin control system. The system includes: a multimodal sensor array for collecting sound waves, spectral data, and radar signals reflecting the evolution of the crispy skin; a semantic feature mapping device for generating a semantic query vector representing the current state; a brain-like integrated knowledge graph database for storing historical production trajectories; an approximate nearest neighbor strategy retrieval device for retrieving successful strategies through similarity matching; a generative counterfactual reasoning decision-making device for performing causal deduction and selecting the optimal strategy to generate control commands; and a spraying actuator for adjusting parameters to achieve precise intervention. This invention, by constructing a brain-like architecture, achieves deep perception and self-evolutionary decision-making regarding the crispy skin state, improving product quality consistency.

Description

Technical Field

[0001] This invention belongs to the field of advanced control technology, specifically relating to an intelligent spray crisping control system based on advanced control. Background Technology

[0002] With the continuous improvement of automation in the food industry, intelligent processing equipment plays a core role in enhancing food production efficiency and product consistency. In the field of intensive meat processing, the surface curing and flavor development of crispy products are key factors determining product quality. To achieve high-quality output on industrial production lines, processing equipment needs to precisely adjust heating, spraying, and environmental parameters to induce Maillard reactions on the surface of ingredients and dynamically control the rate of oil exudation. This makes intelligent control systems with precise intervention capabilities the technological backbone of modern food processing equipment.

[0003] Intelligent spraying technology based on advanced control focuses on the dynamic evolution of the physical properties of food surfaces. It aims to precisely construct a crispy outer layer by real-time adjustment of the spray medium flow rate, atomization particle size, and spray trajectory. This technological direction requires the system to deeply analyze the microscopic physicochemical state of food under continuous production conditions and map complex nonlinear physical characteristics into efficient control decisions in real time. In dynamically changing production scenarios, establishing a precise correlation between the real-time state of the food surface and the spraying strategy, and overcoming control fluctuations caused by lag, is key to achieving ultimate stability in the sensory quality of food.

[0004] Existing technologies mostly employ static control models based on preset curves, making it difficult to perceive the dynamic evolution of surface water activity gradients and microscopic bubbling characteristics of food in real time. This results in a severe disconnect between the spraying strategy and the physicochemical reaction process. Furthermore, traditional control architectures rely excessively on visual sensors for geometric positioning. However, when faced with unstructured disturbances such as food deformation due to heat or oil splattering, feature extraction is often hindered, leading to localized overspray causing soft collapse or underspray causing scorching. This creates a technical bottleneck where control accuracy and system robustness are difficult to balance. Existing systems lack in-depth mining of historical batch production trajectories and the ability to transfer knowledge across scenarios, causing production parameter optimization to heavily rely on human experience and making it impossible to achieve adaptive strategy evolution for complex dynamic conditions. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent spray crisping control system based on advanced control, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The intelligent spray-based crispy skin control system based on advanced control includes a multimodal sensor array, a semantic feature mapping device, a brain-like in-memory computing knowledge graph database, an approximate nearest neighbor strategy retrieval device, a generative counterfactual reasoning decision-making device, and a spray execution mechanism, wherein: The multimodal sensor array is used to collect multidimensional physicochemical state signals of food during the heating and cooking process in real time. The signals include high-frequency acoustic wave oscillation signals, near-infrared spectral reflection signals and millimeter-wave radar echo signals on the surface of the food, in order to obtain raw data streams reflecting the characteristics of skin blistering and rupture, oil seepage rate and geometric deformation state of the food. The semantic feature mapping device is connected to the multimodal sensor array and is used to receive the raw data stream and, through a preset self-supervised comparative learning logic, map the unstructured raw data stream to a high-dimensional semantic feature vector space to generate a semantic query vector that represents the current crispy evolution state of the food ingredient. The brain-like in-memory computing knowledge graph database is used to store multiple batches of trajectory data structures in the historical production process. Each trajectory data structure contains a semantic feature vector at a specific production moment, the executed spraying action instruction, and the corresponding crispy skin quality output effect. The database achieves deep coupling between data storage and logical operation through an in-memory computing architecture. The approximate nearest neighbor strategy retrieval device is connected to the semantic feature mapping device and the brain-like storage and computing integrated knowledge graph database, respectively. It is used to perform high-dimensional similarity matching in the brain-like storage and computing integrated knowledge graph database using the semantic query vector as the search term, and retrieve the set of historical successful strategy trajectories that are closest to the current food state in terms of physical evolution trend. The generative counterfactual reasoning decision-making device is connected to the approximate nearest neighbor strategy retrieval device. It is used to perform causal inference on the retrieved set of historical successful strategy trajectories, calculate the expected benefits and select the optimal control strategy by simulating the evolution path of fragile skin characteristics under different spraying intervention schemes in the semantic space, and encapsulate it into advanced control instructions. The spraying actuator is connected to the generative counterfactual reasoning decision-making device to receive the advanced control commands and adjust the flow rate, pressure, atomization particle size, and nozzle movement trajectory of the spraying medium accordingly, so as to achieve precise intervention on the surface of the food.

[0007] Preferably, the multimodal sensor array includes a high-frequency acoustic acquisition unit, a near-infrared spectral sensing unit, and a millimeter-wave detection unit. The high-frequency acoustic acquisition unit is installed inside the heating environment to capture minute acoustic pulses generated on the food's surface during Maillard reactions and rapid moisture loss; these pulse signals reflect the breaking frequency of the crispy skin's microstructure. The near-infrared spectral sensing unit scans the food surface and, by analyzing the absorption intensity of specific wavelengths, monitors the dynamic amount of oil seepage and the distribution of water activity gradients in real time. The millimeter-wave detection unit penetrates obstructions in high-temperature, oily fume environments to accurately acquire unstructured deformation parameters of the food caused by heat, such as shrinkage, expansion, or twisting.

[0008] Furthermore, the semantic feature mapping device incorporates feature decoupling logic. This logic extracts essential representative feature components from multi-source heterogeneous sensor signals, eliminating environmental noise and random disturbances to ensure that the generated semantic query vector accurately corresponds to the specific physical stage of crisping in food ingredients. The semantic feature mapping device also features dynamic weight allocation logic, which automatically adjusts the contributions of sound waves, spectral data, and radar signals in the vector synthesis process based on the current clarity index of the production environment, ensuring stable semantic representations are output even under visually impaired conditions.

[0009] Furthermore, the brain-inspired in-memory computing knowledge graph database uses a graph network structure to store data. Each trajectory data structure not only contains quantified physical parameters but also contextual semantic tags describing the production scenario. The brain-inspired in-memory computing knowledge graph database has a dynamic evolution function, which is used to encapsulate the current production trajectory, environmental disturbance factors, and final product qualification rate after each production task is completed, and then integrate them into the existing knowledge network after correlation analysis, so as to realize the continuous accumulation and logical reinforcement of production experience.

[0010] Furthermore, the approximate nearest neighbor strategy retrieval device employs a hierarchical navigation small-world algorithm. This algorithm is used to locate the feature cluster closest to the semantic query vector from massive historical trajectory data within a short, predetermined time period. During the retrieval process, the device considers not only the static similarity at the current moment but also, more importantly, the dynamic trend similarity of the food's state evolving over time, ensuring that the retrieved strategy has predictive capabilities.

[0011] Furthermore, the generative counterfactual reasoning decision-making device includes a causal model construction unit and a virtual state inference unit. The causal model construction unit utilizes historical correlation data from the brain-like in-memory computing knowledge graph database to construct a nonlinear causal chain between the spraying action and the crispness quality index. The virtual state inference unit is used to execute simulation experiments; that is, for the current moment, it assumes different gradients of spraying pressure or different frequencies of scanning trajectories to predict the semantic feature vector change trend of the ingredients over a predetermined time period. The generative counterfactual reasoning decision-making device compares the crispness index and color uniformity index under multiple hypothetical paths, selects the optimal path that can achieve the preset quality threshold, and converts it into the advanced control command.

[0012] Furthermore, the spraying actuator includes a multi-degree-of-freedom robotic arm, a high-precision flow proportional valve, and an ultrasonic atomizing nozzle. The multi-degree-of-freedom robotic arm is used to execute complex spatial coverage trajectories according to the advanced control commands. The high-precision flow proportional valve is used to respond to flow adjustment requirements within milliseconds. The ultrasonic atomizing nozzle is used to change the atomized particle size of the spraying medium by adjusting the driving frequency to adapt to the penetration requirements of different food surface porosities.

[0013] Furthermore, the system also includes an environmental perception and correction unit. This unit monitors fluctuations in ambient temperature, humidity, and wind speed within the heating chamber and feeds these environmental disturbances back to the generative counterfactual reasoning decision-making device in real time as correction factors. When the ambient temperature deviates from a preset range, the device automatically adjusts the boundary conditions of the counterfactual reasoning to compensate for the impact of environmental fluctuations on the rate of brittle skin formation.

[0014] Furthermore, the system possesses anomaly detection and robust switching capabilities. When the multimodal sensor array detects a sudden change or absence of a signal in a certain dimension outside a predetermined range, the near nearest neighbor strategy retrieval device automatically switches to robust retrieval mode. In robust retrieval mode, the system retrieves non-precise coverage spraying strategies that have demonstrated robust performance in high-interference and low-visibility scenarios from historical records. By expanding the spray coverage area and increasing the spraying frequency, it replaces point-to-point high-precision operations to ensure that a uniform crisping layer can still form on the surface of the food even under extreme conditions.

[0015] Furthermore, the brain-like in-memory computing knowledge graph database is also equipped with cross-batch transfer learning logic. This logic allows the system to extract similar physicochemical change patterns between different types of ingredients when faced with entirely new types of ingredients, transferring the mature spray control experience of known ingredients to the initial production process of the new ingredients. The system automatically matches the semantic feature evolution path of the new ingredients in the early stage of heating and searches for historical trajectories with similar evolutionary characteristics from the database, reducing the parameter debugging time during the production of new products and enabling rapid iteration of production strategies.

[0016] Furthermore, the spraying actuator is also equipped with a media recovery and circulation filtration unit. This unit is used to collect excess media that does not adhere to the surface of the food during the spraying process, and after multi-stage filtration and component balance adjustment, it is reintroduced into the spraying pipeline. When generating instructions, the generative counterfactual reasoning decision-making device simultaneously considers the media utilization rate index, and optimizes the start-stop sequence and spray angle of the nozzles to minimize the loss of spraying media while ensuring the crispy skin quality.

[0017] Furthermore, the system is also equipped with a quality evaluation feedback interface. This interface is used to receive sensory evaluation data of the product entered by backend testing equipment or manually. The brain-like in-memory computing integrated knowledge graph database adjusts the weights of the corresponding trajectory data structure based on the feedback data. For trajectories evaluated as having excellent quality, their recommendation weight during the retrieval process is increased; for trajectories with defects, negative sample constraints are established, enabling the generative counterfactual reasoning decision-making device to automatically avoid similar paths in subsequent deductions, achieving closed-loop self-evolution of the control logic.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. Achieved deep situational awareness across modalities. This invention captures heterogeneous signals such as acoustic, spectral, and radar signals through a multimodal sensor array, and transforms them into vectors in a high-dimensional semantic space using a semantic feature mapping device. This solves the problem of perception failure in traditional visual solutions under conditions of smoke and oil smudges, and can accurately capture the evolution of the microscopic physicochemical state of the crisping process on the surface of food, achieving a deep analysis of the strong coupling relationship between the crisping process and the spraying strategy.

[0019] 2. An innovative brain-inspired architecture of retrieval as control is introduced. By combining a brain-inspired in-memory computing knowledge graph database with an approximate nearest neighbor strategy retrieval device, the system transforms complex nonlinear control problems into efficient data retrieval and matching problems. This architecture eliminates excessive reliance on precise mathematical models, enabling the system to quickly find the optimal response trajectory from historical experience when facing unstructured chemical conditions such as significant deformation of ingredients, thus improving the system's robustness and dynamic response speed.

[0020] 3. Possesses powerful causal deduction and self-evolving decision-making capabilities. Utilizing a generative counterfactual reasoning decision-making device, the system can perform virtual deductions before executing instructions, predicting the crispness evolution under different intervention strategies. This enables precise pre-adjustment of spray parameters, avoiding soft collapse caused by localized overspray or scorching caused by missed spray. Simultaneously, with the continuous accumulation of production data, the system can achieve continuous knowledge accumulation and cross-batch transfer, improving the yield rate of the first trial production and breaking the long-term dependence of food processing on human experience.

[0021] 4. A balance is achieved between energy conservation and stable quality. This invention can accurately predict the critical point of crispy crust formation. By dynamically adjusting the flow rate, pressure, and atomization particle size of the spraying medium, it achieves significant energy conservation of the spraying medium while ensuring a high level of uniformity in the density of micro-bubbles on the crispy crust surface. The final product possesses excellent qualities of a uniformly crispy crust and tender, juicy interior, improving the consistency of sensory quality in industrialized food production. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the intelligent decision-making mechanism based on generative counterfactual reasoning in this invention; Figure 3 This is a logical flowchart of the multimodal signal perception and high-dimensional semantic feature mapping in this invention; Figure 4 This is a flowchart illustrating the strategy retrieval and knowledge evolution logic of the brain-like storage-computing integrated knowledge graph in this invention. Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between the perception layer, decision layer and execution layer in this invention; Figure 6 This is a schematic diagram comparing the core principle of this invention with existing technologies in terms of the precision of controlling the brittle skin evolution state. Detailed Implementation

[0023] Example 1: Please refer to the appendix Figure 1 To be continued Figure 6 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0024] The intelligent spray crispy skin control system based on advanced control includes a multimodal sensor array, a semantic feature mapping device, a brain-like in-memory computing knowledge graph database, an approximate nearest neighbor strategy retrieval device, a generative counterfactual reasoning decision-making device, and a spraying actuator. The multimodal sensor array is used to collect multidimensional physicochemical state signals of food during the heating and cooking process in real time. The signals include high-frequency acoustic wave oscillation signals, near-infrared spectral reflection signals, and millimeter-wave radar echo signals from the surface of the food, in order to obtain raw data streams reflecting the characteristics of skin blistering and rupture, oil exudation rate, and the geometric deformation state of the food.

[0025] The multimodal sensor array is not a simple sensory simulation, but a deep physical integration designed for extreme conditions in food processing environments. The high-frequency acoustic oscillation signal is collected at a frequency between 20 kHz and 100 kHz to capture ultrasonic pulses generated during the Maillard reaction on the food surface due to rapid localized water vaporization and protein microstructural breakage. The near-infrared spectral reflectance signal covers the characteristic band from 900 nm to 1700 nm; by analyzing the absorption rate of different wavelengths of light on the food surface, the surface water activity and the density of oil molecules are accurately calculated. The millimeter-wave radar echo signal uses 77 GHz frequency-modulated continuous wave technology, capable of penetrating dense oil fumes and water vapor; by analyzing the phase changes and delays of the echo, the three-dimensional deformation contour of the food is constructed in real time.

[0026] The semantic feature mapping device is connected to the multimodal sensor array and is used to receive the raw data stream and, through a preset self-supervised comparative learning logic, map the unstructured raw data stream to a high-dimensional semantic feature vector space to generate a semantic query vector representing the current crispy evolution state of the food ingredient.

[0027] The semantic feature mapping device integrates a high-performance processing chip with a deep residual network architecture specifically designed for feature alignment of multi-source heterogeneous signals. During processing, the device first performs a short-time Fourier transform on the acoustic signal, converting it into spectrogram features; simultaneously, it performs near-infrared spectral reduction to extract feature components reflecting changes in chemical composition; and it converts the millimeter-wave radar point cloud data into geometric features characterizing deformation gradients.

[0028] Subsequently, through self-supervised contrastive learning logic, the system searches for intrinsic correlations between different modal signals in the absence of labels. For example, the enhancement of sound wave pulses is often accompanied by a decrease in the moisture absorption peak in the spectrum. This correlation is solidified into coordinate positions in the semantic feature vector space. The final output semantic query vector highly condenses the current physical cooking stage of the food, rather than simply piling up raw numerical values.

[0029] Specifically, the self-supervised contrastive learning logic uses the unsupervised InfoNCE contrastive loss function for initializing and training the network parameters. During the same batch production process, for the same ingredient at time steps... The acquired multimodal signals are then augmented (e.g., by adding Gaussian white noise or randomly masking frequency bands) to construct positive sample pairs. Construct a negative sample set from the feature vectors of other different ingredients or different time steps within the same batch. The contrastive loss function The formula is as follows:

[0030] in, Indicates the comparison loss value, This represents the function for calculating cosine similarity. This represents the learnable temperature hyperparameter, initialized to 0.07, used to adjust the model's ability to distinguish between difficult negative samples. For indicator functions, when The value is 1 if it is true, and 0 otherwise. This represents the total number of samples in the current batch. By minimizing this loss function, the network is constrained to bring features closer together at the same physical evolution stage and push features further apart at different states in a high-dimensional space.

[0031] The brain-like in-memory computing knowledge graph database is used to store multiple batches of trajectory data structures in the historical production process. Each trajectory data structure contains a semantic feature vector at a specific production moment, the executed spraying action instruction, and the corresponding crispy skin quality output effect. The database achieves deep coupling between data storage and logical operation through an in-memory computing architecture.

[0032] At the physical implementation level, the database adopts a neuromorphic computing architecture based on non-volatile storage media, which enables data reading and preliminary similarity screening to be completed within the same physical unit, completely eliminating the communication bandwidth bottleneck between memory and processor in traditional computer architecture.

[0033] In one preferred embodiment, the neuromorphic in-memory computing knowledge graph database is physically constructed on a cross-array based on resistive random access memory (RRAM). Let the semantic feature vector dimension stored in the historical trajectory database be... The resistive switching memory array includes One storage unit, of which This represents the total number of historical trajectory nodes. The conductivity value of each storage unit. Programmed into the corresponding historical semantic feature vector The numerical value of the dimension. When similarity matching needs to be performed, the semantic feature mapping device will use the semantic query vector generated at the current moment. Transform into input voltage vector (Right now ,in (This is the voltage mapping scaling factor). According to Kirchhoff's current law and Ohm's law, the array's... Converging current of column output The calculation formula is as follows:

[0034] in, Indicates the first Output current values ​​of each historical trajectory node The first element representing the input voltage vector is... One portion, Indicates the first Line number The conductance value of the column storage cell. Due to this current value... In physical essence, it is equivalent to the query vector. With the The system reads the current values ​​of each column through an analog-to-digital converter, using the result of the dot product operation between the historical vectors. The column with the largest current value is selected first. Each node is identified as the set of historical successful strategy trajectories output by the approximate nearest neighbor strategy retrieval device.

[0035] The trajectory data structure is organized in the form of a graph network, where nodes represent the semantic state at a certain moment, and edges represent the transition probabilities between states and the control interventions applied during the transition process. Each successful production trajectory is considered an optimal path in the knowledge graph. In addition, the database also stores environmental interference factors, such as temperature fluctuations and humidity saturation within the heating chamber, which are stored as context labels associated with semantic feature vectors.

[0036] The approximate nearest neighbor strategy retrieval device is connected to both the semantic feature mapping device and the brain-like in-memory computing knowledge graph database. It uses the semantic query vector as the search term to perform high-dimensional similarity matching in the brain-like in-memory computing knowledge graph database, retrieving a set of historical successful strategy trajectories that are closest to the current food state in terms of physical evolution trend. The approximate nearest neighbor strategy retrieval device abandons the traditional global traversal algorithm and instead adopts a hierarchical navigation small-world algorithm. This algorithm, by constructing a multi-layered index structure, enables the retrieval process to be completed in logarithmic time complexity. During the retrieval process, the device calculates the Euclidean distance or cosine similarity between the current semantic query vector and the feature vectors stored in the database.

[0037] To improve the robustness of retrieval, the device not only focuses on the distance at the current single moment, but also introduces a sliding window mechanism to consider the semantic evolution trend within 5 to 10 seconds before the current moment, and to find successful cases that are not only similar in current state, but also highly consistent in historical evolution path.

[0038] The generative counterfactual reasoning decision-making device is connected to the approximate nearest neighbor strategy retrieval device. It is used to perform causal inference on the retrieved set of historical successful strategy trajectories, calculate the expected benefits and select the optimal control strategy by simulating the evolution path of fragile skin characteristics under different spraying intervention schemes in the semantic space, and encapsulate it into advanced control instructions.

[0039] The generative counterfactual reasoning decision-making device is the logical core of the system. Based on a causal model building unit, it uses historical data to establish the intervention relationship between the spraying action and the crispy skin quality effect. When multiple candidate strategies are retrieved, the device performs virtual state inference. It assumes: If the spraying pressure is increased by 10% at this time, how will the crispy skin texture characteristics of the food change in the next 30 seconds? If the current flow rate remains unchanged, will the surface oil seepage rate exceed the preset critical value? By comparing the simulation results under various hypothetical paths, the system selects the strategy path that minimizes the standard deviation of crispy skin crispness and has the highest color uniformity score. This ability to extrapolate the future gives the system strong predictive power and can offset the large physical lag effect in the process of the spraying medium leaving the nozzle and reaching the surface of the food.

[0040] The spraying actuator, connected to the generative counterfactual reasoning decision-making device, receives advanced control commands and adjusts the flow rate, pressure, atomization particle size, and nozzle trajectory of the spraying medium accordingly to achieve precise intervention on the food surface. The spraying actuator includes a high-precision servo drive system. The flow proportional valve has millisecond-level adjustment accuracy and can linearly switch flow rates according to commands. The ultrasonic atomizing nozzle breaks the spraying medium into adjustable droplets with a particle size between 10 and 50 micrometers by adjusting the transducer's drive frequency, ensuring the medium can uniformly penetrate into the micrometer-level pores of the food surface. A multi-degree-of-freedom robotic arm carries the nozzle through complex spatial trajectory movements within the heating chamber. Its trajectory is dynamically corrected based on the real-time deformation of the food to ensure uniform spray coverage.

[0041] The multimodal sensor array includes a high-frequency acoustic acquisition unit, a near-infrared spectral sensing unit, and a millimeter-wave detection unit. The high-frequency acoustic acquisition unit is installed inside the heating environment to capture tiny acoustic pulses generated on the surface of the food during the Maillard reaction and rapid moisture loss. These pulse signals reflect the breaking frequency of the crispy skin's microstructure.

[0042] To cope with high-temperature and high-fume environments, the high-frequency acoustic acquisition unit employs a specially designed acoustic waveguide structure. Sensitive sensor probes are deployed outside the high-temperature zone, and the waveguide transmits internal acoustic vibrations to the sensor without loss. The near-infrared spectral sensing unit scans the food surface, analyzing the absorption intensity of specific wavelengths to monitor the dynamic exudation of oil and the distribution of water activity gradients in real time. The near-infrared spectral sensing unit integrates a spectrometer and an array detector, capable of outputting the compositional distribution map of the food surface at a rate of over 50 frames per second. The millimeter-wave detection unit penetrates obstructions in high-temperature and fume environments to accurately acquire unstructured deformation parameters of food caused by heat, such as contraction, expansion, or twisting. The short wavelength characteristics of millimeter waves enable centimeter-level spatial resolution, capturing minute volumetric changes in food.

[0043] The semantic feature mapping device incorporates feature decoupling logic. This logic extracts essential representative feature components from multi-source heterogeneous sensor signals, eliminates environmental noise and random disturbances, and ensures that the generated semantic query vector accurately corresponds to the specific physical stage of crisping in the food ingredient.

[0044] The feature decoupling logic uses a combination of principal component analysis and independent component analysis to separate temperature background noise, mechanical vibration noise, and actual food reaction signals from the signal. The semantic feature mapping device is also equipped with dynamic weight allocation logic, which automatically adjusts the contributions of sound waves, spectral signals, and radar signals in the vector synthesis process based on the current clarity index of the production environment. For example, when the concentration of cooking fumes in the environment exceeds a preset threshold, causing severe attenuation of the near-infrared spectrum, the dynamic weight allocation logic automatically reduces the weight of the spectral signal and simultaneously increases the weight of the millimeter-wave radar signal to ensure stable semantic representation output even under visually impaired conditions. This dynamic adjustment mechanism gives the system extremely high environmental adaptability in extreme industrial environments.

[0045] In one specific embodiment, the dynamic weight allocation logic is implemented using a softmax adaptive weighting algorithm based on the environmental clarity index. The semantic feature mapping device receives initial feature vectors extracted by the high-frequency acoustic acquisition unit, the near-infrared spectral sensing unit, and the millimeter-wave detection unit, denoted as follows: , , The system calculates the environmental clarity index of each modal signal in real time. , , (The value ranges from [0,1], with larger values ​​indicating less interference). Specifically, for near-infrared spectral sensing units, their sharpness index... This is obtained by calculating the ratio of the current signal-to-noise ratio to the reference signal-to-noise ratio. When the concentration of oil fumes in the cavity causes the signal-to-noise ratio to fall below a preset threshold, the signal-to-noise ratio will be lowered. A significant decrease. Then the fused feature vector... The calculation formula is as follows:

[0046] in, , , These represent the linear mapping layers corresponding to the respective modes. , , This represents the dynamically allocated weighting coefficient. The weighting coefficient is calculated using the following formula:

[0047] in, , This represents the temperature regulation hyperparameter, used to control the smoothness of weight allocation. In this embodiment... Set to 2.0. Using the above formula, when the near-infrared spectral sharpness index... When the weight is reduced due to the obstruction of cooking fumes, its corresponding weight It automatically decreases under the normalization effect of the denominator, and at the same time The corresponding increase ensures the output feature vector Robustness under harsh working conditions.

[0048] The brain-like in-memory computing knowledge graph database uses a graph network structure to store data. Each trajectory data structure not only contains quantified physical parameters but also contextual semantic tags describing the production scenario. These tags include information such as the type of food, initial weight, and cold chain temperature history, providing richer limiting dimensions for subsequent retrieval.

[0049] The database features dynamic evolution capabilities. After each production task is completed, it encapsulates the current production trajectory, environmental disturbance factors, and final product qualification rate, and integrates them into the existing knowledge network after correlation analysis. The specific evolution process includes knowledge compression and importance resampling: for trajectories that produce high-quality products with low energy consumption, the system increases their connection weights; while for paths that produce defective products or are inefficient, the system establishes negative constraint weights. In this way, production experience is continuously accumulated and logically strengthened, making the system increasingly intelligent over time, gradually forming a control knowledge graph with industry-monopolistic characteristics.

[0050] The approximate nearest neighbor strategy retrieval device employs a hierarchical navigation small-world algorithm. This algorithm is used to locate the feature cluster closest to the semantic query vector from massive historical trajectory data within a very short, predetermined time. The retrieval time window is typically strictly limited to meet real-time control requirements. During the retrieval process, the device considers not only the static similarity at the current moment but also, more importantly, the dynamic trend similarity of the food's state evolution over time. By calculating the first and second derivatives of the semantic vector, the device evaluates the acceleration and direction of state changes, ensuring that the retrieved strategy has predictive capabilities and can prevent impending physical state anomalies.

[0051] The generative counterfactual reasoning decision-making device comprises a causal model construction unit and a virtual state deduction unit. The causal model construction unit utilizes historical correlation data from the brain-like in-memory computing knowledge graph database to construct a nonlinear causal chain between the spraying action and the crispness quality indicators. This causal model differs from traditional correlation models; it can distinguish between physical changes caused by spraying and spontaneous changes caused by environmental heating. The virtual state deduction unit is used to execute simulation experiments, that is, for the current moment, assuming different gradients of spraying pressure or different frequencies of scanning trajectories, to predict the trend of semantic feature vector changes in the ingredients over a predetermined time period. The generative counterfactual reasoning decision-making device compares the crispness and color uniformity indicators under multiple hypothetical paths, selects the optimal path that achieves the preset quality threshold, and converts it into the advanced control command. The decision-making process also incorporates the Nash equilibrium concept from game theory, ensuring optimal taste while also saving spraying medium and reducing energy consumption.

[0052] Specifically, the virtual state deduction unit's built-in data-driven dynamics model based on a recurrent neural network adopts a gated recurrent unit (GRU) architecture. At time steps... The network receives the hidden state from the previous moment. and the current semantic feature vector and candidate control commands The hidden state update equation for the concatenated input vector is as follows:

[0053] in, Indicates time step The hidden state of the network, This indicates a feature concatenation operation. This represents a multi-dimensional control command vector containing spray flow rate, pressure, and atomized particle size. Through this hidden state, the network outputs a forecast of future timeframes. Crispy skin feature prediction vector The formula is as follows:

[0054] in, This represents the output layer weight matrix. This represents the output layer bias vector.

[0055] After obtaining the future evolution paths of multiple candidate control sequences, the generative counterfactual reasoning decision-making device selects the optimal strategy by solving the following multi-objective optimization function:

[0056] in, This represents the optimization objective value corresponding to the optimal control sequence. Indicates length is The control sequence, This represents the standard deviation of local crispness extracted from the predicted feature vector. This represents the score for color uniformity as determined by analysis. Indicates execution instruction Medium loss and energy consumption penalty terms during operation. , , These are the non-negative weight coefficients determined under the logic of Nash equilibrium game theory. The system solves the above formula using either gradient descent or particle swarm optimization to find the coefficients that make the weights equal to ... Minimize control sequence The first frame instruction is encapsulated into the advanced control instruction and sent to the spray actuator, and subsequent frames are used as the comparison benchmark for the shadow control module.

[0057] The spraying actuator includes a multi-degree-of-freedom robotic arm, a high-precision flow proportional valve, and an ultrasonic atomizing nozzle. The multi-degree-of-freedom robotic arm executes complex spatial coverage trajectories according to the advanced control commands. A collision detection sensor and position compensation unit are installed at the end of the robotic arm, enabling millimeter-level positioning accuracy within the complex heating chamber space. The high-precision flow proportional valve responds to flow adjustment needs within milliseconds and employs a ceramic valve core structure, exhibiting extremely high wear resistance and corrosion resistance. The ultrasonic atomizing nozzle adjusts the driving frequency to change the atomized particle size of the spray medium, adapting to the penetration requirements of different food surface porosities. For example, for foods with denser surfaces, the system automatically instructs the nozzle to generate finer atomized particles to enhance the crispness through microscopic penetration.

[0058] The system also includes an environmental perception and correction unit. This unit monitors fluctuations in ambient temperature, humidity, and wind speed within the heating chamber and feeds these environmental disturbances back to the generative counterfactual reasoning decision-making device in real time as correction factors. The environmental perception and correction unit comprises a thermocouple array and a humidity sensor array deployed around the chamber. When the ambient temperature deviates from a preset range, the device automatically adjusts the boundary conditions of the counterfactual reasoning. For example, if an unexpected increase in ambient temperature leads to excessively rapid moisture evaporation, the system automatically shortens the decision loop cycle and appropriately increases the spray frequency to compensate for the impact of environmental fluctuations on the rate of brittle crust formation, ensuring consistent quality across batches of products.

[0059] The system features anomaly detection and robust switching capabilities. When the multimodal sensor array detects a sudden change or loss of signal in a certain dimension outside a predetermined range, the near nearest neighbor strategy retrieval device automatically switches to robust retrieval mode. In robust retrieval mode, the system retrieves non-precise coverage spraying strategies that have proven robust in high-interference, low-visibility scenarios from historical records. This strategy typically does not pursue extreme local precision but instead replaces point-to-point high-precision operations by expanding the spray coverage area and increasing the spraying frequency. This ensures that even in extreme conditions such as accidental sensor damage or sudden large-area smoke obstruction, a uniform brittle layer can still form on the surface of the food, preventing the entire batch of material from being scrapped.

[0060] The brain-like in-memory computing knowledge graph database is also equipped with cross-batch transfer learning logic. This logic allows the system to transfer mature spray control experience from known ingredients to the initial production process of new ingredients by extracting similar physicochemical change patterns between different types of ingredients. For example, when processing a new type of poultry, the system automatically matches the semantic feature evolution path of the ingredient in the early stage of heating and searches the database for historical trajectories with similar evolutionary characteristics in terms of protein content, skin thickness, etc. Through this cross-domain feature alignment, the system can quickly generate an initially usable spray strategy, reducing the parameter debugging time during new product production and enabling rapid iteration of production strategies and maximizing the utilization of knowledge assets.

[0061] The spraying actuator is also equipped with a media recovery and circulation filtration unit. This unit collects excess media that does not adhere to the food surface during the spraying process, and after multi-stage filtration and component balance adjustment, it is reintroduced into the spraying pipeline. The recovery unit includes a collection ramp located at the bottom of the heating chamber, a stainless steel primary filter, a high-precision activated carbon adsorption tower, and a component analyzer. The generative counterfactual reasoning decision-making device considers the media utilization rate when generating instructions. By optimizing the start-stop sequence and spray angle of the nozzles, the loss of spraying media is reduced while ensuring the crispy texture. This closed-loop material utilization model reduces production costs and aligns with the industry trend of green manufacturing.

[0062] The system is also equipped with a quality evaluation feedback interface. This interface is used to receive sensory evaluation data of the product entered by backend testing equipment or manually. The evaluation data includes quantitative indicators such as taste score, crust thickness, oil content, and color saturation. The brain-like storage-computing integrated knowledge graph database adjusts the weights of the corresponding trajectory data structure based on the feedback data. For trajectories evaluated as excellent quality, the system not only increases their recommendation weight in the retrieval process, but also uses an adaptive algorithm to mine the key feature combinations behind the trajectory and solidify them into the gold medal strategy in the database. For trajectories with defects, negative sample constraints are established, so that the generative counterfactual reasoning decision-making device automatically avoids similar paths in subsequent deductions, realizing the closed-loop self-evolution of control logic, completely breaking the dependence of traditional production on high-level human experience, and transforming craftsman experience into replicable and evolvable digital assets.

[0063] The internal logic of the semantic feature mapping device is further described in detail. When performing self-supervised contrastive learning, the device employs a pre-defined contrastive loss function. This loss function aims to reduce the Euclidean distance between semantic vectors of the same ingredient observed in different time windows and different modalities, while simultaneously increasing the Euclidean distance between semantic vectors of different types of ingredients or under abnormal conditions.

[0064] The system defines the sound, light, and electrical signals at the current moment as positive sample pairs, and signals from other random historical moments as negative samples. Through thousands of gradient iterations, the network learns a feature extractor that reflects the essence of crispy crust cooking. This feature extractor can not only identify the macroscopic phenomenon of bubbling, but also identify deeper features reflecting physicochemical critical points, such as the decay of bubbling intensity and the approaching saturation of oil exudation rate.

[0065] The hardware implementation details of the described neuromorphic in-memory computing knowledge graph database are further elaborated. This database is physically built on a computing chip based on a resistive random access memory (RRAM) array. Unlike traditional storage architectures, the conductance of the RRAM cells is used to directly represent the weights in the knowledge graph. When semantic query vectors are input to the array in the form of voltage, according to Kirchhoff's current law, the output current directly corresponds to the similarity dot product operation between vectors. This physical-level in-memory computing fusion improves the system's retrieval speed by two orders of magnitude compared to traditional server architectures, while reducing power consumption. This efficient hardware support is the physical foundation for achieving real-time retrieval and control.

[0066] The dynamic trend consideration logic of the approximate nearest neighbor strategy retrieval device is further detailed below. When performing high-dimensional space retrieval, the device first defines the difference between the current vector Q and the vector Q-pre from the previous time step as the evolution velocity vector V. When searching for similar trajectories in the database, the system not only requires the feature point Q to have the minimum distance to point P in the database, but also requires that the angle between V and the tangent direction of the path in the database be minimized. This direction-based retrieval ensures the consistency of the control strategy. If the system finds that the current physical evolution velocity exceeds the coverage of all historical successful trajectories, the retrieval device will automatically trigger an extrapolation alarm and prompt the generative decision-making module to enter a safety protection mode, that is, to prioritize the execution of a conservative strategy aimed at preventing the food from burning.

[0067] The dynamic trend consideration logic of the approximate nearest neighbor strategy retrieval device is further elaborated. After the hierarchical navigation small-world algorithm locates the candidate feature cluster, the system does not simply rely on Euclidean distance, but constructs a comprehensive distance metric function that incorporates dynamic evolution trends. Let the current time be... The semantic query vector is The vector of the historical candidate trajectory in the corresponding time window is ; before the current moment The semantic evolution velocity vector within the time window is denoted as The acceleration vector is denoted as (The speed of historical trajectory) With acceleration Similarly, calculate the comprehensive distance metric function. The formula is as follows:

[0068] in, This represents the overall distance metric. The squared L2 norm of a vector. , , These are the static state deviation penalty coefficient, the evolution velocity deviation penalty coefficient, and the evolution acceleration deviation penalty coefficient, respectively, and they satisfy the following conditions: In this embodiment, to ensure the foresight of the strategy, a setting is made. , , System selection The trajectory with the smallest value is selected as the optimal match, thereby ensuring that the retrieved strategy is not only similar in the current state, but its future physical evolution trend is also highly consistent with the current ingredients.

[0069] The virtual inference algorithm of the generative counterfactual reasoning decision-making device is further described in detail. This device incorporates a lightweight physics simulation engine, but instead of being based on explicit partial differential equations, it is based on a data-driven recurrent neural network. This network learns a nonlinear dynamic model of how semantic feature vectors change with control commands. During counterfactual reasoning, the system generates multiple candidate control sequences in parallel. These candidate sequences are then input into the recurrent neural network to observe whether the predicted fragile feature vectors will enter the overheated or undercooked zone. The final selected advanced control command is a set of multidimensional control parameters, including nozzle spatial coordinates, spray pressure waveform, and pulse width modulation ratio.

[0070] The multi-degree-of-freedom (DOF) collaborative mechanism of the spray actuator is further detailed. The multi-DOF robotic arm employs nonlinear dynamic compensation control, eliminating inertial impact during high-speed reversal and ensuring the nozzle is always aligned with the food surface along the optimal normal direction. The high-precision flow proportional valve features a special anti-clogging design, capable of handling brittle water media containing fine particles or high viscosity. The transducer surface of the ultrasonic atomizing nozzle is covered with an ultra-hard DLC (diamond-like carbon) film to prevent frequency shifts in high-temperature corrosive environments. These three sub-components, synchronized via a fieldbus, form a highly collaborative physical intervention entity, achieving precise conversion from semantic commands to physical intervention.

[0071] Example 2: To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0072] This embodiment describes an intelligent spray crisping control system based on an edge computing distributed architecture and advanced control. This architecture is particularly suitable for large-scale industrial food production lines with multiple parallel heating units. In this embodiment, while maintaining the core logic of Embodiment 1, the system achieves a balance of computational load and improved system reliability through hardware topology reconstruction.

[0073] The intelligent spraying and crisping control system based on advanced control includes a distributed multimodal sensor subsystem, an edge-side semantic processing node, a centralized brain-like knowledge management server, a distributed strategy execution control unit, and a cyclic spraying matrix. The distributed multimodal sensor subsystem is deployed at each independent heating station to collect the raw state signals of the food ingredients within that station. Each subsystem includes an independent signal preprocessing module for anti-aliasing filtering of the acoustic signals and dark current compensation of the spectral signals. This distributed acquisition method avoids noise interference caused by long-distance transmission of analog signals and improves the signal-to-noise ratio of the data.

[0074] The edge-side semantic processing node is installed in an industrial cabinet next to the production line and is connected to the distributed multimodal sensor subsystem via an industrial Ethernet network. This node embeds a compressed, lightweight semantic mapping model, enabling real-time conversion of raw signals into semantic feature vectors using an FPGA acceleration card. Because the processing is completed at the edge, the system can delay the output of the current food ingredient's state representation, thus improving the system's response bandwidth.

[0075] The centralized neuromorphic knowledge management server communicates with multiple edge semantic processing nodes via a high-speed fiber optic network. This server cluster carries a complete neuromorphic in-store computing knowledge graph database. Due to its centralized management model, the server can aggregate production data from all workstations along the entire production line. When a workstation encounters a rare situation, it can instantly share successful response trajectories accumulated by other workstations, achieving collaborative optimization through collective intelligence. Furthermore, the centralized server is also responsible for periodically training and updating the global causal model offline and distributing the updated model parameters to each edge node.

[0076] The distributed strategy execution control unit is configured at each spray station, receiving advanced control commands retrieved and derived from a centralized server. This unit integrates a hardware-level safety closed-loop circuit. When executing the control strategy issued by the server, this unit verifies in real time whether the physical parameters at the execution end are within safety thresholds. If the communication network is interrupted, the distributed strategy execution control unit can immediately switch to a pre-stored offline baseline control strategy to ensure the continuity and safety of the production process.

[0077] The described circulating spray matrix replaces the single robotic arm structure of Example 1, employing a fixed matrix arrangement of ultrasonic nozzles. Each nozzle is equipped with an independent control valve group. The system dynamically activates nozzle combinations at different positions and frequencies via generative decision commands, creating a wave-like or scanning spray coverage effect. While this matrix structure sacrifices spatial mobility flexibility, it significantly increases system throughput, enabling it to adapt to high-speed, fluid industrial production lines.

[0078] In this distributed architecture, the generative counterfactual reasoning decision-making device is divided into a two-level decision architecture. The first level is a rapid response inference at the edge, responsible for compensating for millisecond-level fluctuations; the second level is a global optimal inference at the centralized server, responsible for long-term optimization of the process path throughout the entire production cycle. The two levels of decision-making are integrated through hierarchical scheduling logic, ensuring both local control accuracy and overall quality stability.

[0079] The edge-side semantic processing nodes are also equipped with a local caching mechanism. This mechanism can store production trajectories from the most recent hour. When the communication latency between the edge node and the central server suddenly increases, the node can utilize the local cache to perform a simplified approximate nearest neighbor search, thus achieving robustness against weak networks. This architectural design improves the system's reliability in complex industrial electromagnetic environments.

[0080] The control logic of the described cyclic spray matrix is ​​further detailed. The system utilizes generative counterfactual reasoning to predict which nozzle in the matrix will be within the coverage area of ​​the food ingredient at any given moment as it moves along the conveyor belt. The decision command precisely calculates the activation time and spray pulse width of each nozzle, ensuring that the total amount of spray medium received by the food ingredient during its movement conforms to a preset optimal evolution curve. This time-for-space based control method solves the uniformity problem in large-scale continuous production.

[0081] Furthermore, the quality evaluation feedback interface in this embodiment adopts a combined approach of online visual inspection and offline physicochemical analysis. The online visual inspection system is installed at the production line exit, using a high-speed camera and image analysis module to evaluate the surface brittleness uniformity and color of each product in real time. Offline analysis data is periodically synchronized from the laboratory management system. Both work together on a centralized neuromorphic knowledge management server, continuously optimizing path weights in the knowledge graph through reinforcement learning algorithms, thereby increasing the automation level of the entire production line as operating time increases.

[0082] Example 3: To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0083] This embodiment describes the application of an intelligent spray crisping control system based on advanced control in a mobile flexible processing unit. In this embodiment, the system is highly integrated into a mobile automation module, primarily designed to handle customized catering production scenarios with multiple varieties and small batches, emphasizing the system's rapid deployment capabilities and cross-species adaptive perception.

[0084] The intelligent spraying and crisping control system based on advanced control includes an integrated multimodal perception sphere, an embedded semantic feature engine, a cloud-synchronized knowledge base interface, an adaptive decision-making main control board, and a flexible collaborative spraying robot. The integrated multimodal sensing sphere is installed at the top center of the processing unit. Inside the sphere is a set of miniaturized high-frequency acoustic sensors and a multispectral imager. The sphere's outer shell is made of high-temperature resistant quartz glass and equipped with an automatic air curtain dust removal system to prevent oil fumes from condensing on the glass surface and affecting sensing accuracy. This sphere can monitor the processing area below from a panoramic perspective, acquiring real-time physicochemical signals of the food ingredients.

[0085] The embedded semantic feature engine is implemented using a low-power on-chip system. The engine is pre-installed with general semantic mapping operators for various common food ingredients (such as roast duck, suckling pig, and crispy pigeon). By introducing a domain-adaptive algorithm, the engine can identify common physical property features between new and known food ingredients. For example, it can identify that regardless of the food ingredient, the thermal shrinkage of collagen in the skin exhibits a similar phase shift pattern in millimeter-wave echoes.

[0086] The cloud-based synchronous knowledge base interface is used to upload locally collected semantic vectors to the cloud-based expert system and download the latest optimal control trajectory. This interface supports 5G or satellite communication, enabling processing units located in different geographical locations to share the same massive knowledge base. This single-machine processing, globally intelligent model allows even remote processing sites to produce crispy products of the same quality as those from the headquarters laboratory.

[0087] The adaptive decision-making main control board receives the preliminary strategy from the cloud and performs secondary corrections based on real-time environmental parameters (such as local atmospheric pressure and humidity) obtained from the local sensing sphere. The main control board integrates parameter fine-tuning logic based on counterfactual reasoning, enabling localization of the cloud-based strategy within milliseconds. For example, in high-altitude, low-pressure areas, the system automatically lowers the preset boiling point of the spray medium and adjusts the moisture evaporation constant in the simulation model accordingly.

[0088] The flexible collaborative spraying robot employs a lightweight six-axis flexible collaborative robotic arm. This arm features collision self-detection and force feedback, enabling it to work safely and collaboratively with a human operator within the same space. The end effector of the robotic arm integrates a multi-functional composite nozzle, capable of both wide-area spraying and narrow-beam precision spraying. Based on counterfactual reasoning, the system switches the nozzle's operating mode in real time to adapt to the complex geometric contours of the food surface.

[0089] The automatic calibration logic of the integrated multimodal sensing sphere is further detailed below. Before each startup, the system automatically performs benchmark calibration on its internal sensors. The acoustic sensors perform self-testing by emitting standard audio through an internally integrated miniature speaker; the spectrometer performs wavelength calibration by pointing to a built-in standard color chart. This highly integrated self-testing mechanism ensures consistent sensing even after frequent movements and restarts.

[0090] The adaptive decision-making main control board also features shadow control functionality. While executing control commands, the system runs another pre-set, manually configured empirical scheme in the background. By comparing the expected effects of the retrieval-driven strategy and the empirical scheme in the deduction model, the system can quantify the quality gains brought by the advanced control scheme. If a potential risk is detected in the retrieved strategy under the current environment, the system automatically and smoothly transitions to the shadow scheme, ensuring the flawless execution of the process.

[0091] Furthermore, this embodiment also develops a semantic-based interactive control interface to meet the needs of customized production. Operators only need to input their desired crispier crust or reduced oil content using natural language, and the system automatically translates these sensory requests into target deviation vectors in a high-dimensional semantic space. The near nearest neighbor retrieval device then readjusts the retrieval priority based on this deviation vector, searching for historical trajectories that perform better in relevant dimensions. This direct path from natural language to advanced control parameters lowers the operational threshold of intelligent equipment.

[0092] The system is also equipped with a mobile energy and media management unit, providing a stable power supply and a constant temperature and pressure spraying medium for the entire control system. The management unit integrates an automatic real-time media composition adjustment device, capable of adjusting the proportions of sugar, acetic acid, and spices in the crisping water in real time according to decision commands. This allows the system to not only adjust how it sprays, but also what it sprays, further expanding the dimensions of advanced control and achieving precise construction of both product flavor and physical texture.

[0093] The intelligent spray crispy skin control system based on advanced control described in this invention is not limited to the sensor models, processor architectures, or actuator forms specifically mentioned in the above embodiments. Without departing from the core concept of this invention, any technical solution that transforms the cross-modal physical features of the food surface into semantic vectors and utilizes knowledge graphs for retrieval and causal inference to achieve spray control should be included within the scope of protection of this invention.

[0094] The implementation details of the self-supervised learning, neuromorphic computing, approximate nearest neighbor retrieval, and counterfactual reasoning logic involved in this invention can be dynamically adjusted according to the development of specific computing hardware. The various modules in the system can be connected via wired or wireless means, and their physical locations can be within the same cavity or distributed across the edge and cloud.

[0095] For those skilled in the art, various improvements and modifications can be made without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. Hardware details not detailed in the claims can be implemented using existing standard industrial components. The solutions of this invention are not only applicable to traditional meat processing such as crispy pork and roast duck, but can also be extended to novel food fields such as plant-based meat and artificial protein products with similar surface treatment processes.

Claims

1. An intelligent spray crisping control system based on advanced control, characterized in that, It includes a multimodal sensor array, a semantic feature mapping device, a brain-like in-memory computing knowledge graph database, an approximate nearest neighbor strategy retrieval device, a generative counterfactual reasoning decision-making device, and a spraying actuator; The multimodal sensor array is used to collect multidimensional physicochemical state signals of food during the heating and cooking process in real time, so as to obtain raw data streams that reflect the blistering and rupture characteristics of the food surface, the rate of oil seepage, and the geometric deformation state of the food. The semantic feature mapping device is used to receive the raw data stream of the geometric deformation state of the food, and through a preset self-supervised contrastive learning logic, map the unstructured raw data stream to a high-dimensional semantic feature vector space to generate a semantic query vector representing the current crispy evolution state of the food. The brain-like storage and computing integrated knowledge graph database is used to store multiple batch trajectory data structures in the historical production process. Each trajectory data structure contains a semantic feature vector at the production time, the executed spraying action instruction, and the corresponding crispy skin quality output effect. The approximate nearest neighbor strategy retrieval device is used to perform high-dimensional similarity matching in the brain-like storage and computing integrated knowledge graph database using the semantic query vector as the search term, and retrieve the set of historical successful strategy trajectories. The generative counterfactual reasoning decision-making device is used to perform causal inference on the retrieved set of historical successful strategy trajectories. By simulating the evolution path of fragile skin characteristics under different spraying intervention schemes in the semantic space, the optimal control strategy is selected and encapsulated into advanced control instructions. The spraying actuator is used to receive the advanced control commands and adjust the flow rate, pressure, atomization particle size, and movement trajectory of the spraying medium accordingly to achieve precise intervention on the surface of the food.

2. The intelligent spray crispy skin control system based on advanced control according to claim 1, characterized in that, The multimodal sensor array includes: a high-frequency acoustic acquisition unit, installed inside the heating environment, equipped with an acoustic waveguide structure to transmit internal acoustic vibrations to a sensor probe located outside the high-temperature zone, used to capture tiny acoustic pulses generated on the surface of the food during cooking. These tiny acoustic pulse signals reflect the breaking frequency of the crispy skin's microstructure. The near-infrared spectral sensing unit integrates a spectrometer and an array detector. A millimeter-wave detection unit is used to transmit and receive frequency-modulated continuous wave signals. By analyzing the phase change and delay of the echo signal, it obtains the deformation parameters of the food caused by heat, such as contraction, expansion, or twisting.

3. The intelligent spray crispy skin control system based on advanced control according to claim 2, characterized in that, The semantic feature mapping device has built-in feature decoupling logic and dynamic weight allocation logic: the feature decoupling logic is used to perform short-time Fourier transform and feature alignment on the original data stream, and uses a combination algorithm of principal component analysis and independent component analysis to separate the food reaction signal from the environmental background noise, and extract the essential feature components that reflect the formation of crispy skin. The dynamic weight allocation logic is used to monitor the clarity index of the current production environment. When the concentration of oil fume in the environment exceeds a preset threshold, causing the near-infrared spectral signal to attenuate, the contribution weight of the near-infrared spectral sensing unit in the process of synthesizing the semantic query vector is automatically reduced, and the contribution weight of the millimeter-wave detection unit is increased simultaneously to output a stable semantic representation. The semantic feature mapping device uses a built-in deep residual network architecture to convert the processed signal features into a dense vector as the semantic query vector.

4. The intelligent spray crispy skin control system based on advanced control according to claim 3, characterized in that, The brain-like in-memory computing knowledge graph database is physically built on a computing chip based on a resistive switching memory array: the conductance value of the storage unit in the resistive switching memory array is used to represent the path weight in the knowledge graph, to receive the semantic query vector in the form of voltage, and to perform the similarity dot product operation directly in the storage unit according to Kirchhoff's current law; the trajectory data structure is organized in the form of a graph network, where nodes represent the semantic state at the moment of production, and edges represent the transition probability between states and the control intervention applied during the transition process; The brain-like in-memory computing knowledge graph database is also equipped with cross-batch transfer learning logic, which is used to extract similar physical and chemical change patterns between different types of ingredients when faced with new types of ingredients, match the semantic feature evolution path of the new ingredients in the early stage of heating, and transfer control strategies from the historical trajectory of known ingredients.

5. The intelligent spray crispy skin control system based on advanced control according to claim 4, characterized in that, The approximate nearest neighbor strategy retrieval device employs a hierarchical navigation small-world algorithm: the algorithm is used to construct a multi-layered index structure, and locates the nearest feature cluster by calculating the Euclidean distance or cosine similarity between the semantic query vector and the feature vectors stored in the brain-like storage and computing integrated knowledge graph database; The approximate nearest neighbor strategy retrieval device is equipped with dynamic trend consideration logic. It extracts the semantic evolution trend within a predetermined time period before the current moment through a sliding window mechanism, and calculates the first and second derivatives of the semantic vector to evaluate the acceleration and direction of the change in the state of the food, ensuring that the angle between the retrieved strategy and the current food evolution speed vector is within a preset range. The approximate nearest neighbor strategy retrieval device has a limited retrieval time window to meet the real-time control requirements of the system.

6. The intelligent spray crispy skin control system based on advanced control according to claim 5, characterized in that, The generative counterfactual reasoning decision-making device includes a causal model construction unit, a virtual state deduction unit, and a shadow control module: the causal model construction unit uses historical correlation data in the brain-like storage and computing integrated knowledge graph database to construct a nonlinear causal chain between spraying action commands and crisp skin quality indicators, and distinguishes between physical changes caused by spraying intervention and spontaneous changes caused by environmental heating. The virtual state deduction unit has a built-in data-driven dynamics model based on recurrent neural networks, which is used to generate multiple candidate control sequences in parallel for the current time and to pre-determine the semantic feature vector evolution path of each candidate control sequence in a predetermined time period in the future. The generative counterfactual reasoning decision-making device is used to compare the simulation results under multiple candidate control sequences, select the strategy path that minimizes the standard deviation of crispness and has the highest color uniformity score, and encapsulate it as the advanced control instruction. The shadow control module is used to run manually preset experience schemes in the background, quantify the quality gains brought by the advanced control instructions, and perform a safe and smooth switch when a strategy risk is detected.

7. The intelligent spray crisping control system based on advanced control according to claim 6, characterized in that, The spraying actuator includes a multi-degree-of-freedom robotic arm, a high-precision flow proportional valve, and an ultrasonic atomizing nozzle. The multi-degree-of-freedom robotic arm is equipped with a nonlinear dynamic compensation control program, which is used to execute a spatial coverage trajectory according to the advanced control instructions, and to achieve millimeter-level positioning accuracy at the end of the robotic arm through a collision detection sensor and a position compensation unit. The high-precision flow proportional valve uses a ceramic valve core structure to respond to the flow regulation requirements in the advanced control command within milliseconds. The transducer surface of the ultrasonic atomizing nozzle is covered with a diamond-like film, which is used to change the particle size of the atomized spray medium by adjusting the driving frequency, so as to meet the penetration requirements of different food surface porosities.

8. The intelligent spray crispy skin control system based on advanced control according to claim 7, characterized in that, The system also includes an environmental perception correction unit and a robust switching control unit: the environmental perception correction unit includes a thermocouple array and a humidity sensor array deployed around the heating cavity, used to monitor the fluctuations in ambient temperature, humidity and wind speed inside the cavity, and feeds back the fluctuations as correction factors to the generative counterfactual reasoning decision device in real time, so as to adjust the boundary conditions of counterfactual reasoning and compensate for the impact of environmental fluctuations on the rate of brittle formation. The robust switching control unit is used to control the near nearest neighbor strategy retrieval device to switch to robust retrieval mode when the multimodal sensor array detects a sudden change or absence of a signal in a certain dimension outside a predetermined range. It retrieves non-precise coverage spraying strategies that have performed robustly in high-interference scenarios from historical records and replaces point-to-point high-precision operations by expanding the spraying coverage area and increasing the spraying frequency.

9. The intelligent spray crispy skin control system based on advanced control according to claim 8, characterized in that, The system also includes a media recovery and circulation filtration unit: the media recovery and circulation filtration unit includes a collection slope located at the bottom of the heating chamber, a stainless steel primary filter, an activated carbon adsorption tower and a component analyzer, used to collect excess media that does not adhere to the surface of the food during the spraying process, and after filtration and component balance adjustment, it is sent back into the spraying pipeline. When generating the advanced control commands, the generative counterfactual reasoning decision-making device simultaneously considers the utilization rate of the medium. By optimizing the start-stop sequence and spray angle of the nozzles, it reduces the loss of the spray medium while ensuring the crispness of the coating.

10. The intelligent spray crispy skin control system based on advanced control according to claim 9, characterized in that, The system is also equipped with a quality evaluation feedback interface and an interactive control interface: the quality evaluation feedback interface is used to receive product sensory evaluation data, and the brain-like storage and computing integrated knowledge graph database corrects the weight of the corresponding trajectory data structure according to the feedback data, increases the recommendation weight for trajectories evaluated as excellent quality, and establishes negative sample constraints for trajectories with defects. The interactive control interface is used to receive demand instructions from natural language input and convert the demand instructions into a target deviation vector in a high-dimensional semantic space, so that the approximate nearest neighbor strategy retrieval device can adjust the retrieval priority according to the target deviation vector. The brain-like in-memory computing knowledge graph database has a dynamic evolution function. After each production task is completed, it encapsulates the current production trajectory, environmental disturbance factors and product qualification rate, and integrates them into the existing knowledge network after correlation analysis, so as to realize the closed-loop self-evolution of control logic.

Images