Method for intelligently regulating and controlling production parameters in production process of fruit concentrated juice

Abstract

The invention discloses a method for intelligently regulating and controlling production parameters in a fruit concentrated juice production process, which comprises the following steps of: acquiring multi-dimensional process parameters such as temperature, pressure, flow, concentration, equipment state and the like in real time through a multi-channel sensor network, and forming a standardized data sequence after filtering, normalization and drift correction; extracting stage features by using technologies such as a sliding window and Fourier transform, and inputting the stage features into the lightweight classification model to realize production stage identification; in combination with an identification result, dynamically calling a corresponding multi-target optimization sub-model, and realizing nonlinear prediction and optimal solution selection of process parameter setting by adopting an LSTM and a multi-target genetic algorithm; on the basis of real-time feedback, the performance of the model is automatically evaluated, self-adaptive adjustment and optimization of the optimization algorithm are achieved through reinforcement learning and an incremental updating mechanism, multi-target collaborative optimization, self-adaptive adjustment and model switching in the production process can be achieved, and the consistency of production efficiency and product quality is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial process automation and intelligent optimization control, and particularly relates to a method for intelligent regulation of production parameters in a fruit concentrate production process. BACKGROUND

[0002] With the continuous improvement of the intelligent level of the food processing industry, the automatic parameter regulation of complex multi-stage processing technology such as fruit concentrate has become an important direction of current industry development. The existing technology generally adopts a production parameter regulation scheme based on multi-objective optimization, cooperatively considers key targets such as yield, energy consumption, quality, and equipment wear by means of intelligent process modeling and optimization algorithm, and dynamically predicts and optimizes core process parameters such as temperature, pressure, flow, and concentration. This technical route is based on multi-objective optimization algorithms (such as linear weighting method, Pareto frontier algorithm, genetic algorithm, etc.), combined with an industrial Internet of Things sensor network, to realize automatic monitoring and parameter adjustment of the production process, and has become the mainstream scheme and technical hotspot in the field of intelligent process control of the food processing industry.

[0003] In actual production, intelligent systems including multi-objective optimization parameter regulation based on a unified model, multi-parameter setting of a fixed target weight method, and process anomaly identification through time series analysis have been widely applied to multi-stage process scenarios such as concentrate evaporation, sterilization, and filtration. These technologies usually collect multi-dimensional sensor data such as temperature and pressure, build a unified multi-objective optimization model, distribute targets such as yield, energy consumption, and quality according to a preset weight, and output a set of optimal parameter setting values in a nonlinear dynamic prediction module for closed-loop control instructions to realize automated production.

[0004] However, most existing multi-objective optimization control systems have key technical bottlenecks in parameter weight distribution and target coordination mechanisms. The mainstream technical solutions are mainly based on "fixed target weight" or "unified model direct optimization", such as the automatic process modeling and coordinated parameter regulation method in the existing technical disclosure, which usually sets the weight proportion of each optimization target in advance and always uses a fixed target weight matrix for solving at different production stages, lacking the ability to adaptively respond to the differences between process stages and the dynamic changes of production targets. In actual multi-stage production processes, this mode often fails to balance the target conflicts between stages, for example, the stable heating stage pays more attention to energy consumption and equipment wear, while the concentration stage pays more attention to product quality and concentration accuracy, resulting in a fixed weight model that cannot flexibly balance multiple targets such as yield, cost, and quality.

[0005] In addition, the prior art adopts a multi-objective unified modeling method, ignores the dynamic characteristics and context differences of each stage in the process flow, and is prone to imbalance of target priority or intensification of target conflict in a certain stage, thereby reducing the actual regulation accuracy and system intelligence of the optimization result. Although some intelligent decision systems introduce timing characteristics and stage identification mechanisms, they are mainly based on static mapping rules, and the response to new stage characteristic changes is lagging, lacking the ability of adaptive model switching and dynamic weight adjustment in the context environment. When dealing with production target changes, equipment running state fluctuations and abnormal working condition identification, problems such as model switching delay, reduced optimization effect and poor system stability often occur. SUMMARY

[0006] The present application provides an intelligent production parameter regulation method in a fruit concentrate production process.

[0007] The technical scheme of the present application is implemented as follows: an intelligent production parameter regulation method in a fruit concentrate production process, comprising:

[0008] S1: collecting real-time running parameters of multiple process stages in the fruit concentrate production process, including temperature, pressure, flow, concentration and equipment state signals, and recording the corresponding time stamps and process numbers of each parameter;

[0009] S2: filtering and normalizing the collected multi-dimensional parameter data to eliminate abnormal values and sensor drift errors, to generate a standardized process data sequence;

[0010] S3: based on a time series feature extraction method, identifying the characteristic vector of the current production stage, the characteristic vector including parameter change rate, trend inflection point and periodic fluctuation mode;

[0011] S4: inputting the characteristic vector into a stage identification model to perform lightweight classification reasoning to determine the current production stage and output the corresponding stage label;

[0012] S5: according to the stage label, calling a matching optimization model from a pre-constructed multi-objective optimization sub-model set, the optimization model being configured with a weight distribution and constraint condition suitable for the stage;

[0013] S6: inputting the current standardized parameter sequence and historical data into the called optimization sub-model to perform nonlinear dynamic prediction and generate the optimal process parameter setting value at the next time;

[0014] S7: outputting the optimal process parameter setting value to the control system for dynamic regulation of the actuator to achieve production parameter regulation under multi-objective collaborative optimization;

[0015] S8: Collect the actual process parameter feedback data after regulation, calculate the realization deviation of each optimization target, and generate a model performance evaluation index;

[0016] S9: Based on the model performance evaluation index, the parameter configuration and switching logic of each optimization sub-model are updated online to realize model adaptive optimization and improve stage recognition accuracy.

[0017] The fruit concentrate production process intelligent parameter regulation method provided by the application has the following beneficial effects:

[0018] (1) The present application uses high-precision stage feature extraction and recognition technology to divide the fruit concentrate production process into clear multi-process stages, and then establishes independent multi-objective optimization sub-models for each stage, respectively configures different target weights and constraint conditions. Combined with the feature-driven switching and reinforcement learning online weight optimization of real-time context perception, the system can adaptively trade off energy consumption, quality, efficiency and other optimization targets in different stages, significantly improving the flexibility of multi-objective collaborative optimization compared with the fixed weight strategy, and effectively avoiding the performance loss caused by stage target conflict;

[0019] (2) For the nonlinearity and parameter dynamic fluctuation problem of the concentrate production process, the present application innovates in multiple levels in the model structure: on the one hand, a multi-channel sensor network and distributed data preprocessing are used to ensure high-quality real-time parameter input in each stage, thereby improving data reliability from the source; on the other hand, advanced prediction and optimization algorithms such as LSTM and NSGA-II are used to significantly improve the adaptability of the model to dynamic process. The model switching driven by stage features can switch the optimal prediction and control strategy according to the actual state of the process, so that the response time to production disturbance is greatly shortened, and the prediction accuracy is significantly improved;

[0020] (3) The multi-objective dynamic optimization and closed-loop adaptive control mechanism of the present application effectively realizes the accurate and timely adjustment of temperature, pressure, flow, concentration and other core indicators, reduces the waste of resources and overbudget energy consumption caused by process misjudgment or delayed adjustment. Under the condition of automatic model switching, it can real-time focus on energy saving, quality, equipment wear and other multi-dimensional targets, balance cost and performance, and the product consistency is obviously improved;

[0021] (4) Through a series of fine data processing links such as data filtering, feature normalization, wavelet denoising and principal component dimensionality reduction, the present application greatly improves the accuracy of stage recognition and model switching, has self-adaptive fault tolerance capability to sensor drift, outliers and other interference, and ensures the stable operation of the system under the conditions of different batches, different raw materials and equipment parameter changes. Especially through the incremental learning and online model iteration mechanism, the present application can continuously optimize its own strategy with the change of the production environment, significantly improving the process intelligence and generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0022] Fig. 1 A flow chart of the intelligent production parameter regulation method in the fruit concentrate production process according to the present application;

[0023] Fig. 2 A sub-flow chart of the intelligent production parameter regulation method in the fruit concentrate production process according to the present application;

[0024] Fig. 3 Another sub-flow chart of the intelligent production parameter regulation method in the fruit concentrate production process according to the present application. DETAILED DESCRIPTION

[0025] In order to make the objects and advantages of the present application more clear and obvious, the present application will be further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0026] The preferred implementation methods of the present application will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these implementation methods are only used to explain the technical principles of the present application and do not limit the protection scope of the present application.

[0027] As used herein, the singular forms "a", "an" and "the" also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms "comprise / comprising" or "have / having" specify the presence of stated features, integers, steps, operations, components, parts or combinations thereof, but do not exclude the presence or addition of one or more other features, integers, steps, operations, components, parts or combinations thereof. At the same time, the term "and / or" used in the specification includes any and all combinations of the related listed items.

[0028] Referring to Figs. 1-3 As shown in the drawings, an intelligent production parameter regulation method in a fruit concentrate production process includes:

[0029] S1: Collecting real-time running parameters in multiple process stages in the fruit concentrate production process, including temperature, pressure, flow, concentration and equipment state signals, and recording the time stamps and process numbers corresponding to each parameter;

[0030] S2: Filtering and normalizing the collected multi-dimensional parameter data to eliminate abnormal values and sensor drift errors to generate standardized process data sequences;

[0031] S3: Identifying the feature vector of the current production stage based on the time series feature extraction method, the feature vector including the parameter change rate, trend inflection point and periodic fluctuation mode;

[0032] S4: inputting the feature vector into a stage recognition model, performing lightweight classification inference to determine the current production stage and outputting a corresponding stage label;

[0033] S5: according to the stage label, calling a matched optimization model from a pre-constructed set of multi-objective optimization sub-models, the optimization model being configured with a weight distribution and constraint condition suitable for the stage;

[0034] S6: inputting the current normalized parameter sequence together with historical data into the called optimization sub-model, performing nonlinear dynamic prediction to generate optimal process parameter setting values at the next time;

[0035] S7: outputting the optimal process parameter setting values to a control system for dynamic regulation of an actuator to achieve production parameter adjustment under multi-objective collaborative optimization;

[0036] S8: collecting feedback data of the actual process parameters after regulation and control, and calculating the implementation deviation of each optimization objective to generate a model performance evaluation index;

[0037] S9: based on the model performance evaluation index, online updating the parameter configuration and switching logic of each optimization sub-model to realize model adaptive optimization and improve stage recognition accuracy.

[0038] The step S1: collecting real-time running parameters of multiple process stages in the fruit concentrate production process, including temperature, pressure, flow, concentration and equipment state signals, and recording the time stamps and process numbers corresponding to each parameter. Specifically, it includes:

[0039] S1.1: based on the industrial Internet of Things architecture, deploy a multi-channel sensor network to collect data from temperature sensors, pressure transmitters, flow meters, concentration detectors and equipment state monitoring devices in the fruit concentrate production process to obtain original running parameters of each process stage;

[0040] Based on the industrial Internet of Things architecture, configure multi-channel sensor network node parameters to achieve global coverage of temperature, pressure, flow, concentration and equipment state signals at key process points on the fruit concentrate production line;

[0041] Use a distributed sensor access protocol (parameters: MQTT QoS level set to 1-2, preferably 2, node heartbeat interval 1-20 seconds, preferably 5 seconds) to realize real-time transmission of heterogeneous data from multiple types of sensors and ensure the reliability and sequential consistency of data transmission;

[0042] Further, by deploying the protocol conversion module of Modbus-TCP / IP to industrial Ethernet (parameters: sampling period of 0.1-2 seconds, preferably 1 second, buffer capacity of 256-4096 frames, preferably 1024 frames) through the centralized gateway node, the data channel of fieldbus sensors and the upper computer system is unified, and multi-source raw signal streams are obtained;

[0043] Further, by using the hardware time synchronization module (based on IEEE1588 PTP protocol, clock deviation calibration accuracy better than 1 microsecond), the timestamps of the collected data of each channel are aligned, and time marker signals accurately corresponding to the actual physical events are generated;

[0044] Further, by the sensor health state diagnosis algorithm (parameters: response time 50ms, threshold voltage drift 0.05V), the working state of each collection node is detected online, and a stability diagnosis report is generated;

[0045] Through the architectural deployment and data access mechanism, the collected raw operating data of temperature, pressure, flow, concentration and equipment state are converted into multi-dimensional data packets with complete timestamps, sensor state labels and spatial distribution position information, providing high-precision, traceable structured data sources for subsequent filtering and feature extraction;

[0046] For example, on a certain fruit concentrate evaporation concentration production line, Pt100 platinum resistance sensors are installed for temperature collection in the evaporation section, with a graduation number of A, a temperature measurement range of -50 to 250°C, and a sampling frequency of 10Hz; a pressure transmitter is installed at the inlet of the vacuum pump for pressure collection, with a range of 0 to -0.1MPa, an accuracy level of 0.25, and a sampling frequency of 5Hz; an electromagnetic flowmeter is installed on the feed pipe for flow monitoring, with a range of 0 to 50m 3 / h, a sampling frequency of 2Hz; an online refractometer is installed at the discharge port for concentration detection, with a detection range of Brix 0 to 70%, and a sampling frequency of 1Hz; a current transformer is introduced into the key motor control loop for equipment state monitoring, to realize the state current threshold discrimination of running / stop. Each sensor sends data to the industrial gateway through the MQTT protocol, the gateway performs Modbus and TCP / IP data frame conversion, and the timestamps of each data packet are calibrated through the IEEE1588 synchronization module. In a batch production, the system captures that the temperature interval is stable at 92 to 94°C, the vacuum degree is maintained at -0.09MPa, the flow is 18 to 20m 3 / h, the concentration is stable at Brix 63%, and the equipment state label is always in running state, verifying the stability and high precision of the data collection system, and providing reliable training and operation data for the nonlinear dynamic prediction model;

[0047] S1.2: Perform sliding window filtering on the collected temperature signal, use Butterworth low-pass filter to eliminate high-frequency noise interference, to obtain stable temperature measurement value as input data for stage feature extraction;

[0048] For the original temperature signal from the Pt100 platinum resistance temperature sensor, a sliding window filtering method is used (parameters: window length 50 sampling points, sampling frequency 10 Hz) to realize local smoothing processing of temperature data in continuous time period;

[0049] Further, by constructing a Butterworth low-pass filter (parameters: filter order 4, cutoff frequency 1 Hz), the high-frequency noise components in the temperature signal within the window higher than the cutoff frequency are effectively suppressed;

[0050] Further, based on the bilinear transformation method, the Butterworth filter is discretized, the analog filter transfer function H(s) is converted to digital filter transfer function H(z), and the digital filter coefficient matrix is obtained, which is used for real-time discrete convolution operation in the acquisition system;

[0051] Further, for the temperature signal after digital filtering, the stability coefficient σ of the temperature within the window is calculated using the root mean square (RMS) fluctuation rate index, to quantify the smoothness of the filtering result;

[0052] Further, through the stability coefficient threshold discrimination mechanism (parameters: stability threshold 0.05℃), the online monitoring of the quality of the filtered temperature data is realized, and the abnormal window data with excessive fluctuation is eliminated;

[0053] Through the joint processing method of Butterworth low-pass filter and sliding window method, the original temperature signal of the previous step is converted into stable temperature measurement value with high signal-to-noise ratio, high smoothness and meeting the accuracy requirements of stage feature extraction, realizing the high reliability guarantee of temperature dynamic feature input in the subsequent stage feature extraction module;

[0054] For example, in the temperature monitoring of a certain evaporation concentration section, the Pt100 platinum resistance sensor outputs a raw temperature sequence with a frequency of 10 Hz, after window smoothing processing with a sliding window length of 50 points, more than 94% of the high-frequency noise energy is suppressed by using a 4th order Butterworth low-pass filter with a cutoff frequency of 1 Hz. The filter continuously applies the bilinear transformation method to realize the transfer function

[0055] The conversion from H(s) to H(z) is performed, and the digital filter coefficients b = [0.0048, 0.0192, 0.0288, 0.0192, 0.0048] and a = [1, -2.3695, 2.3140, -1.0547, 0.1874] are derived, which are directly operated in the data acquisition card through a real-time IIR filtering algorithm. The temperature stability coefficient calculation result in the window is 0.012℃, which is lower than the stability threshold of 0.05℃, so the window data is transmitted as valid input to the stage feature extraction module. In the training and verification stage of the stage recognition model, the signal-to-noise ratio of the temperature input is improved by about 18dB, and the stage recognition accuracy is improved by 3.2% compared with the unfiltered processing;

[0056] S1.3: The analog signal output by the pressure transmitter is subjected to analog-to-digital conversion processing, which is converted into digital pressure data based on the ADC module, and the dimensional difference is eliminated by the Z-score standardization method to generate a standardized pressure feature sequence;

[0057] S1.4: The pulse signal output by the flowmeter is subjected to frequency-voltage conversion and linear interpolation processing to obtain a flow variation curve on a continuous time scale, and a mapping relationship between the flow parameter and the process stage is established in combination with the process number;

[0058] S1.5: The refractive index signal output by the concentration detector is subjected to a non-linear compensation algorithm processing, and the temperature drift effect is eliminated based on a polynomial fitting method to generate high-precision concentration measurement values, which are stored in a real-time database in synchronization with the time stamp;

[0059] S1.6: The edge trigger detection processing is performed on the device state signal, and different working condition states such as device running, shutdown, cleaning, etc. are identified based on a state machine model to generate a device state marker sequence, which is used as an important context feature input for stage recognition;

[0060] S1.7: The processed temperature, pressure, flow, concentration and device state data are time-stamped and annotated with process numbers to generate a multi-dimensional parameter time series data set with stage labels, which is used as an input data source for subsequent feature extraction and stage recognition model.

[0061] The step S2: The collected multi-dimensional parameter data is subjected to filtering and normalization processing to eliminate abnormal values and sensor drift errors to generate standardized process data sequences. Specifically, it includes:

[0062] S2.1: The collected multi-dimensional process parameter signals are subjected to sliding window segmentation processing, and the temperature, pressure, flow, concentration and device state signals are time-aligned in time windows to generate structured time series data blocks, thereby improving the time consistency of subsequent filtering processing;

[0063] For temperature, pressure, flow, concentration and equipment state data sequences from multi-channel sensor network, a sliding window segmentation method (parameters: window length 1-30 seconds, preferably 5 seconds, window shift step 0.1-5 seconds, preferably 1 second) is used to extract local data blocks of multi-dimensional process parameter signals on continuous time axis, so as to maintain time synchronization between sampling points in subsequent processing stage;

[0064] Further, through a timestamp-based alignment algorithm (parameters: timestamp resolution 0.1-10 milliseconds, preferably 1 millisecond), the sampling points of temperature, pressure, flow, concentration and equipment state in the sliding window are accurately aligned, the time positions of different sampling frequency sensor data points are unified, and a uniform sampling sequence with equal time intervals is generated;

[0065] Further, a difference complement interpolation method (parameters: interpolation method is cubic spline interpolation) is used to fill in the missing sampling points caused by collection delay or instantaneous interruption, so as to ensure that the length of each parameter data column in the time window is consistent and there is no missing value, thereby enhancing the structural integrity of subsequent filtering processing;

[0066] Further, through a windowed data structured formatting algorithm, the aligned multi-dimensional signals are arranged in matrix form with uniform time index and parameter values, the matrix dimension is defined as N×M, where N is the number of sampling points in the window, and M is the number of process parameter dimensions, so as to be batched into the filtering and normalization module;

[0067] Through the above sliding window segmentation and time sequence alignment, the original process parameter stream from asynchronous multi-source is converted into structured time sequence data block with unified time reference, no missing value and matrix format, realizing the time sequence consistency and calculation accuracy improvement in the subsequent S2.2 wavelet threshold denoising processing stage;

[0068] For example, in a batch evaporation concentration production line, the temperature sampling frequency is 10 Hz, the pressure is 5 Hz, the flow is 2 Hz, the concentration is 1 Hz, and the equipment state is 1 Hz. Through sliding window segmentation with window length 5 seconds and step 1 second, different frequency data are combined into each window containing 50 temperature sampling points, 25 pressure points, 10 flow points, 5 concentration points and 5 state points. Based on cubic spline interpolation, the value of the missing point is obtained by smooth curve fitting of adjacent valid values, and the timestamp alignment accuracy is controlled within 1 millisecond. Finally, a 50×5 matrix format data block is output. It is verified that the time sequence consistency of this method is improved to less than 2 milliseconds of delay tolerance, and the signal reconstruction error mean square value is reduced by about 12.5% in subsequent wavelet denoising calculation;

[0069] S2.2: Perform filtering processing on each dimension parameter signal within the time window based on the wavelet threshold denoising algorithm, use db4 wavelet basis to perform multi-scale decomposition on the signal and set soft threshold to eliminate high frequency noise and sensor transient interference, obtain the reconstructed signal after denoising, and improve the signal-to-noise ratio of parameter data;

[0070] S2.3: Perform Z-score based standardization processing on the filtered parameter signal, calculate the mean and standard deviation of each dimension parameter within the time window, and perform normalization transformation on the original signal to generate a standardized parameter sequence with zero mean and unit variance, eliminating the influence of dimension difference on feature extraction;

[0071] S2.4: Perform drift error identification on the standardized parameter sequence based on the sliding window mean shift detection algorithm, if a significant trend deviation is detected, use difference processing to correct the sequence to eliminate the system error caused by long-term drift of the sensor, and generate standardized data without drift deviation;

[0072] S2.5: Align and splice the de-drifted standardized data sequence according to the process number and time stamp in time sequence to form a multi-dimensional process data matrix in a unified format, and add stage label and equipment state identification to generate a standardized input data set that can be used for subsequent feature extraction and model calling.

[0073] The step S3: based on the time series feature extraction method, identify the feature vector of the current production stage, the feature vector includes parameter change rate, trend inflection point and periodic fluctuation mode. As shown in Fig. 2

[0074] S3.1: Perform sliding window segmentation processing on the standardized process data sequence to extract the local dynamic features of each process parameter in the time dimension, thereby generating the parameter trend sequence within the time window;

[0075] S3.2: Perform first and second derivative calculation on the time window trend sequence of temperature, pressure, flow and concentration parameters based on the difference operation method, to obtain the change rate and inflection point feature vector of each parameter;

[0076] Based on the standardized time window trend sequence input, the numerical difference operation module is called for temperature, pressure, flow and concentration parameters (parameters: difference order = 1, sampling interval based on the unified time step of time window data), which realizes the first order difference processing on the sequence to describe the instantaneous change slope of each parameter in the time dimension, thereby obtaining the change rate data vector;

[0077] ​Furthermore, using a high-precision center difference formula (parameter: the difference step size is the sampling interval Δt (0.01-1 second, preferably 0.1 second)), the rate of change of each parameter at time point t is calculated:

[0078]

[0079] Where x is the time window trend sequence sample value of the parameter to be processed, and Δt is the sampling interval. The rate of change sequence calculated in this way can effectively characterize the dynamic response intensity of the process parameters;

[0080] Furthermore, for sequences with the same trend, the second-order difference operation module is called (parameter: difference order = 2) to calculate the second derivative to quantify the change in the rate of change, i.e., the acceleration characteristic:

[0081]

[0082] The second derivative reflects the curvature of the process parameter variation trend and is used to detect the inflection point of process state change.

[0083] Furthermore, by combining the second derivative sign change and amplitude threshold detection algorithm (parameter: amplitude threshold is set through historical data statistics), sampling points where the positive and negative signs switch within the time window and the amplitude exceeds the threshold are identified, thereby generating an inflection point marker sequence;

[0084] Furthermore, the rate of change data vector and the inflection point marker sequence are indexed and aligned in the time dimension and concatenated in the parameter dimension to generate a multidimensional rate of change-inflection point feature matrix, which provides input feature primitives for subsequent feature fusion steps;

[0085] Through the above calculation and threshold detection processing method based on differential operation, the time window trend sequence is transformed into a quantitative change rate feature vector and a discrete trend inflection point marker, which realizes the accurate characterization of the dynamic change characteristics of process parameters and the early capture of stage state transition signals.

[0086] For example, during the fruit juice evaporation and concentration stage, the first derivative of the central difference is calculated on the standardized temperature trend sequence (sampling frequency 10Hz, Δt = 0.1 seconds). The rate of change is calculated for temperature values ​​of 85.0, 85.4, 85.9, and 86.5, yielding a rate of increase of 2.5℃ / s at t = 0.2 seconds. Second-order difference calculation is performed, and a positive amplitude of 4.0℃ / s² is obtained at the third sampling point of the same sequence, with the sign changing from negative to positive, triggering the inflection point marker determination. The same processing is performed on the pressure trend sequence (sampling frequency 5Hz, Δt = 0.2 seconds), and an inflection point amplitude exceeding the set threshold of 0.15MPa / s² is detected at the evaporation intensity adjustment point. The final output contains a feature matrix containing four sets of rate of change features and corresponding inflection point markers, used as input for subsequent stage identification models. Verification shows that this feature extraction can increase the average lead time for stage switching detection to 1.2 seconds, helping to optimize the real-time performance and stability of model switching.

[0087] S3.3: The Fast Fourier Transform algorithm is used to perform frequency domain analysis on the time window sequence of each parameter to extract the main frequency components and their amplitude distribution, thereby identifying the periodic fluctuation pattern characteristics of the parameters;

[0088] Based on the parameter time window trend series obtained by the S3.2 difference operation, the Fast Fourier Transform (FFT) algorithm (parameter: number of transform points 22) is used. k (Hanning window weighting) enables the conversion of process parameter sequences of various dimensions from time-domain signals to frequency-domain signals in order to extract periodic component features;

[0089] Furthermore, through the spectrum amplitude calculation module (parameter: amplitude) N is the sequence length, Re and Im are the real and imaginary parts of the FFT coefficients, respectively, to realize the quantitative calculation of the amplitude of each frequency component in the frequency domain and obtain the complete amplitude distribution vector;

[0090] Furthermore, by using peak search and main frequency extraction algorithms (parameter: amplitude threshold is 1.5 times the global average amplitude), the frequency component with the largest energy contribution in the spectrum distribution is located, and the set of main frequency components is output.

[0091] Furthermore, the energy contribution rate of each major frequency component is calculated and an energy ratio vector is generated through a spectral energy normalization algorithm, so as to conduct comparative analysis of the intensity of periodic modes across different parameters;

[0092] Furthermore, through periodic pattern recognition rules (parameter: frequency resolution) T is the window duration. It is determined whether the characteristic frequency falls within the preset process characteristic frequency band range (such as the evaporation and concentration stirring impact frequency band, the feed pulsation frequency band), and the energy characteristics of the corresponding frequency band are marked to characterize the typical fluctuation mode of the current stage.

[0093] Through the above-mentioned FFT frequency domain analysis and main frequency pattern extraction processing method, the implicit periodic components of the time domain trend sequence are transformed into quantitative main frequency components, amplitude distribution and energy feature vectors, so as to achieve accurate identification and characteristic expression of periodic fluctuation patterns.

[0094] For example, during the steady-state heating phase of evaporation and concentration, a 64-point FFT is performed on a standardized temperature trend sequence of 50 points (sampling frequency 10Hz) after applying a Hanning window, yielding the real part vector Re and the imaginary part vector Im. The amplitudes A0 to A are then calculated using the formula. 32 The analysis revealed a dominant frequency of 0.5 Hz and an amplitude of 4.2 °C, with a normalized energy value of 0.65, corresponding to the cyclic disturbance frequency band of the stirring blades, characterizing the dominant influence of cyclic stirring on the temperature field in the current stage. For the flow parameter, FFT analysis yielded a dominant frequency of 0.2 Hz and an energy contribution rate of 0.48, corresponding to feed pulsation characteristics. The combined periodic fluctuation feature vector of both parameters will be used as input for the feature fusion step S3.4, effectively distinguishing between the steady-state heating stage and the frequent adjustment stage. Verification results show that the introduction of periodic features can improve the stage identification accuracy by approximately 8.7%.

[0095] S3.4: Multidimensionally integrate the rate of change characteristics, trend inflection point characteristics, and periodic fluctuation characteristics to generate a high-dimensional stage feature vector to characterize the dynamic process state of the current production stage.

[0096] Based on the rate of change feature matrix, trend inflection point marker sequence, and periodic fluctuation feature vector output by S3.2 and S3.3, a feature vector concatenation method (parameters: aligned by time index, parameter dimension order fixed as rate of change - inflection point - periodic feature) is adopted to realize the merging processing of the three types of feature data in the parameter dimension and generate a preliminary fused feature matrix.

[0097] Furthermore, by using a scale normalization algorithm (parameters: normalization range 0 to 1, using minimum-maximum normalization), the numerical range of feature components at different scales is made consistent, thus suppressing the interference of the numerical amplitude of a single type of feature on the overall feature vector weight distribution.

[0098] Furthermore, through feature weight allocation and weighted fusion algorithm (parameters: change rate feature weight 0.4, inflection point feature weight 0.3, periodic feature weight 0.3), weighting factors are applied during the merging process according to the contribution ratio of different feature types to the sensitivity of stage identification, so as to obtain a linearly weighted high-dimensional stage feature vector.

[0099] Furthermore, the Pearson correlation coefficient matrix between the fused feature components is calculated using the feature correlation suppression method (parameter: correlation threshold 0.85), and redundant features of highly correlated feature pairs are removed or combined to reduce feature redundancy and enhance the information independence of feature vectors.

[0100] Through the above-mentioned feature splicing, scale unification, weight fusion and correlation suppression processing, the three types of time series features, namely rate of change, trend inflection point and periodic fluctuation, are integrated into a single high-dimensional stage feature vector, so as to realize a comprehensive dynamic process state characterization covering dynamic change intensity, structural turning signal and periodic energy pattern.

[0101] For example, in the steady-state heating stage of fruit juice evaporation and concentration, the rate of change feature matrix includes 10-dimensional time-series rate of change components for each of the four process parameters (40 dimensions in total). The trend inflection point feature is marked by two main inflection points within the window corresponding to each parameter (4×2=8 dimensions, using 0 / 1 binary encoding). The periodic fluctuation feature consists of two components: the dominant frequency and energy ratio of temperature, pressure, flow rate, and concentration (4×2=8 dimensions). The total dimension of the three types of features is 56. These features are concatenated in the order of [rate of change → inflection point → period] to form a preliminary feature matrix. Min-max normalization is used to unify the numerical distribution to the [0,1] interval. After applying weighting factors, they are fused using the following formula:

[0102] F fusion =0.4×F rate +0.3×F turn +0.3×F freq

[0103] Among them, F rate F turn F freq These represent the normalized rate of change, inflection point, and periodic fluctuation feature vectors, respectively. After Pearson correlation calculation and redundancy compression, a 44-dimensional high-dimensional stage feature vector is finally output. Validation results in the stage identification model show that the stage classification accuracy under this fusion strategy is improved by about 12.3% compared with a single feature input, and the stability index of the feature vector across different production batches is improved by about 9.8%.

[0104] S3.5: Perform principal component analysis to reduce the dimensionality of the generated stage feature vectors, so as to remove redundant information and retain the most discriminative low-dimensional feature subset, and output the standardized feature input for the stage recognition model.

[0105] Step S4: Input the feature vector into the stage recognition model, perform lightweight classification reasoning to determine the current production stage and output the corresponding stage label. For example... Fig. 3 As shown, it specifically includes:

[0106] S4.1: The extracted production stage feature vectors are format adapted to meet the data structure requirements of the stage identification model input interface. The feature vectors include parameter change rate, trend inflection point and periodic fluctuation pattern. The processed data is used as the inference input of the stage identification model to ensure the compatibility and stability of the model inference process.

[0107] After dimensionality reduction processing by S3.5 principal component analysis, the obtained low-dimensional stage feature vectors need to be used as input and adapted to the data interface of the stage recognition model.

[0108] A data structure mapping method (parameter: input tensor shape = (B, F), where B is the batch size and F is the feature dimension) is adopted to adjust the memory layout and dimension order of the dimensionality-reduced feature vectors to match the input tensor requirements of the recognition model;

[0109] Furthermore, an algorithm for rearranging the order of feature components (parameters: a fixed sequence of rate of change components → inflection point components → periodic components) is used to achieve an ordered arrangement of the internal dimensions of the feature vectors and map them to the fixed input index structure of the model to ensure that the semantic consistency of multi-source features and the binding relationship of model weights are not destroyed.

[0110] Furthermore, a floating-point format standardization method (parameters: numerical precision of 32-bit IEEE 754 single-precision floating point, numerical range [0,1]) is used to perform element-by-element format conversion on the rearranged feature data to ensure the numerical precision and operator compatibility of the data during the inference calculation process.

[0111] Furthermore, the batch normalization preprocessing module (parameters: mean vector and variance vector are preloaded based on training set statistics) is used to perform normalization transformation on the low-dimensional feature vectors within the input batch to eliminate the differences in input distribution between batches and improve the robustness of the model during convergence.

[0112] Through the above data structure mapping, component order rearrangement, numerical format standardization and batch normalization preprocessing, the dimensionality-reduced stage feature vector is transformed into an input tensor that meets the interface requirements of the stage recognition model, thus achieving full adaptation of the input data in terms of structure, numerical value and statistical distribution.

[0113] For example, in the steady-state heating stage of the fruit juice evaporation and concentration process, the 44-dimensional reduced feature vector output by S3.5 is assembled into a matrix shape (16, 44) with a batch size B = 16. The feature component order rearrangement algorithm is called to adjust the original [rate of change → period → inflection point] order to [rate of change → inflection point → period] and map it to the model input index table. Subsequently, a numerical format conversion is performed, converting all elements to 32-bit single-precision floating-point numbers, maintaining them within the [0, 1] interval. By loading the mean vector (0.48,) and variance vector (0.12,) collected during the training phase, batch normalization is performed to obtain a uniformly distributed input tensor. After this data input stage, the inference accuracy is improved by approximately 6.5% compared to direct input without adaptation, and the processing latency is controlled within 0.35 milliseconds per batch, meeting the industrial application requirements for real-time multi-objective optimization model switching.

[0114] S4.2: Perform inference calculations on the adapted feature vectors based on a lightweight classification algorithm. The lightweight classification algorithm adopts a compressed neural network structure trained offline and performs layer-by-layer nonlinear transformations on the input features to obtain the preliminary classification confidence distribution of the current production stage.

[0115] After data format adaptation and batch normalization processing by S4.1, the resulting low-dimensional stage feature input tensor is sent to the lightweight classification inference module.

[0116] A compressed convolutional neural network structure (parameters: number of convolutional layers = 3, kernel size = (3,3), channel compression ratio = 0.5, activation function = ReLU) is adopted to achieve joint encoding of spatial and temporal patterns of input features within the local receptive field;

[0117] Furthermore, the multi-channel feature map output by the convolution is reduced in dimensionality by a global average pooling algorithm (parameters: mean along the time dimension and feature dimension) and a fixed-length global feature vector is generated to reduce the number of parameters and retain key information.

[0118] Furthermore, a two-layer fully connected neural network (parameters: number of nodes in the first layer = 64, number of nodes in the second layer = number of stage categories C, activation function = ReLU / linear) is used to perform a layer-by-layer nonlinear mapping on the global feature vector to form a high-level semantic representation;

[0119] Furthermore, a batch normalization layer (parameters: mean and variance are derived from offline training statistics) and a dropout regularization layer (parameters: dropout ratio = 0.3) are applied sequentially to each fully connected output to suppress overfitting and improve the model's generalization performance.

[0120] Furthermore, by applying a linear transformation to the output layer, the unnormalized score vectors z for each stage are obtained. c , where c represents the stage category index, which serves as a representation of the initial classification confidence distribution;

[0121] Through the above lightweight compressed neural network inference processing method, the input low-dimensional stage feature tensor is mapped to a preliminary classification confidence distribution, providing basic data for the Softmax normalization of S4.3, and realizing low-latency and high-efficiency stage recognition calculation.

[0122] For example, in the steady-state heating stage of fruit juice evaporation and concentration, the batch input tensor shape is (16, 44), which is input to a compressed convolutional layer (3×3 convolutional kernel, input channel 1, output channel 8, compressed channel 4), generating a feature map shape of (16, 20, 22, 4). After global average pooling, a global feature of (16, 4) is obtained. This feature is input to a fully connected layer with 64 nodes, with a weight matrix shape of (4, 64). The output is processed by ReLU and Dropout (0.3), and then input to a second fully connected layer with C=5 output nodes (representing 5 production stages), resulting in a stage score matrix z of shape (16, 5). c In this embodiment, the model inference latency is 0.42 milliseconds per batch, and the initial classification accuracy reaches 94.1% on the validation set, providing high-confidence input for subsequent S4.3 probability normalization and stage label determination;

[0123] S4.3: Perform normalized probability calculation on the classification confidence distribution, and use the Softmax function to perform exponential normalization on the output vector to generate probability values ​​corresponding to each production stage, which serve as the quantitative basis for stage identification results;

[0124] The unnormalized stage score vector z output by the S4.2 lightweight classification algorithm c The Softmax normalization method (parameter: temperature coefficient τ = 1.0) is used to realize the mapping transformation from multi-class scores to probability distribution;

[0125] Furthermore, the index score e corresponding to category c at each stage is calculated using the index calculation module. c The calculation formula is as follows:

[0126]

[0127] Among them, z c Let τ be the unnormalized score of stage c, and τ be the temperature adjustment coefficient used to control the smoothness of the probability distribution.

[0128] Furthermore, based on the exponential score vector e c Calculate the index score and S for all categories;

[0129] Furthermore, the normalized probability value p corresponding to stage c is calculated through the normalization operation module. c The calculation formula is as follows:

[0130]

[0131] The above probability values ​​constitute the normalized probability distribution vector P of the entire category, which serves as the quantitative basis for the stage identification results.

[0132] Furthermore, through numerical stabilization, each z is processed before the Softmax calculation. c Subtract the maximum value z in the score vector * This suppresses large value overflow and improves numerical accuracy.

[0133] By using the Softmax normalization calculation method, the unnormalized score vector output by the lightweight classification algorithm is transformed into a stage probability distribution between 0 and 1 with a sum of 1, thereby realizing a quantitative representation of the confidence level of different production stages and providing a comparable probability basis for the subsequent S4.4 stage label determination.

[0134] For example, in the steady-state heating stage of fruit juice evaporation and concentration, the unnormalized score vector output by S4.2 is [2.35, 1.70, -0.85, 0.50, -1.20]. Assuming a temperature coefficient τ = 1.0, numerical stabilization is first performed, with a maximum value of 2.35. The differences between each element and this value are [0, -0.65, -3.20, -1.85, -3.55]. The calculated index score is [1.000, 0.524, 0.041, 0.157, 0.028], and the summation S = 1.750. The normalized probability distribution is [0.571, 0.299, 0.023, 0.090, 0.016]. The results show that the confidence level is the highest in the steady-state heating stage, which is 0.571, significantly higher than other stages. This provides a clear basis for the maximum probability tag determination in S4.4. It has been verified that this normalized probability can improve the stability of tag determination by about 7.8% and reduce stage switching jitter by about 12%.

[0135] S4.4: Execute stage label determination logic based on normalized probability values, select the production stage category with the highest confidence using the maximum probability principle, and generate the corresponding stage label output as a control signal for switching multi-objective optimization sub-models;

[0136] S4.5: Perform consistency verification between the stage label and the historical stage identification results, and use the stage transition rules within the sliding window to perform stage state stability judgment. If the identification results show a sudden change trend in the time dimension, trigger the model output caching mechanism to prevent frequent switching of the optimization model caused by misidentification.

[0137] Step S5: Based on the stage label, a matching optimization model is called from a pre-built set of multi-objective optimization sub-models. This optimization model is configured with weight distributions and constraints adapted to the stage. Specifically, this includes:

[0138] S5.1: Based on the stage labels output by the stage identification model, perform an index query operation on the set of multi-objective optimization sub-models to locate the optimization sub-model that matches the current stage label;

[0139] S5.2: Perform parameter loading operation on the located optimization sub-model to obtain the multi-objective weight distribution matrix and process constraint set adapted to this production stage;

[0140] Based on the S5.1 positioning, a multi-objective optimization sub-model matching the current stage label is located. It is necessary to load the multi-objective weight distribution matrix and process constraint set customized for this stage from the sub-model.

[0141] A parameter binding parsing method is adopted (parameter: index key is stage label ID, binding mode is hash mapping) to realize the mapping and parsing between the weight matrix inside the sub-model and the constraint condition file path;

[0142] Furthermore, through a matrix deserialization algorithm (parameters: storage format is binary floating-point stream, element precision is 32-bit IEEE 754 single-precision floating-point), the multi-objective weight distribution matrix is ​​efficiently restored from persistent storage to a memory tensor, and a stage-specific weight matrix W is obtained. stage ;

[0143] Furthermore, a checksum verification method (parameter: CRC-32 polynomial is 0x04C11DB7) is used to verify the integrity of the weight matrix data stream and generate a verification pass flag to confirm that the loaded data is correct.

[0144] Furthermore, through the constraint parsing engine (parameter: constraint description language is JSON Schema 2020-12), the structured parsing of the process constraint set is realized, transforming the description containing physical constraints, equipment capacity limitations, and safety boundaries into an executable constraint rule object set C. stage ;

[0145] Furthermore, a normalized mapping method (parameters: target interval is [0,1], mapping function is minimum-maximum scaling) is used to perform a linear transformation on each element in the weight matrix to eliminate the absolute value difference between different target weighting coefficients and retain their relative proportions;

[0146] Through the above parameter binding, matrix deserialization, data verification, constraint parsing and normalization mapping processing methods, the weight distribution matrix and process constraint set in the stage matching sub-model are transformed into memory objects with valid structure and consistent values ​​that can be directly called by the optimization solution module, so as to realize the fast and accurate configuration of the optimization model during the stage switching process.

[0147] For example, in the steady-state heating stage of fruit juice evaporation and concentration, the stage label ID is 3, the corresponding weight matrix storage path is " / model / stage3 / weights.bin", and the process constraint file path is " / model / stage3 / constraints.json". The system parses the hash mapping table of stage label ID=3 to obtain the above file paths.

[0148] The matrix deserialization algorithm is used to restore "weights.bin" to a weight matrix W of shape (4,1). stage Its initial value is [0.25, 0.45, 0.20, 0.10].

[0149] Performing a CRC-32 check calculates the checksum 0x8F3A2C17, which matches the stored value, indicating that the data is complete.

[0150] After parsing "constraints.json", the constraint set C is obtained. stage It includes rules such as temperature range [85.0, 90.0] degrees Celsius, pressure limit of 1.5 MPa, and flow fluctuation range of ±5%.

[0151] After mapping the weight matrix to the [0,1] interval using the minimum-maximum scaling formula, we obtain [0.25,1.00,0.00,0.50].

[0152] The final output W stage With C stage The data is stored in memory and passed to the multi-objective optimization solver initialization module of S5.3 to achieve rapid configuration of the steady-state heating stage and efficient execution of subsequent predictive optimization.

[0153] S5.3: Input the multi-objective weight distribution matrix adapted to this stage into the multi-objective optimization solver to initialize the objective priority configuration and trade-off strategy under this stage;

[0154] The stage-specific multi-objective weight distribution matrix W output by S5.2 stage As input, a multi-objective optimization solver initialization algorithm (parameters: solution mode = weighted method, numerical precision = 1e-6, iteration limit = 5000) is used to load the target priority configuration in this stage;

[0155] Furthermore, W is processed using a weight normalization method (parameter: normalization interval [0,1], keeping the proportion unchanged). stage Each row of elements undergoes a linear mapping to generate a weight matrix W with a uniform numerical range. stage' This ensures the numerical stability of the solver under multi-objective weighting.

[0156] Furthermore, a weight vector sparsification algorithm (parameter: L1 norm constraint coefficient is 0.01-0.1, preferably 0.05) is used to reduce small weight elements to zero, thereby removing low-contribution target weights and obtaining a set of non-zero weight distributions as the weighted input for the subsequent objective function;

[0157] Furthermore, through a matrix dimension mapping algorithm (parameter: target dimension = number of multi-objective functions), the sparsified weight matrix is ​​bound one-to-one with the solver's built-in objective function indices, generating a weight-objective mapping table structure M. map .

[0158] Furthermore, based on the mapping table structure, the weight vector is embedded into the multi-objective cost function definition of the solver, forming a stage-specific weighted objective function F. stage Its form is:

[0159]

[0160] Among them, w stage,i f is the weighting coefficient for the i-th objective in this stage. i Here, n represents the corresponding objective function value, and n is the total number of objective functions.

[0161] By using the above chain-style initialization process, the stage-specific weight distribution matrix is ​​transformed into a multi-objective cost function with binding relationships, thereby enabling the multi-objective optimization solver to be loaded quickly and configured with priority at that stage.

[0162] For example, during the steady-state heating stage of fruit juice evaporation and concentration, the output W of S5.2 stage The weights are [0.25, 1.00, 0.00, 0.50], and the relative proportions remain unchanged after normalization. Sparsening is applied with an L1 norm of 0.05, and the third objective weight is set to zero, resulting in a sparse weight vector of [0.25, 1.00, 0.00, 0.50]. Matrix dimension mapping table M map Weights are indexed as follows: 0 → energy consumption target, 1 → product quality target, 3 → equipment wear target. After embedding the solver, the following is generated:

[0163] F stage =0.25f1 + 1.00f2 + 0.50f4

[0164] Where f1 is the energy consumption function, f2 is the mass function, and f4 is the wear function. After loading this weighted combination, the optimization solver can converge to the solution set within the Pareto front curve region within 5000 iterations, and the weight and target priority configuration time is controlled within 0.12 seconds, realizing the ability to dynamically balance energy consumption, mass, and wear targets during the steady-state heating stage;

[0165] S5.4: Boundary constraints are applied to the feasible solution space of the optimization sub-model based on the set of process constraints to generate a constrained optimization problem expression suitable for this stage;

[0166] S5.5: Perform a preprocessing transformation on the constrained optimization problem expression adapted to this stage to generate an optimization model interface description file suitable for the input format of the nonlinear dynamic prediction module.

[0167] Step S6: Input the current standardized parameter sequence and historical data into the invoked optimization sub-model, perform nonlinear dynamic prediction, and generate the optimal process parameter setpoints for the next time step. Specifically, this includes:

[0168] S6.1: Perform sliding window stitching on the current standardized parameter sequence and historical process data to construct an input data matrix containing time dynamic features;

[0169] S6.2: Based on the called optimized sub-model, a Long Short-Term Memory (LSTM) network is used to extract nonlinear features from the input data matrix to obtain a hidden state vector containing multi-stage dynamic characteristics;

[0170] S6.3: Input the hidden state vector into the multi-objective cost function module, calculate the weight fitness of each optimization objective based on the Pareto front theory, and generate a multi-dimensional objective priority distribution vector;

[0171] S6.4: The non-dominated sorting genetic algorithm (NSGA-II) is used to solve the multi-objective optimization problem of the multi-dimensional objective priority distribution vector and system constraints to generate multiple candidate process parameter setting schemes;

[0172] S6.5: Based on the fuzzy comprehensive evaluation method, the merits of multiple candidate process parameter setting schemes are ranked, and the optimal set value in the Pareto optimal solution set is selected as the control command output for the next moment.

[0173] Step S7: The optimal process parameter setpoint is output to the control system for dynamic adjustment of the actuator, thereby achieving production parameter adjustment under multi-objective collaborative optimization. Specifically, this includes:

[0174] S7.1: Convert the format and adapt the communication protocol of the optimal process parameter setpoints output by the nonlinear dynamic prediction module to generate control command data packets that the control system can recognize.

[0175] S7.2: Based on industrial Ethernet or fieldbus protocol, the control command data packet is transmitted to the PLC control unit to realize real-time data interaction with the actuator;

[0176] S7.3: In the PLC control unit, the received control command data packet is parsed and verified to extract setpoint parameters for various dimensions such as temperature, pressure, flow rate, and concentration.

[0177] S7.4: Based on the PID control algorithm, the extracted setpoint parameters are compared with the actual process parameters collected by the current sensor, the control deviation is calculated and the corresponding control output signal is generated;

[0178] S7.5: The control output signal is sent to the actuator to drive the regulating valve, frequency converter pump, heating element and other devices to achieve dynamic closed-loop control of various process parameters in the production of fruit concentrate.

[0179] Step S8: Collect feedback data of the actual process parameters after adjustment, calculate the achievement deviation of each optimization objective, and generate model performance evaluation indicators. Specifically, this includes:

[0180] S8.1: Real-time acquisition of actual process parameters after control execution, including temperature, pressure, flow rate, concentration and equipment status signals, to obtain process data after control system response;

[0181] S8.2: Filter and normalize the collected actual process parameter data to eliminate sensor noise and dimensional differences in order to generate a standardized sequence of actual process data.

[0182] S8.3: Based on the preset optimization objective function, compare the standardized actual process data sequence with the target set value output by the optimization sub-model to calculate the achievement deviation of each optimization objective;

[0183] For the standardized actual process data sequence after filtering and normalization, a deviation calculation method based on Euclidean distance is used to quantify the instantaneous deviation between each process parameter and the corresponding target set value.

[0184] Furthermore, by performing point-by-point difference operations, the difference between the actual value and the set value of the k-th dimension parameter under a certain optimization objective is extracted, and the instantaneous error vector at the parameter level of the optimization objective is obtained.

[0185] Furthermore, based on the root mean square deviation calculation method, an energy-based measure of the parameter deviation corresponding to each optimization objective is realized, and the comprehensive deviation amplitude of a single objective is obtained.

[0186] Furthermore, a time-weighted moving average method is adopted (parameter: the time weight decay coefficient λ is set according to the dynamic change rate of the stage) to achieve smoothing of short-term drastic fluctuations and generate a smoothing deviation index of the optimization target to adapt to the dynamic characteristics of parameter changes in the production process.

[0187] Furthermore, by matching the smoothing deviation index of each optimization objective with the normalization coefficient of its objective function, a unified error characterization under different dimensions and tolerances can be achieved.

[0188] By using the above-mentioned combined algorithm of difference-root mean square-smoothing-normalization based on the objective function, the standardized actual process data sequence is compared with the target set value output by the optimization sub-model, and transformed into a quantifiable achievement deviation reflecting the achievement of each optimization objective, providing a standardized input for subsequent multi-objective comprehensive deviation calculation;

[0189] For example, during the steady-state heating stage of evaporation and concentration, the temperature setpoint is 85.0℃, the pressure setpoint is 0.30MPa, and the flow rate setpoint is 5.5m³. 3 Using a concentration setpoint of 62.0% per hour as a reference benchmark for optimization, the actual data sequence processed by S8.2 was subjected to deviation calculations: instantaneous temperature deviation 0.3℃, instantaneous pressure deviation 0.02MPa, and instantaneous flow rate deviation 0.1m³. 3 / h, instantaneous concentration deviation 0.4%. Root mean square deviation calculated for the temperature series: Similarly, the root mean square deviations for pressure (0.02 MPa), flow rate (0.1), and concentration (0.4) were obtained. A weighted moving average with λ = 0.8 was used to suppress short-term noise. After normalization to tolerances of 5℃ (temperature), 0.05 MPa (pressure), 1.0 (flow rate), and 1.0 (concentration), the deviations for temperature (0.06), pressure (0.4), flow rate (0.1), and concentration (0.4) were obtained. These results were used as input components of the multi-objective comprehensive deviation index in subsequent section S8.4, effectively reflecting the degree of agreement between the current control effect and the model predictions.

[0190] S8.4: Weighted summation of the deviations from the achievement of each optimization objective to generate a multi-objective comprehensive deviation index, in order to quantify the prediction and control performance of the current optimization sub-model;

[0191] S8.5: Based on the multi-objective comprehensive deviation index and combined with historical model performance data, calculate the model performance evaluation index, which includes the mean prediction error, deviation volatility, and target achievement rate.

[0192] Step S9: Based on the model performance evaluation metrics, the parameter configuration and switching logic of each optimized sub-model are updated online to achieve adaptive model optimization and improve stage recognition accuracy. Specifically, this includes:

[0193] S9.1: Normalize the model performance evaluation index to eliminate the dimensional differences between different optimization objectives and generate a standardized performance deviation vector;

[0194] S9.2: Based on the standardized performance deviation vector, calculate the fitness score of each optimization sub-model at the current stage. The fitness score reflects the actual performance of the model in multi-objective collaborative optimization.

[0195] S9.3: The reinforcement learning algorithm is used to adjust the weight distribution of the optimization sub-model online, and the dynamic priority of each optimization objective is updated based on the fitness score feedback to generate an updated multi-objective trade-off strategy.

[0196] S9.4: The classification parameters of the stage identification model are fine-tuned using an incremental learning mechanism. Based on the comparison error between historical identification results and actual production stages, the accuracy and robustness of stage identification are improved.

[0197] S9.5: Based on the updated multi-objective trade-off strategy and the optimized stage identification model, the constraints and switching thresholds in the multi-objective optimization sub-model set are dynamically adjusted to generate adaptive model switching logic.

[0198] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0199] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent control of production parameters in the production process of fruit concentrate, characterized in that, Includes the following steps: S1: Collect real-time operating parameters of multiple process stages in the production of fruit concentrate, and record the timestamp and process number corresponding to each parameter; S2: Filter and normalize the collected multidimensional parameter data to generate a standardized process data sequence; S3: Based on the time series feature extraction method, identify the feature vector of the current production stage, the feature vector including parameter change rate, trend inflection point and periodic fluctuation pattern; S4: Input the feature vector into the stage recognition model, perform lightweight classification reasoning, determine the current production stage, and output the corresponding stage label; S5: Based on the stage label, call the matching optimization model from the pre-built multi-objective optimization sub-model set, wherein the optimization model is configured with weight distribution and constraints adapted to the stage; S6: Input the current standardized parameter sequence and historical data into the called optimization sub-model, perform nonlinear dynamic prediction, and generate the optimal process parameter setpoint for the next moment; S7: Output the optimal process parameter setpoint to the control system for dynamic control of the actuator; S8: Collect feedback data of actual process parameters after adjustment, calculate the achievement deviation of each optimization objective, and generate model performance evaluation indicators; S9: Based on the model performance evaluation index, update the parameter configuration and switching logic of each optimized sub-model online.

2. The intelligent control method for production parameters in the production process of fruit concentrate according to claim 1, characterized in that, Step S1 specifically includes: A multi-channel sensor network is deployed based on an industrial IoT architecture to collect data on the fruit concentrate production process and obtain the original operating parameters of each process stage. The acquired temperature signal is subjected to sliding window filtering to obtain a stable temperature measurement value; The analog signal output by the pressure transmitter is converted into digital pressure data by analog-to-digital conversion, and the dimensional differences are eliminated by Z-score normalization to generate a standardized pressure characteristic sequence. The pulse signal output by the flow meter is subjected to frequency-to-voltage conversion and linear interpolation to obtain the flow rate change curve on a continuous time scale, and the mapping relationship between flow parameters and process stages is established by combining the process number. The refractive index signal output by the concentration detector is processed by a nonlinear compensation algorithm to generate a high-precision concentration measurement value, which is then stored in a real-time database in sync with the timestamp. Edge-triggered detection processing is performed on the equipment status signals, and different operating states of the equipment are identified based on the state machine model to generate a sequence of equipment status markers; The processed temperature, pressure, flow rate, concentration, and equipment status data are timestamped and labeled with process numbers to generate a multidimensional parameter time series dataset.

3. The intelligent control method for production parameters in the production process of fruit concentrate according to claim 2, characterized in that, In step S1, the multi-channel sensor network uses the MQTT protocol for real-time data transmission, with a QoS level of 1-2 and a node heartbeat interval of 1-20 seconds. The gateway performs the conversion between Modbus-TCP / IP and industrial Ethernet protocols, with a sampling period of 0.1-2 seconds and a buffer capacity of 256-4096 frames. Hardware time synchronization based on the IEEE1588 PTP protocol is used, with a time calibration accuracy better than 1 microsecond.

4. The intelligent control method for production parameters in the production process of fruit concentrate according to claim 1, characterized in that, Step S2 specifically includes: The collected multidimensional parameter time series data is processed by sliding window segmentation. The temperature, pressure, flow rate, concentration and equipment status signals are time-series aligned in units of time window to generate structured time series data blocks. The parameters of each dimension within the time window are filtered, the signal is decomposed into multiple scales and a soft threshold is set to obtain the reconstructed signal after noise reduction. The filtered reconstructed signal is subjected to standardization processing, the mean and standard deviation of each dimension parameter within the time window are calculated, and the original signal is normalized to generate a standardized parameter sequence. The standardized parameter sequence is subjected to drift error identification. If a significant trend shift is detected, the sequence is detrended and corrected by differential processing to generate standardized data without drift bias. The standardized data without drift bias is aligned and stitched together according to the process number and timestamp to form a multi-dimensional process data matrix in a unified format. Stage labels and equipment status identifiers are added to generate a standardized input dataset.

5. The intelligent control method for production parameters in the production process of fruit concentrate according to claim 4, characterized in that, In step S2, the sliding window is segmented and time-series aligned. The window length is 1-30 seconds, the shift step is 0.1-5 seconds, the timestamp resolution is 0.1-10 milliseconds, and the interpolation method is cubic spline interpolation.

6. The intelligent control method for production parameters in the production process of fruit concentrate according to claim 1, characterized in that, Step S3 specifically includes: A sliding window segmentation process is performed on the standardized input dataset to extract the local dynamic features of each process parameter in the time dimension and generate a sequence of parameter change trends within the time window. Based on the difference operation method, the first and second derivatives of the time window trend series of temperature, pressure, flow rate and concentration parameters are calculated to obtain the rate of change and inflection point feature vector of each parameter. The Fast Fourier Transform algorithm is used to perform frequency domain analysis on the time window sequences of each parameter, extract the main frequency components and their amplitude distribution, and identify the periodic fluctuation pattern characteristics of the parameters. The rate of change and inflection point feature vector are fused with the periodic fluctuation feature in multiple dimensions to generate a high-dimensional stage feature vector. Principal component analysis is performed on the high-dimensional stage feature vector to reduce its dimensionality, and standardized feature input is output.

7. The intelligent control method for production parameters in the production process of fruit concentrate according to claim 6, characterized in that, In step S3, the time series feature extraction includes performing first-order and second-order central difference on temperature, pressure, flow rate, and concentration, with a sampling interval of 0.01-1 seconds, and identifying the change point of the rate of change sign through the trend inflection point amplitude threshold, with the threshold determined based on historical statistics.

8. The method for intelligent control of production parameters in the production process of fruit concentrate according to claim 1, characterized in that, Step S4 specifically includes: The extracted production stage feature vectors are processed for format adaptation. The feature vectors include parameter change rate, trend inflection point and periodic fluctuation pattern. Based on a lightweight classification algorithm, inference calculations are performed on the adapted feature vectors, and the input features are transformed nonlinearly layer by layer to obtain the preliminary classification confidence distribution of the current production stage. The preliminary classification confidence distribution is normalized and probability is calculated. The output vector is exponentially normalized using the Softmax function to generate probability values ​​corresponding to each production stage. The stage label determination logic is executed based on the normalized probability value. The production stage category with the highest confidence is selected by adopting the maximum probability principle, and the corresponding stage label output is generated. The consistency between the stage labels and the historical stage identification results is verified. The stage state stability is judged by the stage transition rules in the sliding window. If the identification results show a sudden change trend in the time dimension, the model output caching mechanism is triggered.

9. The intelligent control method for production parameters in the production process of fruit concentrate according to claim 1, characterized in that, Step S5 specifically includes: Based on the stage labels output by the stage identification model, an index query operation is performed on the set of multi-objective optimization sub-models to locate the optimization sub-model that matches the current stage label; Perform a parameter loading operation on the optimized sub-model to obtain a multi-objective weight distribution matrix and a set of process constraints adapted to this production stage; The multi-objective weight distribution matrix adapted to this stage is input into the multi-objective optimization solver to initialize the objective priority configuration and trade-off strategy under this stage; Based on the set of process constraints, the feasible solution space of the optimization sub-model is bounded to generate a constrained optimization problem expression adapted to this stage; The constrained optimization problem expression adapted to this stage is preprocessed and transformed to generate an optimization model interface description file.

10. The method for intelligent control of production parameters in the production process of fruit concentrate according to claim 9, characterized in that, In step S5, the weight distribution matrix of the multi-objective optimization sub-model is stored in 32-bit single-precision floating-point format, normalized to the [0,1] interval, and the sparsity threshold L1 norm is 0.01-0.1.

Images