Coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion

Abstract

The invention relates to the technical field of industrial detection data processing, and discloses a coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion, which comprises a priori constraint domain construction module, a state space mapping module, a multi-agent cooperative calculation module and a strategy game and fusion output module, according to the system, a calculation framework embedded with a coal physical and chemical topological graph is constructed, the dynamic game of a sensing agent and a cognitive agent is utilized, and feature reweighting and graph path searching are executed under the constraint of a physical rule. Based on the verification capability of the system on the causality of a reasoning path, the problems of logic collapse and pseudo-correlation reasoning caused by lack of physical constraints of a data-driven model under the working conditions of strong noise and small samples are effectively solved, and the logic self-consistency and environmental stability of heterogeneous material detection are improved.

Description

Technical Field

[0001] This invention relates to a coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion, belonging to the field of industrial detection data processing technology. Background Technology

[0002] Current mainstream data analysis for heterogeneous materials uses end-to-end computational models based on statistical regression or deep neural networks. These models are based on the assumption of independent and identically distributed data, fitting the distribution patterns of historical data to construct numerical mapping relationships between sensor signals and physicochemical indicators. However, the heterogeneity of the measured objects in industrial settings and the time-varying nature of environmental parameters lead to nonlinear drift in the characteristic distribution. Purely data-driven models lack an understanding of physical laws and rely on numerical correlation fitting. They are prone to outputting high-confidence but incorrect predictions that violate basic chemical laws under long-tail conditions or strong noise interference, resulting in insufficient logical consistency and stability of the computational system.

[0003] Existing technologies attempt to address data credibility and timeliness issues by introducing distributed ledgers and edge processing mechanisms. For example, Chinese invention patent CN116911670A discloses a coal testing data processing method and system. This method encrypts testing data at the edge and constructs blockchain transactions. It uses a pre-trained machine learning model to extract features and obtain quality analysis data. Although this improves the data transmission and storage anti-tampering capabilities and security, the core analysis process still uses the traditional black-box machine learning approach. It fails to solve the problem of the disconnect between the reasoning model and the physicochemical mechanism. It cannot establish endogenous constraints based on the conservation of material components or the symbiotic relationship of elements within the model when processing high-dimensional feature data.

[0004] Therefore, how to construct a computational analysis system that embeds deterministic domain knowledge logic into a stochastic numerical reasoning process and achieves the ability to verify physical constraints and logical consistency under non-stationary conditions has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of this invention is as follows: A coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion, the system comprising:

[0006] The knowledge graph storage module is used to store the coal quality attribute relationship graph. The graph includes entity nodes that characterize the physical and chemical properties of coal and the associated edges that define the positive promotion or negative inhibition relationships between entity nodes.

[0007] The data preprocessing module is used to acquire the detection and sensing data stream and map it into an input feature tensor;

[0008] The multi-agent collaborative computing module includes mutually coupled perceptual agents, cognitive agents, and a logical consistency verification unit;

[0009] The perceptual agent is used to calculate the gain weight vector of the feature channels based on the input feature tensor, and then uses the gain weight vector to perform weighting on the perceptual data stream to generate the target feature vector.

[0010] A cognitive agent is used to perform path search in the coal quality attribute relationship graph using the target feature vector as an index, and generate candidate semantic paths connecting the input node to the output node;

[0011] The logical consistency verification unit performs feature mask verification before outputting the results: it locks the dominant feature dimension in the subset of input features corresponding to the candidate semantic path; it performs a masking operation on the input feature tensor to set the value of the dominant feature dimension to zero and generates a mask feature tensor; it calls the cognitive agent to perform a secondary path search based on the mask feature tensor and generate a perturbation semantic path; it calculates the structural similarity between the candidate semantic path and the perturbation semantic path; in response to the structural similarity being lower than a preset consistency threshold, it confirms the validity of the candidate semantic path and calculates the output analysis results; in response to the structural similarity being higher than the consistency threshold, it determines the candidate semantic path as a pseudo-association path and triggers a negative reward update for the cognitive agent.

[0012] Preferably, the associated edges in the knowledge graph storage module include proportionally constant edges connecting two entity nodes with a chemical symbiotic relationship; the perceptual agent is also used to perform drift compensation operations: calculating the ratio fluctuation index of the feature dimensions of the entity nodes corresponding to the two ends of the proportionally constant edge in the perceptual data stream; in response to the ratio fluctuation index being in a preset stable range and the absolute amplitude of the feature dimension changing monotonically, determining it as a systematic drift and calculating a global gain compensation factor; using the global gain compensation factor to perform multiplicative correction on the perceptual data stream; in response to the ratio fluctuation index exceeding the stable range, determining it as a change in component structure and triggering the cognitive agent to adjust the search strategy.

[0013] Preferably, the multi-agent collaborative computing module is also used to perform cold start acceleration operations; the perceptual agent is used to calculate the outlier index of the input feature tensor relative to historical experience data; in response to the outlier index being higher than a preset threshold, the cognitive agent is used to retrieve the reference semantic path with the highest morphological similarity to the target feature vector from the historical path database; the path skeleton of the reference semantic path is extracted to generate an initialized path search probability distribution matrix; and the path search probability distribution matrix is ​​used to constrain the node transition probability of the first round of path search.

[0014] Preferably, the associated edges have variable transfer weight attributes; the system also includes an environmental adaptive adjustment unit; the environmental adaptive adjustment unit is used to obtain the environmental state vector at the current moment, calculate the weight correction coefficient for different types of associated edges using a pre-trained mapping model, and dynamically adjust the transfer weight value of the associated edges using the weight correction coefficient; the cognitive agent is used to calculate the transfer probability between nodes based on the adjusted transfer weight value, wherein when the environmental state vector represents a strong interference condition, the interference source corresponding to the strong interference condition is identified, and the transfer weight value of the associated edge physically related to the interference source is increased to reduce the selection probability of the path.

[0015] Preferably, the perceptual agent adopts a deep neural network structure based on policy gradient; the perceptual agent outputs a set of feature mask vectors as the action space, and the dimension of the feature mask vectors is consistent with the number of spectral channels of the perceptual data stream; the logical reward function includes a sparsity penalty term to constrain the number of non-zero elements in the feature mask vectors, so that the perceptual agent selects the smallest complete feature subset.

[0016] Preferably, when calculating structural similarity, the logical consistency verification unit performs the following operations: extracting the first node sequence contained in the candidate semantic path and the second node sequence contained in the perturbed semantic path; calculating the length of the longest common subsequence of the first node sequence and the second node sequence; determining the structural similarity based on the ratio of the length of the longest common subsequence to the total path length; if the ratio is greater than a preset judgment threshold, it is determined that the reasoning logic is not blocked due to feature loss, and the original reasoning is determined to be a pseudo-association that depends on non-key features.

[0017] Preferably, the ratio fluctuation index calculated during drift compensation operation Satisfies the following definition: ,in, and These are the real-time observations of the feature dimensions corresponding to the entity nodes at both ends of the proportionally constant edge. Represents variance. Represents the mean; the perceptual agent only... When the drift is less than the preset drift threshold, the global gain compensation factor is calculated.

[0018] Preferably, the strategy game and fusion output module is used to perform path weighted calculation; the cognitive agent is used to output multiple candidate semantic paths and their corresponding path confidence scores; the strategy game and fusion output module is used to extract the attribute values ​​of the activated nodes on each candidate semantic path, and use the path confidence scores to perform weighted summation of the attribute values ​​to obtain the final analysis results; the path confidence scores are jointly calculated based on the weight product of all associated edges on the path and the feedback value of the logical reward function.

[0019] Preferably, the knowledge graph storage module is also used to perform dynamic graph labeling; in response to the continuous triggering of negative reward updates for specific types of working conditions by the logical consistency verification unit, the system automatically marks the invalid subgraph region under the working condition in the coal quality attribute relationship graph; in subsequent reasoning, the cognitive agent is prohibited from wandering to the invalid subgraph region or the transition probability of entering the invalid subgraph region is reduced.

[0020] Preferably, the data preprocessing module is used to perform multimodal feature stitching; the data preprocessing module acquires spectral detection data and environmental sensor data; the first encoder maps the spectral detection data into spectral feature vectors, and the second encoder maps the environmental sensor data into environmental embedding vectors; the environmental embedding vectors are expanded through a broadcast operation and then stitched with the spectral feature vectors in the channel dimension to generate an input feature tensor, so that the perceptual agent introduces the nonlinear influence of the environmental state on the spectral features during weighted calculation.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. In coal detection data calculation, a two-layer intelligent agent game computing architecture constrained by a static coal quality physicochemical topology map is constructed to solve the problems of data-driven model logic collapse and small sample generalization failure in the analysis of heterogeneous material detection data. By utilizing the strong coupling chemical constraint relationship between element nodes in the map, the spatial boundary of the probabilistic walk action of the cognitive agent is defined. The perceptual agent's reweighting strategy for the features of the input heterogeneous data stream forms a dynamic game interlock with the cognitive agent's map path search strategy, so that the final generated analysis result is the optimal semantic interpretation path under the constraint of physical laws. The deterministic symbolic logic is embedded in the stochastic numerical reasoning process calculation method to eliminate erroneous predictions that violate the basic chemical laws due to sensor drift or sample distribution deviation, and ensure the self-consistency of the physical meaning of the system output.

[0023] 2. The system utilizes a logic stability verification unit to perform counterfactual perturbation-based topology comparison. Based on the computational system's endogenous verification capability of the causality of inference paths, it locks the optimal semantic path and corresponding key input feature subset. It performs a forced masking operation on the dominant feature dimension of the global state tensor and triggers a secondary path search to construct a counterfactual computational state. The system compares the topological similarity between the original path and the perturbation path, qualitatively identifies and eliminates pseudo-associative inference paths induced by random fluctuations in background noise and accidental strong correlations between target index values. It relies solely on internal logical adversarial verification within the computational model rather than external physical standard verification methods, without adding additional hardware detection steps, thus improving the system's feature validity identification accuracy and logic stability under strong noise interference conditions.

[0024] 3. By using the constant-proportion edges in the topological graph to characterize the chemical invariant features, a calculation benchmark is constructed to distinguish between instrument drift in the physical domain and component changes in the chemical domain. The sensing agent calculates the stability index of the feature dimension ratio of the corresponding symbiotic entity node in real time. In the absence of an external standard source, the overall amplitude attenuation of the spectrum caused by hardware aging or ash accumulation is decoupled from the component structure change caused by coal type changes. For scenarios identified as physical drift, a global gain compensation factor is generated to correct the input tensor multiplication, realizing adaptive calibration of the calculation benchmark. The time-varying calculation strategy of the measurement equipment is corrected by using the chemical structure invariance of the measured object itself, ensuring the consistency and long-term stability of the inference model in the face of nonlinear environmental interference throughout its entire life cycle. Attached Figure Description

[0025] Figure 1 This is a system logic architecture diagram of the multi-agent collaborative computing and knowledge graph fusion of the present invention;

[0026] Figure 2 This is a trend diagram showing the impact of the structural similarity threshold of the logic verification unit of this invention on performance indicators;

[0027] Figure 3 This is a diagram showing the physical deployment and multi-level hardware topology of the present invention in the industrial field. Detailed Implementation

[0028] The embodiments of the present invention are intended to provide a clear and complete description of the technical solutions defined in this application, but these embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0029] A coal detection data computation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion comprises a knowledge graph storage module, a data preprocessing module, a multi-agent collaborative computing module, and a strategy game and fusion output module. The knowledge graph storage module stores a pre-constructed coal quality attribute relationship graph, which serves as a prior constraint domain. It includes entity nodes characterizing the physicochemical properties of coal and association edges defining positive or negative facilitation relationships between these entity nodes. Entity nodes include physicochemical index objects such as carbon nodes, hydrogen nodes, ash composition nodes, and calorific value nodes. The association edges include a type of pre-labeled constant-proportion edge, which connects two entity nodes with a chemical symbiotic relationship, such as connecting silicon nodes and aluminum nodes in an aluminosilicate mineral structure. Element nodes are used to characterize the fixed molar ratio relationship between two elements constrained by the crystal structure. The data preprocessing module is used to perform feature mapping and fusion of multimodal data. This module acquires spectral detection data and environmental sensor data, including temperature, humidity, and specular contamination data. This module uses a first encoder, such as a one-dimensional convolutional neural network, to map the spectral detection data into spectral feature vectors, and uses a second encoder, such as a multilayer perceptron, to map the environmental sensor data into environmental embedding vectors. This module performs a broadcast operation to extend the environmental embedding vectors to the same dimension as the spectral feature vectors, and concatenates the two in the channel dimension to generate an input feature tensor. This input feature tensor explicitly contains information on the nonlinear influence of environmental state on spectral features.

[0030] The multi-agent collaborative computing module comprises mutually coupled perceptual agents, cognitive agents, and a logical consistency verification unit. The perceptual agent employs a deep neural network structure based on policy gradients. It takes an input feature tensor as input and outputs a set of feature mask vectors as its action space. The dimension of the feature mask vectors is consistent with the number of spectral channels in the perceptual data stream, representing the gain weights of each channel. The perceptual agent performs point-by-point multiplication weighting on the perceptual data stream using the feature mask vectors to generate a target feature vector. This process suppresses feature bands affected by environmental noise and enhances effective feature bands. The logical reward function includes a sparsity penalty term, such as an L1 norm regularization term, to constrain the number of non-zero elements in the feature mask vector, enabling the perceptual agent to select the minimum complete feature subset. The cognitive agent performs graph path search. Using the target feature vector output by the perceptual agent as an index, it starts from the input node in the coal quality attribute relationship graph and performs probability transfer based on the weights of the associated edges, generating several candidate semantic paths connecting the input node to the output node. The logical consistency verification unit... This unit is used to perform topological structure comparison based on counterfactual masks to verify the causality of inference paths. Before outputting the analysis results, it performs the following procedures: locking the dominant feature dimension corresponding to the candidate semantic path generated by the cognitive agent in the input space; performing a masking operation on the input feature tensor, setting the value of the dominant feature dimension to zero, and generating a counterfactual mask feature tensor; keeping the agent's policy parameters unchanged, calling the cognitive agent to perform a secondary path search based on the mask feature tensor to generate a perturbed semantic path; calculating the structural similarity between the candidate semantic path and the perturbed semantic path. The structural similarity is calculated as follows: extracting the first node sequence of the candidate semantic path and the second node sequence of the perturbed semantic path, calculating the length of their longest common subsequence, and using the ratio of this length to the total path length as the structural similarity. If the structural similarity is lower than a preset consistency threshold, such as 0.4, it is confirmed that the original inference logic has a strong causal dependency, and the candidate semantic path is determined to be valid; if the structural similarity is higher than the consistency threshold, it is determined that the original inference depends on a pseudo-association of non-key features, triggering a negative reward update for the cognitive agent.

[0031] The perceptual agent is also used to perform drift compensation operations based on chemical invariants. The system monitors the feature dimensions of entity nodes corresponding to the two ends of the proportional constant edge in the perceptual data stream in real time, and the perceptual agent calculates the ratio fluctuation index. Its calculation formula is: where, and These are the real-time observations of the feature dimensions corresponding to the entity nodes at both ends of the proportional constant edge, where Var represents the variance and Mean represents the mean. Less than a preset drift detection threshold, for example, 0.05, and and The absolute amplitude shows a monotonic trend, which the system determines is a systematic drift caused by physical aging of the instrument, and calculates a global gain compensation factor to multiply and correct the sensed data stream; if If the drift threshold is exceeded, the system determines it as a change in the coal composition structure, maintains the global gain unchanged, and triggers the cognitive agent to adjust the search strategy. The system also includes an environmental adaptability adjustment unit to realize the dynamic correction of the graph logic. This unit obtains the environmental state vector at the current moment, uses a pre-trained mapping model to calculate the weight correction coefficients for different types of associated edges, and uses the coefficients to dynamically adjust the transfer weight values ​​of the associated edges in the coal quality attribute relationship graph. When the environmental state vector represents a high humidity condition, this unit automatically increases the transfer impedance of all associated edges that are physically coupled with moisture and reduces their transfer weights. The cognitive agent calculates the transfer probability between nodes based on the adjusted transfer weights, thereby automatically avoiding logical paths affected by environmental interference.

[0032] The strategy game and fusion output module is used to calculate the final analysis results. This module extracts the attribute values ​​of all activated nodes on the effective candidate semantic paths output by the cognitive agent, and uses the path confidence score to perform a weighted summation of the attribute values ​​to obtain the final coal physicochemical index values. The path confidence score is jointly calculated and determined by the weight product of all associated edges on the path and the feedback value of the logical reward function. If the logical consistency verification unit continuously triggers negative rewards for a specific working condition, the knowledge graph storage module will mark the invalid subgraph region under that working condition in the graph, and prohibit the agent from entering the region in subsequent inference. The establishment of constant ratio edges and the initial weight assignment in the knowledge graph follow the chemometric variance analysis procedure based on historical data. It retrieves historical offline test datasets covering the target coal type to calculate the molar ratio variance between pairs of silicon, aluminum, and iron rock-forming elements. Pairs of elements with variance values ​​lower than the chemometric convergence limit of 0.02 are judged as co-existing nodes constrained by crystal structure. The absolute value of the coefficient is set as the initial rigid weight of the proportional constant edge as a strong constraint benchmark for inference; the consistency threshold of the logical consistency verification unit is determined through an optimization process based on the subject working characteristic curve. A positive sample set containing expert-reviewed true causal paths and a negative sample set of pseudo-association paths generated by random feature masks are constructed. The structural similarity distribution of the two types of samples is calculated, and the similarity value corresponding to the maximum value of the Youden index is selected as the online operation judgment boundary to quantitatively distinguish between causal inference and pseudo-association matching; the weight correction coefficient of the environmental adaptability adjustment unit is preset using a gradient-sensitive environmental baseline calibration method. The full spectrum of standard coal samples is scanned in a standard temperature and humidity control box, and the nonlinear decay rate of the spectral feature vector magnitude is recorded as the environmental humidity increases at a gradient of 10%. The environmental state vector is fitted to the mapping function of the logical edge impedance coefficient. In real-time inference, the environmental sensor readings are directly converted into the suppression gain for moisture-related association edges; for drift compensation operations, the sensing agent only calculates the ratio fluctuation index in real time. Gain correction is initiated when the confidence interval is statistically determined under normal historical operating conditions. The confidence interval is the mean of the characteristic ratio of the elements at both ends of the constant ratio side plus or minus three standard deviations during the stable operating period. In engineering, this defines the numerical boundary between the physical aging of the instrument and the changes in the chemical composition of the coal.

[0033] Example 1: In an industrial application scenario for real-time online monitoring of coal quality entering a thermal power plant, the monitoring system is deployed at the coal conveyor belt site. This environment is characterized by high dust, high vibration, and drastic seasonal fluctuations in temperature and humidity. When the system faces continuous rainfall causing the relative humidity to consistently exceed [a certain value], [further challenges arise]. Furthermore, when the surface moisture content of coal increases, the data preprocessing module receives noisy spectral data from the spectrometer and environmental state data from the environmental sensor in parallel. The infrared baseline in the spectral data exhibits nonlinear drift, and the intensity of the OH bond characteristic peak corresponding to moisture tends to mask the characteristic peaks caused by the organic components of the coal. The data preprocessing module maps the environmental state data into an environmental embedding vector and injects it into the input feature tensor. Based on the high humidity environmental information explicitly contained in the input feature tensor, the sensing agent dynamically adjusts the feature mask vector output by the strategy network, reducing the gain weight of the feature channels corresponding to the moisture interference frequency band while amplifying the frequency band corresponding to the hydrocarbon skeleton vibration. Gain weights are used to generate target feature vectors that suppress environmental background noise. At the same time, the environmental adaptive adjustment unit identifies the current high humidity state and automatically calculates correction coefficients to increase the transfer impedance of all associated edges connected to moisture entity nodes in the coal quality attribute relationship graph. When the cognitive agent performs path search using the target feature vector as an index, it is constrained by the transfer impedance to avoid over-reliance on the logical path of deriving calorific value from moisture features. Instead, it searches for alternative semantic paths that point to low calorific value nodes through ash and fixed carbon nodes. This mechanism ensures that under the condition of abnormally enhanced moisture features, the reasoning process still follows the material conservation logic of coal combustion science, rather than simply fitting pseudo-correspondences in the data.

[0034] During long-term continuous operation of the system, due to light source attenuation or micro-contamination of the mirror surface, the signal amplitude acquired by the spectrometer experiences overall attenuation. The system monitors in real time the characteristic responses of silicon and aluminum nodes corresponding to the two ends of the constant ratio edge in the spectrum, and senses the ratio fluctuation index calculated by the intelligent agent. Maintaining a value below 0.05 indicates that the molar ratio of silicon to aluminum has not changed, while the characteristic amplitudes of both remain unchanged. and All signals showed a monotonically decreasing trend, indicating that the current state was a systematic drift caused by the physical aging of the instrument, rather than a change in coal type. The system then calculated a global gain compensation factor that could pull the current signal amplitude back to the baseline level and performed a multiplicative correction on the perceived data stream. The logical consistency verification unit performed causal verification on the candidate semantic paths generated by the cognitive agent. The system locked the carbon element feature dimension that played a dominant role in the path and performed a zero-masking operation on it at the input layer to generate a counterfactual mask feature tensor. The perturbation semantic path re-searched and generated by the cognitive agent based on this mask input had a structural similarity to the original path that was lower than a preset threshold, indicating that the original inference path had a strong causal dependence on the key carbon element features, rather than a random fitting based on environmental noise. The strategy game and fusion output module finally calculated the coal quality physicochemical indicators with deviations from the offline test results within the allowable range based on the confidence scores of each verified effective path.

[0035] Example 2: This example aims to verify the effectiveness and engineering feasibility of the coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion proposed in this invention in terms of nonlinear interference suppression, pseudo-association identification, and drift compensation. The experimental platform was built in the sampling room of a coal conveyor belt in a coal-fired power plant to simulate a real online detection environment. The data acquisition system includes an industrial-grade near-infrared spectrometer (spectral range 900). Up to 1700 Spectral resolution 5 In addition to an environmental sensing unit integrating temperature and humidity sensors and a dust concentration monitor, this embodiment collects 1200 sets of raw spectral data and their corresponding offline physicochemical index true values ​​(calorific value, moisture, ash content) for three commonly used coal types (bituminous coal A, lean coal B, and lignite C) to construct a gradient control test system. To verify the system's feature reweighting capability under strong noise conditions, a noise injection control group is designed, selecting 100 sets of reference spectral data as raw input and superimposing a signal-to-noise ratio (SNR) of 15. Gaussian white noise and 50 simulating belt motor vibration A sinusoidal interference signal was used to construct a contaminated test set. This test set was then input into a baseline model (comparison sample 1) without the weighting mechanism of the perceptual agent and a system with the perceptual agent of the present invention enabled (the present invention sample). The comparison results are shown in Table 1.

[0036] Table 1: Comparison of Signal-to-Noise Ratio and Key Band Energy Ratio Before and After Feature Weighting

[0037]

[0038] Data shows that the average signal-to-noise ratio of the target feature vector output by the present invention is improved to 28.5. More importantly, the sensing agent identifies and suppresses the OH bond frequency band dominated by ambient humidity (the energy proportion of which is determined by...). Down to At the same time, it enhances the CH bond frequency band, which characterizes the core properties of coal (energy proportion increased by...). Upgraded to This indicates that the perceptual agent can dynamically adjust the feature channel gain based on the environmental embedding vector, extracting effective features with high chemical fingerprint value from a noisy background. To verify the ability of the logical consistency verification unit to identify pseudo-associations, a counterfactual interference control group was constructed. Fifty high-calorific-value coal sample data were selected, and the corresponding carbon element feature dimension values ​​were artificially set to zero to generate a counterfactual mask input. These were fed into a traditional deep learning model without verification mechanism (comparison sample 2) and the system of this invention (the present invention sample group). In the absence of carbon features, comparison sample 2 still outputs high-calorific-value prediction results based on other secondary features, such as auxiliary elements that are weakly correlated with calorific value, showing obvious pseudo-association dependence. However, the logical consistency verification unit of the present invention sample group detected that the structural similarity between the perturbation semantic path generated by the cognitive agent and the original path dropped to 0.25 (below the preset threshold of 0.4), and then determined that the original inference path was invalid and triggered a negative reward, causing the system to correct its inference logic in subsequent iterations. Finally, the present invention sample group outputs results that cannot be determined or have low confidence.

[0039] In addition, for the drift compensation verification of long-term instrument operation, an experimental design was made to monitor the control group for a long period of time. During 30 days of continuous operation, the intensity of the monitored light source naturally decreased by approximately The process involves the system of this invention calculating the ratio fluctuation index in real time. The value remained consistently between 0.02 and 0.04 (below the judgment threshold of 0.05), accurately identifying the change as a systematic drift rather than a change in coal type. The system's automatically generated global gain compensation factor showed a monotonically increasing trend over time, effectively offsetting the signal amplitude decrease caused by light source attenuation. The test results on the 30th day showed that the calorific value prediction error (RMSE) after drift compensation was 0.45. This is better than the 1.82 in the uncompensated group. .

[0040] Example 3: This example combines Figures 1 to 3 This document describes a coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion. Figure 1As shown, the knowledge graph storage module is responsible for storing the coal quality attribute relationship graph containing entity nodes and positive and negative correlation edges, providing a basic graph topology structure for the cognitive agent. Simultaneously, the data preprocessing module acquires the detection and perception data stream and maps it into an input feature tensor. The perception agent receives this input feature tensor and calculates the feature channel gain weight vector, thereby generating a target feature vector as an index. The cognitive agent uses this target feature vector as an index to perform path search, generating candidate semantic paths connecting the input to the output and passing them to the logical consistency verification unit. This verification unit performs feature mask verification and counterfactual perturbation operations, eliminating false associations by comparing structural similarity. If the verification passes, it outputs the valid path and confidence score to the strategy game and fusion output module, calculates the candidate path confidence score, and performs a weighted sum to output the final analysis result. If the verification fails, it triggers a negative reward and instructs the cognitive agent to perform a re-search.

[0041] like Figure 2 As shown in the figure, the horizontal axis represents the structural similarity threshold, ranging from 0.1 to 0.9, and the vertical axis represents the ratio, ranging from 0 to 1.0. The chart plots three curves: the solid line represents the true positive rate (recall rate), which gradually decreases as the threshold increases; the dashed line represents the false positive rate, which decreases as the threshold increases; and the dotted line represents the precision rate, which initially increases and then slowly decreases. This chart visually reflects the quantitative impact of the structural similarity threshold selection on the system's ability to distinguish between causal and pseudo-associative features during counterfactual mask verification. Figure 3 As shown, the industrial field data acquisition workstation located at the coal conveyor belt site integrates an industrial-grade near-infrared spectrometer for acquiring spectral detection data, a multi-dimensional environmental sensor group for monitoring parameters such as temperature and humidity, and a data acquisition gateway responsible for signal buffering and analog-to-digital conversion. The workstation is connected to a high-performance computing and analysis server located in the central control room via an industrial Ethernet TCP / IP protocol. As the operating environment for the coal detection data computing and analysis system, it deploys a data preprocessing module, an environmental adaptability adjustment unit, a multi-agent collaborative computing module, a logical consistency verification unit, drift compensation computing logic, and a strategy game and fusion output module. The server is further connected to a knowledge graph storage server located in the data center via a high-speed local area network for accessing the coal quality attribute relationship graph library and historical path and experience database.

[0042] Example 4: This example aims to provide targeted repair and in-depth explanation of key parameter calibration and technical black boxes in the knowledge graph construction and adaptive update process, ensuring the system's sufficiency of disclosure under complex working conditions. Regarding the construction logic and weight initialization of constant-proportion edges in the coal quality attribute relationship graph, this invention executes a priori calibration procedure based on chemometrics. The system acquires historical offline test datasets covering the target coal type, extracts the content data of rock-forming elements such as silicon (Si), aluminum (Al), and iron (Fe), and calculates the molar ratio variance between each element pair. If the variance of an element pair is lower than a preset chemometric threshold, such as 0.02, it is determined that there is a symbiotic relationship between the element pairs strongly constrained by the crystal structure. The system establishes a constant-proportion edge for this node pair in the graph and sets its initial weight to the absolute value of the Pearson correlation coefficient of the element pair in historical data, serving as a strong constraint benchmark for inference. In the environmental adaptability adjustment unit, the method for calculating correction coefficients by the environmental impedance mapper is discussed. The algorithm path in this embodiment adopts a nonlinear mapping mechanism based on fuzzy logic, and the system defines an environment state vector. The humidity dimension is used as the input variable, and three fuzzy sets are set: dry, moderate, and high humidity. Each set is based on a Gaussian membership function, and correction coefficients are defined for logical edges related to HO bonds. For the output variable, three fuzzy sets are set: maintain, slightly suppress, and strongly suppress. The system has a built-in fuzzy inference rule base. For example, if the humidity is high, the correction coefficient is strongly suppressive. During the inference process, the system calculates the membership degree of the current humidity value to each fuzzy set, performs fuzzy inference using the maximum-minimum synthesis method, and finally uses the centroid method to defuzzify and output the accurate correction coefficient. This coefficient is used to multiply and correct the original weights of all HO-bond related edges in the graph.

[0043] Regarding the determination of the sparse penalty term coefficient in the perceptual agent, this invention follows an adaptive adjustment strategy based on information gain rate. In the early stages of model training, to encourage the agent to explore more potential features, the penalty term coefficient is adjusted accordingly. Set to a smaller value, for example As training progresses, the system periodically calculates the information gain rate of the current feature subset for predicting physicochemical indicators. If the growth of the information gain rate tends to level off, for example, if the growth rate is lower than [a certain value] for five consecutive epochs, the system will [follow the previous calculation]. The system gradually increases The value increases each time, for example. This process enables the agent to eliminate redundant features until the dimension of the feature subset shrinks to within the preset sparsity target range. For the calculation logic of the outlier index in cold start scenarios, this system adopts a quantitative evaluation method based on Mahalanobis distance. The system maintains a benchmark covariance matrix composed of target feature vectors from historical normal operating conditions. and mean vector For a new input feature vector x, the system calculates its Mahalanobis distance to the baseline distribution. ,like Exceeding a critical threshold determined based on the chi-square distribution, such as confidence level. If the corresponding critical value is reached, the current sample is determined to be an outlier, triggering a heuristic search mode.

[0044] Example 5: This example provides offline calibration and on-site deployment pre-calibration procedures for the initial construction of the knowledge graph, parameter calibration of the environmental adaptability adjustment unit, and determination of the logical consistency verification threshold, in order to solve potential black box problems in engineering implementation. For the initialization of the coal quality attribute relationship graph, this invention executes a statistical construction procedure based on historical big data. The system collects at least 5000 sets of historical offline test data in the target application scenario, covering indicators such as total moisture, ash, volatile matter, fixed carbon, total sulfur, and calorific value. For any two indicator nodes A and B, the system calculates their Pearson correlation coefficient in the historical data. and mutual information value I(A;B), if If the relevance is greater than a preset threshold (e.g., 0.6) and I(A;B) is greater than an information content threshold (e.g., 0.5), then an edge connecting A and B is established in the graph, and the initial weights are set to... For strong constraints confirmed by chemical principles, such as the negative correlation between ash content and calorific value, an association edge is established regardless of the statistical value, and an initial weight of 0.9 with high confidence is applied.

[0045] Prior to on-site deployment, parameter calibration of the fuzzy inference rule base in the environmental adaptability adjustment unit was performed, and an environmental sensitivity test procedure was executed. The spectral acquisition equipment was placed in a temperature and humidity controlled environmental test chamber, and the temperature was set at [temperature range missing]. to Relative humidity at to Within a range, spectral data of standard coal samples are collected at each environmental gradient point according to the gradient variation. The Euclidean distance deviation between the spectral eigenvector and the standard value is calculated. Based on the rate of change of the deviation with humidity, the center value and width of the membership function of the fuzzy set "high humidity" are determined. If the humidity increases by a certain amount... This leads to an increase in deviation exceeding Then, the humidity range is defined as the strong membership interval of the high humidity fuzzy set, and the corresponding impedance correction coefficient is set accordingly. The strong suppression level is set to, for example, 0.1, to ensure effective blocking of logical paths under high humidity interference. Regarding the determination of the structural similarity threshold in the logical consistency verification unit, this invention adopts an optimized procedure based on ROC curves to construct a verification set containing positive samples (true causal paths) and negative samples (pseudo-association paths). Positive samples consist of standard inference paths generated by an expert knowledge base, while negative samples are generated by introducing random noise or forcing erroneous associations into positive samples. The system traverses the judgment thresholds in the range of 0 to 1 with a step size of 0.05, calculates the classification accuracy, recall, and false positive rate of the verification unit for positive and negative samples at each threshold, plots ROC curves, and selects the threshold corresponding to the maximum Youden exponent as the default judgment threshold for field deployment, for example, 0.4.

[0046] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion, characterized in that, The system includes: The knowledge graph storage module is used to store the coal quality attribute relationship graph. The graph includes entity nodes that characterize the physical and chemical properties of coal and the associated edges that define the positive promotion or negative inhibition relationships between entity nodes. The data preprocessing module is used to acquire the detection and sensing data stream and map it into an input feature tensor; The multi-agent collaborative computing module includes mutually coupled perceptual agents, cognitive agents, and a logical consistency verification unit; The perceptual agent is used to calculate the gain weight vector of the feature channels based on the input feature tensor, and then uses the gain weight vector to perform weighting on the perceptual data stream to generate the target feature vector. A cognitive agent is used to perform path search in the coal quality attribute relationship graph using the target feature vector as an index, and generate candidate semantic paths connecting the input node to the output node; The logical consistency verification unit performs feature mask verification before outputting the results: it locks the dominant feature dimension in the subset of input features corresponding to the candidate semantic path; it performs a masking operation on the input feature tensor to set the value of the dominant feature dimension to zero and generates a mask feature tensor; it calls the cognitive agent to perform a secondary path search based on the mask feature tensor and generate a perturbation semantic path; it calculates the structural similarity between the candidate semantic path and the perturbation semantic path; in response to the structural similarity being lower than a preset consistency threshold, it confirms the validity of the candidate semantic path and calculates the output analysis results; in response to the structural similarity being higher than the consistency threshold, it determines the candidate semantic path as a pseudo-association path and triggers a negative reward update for the cognitive agent.

2. The coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion as described in claim 1, characterized in that, The associated edges in the knowledge graph storage module include proportionally constant edges connecting two entity nodes with a chemical symbiotic relationship; the perceptual agent is also used to perform drift compensation operations: calculating the ratio fluctuation index of the feature dimensions of the entity nodes corresponding to the two ends of the proportionally constant edge in the perceptual data stream; in response to the ratio fluctuation index being in a preset stable range and the absolute amplitude of the feature dimension changing monotonically, it is determined to be a systematic drift and a global gain compensation factor is calculated; the perceptual data stream is multiplicatively corrected using the global gain compensation factor; in response to the ratio fluctuation index exceeding the stable range, it is determined to be a change in component structure and the cognitive agent is triggered to adjust the search strategy.

3. The coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion according to claim 1, characterized in that, The multi-agent collaborative computing module is also used to perform cold start acceleration operations; the perceptual agent is used to calculate the outlier index of the input feature tensor relative to historical experience data; in response to the outlier index being higher than a preset threshold, the cognitive agent is used to retrieve the reference semantic path with the highest morphological similarity to the target feature vector from the historical path database; the path skeleton of the reference semantic path is extracted to generate an initialized path search probability distribution matrix; and the path search probability distribution matrix is ​​used to constrain the node transition probability of the first round of path search.

4. The coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion as described in claim 1, characterized in that, The associated edges have variable transfer weight attributes; the system also includes an environmental adaptive adjustment unit; the environmental adaptive adjustment unit is used to obtain the environmental state vector at the current moment, calculate the weight correction coefficient for different types of associated edges using a pre-trained mapping model, and dynamically adjust the transfer weight value of the associated edges using the weight correction coefficient; the cognitive agent is used to calculate the transfer probability between nodes based on the adjusted transfer weight value, wherein when the environmental state vector represents a strong interference condition, the interference source corresponding to the strong interference condition is identified, and the transfer weight value of the associated edge physically related to the interference source is increased to reduce the selection probability of the path.

5. The coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion according to claim 1, characterized in that, The perceptual agent adopts a deep neural network structure based on policy gradient; the perceptual agent outputs a set of feature mask vectors as the action space, and the dimension of the feature mask vectors is consistent with the number of spectral channels of the perceptual data stream; the logical reward function includes a sparsity penalty term, which is used to constrain the number of non-zero elements in the feature mask vectors, so that the perceptual agent selects the smallest complete feature subset.

6. The coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion according to claim 1, characterized in that, When calculating structural similarity, the logical consistency verification unit performs the following operations: extracting the first node sequence contained in the candidate semantic path and the second node sequence contained in the perturbed semantic path; calculating the length of the longest common subsequence of the first node sequence and the second node sequence; determining the structural similarity based on the ratio of the length of the longest common subsequence to the total path length; if the ratio is greater than the preset judgment threshold, it is determined that the reasoning logic is not blocked due to feature loss, and the original reasoning is determined to be a pseudo-association that depends on non-key features.

7. The coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion according to claim 2, characterized in that, Ratio fluctuation index calculated during drift compensation operation Satisfy the following definition: ,in, and These are the real-time observations of the feature dimensions corresponding to the entity nodes at both ends of the proportionally constant edge. Represents variance. Represents the mean; the perceptual agent only... When the drift is less than the preset drift threshold, the global gain compensation factor is calculated.

8. The coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion according to claim 1, characterized in that, The strategy game and fusion output module is used to perform path weighted calculation; the cognitive agent is used to output multiple candidate semantic paths and their corresponding path confidence scores; the strategy game and fusion output module is used to extract the attribute values ​​of the activated nodes on each candidate semantic path, and use the path confidence scores to perform weighted summation of the attribute values ​​to obtain the final analysis results; the path confidence score is jointly calculated based on the weight product of all associated edges on the path and the feedback value of the logical reward function.

9. The coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion according to claim 1, characterized in that, The knowledge graph storage module is also used to perform dynamic graph labeling; In response to the continuous triggering of negative reward updates for specific types of working conditions by the logical consistency verification unit, the system automatically marks the invalid subgraph region under that working condition in the coal quality attribute relationship graph; in subsequent reasoning, the cognitive agent is prohibited from wandering to the invalid subgraph region or the transition probability of entering the invalid subgraph region is reduced.

10. A coal detection data calculation and analysis system based on multi-agent reinforcement learning and knowledge graph fusion as described in claim 1, characterized in that, The data preprocessing module is used to perform multimodal feature stitching; The data preprocessing module acquires spectral detection data and environmental sensor data; it uses a first encoder to map the spectral detection data into spectral feature vectors and a second encoder to map the environmental sensor data into environmental embedding vectors; the environmental embedding vectors are expanded through a broadcast operation and then concatenated with the spectral feature vectors in the channel dimension to generate an input feature tensor, so that the perceptual agent can introduce the nonlinear influence of the environmental state on the spectral features during weighted calculation.

Images