Heat treatment process matching regulation and control method and system based on dewatered sludge calorific value estimation

Abstract

The invention provides a heat treatment process matching regulation and control method and system based on dewatered sludge calorific value estimation, and the method comprises the steps: collecting physicochemical property data of dewatered sludge from different sources in batches, pairing the physicochemical property data with optimal operation parameters under three typical heat treatment processes, and carrying out multi-layer statistical screening and characteristic standardization; and training by adopting a deep neural network to form a calorific value-process coupling mapping model. The model can automatically calculate the feeding rate, the air supply amount, the hearth temperature gradient and the residence time matched with the current equipment working condition according to the sludge characteristics detected in real time. And online feedback error monitoring and model adaptive fine tuning are realized, and the feature contribution degree is quantified through SHAP. According to the method, the accuracy and the consistency of parameter recommendation are remarkably improved, the characteristic mode can be quickly identified and the optimal regulation and control strategy can be output especially for dewatered sludge with large component fluctuation and complex sources, the system response time is greatly shortened, and the heat treatment process is always kept in an efficient and stable operation interval.

Description

Technical Field

[0001] This invention relates to the field of solid waste thermal treatment and intelligent control technology, and in particular to a thermal treatment process matching and control method and system based on the estimation of the calorific value of dewatered sludge. Background Technology

[0002] In the field of solid waste thermal treatment and intelligent control technology, especially in the area of ​​sludge calorific value estimation and process control systems based on nonlinear mapping spectra, existing systems are typically divided into two main parts: a calorific value estimation module and a process control module. Mainstream solutions generally employ empirical regression models based on physicochemical testing data to estimate the calorific value of dewatered sludge, relying on manual rules or simplified linear formulas to determine the operating parameters of the thermal treatment process. For example, the commonly used calorific value estimation method in the industry collects data such as sludge moisture content, organic matter content, and ash ratio, and uses regression analysis or empirical formulas based on previously obtained statistical distribution parameters for rapid calorific value estimation. Subsequently, process control parameters such as feed rate, air supply, furnace temperature gradient, and residence time are often set based on the experience of on-site operators, or adjusted using fuzzy control or a single closed-loop feedback method. Some systems have preliminary data-driven capabilities, but these are often limited to process adaptive parameter tuning and lack a precise end-to-end closed-loop mapping mechanism.

[0003] In recent years, with the increasing demands for sludge energy utilization and green, low-carbon emissions, the need for intelligent sensing of feed material characteristics and parameter coordination in thermal treatment processes has continued to grow. Technological development trends are gradually shifting from offline decision-making based on static rules to dynamic response and intelligent closed-loop control. Industry advancements have led to the application of multivariate regression models, such as Support Vector Machine Regression (SVR) and Extreme Learning Machine (ELM), in sludge calorific value estimation and energy consumption optimization. However, these models largely remain at the level of simple parameter mapping and single-condition optimization. When faced with scenarios involving diverse sludge characteristics from different sources and complex and variable operating conditions of thermal treatment equipment, the adaptability and predictive accuracy of these decision-making models are insufficient.

[0004] Currently, most representative technological achievements rely on regression formulas or data-driven shallow learning models to estimate calorific value and recommend process parameters. Their respective technical characteristics and applicability are limited as follows: In the calorific value estimation stage, mainstream models primarily use univariate or low-dimensional features, failing to characterize the coupling effects between sludge physicochemical indicators and lacking the ability to uncover nonlinear relationships. In the process control stage, manual decision support or single-point feedback adjustment are frequently used, lacking system-level linkage and multi-objective collaborative optimization mechanisms. These technologies are mainly suitable for relatively stable sludge sources and routine operation under single-process equipment conditions, and are ill-suited to address the complex needs of highly heterogeneous sludge sources, parallel processes, and collaborative optimization of energy consumption and emission indicators in fields such as urban wastewater treatment and industrial wastewater treatment.

[0005] The existing technical system has obvious deficiencies, mainly manifested in the lack of accuracy and intelligence in the parameter mapping relationship between the calorific value estimation module and the process control module. Most current systems rely on serial transmission with the calorific value as an intermediate variable, taking the calorific value estimation result as an independent input for the process control link. However, due to the nonlinearity, diversity of process responses, and fluctuations in sludge physical properties, it is difficult to directly convert the estimation result into reasonable control instructions. Most existing solutions adopt empirical rules, linear approximation, or manually set thresholds, which cannot effectively adapt to the complex nonlinear mapping between calorific value characteristics and the operating parameters of various types of equipment, resulting in large parameter recommendation deviations and making it difficult to ensure optimal energy consumption or safe operation. At the same time, there is a lack of closed-loop adaptive learning and model online optimization mechanisms, showing problems of response lag and poor adaptability in the face of dynamic fluctuations in sludge characteristics and operating conditions.

[0006] In addition, the intelligent control systems widely concerned in the current industry mainly focus on parameter optimization or local feedback of a single link, and lack attention to the full-process information closed-loop of calorific value estimation and process matching, which directly affects the overall coordination efficiency and intelligent level of the system. From an operational perspective, existing models are often black-box type shallow algorithms, lacking transparency and interpretability in the control process, making it difficult to gain sufficient trust from on-site operators and restricting their promotion and application in large-scale automated thermal treatment lines.

[0007] Therefore, there is an urgent need to provide an intelligent mapping model that can achieve the nonlinear, end-to-end mapping between the physicochemical properties of sludge and the key control parameters of various thermal treatment processes. Summary of the Invention

[0008] The present invention provides a thermal treatment process matching control method and system based on the calorific value estimation of dewatered sludge, aiming to solve the problems existing in the prior art mentioned in the above background technology.

[0009] The technical solution of the present invention: A thermal treatment process matching control method based on the calorific value estimation of dewatered sludge includes the following steps: S1: Collect the physicochemical property data of dewatered sludge and simultaneously record the optimal operating parameter combinations to obtain an input-output paired sample set; S2: Preprocess the input-output paired sample set; S3: Use a deep neural network model to train and generate a calorific value-process coupling mapping graph; S4: Input the physicochemical indexes obtained by rapid detection of the sludge to be processed in real time into the trained calorific value-process coupling mapping graph to obtain a set of target setting values for the feed rate, air supply volume, furnace temperature gradient, and residence time under the corresponding working conditions; S5: Generate a control instruction sequence based on the set of target setting values and send it to the actuator of the thermal treatment system to drive each process unit to operate according to the recommended parameter combination; S6: During the process execution, continuously collect the actual operating data fed back by the sensor, calculate the residual sequence between the actual operating parameters and the target set value set, and determine whether the residual sequence continuously exceeds the preset threshold range. S7: If it is determined that the residual sequence continuously exceeds the preset threshold range, the online learning mechanism is triggered. The network weights of the calorific value-process coupling mapping map are locally fine-tuned using the latest input-output data to generate an updated mapping map to adapt to the current operating condition deviation. S8: Combine the SHAP value analysis method to quantitatively evaluate the contribution of each input feature in the updated calorific value-process coupling mapping to the output control parameters.

[0010] The present invention also provides a thermal treatment process matching and control system based on the estimation of the calorific value of dewatered sludge, wherein the system utilizes the above method to achieve thermal treatment process matching and control.

[0011] The beneficial effects of this invention are as follows: 1. This invention achieves end-to-end intelligent generation of key operating parameters from multi-dimensional physicochemical characteristics of sludge to key operating parameters by constructing a calorific value-process coupling mapping map. This invention utilizes a deep neural network to perform nonlinear fitting on a large amount of paired data, directly mapping multi-source features such as moisture content, organic matter content, and particle size distribution into control commands such as feed rate, air supply, and furnace temperature gradient. This effectively avoids the cumulative error and decision-making gap in the serial structure, significantly improving the accuracy and consistency of parameter recommendations. Especially when dealing with dewatered sludge with large compositional fluctuations and complex sources, it can quickly identify characteristic patterns and output the optimal control strategy, greatly shortening the system response time and keeping the thermal treatment process in a highly efficient and stable operating range. 2. This invention introduces an online learning mechanism and an interpretability analysis module, forming a closed-loop optimization system with continuous evolution characteristics. In each run, this invention collects actual feedback data from the sensors and performs residual analysis with the model's predicted values. Once a continuous deviation of the system is detected, local weight fine-tuning is triggered, and the mapping spectrum is dynamically updated to ensure that the model always fits the current equipment status and the changing trend of raw material characteristics. At the same time, the contribution of each input feature in parameter decision-making is quantitatively analyzed by combining the SHAP value, realizing the visualization of the control logic. 3. This invention constructs a comprehensive decision-making framework that integrates multi-dimensional perception, deep modeling, dynamic evolution, and interpretable output. It fundamentally solves the technical problem that the calorific value estimation results cannot be effectively converted into precise control commands. By establishing a deep nonlinear correlation between high-dimensional inputs and multi-objective outputs, the system can not only meet the optimization needs of a single process path, but also be horizontally extended to the unified modeling and intelligent switching of various heat treatment technology routes such as rotary kiln drying, fluidized bed incineration, and pyrolysis carbonization. It has good versatility and scalability, and does not rely on cumbersome manual parameter tuning or preset threshold rules, which significantly reduces the dependence on expert experience and improves the level of automation and operating efficiency. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention; Figure 2 This is a schematic diagram of the generation process of the calorific value-process coupling mapping map in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the process of generating the target set of values ​​in an embodiment of the present invention. Detailed Implementation

[0013] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0014] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0015] like Figure 1 As shown in the figure, this embodiment provides a method for matching and controlling a thermal treatment process based on the estimation of the calorific value of dewatered sludge, which specifically includes the following steps: S1: Collect physicochemical property data of multiple batches of dewatered sludge from different sources. The physicochemical property data includes moisture content, organic matter content, ash ratio, particle size distribution and measured lower heating value. Simultaneously record the optimal combination of operating parameters for each batch of sludge in rotary kiln drying, fluidized bed incineration and pyrolysis carbonization processes to form an input-output paired sample set with operating condition labels. S2: Perform data preprocessing on the input-output paired sample set, including outlier removal, missing value imputation and feature standardization, to generate standardized sludge feature vectors and normalized process control parameter vectors, which serve as the training basis for constructing nonlinear mapping relationships. S3: Based on the pre-processed standardized sludge feature vector and normalized process control parameter vector, a calorific value-process coupling mapping spectrum is generated by training a deep neural network model. The calorific value-process coupling mapping spectrum represents the nonlinear functional relationship between the multidimensional characteristics of sludge and the key control parameters of the thermal treatment equipment. S4: Input the physicochemical indicators obtained from the rapid detection of the sludge to be treated in real time into the trained calorific value-process coupling mapping map to obtain the target set of feed rate, air supply, furnace temperature gradient and residence time under the corresponding working conditions. S5: Generate a control instruction sequence based on the target set of values, and send it to the actuator of the heat treatment system to drive each process unit to operate according to the recommended parameter combination, thereby realizing closed-loop process matching based on the estimated calorific value. S6: During the process execution, continuously collect actual operating data fed back by sensors, calculate the residual sequence between the actual operating parameters and the target set of values, and determine whether the residual sequence continuously exceeds the preset threshold range. S7: If it is determined that the residual sequence continuously exceeds the preset threshold range, the online learning mechanism is triggered. The network weights of the calorific value-process coupling mapping map are locally fine-tuned using the latest input-output data to generate an updated mapping map to adapt to the current operating condition deviation. S8: Combining the SHAP value analysis method, the contribution of each input feature in the updated calorific value-process coupling mapping to the output control parameters is quantitatively evaluated, and an interpretability report is generated to help operators understand the control logic and verify the consistency of model behavior.

[0016] In this embodiment, step S1 involves collecting physicochemical property data of multiple batches of dewatered sludge from different sources. This data includes moisture content, organic matter content, ash content, particle size distribution, and measured lower heating value. Simultaneously, the optimal operating parameter combinations for each batch of sludge in rotary kiln drying, fluidized bed incineration, and pyrolysis carbonization processes are recorded to form an input-output paired sample set with operating condition labels. Specifically, this includes the following steps: S1.1: Obtain multiple batches of dewatered sludge samples from different sources such as municipal sewage treatment plants and industrial wastewater treatment stations. Based on laboratory testing equipment, each batch of samples is tested for moisture content, organic matter content, ash ratio, and particle size distribution. Simultaneously, the lower heating value is measured using an oxygen bomb calorimeter to generate a raw physicochemical property dataset to comprehensively characterize the energy properties and physicochemical heterogeneity of the sludge. S1.2: Based on each batch of dewatered sludge samples mentioned above, small-scale or pilot-scale experiments were carried out in three typical thermal treatment process units: rotary kiln drying, fluidized bed incineration, and pyrolysis carbonization. The response surface methodology was used to optimize the operating conditions of each process and determine the optimal combination of operating parameters that maximizes energy utilization efficiency and meets pollutant emission standards under the corresponding operating conditions. These parameters include feed rate, air supply, furnace temperature gradient, and residence time, forming the baseline data for process-level control parameters. S1.3: Spatiotemporally align the original physicochemical properties data of each batch of sludge with the optimal combination of operating parameters obtained in the three thermal treatment processes. Establish a one-to-one input-output pairing relationship based on the batch number and treatment timestamp, and attach operating condition labels to each set of data to identify the type of thermal treatment process to which it is applicable. Generate an input-output pairing sample set with operating condition labels to support subsequent multimodal mapping modeling.

[0017] S1.4: Perform a preliminary quality assessment on the generated input-output paired sample set, use box plots to identify outliers in physicochemical indicators, combine the Kolmogorov-Smirnov test to judge the consistency of data distribution, and remove outliers that deviate significantly from the overall trend to ensure the representativeness and reliability of the training data and obtain a clean sample subset for subsequent modeling. S1.5: Normalize the physicochemical properties and process control parameters in the clean sample subset, use the Z-score standardization method to transform the input feature vector into a dimensionless form, and compress the output control parameter vector in the [0,1] interval based on the maximum-minimum method to generate a standardized sludge feature vector and a normalized process control parameter vector, which serve as the basis for a unified data format for constructing nonlinear mapping relationships.

[0018] In this embodiment, step S2 involves preprocessing the input-output paired sample set, including outlier removal, missing value imputation, and feature standardization, to generate a standardized sludge feature vector and a normalized process control parameter vector, which serve as the training basis for constructing the nonlinear mapping relationship. Specifically, this includes the following steps: S2.1: Obtain physicochemical property data of multiple batches of dewatered sludge and the optimal combination of operating parameters under corresponding working conditions. The physicochemical property data includes moisture content, organic matter content, ash ratio, particle size distribution, and measured lower heating value. The combination of operating parameters includes feed rate, air supply, furnace temperature gradient, and residence time. Based on industrial sensor measurement standards, unify the units and correct the dimensions of each field data to generate a preliminary structured input-output paired sample set to eliminate systematic bias caused by differences in acquisition equipment. S2.2: Based on statistical criteria, outlier detection is performed on the preliminary structured input-output paired sample set. An improved interquartile range (IQR) method combined with Mahalanobis distance algorithm is used to identify outlier sample points that deviate from the normal distribution pattern. The identified outlier data is removed to generate a clean data subset to prevent outlier samples from introducing misleading gradient directions for subsequent model training. S2.3: For missing fields in the clean data subset, the missing value imputation method (MICE) is used based on the correlation structure between variables to perform missing value imputation; specifically, based on the strong correlation between organic matter content and measured lower heating value, a regression prediction equation is constructed to fill the missing heating value data and generate a repaired sample set with 100% integrity to ensure data continuity in the high-dimensional feature space. S2.4: Perform feature standardization on all input feature variables in the repaired sample set. Use Z-score normalization to convert the original indicators such as moisture content, organic matter content, ash ratio, and particle size distribution into a standard normal distribution with a mean of 0 and a standard deviation of 1, generating a standardized sludge feature vector. At the same time, perform Min-Max normalization on the output control parameters such as feed rate, air supply, furnace temperature gradient, and residence time, compressing them to the [0,1] interval to generate a normalized process control parameter vector, so as to eliminate the interference of parameters of different magnitudes on the optimization process of the deep neural network model. S2.5: Align the standardized sludge feature vector and the normalized process control parameter vector according to the batch label and encapsulate them into a high-dimensional tensor format data training package. The data training package serves as the input basis for constructing nonlinear mapping relationships. The training package is divided into training set, validation set and test set according to the proportion through a data partitioning strategy, and a data loader is configured to support the batch iterative training and performance evaluation of subsequent deep neural network models. In this embodiment, in step S3, based on the preprocessed standardized sludge feature vector and the normalized process control parameter vector, a deep neural network model is used to train and generate a calorific value-process coupling mapping map. This calorific value-process coupling mapping map represents the nonlinear functional relationship between the multidimensional characteristics of sludge and the key control parameters of the thermal treatment equipment, such as... Figure 2 As shown, the specific steps include the following: S3.1: Based on the standardized sludge feature vector and normalized process control parameter vector output from the previous step S2, a high-dimensional labeled training dataset is constructed. The input variables are a five-dimensional feature vector consisting of moisture content, organic matter content, ash ratio, particle size distribution, and measured lower heating value. The output variables are a four-dimensional control parameter vector consisting of feed rate, air supply, furnace temperature gradient, and residence time. This forms supervised learning sample pairs for deep neural network training, obtaining original training inputs with working condition adaptability. Based on the standardized sludge feature vector and normalized process control parameter vector output from the preceding step S2, a label alignment processing algorithm is used to achieve consistent mapping between input features and output parameters in terms of time series and operating conditions.

[0019] Furthermore, the five-dimensional feature vector consisting of moisture content, organic matter content, ash ratio, particle size distribution, and measured lower heating value is matched to the four-dimensional control parameter vector consisting of feed rate, air supply, furnace temperature gradient, and residence time through the dimension mapping construction method, while maintaining index consistency.

[0020] Furthermore, a supervised learning sample generation algorithm is used to combine the above input features and output parameters into a high-dimensional labeled training dataset, ensuring that each sample pair has both standardized input features and normalized output parameters.

[0021] Furthermore, a data integrity verification method is used to perform structured verification on the high-dimensional labeled training dataset, eliminating sample pairs with missing labels or mismatched parameters, and generating the original training input dataset for working condition adaptability.

[0022] S3.2: Design a feedforward deep neural network model architecture with a multi-hidden-layer structure, wherein the number of neurons in the input layer matches the five-dimensional normalized sludge feature vector, the number of neurons in the output layer corresponds to the four-dimensional normalized process control parameter vector, and no less than three hidden layers are set in the middle and ReLU activation function is adopted. The initial connection weights are configured through the Xavier initialization method to build a basic calculation framework for the calorific value-process coupling mapping spectrum that can fit complex nonlinear mapping relationships, and generate an initial model structure with deep feature extraction capability. Furthermore, the deep feature extraction function of the intermediate hidden layer is realized through the architecture planning algorithm, and the network topology is constructed by fully connecting each neuron in the layer with the adjacent layer to ensure the integrity of feature transmission and nonlinear combination capability.

[0023] Furthermore, the nonlinear transformation function of the hidden layer output is realized through activation function configuration method, thereby enhancing the network's ability to respond sparsely to high-dimensional input features.

[0024] Furthermore, a weight initialization method is used to achieve initial distribution balancing of each connection weight, and each weight is calculated. The initial range of values The formula is:

[0025] in, The number of input neurons for the current layer. This represents the number of output neurons in the current layer.

[0026] Furthermore, the basic computational function of generating a normalized process control parameter vector after the input feature vector is mapped through multiple hidden layers is realized by encapsulating the network computing framework, and an initial model structure with deep feature extraction capability is output.

[0027] S3.3: Perform batch training on the deep neural network model, use the mean squared error loss function to quantify the deviation between the predicted control parameter vector and the true normalized process parameter vector, implement the back propagation algorithm based on the Adam optimizer to perform weight iterative updates, if the loss decrease on the validation set after each round of training is less than the preset convergence threshold for 5 consecutive rounds, the model is determined to be convergent, and finally output the parameterized calorific value-process coupling mapping map model after training, whose internal weight matrix encodes the nonlinear mapping law from the sludge feature space to the process parameter space; The batch gradient descent training method is employed, using mean squared error (MSE) as the loss function to quantify the difference between the predicted output vector and the true normalized process parameter vector. Defined as:

[0028] in, This refers to the batch sample size. Predict the output vector for the model. This is the true normalized process parameter vector.

[0029] Furthermore, the backpropagation algorithm is implemented through the Adam optimizer, and the gradient calculation and iterative update of the connection weights and bias parameters of each layer are performed using the loss function to ensure that the deep feature extraction and parameter mapping capabilities are optimized under the action of the nonlinear activation function ReLU.

[0030] Furthermore, convergence monitoring is achieved by calculating the magnitude of loss change on the validation set after each training round. When the magnitude of the loss decrease on the validation set is lower than the preset convergence threshold for five consecutive rounds, the training termination condition is triggered to prevent overfitting and improve generalization ability.

[0031] Furthermore, by saving the snapshot of the parameter weights with the best performance on the current validation set, a trained parameterized calorific value-process coupling mapping model is formed, whose internal weight matrix encodes the high-dimensional nonlinear mapping relationship from the sludge feature space to the process parameter space.

[0032] S3.4: Perform offline inference testing on the trained 'calorific value-process coupling mapping map' model. Input the standardized sludge feature vector from the independent validation batch, obtain the corresponding predicted process control parameter vector, and compare it with the actual recorded optimal operating parameters. Calculate the mean absolute percentage error (MAPE) of each control dimension. If the MAPE of all dimensions is less than 8%, the mapping map is determined to meet the accuracy requirements for engineering applications, and a qualified model instance that can be used for online control is generated.

[0033] S3.5: The 'calorific value-process coupling mapping map' model that has passed accuracy verification is solidified into a deployable industrial control model file. Its network structure and training weights are encapsulated in ONNX format and embedded into the central controller of the thermal treatment system. As the core decision-making unit connecting the calorific value estimation results and the generation of control instructions, it realizes the functional closed loop from the input of multi-dimensional sludge characteristics to the recommended output of key control parameters, and provides a reliable mapping basis for the real-time parameter generation in the subsequent S4.

[0034] Furthermore, the ONNX (Open Neural Network Exchange) formatting and encapsulation method is used to achieve cross-platform deployment capability of deep neural network topology and corresponding weights, and to obtain model files that conform to the loading specifications of industrial control environments.

[0035] Furthermore, the ONNX file is embedded, loaded, and initialized in the central controller of the heat treatment system through a model interface adaptation algorithm, and an inference engine instance that can directly respond to feature vector input is generated.

[0036] Furthermore, a security verification mechanism is used to confirm the integrity and version consistency of the model files during the deployment process, and a qualified loading confirmation signal is generated.

[0037] In this embodiment, in step S4, the physicochemical indicators obtained from the rapid detection of the sludge to be treated in real time are input into the trained calorific value-process coupling mapping map to obtain the target set of feed rate, air supply, furnace temperature gradient and residence time under the corresponding operating conditions, such as... Figure 3 As shown, the specific steps include the following: S4.1: Acquire real-time rapid detection data of the dewatered sludge to be treated. The data includes moisture content, organic matter content, ash ratio, particle size distribution, and near-infrared spectral characteristics, which serve as input conditions for the calorific value-process coupling mapping spectrum. The above physicochemical indicators are collected by a multimodal sensor system deployed on-site in the industrial field, and the original signals are filtered, denoised, and baseline corrected to eliminate measurement deviations caused by environmental interference and generate standardized sludge feature vectors. The input conditions are the real-time rapid detection raw signals of the dewatered sludge to be treated at the industrial site, including moisture content, organic matter content, ash ratio, particle size distribution and near-infrared spectral characteristics, which are collected by a multimodal sensor system.

[0038] A multimodal sensor data fusion method is used to achieve synchronous acquisition and timestamp alignment of different types of signals such as moisture content, near-infrared spectrum and particle distribution.

[0039] Furthermore, noise suppression is achieved on the original electrical signals of moisture content and organic matter content through digital filtering algorithms, and preliminary smoothed physicochemical parameter time series data are obtained.

[0040] Furthermore, by using a baseline correction method, background drift of the near-infrared spectral signal is eliminated, and net spectral curve data of absorbance is generated.

[0041] Furthermore, the particle size distribution is accurately extracted using a standard particle size statistical algorithm, and particle size histogram vector data is generated.

[0042] The smoothed physicochemical parameters, organic matter content, ash ratio, and particle size distribution vector are transformed into dimensionless values ​​using the Z-score normalization method. The Z-score normalization formula is as follows:

[0043] in, These are the normalized sample values. These are the original sample values. The sample mean. To generate a standardized sludge feature vector with a mean of zero and a variance of one, we use the sample standard deviation.

[0044] S4.2: Perform dimension alignment on the generated standardized sludge feature vector to ensure that its input dimension is consistent with the input layer structure of the calorific value-process coupling mapping model; based on the preset feature mapping rules, replace missing or abnormal feature terms with sliding window mean interpolation results to ensure the integrity of input data and the stability of model inference, and output a normalized feature tensor that adapts to the model input interface; S4.3: Input the normalized feature tensor into the trained calorific value-process coupling map, which is a nonlinear function mapping model built based on a deep neural network, containing multiple fully connected hidden layers and nonlinear activation units; use the forward propagation algorithm to perform feature space transformation, calculate the output response of the hidden layer layer by layer, and finally generate an unnormalized process control parameter prediction value sequence in the output layer as the original control suggestion set.

[0045] The normalized feature tensor obtained by S4.2 dimension alignment and missing term imputation is input into the calorific value-process coupling mapping map model deployed in the central controller, and the nonlinear spatial mapping calculation of the input features is realized by the forward propagation algorithm of deep neural network.

[0046] Furthermore, through fully connected operations from the input layer to each hidden layer, the net input value of each hidden layer is calculated based on matrix multiplication and addition rules. The formula is as follows: , in This is the weight matrix. The current layer's input vector, The bias vector is used to input the resulting net input value into the nonlinear activation function. .

[0047] Furthermore, a nonlinear transformation is performed using the ReLU activation function, as shown in the following formula: , This enhances the model's ability to fit nonlinearly in a multidimensional feature space and generates the output response vectors of each hidden layer, providing input conditions for the next layer's computation.

[0048] Furthermore, by iteratively calculating layer by layer up to the output layer, a fully connected operation is used to generate an unnormalized prediction vector for the output layer. The formula is: ; in and These are the output layer weight matrix and bias vector, respectively. The output response of the final hidden layer is used to generate the original predicted values ​​of the process control parameters.

[0049] Furthermore, through batch tensor quantization, the above-mentioned output layer prediction values ​​are integrated into a four-dimensional process parameter prediction value sequence, which corresponds to the feed rate, air supply, furnace temperature gradient and residence time in sequence, forming the original control suggestion set.

[0050] S4.4: Perform inverse normalization on the unnormalized predicted value sequence of process control parameters, and restore the physical dimensions based on the maximum-minimum scaling parameters used in the training phase; perform feasibility verification and saturation correction on the restored values ​​according to the engineering constraint boundaries of each target parameter (such as the maximum feed rate limit and the minimum air supply threshold), and generate a set of target set values ​​for feed rate, air supply, furnace temperature gradient and residence time that conform to the operating specifications of the heat treatment equipment.

[0051] S4.5: Encapsulate the target setpoint set into a structured control data package, attach timestamps and operating condition tags, and temporarily store it in the real-time database as the input basis for generating the control instruction sequence in the next stage; at the same time, trigger the status synchronization mechanism to notify the monitoring system that the calorific value-driven process parameter recommendation has been completed, realizing information bridging and functional decoupling between the estimation module and the control module.

[0052] The target setpoints for feed rate, air supply, furnace temperature gradient, and residence time, which have been denormalized and verified by engineering constraints, are input into the control data encapsulation module as the original payload for data packet construction.

[0053] The structured data encapsulation method is used to embed the above four target settings into a JSON or binary structure according to a unified field format, and to add extended metadata such as batch identifiers and device numbers to form a parsable control data structure.

[0054] Furthermore, a timestamp field accurate to the millisecond level is added to the structured control data packet through a timestamp generation algorithm, and a label value reflecting the heat treatment process mode is added to the data packet in combination with the working condition label generation module to support the control timing matching in the subsequent state machine model.

[0055] Furthermore, the structured control data packets with added timestamps and operating condition tags are persistently stored in the specified tablespace of the real-time database through the real-time database temporary storage interface, and a return code reflecting the data entry status is generated.

[0056] The state synchronization mechanism triggers the monitoring of data packet entry events and sends a state synchronization message to the upper-level monitoring system after the event is triggered. The message body contains data packet summary information and entry confirmation identifier, realizing the result notification and functional decoupling between the estimation module and the control module.

[0057] In this embodiment, in step S5, a control command sequence is generated based on the target set of values ​​and sent to the actuator of the heat treatment system to drive each process unit to operate according to the recommended parameter combination, thereby realizing closed-loop process matching based on the estimated calorific value. Specifically, this includes the following steps: S5.1: Obtain the set of target set values ​​for feed rate, air supply, furnace temperature gradient and residence time output from the calorific value-process coupling mapping map, and use them as the input reference for generating control commands to ensure that subsequent control actions are adapted to the current calorific value characteristics of sludge; Based on the structured control data packets stored in the real-time database, a data parsing module is used to extract and convert the control parameters of feed rate, air supply, furnace temperature gradient and residence time.

[0058] The data integrity verification method is used to verify the time consistency and operating condition adaptability of the extracted parameter set, and to obtain the target set of values ​​that meet the time sequence and operating condition prerequisites.

[0059] A semantic binding algorithm is used to semantically link the current sludge calorific value feature field with the corresponding control parameter target value, forming a binding structure that can be used for control logic calls.

[0060] Furthermore, a high-precision numerical analysis method is used to convert the target setpoint in the binding structure into a numerical value in a unified dimensional format, thereby providing a directly calculable data foundation for the subsequent generation of control sequences.

[0061] S5.2: Perform unit normalization and range verification on the target set of values, determine whether the target parameters are within the safe operating range based on the physical control range of each actuator, and if there is an over-limit, perform boundary clamping correction to generate a compliant control parameter vector to ensure the safety of equipment operation. S5.3: Based on the compliant control parameter vector and combined with the control timing logic of each process unit of the heat treatment system, a sequence of control instructions with priority tags is generated using a preset state machine model. The sequence of control instructions includes the start-up order, adjustment step size and cooperative action flag, so as to realize coordinated control among multiple actuators. S5.4: The generated control command sequence is encapsulated into a standard control message through the industrial communication protocol and sent to the corresponding actuators via the PLC gateway module, including the variable frequency feed motor, the fan regulating valve, the burner temperature control unit and the rotary kiln drive device, so as to drive each process unit to start operation according to the recommended parameter combination; The control command sequence with priority markers generated by S5.3 is used as the encapsulation object, and the industrial communication protocol encapsulation method is used to convert the control parameter vector into a recognizable standard control message structure.

[0062] Furthermore, by using a message encoding algorithm, the values ​​of each control parameter in the control instruction sequence are filled into the data field in the order defined by the protocol, generating a binary frame structure with complete instruction semantics and physical quantity data, and obtaining a communication data packet that can be directly parsed by the PLC gateway module.

[0063] Furthermore, the checksum of the message is calculated using the CRC-16 redundancy check generation method, and the checksum is inserted into the end of the message to realize transmission integrity verification and generate a final encapsulated control message with error prevention capability.

[0064] Furthermore, the PLC gateway module's message routing mechanism distributes the encapsulated control messages to the corresponding hardware interface channels and executes the network layer handshake protocol to confirm the receiving status of the other end, ensuring that each actuator can receive and buffer control messages within a specified time window.

[0065] Furthermore, the parser module at the actuator end parses the received control messages according to the field order of the industrial communication protocol, and loads the feed rate command into the variable frequency feed motor drive unit, the air supply command into the fan regulating valve servo controller, the furnace temperature gradient command into the burner temperature control unit control board, and the residence time command into the internal control logic of the rotary kiln drive device, so as to realize that each process unit is put into operation according to the recommended parameter combination.

[0066] S5.5: After the command is issued, the feedback confirmation mechanism is activated to read the response signal returned by the actuator, verify the actual loading status of the control command sequence, and if an abnormal response is detected, an alarm is triggered and the previous stable operating condition is maintained to ensure the reliability and fault tolerance of the closed-loop control process.

[0067] In this embodiment, step S6, during the process execution, continuously collects actual operating data fed back by sensors, calculates the residual sequence between the actual operating parameters and the target set of values, and determines whether the residual sequence continuously exceeds a preset threshold range, specifically including the following steps: S6.1: Acquire real-time sensor feedback data from each actuator in the heat treatment system during operation. The sensor feedback data includes the actual measured value of the feed rate, the actual flow signal of the air supply, the multi-point temperature measurement sequence of the furnace temperature gradient, and the monitoring record of the sludge residence time, which serve as input conditions for evaluating control accuracy. Based on industrial communication protocols, periodically collect time series data of the above operating parameters from the distributed control system (DCS) or programmable logic controller (PLC) to generate an actual operating parameter vector, thereby constructing a basis for comparison with the target set of values. S6.2: Align the actual operating parameter vector with the target set generated in S5 dimension by dimension, calculate the instantaneous deviation between the two, and obtain a residual vector composed of feed rate residual, air supply residual, furnace temperature gradient residual, and residence time residual; perform a time-series splicing operation on the residual vectors in multiple consecutive sampling periods based on a sliding time window to form a residual sequence matrix to reflect the evolution trend of tracking error of process parameters under dynamic operation; S6.3: Perform statistical analysis on each parameter channel in the residual sequence matrix, calculate its mean and standard deviation within a preset time window, and set a dual threshold criterion in conjunction with the process safety boundary: when the residual mean of a certain parameter channel continuously exceeds ±10% of the set value or its fluctuation amplitude is greater than 3σ, mark the channel as entering an abnormal state; use Boolean logic to combine the judgment results of each channel to generate a comprehensive deviation flag bit, which serves as an intermediate judgment basis for whether it deviates from the normal control range; The input condition is the residual sequence matrix generated by S6.2, which contains continuous sampling data of the feed rate residual channel, air supply residual channel, furnace temperature gradient residual channel and residence time residual channel.

[0068] A sliding window statistical analysis method is used to quantify the dynamic characteristics of each parameter channel in the residual sequence matrix.

[0069] Furthermore, the average deviation of each channel within the window is calculated using the mean calculation formula:

[0070] in, The mean value reflects the central tendency of the residuals; The total number of samples within the window. These are the residual sample values.

[0071] Furthermore, based on a dual-threshold criterion preset within the process safety boundary, abnormal state identification is achieved, specifically: mean threshold... Set value and fluctuation range threshold When the mean continuously exceeds the set value and the standard deviation exceeds the threshold, the corresponding channel is marked as an abnormal state.

[0072] Furthermore, a Boolean logic combination algorithm is used to jointly determine cross-channel abnormal states and generate a comprehensive deviation flag, which serves as an intermediate basis for determining whether the current operating condition deviates from the normal control range.

[0073] S6.4: Perform a timing consistency check based on the comprehensive deviation flag bit, and use a finite state machine model to track the duration of its continuous true state; if the state continues for more than three complete control cycles (i.e. the residual sequence continuously exceeds the preset threshold range), it is determined that the target set value set cannot effectively guide the actual operation and there is a risk of model mismatch; output the deviation persistence judgment signal as the start condition for triggering the online learning mechanism in S7. S6.5: Package the residual sequence matrix and its corresponding input sludge feature vector and target set value set into a monitoring log entry and store it in the system diagnostic database; at the same time, push early warning information to the human-machine interface (HMI) to prompt operators to pay attention to the abnormal situation of the current working condition, and provide data traceability support for subsequent model fine-tuning and manual intervention.

[0074] In this embodiment, in step S7, if it is determined that the residual sequence continuously exceeds the preset threshold range, an online learning mechanism is triggered. The network weights of the calorific value-process coupling mapping spectrum are locally fine-tuned using the latest input-output data to generate an updated mapping spectrum to adapt to the current operating condition deviation. Specifically, the following steps are included: S7.1: Based on the judgment result that the residual sequence continuously exceeds the preset threshold range, an online learning trigger signal is generated as a control command to start the model update mechanism, indicating that the current calorific value-process coupling mapping spectrum can no longer accurately reflect the actual operating condition response characteristics, and new samples need to be introduced for incremental optimization. S7.2: Obtain the standardized sludge feature vector of the latest batch of sludge to be treated and its corresponding actual process execution parameter feedback data to form a new input-output training sample pair, and perform consistency verification and anomaly detection on the sample pair to prevent noisy data or sensor failures from interfering with the model update process, and generate a reliable incremental training set. S7.3: Input the reliable incremental training set into the current version of the thermal value-process coupling mapping map, calculate the gradient of the loss function based on the backpropagation algorithm of the deep neural network, identify the source of deviation between the predicted and measured values ​​of the output control parameters, and generate the corrected gradient vector of the connection weights of each layer as the basis for weight adjustment. A reliable incremental training set is loaded into the input interface of the current version of the calorific value-process coupling map model. Batch tensor data reading is used to achieve dimensionality matching with the neuron structure of the model input layer. The activation response of the standardized sludge feature vector is calculated layer by layer in the network through the forward propagation algorithm, and the predicted normalized process control parameter vector is output.

[0075] Furthermore, the backpropagation algorithm of the deep neural network is invoked to calculate the partial derivatives of each connection weight sequentially from the output layer to the input layer. The gradient propagation is then carried out using the chain rule to generate the set of gradient tensors corresponding to the weight matrix of each layer.

[0076] Furthermore, for the gradient tensor of the connection weights of each layer, the weight correction vector is calculated in combination with the current model parameter values ​​and the learning rate to ensure that the correction amount meets the magnitude constraint preset by the local update strategy, so as to prevent excessive perturbation to the stable feature extraction layer.

[0077] S7.4: The mini-batch gradient descent method is used to perform local fine-tuning of the weights of the last three hidden layers of the calorific value-process coupling mapping map, and the weight parameters of the input layer and the first few feature extraction layers are frozen to retain the stable cross-condition generalization ability of the original model. Only the fine mapping ability for the current operating conditions is optimized to generate a locally updated deep neural network model.

[0078] S7.5: Perform inference verification on the locally updated deep neural network model, input multiple historical test samples and compare the consistency of its output parameters with the target set of values. If the average residual decreases and the convergence criterion is met, the updated model is deployed as a new calorific value-process coupling mapping map to complete the online learning cycle and improve the system's adaptability to unsteady conditions.

[0079] In this embodiment, step S8 involves using the SHAP value analysis method to quantify the contribution of each input feature in the updated calorific value-process coupling mapping to the output control parameters, generating an interpretability report to assist operators in understanding the control logic and verifying the consistency of model behavior. Specifically, this includes the following steps: S8.1: Based on the updated calorific value-process coupling mapping model structure, obtain its predicted output of standardized sludge feature vectors under the current operating conditions, and use it as a benchmark model instance for SHAP value analysis to establish the technical basis for interpretable analysis.

[0080] S8.2: Construct a background dataset by randomly sampling multiple sets of labeled standardized sludge feature vectors and their corresponding normalized process control parameter vectors from the historical operation database to form a reference input set for SHAP value calculation, ensuring that the interpretation results cover typical operating conditions.

[0081] S8.3: The KernelSHAP algorithm is used to perform a local approximate interpretation of the calorific value-process coupling mapping model. Based on the prediction difference between the reference input set and the sample to be interpreted, the SHAP value of each input feature (moisture content, organic matter content, ash ratio, particle size distribution) is calculated for each output control parameter (feed rate, air supply, furnace temperature gradient, residence time) to obtain the feature-level contribution quantification index.

[0082] S8.4: Aggregate the calculated SHAP values ​​according to the output control parameter dimension to generate a ranking list of key influencing factors for each process parameter, and construct a global feature importance map by weighting by absolute mean to form an overall cognitive framework for the model decision-making mechanism.

[0083] S8.5: Generate interpretability reports based on SHAP value visualization rules, including force plots for individual samples and beeswarm plots for population data, and push the reports to the human-computer interaction interface for operators to verify whether the model behavior conforms to prior process knowledge, thereby achieving traceability and consistency verification of control logic.

[0084] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0085] Unless otherwise defined, the technical or scientific terms used herein should be understood in their ordinary sense by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” involved in the embodiments of this invention refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0086] The above description is merely an exemplary embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for matching and controlling thermal treatment processes based on the estimation of the calorific value of dewatered sludge, characterized in that, Specifically, it includes: S1: Collect physicochemical property data of dewatered sludge and simultaneously record the optimal combination of operating parameters to obtain an input-output paired sample set; S2: Preprocess the input-output paired sample set; S3: Use a deep neural network model to train and generate a calorific value-process coupling mapping map; S4: Input the physicochemical indicators obtained from the rapid detection of the sludge to be treated in real time into the trained calorific value-process coupling mapping map to obtain the target set of feed rate, air supply, furnace temperature gradient and residence time under the corresponding working conditions. S5: Generates a sequence of control instructions based on the target setpoints and sends them to the actuators of the heat treatment system to drive each process unit to operate according to the recommended parameter combination; S6: During the process execution, continuously collect the actual operating data fed back by the sensor, calculate the residual sequence between the actual operating parameters and the target set value set, and determine whether the residual sequence continuously exceeds the preset threshold range. S7: If it is determined that the residual sequence continuously exceeds the preset threshold range, the online learning mechanism is triggered. The network weights of the calorific value-process coupling mapping map are locally fine-tuned using the latest input-output data to generate an updated mapping map to adapt to the current operating condition deviation. S8: Combine the SHAP value analysis method to quantitatively evaluate the contribution of each input feature in the updated calorific value-process coupling mapping to the output control parameters.

2. The thermal treatment process matching and control method and system based on the estimation of the calorific value of dewatered sludge as described in claim 1, characterized in that, In step S1, the original physicochemical property data are spatiotemporally aligned with the optimal operating parameter combinations obtained in rotary kiln drying, fluidized bed incineration and pyrolysis carbonization to establish input-output pairing relationships, and operating condition labels are attached to each set of data to generate an input-output pairing sample set with operating condition labels.

3. The thermal treatment process matching and control method and system based on the estimation of the calorific value of dewatered sludge as described in claim 1, characterized in that, In step S2, the preprocessing includes outlier removal, missing value imputation, and feature standardization to generate a standardized sludge feature vector and a normalized process control parameter vector.

4. The method and system for matching and controlling thermal treatment processes based on the estimation of the calorific value of dewatered sludge as described in claim 1, characterized in that, In step S3, a training dataset is constructed based on the standardized sludge feature vector and the normalized process control parameter vector. The input variables are moisture content, organic matter content, ash ratio, particle size distribution and measured lower heating value, and the output variables are feed rate, air supply, furnace temperature gradient and residence time. Design a feedforward deep neural network model architecture with a multi-hidden-layer structure, wherein the number of neurons in the input layer matches the standardized sludge feature vector, the number of neurons in the output layer corresponds to the normalized process control parameter vector, and no less than three hidden layers are set in the middle and ReLU activation function is adopted. The initial connection weights are configured through the Xavier initialization method to build a basic calculation framework for the calorific value-process coupling mapping graph that can fit complex nonlinear mapping relationships. A batch training process is performed on the deep neural network model. The mean squared error loss function is used to quantify the deviation between the predicted control parameter vector and the actual normalized process parameter vector. The backpropagation algorithm is implemented based on the Adam optimizer to perform iterative weight updates.

5. The thermal treatment process matching and control method and system based on the estimation of the calorific value of dewatered sludge as described in claim 1, characterized in that, In step S4, real-time rapid detection data of the dewatered sludge to be treated is obtained. The data includes moisture content, organic matter content, ash ratio, particle size distribution and near-infrared spectral characteristics. The original signal is filtered and denoised and baseline corrected to generate a standardized sludge feature vector.

6. The thermal treatment process matching and control method and system based on the estimation of the calorific value of dewatered sludge as described in claim 5, characterized in that, In step S4, a dimension alignment operation is performed on the standardized sludge feature vector to make its input dimension consistent with the input layer structure of the calorific value-process coupling mapping model; based on the preset feature mapping rules, missing or abnormal feature terms are replaced with sliding window mean interpolation results, and a normalized feature tensor is output.

7. The thermal treatment process matching and control method and system based on the estimation of the calorific value of dewatered sludge as described in claim 6, characterized in that, In step S4, the normalized feature tensor is input into the trained calorific value-process coupling map, which is a nonlinear function mapping model built based on a deep neural network, containing multiple fully connected hidden layers and nonlinear activation units; the feature space transformation is performed using the forward propagation algorithm, the output response of the hidden layer is calculated layer by layer, and finally an unnormalized process control parameter prediction value sequence is generated in the output layer. The predicted value sequence is subjected to inverse normalization, and the physical dimensions are restored based on the maximum-minimum scaling parameters used in the training phase. The restored values ​​are then subjected to feasibility verification and saturation correction according to the engineering constraint boundaries of each target parameter, generating a set of target set values ​​for feed rate, air supply, furnace temperature gradient and residence time that conform to the operating specifications of the heat treatment equipment.

8. The method and system for matching and controlling thermal treatment processes based on the estimation of the calorific value of dewatered sludge as described in claim 1, characterized in that, In step S5, the target set value set is normalized and range is checked. Based on the physical control range of each actuator, it is determined whether the target parameters are within the safe operating range. If there is an over-limit, boundary clamping correction is performed to generate a compliant control parameter vector.

9. The thermal treatment process matching and control method and system based on the estimation of the calorific value of dewatered sludge as described in claim 8, characterized in that, In step S5, based on the compliant control parameter vector and combined with the control timing logic of each process unit of the heat treatment system, a sequence of control instructions with priority tags is generated using a preset state machine model to achieve coordinated control among multiple actuators.

10. The method and system for matching and controlling thermal treatment processes based on the estimation of the calorific value of dewatered sludge as described in claim 9, characterized in that, In step S5, the control command sequence is encapsulated and sent to the corresponding actuator to drive each process unit to start operation according to the recommended parameter combination; After the command is issued, a feedback confirmation mechanism is activated to read the response signal returned by the actuator, verify the actual loading status of the control command sequence, and trigger an alarm if an abnormal response is detected and maintain the previous stable operating condition.

Images