A machine learning optimization method for sorting ash loading ship parameters

Abstract

The application discloses a kind of sorting ash conveying ship loading parameter machine learning optimization methods, comprising the following steps: S1, multiple-source data is collected and data pre-processing, generate space-time working condition sequence;S2, space-time working condition sequence is executed spectral domain decomposition, extract multi-scale frequency component;S3, to working condition spectrum tensor executes graph structure mapping, constructs dynamic association graph;S4, to topological feature matrix executes time series prediction, extracts long-time dependent feature;S5, to parameter prediction sequence executes multi-objective reinforcement optimization, constructs reward constraint function, generates parameter control sequence;S6, according to parameter control sequence adjusts equipment operating state, executes error reverse mapping processing, updates node coupling weight and reward constraint function.The application realizes the intelligent prediction control and adaptive optimization of ash conveying and ship loading process by multi-source data fusion, frequency domain decomposition, graph structure modeling and reinforcement learning optimization.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a machine learning optimization method for sorting, conveying and loading parameters for ash shipment. Background Technology

[0002] In port bulk material transportation and loading operations, ash sorting and loading systems typically involve multiple continuous processes, including ash pipeline transportation, ash sorting and processing, and ship loading. Existing technologies often employ experience-based settings or manual adjustments to control key operating parameters such as ash conveying pressure, fan frequency, valve opening, and loading flow rate. These adjustments are made manually or through simple rules to maintain basic system stability.

[0003] As the scale of ash conveying increases and the fluctuations in ash material properties intensify, existing control methods are gradually revealing their insufficient adaptability when facing complex operating conditions. Due to the obvious time-varying and nonlinear characteristics of the ash conveying process, there is a strong coupling relationship between ash conveying pressure fluctuations, ash flow concentration changes, and sorting particle size distribution changes. Traditional methods are difficult to effectively model the dynamic correlation between multiple parameters, resulting in parameter adjustment lags and problems such as uneven conveying, sorting deviations, and discontinuous loading cycles.

[0004] Meanwhile, existing technologies have limited utilization of multi-source data, with most control decisions based solely on single sensor data or local monitoring data. They lack the ability to model the global collaborative processes between the ash conveying system, sorting system, and loading system, and cannot fully exploit the frequency domain features, topological features, and dynamic evolution characteristics in time-series data, making it difficult to achieve global optimization in control strategies. Furthermore, existing methods typically consider only a single indicator during optimization, such as conveying stability or energy consumption control, lacking a comprehensive balancing optimization mechanism among ash conveying stability, sorting purity, loading continuity, and equipment energy consumption. This can easily lead to improvements in local performance at the expense of overall efficiency.

[0005] Furthermore, existing rule-based or traditional optimization algorithm-based control methods lack continuous learning capabilities and are difficult to adaptively update according to changes in operating conditions. They are prone to model failure under conditions such as equipment aging, changes in material sources, and external disturbances, thereby affecting long-term operational stability and economy.

[0006] Therefore, how to provide a machine learning optimization method for sorting, conveying, and loading parameters is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a machine learning optimization method for sorting, ash conveying, and loading parameters. This invention fully utilizes multi-source time-series data processing, discrete wavelet transform spectral domain analysis, graph attention network topology modeling, bidirectional recurrent neural network time-series prediction, and multi-objective reinforcement learning strategy iteration techniques to dynamically model and collaboratively optimize the operating parameters of the entire process of ash conveying, sorting, and loading. It has the advantages of strong parameter adaptive optimization capability, high system collaborative accuracy, and excellent operational stability and energy consumption optimization effect.

[0008] A machine learning optimization method for sorting, conveying, and loading parameters according to an embodiment of the present invention includes the following steps: S1. Collect multi-source data during the sorting, ash conveying, and loading process, and perform time synchronization processing, working condition slicing processing, and multi-dimensional indexing and encoding processing to generate a spatiotemporal working condition sequence. S2. Perform spectral domain decomposition on the spatiotemporal operating condition sequence, extract multi-scale frequency components using discrete wavelet transform, and generate the operating condition spectrum tensor by combining sliding window statistical results and ash flow fluctuation amplitude. S3. Perform graph structure mapping processing on the working condition spectrum tensor to construct a dynamic association graph of the ash conveying pipeline nodes, sorting nodes and loading nodes, and use graph attention network to calculate node coupling weights to generate a topology feature matrix. S4. Perform time-series prediction processing on the topological feature matrix, extract long-term dependent features using a bidirectional gated recurrent network, and calculate the parameter correlation strength by combining a multi-head attention mechanism to generate a parameter prediction sequence including ash conveying pressure parameters, fan frequency parameters, and loading flow rate. S5. Perform multi-objective reinforcement optimization processing on the parameter prediction sequence, construct a reward constraint function with ash conveying stability, sorting purity, loading continuity and equipment energy consumption, and use a policy iterative network to generate parameter control sequence. S6. Adjust the operating status of the ash conveying equipment and the ship loading equipment according to the parameter control sequence, perform error reverse mapping processing on the real-time operating results, and update the node coupling weight and reward constraint function in combination with the working condition deviation value.

[0009] Optionally, S1 specifically includes: S11. Multi-source data includes: Data on ash conveying pipeline pressure, fan frequency, valve opening, ash concentration, sorting particle size, loading flow rate, and silo status; S12. Time synchronization processing includes: The timestamps of each acquisition node are calibrated based on a unified master clock reference source. The clock offset between each data acquisition node is calculated by bidirectional timestamp alignment. The timestamps of each data sequence are corrected based on the clock offset. The corrected discrete sampling points are mapped to a unified time axis scale. Linear interpolation reconstruction is performed on the unaligned sampling points to generate a continuous time series. S13. Working condition slice processing includes: Sliding segmentation is performed on a continuous time series based on a fixed time window, dividing the continuous time series into time segments of equal length and generating time segment numbers; S14. Multidimensional index encoding processing includes: A composite index key consisting of device identifier code, node identifier code, time segment number and spatial location coordinates is constructed. Vectorized hash encoding is performed on the composite index key to generate an index vector. The index vector is then bound and mapped with multi-source data within the corresponding time segment to generate a spatiotemporal operating condition sequence.

[0010] Optionally, S2 specifically includes: S21. Perform spectral domain decomposition on the spatiotemporal operating condition sequence, including: The pressure sequence of ash conveying pipeline, the frequency sequence of fan, the valve opening sequence, the ash flow concentration sequence, the sorting particle size sequence, and the loading flow rate sequence in the spatiotemporal operating condition sequence are rearranged according to a unified time axis. S22. Perform discrete wavelet transform decomposition processing, and use preset wavelet basis functions to perform multi-scale decomposition on each variable sequence to generate multi-scale frequency sub-sequences containing low-frequency approximation components and high-frequency detail components. S23. Reconstruct and splice the decomposition results at different scales according to the scale hierarchy to form a set of frequency components; S24. Perform statistical feature extraction processing on the frequency component set within a sliding time window, calculate the mean, variance, kurtosis and skewness features in each window, and generate a sliding window statistical feature set. S25. Perform fluctuation amplitude calculation on the instantaneous difference between the ash flow concentration sequence and the ash conveying flow rate sequence to generate the ash flow fluctuation amplitude sequence; S26. Tensile splicing is performed on the frequency component set, the sliding window statistical feature set, and the gray flow fluctuation amplitude sequence according to the time alignment rule to construct a three-dimensional working condition spectrum tensor. Among them, the dimensions of the operating condition spectrum tensor correspond to the time dimension, frequency scale dimension, and feature channel dimension, respectively.

[0011] Optionally, S3 specifically includes: S31. Perform graph structure mapping processing on the operating condition spectrum tensor to map the time dimension feature slices in the operating condition spectrum tensor to node attribute inputs. S32. Construct an initial directed topology graph structure for the ash conveying pipeline nodes, sorting nodes, and loading nodes according to their physical connection relationships and material flow direction relationships, including: For each node, a node feature set is defined, which includes pressure spectrum feature vector, frequency component feature vector, ash flow fluctuation amplitude feature vector, and statistical feature vector. Define a set of edge features for each edge, including the difference in transport throughput, the time delay offset, and the equipment linkage response parameters; S33. Input the node feature set and edge feature set into the graph attention network to generate a topological feature matrix, including: Perform linear mapping on each node feature vector in the node feature set to generate the corresponding query vector matrix, key vector matrix and value vector matrix respectively; The key vector matrix and the value vector matrix are index-matched according to the node adjacency relationship to form a neighborhood feature alignment structure; Based on the adjacency constraint, the inner product similarity is calculated for the query vector and key vector of each central node and its neighboring nodes. The transport throughput difference feature, time delay offset feature and equipment linkage response parameter feature in the edge feature set are introduced into the similarity correction process to generate the edge constraint attention coefficient. The attention coefficients of the edge constraints are normalized to generate attention weights, and the value vectors of the corresponding neighboring nodes are weighted and aggregated according to the attention weights to generate the node embedding representation at a single time step. Perform time-dimensional stacking processing on the node embedding representations corresponding to different time segments; Cross-temporal dimension aggregation and splicing processing is performed on the node embedding representations at each time step to generate a topological feature matrix that integrates spatial topological dependencies and spectral evolution features.

[0012] Optionally, S4 specifically includes: S41. Perform temporal prediction processing on the topological feature matrix, split the topological feature matrix into input sequences according to the time dimension, and construct a temporal input tensor. S42. Input the temporal input tensor into the forward gated recurrent unit and the reverse gated recurrent unit respectively for bidirectional state encoding processing to generate a long-term dependency feature representation containing the forward hidden state sequence and the reverse hidden state sequence. S43. Concatenate the forward hidden state sequence and the reverse hidden state sequence in the time dimension to form a bidirectional temporal feature vector; S44. Construct a query matrix, key matrix, and value matrix based on bidirectional temporal feature vectors, and rearrange and match the key matrix and value matrix according to the time step alignment rule to form a cross-time association structure. S45. Utilize a multi-head attention mechanism to perform independent feature subspace mapping processing on different attention heads, and calculate the similarity between the query matrix and the key matrix within each attention head to generate an attention weight matrix, including: The bidirectional time series feature vector is divided into multiple independent subspaces according to the feature dimension, and an independent linear mapping matrix is ​​configured for each subspace to generate the corresponding query submatrix, key submatrix and value submatrix; Perform a dot product operation on the query submatrix and the key submatrix within each attention head; The time position encoding bias term and the node topological distance penalty term are superimposed on the dot product operation result, and a normalized exponential transformation is performed to generate the attention weight matrix; S46. The ash conveying pressure characteristics, fan frequency characteristics, and ship loading flow characteristics are introduced as constraint modulation factors into the attention weight correction process to generate the parameter correlation strength matrix. S47. Perform weighted fusion processing on the parameter correlation strength matrix and map it to generate a parameter prediction sequence containing the ash conveying pressure parameter sequence, the fan frequency parameter sequence, and the ship loading flow rate parameter sequence.

[0013] Optionally, S5 specifically includes: S51. Perform multi-objective enhancement optimization processing on the parameter prediction sequence, and expand the ash conveying pressure parameter sequence, fan frequency parameter sequence and ship loading flow parameter sequence in the parameter prediction sequence according to the time step to construct the state action sequence. S52. The current working condition state vector is concatenated with the parameter control vector of the previous time to form a reinforcement learning state vector. The reinforcement learning state vector is then input into the policy iterative network for policy evaluation and policy improvement. S53. In the process of strategy evaluation, construct a multi-objective reward constraint function consisting of ash conveying stability index, sorting purity index, loading continuity index and equipment energy consumption index. S54. Perform joint quantitative calculations on the pressure fluctuation amplitude, ash concentration deviation, ship loading flow fluctuation rate, and equipment power consumption of the ash conveying pipeline to generate a time-step reward feedback value sequence, and perform discount cumulative calculations on the reward feedback value sequence to generate a state value function. S55. In the policy improvement process, the gradient update and adjustment of the action probability distribution is performed based on the state value function. The execution probability of actions controlled by different parameters is reweighted and a constraint projection operator is introduced to perform clipping and mapping on the out-of-limit action space to generate a constrained policy distribution. S56. Perform sampling decision processing on the constrained policy distribution to generate the optimal action sequence, and reconstruct the optimal action sequence into a parameter control sequence by aligning it in time.

[0014] Optionally, S6 specifically includes: S61. Based on the parameter control sequence, the operating status of the ash conveying equipment and the loading equipment is adjusted in a closed loop. The ash conveying pressure control command, fan frequency control command and loading flow control command in the parameter control sequence are mapped to the corresponding actuator control quantities and issued synchronously according to the time step to form a real-time control command stream. S62. Synchronously collect and construct a real-time operation result vector of the pressure feedback value, ash concentration feedback value, sorting particle size feedback value and loading flow rate feedback value of the ash conveying pipeline. S63. Calculate the error vector by performing a dimension-by-dimensional difference calculation between the real-time running result vector and the target value of the corresponding time step of the parameter control sequence, and perform backpropagation mapping processing on the error vector to generate the control deviation gradient matrix. S64. Based on the control deviation gradient matrix, the node coupling weights in the topological feature matrix are recursively corrected and updated, and the edge weight parameters between the ash conveying node and the sorting node and the edge weight parameters between the sorting node and the loading node are synchronously adjusted. S65. Map the operating condition deviation value to the reward constraint function correction factor, and perform dynamic recalibration on the ash conveying stability constraint, sorting purity constraint, loading continuity constraint and equipment energy consumption constraint to generate the updated reward constraint function. S66. The updated node coupling weights and reward constraint functions are written back to the policy iteration network and graph attention network to form a closed-loop learning and update process.

[0015] The beneficial effects of this invention are: First, by performing unified time-series alignment, spectral decomposition, and graph structure modeling on multi-source heterogeneous data during the sorting, ash conveying, and loading process, a refined expression of the entire process operation status and multi-scale feature fusion were achieved, thereby improving the accuracy and stability of parameter modeling under complex working conditions.

[0016] Secondly, based on the bidirectional time-series prediction model and the multi-head attention mechanism, dynamic correlation learning of topological features is carried out, which effectively characterizes the nonlinear coupling relationship between ash conveying pressure, fan frequency and loading flow rate, and realizes the forward prediction of operating parameters and multi-dimensional collaborative optimization control.

[0017] Finally, by introducing a multi-objective reinforcement learning strategy iteration mechanism, a unified constraint optimization system was established among ash conveying stability, sorting purity, loading continuity and equipment energy consumption, thereby realizing adaptive dynamic adjustment of operating parameters and continuous optimization of long-term operating performance. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a machine learning optimization method for sorting, conveying, and loading parameters proposed in this invention. Figure 2 This is a flowchart of multi-source spatiotemporal data modeling and spectral tensor generation for a machine learning optimization method for sorting, conveying and loading parameters proposed in this invention; Figure 3 This is a flowchart of the graph structure mapping and attention topology modeling process for a machine learning optimization method for sorting, conveying and loading parameters proposed in this invention. Figure 4 This is a flowchart of the time-series prediction and reinforcement learning optimization process for a machine learning optimization method for sorting, conveying, and loading parameters proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-4 A machine learning optimization method for sorting, conveying, and loading parameters of ash onto ships includes the following steps: S1. Collect multi-source data during the sorting, ash conveying, and loading process, and perform time synchronization processing, working condition slicing processing, and multi-dimensional indexing and encoding processing to generate a spatiotemporal working condition sequence. S2. Perform spectral domain decomposition on the spatiotemporal operating condition sequence, extract multi-scale frequency components using discrete wavelet transform, and generate the operating condition spectrum tensor by combining sliding window statistical results and ash flow fluctuation amplitude. S3. Perform graph structure mapping processing on the working condition spectrum tensor to construct a dynamic association graph of the ash conveying pipeline nodes, sorting nodes and loading nodes, and use graph attention network to calculate node coupling weights to generate a topology feature matrix. S4. Perform time-series prediction processing on the topological feature matrix, extract long-term dependent features using a bidirectional gated recurrent network, and calculate the parameter correlation strength by combining a multi-head attention mechanism to generate a parameter prediction sequence including ash conveying pressure parameters, fan frequency parameters, and loading flow rate. S5. Perform multi-objective reinforcement optimization processing on the parameter prediction sequence, construct a reward constraint function with ash conveying stability, sorting purity, loading continuity and equipment energy consumption, and use a policy iterative network to generate parameter control sequence. S6. Adjust the operating status of the ash conveying equipment and the ship loading equipment according to the parameter control sequence, perform error reverse mapping processing on the real-time operating results, and update the node coupling weight and reward constraint function in combination with the working condition deviation value.

[0021] In this embodiment, S1 specifically includes: S11. Multi-source data includes: Data on ash conveying pipeline pressure, fan frequency, valve opening, ash concentration, sorting particle size, loading flow rate, and silo status; S12. Time synchronization processing includes: The timestamps of each acquisition node are calibrated based on a unified master clock reference source. The clock offset between each data acquisition node is calculated by bidirectional timestamp alignment. The timestamps of each data sequence are corrected based on the clock offset. The corrected discrete sampling points are mapped to a unified time axis scale. Linear interpolation reconstruction is performed on the unaligned sampling points to generate a continuous time series. S13. Working condition slice processing includes: Sliding segmentation is performed on a continuous time series based on a fixed time window, dividing the continuous time series into time segments of equal length and generating time segment numbers; S14. Multidimensional index encoding processing includes: A composite index key consisting of device identifier code, node identifier code, time segment number and spatial location coordinates is constructed. Vectorized hash encoding is performed on the composite index key to generate an index vector. The index vector is then bound and mapped with multi-source data within the corresponding time segment to generate a spatiotemporal operating condition sequence.

[0022] In this embodiment, S2 specifically includes: S21. Perform spectral domain decomposition on the spatiotemporal operating condition sequence, including: The pressure sequence of ash conveying pipeline, the frequency sequence of fan, the valve opening sequence, the ash flow concentration sequence, the sorting particle size sequence, and the loading flow rate sequence in the spatiotemporal operating condition sequence are rearranged according to a unified time axis. S22. Perform discrete wavelet transform decomposition processing, and use preset wavelet basis functions to perform multi-scale decomposition on each variable sequence to generate multi-scale frequency sub-sequences containing low-frequency approximation components and high-frequency detail components. S23. Reconstruct and splice the decomposition results at different scales according to the scale hierarchy to form a set of frequency components; S24. Perform statistical feature extraction processing on the frequency component set within a sliding time window, calculate the mean, variance, kurtosis and skewness features in each window, and generate a sliding window statistical feature set. S25. Perform fluctuation amplitude calculation on the instantaneous difference between the ash flow concentration sequence and the ash conveying flow rate sequence to generate the ash flow fluctuation amplitude sequence; S26. Tensile splicing is performed on the frequency component set, the sliding window statistical feature set, and the gray flow fluctuation amplitude sequence according to the time alignment rule to construct a three-dimensional working condition spectrum tensor. Among them, the dimensions of the operating condition spectrum tensor correspond to the time dimension, frequency scale dimension, and feature channel dimension, respectively.

[0023] In this embodiment, S3 specifically includes: S31. Perform graph structure mapping processing on the operating condition spectrum tensor to map the time dimension feature slices in the operating condition spectrum tensor to node attribute inputs. S32. Construct an initial directed topology graph structure for the ash conveying pipeline nodes, sorting nodes, and loading nodes according to their physical connection relationships and material flow direction relationships, including: For each node, a node feature set is defined, which includes pressure spectrum feature vector, frequency component feature vector, ash flow fluctuation amplitude feature vector, and statistical feature vector. Define a set of edge features for each edge, including the difference in transport throughput, the time delay offset, and the equipment linkage response parameters; S33. Input the node feature set and edge feature set into the graph attention network to generate a topological feature matrix, including: Perform linear mapping on each node feature vector in the node feature set to generate the corresponding query vector matrix, key vector matrix and value vector matrix respectively; The key vector matrix and the value vector matrix are index-matched according to the node adjacency relationship to form a neighborhood feature alignment structure; Based on the adjacency constraint, the inner product similarity is calculated for the query vector and key vector of each central node and its neighboring nodes. The transport throughput difference feature, time delay offset feature and equipment linkage response parameter feature in the edge feature set are introduced into the similarity correction process to generate the edge constraint attention coefficient. The attention coefficients of the edge constraints are normalized to generate attention weights, and the value vectors of the corresponding neighboring nodes are weighted and aggregated according to the attention weights to generate the node embedding representation at a single time step. Perform time-dimensional stacking processing on the node embedding representations corresponding to different time segments; Cross-temporal dimension aggregation and splicing processing is performed on the node embedding representations at each time step to generate a topological feature matrix that integrates spatial topological dependencies and spectral evolution features.

[0024] In this embodiment, S4 specifically includes: S41. Perform temporal prediction processing on the topological feature matrix, split the topological feature matrix into input sequences according to the time dimension, and construct a temporal input tensor. S42. Input the temporal input tensor into the forward gated recurrent unit and the reverse gated recurrent unit respectively for bidirectional state encoding processing to generate a long-term dependency feature representation containing the forward hidden state sequence and the reverse hidden state sequence. S43. Concatenate the forward hidden state sequence and the reverse hidden state sequence in the time dimension to form a bidirectional temporal feature vector; S44. Construct a query matrix, key matrix, and value matrix based on bidirectional temporal feature vectors, and rearrange and match the key matrix and value matrix according to the time step alignment rule to form a cross-time association structure. S45. Utilize a multi-head attention mechanism to perform independent feature subspace mapping processing on different attention heads, and calculate the similarity between the query matrix and the key matrix within each attention head to generate an attention weight matrix, including: The bidirectional time series feature vector is divided into multiple independent subspaces according to the feature dimension, and an independent linear mapping matrix is ​​configured for each subspace to generate the corresponding query submatrix, key submatrix and value submatrix; Perform a dot product operation on the query submatrix and the key submatrix within each attention head; The time position encoding bias term and the node topological distance penalty term are superimposed on the dot product operation result, and a normalized exponential transformation is performed to generate the attention weight matrix; S46. The ash conveying pressure characteristics, fan frequency characteristics, and ship loading flow characteristics are introduced as constraint modulation factors into the attention weight correction process to generate the parameter correlation strength matrix. S47. Perform weighted fusion processing on the parameter correlation strength matrix and map it to generate a parameter prediction sequence containing the ash conveying pressure parameter sequence, the fan frequency parameter sequence, and the ship loading flow rate parameter sequence.

[0025] In this embodiment, S5 specifically includes: S51. Perform multi-objective enhancement optimization processing on the parameter prediction sequence, and expand the ash conveying pressure parameter sequence, fan frequency parameter sequence and ship loading flow parameter sequence in the parameter prediction sequence according to the time step to construct the state action sequence. S52. The current working condition state vector is concatenated with the parameter control vector of the previous time to form a reinforcement learning state vector. The reinforcement learning state vector is then input into the policy iterative network for policy evaluation and policy improvement. S53. In the process of strategy evaluation, construct a multi-objective reward constraint function consisting of ash conveying stability index, sorting purity index, loading continuity index and equipment energy consumption index. S54. Perform joint quantitative calculations on the pressure fluctuation amplitude, ash concentration deviation, ship loading flow rate fluctuation rate, and equipment power consumption of the ash conveying pipeline to generate a time-step reward feedback value sequence. Then, perform a discount and cumulative calculation on the reward feedback value sequence to generate a state value function, including: The pressure fluctuation amplitude sequence of the ash conveying pipeline is calculated by differential calculation of adjacent time steps. A concentration deviation value sequence is generated by calculating the point-by-point deviation between the ash flow concentration sequence and the target concentration benchmark value. The flow volatility series is calculated based on the ratio of the standard deviation to the mean within the sliding window for the loading flow series; The power consumption sequence of the equipment is normalized to generate a power consumption sequence. The pressure fluctuation amplitude sequence, concentration deviation value sequence, flow fluctuation rate sequence and power consumption sequence are vectorized and concatenated step by step according to the time alignment rule to generate the original cost vector. The original cost vector is linearly weighted and summed according to preset weight coefficients, and the state value function is generated by recursive cumulative calculation according to the time discount factor; S55. In the policy improvement process, the gradient update and adjustment of the action probability distribution is performed based on the state value function. The execution probability of actions controlled by different parameters is reweighted and a constraint projection operator is introduced to perform clipping and mapping on the out-of-limit action space to generate a constrained policy distribution. S56. Perform sampling decision processing on the constrained policy distribution to generate the optimal action sequence, and reconstruct the optimal action sequence into a parameter control sequence by aligning it in time.

[0026] In this embodiment, S6 specifically includes: S61. Based on the parameter control sequence, the operating status of the ash conveying equipment and the loading equipment is adjusted in a closed loop. The ash conveying pressure control command, fan frequency control command and loading flow control command in the parameter control sequence are mapped to the corresponding actuator control quantities and issued synchronously according to the time step to form a real-time control command stream. S62. Synchronously collect and construct a real-time operation result vector of the pressure feedback value, ash concentration feedback value, sorting particle size feedback value and loading flow rate feedback value of the ash conveying pipeline. S63. Calculate the error vector by performing a dimension-by-dimensional difference calculation between the real-time running result vector and the target value of the corresponding time step of the parameter control sequence, and perform backpropagation mapping processing on the error vector to generate the control deviation gradient matrix. S64. Based on the control deviation gradient matrix, the node coupling weights in the topological feature matrix are recursively corrected and updated, and the edge weight parameters between the ash conveying node and the sorting node and the edge weight parameters between the sorting node and the loading node are synchronously adjusted. S65. Map the operating condition deviation value to the reward constraint function correction factor, and perform dynamic recalibration on the ash conveying stability constraint, sorting purity constraint, loading continuity constraint and equipment energy consumption constraint to generate the updated reward constraint function. S66. The updated node coupling weights and reward constraint functions are written back to the policy iteration network and graph attention network to form a closed-loop learning and update process.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a large-scale continuous operation scenario for bulk material sorting, ash conveying, and loading. In this scenario, the conveyed material is a mixture of high-ash powder and fine particles, characterized by uneven particle size distribution, significant moisture content fluctuations, and susceptibility to local blockages and sorting efficiency fluctuations during conveying. Furthermore, the ash conveying pipeline is long, and the conveying path includes multi-stage sorting devices and multi-compartment loading structures. The overall system exhibits strong coupling, multivariate disturbances, and nonlinear dynamic changes. Under traditional control methods, this scenario typically relies on experience to set the ash conveying pressure and fan frequency parameters. When the material particle size changes or the moisture content increases, the system is prone to problems such as increased conveying pressure fluctuations, decreased sorting accuracy, and discontinuous loading cycles. Simultaneously, during the collaborative operation of multiple devices, the lack of a unified optimization mechanism among the various stages leads to a significant situation where some devices operate stably but overall efficiency declines.

[0028] In this embodiment, the pressure sensor of the ash conveying pipeline, the fan frequency acquisition unit, the valve opening detection unit, the ash flow concentration detection unit, the particle size online analysis unit, and the ship loading flow metering unit are uniformly connected to construct a multi-source time-series data acquisition link. The acquired data is then processed for time synchronization, mapping data from different sampling frequencies onto a unified time axis to form a continuous time-series sequence. Based on this, a sliding window slicing process is performed on the time-series data, forming a working condition segment every 5 seconds. Within each working condition segment, the pressure fluctuation amplitude, flow fluctuation rate, concentration deviation, and equipment power change characteristics are extracted to construct an initial working condition spectrum representation.

[0029] The aforementioned spectral data is then input into a graph structure mapping unit, constructing a dynamic topology structure for the ash conveying nodes, sorting nodes, and loading nodes. A graph attention mechanism is used to calculate the coupling weights between nodes, enabling the system to characterize the transmission effect of ash conveying pressure changes on sorting accuracy and loading flow rate. Furthermore, the topological features are input into a bidirectional cyclic time-series prediction network to jointly predict ash conveying pressure, fan frequency, and loading flow rate in the short term. A multi-head attention mechanism is then used to calculate the correlation strength between different parameters, thereby generating a parameter prediction sequence.

[0030] During the parameter optimization phase, a multi-objective reinforcement learning strategy is introduced, using ash conveying stability, sorting purity, loading continuity, and equipment energy consumption as a unified reward constraint system to dynamically evaluate different control actions. The optimal parameter control sequence is then generated through a strategy iterative network. During execution, the system continuously updates node coupling weights and reward functions based on real-time feedback data, enabling the control strategy to have adaptive evolution capabilities and thus achieving long-term stable optimization.

[0031] In actual operation comparison tests, statistical analysis was conducted using 72 hours of continuous operation data under the same transport load conditions. The comparison between the traditional control method and the method of this invention is shown in the table below: Table 1 Comparison of Optimization Effects of Sorting, Ash Conveying, and Ship Loading Parameters

[0032] As shown in Table 1, without changing the hardware structure, this invention significantly improves the overall stability of the ash conveying system through time-series modeling of multi-source data, graph structure coupling analysis, and reinforcement learning-based optimized control. Pressure and flow fluctuations during the conveying process are significantly reduced, while sorting accuracy and loading continuity are simultaneously improved, and the unit energy consumption of the equipment is effectively reduced. Even under complex conditions such as sudden increases in ash moisture content or large changes in particle size distribution, the system can still quickly adjust control parameters through online strategy updates, avoiding the lag and local instability problems common in traditional methods, thereby significantly improving overall conveying efficiency and operational reliability.

[0033] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine learning optimization method for sorting, conveying, and loading parameters of ash onto ships, characterized in that, Includes the following steps: S1. Collect multi-source data during the sorting, ash conveying, and loading process, and perform time synchronization processing, working condition slicing processing, and multi-dimensional indexing and encoding processing to generate a spatiotemporal working condition sequence. S2. Perform spectral domain decomposition on the spatiotemporal operating condition sequence, extract multi-scale frequency components using discrete wavelet transform, and generate the operating condition spectrum tensor by combining the sliding window statistical results and the ash flow fluctuation amplitude. S3. Perform graph structure mapping processing on the working condition spectrum tensor to construct a dynamic association graph of the ash conveying pipeline nodes, sorting nodes and loading nodes, and use graph attention network to calculate node coupling weights to generate a topology feature matrix. S4. Perform time-series prediction processing on the topological feature matrix, extract long-term dependent features using a bidirectional gated recurrent network, and calculate the parameter correlation strength by combining a multi-head attention mechanism to generate a parameter prediction sequence including ash conveying pressure parameters, fan frequency parameters, and loading flow rate. S5. Perform multi-objective reinforcement optimization processing on the parameter prediction sequence, construct a reward constraint function with ash conveying stability, sorting purity, loading continuity and equipment energy consumption, and use a policy iterative network to generate parameter control sequence. S6. Adjust the operating status of the ash conveying equipment and the ship loading equipment according to the parameter control sequence, perform error reverse mapping processing on the real-time operating results, and update the node coupling weight and reward constraint function in combination with the working condition deviation value.

2. The machine learning optimization method for sorting, conveying, and loading parameters of ash cargo according to claim 1, characterized in that, S1 specifically includes: S11. Multi-source data includes: Data on ash conveying pipeline pressure, fan frequency, valve opening, ash concentration, sorting particle size, loading flow rate, and silo status; S12. Time synchronization processing includes: The timestamps of each acquisition node are calibrated based on a unified master clock reference source. The clock offset between each data acquisition node is calculated by bidirectional timestamp alignment. The timestamps of each data sequence are corrected based on the clock offset. The corrected discrete sampling points are mapped to a unified time axis scale. Linear interpolation reconstruction is performed on the unaligned sampling points to generate a continuous time series. S13. Working condition slice processing includes: Sliding segmentation is performed on a continuous time series based on a fixed time window, dividing the continuous time series into time segments of equal length and generating time segment numbers; S14. Multidimensional index encoding processing includes: A composite index key consisting of device identifier code, node identifier code, time segment number and spatial location coordinates is constructed. Vectorized hash encoding is performed on the composite index key to generate an index vector. The index vector is then bound and mapped with multi-source data within the corresponding time segment to generate a spatiotemporal operating condition sequence.

3. The machine learning optimization method for sorting, conveying, and loading parameters of ash as described in claim 1, characterized in that, S2 specifically includes: S21. Perform spectral domain decomposition on the spatiotemporal operating condition sequence, including: The pressure sequence of ash conveying pipeline, the frequency sequence of fan, the valve opening sequence, the ash flow concentration sequence, the sorting particle size sequence, and the loading flow rate sequence in the spatiotemporal operating condition sequence are rearranged according to a unified time axis. S22. Perform discrete wavelet transform decomposition processing, and use preset wavelet basis functions to perform multi-scale decomposition on each variable sequence to generate multi-scale frequency sub-sequences containing low-frequency approximation components and high-frequency detail components. S23. Reconstruct and splice the decomposition results at different scales according to the scale hierarchy to form a set of frequency components; S24. Perform statistical feature extraction processing on the frequency component set within a sliding time window, calculate the mean, variance, kurtosis and skewness features in each window, and generate a sliding window statistical feature set. S25. Perform fluctuation amplitude calculation on the instantaneous difference between the ash flow concentration sequence and the ash conveying flow rate sequence to generate the ash flow fluctuation amplitude sequence; S26. Tensile splicing is performed on the frequency component set, the sliding window statistical feature set, and the gray flow fluctuation amplitude sequence according to the time alignment rule to construct a three-dimensional working condition spectrum tensor. Among them, the dimensions of the operating condition spectrum tensor correspond to the time dimension, frequency scale dimension, and feature channel dimension, respectively.

4. The machine learning optimization method for sorting, conveying, and loading parameters of ash cargo according to claim 1, characterized in that, S3 specifically includes: S31. Perform graph structure mapping processing on the operating condition spectrum tensor to map the time dimension feature slices in the operating condition spectrum tensor to node attribute inputs. S32. Construct an initial directed topology graph structure for the ash conveying pipeline nodes, sorting nodes, and loading nodes according to their physical connection relationships and material flow direction relationships, including: For each node, a node feature set is defined, which includes pressure spectrum feature vector, frequency component feature vector, ash flow fluctuation amplitude feature vector, and statistical feature vector; Define a set of edge features for each edge, including the difference in transport throughput, the time delay offset, and the equipment linkage response parameters; S33. Input the node feature set and edge feature set into the graph attention network to generate a topological feature matrix, including: Perform linear mapping on each node feature vector in the node feature set to generate the corresponding query vector matrix, key vector matrix and value vector matrix respectively; The key vector matrix and the value vector matrix are index-matched according to the node adjacency relationship to form a neighborhood feature alignment structure; Based on the adjacency constraint, the inner product similarity is calculated for the query vector and key vector of each central node and its neighboring nodes; The transport throughput difference feature, time delay offset feature and equipment linkage response parameter feature in the edge feature set are introduced into the similarity correction process to generate the edge constraint attention coefficient. The attention coefficients of the edge constraints are normalized to generate attention weights, and the value vectors of the corresponding neighboring nodes are weighted and aggregated according to the attention weights to generate the node embedding representation at a single time step. Perform time-dimensional stacking processing on the node embedding representations corresponding to different time segments; The node embedding representations at each time step are aggregated and spliced ​​across time dimensions to generate a topological feature matrix that integrates spatial topological dependencies and spectral evolution features.

5. The machine learning optimization method for sorting, conveying, and loading parameters of ash cargo according to claim 1, characterized in that, S4 specifically includes: S41. Perform temporal prediction processing on the topological feature matrix, split the topological feature matrix into input sequences according to the time dimension, and construct a temporal input tensor. S42. Input the temporal input tensor into the forward gated recurrent unit and the reverse gated recurrent unit respectively for bidirectional state encoding processing to generate a long-term dependency feature representation containing the forward hidden state sequence and the reverse hidden state sequence. S43. Concatenate the forward hidden state sequence and the reverse hidden state sequence in the time dimension to form a bidirectional temporal feature vector; S44. Construct a query matrix, key matrix, and value matrix based on bidirectional temporal feature vectors, and rearrange and match the key matrix and value matrix according to the time step alignment rule to form a cross-time association structure. S45. Utilize a multi-head attention mechanism to perform independent feature subspace mapping processing on different attention heads, and perform similarity calculation on the query matrix and key matrix within each attention head to generate an attention weight matrix; S46. The ash conveying pressure characteristics, fan frequency characteristics, and ship loading flow characteristics are introduced as constraint modulation factors into the attention weight correction process to generate the parameter correlation strength matrix. S47. Perform weighted fusion processing on the parameter correlation strength matrix and map it to generate a parameter prediction sequence containing the ash conveying pressure parameter sequence, the fan frequency parameter sequence, and the ship loading flow rate parameter sequence.

6. The machine learning optimization method for sorting, conveying, and loading parameters of ash as described in claim 1, characterized in that, S5 specifically includes: S51. Perform multi-objective enhancement optimization processing on the parameter prediction sequence, and expand the ash conveying pressure parameter sequence, fan frequency parameter sequence and ship loading flow parameter sequence in the parameter prediction sequence according to the time step to construct the state action sequence. S52. The current working condition state vector is concatenated with the parameter control vector of the previous time to form a reinforcement learning state vector. The reinforcement learning state vector is then input into the policy iterative network for policy evaluation and policy improvement. S53. In the process of strategy evaluation, construct a multi-objective reward constraint function consisting of ash conveying stability index, sorting purity index, loading continuity index and equipment energy consumption index. S54. Perform joint quantitative calculations on the pressure fluctuation amplitude, ash concentration deviation, ship loading flow fluctuation rate, and equipment power consumption of the ash conveying pipeline to generate a time-step reward feedback value sequence, and perform discount cumulative calculations on the reward feedback value sequence to generate a state value function. S55. In the policy improvement process, the gradient update and adjustment of the action probability distribution is performed based on the state value function. The execution probability of actions controlled by different parameters is reweighted and a constraint projection operator is introduced to perform clipping and mapping on the out-of-limit action space to generate a constrained policy distribution. S56. Perform sampling decision processing on the constrained policy distribution to generate the optimal action sequence, and reconstruct the optimal action sequence into a parameter control sequence by aligning it in time.

7. The machine learning optimization method for sorting, conveying, and loading parameters of ash cargo according to claim 1, characterized in that, S6 specifically includes: S61. Based on the parameter control sequence, the operating status of the ash conveying equipment and the loading equipment is adjusted in a closed loop. The ash conveying pressure control command, fan frequency control command and loading flow control command in the parameter control sequence are mapped to the corresponding actuator control quantities and issued synchronously according to the time step to form a real-time control command stream. S62. Synchronously collect and construct a real-time operation result vector of the pressure feedback value, ash concentration feedback value, sorting particle size feedback value and loading flow rate feedback value of the ash conveying pipeline. S63. Calculate the error vector by performing a dimension-by-dimensional difference calculation between the real-time running result vector and the target value of the corresponding time step of the parameter control sequence, and perform backpropagation mapping processing on the error vector to generate the control deviation gradient matrix. S64. Based on the control deviation gradient matrix, the node coupling weights in the topological feature matrix are recursively corrected and updated, and the edge weight parameters between the ash conveying node and the sorting node and the edge weight parameters between the sorting node and the loading node are synchronously adjusted. S65. Map the operating condition deviation value to the reward constraint function correction factor, and perform dynamic recalibration on the ash conveying stability constraint, sorting purity constraint, loading continuity constraint and equipment energy consumption constraint to generate the updated reward constraint function. S66. The updated node coupling weights and reward constraint functions are written back to the policy iteration network and the graph attention network to form a closed-loop learning and update process.

Images