Straw resource processing method based on intelligent sensing

Abstract

The application relates to the technical field of agricultural intelligent equipment, in particular to a straw resource processing method based on intelligent sensing, which comprises the following steps: collecting raw material characteristic data, operation environment data and soil basic parameter data of to-be-processed straw; performing feature fusion processing on the raw material characteristic data, the operation environment data and the soil basic parameter data to construct a straw basic state vector; inputting the straw basic state vector into a pre-trained straw biological conversion pre-response model to obtain a pre-response characteristic parameter set; generating a processing control parameter set through a pre-trained process execution mapping model based on the pre-response characteristic parameter set, and controlling a straw resource processing execution equipment based on the processing control parameter set; the processing control parameter set comprises flame consumption nozzle control parameters and plowing equipment control parameters.

Description

Technical Field

[0001] This application relates to the field of agricultural intelligent equipment technology, and in particular to a method for the resource utilization of straw based on intelligent sensing. Background Technology

[0003] Current straw processing methods typically rely on experience to set operational parameters, such as combustion intensity, flame spray conditions, and tillage depth. They lack the comprehensive perception and dynamic analysis capabilities regarding straw raw material characteristics, operating environment, and soil conditions. This makes it difficult for the processing to adaptively adjust to different scenarios, thus affecting the consistency and stability of straw processing results. Furthermore, traditional control methods are mostly based on single parameters or fixed rules for equipment control, making it difficult to achieve optimal parameter matching under multi-factor coupling conditions. Especially in complex agricultural environments, there is a clear synergistic relationship between combustion and tillage operations. Without a unified intelligent decision-making mechanism, problems such as low energy utilization efficiency or excessive soil disturbance can easily occur. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides a straw resource utilization method based on intelligent sensing, which solves the technical problems in the prior art that lack the ability to fuse and analyze multi-source information on straw raw material characteristics, working environment and soil conditions, making it difficult to achieve adaptive prediction and collaborative optimization control of straw burning and tillage parameters, thus leading to the dependence of working parameters on experience setting, low control accuracy and unstable treatment effect in the straw resource utilization process.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the main technical solution adopted in this application includes a straw resource utilization method based on intelligent sensing. The method includes: S1, collecting raw material characteristic data, operating environment data, and soil basic parameter data of the straw to be processed; S2, performing feature fusion processing on the raw material characteristic data, operating environment data, and soil basic parameter data to construct a straw basic state vector; wherein the straw basic state vector includes a straw physical structure feature sub-vector corresponding to the raw material characteristic data, an environmental state feature sub-vector corresponding to the operating environment data, and a soil state feature sub-vector corresponding to the soil basic parameter data; S3, inputting the straw basic state vector into a pre-trained straw biotransformation pre-response model to obtain a pre-response feature parameter set; wherein the straw biotransformation... The conversion pre-response model is trained based on a straw basic state vector training dataset; the pre-response feature parameter set includes target values ​​for straw combustion temperature and tillage depth; the straw basic state vector training dataset includes historical straw basic state vectors and corresponding pre-set straw combustion temperature and tillage depth; S4, based on the pre-response feature parameter set, a processing control parameter set is generated through a pre-trained process execution mapping model, and the straw resource utilization processing equipment is controlled based on the processing control parameter set; wherein, the processing control parameter set includes flame scavenging nozzle control parameters and tillage equipment control parameters; the flame scavenging nozzle control parameters include injection temperature, ground clearance, fuel flow rate, and nozzle angle; the tillage equipment control parameters include soil penetration depth and crushing intensity;

[0008] The process execution mapping model is a multivariable nonlinear parameter mapping model based on a multilayer perceptron.

[0009] (III) Beneficial Effects

[0010] The intelligent sensing-based straw resource utilization method provided in this application collects raw material characteristic data, operational environment data, and basic soil parameter data of the straw to be processed. It then constructs a basic straw state vector containing sub-vectors of straw physical structure characteristics, environmental state characteristics, and soil state characteristics. This achieves unified representation and fusion analysis of multi-source heterogeneous information in the straw processing scenario, providing fundamental data support for intelligent decision-making on subsequent processing parameters. Furthermore, by inputting the straw basic state vector into a straw bioconversion pre-response model, this application obtains a pre-response feature parameter set containing target values ​​for straw combustion temperature and tillage depth. This allows straw combustion processing parameters and tillage operation parameters to be jointly predicted within the same model framework, thereby achieving collaborative modeling and linkage optimization between the combustion and tillage processes. This effectively avoids the control fragmentation problem caused by independently setting combustion and tillage parameters in traditional technologies. Furthermore, this application generates a set of processing control parameters using a process execution mapping model based on a pre-response characteristic parameter set. It then performs unified mapping control on the control parameters of the flame-absorbing nozzle and the tillage equipment, establishing a coordinated relationship between injection temperature, ground clearance, fuel flow rate, nozzle angle, soil penetration depth, and crushing intensity. This achieves synergistic control of flame-absorbing treatment and mechanical tillage, improving the completeness of straw combustion and the uniformity of tillage. Simultaneously, this application models the complex nonlinear relationships in the straw treatment process using a multi-layer perceptron multivariate nonlinear parameter mapping model. This allows the target values ​​for combustion temperature and tillage depth to be adaptively adjusted under the drive of multi-source sensing data, thereby improving the control accuracy and environmental adaptability of the straw resource utilization process and reducing fluctuations in treatment effects caused by empirically set operating parameters. Therefore, this application can achieve synergistic optimization control of straw combustion parameters and tillage parameters in the straw resource utilization process, improving the intelligence level, operational stability, and resource utilization efficiency of the process, and enhancing adaptability to different operating environments and soil conditions. Attached Figure Description

[0011] Figure 1 This is a schematic flowchart of a straw resource utilization method based on intelligent sensing according to an embodiment of this application;

[0012] Figure 2 This is a flowchart illustrating a straw resource utilization method based on intelligent sensing according to another embodiment of this application. Detailed Implementation

[0013] To better explain and facilitate understanding of this application, the following detailed description of the application is provided in conjunction with the accompanying drawings and specific embodiments.

[0014] In related technologies, the control of operational parameters during straw resource utilization is mainly based on experience-based straw treatment control schemes. These schemes typically involve independently setting key parameters such as straw combustion temperature, flame injection conditions, and tillage depth based on human experience, and then separately driving the flame suppression device and the tillage and covering mechanism during operation. However, this scheme lacks a linkage mechanism between the straw combustion process and the tillage process. The combustion and tillage parameters are adjusted independently, lacking coordinated control and linkage optimization capabilities. This can easily lead to a mismatch between combustion intensity and tillage depth, resulting in incomplete straw combustion or excessive soil disturbance, affecting the consistency and stability of the overall treatment effect.

[0015] To address this, the intelligent sensing-based straw resource utilization method provided in this application collects straw raw material characteristic data, operational environment data, and soil basic parameter data to construct a basic straw state vector. Within a unified state representation space, it integrates and models the influencing factors related to straw combustion and tillage. Furthermore, this application synchronously outputs target values ​​for straw combustion temperature and tillage depth through a straw bioconversion pre-response model, enabling the joint generation of combustion and tillage parameters within the same prediction model. A process execution mapping model maps these parameters to flame scavenging nozzle control parameters and tillage equipment control parameters, thereby achieving coordinated control of injection temperature, ground clearance, fuel flow rate, nozzle angle, soil penetration depth, and crushing intensity. Therefore, this application effectively solves the problem of independently set and uncoordinated control of straw combustion and tillage parameters in existing technologies, achieving coordinated optimization control of combustion and tillage operations, and improving the overall consistency, operational stability, and processing efficiency of straw resource utilization.

[0016] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0017] Example 1

[0018] Figure 1 This is a schematic flowchart of a straw resource utilization method based on intelligent sensing according to an embodiment of this application. Figure 1As shown, the intelligent sensing-based straw resource utilization method includes: S1, collecting raw material characteristic data, operating environment data, and soil basic parameter data of the straw to be processed; wherein, the raw material characteristic data includes straw type, straw moisture content, fiber structure parameters, lignin content, etc.; the operating environment data includes ambient temperature, ambient humidity, oxygen concentration, and microbial activity parameters; the soil basic parameter data includes soil hardness, soil moisture content, and soil density;

[0019] S2. Perform feature fusion processing on the raw material characteristic data, working environment data, and soil basic parameter data to construct a straw basic state vector; wherein the straw basic state vector includes a straw physical structure feature sub-vector corresponding to the raw material characteristic data, an environmental state feature sub-vector corresponding to the working environment data, and a soil state feature sub-vector corresponding to the soil basic parameter data.

[0020] Optionally, in some embodiments of this application, step S2 includes:

[0021] S21. Normalize and scale the raw material characteristic data, working environment data and soil basic parameter data, and construct straw physical structure feature sub-vector, environmental state feature sub-vector and soil state feature sub-vector respectively based on preset feature coding rules.

[0022] S22. The straw physical structure feature vector, environmental state feature vector, and soil state feature vector are fused to generate the straw basic state vector.

[0023] For example, the feature fusion processing of raw material characteristic data, working environment data, and soil basic parameter data in step S2 is as follows: First, the collected raw material characteristic data, such as straw type, straw moisture content, fiber structure parameters, and lignin content, are standardized. For example, the straw type is encoded and converted, continuous parameters such as straw moisture content and lignin content are normalized, and fiber structure parameters are converted into structural strength indicators or fiber density indicators, thereby forming a straw physical structure feature sub-vector. For example, for a corn straw sample, its straw moisture content is 35%, its lignin content is 18%, and the tensile strength corresponding to the fiber structure parameters is at a medium level, which can be mapped to the first feature sub-vector represented as X1. Second, the working environment data, such as ambient temperature, ambient humidity, oxygen concentration, and microbial activity parameters, are feature-processed. For example, ambient temperature and ambient humidity are jointly normalized, oxygen concentration is converted into an oxidation reaction influence coefficient, and microbial activity parameters are converted into a decomposition activity index, thereby forming an environmental state feature sub-vector. For example, in an operating environment with an ambient temperature of 28℃, humidity of 65%, oxygen concentration of 21%, and high microbial activity, this can be represented as the second feature vector X2. Further, the soil basic parameter data, such as soil hardness, soil moisture content, and soil density, are structured. For example, soil hardness is converted into a mechanical resistance coefficient, soil moisture content is normalized, and soil density is converted into a compaction degree index, thus forming a soil state feature vector. For example, for medium-hard soil, which has high hardness, a moisture content of approximately 25%, and high density, it can be represented as the third feature vector X3. After completing the above processing, the straw physical structure feature vector X1, the environmental state feature vector X2, and the soil state feature vector X3 are spliced ​​or weighted according to a preset fusion rule to generate the straw basic state vector X. The preset fusion rules may include vector-level concatenation, weighted summation, or linear transformation based on the feature weight matrix, so that the straw basic state vector can simultaneously represent the characteristics of straw raw materials, working environment conditions, and soil basic state, providing a unified data expression basis for the input of the subsequent straw bioconversion pre-response model.

[0024] In one optional implementation, the preset feature encoding rules are used to uniformly quantify the collected raw heterogeneous data, converting different types of data into feature vector expressions that can be used for subsequent model calculations. Specifically, for numerical data, such as parameters like straw moisture content, ambient temperature, and soil hardness, normalization or standardization methods are used to map them to a unified numerical range to eliminate differences between different physical dimensions. For categorical data, such as parameters like straw type, one-heat encoding or vector embedding is used for numerical representation to form categorical feature vectors with discriminative capabilities. For structural or mechanism-related parameters, such as fiber structure parameters, lignin content, and microbial activity, a preset mapping function is used to convert them into structural strength feature values ​​or process influence coefficients based on their influence on the straw decomposition and combustion process, to characterize their effect on subsequent processing. Furthermore, for environmental and soil-related physical parameters, such as oxygen concentration, soil density, and soil moisture content, preset feature functions can be used to convert them into environmental impact factors or soil action weight parameters to reflect the influence of different operating conditions on the straw combustion and tillage process. Through the above encoding process, the original multi-source heterogeneous data are uniformly mapped to the same feature space, thereby forming a feature expression with consistent structure and scale, providing a data foundation for the subsequent construction of straw basic state vectors and model calculation.

[0025] S3. Input the straw basic state vector into the pre-trained straw bioconversion pre-response model to obtain a pre-response feature parameter set; wherein, the straw bioconversion pre-response model is trained based on the straw basic state vector training dataset; the pre-response feature parameter set includes the target value of straw combustion temperature and the target value of tillage depth;

[0026] The training dataset for straw basic state vectors includes historical straw basic state vectors and corresponding pre-set straw burning temperatures and tillage depths.

[0027] For example, during historical operations, a basic state vector of straw collected at a certain time point might contain information such as corn straw type, high moisture content, high ambient humidity, and moderately hard soil. The corresponding historical control parameters would be a combustion temperature setpoint T1 of 750℃ and a tillage depth setpoint D1 of 18cm. For a basic state vector of wheat straw collected at another time point, with lower straw moisture content, higher ambient temperature, and looser soil, the corresponding historical combustion temperature setpoint T2 would be 680℃ and a tillage depth D2 of 15cm. During model training, these historical straw basic state vectors are used as input samples, and the corresponding combustion temperature and tillage depth setpoints are used as supervised learning labels. These are then input into the straw bioconversion pre-response model for parameter optimization. This allows the model to learn the nonlinear mapping relationship between straw raw material characteristics, operational environmental conditions, soil basic parameters, and target treatment process parameters, thereby forming a predictive model structure. In practical applications, when a new basic state vector of straw is input, such as the current straw type being rice straw, moisture content being 30%, ambient temperature being 26℃, and soil density being relatively high, the straw bioconversion pre-response model performs forward calculations on it and outputs the corresponding pre-response feature parameter set, including the target value T of straw combustion temperature. new and the target value of tillage depth D new For example, output T respectively new 720℃, D new The value is 17cm. Through the above method, the straw bioconversion pre-response model can jointly predict the target values ​​of straw combustion temperature and tillage depth under different operating conditions, based on the learning results between the historical straw basic state vector and the corresponding process parameters.

[0028] S4. Based on the pre-response feature parameter set, a processing control parameter set is generated through a pre-trained process execution mapping model, and the straw resource utilization processing equipment is controlled based on the processing control parameter set; wherein, the processing control parameter set includes flame scavenging nozzle control parameters and tillage equipment control parameters; the flame scavenging nozzle control parameters include injection temperature, ground clearance, fuel flow rate, and nozzle angle; the tillage equipment control parameters include soil penetration depth and crushing intensity; the process execution mapping model is a multivariate nonlinear parameter mapping model based on a multilayer perceptron.

[0029] For example, when the pre-response feature parameter set outputs a target value of 720℃ for straw combustion temperature and 17cm for tillage depth, the process execution mapping model first performs feature analysis and standardization on the above target parameters to form a process control feature vector, which is then input into a multilayer perceptron network for nonlinear mapping calculation, thereby generating a corresponding equipment operating condition control vector. In a specific embodiment, for the target value of 720℃ for straw combustion temperature, the process execution mapping model outputs flame scavenging nozzle control parameters with the injection temperature set, for example, between 700℃ and 750℃, a ground clearance set, for example, 0.8m, a fuel flow rate set, for example, at a medium-high output level, and a nozzle angle set, for example, at a tilted injection angle of 45° to 60°, to ensure uniform flame coverage and complete combustion; for the target value of 17cm for tillage depth, the model outputs tillage equipment control parameters with the soil penetration depth set, for example, at 17cm, and generates corresponding crushing strength control parameters based on soil resistance, for example, set to a medium-high crushing strength, to improve the straw burial effect and soil mixing uniformity. Furthermore, the processing control parameter set, after being generated, is directly used to drive the flame extinguishing nozzle and the tillage and covering mechanism to perform operation control, so that the flame extinguishing device and the tillage equipment can be coordinated and linked under the constraints of a unified process objective, thereby achieving synchronous optimization between straw combustion treatment and mechanical tillage treatment.

[0030] By employing the above methods, the pre-response characteristic parameter set can be effectively mapped into executable equipment control parameters, realizing a nonlinear transformation from the process target layer to the equipment execution layer, thereby improving the automation control accuracy and synergy of the straw resource utilization process.

[0031] Optionally, in some embodiments of this application, the straw bioconversion pre-response model includes: a feature mapping module, a state fusion module, and a response prediction module;

[0032] The feature mapping module is used to perform nonlinear feature transformation processing on each feature sub-vector to obtain the corresponding embedded feature representation; the feature mapping module includes a multi-branch feature embedding network structure, which includes a first feature mapping branch corresponding to the straw physical structure feature sub-vector, a second feature mapping branch corresponding to the environmental state feature sub-vector, and a third feature mapping branch corresponding to the soil state feature sub-vector.

[0033] The first feature mapping branch includes a fully connected mapping layer, a feature recalibration layer, and a nonlinear activation layer arranged sequentially.

[0034] The fully connected mapping layer is used to perform high-dimensional space expansion processing on the input straw physical structure feature sub-vectors to generate an initial structure embedding representation;

[0035] The feature recalibration layer is used to perform weight allocation processing on the initial structural embedding representation based on the channel attention mechanism to generate a structural recalibrated embedding feature representation;

[0036] The nonlinear activation layer is used to perform nonlinear mapping transformation on the structural recalibrated embedded feature representation and output the structural nonlinear embedded feature representation.

[0037] In this embodiment, the first feature mapping branch first performs high-dimensional space expansion processing on the original features such as straw moisture content, fiber structure parameters, and lignin content through a fully connected mapping layer, so that the low-dimensional physical parameters are mapped to the high-dimensional semantic space to enhance their separability. Then, the feature recalibration layer performs adaptive weight allocation on different structural features based on the channel attention mechanism, so that features that have a greater impact on combustion resistance and decomposition difficulty (such as lignin content and fiber density) receive higher response weights, thereby enhancing key structural information. Finally, the recalibrated features are nonlinearly transformed through a nonlinear activation layer, so that the structural features form a stable nonlinear expression distribution in the high-dimensional space, thereby improving the contribution of straw physical structure to the prediction of combustion temperature target value.

[0038] The second feature mapping branch includes an environmental feature encoding layer, a gated feature selection unit, and a nonlinear transformation layer;

[0039] The environmental feature encoding layer is used to perform high-dimensional feature encoding on the environmental state feature sub-vectors to generate a basic environmental embedding representation.

[0040] The gated feature selection unit is used to perform information filtering and enhancement processing on the environmental basic embedding representation based on feature relevance weights to generate an environmental filtering embedding representation.

[0041] The nonlinear transformation layer is used to perform nonlinear mapping processing on the filtered environmental embedding features to obtain an environmental nonlinear embedding feature representation.

[0042] In this embodiment, the second feature mapping branch performs unified high-dimensional encoding of environmental temperature, humidity, oxygen concentration, and microbial activity parameters through an environmental feature encoding layer, transforming multivariate environmental information into a continuous and learnable representation. Subsequently, a gated feature selection unit dynamically filters and suppresses the environmental embedding representation based on feature relevance weights to weaken the adverse effects of environmental noise (such as short-term temperature fluctuations or local humidity anomalies) on the model output, while strengthening environmental factors that play a dominant role in straw decomposition activity and combustion efficiency. Then, a nonlinear transformation layer further enhances the expression complexity of environmental features, enabling environmental factors to participate in subsequent fusion calculations in a nonlinear manner, thereby improving the model's robustness to complex climatic conditions.

[0043] The third feature mapping branch includes a graph structure encoding layer and a neighborhood feature aggregation layer;

[0044] The graph structure encoding layer is used to perform structured encoding processing on soil state feature subvectors, convert them into node feature representations, and construct a soil state association graph based on the correlation between soil physical properties to form soil node feature representations and corresponding graph structure topology relationships.

[0045] The neighborhood feature aggregation layer is used to perform weighted aggregation calculation on the soil node feature representation based on preset association rules, so as to fuse the feature information between mutually related nodes and generate soil association embedded feature representation;

[0046] Specifically, the third feature mapping branch converts parameters such as soil hardness, soil moisture content, and soil density into node feature representations through a graph structure encoding layer, and constructs a soil state association graph based on the coupling relationship between soil physical properties, so that the spatial and mechanical relationship between different soil parameters is explicitly and structurally expressed. Subsequently, the neighborhood feature aggregation layer performs weighted aggregation of adjacent node features based on preset association rules, so that the structural consistency and non-uniformity differences of local soil areas can be effectively modeled, thereby avoiding the structural information loss problem caused by the "independent processing" of soil information in traditional vector representation, and enabling the model to more accurately reflect the resistance changes and soil disturbance characteristics during tillage operations.

[0047] Through the aforementioned multi-branch feature mapping structure, this application achieves a decoupled modeling and hierarchical representation mechanism for straw physical structure features, environmental state features, and soil state features, avoiding mutual interference between different types of features in a unified mapping space, thereby significantly improving the stability and discriminative ability of feature representation. Simultaneously, by introducing channel attention mechanism, gating selection mechanism, and graph structure neighborhood aggregation mechanism, various features complete information enhancement and noise suppression within their respective optimal structural paths, significantly improving the model's sensitivity to key influencing factors.

[0048] In one optional implementation, the calculation process of the first feature mapping branch, the second feature mapping branch, and the third feature mapping branch is explained in conjunction with a specific operational scenario: For example, in a certain straw resource utilization operation, the input straw physical structure feature sub-vector includes straw type as corn straw, moisture content as 32%, fiber structure strength as medium to high, and lignin content as 18%; environmental state feature sub-vector includes environmental temperature as 28℃, environmental humidity as 70%, oxygen concentration as 21%, and high microbial activity; soil state feature sub-vector includes high soil hardness, soil moisture content as 25%, and high soil density.

[0049] In the first feature mapping branch, the straw physical structure feature vectors are first extended to a higher dimension through a two-layer cascaded fully connected mapping structure. The first layer maps the original structural features to a 128-dimensional feature space to enhance basic expressive power, and the second layer further maps them to a 256-dimensional space, thereby generating an initial structural embedding representation. This allows factors such as straw moisture content, fiber structure, and lignin to form separable feature distributions in the high-dimensional space. Subsequently, in the feature recalibration layer, global average pooling is used to perform global feature compression on the initial structural embedding representation to obtain overall structural statistical information. This information is then input into the two fully connected sub-networks (128-dimensional and 256-dimensional) to generate channel weight coefficients. These weight coefficients are normalized to the range of 0 to 1 using the Sigmoid function. For example, a weight of 0.85 is assigned to the lignin-related channel, and a weight of 0.65 is assigned to the moisture content-related channel. This allows for channel-by-channel weighting of the initial embedding features, resulting in a structurally recalibrated embedding feature representation. Finally, the recalibrated features are subjected to element-wise nonlinear transformation using ReLU or LeakyReLU activation functions, making the structural feature representation sparser and more discriminative, thereby outputting a structural nonlinear embedding feature representation.

[0050] In the second feature mapping branch, environmental state features such as temperature, humidity, and oxygen concentration are first encoded using a two-layer fully connected encoding structure (64-dimensional and 128-dimensional), converting them into environmental basic embedding representations, allowing environmental factors to be expressed in a unified feature space. Subsequently, in the gated feature selection unit, a gating coefficient consistent with the embedding dimension is generated using the Sigmoid function. For example, a higher gating value of 0.9 is assigned to high humidity conditions, and 0.7 is assigned to oxygen concentration fluctuations. The environmental embedding representation is filtered and enhanced through element-wise multiplication, thereby suppressing the impact of short-term environmental noise and highlighting the role of key environmental factors, resulting in a filtered environmental embedding representation. Finally, the filtered features are nonlinearly mapped using the Tanh or GELU activation function to stabilize their distribution in a unified expression space, outputting a nonlinear environmental embedding feature representation.

[0051] In the third feature mapping branch, soil hardness, moisture content, and density parameters are first converted into node feature representations through a two-layer graph structure coding network (64-dimensional and 128-dimensional), and a soil state association graph is constructed. For example, the high correlation between soil hardness and density is modeled as an adjacency edge relationship, thus forming a graph model representation with structural topological relationships. Subsequently, in the neighborhood feature aggregation layer, the influence weights between adjacent nodes are normalized through the Softmax function. For example, nodes in high-density areas are given higher normalization weights. Then, the neighborhood node features are aggregated and calculated through a weighted summation unit, so that the structural differences of local soil areas can be fully expressed, and finally, the soil association embedded feature representation is obtained.

[0052] Through the above specific implementation methods, different types of features are expressed and enhanced in the structure mapping layer, environment mapping layer and graph structure mapping layer, respectively, thereby effectively improving the refinement and stability of feature expression, and enabling the model to more accurately reflect the complex nonlinear relationship between the straw burning process and the tillage operation process, which is affected by physical structure, environmental conditions and soil conditions.

[0053] The state fusion module is used to perform fusion processing on each of the embedded feature representations to generate a fused state feature vector; specifically, the state fusion module performs fusion processing on the structural nonlinear embedded feature representation, the environmental nonlinear embedded feature representation, and the soil-related embedded feature representation to generate a fused state feature vector.

[0054] The response prediction module is used to perform nonlinear regression calculation on the fused state feature vector and output the corresponding pre-response feature parameter set.

[0055] The response prediction module includes a feature compression mapping unit, a nonlinear regression prediction unit, and a constraint mapping unit connected in sequence.

[0056] The feature compression mapping unit adopts a two-layer fully connected network structure, wherein the first fully connected layer has a dimension of 256 and the second fully connected layer has a dimension of 128. The feature compression mapping unit is used to perform feature compression and high-dimensional semantic reconstruction processing on the fused state feature vector to generate intermediate state feature representation.

[0057] The nonlinear regression prediction unit includes a dual-branch regression structure, wherein the first regression branch is used to output the prediction result of the target value of straw burning temperature, and the second regression branch is used to output the prediction result of the target value of tillage depth.

[0058] Each regression branch consists of a three-layer cascaded fully connected network, with the first layer having a dimension of 128, the second layer having a dimension of 64, and the third layer being a 1-dimensional output layer, with the ReLU activation function introduced in the second layer;

[0059] The constraint mapping unit is used to perform constraint mapping processing on the predicted results of the straw burning temperature target value and the predicted results of the tillage depth target value to generate the pre-response feature parameter set.

[0060] For example, in a straw resource utilization operation, the fused state feature vector output by the state fusion module simultaneously represents the physical structure characteristics of the straw, the characteristics of the operating environment, and the characteristics of the soil state. This includes comprehensive information such as high moisture content in corn straw, high ambient humidity, and high soil density. In the response prediction module, the fused state feature vector is first processed by a feature compression mapping unit for feature compression and semantic reconstruction. Specifically, a two-layer fully connected network structure is adopted. The first layer maps the high-dimensional fused features to a 256-dimensional space to initially compress redundant feature information and retain the main influencing factors. The second layer further maps the features to a 128-dimensional space, thereby forming an intermediate state feature representation, enabling different types of features to be reconstructed and expressed in a unified low-dimensional semantic space. For example, in this process, information such as moisture content, lignin content, and soil density, which were originally expressed separately, are compressed into feature vectors with unified semantic expression capabilities, enhancing the stability and computational efficiency of subsequent regression predictions.

[0061] Subsequently, in the nonlinear regression prediction unit, the intermediate state feature representation is input into a two-branch regression structure for prediction calculation. The first regression branch outputs the predicted target value of straw combustion temperature. This branch consists of a three-layer cascaded fully connected network with 128-dimensional, 64-dimensional, and 1-dimensional output structures, respectively. A ReLU activation function is introduced in the second layer to enhance the model's ability to fit nonlinear combustion characteristics (such as the effect of moisture content on combustion efficiency). The second regression branch has the same structure as the first branch and outputs the predicted target value of tillage depth, thus achieving parallel prediction of combustion and tillage parameters. For example, under the current operating conditions, the model might output a target combustion temperature of 730℃ and a target tillage depth of 18cm to adapt to the treatment requirements of high-moisture straw and highly compacted soil.

[0062] Finally, in the constraint mapping unit, the predicted results of the straw combustion temperature target value and the predicted results of the tillage depth target value are processed by engineering constraint mapping to generate the final pre-response feature parameter set. Specifically, when the predicted combustion temperature value exceeds the upper limit of the preset safety range, it is mapped to the upper limit value; when it is lower than the minimum combustion efficiency threshold, it is mapped to the minimum effective value range. Similarly, the predicted tillage depth value is constrained and adjusted according to the soil mechanical bearing capacity and the safe operating range. For example, when the model outputs a combustion temperature of 780℃, it can be corrected to a safe control value within 750℃ through the constraint mapping unit, thereby avoiding equipment overload or excessive combustion. Through the above processing method, the response prediction module can realize a three-level processing flow of feature compression, nonlinear regression prediction, and engineering constraint correction based on the fused state features, thereby effectively improving the accuracy and stability of the predicted straw combustion temperature target value and tillage depth target value.

[0063] Optionally, in some embodiments of this application, the constraint mapping process includes:

[0064] The target value of straw combustion temperature is matched with a preset temperature constraint range, which is denoted as [Tmin, Tmax], and the target value of straw combustion temperature is recorded as T.

[0065] The target tillage depth is matched with a preset operating depth constraint interval, which is denoted as [Dmin, Dmax], and the target tillage depth is recorded as D.

[0066] When T satisfies Tmin≤T≤Tmax, the target value of straw combustion temperature T is kept unchanged as the constraint output result.

[0067] When T is less than Tmin, the target value of straw combustion temperature T is mapped to Tmin;

[0068] When T is greater than Tmax, the target value of straw combustion temperature T is mapped to Tmax;

[0069] When D satisfies Dmin≤D≤Dmax, the target value of tillage depth D remains unchanged as the constraint output result;

[0070] When D is less than Dmin, the target value of tillage depth D is mapped to Dmin, where Dmin≥0;

[0071] When D is greater than Dmax, the target value of tillage depth D is mapped to Dmax.

[0072] In one optional implementation, the constraint mapping process is illustrated with the following specific example: For instance, in a straw resource utilization operation, the target value T for straw combustion temperature output by the response prediction module is 780℃, and the target value D for tillage depth is 22cm. Based on preset process safety and equipment compatibility requirements, the system sets the temperature constraint range as [Tmin, Tmax] = [650℃, 750℃], and the tillage depth constraint range as [Dmin, Dmax] = [10cm, 20cm]. During the constraint mapping process, the target value T for combustion temperature is first matched against the range. Since T = 780℃ is greater than Tmax = 750℃, it is mapped to T = 750℃ according to the constraint rules. This means that the target value exceeding the safety upper limit is truncated and corrected, returning it to the safe control range where the equipment can operate stably. If, in another case, T is 700℃, since the range of 650℃ ≤ T ≤ 750℃ is satisfied, the original value is kept unchanged as the constraint output result, thus ensuring that the normal prediction result is not interfered with within a reasonable process range. Similarly, a constraint matching judgment is performed on the target value D of tillage depth. Since D = 22cm is greater than Dmax = 20cm, it is mapped and corrected to D = 20cm to avoid excessive soil disturbance or excessive equipment load caused by excessive tillage depth. When D is at 12cm or other conditions satisfying 10cm ≤ D ≤ 20cm, the original value is kept unchanged to ensure a balance between operational effectiveness and soil disturbance. Through the above constraint mapping process, both the target values ​​of combustion temperature and tillage depth are limited to the preset safe range, thus forming a final stable and executable pre-response characteristic parameter set.

[0073] Optionally, in some other embodiments of this application, the straw bioconversion pre-response model is a time series regression model based on a one-dimensional convolutional neural network, used to extract local time series features from the straw basic state vector and output the target values ​​of straw combustion temperature and tillage depth.

[0074] In one optional implementation, the straw bioconversion pre-response model is a prediction model based on time series modeling, which is used to extract time dependencies and local or global time series features based on the historical or serialized feature information contained in the straw basic state vector, and output the corresponding target values ​​for straw combustion temperature and tillage depth.

[0075] Specifically, the straw bioconversion pre-response model can be a time series regression model based on a one-dimensional convolutional neural network (1D-CNN), which uses convolutional kernels to extract local receptive fields of the sequence features of the straw basic state vector to obtain local temporal change features and achieve regression prediction; or, the model can be a time series prediction model with a recurrent neural network structure, including a long short-term memory network (LSTM) or a gated recurrent unit (GRU), used to model the long-term dependency and dynamic evolution characteristics of the straw basic state vector in the time dimension; or, the straw bioconversion pre-response model can be a Transformer structure prediction model based on an attention mechanism, which uses a self-attention mechanism to weight the correlation between different time steps or different feature dimensions to enhance the global temporal dependency expression ability; or, the straw bioconversion pre-response model can also be a machine learning prediction model based on gradient boosting decision tree (GBDT), used to perform nonlinear regression fitting on the straw basic state vector to achieve the predicted output of the target values ​​of straw combustion temperature and tillage depth.

[0076] Specifically, the process execution mapping model includes a parameter parsing unit, a feature mapping unit, and a constraint correction unit connected in sequence;

[0077] The parameter parsing unit is used to perform structured parsing of the pre-response feature parameter set, obtain the target value of straw burning temperature and the target value of tillage depth, and construct a standardized process control feature vector.

[0078] The feature mapping unit is a nonlinear regression mapping network based on a multilayer perceptron, used to perform multivariate mapping processing on the standardized process control feature vector to generate equipment operating condition control vector; wherein, the equipment operating condition control vector includes thermal control sub-vector and mechanical operation control sub-vector;

[0079] The thermal control sub-vector is obtained by mapping the target value T of straw combustion temperature, and correspondingly generates injection temperature control quantity, ground clearance control quantity, fuel flow control quantity and nozzle angle control quantity;

[0080] The mechanical operation control sub-vector is obtained by mapping the tillage depth target value, and correspondingly generates the soil penetration depth control quantity and the crushing intensity control quantity.

[0081] The constraint correction unit is used to perform engineering constraint correction processing on the equipment operating condition control vector to generate a set of processing control parameters;

[0082] The engineering constraint correction process includes: limiting the injection temperature control quantity to a preset safe temperature range; limiting the ground clearance control quantity to the allowable range for equipment installation; limiting the fuel flow control quantity to the maximum output power; limiting the nozzle angle control quantity to the mechanical rotation angle range; constraining the soil bearing capacity of the penetration depth control quantity; and constraining the equipment load capacity of the crushing strength control quantity.

[0083] For example, in a straw resource utilization operation, suppose the pre-response feature parameter set output in step S3 includes a target straw combustion temperature T of 720℃ and a target tillage depth D of 18cm. First, in the parameter parsing unit, the pre-response feature parameter set is structured and parsed. T and D are extracted and standardized, for example, through interval normalization or scaling, to convert them into a process control feature vector of a uniform scale, thereby eliminating the influence of differences in the dimensions of different physical quantities on subsequent mapping calculations.

[0084] Subsequently, in the feature mapping unit, a nonlinear regression mapping network based on a multilayer perceptron is used to perform multivariate mapping processing on the standardized process control feature vector to generate the equipment operating condition control vector. This equipment operating condition control vector can be further divided into thermal control sub-vectors and mechanical operation control sub-vectors. Specifically, for the target value T of straw combustion temperature, the model outputs a thermal control sub-vector, corresponding to the generation of injection temperature control quantities (e.g., set to around 700℃), ground clearance control quantities (e.g., set within the range of 0.7–0.9m), fuel flow control quantities (e.g., corresponding to medium to high output power), and nozzle angle control quantities (e.g., set to a 45° tilt angle). For the target value D of tillage depth, the model outputs a mechanical operation control sub-vector, corresponding to the generation of soil penetration depth control quantities (e.g., set to 18cm) and crushing intensity control quantities (e.g., set to a medium to high intensity level based on soil compaction). Through this multivariate mapping process, combustion process parameters and tillage process parameters can be jointly mapped within the same model framework, achieving coordinated control of the thermal process and the mechanical operation process.

[0085] Furthermore, in the constraint correction unit, the equipment operating condition control vector is subjected to engineering constraint correction processing to ensure that each control parameter meets the limitations of equipment operation safety and working environment. For example, a safe temperature range limit (e.g., 650℃~750℃) is imposed on the spray temperature control quantity to avoid equipment damage or safety risks caused by excessive temperature; a limit on the allowable range of equipment installation structure (e.g., 0.5m~1.2m) is imposed on the ground clearance control quantity; a maximum output power limit is imposed on the fuel flow control quantity to prevent energy system overload; a mechanical rotation angle range limit (e.g., 0°~90°) is imposed on the nozzle angle control quantity; simultaneously, soil bearing capacity constraints are imposed on the soil penetration depth control quantity to avoid soil structure damage caused by excessive plowing; and equipment load capacity constraints are imposed on the crushing strength control quantity to prevent overload wear of mechanical parts. After the above constraint correction processing, a set of processing control parameters that meets the actual engineering requirements is finally generated and used to drive the straw resource utilization processing execution equipment for operation control.

[0086] This application achieves a standardized conversion path from process target parameters to equipment execution parameters by setting up a hierarchical processing structure consisting of a parameter parsing unit, a feature mapping unit, and a constraint correction unit. On the one hand, parameter parsing and standardization eliminate the dimensional differences between different physical parameters, improving the stability of subsequent mapping calculations. On the other hand, nonlinear mapping achieved through a multilayer perceptron enables the simultaneous generation of thermal control parameters and mechanical operation parameters by the target values ​​of straw combustion temperature and tillage depth, thereby achieving coordinated control between the straw combustion and tillage processes and avoiding the mismatch problem caused by setting combustion and tillage separately in traditional technologies. In addition, the constraint correction unit imposes engineering boundary restrictions on each control variable, ensuring that the output parameters meet the optimal process target while possessing good engineering feasibility and safety.

[0087] See Figure 2 Optionally, in some embodiments of this application, after controlling the straw resource utilization processing equipment based on the processing control parameter set, the method further includes:

[0088] S5. Obtain information on the effectiveness of straw treatment, including combustion completeness, straw residue rate, soil disturbance degree, and changes in soil organic matter.

[0089] For example, after the operation was completed on a certain plot of land, image recognition and sensor detection revealed that: the straw combustion completeness was 92%, the straw residue rate was 6%, the soil disturbance level was moderate to high, and the soil organic matter change increased by approximately 2.3%. Among these, the combustion completeness can be obtained through infrared imaging or ash analysis, the straw residue rate can be obtained through visual detection or sampling statistics, the soil disturbance level can be obtained through soil structure sensors or compaction changes, and the change in soil organic matter can be assessed through pre- and post-sampling detection.

[0090] S6. Construct a three-dimensional response relationship map based on the pre-response feature parameter set and the effect information. The three-dimensional response relationship map uses the target value of straw burning temperature, the target value of tillage depth, and the treatment effect index as three dimensions.

[0091] Specifically, a three-dimensional response space is formed based on the accumulation of historical operation data, with the target combustion temperature T as the first dimension, the target tillage depth D as the second dimension, and the treatment effect indicators (such as combustion completeness, residue rate, and other comprehensive indicators) as the third dimension. For example, in this space, it can be observed that when the target combustion temperature T is between 700℃ and 730℃ and the target tillage depth D is between 16cm and 18cm, the corresponding combustion completeness is high and the residue rate is low, thus forming a high-quality response area.

[0092] S7. Based on the three-dimensional response relationship map, a multi-channel feature evaluation network is used to evaluate the pre-response feature parameter set to obtain an evaluation score;

[0093] The multi-channel feature evaluation network includes:

[0094] The process parameter channel is used to perform feature weighting processing on the target values ​​of straw combustion temperature and tillage depth based on the attention mechanism to obtain the process parameter feature representation, and to calculate the process evaluation score based on the process parameter feature representation;

[0095] For example, in a specific operational scenario, based on the basic state vector of straw collected from the current plot, a set of pre-response feature parameters is predicted by the straw biotransformation pre-response model. The target value for straw combustion temperature T is 720℃, and the target value for tillage depth D is 18cm. Simultaneously, the effect information collected after the operation is as follows: combustion completeness is 92%, straw residue rate is 6%, soil disturbance is moderate to high (quantifiable as 0.7), and soil organic matter change is +2.3%. Furthermore, the current operational environment data includes ambient temperature of 28℃, ambient humidity of 70%, soil moisture content of 25%, and relatively high soil density. In the process parameter channel, the target value for combustion temperature T and the target value for tillage depth D are first combined to form a process parameter vector, for example, represented as [720, 18], and then input into the feature weighting model based on an attention mechanism. The model learns the importance weights of different parameters under current operating conditions. For example, under high humidity and high-density soil conditions, the model automatically assigns a higher weight to combustion temperature (e.g., 0.65) and a weight to tillage depth (0.35), thereby generating a weighted feature representation of the process parameters, such as [0.65×720, 0.35×18]. Subsequently, this feature representation is input into a scoring layer (such as a fully connected network) to calculate a process evaluation score, for example, 0.87, which reflects the degree of matching of the parameter combination in terms of process rationality.

[0096] The operation effect channel is used to perform nonlinear mapping on the effect information based on a deep feedforward neural network to obtain an effect feature representation, and to calculate an effect evaluation score based on the effect feature representation. For example, in the operation effect channel, the effect information is vectorized and represented as [0.92, 0.06, 0.7, 2.3], and input into the deep feedforward neural network for nonlinear mapping. This network models the nonlinear relationship between different effect indicators through a multi-layer fully connected structure. For example, it can identify "high combustion completeness + low residue rate" as a high-quality combination, and strengthen the expression of "moderate soil disturbance + increased organic matter" as a positive factor. After feature extraction, the effect feature representation is obtained, and the effect evaluation score is calculated through the output layer, for example, 0.91, to characterize the degree of superiority or inferiority of the current parameter combination in terms of actual treatment effect.

[0097] The environmental adaptation channel is used to extract features from the operational environment data using a graph-structured neural network, obtaining an environmental feature representation, and calculating an environmental evaluation score based on this representation. Specifically, in the environmental adaptation channel, a graph structure is constructed by combining current operational environment data with historical environmental data and input into the graph neural network. For example, association edges are established between the current environmental node (temperature 28℃, humidity 70%, high soil density) and multiple historical operational environment nodes, with higher weights for connections between similar environmental conditions (such as high-humidity, high-density soil). After multiple rounds of neighborhood feature aggregation through the graph neural network, an environmental feature representation is extracted, reflecting the degree of matching between the current environment and historically efficient operational environments. Subsequently, an environmental evaluation score, for example, 0.84, is calculated through a scoring layer to reflect the adaptability of the pre-response feature parameter set under the current environmental conditions.

[0098] S8. The process evaluation score, effect evaluation score, and environmental evaluation score are input into the weighted fusion module for weighted fusion processing to obtain a comprehensive evaluation score. Specifically, the process evaluation score, effect evaluation score, and environmental evaluation score are input into the weighted fusion module for comprehensive calculation, for example, weighted fusion is performed according to preset weights of 0.4, 0.4, and 0.2 to obtain a comprehensive evaluation score (e.g., 0.886). This comprehensive score reflects the overall performance of the current pre-response characteristic parameter set in terms of process rationality, operational effectiveness, and environmental adaptability.

[0099] S9. Determine the optimal pre-response feature parameter set based on the comprehensive evaluation score, and add the optimal pre-response feature parameter set and the corresponding straw basic state vector to the straw basic state vector training dataset to iteratively update the straw bioconversion pre-response model.

[0100] The optimal pre-response feature parameter set is determined based on the maximum value of the comprehensive evaluation score or by satisfying a preset threshold condition.

[0101] Specifically, in step S9, different parameter combinations are optimized based on the comprehensive evaluation score. When the score is the maximum value in the current candidate parameter set or exceeds a preset threshold (e.g., 0.85), the corresponding pre-response feature parameter set is determined as the optimal pre-response feature parameter set. This optimal parameter set and the corresponding straw basic state vector are added to the training dataset as new high-quality samples for incremental training or periodic updates of the straw bioconversion pre-response model. For example, after multiple operations, the model gradually learns that under high humidity and high-density soil conditions, the combustion temperature should be appropriately increased and the tillage depth increased to achieve better treatment results.

[0102] In this embodiment of the application, the straw resource utilization treatment execution equipment includes: a flame extinguishing device and a tillage and covering mechanism, which are respectively used to perform straw burning treatment and tillage and covering treatment according to the treatment control parameter set.

[0103] The flame suppression device includes a matrix propane nozzle assembly, a fuel supply module, and a nozzle angle and height adjustment mechanism. The matrix propane nozzle assembly consists of at least nine nozzles arranged in an array. Each nozzle is connected to the fuel supply module via an independent electromagnetic flow valve. The nozzle angle adjustment mechanism includes a servo motor drive assembly, and the height adjustment mechanism includes an electric push rod or a hydraulic lifting unit. The flame suppression device receives and processes flame suppression nozzle control parameters from a set of control parameters and adjusts the injection angle, ground clearance, and fuel flow rate of each nozzle according to these parameters.

[0104] The tillage and covering mechanism includes a spiral-arranged plow body assembly, a soil-crushing roller assembly, and a hydraulic system for adjusting the plow blade depth. The spiral-arranged plow body assembly consists of at least two sets of staggered plow bodies, which are connected to the frame via a suspension beam. The soil-crushing roller is located behind the plow body and driven by a geared motor. The hydraulic system for adjusting the plow blade depth includes a hydraulic cylinder and a displacement feedback sensor, used to control the depth of the plow body into the soil. The tillage and covering mechanism receives control parameters from the tillage equipment and adjusts the depth of the plow body into the soil and the rotation speed of the soil-crushing roller (to reflect the crushing intensity) based on the control parameters.

[0105] In some other embodiments of this application, the straw resource utilization processing equipment further includes a suspended pollutant control device. The suspended pollutant control device includes a fire-blocking dust collection plate assembly, a smoke-gathering shell, and a water mist spraying system. The fire-blocking dust collection plate assembly is an inclined, high-temperature resistant metal plate structure. The smoke-gathering shell is a closed or semi-closed shell structure with an inner wall guiding curved surface. The water mist spraying system includes a high-pressure water pump and a multi-nozzle atomizing spray unit. At least three laser-scattering smoke sensors distributed along the height direction are installed inside the smoke-gathering shell. The suspended pollutant control device is used to adjust the water mist spraying duration, spraying pressure, and the opening of the shell's air outlet according to pollution control parameters to achieve coordinated control of PM2.5 and smoke.

[0106] Example 2

[0107] In this embodiment, a method for the resource utilization of straw after the harvest of crops such as corn, wheat, and soybeans in farmland is provided based on intelligent sensing, along with an integrated processing device, to achieve efficient straw disposal, pollution control, and soil improvement. This method uses an intelligent control center as its core, acquiring multi-source data through an intelligent sensing system, generating control parameters through model calculations, and driving the coordinated operation of the flame suppression device, the tillage and covering mechanism, and the pollutant control device.

[0108] The straw resource utilization method begins with data acquisition. A high-resolution infrared spectroscopy scanner at the front of the equipment continuously scans the work area at a frequency of 10 scans per second within a scanning range of 0.8-1.5 μm. By analyzing the difference in reflectance between straw and soil in the near-infrared band, information on straw cover density, thickness, and distribution uniformity is obtained. Combined with a calibration model trained based on soybean, corn, and wheat straw spectral databases, the amount of straw per unit area is calculated with an error controlled within 5%. Simultaneously, the straw type is identified by recognizing the difference in characteristic absorption peak intensity between the 1.2 μm and 1.4 μm bands. A soil parameter sensor located below the scanner simultaneously acquires soil hardness data within the 0-300 mm range and moisture content data in the top 10 cm layer to determine soil bearing capacity and tillage resistance. An environmental sensor at the top of the equipment collects wind speed, wind direction, and temperature and humidity parameters in real time to assess the impact of the external environment on flame propagation and smoke diffusion.

[0109] After acquiring raw material characteristic data, operational environment data, and basic soil parameter data, the data are normalized and scaled. Based on preset coding rules, sub-vectors representing the physical structure of straw, environmental state, and soil state are constructed. These are then fused to generate a basic straw state vector. This state vector serves as input for subsequent model calculations, comprehensively representing the overall state of the current operational area.

[0110] The basic state vector of straw is input into a pre-trained straw bioconversion pre-response model for processing. This model employs a multi-branch feature embedding structure, performing nonlinear mapping and embedding representation on different types of features. Straw features are enhanced with structural information through fully connected mapping and channel attention mechanisms; environmental features are filtered and strengthened through gating units; and soil features are aggregated through graph structure modeling to achieve correlation information. After fusion, a unified state feature vector is formed, and the target values ​​for straw combustion temperature and tillage depth are output through a nonlinear regression prediction module. Subsequently, a constraint mapping unit applies interval restrictions to the prediction results to ensure they meet preset temperature and tillage depth constraints, thereby obtaining the pre-response feature parameter set.

[0111] Based on the aforementioned pre-response characteristic parameter set, equipment control parameters are further generated through a process execution mapping model. This model employs a multi-layer perceptron structure, mapping the target value of straw combustion temperature to thermal control parameters such as injection temperature, nozzle angle, ground clearance, and fuel flow rate. Simultaneously, the target value of tillage depth is mapped to mechanical control parameters such as soil penetration depth and crushing strength. A constraint correction unit performs engineering range correction on each control variable to ensure safe and reliable equipment operation. The control center then adjusts the actuators in real time based on the generated control parameters.

[0112] During the high-temperature combustion phase, the combustion device operates using a matrix structure of nine propane nozzles, spaced 400mm x 400mm apart, with a total power of 135kW. Initial parameters are set according to the type and density of the straw; for example, a larger spray angle and higher fuel flow rate are set for areas with corn stalks to enhance penetration, while the flame intensity is appropriately reduced for areas with soybean or wheat stalks to avoid energy waste. An infrared temperature sensor is installed below each nozzle to monitor the surface temperature of the burned straw in real time, and the fuel flow rate, nozzle angle, and ground clearance are dynamically adjusted based on temperature deviations. When the temperature is below a set threshold, the flame intensity is increased; when the temperature is too high, the flame intensity is reduced, ensuring that the straw is fully carbonized within a suitable temperature range. This dynamic adjustment mechanism ensures a stable straw combustion rate of over 92%.

[0113] After flame treatment, the equipment enters the tillage, crushing, and simultaneous covering stage. The tillage mechanism employs a spirally arranged plow body structure, coupled with a soil-crushing roller to loosen the soil and mix it with straw residue. Based on real-time straw quantity data acquired by front-end infrared scanning, the control center dynamically adjusts the plow blade depth. When an increase in straw quantity is detected, the tillage depth is appropriately increased to ensure the straw is fully buried in the soil; when the straw quantity decreases, the tillage depth is decreased to avoid excessive disturbance to the soil structure. Simultaneously, the soil-crushing roller breaks up the turned-up clods, controlling the soil particle size within a reasonable range, thereby improving mixing uniformity and soil looseness.

[0114] Smoke and suspended particles generated during operation are treated by a pollutant control device. The fire-blocking dust collection plate intercepts large particles through inertial impaction and gravity settling, the smoke-gathering shell guides and collects the smoke, and the water mist spraying system automatically activates based on the concentration monitored by the smoke sensor. When the smoke concentration exceeds a set threshold, water mist spraying is activated, using fine water droplets to adsorb PM2.5 particles for wet purification; spraying automatically stops when the concentration drops to a safe range. Through these multi-stage control measures, pollutant emissions can be effectively reduced, achieving centralized control and efficient purification of smoke.

[0115] After the pollutants are controlled, the carbonized straw residue and the intercepted pollutants are buried together in the soil to a depth of approximately 150 mm using a soil covering mechanism. This creates a low-oxygen environment, promoting the decomposition of organic pollutants by anaerobic microorganisms. This deep burial process is monitored in real time by depth sensors, and the soil covering angle is dynamically adjusted to ensure the burial depth remains stable within the target range. Under anaerobic conditions, pollutants can be decomposed into small molecules and converted into soil organic matter, thereby achieving the harmless treatment of pollutants and improving soil fertility.

[0116] After the operation is completed, the system further acquires information on the treatment effect, including the completeness of straw combustion, residual rate, degree of soil disturbance, and change in organic matter. Based on the pre-response feature parameter set and effect information, a three-dimensional response relationship map is constructed. A multi-channel feature evaluation network is used to evaluate the process parameters, operational effect, and environmental adaptability, and a comprehensive evaluation score is obtained through weighted fusion. The optimal parameter combination is selected based on this score and added to the training dataset along with the corresponding straw basic state vector. This allows for iterative updates to the straw bioconversion pre-response model, continuously improving the system's prediction accuracy and control effect.

[0117] The above-described implementation methods enable intelligent and precise control of the entire straw treatment process, which not only significantly improves straw disposal efficiency but also effectively reduces environmental pollution and improves soil structure and fertility, thus possessing significant potential for widespread application.

[0118] Example 3

[0119] This embodiment provides a straw resource utilization system and its operation method based on intelligent sensing. The system constructs a closed-loop control mechanism of "sensing-modeling-decision-execution-feedback" to achieve efficient straw disposal, effective control of suspended pollutants, and synergistic improvement of soil structure and fertility, thereby improving agricultural efficiency while reducing environmental pollution.

[0120] Before operation, this system first acquires a set of processing control parameters and then controls the straw resource utilization processing equipment based on this set of parameters.

[0121] The process of obtaining the processing control parameter set includes the following steps:

[0122] First, collect data on the raw material characteristics of the straw to be processed, the working environment, and basic soil parameters.

[0123] Secondly, feature fusion processing is performed on the above multi-source data to construct a basic state vector of straw. This state vector includes: a sub-vector of straw physical structure features corresponding to raw material feature data, a sub-vector of environmental state features corresponding to working environment data, and a sub-vector of soil state features corresponding to basic soil parameter data;

[0124] Then, the straw basic state vector is input into a pre-trained straw bioconversion pre-response model to obtain a pre-response feature parameter set. The pre-response feature parameter set includes target values ​​for straw combustion temperature and tillage depth. The model is trained based on historical straw basic state vectors and their corresponding preset combustion temperature and tillage depth data.

[0125] Finally, the pre-response feature parameter set is input into the pre-trained process execution mapping model to generate a processing control parameter set, which is then used to control the straw resource utilization processing equipment. The processing control parameter set includes flame scavenging nozzle control parameters and tillage equipment control parameters; the flame scavenging nozzle control parameters include injection temperature, ground clearance, fuel flow rate, and nozzle angle; the tillage equipment control parameters include soil penetration depth and crushing intensity. The process execution mapping model is a multivariate nonlinear mapping model based on a multilayer perceptron.

[0126] During execution, the system combines real-time on-site sensing data for dynamic control and adjustment.

[0127] A high-resolution infrared spectral scanner is installed at the front end of the equipment's travel path. Its scanning range is 0.8–1.5 μm, its scanning width is 1.2–1.5 m, and its scanning frequency is 10 scans per second. This scanner detects the difference in reflectivity between straw and soil in the near-infrared band and, combined with a calibration model trained on a database of soybean, corn, and wheat straw spectral data, calculates the straw cover density, thickness, and uniformity of distribution. The calculation error per unit area of ​​straw amount does not exceed 5%, and a straw distribution heat map is generated.

[0128] For example, wheat straw has a reflectance 15% to 20% lower than that of soil at 1.2 μm, and the absorption peak intensity of corn straw at 1.4 μm is 1.3 to 1.5 times that of soybean straw. The system identifies the type and density of straw based on this.

[0129] A soil hardness sensor and a moisture content detector are installed below the scanner to measure soil hardness (0–300 mm, 0–5 MPa, accuracy ±0.1 MPa) and top 10 cm moisture content (5%–30%, accuracy ±1%). When soil hardness exceeds 3 MPa, the plow drive torque is increased; when moisture content exceeds 25%, the operating speed is reduced.

[0130] A wind speed and direction sensor, as well as a temperature and humidity sensor, are installed on the top of the device. When the wind speed exceeds 5 m / s, the flame power is reduced and the smoke casing's guide angle is adjusted.

[0131] Based on the above-mentioned sensing data, the control center generates an initial parameter scheme, for example: when the amount of straw is greater than 1.2 kg / m², the initial depth of the plow blade entering the ground is 25 mm; the initial temperature of the flame in the corn straw area is 1350℃; and the smoke shell sealing degree is increased by 20% under high wind speed conditions.

[0132] The flame suppression device adopts a matrix structure consisting of 9 propane nozzles (400mm×400mm spacing), with a single nozzle power of 15kW and a total power of 135kW, ensuring a flame overlap rate of not less than 30%.

[0133] The nozzle angle can be adjusted from 30° to 45°, the ground clearance is 120 to 180 mm, and the propane flow rate is 200 to 300 g / min. Initial parameters are set according to the type of straw: 45° angle, 150 mm height, and 280 g / min flow rate for corn straw areas; 35° angle, 160 mm height, and 220 g / min flow rate for soybean straw areas.

[0134] Each nozzle is equipped with an infrared temperature sensor (500–2000℃, sampling frequency 5 times / second) for real-time detection of straw carbonization temperature. The system dynamically adjusts based on temperature deviation.

[0135] When the temperature is more than 100°C below the threshold, increase the fuel flow rate, lower the altitude, and increase the injection angle.

[0136] When the temperature exceeds the threshold of 200°C, reduce the fuel flow rate, increase the altitude, and decrease the angle.

[0137] Fine-tuning is performed when the temperature is within ±50℃.

[0138] Through this regulation mechanism, corn stalks can be carbonized in 0.8 to 1.0 seconds at 1300 to 1350℃, and soybean and wheat stalks can be carbonized in 0.5 to 0.7 seconds at 1200 to 1300℃. The straw utilization rate reaches more than 92.3%, and the residual particle size is no more than 5 mm.

[0139] The tillage mechanism uses a spirally arranged plowshare (370mm spacing, 120mm lateral offset), with the following parameters: 42° circumference angle, 580mm radius of curvature, and 28° blade inclination angle. The soil-crushing roller has a diameter of 500mm and a rotation speed of 300r / min.

[0140] The initial depth of entry into the ground is set to 15-25 mm based on the amount of straw, and the operating speed is 5.15 km / h.

[0141] The system dynamically adjusts the plow depth based on real-time feedback of straw quantity from infrared scanning.

[0142] When the amount of straw increases by more than 10%, the depth increases by 2-3 mm;

[0143] When the reduction is more than 10%, the depth decreases by 2-3 mm;

[0144] For areas with uneven distribution, precise local control can be achieved by independently adjusting the left and right plow bodies.

[0145] This stage can achieve a soil fragmentation rate of over 85.6%, a stable tillage depth of 250 mm (±5 mm), and a straw-soil mixing uniformity of over 90.2%.

[0146] The pollution control device includes a fire-blocking dust collection plate, a smoke-gathering shell, and a water mist spraying system.

[0147] The fire-blocking dust collection plate is made of high-temperature resistant alloy steel plate arranged at a 45° angle, which can intercept particles with a diameter greater than 1μm.

[0148] The smoke-gathering shell uses an optimized airflow structure to concentrate the smoke;

[0149] The water mist spraying system can produce water mist particles of 5–10 μm.

[0150] The system monitors the concentration in real time through a smoke sensor. When the concentration exceeds the threshold, it triggers water mist spraying and dynamically adjusts the spraying time and the degree of enclosure sealing according to the concentration changes.

[0151] Through this collaborative control, PM 2.5 The capture rate reached over 85%, the smoke concentration rate reached over 97.08%, and the emission reduction exceeded 90%.

[0152] The mixture is buried into a 150mm deep soil layer by a soil covering mechanism, and the burial depth is controlled within 150mm ± 5mm by dynamically adjusting the angle of the plow blade through a depth sensor.

[0153] This depth creates a low-oxygen environment (oxygen content below 5%), which promotes the decomposition of pollutants by anaerobic microorganisms, converting organic pollutants into CO2 and CH4, etc.

[0154] After 3 months of degradation, the pollutant degradation rate exceeded 99%, the polycyclic aromatic hydrocarbon degradation rate exceeded 99.5%, and the soil organic matter content increased by 15% to 20%.

[0155] This system comprises an intelligent sensing system, a flame suppression device, a tillage and covering mechanism, a pollution control device, and an intelligent control center. The intelligent sensing system includes an infrared spectral scanner, temperature sensors, smoke sensors, soil parameter sensors, and environmental sensors; the flame suppression device includes a matrix-type propane nozzle, a fuel supply system, and an adjustment mechanism; the tillage and covering mechanism includes a auger plow, a soil-crushing roller, and a hydraulic adjustment system; the pollution control device includes a fire-blocking dust collection plate, a smoke enclosure, and a water mist system; and the intelligent control center uses a PLC controller for data processing and execution control.

[0156] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for intelligent sensing-based straw resource processing, characterized in that, include: S1. Collect raw material characteristic data, working environment data, and basic soil parameter data of the straw to be processed; S2. Perform feature fusion processing on the raw material characteristic data, working environment data, and soil basic parameter data to construct a straw basic state vector; wherein the straw basic state vector includes a straw physical structure feature sub-vector corresponding to the raw material characteristic data, an environmental state feature sub-vector corresponding to the working environment data, and a soil state feature sub-vector corresponding to the soil basic parameter data. S3. Input the straw basic state vector into the pre-trained straw bioconversion pre-response model to obtain a pre-response feature parameter set; wherein, the straw bioconversion pre-response model is trained based on the straw basic state vector training dataset; the pre-response feature parameter set includes the target value of straw combustion temperature and the target value of tillage depth; The training dataset for straw basic state vectors includes historical straw basic state vectors and corresponding pre-set straw burning temperatures and tillage depths. S4. Based on the pre-response feature parameter set, a processing control parameter set is generated through a pre-trained process execution mapping model, and the straw resource utilization processing equipment is controlled based on the processing control parameter set; wherein, the processing control parameter set includes flame scavenging nozzle control parameters and tillage equipment control parameters; the flame scavenging nozzle control parameters include injection temperature, ground clearance, fuel flow rate, and nozzle angle; the tillage equipment control parameters include soil penetration depth and crushing intensity; The process execution mapping model is a multivariable nonlinear parameter mapping model based on a multilayer perceptron.

2. The smart sensing based straw resource processing method according to claim 1, characterized in that, Step S2 includes: S21. Normalize and scale the raw material characteristic data, working environment data and soil basic parameter data, and construct straw physical structure feature sub-vector, environmental state feature sub-vector and soil state feature sub-vector respectively based on preset feature coding rules. S22. The straw physical structure feature vector, environmental state feature vector, and soil state feature vector are fused to generate the straw basic state vector.

3. The smart sensing based straw resource processing method according to claim 2, characterized in that, The straw bioconversion pre-response model includes: a feature mapping module, a state fusion module, and a response prediction module; The feature mapping module is used to perform nonlinear feature transformation processing on each feature sub-vector to obtain the corresponding embedded feature representation; The state fusion module is used to perform fusion processing on each of the embedded feature representations to generate a fused state feature vector; The response prediction module is used to perform nonlinear regression calculation on the fused state feature vector and output the corresponding pre-response feature parameter set.

4. The smart sensing based straw resource processing method according to claim 3, characterized in that, The feature mapping module includes a multi-branch feature embedding network structure, which includes a first feature mapping branch corresponding to the straw physical structure feature vector, a second feature mapping branch corresponding to the environmental state feature vector, and a third feature mapping branch corresponding to the soil state feature vector. The first feature mapping branch includes a fully connected mapping layer, a feature recalibration layer, and a nonlinear activation layer arranged sequentially. The fully connected mapping layer is used to perform high-dimensional space expansion processing on the input straw physical structure feature sub-vectors to generate an initial structure embedding representation; The feature recalibration layer is used to perform weight allocation processing on the initial structural embedding representation based on the channel attention mechanism to generate a structural recalibrated embedding feature representation; The nonlinear activation layer is used to perform nonlinear mapping transformation on the structural recalibrated embedded feature representation and output the structural nonlinear embedded feature representation. The second feature mapping branch includes an environmental feature encoding layer, a gated feature selection unit, and a nonlinear transformation layer; The environmental feature encoding layer is used to perform high-dimensional feature encoding on the environmental state feature sub-vectors to generate a basic environmental embedding representation. The gated feature selection unit is used to perform information filtering and enhancement processing on the environmental basic embedding representation based on feature relevance weights to generate an environmental filtering embedding representation. The nonlinear transformation layer is used to perform nonlinear mapping processing on the filtered environmental embedding features to obtain an environmental nonlinear embedding feature representation. The third feature mapping branch includes a graph structure encoding layer and a neighborhood feature aggregation layer; The graph structure encoding layer is used to perform structured encoding processing on soil state feature subvectors, convert them into node feature representations, and construct a soil state association graph based on the correlation between soil physical properties to form soil node feature representations and corresponding graph structure topology relationships. The neighborhood feature aggregation layer is used to perform weighted aggregation calculation on the soil node feature representation based on preset association rules, so as to fuse the feature information between mutually related nodes and generate soil association embedded feature representation; The state fusion module performs fusion processing on the structural nonlinear embedding feature representation, the environmental nonlinear embedding feature representation, and the soil-related embedding feature representation to generate a fused state feature vector.

5. The smart sensing based straw resource processing method according to claim 4, characterized in that, The response prediction module includes a feature compression mapping unit, a nonlinear regression prediction unit, and a constraint mapping unit connected in sequence. The feature compression mapping unit adopts a two-layer fully connected network structure, wherein the first fully connected layer has a dimension of 256 and the second fully connected layer has a dimension of 128. The feature compression mapping unit is used to perform feature compression and high-dimensional semantic reconstruction processing on the fused state feature vector to generate intermediate state feature representation. The nonlinear regression prediction unit includes a dual-branch regression structure, wherein the first regression branch is used to output the prediction result of the target value of straw burning temperature, and the second regression branch is used to output the prediction result of the target value of tillage depth. Each regression branch consists of a three-layer cascaded fully connected network, with the first layer having a dimension of 128, the second layer having a dimension of 64, and the third layer being a 1-dimensional output layer, with the ReLU activation function introduced in the second layer; The constraint mapping unit is used to perform constraint mapping processing on the predicted results of the straw burning temperature target value and the predicted results of the tillage depth target value to generate the pre-response feature parameter set.

6. The smart sensing based straw resource processing method according to claim 5, characterized in that, in, The constraint mapping process includes: The target value of straw combustion temperature is matched with a preset temperature constraint range, which is denoted as [Tmin, Tmax], and the target value of straw combustion temperature is recorded as T. The target tillage depth is matched with a preset operating depth constraint interval, which is denoted as [Dmin, Dmax], and the target tillage depth is recorded as D. When T satisfies Tmin≤T≤Tmax, the target value of straw combustion temperature T is kept unchanged as the constraint output result. When T is less than Tmin, the target value of straw combustion temperature T is mapped to Tmin; When T is greater than Tmax, the target value of straw combustion temperature T is mapped to Tmax; When D satisfies Dmin≤D≤Dmax, the target value of tillage depth D remains unchanged as the constraint output result; When D is less than Dmin, the target value of tillage depth D is mapped to Dmin, where Dmin≥0; When D is greater than Dmax, the target value of tillage depth D is mapped to Dmax.

7. The method for resource utilization of straw based on intelligent sensing according to claim 1, characterized in that, The straw bioconversion pre-response model is a time series regression model based on a one-dimensional convolutional neural network, used to extract local time series features from the basic state vector of straw and output the target values ​​of straw combustion temperature and tillage depth.

8. The smart sensing based straw resource processing method according to claim 6 or 7, characterized in that, The process execution mapping model includes a parameter parsing unit, a feature mapping unit, and a constraint correction unit connected in sequence. The parameter parsing unit is used to perform structured parsing of the pre-response feature parameter set, obtain the target value of straw burning temperature and the target value of tillage depth, and construct a standardized process control feature vector. The feature mapping unit is a nonlinear regression mapping network based on a multilayer perceptron, used to perform multivariate mapping processing on the standardized process control feature vector to generate equipment operating condition control vector; wherein, the equipment operating condition control vector includes thermal control sub-vector and mechanical operation control sub-vector; The thermal control sub-vector is obtained by mapping the target value T of straw combustion temperature, and correspondingly generates injection temperature control quantity, ground clearance control quantity, fuel flow control quantity and nozzle angle control quantity; The mechanical operation control sub-vector is obtained by mapping the tillage depth target value, and correspondingly generates the soil penetration depth control quantity and the crushing intensity control quantity. The constraint correction unit is used to perform engineering constraint correction processing on the equipment operating condition control vector to generate a set of processing control parameters; The engineering constraint correction process includes: limiting the injection temperature control quantity to a preset safe temperature range; limiting the ground clearance control quantity to the allowable range for equipment installation; limiting the fuel flow control quantity to the maximum output power; limiting the nozzle angle control quantity to the mechanical rotation angle range; constraining the soil bearing capacity of the penetration depth control quantity; and constraining the equipment load capacity of the crushing strength control quantity.

9. The smart sensing based straw resource processing method according to claim 8, characterized in that, After controlling the straw resource utilization processing equipment based on the aforementioned set of processing control parameters, the method further includes: S5. Obtain information on the effectiveness of straw treatment, including combustion completeness, straw residue rate, soil disturbance degree, and changes in soil organic matter. S6. Construct a three-dimensional response relationship map based on the pre-response feature parameter set and the effect information. The three-dimensional response relationship map uses the target value of straw burning temperature, the target value of tillage depth, and the treatment effect index as three dimensions. S7. Based on the three-dimensional response relationship map, a multi-channel feature evaluation network is used to evaluate the pre-response feature parameter set to obtain an evaluation score; The multi-channel feature evaluation network includes: The process parameter channel is used to perform feature weighting processing on the target values ​​of straw combustion temperature and tillage depth based on the attention mechanism to obtain the process parameter feature representation, and to calculate the process evaluation score based on the process parameter feature representation; The task performance channel is used to perform nonlinear mapping on the performance information based on a deep feedforward neural network to obtain performance feature representation, and to calculate the performance evaluation score based on the performance feature representation; An environmental adaptation channel is used to extract features from work environment data using a neural network based on graph structure modeling, obtain an environmental feature representation, and calculate an environmental evaluation score based on the environmental feature representation. S8. Input the process evaluation score, effect evaluation score and environmental evaluation score into the weighted fusion module for weighted fusion processing to obtain the comprehensive evaluation score; S9. Determine the optimal pre-response feature parameter set based on the comprehensive evaluation score, and add the optimal pre-response feature parameter set and the corresponding straw basic state vector to the straw basic state vector training dataset to iteratively update the straw bioconversion pre-response model. The optimal pre-response feature parameter set is determined based on the maximum value of the comprehensive evaluation score or by satisfying a preset threshold condition.

10. The smart sensing based straw resource processing method according to claim 8, wherein, The straw resource utilization treatment equipment includes a flame extinguishing device and a tillage and covering mechanism, which are used to perform straw burning treatment and tillage and covering treatment according to the treatment control parameter set, respectively.

Images