Intelligent regulation and control method and system for solid waste utilization

Abstract

The invention relates to the technical field of data processing, and discloses an intelligent regulation and control method and system for solid waste utilization. The method comprises the following steps: deploying a sensor array at a solid waste treatment key node to acquire data and preprocessing the data to form a unified matrix; inputting CNN-LSTM, GRU and LSTM three networks for parallel analysis, and outputting category probability, process parameters and effect prediction; and a temperature, time, catalyst, pH and rotating speed comprehensive regulation and control instruction is generated by adopting a weighted fusion algorithm. The waste component identification accuracy and the resource utilization efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to an intelligent regulation and control method and system for solid waste utilization. BACKGROUND

[0002] With the acceleration of industrialization and rapid development of urbanization, the amount of solid waste is increasing, and its resource utilization has become an important way to solve environmental pollution and resource shortage. Existing solid waste utilization technologies mainly include traditional landfill, incineration treatment methods, and resource utilization technologies such as biological treatment and chemical conversion that have emerged in recent years. These technologies have achieved waste reduction and resource utilization to some extent. At the same time, with the development of sensor technology and data analysis methods, some solid waste treatment systems have begun to use single sensor monitoring and simple statistical analysis methods to monitor the treatment process, and process parameters are set by artificial experience to maintain system operation.

[0003] However, the existing technology has significant deficiencies in the solid waste utilization process. First, the traditional artificial experience regulation and control method is difficult to cope with the complex and variable characteristics of solid waste, resulting in low processing efficiency and unstable resource recovery rate. Second, the existing single sensor monitoring technology cannot accurately identify the multi-dimensional feature information of waste, especially when waste sources are diversified, the component recognition accuracy is seriously insufficient. Third, the existing simple statistical analysis method lacks deep learning ability for complex time sequence characteristics of the treatment process, and cannot predict and optimize process parameters. Finally, the existing technology lacks an effective multi-source information fusion mechanism, and cannot realize intelligent integrated decision-making of different analysis results. SUMMARY

[0004] The present application provides an intelligent regulation and control method and system for solid waste utilization, which solves the technical problems of difficult effective fusion of multi-source heterogeneous data, insufficient intelligent analysis accuracy, and lack of self-adaptive regulation and control mechanism in the solid waste utilization process, and improves the waste component recognition accuracy and resource utilization efficiency.

[0005] In a first aspect, the present application provides an intelligent regulation and control method for solid waste utilization, which comprises:

[0006] Deploying a multi-type sensor array at a key process node of a solid waste treatment system to obtain real-time monitoring data of solid waste, and preprocessing the real-time monitoring data to form a unified data matrix;

[0007] The preprocessed unified data matrix is input into a multi-model integrated network for intelligent analysis, the multi-model integrated network comprising a trained CNN-LSTM fusion network, a trained GRU network and a trained LSTM dynamic modeling network, the CNN-LSTM fusion network being trained by historical waste data to output a waste type probability distribution, the GRU network being trained by historical process parameter data to output optimal process parameters, and the LSTM dynamic modeling network being trained by historical processing effect data to output processing effect prediction indexes;

[0008] The output results of the trained CNN-LSTM fusion network, the trained GRU network and the trained LSTM dynamic modeling network are integrated by using a weighted fusion voting decision algorithm to generate comprehensive regulation and control instructions and perform specific regulation and control actions through an adaptive feedback control system, thereby realizing intelligent regulation and control of the solid waste treatment process.

[0009] In a second aspect, the present application provides an intelligent regulation and control system for solid waste utilization, comprising:

[0010] A multi-source perception module is arranged at key process nodes of a solid waste treatment system to obtain real-time monitoring data of the solid waste, and the real-time monitoring data is preprocessed to form a unified data matrix;

[0011] A fusion decision module is configured to input the preprocessed unified data matrix into a multi-model integrated network for intelligent analysis, the multi-model integrated network comprising a trained CNN-LSTM fusion network, a trained GRU network and a trained LSTM dynamic modeling network, the CNN-LSTM fusion network being trained by historical waste data to output a waste type probability distribution, the GRU network being trained by historical process parameter data to output optimal process parameters, and the LSTM dynamic modeling network being trained by historical processing effect data to output processing effect prediction indexes;

[0012] A closed-loop execution module is configured to integrate the output results of the trained CNN-LSTM fusion network, the trained GRU network and the trained LSTM dynamic modeling network by using a weighted fusion voting decision algorithm to generate comprehensive regulation and control instructions and perform specific regulation and control actions through an adaptive feedback control system, thereby realizing intelligent regulation and control of the solid waste treatment process.

[0013] In a third aspect, an intelligent regulation device for solid waste utilization is provided, comprising a memory and at least one processor, the memory storing instructions; the at least one processor invoking the instructions in the memory to cause the intelligent regulation device for solid waste utilization to perform the intelligent regulation method for solid waste utilization described above.

[0014] In a fourth aspect, a computer-readable storage medium is provided, the computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the intelligent regulation method for solid waste utilization described above.

[0015] In the technical solutions provided in the present application, by deploying multiple types of sensor arrays at key process nodes of the solid waste treatment system, comprehensive collection of multi-dimensional feature information of the waste is achieved, which significantly improves the completeness and accuracy of data acquisition compared to the single sensor monitoring method of the prior art. Through sliding window average noise filtering, Z-score standardization processing and timestamp synchronization technology processing of real-time monitoring data, a unified data matrix is formed, effectively solving the technical problem of multi-source heterogeneous data fusion and laying a data foundation for subsequent intelligent analysis. The preprocessed unified data matrix is input into a multi-model integrated network including a CNN-LSTM fusion network, a GRU network and an LSTM dynamic modeling network for intelligent analysis, which fully utilizes the advantages of different deep learning models. The CNN-LSTM fusion network extracts spatial features through a convolution layer and captures time series rules through an LSTM layer, achieving accurate output of waste type probability distribution; the GRU network effectively handles the gradient vanishing problem in process parameter optimization through a gating mechanism, and the output optimal process parameters are more stable and reliable; the LSTM dynamic modeling network accurately predicts the treatment effect index through a long-term memory mechanism, providing forward-looking guidance for regulation and decision-making. The output results of the three networks are integrated by using a weighted fusion voting decision algorithm, and through dynamic weight allocation and integrated decision processing, the comprehensive regulation instruction generated considers multi-dimensional information of component identification, parameter optimization and effect prediction, significantly improving the accuracy and reliability of the decision.

[0016] The multi-model integration algorithm adopted by the present application embodies unique technical contributions. The algorithm characteristics of the CNN-LSTM fusion network enable it to simultaneously process the spatial distribution characteristics and time series changes of sensor data, which is of great significance for identifying solid waste with complex composition and dynamic changes over time. The combination of the local receptive field characteristics of the convolutional layer and the long-term dependency learning ability of LSTM enables deep mining of multi-level features of waste. The gating algorithm characteristics of the GRU network exhibit good computational efficiency and convergence stability when processing process parameter optimization. The design of its update gate and reset gate enables the network to adaptively select and forget historical information, making it particularly suitable for processing scenarios where process conditions frequently change during solid waste treatment. The memory cell algorithm characteristics of the LSTM dynamic modeling network can effectively capture the long-term dependencies of the treatment process, providing a significant advantage for predicting treatment effectiveness. The weighted fusion voting decision algorithm can automatically optimize the decision weights according to the performance of each model under different processing conditions through a dynamic weight adjustment mechanism. This adaptive feature enables the entire system to continuously learn and improve. Through the adaptive feedback control system, specific control actions are executed, realizing a complete closed loop from data acquisition, intelligent analysis to control execution, ensuring that the solid waste treatment process is always in an optimal state, and fundamentally solving the problems of regulation lag and insufficient precision in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor based on these drawings.

[0018] Figure 1 An embodiment schematic diagram of the intelligent regulation and control method for solid waste utilization in the embodiments of the present application;

[0019] Figure 2 An adaptive process parameter optimization flowchart based on the GRU network and effect feedback in the embodiments of the present application;

[0020] Figure 3 An embodiment schematic diagram of the intelligent regulation and control system for solid waste utilization in the embodiments of the present application;

[0021] Figure 4 A structural schematic block diagram of the intelligent regulation and control device for solid waste utilization in the embodiments of the present application. DETAILED DESCRIPTION

[0022] The embodiments of the present application provide an intelligent regulation method and system for solid waste utilization. The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" or "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0023] For ease of understanding, the specific processes of the embodiments of the present application are described below, please refer to Figure 1 One embodiment of the intelligent regulation method for solid waste utilization in the embodiments of the present application includes:

[0024] S1, deploying a multi-type sensor array at a key process node of a solid waste treatment system, acquiring real-time monitoring data of the solid waste, and pre-processing the real-time monitoring data to form a unified data matrix.

[0025] Specifically, a multi-type sensor array is deployed at a key process node of a solid waste treatment system, the spectral characteristics of organic matter, cellulose and lignin are collected by a near-infrared spectrum sensor, the concentration distribution of heavy metal elements such as lead, mercury, cadmium and chromium is detected by an X-ray fluorescence sensor, the mass change curve in different temperature intervals is monitored by a thermogravimetric analysis sensor, and the environmental conditions are recorded in real time by a gas chromatograph, a temperature and humidity sensor and a pressure sensor, so as to form multi-dimensional real-time monitoring data; after data collection is completed, the original signal is first subjected to sliding window average filtering to remove transient noise, then standardized processing is realized by subtracting the historical mean value and dividing by the standard deviation, thereby eliminating the dimension difference of different sensors, aligning and synchronously resampling according to the time stamp, to ensure the consistency and comparability of data from various sensors in time series, and obtaining a unified data matrix with uniform structure, dimension and direct input deep learning model, providing a data basis for solid waste component identification, process parameter prediction and treatment effect modeling.

[0026] S2, input the pretreated unified data matrix into a multi-model integrated network for intelligent analysis, the multi-model integrated network including a trained CNN-LSTM fusion network, a trained GRU network and a trained LSTM dynamic modeling network, the CNN-LSTM fusion network being trained by historical waste data for outputting a waste type probability distribution, the GRU network being trained by historical process parameter data for outputting optimal process parameters, and the LSTM dynamic modeling network being trained by historical processing effect data for outputting processing effect prediction indexes;

[0027] Specifically, the multi-model integrated network needs to be trained offline before actual application. The training process uses a labeled data set accumulated by a historical solid waste treatment system, which contains unified data matrices obtained after sliding average filtering, standardization and resampling steps on historical sensor data of multiple batches of waste, and corresponding waste type labels, actually executed process parameters and final processing effect indexes.

[0028] The CNN-LSTM fusion network uses historical waste data as training samples, inputs unified data matrices obtained after preprocessing of sensor data such as spectral features, element composition and thermogravimetric characteristics of different types of waste, and their corresponding type labels (such as high-organic waste, mixed metal waste, plastic waste, etc.) into the network, and trains the network parameters through supervised learning, so that it can accurately identify the type of waste and output the probability distribution of each type according to the input pretreated unified data matrix;

[0029] The GRU network is trained using historical process parameter data, inputs unified data matrices obtained after preprocessing of sensor data of different batches of waste, combined with running records under different process conditions (such as temperature, time, catalyst dosage, pH value, rotation speed, etc. Parameter set value) and corresponding processing effect evaluation into the network, and learns the mapping relationship between unified data matrix and historical optimal parameter combination, so that it can predict the optimal process parameters for the current input unified data matrix;

[0030] The LSTM dynamic modeling network is trained using historical treatment effect data, and a unified data matrix obtained by preprocessing sensor data of historical waste is input into the network in combination with actual treatment effect measurement values (such as gas production rate, pollutant removal rate, energy consumption, residue characteristics and other indicators) under different process conditions, so that the network can predict the treatment effect that can be achieved under the current input unified data matrix and process conditions. After the three networks are respectively trained to converge to their respective loss functions, they are integrated and deployed to an actual control system. In actual operation, the unified data matrix obtained by preprocessing is input into the trained multi-model integrated network for joint analysis. The CNN-LSTM fusion network extracts the spatial local features of the sensor data in the unified data matrix through the convolution layer, and combines the LSTM structure to model the time series features in depth, so as to output the type probability distribution of solid waste; the GRU network effectively captures the short-term and medium and long-term dependence relationship in the change of process parameters by using its gating structure, avoids the gradient disappearance problem and improves the calculation efficiency, and then generates the prediction result of the optimal process parameters; the LSTM dynamic modeling network models the dynamic characteristics of long time series through the memory unit, and accurately predicts the key effect indicators of the waste treatment process, such as gas production rate, pollutant removal rate or resource utilization efficiency. The three types of networks play their respective advantages in parallel operation, which not only ensures the accuracy of waste component identification, but also improves the stability of process parameter optimization.

[0031] S3, the output results of the trained CNN-LSTM fusion network, the trained GRU network and the trained LSTM dynamic modeling network are integrated by using a weighted fusion voting decision algorithm, the outputs of each network are respectively mapped to five control variable dimensions of temperature, time, catalyst dosage, pH adjustment amount and speed, and a weighted sum calculation is performed on each control variable dimension according to a dynamic weight coefficient to generate a comprehensive control instruction containing temperature adjustment amount, time adjustment amount, catalyst dosage adjustment amount, pH adjustment amount and speed adjustment amount.

[0032] Specifically, the historical prediction accuracy of the CNN-LSTM fusion network, the GRU network and the LSTM dynamic modeling network in the rolling validation window is counted, and the accuracy of each network is divided by the sum of the accuracies of the three networks to obtain the dynamic weight coefficient. For example, if the accuracy of CNN-LSTM is 0.88, the accuracy of GRU is 0.90, and the accuracy of LSTM is 0.85, the total accuracy is 2.63, and the corresponding dynamic weight coefficients are =0.88 / 2.63=0.335, =0.90 / 2.63=0.342, =0.85 / 2.63=0.323.

[0033] After obtaining the dynamic weight coefficients, the system maps the outputs of the three networks to a unified control variable representation. Specifically, the waste type probability distribution output by the CNN-LSTM is converted to recommended control parameters for the type of waste through a pre-set type-process parameter mapping table. For example, when a high-organic waste is identified, the mapping table gives a temperature recommendation of 600-650°C, a time recommendation of 40-45 min, a catalyst dosage recommendation of 3.0-3.5 g / L, a pH recommendation of 7.0-7.5, and a rotation speed recommendation of 110-130 rpm. The optimal process parameters output by the GRU network directly correspond to specific numerical recommendations for temperature, time, catalyst dosage, pH adjustment, and rotation speed, such as temperature 620°C, time 42 min, catalyst 3.2 g / L, pH 7.3, and rotation speed 120 rpm. The processing effect prediction index (e.g., predicted gas production rate is 5% lower) output by the LSTM network is converted to incremental adjustment recommendations for each control variable through an effect-parameter correction table, such as temperature +10°C, time +2 min, catalyst +0.3 g / L, pH +0.2, and rotation speed +5 rpm. For each control variable (e.g., temperature), the system sums the output values for that control variable from the three networks (denoted as ) multiplied by the corresponding dynamic weight coefficients ( ) to obtain the final adjustment amount for that control variable: temperature adjustment amount = . For example, if = 620°C, = 615°C, = 625°C, then the temperature adjustment amount = 0.335 x 620 + 0.342 x 615 + 0.323 x 625 = 619.6°C. Similarly, the time adjustment amount, catalyst dosage adjustment amount, pH adjustment amount, and rotation speed adjustment amount are calculated to obtain the target set values for the five control variables.

[0034] After generating the five adjustment amounts, each adjustment amount is compared with the current actual operating parameters to calculate the deviation amplitude of each control parameter. For example, if the temperature adjustment amount is 619.6°C and the current operating temperature is 600°C, the temperature deviation amplitude is 19.6°C. The system presets the influence weight coefficients of each control variable on the treatment effect, which represent the influence degree of each control variable on the treatment effect (gas production rate, pollutant removal rate, energy consumption, etc. comprehensive indicators) and are determined through correlation analysis of historical data or expert experience. For example, temperature has the greatest influence on gas production rate, with an influence weight coefficient of 0.35; catalyst dosage is second, with a weight coefficient of 0.30; time weight coefficient is 0.20; pH adjustment amount weight coefficient is 0.10; and speed weight coefficient is 0.05. The calculation of the urgency score uses a weighted sum method, i.e. multiplying the deviation amplitude of each parameter by the corresponding influence weight coefficient. Taking temperature as an example: temperature urgency score = 19.6°C x 0.35 = 6.86. After calculating the urgency scores of the five parameters, the execution priority is determined by sorting the scores from high to low. After sorting, the system encapsulates the sorting results, adjustment amounts, and execution priority into a comprehensive control instruction.

[0035] The data structure of the instruction includes: (1) control parameter name (temperature, time, catalyst dosage, pH adjustment amount, speed); (2) corresponding target set value (619.6°C, 42min, 3.2g / L, 7.3, 120rpm); (3) adjustment step (i.e. the increment from the current value to the target value, such as +19.6°C from 600°C to 619.6°C); (4) execution priority order (1, 2, 3, 4, 5).

[0036] After the comprehensive control instruction is issued to the execution module, the controller converts the instruction into device driving signals. Temperature adjustment is achieved by controlling the heating device power, time adjustment is achieved by setting the reaction residence time, catalyst dosage and pH adjustment amount are achieved by controlling the flow of the dosing device, which includes the flow of the catalyst dosing pump and the flow of the pH regulator dosing pump, and speed adjustment is achieved by controlling the stirring device speed. The system uses PID control algorithm or model predictive control algorithm to adjust the heating device power, stirring device speed, and flow of catalyst dosing pump and pH regulator dosing pump in real time.

[0037] In the process of executing the regulation action, the system continuously collects actual feedback values (including actual temperature, actual treatment time, actual catalyst dosage, actual pH value and actual rotating speed) and compares them with the target set value in the comprehensive regulation instruction to calculate the effect deviation. When the effect deviation exceeds the preset threshold (for example, the actual temperature deviates from the target temperature by more than 5°C), or when abnormal conditions such as sensor drift and sudden change of feed components are detected, the system triggers the weight adaptive updating mechanism, recalculates the dynamic weight coefficients of the three networks, and performs secondary weighted fusion calculation to generate a comprehensive regulation instruction.

[0038] In a specific embodiment, the process of pre-processing the real-time monitoring data to form a unified data matrix can specifically include the following steps:

[0039] (1) For each data point in the real-time monitoring data, take the arithmetic mean of the previous 7 data points, the data point itself, and the next 7 data points, a total of 15 data points, to obtain the filtered data point. All filtered data points form the denoised sensor data;

[0040] (2) Subtract the mean of the historical data of the sensor from each value in the denoised sensor data, and then divide by the standard deviation of the historical data of the sensor to obtain standardized data;

[0041] (3) Sort and align the standardized data from different sensors according to the time stamp, resample and interpolate at an interval of 100ms to obtain a unified data matrix.

[0042] Specifically, in order to convert multi-source heterogeneous sensor raw signals into an input form that can be uniformly processed by a deep learning model, the real-time monitoring data is smoothed and denoised, that is, for each sampling point, the arithmetic mean of the values of the 15 adjacent points before and after it is selected to weaken the interference of random noise on the signal and obtain a more stable trend sequence, thereby forming denoised sensor data; subtract the mean of the long-term historical data of each sensor from the sampling value, and divide by the standard deviation of the historical data to eliminate the differences in dimension and scale between different sensors, achieve consistency and comparability of cross-source data, and obtain standardized results; to ensure the synchronization and continuity of multi-type sensor signals in the time dimension, strictly align the standardized data of various types according to the acquisition time stamp, and perform resampling and interpolation at a uniform interval of 100ms, so that data of different sampling frequencies can be integrated into the same time sequence framework, obtaining a unified data matrix that is complete in dimension, uniform in format and consistent in time sequence, providing data input for convolution-recurrent neural network modeling and intelligent regulation decision-making.

[0043] For example, during the operation of a certain waste co-processing plant, the system scans and detects the raw materials at the inlet through a near-infrared spectrum sensor, and the obtained spectrum curve shows that the organic matter content is about 48%, the cellulose content is about 32%, the lignin content is about 15%, and a small amount of other impurities are accompanied; Subsequently, the X-ray fluorescence sensor detects that the concentrations of lead and cadmium in the batch of materials are 1.8 ppm and 0.9 ppm, respectively, both of which are within the controllable range; The thermogravimetric analysis sensor monitors the mass loss curve of the materials during the temperature rise from 25°C to 800°C, which presents three obvious steps, indicating that the pyrolysis characteristics of the materials exist in multiple stages of reaction; At the same time, the gas chromatograph and environmental sensor record that the VOC concentration in the environment of the plant is 12 mg / m³, the temperature is maintained at 60°C, the humidity is about 55%, and the system pressure is stable at 1.02 atm. After entering the pretreatment link, all these data are first filtered by a moving average filter to obtain a smoothed signal, then standardized by Z-score and uniformly resampled to a time interval of 100 ms, thereby forming a unified data matrix containing multi-dimensional characteristics of spectrum, elements, thermogravimetric analysis and environment; After the matrix is input into the multi-model integrated network, the CNN-LSTM sub-model outputs that the batch of raw materials belongs to "mixed organic waste" with a probability of 92%, the GRU sub-model predicts the optimal process parameters as heating temperature 720°C, holding time 42 minutes, catalyst dosage 3.5 g / L, and stirring speed 120 rpm, and the LSTM sub-model predicts the treatment effect as an increase in gas production rate by 8.7% and a decrease in heavy metal leaching toxicity in residues by 15%; After the outputs of the three models are weighted and fused to generate comprehensive control instructions, the instructions are issued to the execution module to automatically adjust the heating, stirring and dosing equipment, and in the feedback link, the target indicators are compared with the actual execution results in real time, and when it is detected that the gas production rate deviates from the target by more than 2%, the system will trigger recalculation and correct the control signal.

[0044] In a specific embodiment, the process of acquiring real-time monitoring data of solid waste can specifically include the following steps:

[0045] (1) The near-infrared spectrum sensor is used to scan and detect the spectrum of the solid waste, and the wavelength range is 1000-2500 nm. The organic matter content, cellulose content and lignin content data are obtained;

[0046] (2) The X-ray fluorescence sensor is used to detect the elemental composition of the solid waste, and the concentration distribution of heavy metal elements is identified; The thermogravimetric analysis sensor is used to detect the thermal stability of the solid waste, and the mass change characteristic curve is monitored in the temperature range of 25-1000°C. The mass change characteristic parameters are extracted from the mass change characteristic curve, including the initial mass loss temperature, the temperature corresponding to the maximum mass loss rate, the mass loss rate in a specific temperature range, and the final residual mass percentage;

[0047] (3) Real-time monitoring of the processing environment by gas chromatographic sensors, temperature sensors, humidity sensors, and pressure sensors to obtain volatile organic compound concentrations, environmental temperature and humidity, and system pressure parameters;

[0048] (4) The obtained organic matter content, cellulose content, lignin content, heavy metal element concentration distribution, mass change characteristic parameters, volatile organic compound concentration, environmental temperature and humidity, and system pressure parameters together constitute real-time monitoring data.

[0049] Specifically, in practical application, the system uses a near-infrared spectrum sensor to quickly scan the solid waste entering the processing device, with a wavelength range of 1000-2500 nm. From the spectrum curve, it can be determined that the organic matter content of the batch of material is about 50%, the cellulose content is about 30%, and the lignin content is about 12%, providing original ingredient information for the process. An X-ray fluorescence sensor detects the elements of the sample, and the concentration distribution of heavy metal elements is specifically the concentration distribution of lead, mercury, cadmium, and chromium. The results show that the concentrations of lead, mercury, cadmium, and chromium heavy metal elements are identified, among which the concentration of lead is 1.2 ppm and the concentration of cadmium is 0.8 ppm. The detection accuracy is controlled at the order of 1 ppm, ensuring sensitive monitoring of harmful ingredients. Next, a thermogravimetric analysis sensor is used to continuously heat the sample in the temperature range of 25-1000°C. The weight loss curve of the sample is recorded and divided into three stages. The first stage is within 200°C, with a weight loss of about 8%, mainly corresponding to water evaporation. The second stage is at 350-550°C, with a weight loss of about 40%, corresponding to the decomposition of organic ingredients. The third stage is the slow decomposition process after 600°C, indicating the thermal stability and combustible characteristics of the waste. In order to integrate the thermogravimetric curve data into the real-time monitoring data matrix, the system extracts features from the curve, including: initial weight loss temperature, which is the temperature at which 5% weight loss occurs, about 120°C in this example; maximum weight loss rate corresponding temperature, which is the temperature point at which the weight loss rate is the fastest, about 420°C in this example, corresponding to the main organic matter decomposition stage; specific temperature interval weight loss rate, including 100-200°C interval weight loss rate of about 8% (indicating water content), 200-600°C interval weight loss rate of about 40% (indicating volatile matter and organic matter content); final residual mass percentage, which is the percentage of residual mass at 800°C relative to the initial mass, about 32% in this example (indicating ash and fixed carbon content). These four types of characteristic parameters can effectively represent the thermal decomposition behavior of the batch of waste and be input into the unified data matrix as fixed-dimensional numerical values. In addition, a gas chromatography sensor and environmental temperature, humidity, and pressure sensors simultaneously monitor the reaction environment in real time, obtaining a volatile organic compound concentration of 11 mg / m³, a temperature of 58°C, a humidity of 52%, and a system pressure of 1.03 atm. The organic matter, cellulose, and lignin contents obtained by the near-infrared spectrum sensor, the heavy metal element concentrations detected by the X-ray fluorescence sensor, the mass change characteristic parameters extracted by the thermogravimetric analysis sensor, and the volatile organic compound concentration, temperature, humidity, and pressure parameters obtained by the gas chromatography and environmental sensors constitute the multi-dimensional real-time monitoring data.

[0050] In a specific embodiment, the process of outputting the waste type probability distribution by the CNN-LSTM fusion network can specifically include the following steps:

[0051] (1) The input unified data matrix is processed by three one-dimensional convolution layers for spatial feature extraction. Each convolution layer uses a 3x1 convolution kernel with a step size of 1. The convolution output is sequentially processed by batch normalization and ReLU activation function;

[0052] (2) The convolution features are processed by a max pooling layer for feature dimension reduction. The pooling window size is 2x1, and the maximum feature value in each pooling region is retained;

[0053] (3) The pooled feature sequence is processed by a bidirectional LSTM layer for time series modeling. The bidirectional LSTM layer contains 128 hidden units, and the dropout ratio is set to 0.3 to prevent overfitting;

[0054] (4) The LSTM output is processed by a softmax activation function for probability normalization. The output is the probability distribution of 12 main solid waste types.

[0055] Specifically, the unified data matrix is sequentially sent to a spatial feature extraction module composed of three one-dimensional convolution layers. Each layer has a convolution kernel size of 3x1 and a step size of 1. Batch normalization and ReLU activation are performed after each convolution to stabilize training and enhance nonlinear representation. The convolution output is reduced in dimension by max pooling with a size of 2x1, and the dominant response in each pooling region is retained to suppress redundancy and noise. A bidirectional LSTM is introduced for time series modeling on the obtained compact feature sequence. The number of hidden units is set to 128, and a dropout of 0.3 is applied to reduce the risk of overfitting, thereby simultaneously capturing context dependence and long-term dynamics. The output of the bidirectional LSTM is connected to a classification layer with softmax to complete probability normalization, outputting a class probability distribution covering 12 main solid waste types to characterize the confidence of the current batch of samples in each type, providing high-recognizability component prior for process parameter prediction and fusion decision.

[0056] For example, in a waste treatment plant, the system scanned a new batch of solid waste using a near-infrared spectral sensor, obtaining spectral data showing that the material contained 52% organic matter, 28% cellulose, and 15% lignin, with low impurity levels. Next, an X-ray fluorescence sensor performed elemental analysis on the batch of waste, revealing a lead concentration of 1.5 ppm and a cadmium concentration of 0.5 ppm, both within safe limits. A thermogravimetric analysis sensor recorded the thermogravimetric curve of the batch of waste. The first weight loss occurred around 200°C, with a weight loss rate of 10%, mainly caused by moisture evaporation. The second weight loss occurred between 350°C and 600°C, with a weight loss rate of 35%, representing the decomposition of organic matter. To further monitor the treatment environment, a gas chromatography sensor detected a volatile organic compound concentration of 10 mg / m³, a temperature of 60°C, a humidity of 50%, and a pressure of 1.01 atm in the area. These data collectively constituted the real-time monitoring data.

[0057] After preprocessing, this data is converted into a unified data matrix. Following processing by a three-layer convolutional neural network, the convolutional layers extract the spatial features of the materials and perform dimensionality reduction. Using a bidirectional LSTM network, the system captures the dynamic changes in the time series and calculates the probability distribution of each waste type using the softmax activation function. The system predicts that the batch of waste has an 85% probability of being "mixed organic waste," a 10% probability of being "light metal waste," and low probabilities for other types. Based on this result, the system uses a weighted fusion algorithm to generate comprehensive control commands from the predicted waste type, optimal treatment process, and treatment effect. An adaptive feedback control system adjusts the equipment, monitors and corrects various operating parameters in real time, ensuring the entire solid waste treatment process is stable and efficient.

[0058] In one specific embodiment, the process of adjusting the GRU network based on the predicted performance metrics and re-predicting the optimal process parameters may specifically include the following steps:

[0059] (1) Calculate the ratio of the deviation between the predicted treatment effect index and the benchmark effect index to obtain the adjustment factor;

[0060] (2) The learning rate adjustment value of the trained GRU network is calculated based on the adjustment factor. If it is less than the minimum learning rate value, the minimum learning rate value is taken as the learning rate value of the trained GRU network.

[0061] (3) Update the trained GRU network with the adjusted learning rate and use the updated trained GRU network to re-predict the time series of the optimal process parameters.

[0062] Specifically, after obtaining the processing effect prediction index of the current batch, the deviation ratio between the prediction index and the preset reference effect index is calculated as an adjustment factor for adaptive adjustment of the learning rate; then the learning rate of the GRU network is updated according to the adjustment factor, and when the calculated learning rate is lower than the minimum learning rate threshold allowed by the system, the minimum learning rate is taken as the actual effective value to avoid training stagnation or numerical instability; the updated learning rate is used for online fine-tuning training of the GRU network, so that the parameters of the network can quickly converge to the latest working condition and data distribution, and after fine-tuning, the time series of the optimal process parameters are re-predicted by using the corrected GRU network, so as to dynamically correct the target trajectories of temperature, time, catalyst dosage, pH adjustment and speed, etc., so that the control strategy can respond to the changes of the processing effect in time, and realize continuous adaptive optimization and stable operation.

[0063] For example, in a pyrolysis process of organic waste, the system compares the prediction result of the current batch with the preset reference index, finds that there is a certain deviation, and calculates a deviation ratio as an adjustment factor; the system corrects the learning rate of the GRU network according to the adjustment factor, and calculates the deviation ratio of the prediction index of the current batch and the reference effect index, which reflects the deviation degree of the prediction result and the actual effect. For example, if the reference gas production rate is 82 cubic meters per ton, and the current predicted gas production rate is 78 cubic meters per ton, the deviation ratio is (82-78) / 82=0.049, i.e. the deviation is about 4.9%. The deviation ratio is used as an adjustment factor to correct the learning rate of the GRU network according to a certain mapping relationship. The correction rule is: when the deviation ratio is large, it means that the model prediction error is large, and the learning speed needs to be accelerated, so the learning rate is increased; when the deviation ratio is small, it means that the model has approached the optimal state, and the learning rate should be reduced to avoid overshoot. The correction formula is: new learning rate= current learning rate x (1+ a x deviation ratio), where a is the adjustment coefficient, usually taking a value of 0.5 to 2.0. For example, if the current learning rate is 0.001 and the deviation ratio is 0.049, and the adjustment coefficient a is 1.0, then the new learning rate is 0.001 x (1+1.0 x 0.049)=0.00105. In order to avoid the learning rate being too large to cause unstable training or too small to cause slow convergence, the system sets the upper and lower limits of the learning rate: the maximum learning rate is usually set to 5 times the initial learning rate, and the minimum learning rate is set to 0.1 times the initial learning rate. If the calculated new learning rate exceeds the maximum learning rate, the maximum learning rate value is taken; if it is lower than the minimum learning rate, the minimum learning rate value is taken. The corrected learning rate is used for online fine-tuning training of the GRU network, and the latest working condition data is used to update the network parameters, so that the network can quickly adapt to the changes of the characteristics of the waste and the processing conditions, thereby improving the accuracy and robustness of the prediction of the optimal process parameters.

[0064] When the modified learning rate is lower than the preset minimum threshold, the system automatically adopts the minimum learning rate to ensure stable update of the model; after completing the adaptive adjustment of the learning rate, the GRU network is retrained and quickly adapts to new working condition characteristics, and then the process parameter sequence is predicted again, and the generated parameter result is fed back to the control module for execution, so that the heating, stirring and dosing operations are more accurate, thereby effectively reducing the gap between the prediction effect and the target effect, and realizing the continuous optimization and steady-state operation of the waste treatment process. Reference Figure 2 The figure shows an adaptive process parameter optimization process based on a GRU network and effect feedback.

[0065] In a specific embodiment, the process of generating a comprehensive control instruction by integrating the output results of the CNN-LSTM fusion network, the GRU network and the LSTM dynamic modeling network using a weighted fusion voting decision algorithm can specifically include the following steps:

[0066] (1) The historical prediction accuracy of the trained CNN-LSTM fusion network, the trained GRU network and the trained LSTM dynamic modeling network is calculated and processed to obtain dynamic weight coefficients by dividing the accuracy of each network by the sum of the total accuracy;

[0067] (2) The waste type probability distribution, the optimal process parameter and the treatment effect prediction index are respectively mapped to the temperature, time, catalyst dosage, pH adjustment amount and speed five control variable dimensions, and each control variable dimension is respectively subjected to weighted fusion processing, the mapping values of the control variable from the three networks are multiplied by the corresponding dynamic weight coefficients and summed to generate the adjustment amount of the control variable, thereby obtaining the temperature adjustment amount, the time adjustment amount, the catalyst dosage adjustment amount, the pH adjustment agent dosage adjustment amount and the speed adjustment amount;

[0068] (3) The temperature adjustment amount, the time adjustment amount, the catalyst dosage adjustment amount, the pH adjustment agent dosage adjustment amount and the speed adjustment amount are compared with the current actual running temperature, time, catalyst dosage, pH value and speed, respectively, and the deviation amplitude of each control parameter is calculated, wherein the deviation amplitude is the absolute value of the adjustment amount;

[0069] (4) The urgency score is calculated according to the deviation amplitude and the preset influence weight coefficient, and the urgency score is equal to the product of the deviation amplitude and the influence weight coefficient, and the influence weight coefficient represents the influence degree of each control variable on the treatment effect;

[0070] (5) The temperature, time, catalyst dosage, pH adjustment amount and speed five control parameters are sorted in descending order of the urgency score, and the sorted control parameters, the corresponding adjustment amounts and the execution priority order constitute a comprehensive control instruction.

[0071] Specifically, in the fusion decision process, the historical prediction accuracy of the CNN-LSTM fusion network, the GRU network and the LSTM dynamic modeling network on the recent samples is normalized to obtain dynamic weight coefficients, and the calculation method can be expressed as: ; Among them, represents the dynamic weight coefficient of the i-th model; represents the prediction accuracy of the i-th model in the validation sample; is the sum of the accuracy of the three models, which is used to ensure that the normalization result satisfies the constraint that the weight sum is 1. Through this calculation method, the reliability of each model under different working conditions can be dynamically reflected, and reasonable weights can be allocated accordingly.

[0072] After obtaining the weight coefficients, the output results of each model are weighted and fused to generate a comprehensive control parameter vector, and the calculation formula is: ; Among them, represents the comprehensive control parameter vector, which includes temperature adjustment amount, time adjustment amount, catalyst dosage, pH adjustment agent dosage and speed adjustment amount, etc. represents the control suggestion result output by the i-th model; represents the product of the model output and its dynamic weight, reflecting the contribution of the model in the comprehensive decision. Through the above weighted fusion calculation, the advantages of different networks can be complementary, and the uncertainty caused by the prediction deviation of a single model can be avoided, so that more stable and accurate comprehensive control instructions can be generated.

[0073] It should be noted that U and are not single values, but vectors containing multiple control variables. U represents the final generated comprehensive control parameter set, including temperature adjustment amount, time adjustment amount, catalyst dosage, pH adjustment agent dosage and speed adjustment amount, which are five specific numerical values. represents the control parameter suggestion set output by the i-th model, which also contains the suggested values of the five control variables. Specifically, after mapping processing, each network will give a suggested value for each of the five control variables. For example, the CNN-LSTM network outputs a temperature suggestion value of 620°C, a time suggestion value of 40 min, a catalyst suggestion value of 3.0 g / L, a pH suggestion value of 7.2, and a speed suggestion value of 115 rpm according to the waste type identification result, and these five values constitute ​​​; GRU network directly outputs temperature 615℃, time 42min, catalyst 3.2g / L, pH value 7.3, and rotation speed 120rpm, which constitutes ; LSTM network outputs temperature 625℃, time 41min, catalyst 3.1g / L, pH value 7.4, and rotation speed 118rpm, which constitutes When the weighted fusion is performed, the formula is actually calculated respectively for each control variable dimension.

[0074] For the temperature dimension, the calculation formula is: temperature adjustment amount = ×( temperature value in the first model) + ×( temperature value in the second model) + ×( temperature value in the third model). Similarly, the same weighted summation calculation is performed for the four dimensions of time, catalyst, pH, and rotation speed, and finally five independent adjustment amount values are obtained. Taking the calculation of the temperature adjustment amount as an example: is 0.335, the temperature suggestion in the first model is 620℃; is 0.342, the temperature suggestion in the second model is 615℃; is 0.323, the temperature suggestion in the third model is 625℃. According to the weighted summation formula: temperature adjustment amount = 0.335×620 + 0.342×615 + 0.323×625 = 619.6℃. The same way is used to calculate the other four control variables, and finally a comprehensive control parameter vector U containing five values is obtained. Therefore, although the formula looks like multiplying three outputs by weights and summing them up, since each output itself contains the suggested values of five control variables, the actual calculation process is to perform weighted summation on five control variable dimensions respectively, thereby generating five independent adjustment amounts, rather than a single value. This vectorized weighted fusion method not only realizes the comprehensive decision of multi-model results, but also guarantees the independence and accuracy of each control variable.

[0075] It should be noted that in the fusion decision process, the historical prediction accuracy of the CNN-LSTM fusion network, the GRU network and the LSTM dynamic modeling network on the recent samples is normalized to obtain dynamic weight coefficients. After obtaining the dynamic weight coefficients, the system maps the outputs of the three networks to a unified control variable representation. Specifically, the waste type probability distribution output by the CNN-LSTM is converted into recommended control parameters for the type of waste through a pre-set type-process parameter mapping table. For example, when high-organic waste is identified, the mapping table gives a temperature recommendation value of 600-650℃, a time recommendation value of 40-45min, a catalyst dosage recommendation value of 3.0-3.5g / L, a pH recommendation value of 7.0-7.5, and a rotation speed recommendation value of 110-130rpm. The optimal process parameters output by the GRU network directly correspond to specific numerical suggestions for temperature, time, catalyst dosage, pH adjustment amount and rotation speed, such as temperature 620℃, time 42min, catalyst 3.2g / L, pH value 7.3, and rotation speed 120rpm. The processing effect prediction index output by the LSTM network, such as a predicted gas production rate that is 5% lower, is converted into incremental adjustment suggestions for each control variable through an effect-parameter correction table, such as temperature +10℃, time +2min, catalyst +0.3g / L, pH +0.2, and rotation speed +5rpm. For each control variable, such as temperature, the system records the output values of the control variable from the three networks as respectively multiplied by the corresponding dynamic weight coefficients and summed to obtain the final adjustment amount of the control variable, i.e. the temperature adjustment amount . For example, if = 620℃, = 615℃, =625℃ then the temperature adjustment amount = 0.335x620 + 0.342x615 + 0.323x625 = 619.6℃, and so on to calculate the time adjustment amount, catalyst dosage adjustment amount, pH adjustment amount and rotation speed adjustment amount to obtain the target setting values of the five control variables. After generating the five adjustment amounts, the system compares each adjustment amount with the current actual operating parameters to calculate the deviation amplitude of each control parameter, for example, if the temperature adjustment amount is 619.6℃ and the current operating temperature is 600℃, then the temperature deviation amplitude is 19.6℃. The system presets the influence weight coefficients of each control variable on the treatment effect, which represent the influence degree of each control variable on the comprehensive indicators including gas production rate, pollutant removal rate and energy consumption, etc. The influence weight coefficients are determined through correlation analysis of historical data or expert experience, for example, the influence of temperature on gas production rate is the largest, with an influence weight coefficient of 0.35, the influence of catalyst dosage is the second, with a weight coefficient of 0.30, the time weight coefficient is 0.20, the pH adjustment weight coefficient is 0.10, and the rotation speed weight coefficient is 0.05. The calculation of the urgency score adopts the weighted summation method, that is, the deviation amplitude of each parameter is multiplied by the corresponding influence weight coefficient, for example, the temperature urgency score = 19.6℃x0.35 = 6.86. After calculating the urgency scores of the five parameters, the execution priority is determined according to the order from high to low. After the sorting is completed, the system encapsulates the sorting results, adjustment amounts and execution priority into a comprehensive control instruction. The data structure of the instruction includes the control parameter name, i.e. temperature, time, catalyst dosage, pH adjustment amount and rotation speed, the corresponding target setting value, i.e. 619.6℃, 42min, 3.2g / L, 7.3 and 120rpm, the adjustment step, i.e. the increment from the current value to the target value, such as +19.6℃ for temperature from 600℃ to 619.6℃, and the execution priority order, i.e. 1, 2, 3, 4 and 5. After the comprehensive control instruction is issued to the execution module, the controller converts the instruction into device driving signals. Specifically, temperature adjustment is realized by controlling the power of the heating device, time adjustment is realized by setting the reaction residence time, catalyst dosage and pH adjustment amount are realized by controlling the flow rate of the dosing device, including the flow rate of the catalyst dosing pump and the flow rate of the pH adjustment agent dosing pump, and rotation speed adjustment is realized by controlling the rotation speed of the stirring device. The system adopts PID control algorithm or model predictive control algorithm to adjust the power of the heating device, the rotation speed of the stirring device and the flow rate of the catalyst dosing pump and the pH adjustment agent dosing pump in real time.In the process of performing the regulation action, the system continuously collects actual feedback values including actual temperature, actual processing time, actual catalyst dosage, actual pH value and actual rotation speed and compares them with the target set values in the integrated regulation instruction to calculate the effect deviation. When the effect deviation exceeds a preset threshold, for example, the actual temperature deviates from the target temperature by more than 5°C, or an abnormal working condition such as sensor drift or sudden change in feed component is detected, the system triggers the weight adaptive updating mechanism to recalculate the dynamic weight coefficients of the three networks and perform secondary weighted fusion calculation to generate a new integrated regulation instruction. If necessary, the system will also fine-tune the learning rate and prediction window of the GRU network online, such as adjusting the learning rate according to the current deviation ratio to speed up parameter convergence and improve robustness.

[0076] In a specific embodiment, the process of performing a specific regulation action by the adaptive feedback control system can specifically include the following steps:

[0077] (1) Convert the integrated regulation instruction into a device control signal, calculate the control output by the PID control algorithm, and the control output is the heating device power adjustment, the stirring device rotation speed adjustment and the dosage device flow adjustment. Adjust the heating device power, stirring device rotation speed and dosage device flow according to the deviation between the adjustment amount of each regulation parameter in the integrated regulation instruction and the actual processing parameter;

[0078] (2) Real-time monitoring of the execution effect of heating device power adjustment, stirring device rotation speed adjustment and dosage device flow adjustment, comparing and analyzing the actual processing parameters with the adjustment amount of each regulation parameter in the integrated regulation instruction, and calculating the effect deviation value. The actual processing parameters include actual temperature, actual processing time, actual catalyst dosage, actual pH value and actual rotation speed;

[0079] (3) When the effect deviation value exceeds the preset threshold, re-trigger the weighted fusion voting decision algorithm to generate a new integrated regulation instruction.

[0080] Specifically, the system converts the integrated regulation instruction generated by the multi-model fusion into a device control signal that can be recognized by the executing mechanism, and calculates it through the Proportional-Integral-Derivative (PID) control algorithm. Its control law can be expressed as: ; Where, represents the control signal output at time ; represents the instantaneous deviation at a certain time ; e(t) is the deviation between the set value and the actual feedback value; 、 、 The proportional, integral and differential coefficients are respectively P, I and D. Through the above PID operation, the system can adjust the power of the heating device, the rotating speed of the stirring device and the flow of the adding unit in real time to ensure that each operating parameter approaches the target value.

[0081] The deviation e(t) is derived from the difference between the target set value in the comprehensive control instruction and the actual feedback value. Specifically, the target set value of temperature, time, catalyst addition amount, pH adjustment amount and rotating speed is included in the comprehensive control instruction U, for example, the temperature target is 620°C. During execution, the system collects actual operating parameters in real time through sensors, such as the actual temperature is 600°C. At this time, the temperature dimension deviation e temperature(t) = 620-600 = 20°C. Similarly, the deviation values of other control variables can be calculated. Therefore, the deviation e(t) in the PID formula is directly related to the comprehensive control parameter vector U obtained by the above-mentioned weighted fusion calculation: each element in U (such as the temperature adjustment amount 619.6°C) is the target set value of PID control, and e(t) is the difference between the target set value and the current actual measurement value. In other words, U generated by weighted fusion determines what value should be adjusted to, and PID determines how to adjust the equipment to achieve the target value by calculating the deviation e. The control output u(t) calculated by the PID formula is the control signal sent to the execution device. The control signal corresponds to different physical quantities for different control variables: for temperature control, u(t) represents the power adjustment amount of the heating device, which may be in kilowatts or power percentage; for catalyst addition amount and pH adjustment amount control, u(t) represents the flow adjustment amount of the corresponding adding pump, wherein the catalyst addition amount corresponds to the flow of the catalyst adding pump, and the pH adjustment amount corresponds to the flow of the pH adjusting agent adding pump, which is in units of liters per minute or milliliters per second; for rotating speed control, u(t) represents the rotating speed adjustment amount of the stirring motor, which is in units of revolutions per minute. In actual execution, the system applies the PID control algorithm to each control variable such as temperature, catalyst addition amount, pH adjustment amount and rotating speed. Taking temperature as an example: the target temperature is 620°C (from U), the actual temperature is 600°C (measured by the sensor), the deviation e is calculated to be 20°C, and the amount of heating power that should be increased (such as 15% power) is calculated by the PID formula. The power adjustment instruction is the control signal u(t) sent to the heating device for execution. The PID control process of other control variables is similar and is calculated and executed independently.

[0082] In the process of executing the control action, the system synchronously monitors the actual processing parameters and compares them with the target values corresponding to the comprehensive control instructions to obtain an effect deviation value. When the effect deviation value does not exceed a threshold value, the current control strategy is maintained to ensure system stability; when the deviation exceeds the preset threshold value, the system immediately triggers the weighted fusion voting decision algorithm to dynamically weight and fuse the outputs of the multiple models and generate new comprehensive control instructions and reissue them for execution. Through the above closed-loop feedback mechanism, the system can realize adaptive optimization when the waste composition fluctuates or the working condition suddenly changes, so that the entire processing process is always in a near-optimal operating state.

[0083] Taking a pyrolysis gasification process for solid waste utilization as an example, the system collects raw material characteristic data through multiple types of sensors in the feeding link. The near-infrared spectrum sensor detects that the organic matter ratio is about 65%, and the cellulose content is about 25%, while the X-ray fluorescence sensor identifies that this batch of material contains trace heavy metal elements, with a lead content of about 120 ppm. After preprocessing and uniform data matrix construction, the above multi-source data are input into the multi-model integrated network for analysis. The CNN-LSTM subnetwork identifies that the batch of waste belongs to high-combustible organic matter, with a confidence of 92%; the GRU subnetwork predicts that the optimal process parameter is medium-high temperature pyrolysis, with a recommended temperature of about 580°C, and is supplemented with a catalyst, with a dosage ratio of 1.5% of the raw material mass; the LSTM subnetwork predicts the processing effect index, showing that the gas production rate is expected to reach 85 m³ / ton of raw material, and the ratio of CO and H2 in the synthesis gas can reach 45%. After weighted fusion voting (weight distribution: CNN-LSTM: 0.4, GRU: 0.4, LSTM: 0.2), the comprehensive control instructions are generated and issued to the execution unit, setting the target temperature of the heating device to 580°C, the rotation speed of the stirring equipment to 45 rpm, and the flow rate of the catalyst dosing system to 150 g / min (corresponding to a feeding rate of 10 tons / hour). During the operation, the PID controller adjusts the parameters in real time, for example, when there is a 5°C deviation between the set temperature of 580°C and the actual temperature of 575°C, the heating power is automatically adjusted to 85% to quickly approach the target. After about 15 minutes, the online gas analyzer monitors that the CO+H2 concentration in the actual synthesis gas is only 38%, which is lower than the target of 45%, and the deviation exceeds the tolerance threshold of 5%, so the system immediately triggers the weighted fusion voting algorithm for secondary calculation. The GRU subnetwork suggests increasing the temperature to 600°C, and the LSTM subnetwork suggests increasing the catalyst ratio to 1.8%. After re-fusion, the new control instructions are generated to increase the heating temperature set value to 600°C and adjust the catalyst dosing rate to 180 g / min. After the instructions are issued, the processing effect improves, the gas production rate stabilizes at 82 m³ / ton, the effective component CO+H2 concentration in the synthesis gas increases to 44%, and the heavy metals are effectively solidified under high temperature conditions, with significantly reduced impurity emissions, realizing efficient utilization and resource conversion of solid waste.

[0084] The intelligent regulation method for solid waste utilization in the embodiments of the present application is described above, and the intelligent regulation system for solid waste utilization in the embodiments of the present application is described below. Please refer to Figure 3 An embodiment of the intelligent regulation system for solid waste utilization in the embodiments of the present application includes:

[0085] A multi-source perception module, which is deployed with a multi-type sensor array at a key process node of a solid waste treatment system, acquires real-time monitoring data of the solid waste, and pre-processes the real-time monitoring data to form a unified data matrix;

[0086] A fusion decision module, which is configured to input the pre-processed unified data matrix into a multi-model integrated network for intelligent analysis, the multi-model integrated network including a CNN-LSTM fusion network, a GRU network and an LSTM dynamic modeling network, the CNN-LSTM fusion network outputting a waste type probability distribution, the GRU network outputting optimal process parameters, and the LSTM dynamic modeling network outputting a processing effect prediction index;

[0087] A closed-loop execution module, which is configured to integrate the output results of the CNN-LSTM fusion network, the GRU network and the LSTM dynamic modeling network by using a weighted fusion voting decision algorithm, generate comprehensive regulation instructions and execute specific regulation actions through an adaptive feedback control system, and realize intelligent regulation of the solid waste treatment process.

[0088] Through the synergistic cooperation of the above-mentioned components, the system realizes an end-to-end integrated control process from multi-source data acquisition, standardized fusion, parallel intelligent analysis to weighted decision and closed-loop execution, which can quickly complete waste type identification and process parameter adaptive setting under complex and fluctuating feeding conditions, and can also dynamically correct control variables according to real-time feedback to ensure that key variables such as temperature, time, catalyst, pH and rotation speed are stably converged to the target interval; on this basis, the system further prioritizes multiple parameters according to the urgency evaluation, taking into account processing efficiency, energy consumption and emission constraints, to realize high-reliability, scalable and landable intelligent regulation of solid waste resource utilization.

[0089] The intelligent regulation system for solid waste utilization in the embodiments of the present application is described above Figure 3 The intelligent regulation system for solid waste utilization in the embodiments of the present application is described above

[0090] Referring to Figure 4 In addition, the present application also provides an intelligent regulation device for solid waste utilization, which can be a server, and its internal structure can be as followsFigure 4 The intelligent control device for solid waste utilization includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. The processor of the computer is used to provide computing and control capabilities. The memory of the intelligent control device for solid waste utilization includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the intelligent control device for solid waste utilization is used to store the corresponding data in the embodiment. The network interface of the intelligent control device for solid waste utilization is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement the above method.

[0091] Those skilled in the art can understand that, Figure 4 The structure shown in the figure is only a block diagram of part of the structure related to the scheme, and does not constitute a limitation on the intelligent control device for solid waste utilization to which the scheme is applied.

[0092] The present application also provides a computer readable storage medium, which can be a non-volatile computer readable storage medium, and can also be a volatile computer readable storage medium. The computer readable storage medium stores instructions, and when the instructions run on a computer, the computer executes the steps of the intelligent control method for solid waste utilization.

[0093] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-mentioned system, system and unit can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0094] The integrated unit, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical scheme of the present application or the whole or part of the technical scheme which essentially contributes to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) for solid waste utilization to execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0095] The above examples are only used to illustrate the technical solutions of the present application, but not to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those ordinarily skilled in the art should understand: the technical solutions recorded in the foregoing examples can still be modified, or some technical features can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for intelligent regulation of solid waste utilization, characterized by, The application relates to a solid waste treatment system and a method for dynamically adjusting a treatment process of the solid waste treatment system. A plurality of types of sensor arrays are arranged at key process nodes of a solid waste treatment system to obtain real-time monitoring data of the solid waste, and the real-time monitoring data is preprocessed to form a unified data matrix. The preprocessed unified data matrix is input into a multi-model integrated network for intelligent analysis, the multi-model integrated network comprising a trained CNN-LSTM fusion network, a trained GRU network and a trained LSTM dynamic modeling network, the CNN-LSTM fusion network being trained by historical waste data to output a waste type probability distribution, the GRU network being trained by historical process parameter data to output optimal process parameters, and the LSTM dynamic modeling network being trained by historical treatment effect data to output treatment effect prediction indexes. A weighted fusion voting decision algorithm is used to integrate the output results of the trained CNN-LSTM fusion network, the trained GRU network and the trained LSTM dynamic modeling network, and the outputs of the networks are respectively mapped to five control variable dimensions of temperature, time, catalyst dosage, pH adjustment amount and rotating speed, each control variable dimension is weighted and summed according to a dynamic weight coefficient to generate comprehensive control instructions comprising temperature adjustment amount, time adjustment amount, catalyst dosage adjustment amount, pH adjustment amount and rotating speed adjustment amount.

2. The method of claim 1, wherein, The preprocessing of the real-time monitoring data to form the unified data matrix comprises the following steps. For each data point in the real-time monitoring data, the arithmetic mean of 15 data points including the previous 7 data points, the data point itself and the next 7 data points of the data point is obtained to obtain a filtered data point, and all the filtered data points are combined to obtain denoised sensor data. Each value in the denoised sensor data is subtracted from the mean of historical sensor data and then divided by the standard deviation of the historical sensor data to obtain standardized data. The standardized data from different sensors are sorted and aligned according to timestamps, resampled and interpolated at a time interval of 100 ms to obtain the unified data matrix.

3. The method of claim 1, wherein, The real-time monitoring data of the solid waste comprises the following steps. The solid waste is detected by a near-infrared spectrum sensor through spectrum scanning, spectrum analysis in a wavelength range of 1000-2500 nm is adopted to obtain organic matter content, cellulose content and lignin content data; The solid waste is detected by an X-ray fluorescence sensor to identify the concentration distribution of heavy metal elements, and the solid waste is detected by a thermal gravimetric analysis sensor to monitor a mass change characteristic curve in a temperature range of 25-1000 DEG C, and characteristic parameters of mass change are extracted from the mass change characteristic curve, the characteristic parameters of mass change comprising an initial weight loss temperature, a temperature corresponding to a maximum weight loss rate, a weight loss rate in a specific temperature range and a final residual mass percentage; The treatment environment is monitored in real time by a gas chromatograph sensor, a temperature sensor, a humidity sensor and a pressure sensor to obtain volatile organic compound concentration, environmental temperature and humidity and system pressure parameters. The obtained organic matter content, cellulose content, lignin content, heavy metal element concentration distribution, mass change characteristic parameter, volatile organic compound concentration, environmental temperature and humidity, and system pressure parameter jointly constitute the real-time monitoring data.

4. The method of claim 1, wherein, The trained CNN-LSTM fusion network outputs a waste type probability distribution, including: The input unified data matrix is subjected to spatial feature extraction processing by three one-dimensional convolution layers, each convolution layer uses a convolution kernel with a size of 3x1, and the step is set to 1, and the convolution output is sequentially subjected to batch normalization and ReLU activation function processing; The convolution features are subjected to feature dimension reduction processing by a max-pooling layer, the pooling window size is 2x1, and the maximum feature value in each pooling region is retained; The feature sequence after pooling is subjected to time series modeling processing by a bidirectional LSTM layer, the bidirectional LSTM layer contains 128 hidden units, and the dropout ratio is set to 0.3 to prevent overfitting; The LSTM output is subjected to probability normalization processing by a softmax activation function, and the waste type probability distribution of 12 main solid waste types is output.

5. The method of claim 1, wherein, Also includes: According to the processing effect prediction index, the trained GRU network is adjusted and the optimal process parameters are re-predicted, specifically including: calculating the deviation ratio of the processing effect prediction index and the benchmark effect index to obtain an adjustment factor; calculating the learning rate adjustment value of the trained GRU network according to the adjustment factor, if it is less than the minimum learning rate value, taking the minimum learning rate value as the learning rate value of the trained GRU network; updating the GRU network using the adjusted learning rate, and re-predicting the time series of the optimal process parameters using the updated trained GRU network.

6. The method of claim 1, wherein, The output results of the trained CNN-LSTM fusion network, the trained GRU network and the trained LSTM dynamic modeling network are integrated by using a weighted fusion voting decision algorithm, the outputs of each network are mapped to the control variable space of temperature, time, catalyst dosage, pH adjustment amount and speed, and are weighted and summed according to dynamic weight coefficients to generate comprehensive control instructions including temperature adjustment amount, time adjustment amount, catalyst dosage adjustment amount, pH adjustment agent dosage adjustment amount and speed adjustment amount, including: The historical prediction accuracy of the trained CNN-LSTM fusion network, the trained GRU network and the trained LSTM dynamic modeling network is calculated by weight, and the accuracy of each network is divided by the sum of the total accuracy to obtain a dynamic weight coefficient; The waste type probability distribution, the optimal process parameters and the processing effect prediction index are respectively mapped to the five control variable dimensions of temperature, time, catalyst dosage, pH adjustment amount and speed, and each control variable dimension is subjected to weighted fusion processing, the mapping values of the control variable from the three networks are multiplied by the corresponding dynamic weight coefficients and summed to generate the adjustment amount of the control variable, thereby obtaining the temperature adjustment amount, the time adjustment amount, the catalyst dosage adjustment amount, the pH adjustment agent dosage adjustment amount and the speed adjustment amount. The temperature adjustment amount, time adjustment amount, catalyst addition adjustment amount, pH adjustment amount, and rotation speed adjustment amount are compared with the current actual running temperature, time, catalyst addition amount, pH value, and rotation speed, respectively, to calculate the deviation amplitude of each control parameter, wherein the deviation amplitude is the absolute value of the adjustment amount; An urgency score is calculated according to the deviation amplitude and a preset influence weight coefficient, the urgency score being equal to the product of the deviation amplitude and the influence weight coefficient, and the influence weight coefficient representing the influence degree of each control variable on the treatment effect; The temperature, time, catalyst addition amount, pH adjustment amount, and rotation speed are sorted in descending order of the urgency score, and the sorted control parameters, corresponding adjustment amounts, and execution priority order are combined to form the comprehensive control instruction.

7. The method of claim 1, wherein, Further comprising: A specific control action is performed through an adaptive feedback control system, including: The comprehensive control instruction is converted into a device control signal, a control output quantity is calculated through a PID control algorithm, the control output quantity being a heating device power adjustment amount, a stirring device rotation speed adjustment amount, and a dosing device flow adjustment amount, and the heating device power, stirring device rotation speed, and dosing device flow are adjusted according to the deviation between the adjustment amount of each control parameter in the comprehensive control instruction and the actual treatment parameter; The execution effect of the heating device power adjustment, stirring device rotation speed adjustment, and dosing device flow adjustment is monitored in real time, the actual treatment parameters are compared and analyzed with the adjustment amount of each control parameter in the comprehensive control instruction, and an effect deviation value is calculated, the actual treatment parameters including actual temperature, actual treatment time, actual catalyst addition amount, actual pH value, and actual rotation speed; When the effect deviation value exceeds a preset threshold value, the weighted fusion voting decision algorithm is triggered again to generate a new comprehensive control instruction.

8. An intelligent regulation system for solid waste utilization, characterized by, The intelligent control system for solid waste utilization is used to implement the intelligent control method for solid waste utilization as claimed in any one of claims 1-7, and the intelligent control system for solid waste utilization comprises: A multi-source perception module, which is arranged at key process nodes of a solid waste treatment system and comprises a plurality of types of sensor arrays, acquires real-time monitoring data of the solid waste, and pre-processes the real-time monitoring data to form a unified data matrix; A fusion decision module, which is configured to input the pre-processed unified data matrix into a multi-model integrated network for intelligent analysis, the multi-model integrated network comprising a trained CNN-LSTM fusion network, a trained GRU network, and a trained LSTM dynamic modeling network, the CNN-LSTM fusion network being trained by historical waste data to output a waste type probability distribution, the GRU network being trained by historical process parameter data to output optimal process parameters, and the LSTM dynamic modeling network being trained by historical treatment effect data to output treatment effect prediction indexes; A multi-source perception module, which is arranged at key process nodes of a solid waste treatment system and comprises a plurality of types of sensor arrays, acquires real-time monitoring data of the solid waste, and pre-processes the real-time monitoring data to form a unified data matrix; The closed loop execution module is configured to integrate the output results of the trained CNN-LSTM fusion network, the trained GRU network and the trained LSTM dynamic modeling network by using a weighted fusion voting decision algorithm, map each network output to five control variable dimensions of temperature, time, catalyst dosage, pH adjustment amount and rotating speed, respectively, and perform weighted summation calculation on each control variable dimension according to a dynamic weight coefficient to generate a comprehensive control instruction including temperature adjustment amount, time adjustment amount, catalyst dosage adjustment amount, pH adjustment amount and rotating speed adjustment amount.

9. An intelligent regulating device for solid waste utilization, characterized in that, The computer program is stored in the memory and executable on the processor, and the processor executes the computer program to implement the intelligent control method for solid waste utilization according to any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is stored in the memory and executable on the processor, and the processor executes the computer program to implement the intelligent control method for solid waste utilization according to any one of claims 1 to 7.

Images