Mining cable state evaluation method and system based on knowledge and deep Q network fusion

Abstract

The invention relates to the technical field of mining cable state evaluation, and provides a mining cable state evaluation method and system based on knowledge and deep Q network fusion. The method comprises the steps that a fuzzy multilevel analytic hierarchy process and an entropy weight method in a traditional empirical knowledge algorithm are adopted, and subjective weights and objective weights of cable indexes are obtained through calculation. And carrying out weight heterogeneous fusion through an improved D-S evidence theory, and carrying out fuzzy comprehensive evaluation on the cable. According to the embodiment, the graph neural network is utilized to construct the graph structure of the cable data, the historical data characteristics of the cable are deeply mined, a large amount of real and reliable cable data are constructed, and cable data enhancement is realized. According to the embodiment of the invention, a cable evaluation problem is modeled as a Markov decision process, under a graph neural network enhanced data set, a cable evaluation intelligent agent of a depth Q network based on uncertainty perception is trained, continuous trial and error and learning are carried out in an environment, an optimal evaluation strategy is finally grasped, and accurate evaluation of a cable operation state is realized.

Description

Technical Field

[0001] This invention relates to the field of condition assessment technology for mining cables, and in particular to a method and system for condition assessment of mining cables based on the fusion of knowledge and deep Q-networks. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Mining cables are the "arteries" for transmitting electrical energy and control signals in mines, and their safe operation is crucial to the stability of the mine's power system. Factors such as overvoltage, overcurrent, operating temperature, and ambient humidity can accelerate cable aging, causing them to deteriorate from "healthy" to "sub-healthy," making them prone to failure and even leading to power system paralysis over long-term operation. Therefore, condition assessment of mining cables and the implementation of anomaly early warning systems are a research hotspot in the electrical field.

[0004] Existing research includes methods such as traditional empirical knowledge assessment, analytic hierarchy process (AHP) combined with fuzzy comprehensive evaluation to assess urban cable risk, and fusion AHP combined with entropy weight method to evaluate cable quality. However, current research suffers from the following problems: First, traditional empirical formula methods rely excessively on expert subjective judgment and experience. The simple fusion of subjective weights in AHP and objective weights in entropy weight method makes it difficult to ensure the objectivity of assessments for mining cables with multiple parameters. When experts lack experience and face complex, variable, and multi-indicator new fault modes, the objectivity and reference value of cable assessments are limited. Second, in cable assessments, imbalanced datasets and a few problems can lead to a lack of generalization and robustness in the assessment of a few fault categories. Third, existing research on machine learning-based assessment methods relies on subjectively established indicator systems to learn and train qualitative data labels. This results in intelligent algorithms constantly approximating the results of established subjective evaluation systems, failing to truly break away from traditional knowledge-based assessment algorithms and thus failing to eliminate the subjectivity of the assessment. Summary of the Invention

[0005] To address the technical problems mentioned above, this invention provides a method and system for assessing the condition of mining cables based on the fusion of knowledge and deep Q-networks. The multi-level, multi-source assessment system designed in this invention, combined with comprehensive weight determination, eliminates reliance on subjective expert judgment and ensures the objectivity of the assessment through data-driven approaches. New samples are generated through a generator network to solve the problem of imbalanced data and enhance the model's generalization and robustness to various faults. This invention can significantly improve the quality and efficiency of mining cable assessment.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a method for assessing the condition of mining cables based on the fusion of knowledge and deep Q-networks.

[0007] A method for assessing the condition of mining cables based on the fusion of knowledge and deep Q-networks includes: Obtain data on mining cable indicators, establish a multi-level, multi-source data evaluation system, and determine the comprehensive weight of mining cable indicators; The data of mining cable indicators are divided into different evaluation levels, and different thresholds are set for each level. A membership matrix is ​​constructed using a piecewise linear membership function. Based on the comprehensive weight of mining cable indicators and the membership matrix, the comprehensive evaluation vector set of mining cable indicators is determined. Based on the comprehensive evaluation vector set of mining cable indicators, a mining cable graph structure is constructed; based on the mining cable graph structure, an encoder is used to obtain the mining cable depth feature set; based on the mining cable depth feature set, a generator network is used to generate new mining cable samples. Based on the comprehensive evaluation vector set of mining cable indicators and new mining cable samples, a mining cable evaluation dataset is constructed. The mining cable evaluation problem is modeled as a Markov decision process. Based on the mining cable evaluation dataset, a mining cable evaluation agent based on an uncertainty-aware deep Q-network is trained. After training and learning, the agent learns the optimal evaluation strategy and outputs the mining cable state evaluation results.

[0008] Furthermore, mining cable index data is obtained, a multi-level, multi-source data evaluation system is established, and the comprehensive weight of mining cable indexes is determined. The methods include: using fuzzy hierarchical analysis and introducing an expert evaluation matrix to determine the subjective weight of mining cable indexes; using entropy weight method and introducing a mining cable index data evaluation matrix to determine the objective weight of mining cable indexes; and integrating the subjective and objective weights of mining cable indexes to obtain the comprehensive weight of mining cable indexes.

[0009] Furthermore, the mining cable index data is divided into different evaluation levels, and different thresholds are set for each level. A membership matrix is ​​constructed using a piecewise linear membership function. Based on the comprehensive weights of the mining cable indexes and the membership matrix, the comprehensive evaluation vector set for the mining cable indexes is determined. The method includes: The performance data of mining cables are classified into four levels of evaluation: normal, caution, abnormal and serious, and then standardized. By setting thresholds for different levels and using piecewise linear membership functions, the standardized mining cable index data are calculated using piecewise linear membership functions to obtain membership vectors of the indexes belonging to different comments, forming a membership matrix. Calculate the secondary evaluation matrix based on the comprehensive weight and membership matrix of mining cable indicators; Based on the second-level evaluation matrix, determine the second-level fuzzy relation matrix; Based on the comprehensive weight of mining cable indicators and the second-level fuzzy relation matrix, calculate the comprehensive evaluation vector set of mining cable indicators.

[0010] Furthermore, based on the comprehensive evaluation vector set of mining cable indicators, a mining cable diagram structure is constructed; the method includes: The vectors in the comprehensive evaluation vector set of mining cable indicators are regarded as nodes in the mining cable diagram structure; Calculate the cosine similarity between all nodes to obtain the similarity matrix; use Nearest neighbors: Find the most similar nodes in the feature space for each node. For each neighbor, construct an adjacency matrix; based on the similarity matrix and the adjacency matrix, generate a weighted adjacency matrix; where the weight of the edge is the cosine similarity between nodes.

[0011] Furthermore, based on the mining cable graph structure, an encoder is used to obtain the mining cable depth feature set; the method includes: based on the mining cable graph structure, using a GCN encoder, aggregating the information of neighboring nodes layer by layer to update the node representation, and considering both the node's own information and neighboring features when the node state propagates at each layer, thereby realizing deep modeling of the cable feature distribution and obtaining the mining cable depth feature set.

[0012] Furthermore, based on the depth feature set of mining cables, a generator network is used to generate new samples of mining cables; the method includes: A vector is randomly selected from the depth feature set of mining cables. Generate a vector Gaussian noise vectors of the same dimension ; Generate category labels based on the target, and create a corresponding one-hot encoded vector. splicing vectors ; vector The input to the generator network, after passing through a fully connected layer and activation function transformation, generates new samples of mining cables. .

[0013] Furthermore, a mining cable evaluation agent based on an uncertainty-aware deep Q-network includes an input layer, a hidden layer, and an output layer, with an MC Dropout layer integrated after each hidden layer to maintain activation during both training and inference phases. During the process of continuous trial and error and learning in the environment, a learnable temperature parameter T is introduced, which transforms the Q-value output by the network into the action selection probability. :

[0014]

[0015] in, and These are the mean and standard deviation of the Q-value distribution obtained through MC Dropout sampling, respectively. It is a preset target temperature value.

[0016] Furthermore, during the training and learning process, a composite loss function is designed:

[0017]

[0018]

[0019]

[0020]

[0021]

[0022] in, L total Represents the composite loss function. The core issue is the Q-value loss. , and These are uncertainty loss, entropy regularization loss, and temperature regularization loss, respectively. , and These are the corresponding weight hyperparameters. It is the main network's view of the current state. Next action Q-value prediction, It is a time-difference objective. It is in state Next action The reward received later; It is a discount factor. It is an indicator that is 1 if the round is over, otherwise it is 0; In the next state The optimal action is selected based on the probability distribution predicted by the target network and calibrated for uncertainty. It is the Q-value evaluation of the target network for the next state-action pair. It is the standard deviation of the main network's prediction of the Q-value for the selected action. It is the average information value of the probability distribution obtained after the Q value output by the main network is transformed by the Softmax function; It is a target entropy value. It is the currently learnable temperature parameter in the network. It is a preset target temperature value.

[0023] A second aspect of the present invention provides a mining cable condition assessment system based on the fusion of knowledge and deep Q-networks.

[0024] A condition assessment system for mining cables based on the fusion of knowledge and deep Q-networks includes: The weight determination module is configured to: acquire mining cable index data, establish a multi-level, multi-source data evaluation system, and determine the comprehensive weight of mining cable indexes; The comprehensive evaluation module is configured to: divide the mining cable index data into different evaluation levels and set thresholds for different levels; construct a membership matrix using a piecewise linear membership function; and determine the comprehensive evaluation vector set of mining cable indexes based on the comprehensive weights of the mining cable indexes and the membership matrix. The graph neural network module is configured to: construct a graph structure for mining cables based on a comprehensive evaluation vector set of mining cable indicators; obtain a depth feature set for mining cables using an encoder based on the graph structure; and generate new samples of mining cables using a generator network based on the depth feature set. The decision evaluation module is configured to: construct a mining cable evaluation dataset based on the comprehensive evaluation vector set of mining cable indicators and new mining cable samples; model the mining cable evaluation problem as a Markov decision process; and train a mining cable evaluation agent based on an uncertainty-aware deep Q-network based on the mining cable evaluation dataset. After training, the agent learns the optimal evaluation strategy and outputs the mining cable status evaluation results.

[0025] A third aspect of the present invention provides a computer device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the mining cable condition assessment method based on knowledge and deep Q-network fusion as described in the first aspect above.

[0026] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a fully intelligent evaluation chain from indicator acquisition to result output, which has stronger scenario adaptability and evaluation efficiency compared with traditional methods. First, the multi-level, multi-source data evaluation system can comprehensively cover multiple dimensions of parameters such as electrical performance, operating environment, and aging status of mining cables, solving the problem of the one-sidedness of traditional single-indicator systems. Second, by constructing a membership matrix through piecewise linear membership functions, it achieves accurate quantification and unified representation of indicator data of different dimensions, providing a high-quality data foundation for subsequent intelligent learning. Finally, the evaluation problem is modeled as a Markov decision process, and the agent is trained using a deep Q-network with uncertainty awareness, enabling the model to dynamically perceive uncertainties in the evaluation process (such as environmental interference and indicator fluctuations) and achieve rapid response by learning the optimal evaluation strategy. This fully intelligent design not only reduces the cost of manual intervention but also makes the evaluation process adaptable to the multi-parameter and dynamic characteristics of mining scenarios, significantly improving evaluation efficiency and real-time performance compared to traditional methods.

[0027] This invention establishes a multi-level, multi-source data evaluation system and determines comprehensive weights, abandoning the weight setting mode that relies solely on expert experience. Instead, it forms a weight allocation logic based on the inherent correlation of multi-dimensional indicator data, reducing subjective intervention from the root of the evaluation system. On the other hand, the intelligent agent constructed through a deep Q-network uses the evaluation vector set and expanded samples as its learning basis, breaking away from the dependence on "subjective qualitative labeling of the indicator system." It no longer aims to approximate traditional subjective evaluation results, but instead mines evaluation patterns from the essence of the data. Even when expert experience is lacking or when facing new failure modes, it can output objective evaluation results with high reference value, effectively solving the problem of limited objectivity of traditional methods in complex scenarios.

[0028] This invention introduces a generator network to generate new samples based on the deep feature set of mining cables, effectively expanding the sample size for a few fault categories and scarce scenarios, and constructing a balanced mining cable evaluation dataset. This sample expansion method is not a simple data replication, but rather generates effective samples with real data distribution characteristics based on the deep features of cable indicators. This allows the intelligent evaluation model to fully learn the characteristic patterns of various fault modes (especially categories with insufficient sample size in the past), significantly improving the model's adaptability to different fault scenarios. The trained evaluation agent exhibits significantly enhanced generalization performance when facing complex and ever-changing operating states of mining cables. Even when encountering rare faults or marginal indicator data, it maintains stable evaluation accuracy, solving the problem of insufficient robustness caused by limited samples in traditional methods. Attached Figure Description

[0029] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0030] Figure 1 This is a flowchart illustrating the mining cable condition assessment method based on the fusion of knowledge and deep Q-networks, as shown in an embodiment of the present invention. Figure 2 This is an architecture diagram of a mining cable condition assessment method based on the fusion of knowledge and deep Q-networks, as shown in an embodiment of the present invention. Figure 3 This is a diagram illustrating the three-layer progressive structure of the mining cable evaluation system as shown in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the generation of new data by a GNN according to an embodiment of the present invention; Figure 5 This is a flowchart of another embodiment of the mining cable condition assessment method based on the fusion of knowledge and deep Q-networks shown in this invention. Figure 6(a) is a graph showing the change in evaluation accuracy of the agent during the training process, as illustrated in an embodiment of the present invention. Figure 6(b) is a training curve diagram of the composite loss function designed in this invention, as shown in an embodiment of the invention; Figure 6(c) is an average estimation curve of the state-action value of the agent during the training process, as shown in an embodiment of the present invention. Figure 6(d) is a graph illustrating the change in average cumulative reward according to an embodiment of the present invention; Figure 7 This is a structural diagram of a mining cable condition assessment system based on the fusion of knowledge and deep Q-networks, as shown in an embodiment of the present invention. Figure 8 This is a structural diagram of a computer device shown in an embodiment of the present invention. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0033] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0034] To address the problems of traditional assessment methods, such as over-reliance on expert subjective experience, difficulty in handling multi-source data, and imbalanced monitoring data samples, this invention proposes a method and system for assessing the condition of mining cables based on the fusion of knowledge and deep Q-networks. The following describes the solution in detail through several embodiments.

[0035] Figure 1 This is a flowchart illustrating a mining cable condition assessment method based on the fusion of knowledge and deep Q-networks, as shown in an embodiment of the present invention; see also... Figure 1 The method includes: Obtain data on mining cable indicators, establish a multi-level, multi-source data evaluation system, and determine the comprehensive weight of mining cable indicators; The data of mining cable indicators are divided into different evaluation levels, and different thresholds are set for each level. A membership matrix is ​​constructed using a piecewise linear membership function. Based on the comprehensive weight of mining cable indicators and the membership matrix, the comprehensive evaluation vector set of mining cable indicators is determined. Based on the comprehensive evaluation vector set of mining cable indicators, a mining cable graph structure is constructed; based on the mining cable graph structure, an encoder is used to obtain the mining cable depth feature set; based on the mining cable depth feature set, a generator network is used to generate new mining cable samples. Based on the comprehensive evaluation vector set of mining cable indicators and new mining cable samples, a mining cable evaluation dataset is constructed. The mining cable evaluation problem is modeled as a Markov decision process. Based on the mining cable evaluation dataset, a mining cable evaluation agent based on an uncertainty-aware deep Q-network is trained. After training and learning, the agent learns the optimal evaluation strategy and outputs the mining cable state evaluation results.

[0036] This invention first constructs a fusion evaluation system based on traditional empirical knowledge. By improving the DS evidence theory, it fuses the weights of the fuzzy multi-level hierarchical analysis method and the entropy weight method in traditional knowledge evaluation to obtain the subjective comprehensive weights of various cable evaluation indicators. Furthermore, it labels the operational status of some real samples for reference in intelligent algorithm-assisted learning. Then, to address the problems of scarce fault samples and data imbalance in practical engineering, a graph neural network is used to learn the inherent distribution and topological structure of real sample data, constructing a large-scale, high-fidelity synthetic data sample. Finally, the cable evaluation problem is constructed as a Markov decision process, and an uncertainty-aware Deep Q-Network (U-DQN) model is proposed. This model is trained on a balanced dataset generated by the GNN, obtains the optimal evaluation strategy through interactive trial and error with the environment, and utilizes a composite loss function to collaboratively optimize value assessment and uncertainty quantification. Experimental results demonstrate the effectiveness and reliability of the proposed evaluation method for evaluating test samples, verifying its suitability for mining cable evaluation tasks.

[0037] Figure 2 This is an architecture diagram of a mining cable condition assessment method based on the fusion of knowledge and deep Q-networks, as illustrated in an embodiment of the present invention; see reference. Figure 2 The following is a detailed description of the solution in this embodiment: 1. Cable evaluation based on traditional experience-based knowledge assessment methods The operating status of mining cables is affected by a variety of factors. Therefore, how to integrate multiple parameters to comprehensively evaluate the overall operating status of mining cables is a fundamental research question. Firstly, based on multi-source data of 10kV mining cables obtained from actual monitoring in a mine, and employing subjective expert experience and objective data feature analysis methods, a traditional experience-based knowledge fusion evaluation method was proposed. A traditional experience-based knowledge fusion status evaluation system and weighted fusion model for 10kV mining cables were constructed, providing auxiliary learning basis for subsequent intelligent learning algorithms. 1.1 Subjective and Objective Empowerment Based on Improved DS Evidence Theory Based on the performance characteristics of 10kV mining cables, the maximum values ​​of the three-phase voltage and current, as well as the three-phase current imbalance, reflect the cable's insulation capacity and thermal stress. The three-phase voltage imbalance reflects the power quality of the mine's power grid. The cable's surface temperature, as a direct thermal characterization of its operating state, comprehensively reflects the cable's load current, environmental conditions, and heat dissipation. Zero-sequence current and high-frequency leakage current directly reflect the cable's fault state, and monitoring them can provide early warnings of cable faults. The cable's insulation resistance is a direct indicator of its preventative testing, directly representing its health status. The cable's inspection cycle reflects the level of maintenance and protection it receives. In mines, the concentration of flammable and explosive gases in the cable's operating environment is an environmental safety indicator. Excessive humidity in the surrounding environment can induce "water treeing," leading to a decrease in the cable's insulation performance.

[0038] Based on the analysis, a multi-level, multi-source data evaluation system was established, adopting a three-level progressive structure, such as... Figure 3 As shown, the overall health status of the cable is used as the target layer, and its main purpose is to assess the health status of the cable. Operating conditions, preventive tests and environmental factors are used as intermediate links. Various indicators of the cable (including: maximum three-phase voltage, minimum three-phase voltage, maximum three-phase current, three-phase voltage imbalance, three-phase current imbalance, zero-sequence current, cable surface temperature, high-frequency leakage current, insulation resistance, inspection cycle, humidity around the cable and concentration of flammable and explosive gases) are used as the indicator layer.

[0039] For the mining cable, the evaluation object in this embodiment, a factor set is established. ,include These are the 12 indicator parameters of the cable evaluation system index layer. Based on the theory of fuzzy hierarchical analysis, the 12 indicator factors are divided into three layers according to operating conditions, preventive testing, and environmental factors. The specific parameters are shown in Table 1.

[0040] Table 1. Evaluation System for 110kV Mining Cables

[0041] To balance interpretability and learnability, fuzzy multi-level hierarchical weight calculation and entropy weight method were employed to obtain subjective and objective weights, respectively. For the subjective weights, the concentration of factors in the mining cable index layer was compared using a triangular fuzzy numerical scale. The importance of each factor is used to form an expert evaluation matrix. Elements in the matrix , This indicates that the evaluation experts believe the indicators relative to indicators The importance of obtaining the expert evaluation matrix for the three criterion layers in the evaluation system: operating condition matrix Preventive Trial Matrix Environmental factor matrix .

[0042] (1) After obtaining the evaluation matrix, Buckley's geometric mean method is used to solve for the index weights, and the evaluation matrix is ​​then calculated. The fuzzy mean of each indicator is used to obtain the indicator weight value after defuzzification. .

[0043] (2) (3) (4) To verify the consistency of the expert evaluation matrix, a consistency check must be performed on it. If the obtained expert evaluation matrix meets the consistency verification requirements, then the fuzzy evaluation matrix needs to be appropriately modified.

[0044] (5) After verifying the consistency of the evaluation matrix, the FAHP weights of the indicator layer are obtained by fuzzy hierarchical analysis. and criterion layer weights And all satisfy .

[0045] Furthermore, this embodiment uses the entropy weight method to obtain the objective weight values ​​of the cable indicators. First, an evaluation matrix is ​​constructed using the dataset, with the cable data sample size being [number missing]. The evaluation indicators are The original data matrix is ​​obtained as follows: (6) The range method is used to eliminate the dimensional limitations of cable parameters and transform the original data matrix into a dimensionless index matrix.

[0046] (7) In the formula, For the first In the nth sample The normalized data of each evaluation indicator For the first In the nth sample The raw data for each evaluation indicator , These are the maximum and minimum values ​​in the original data.

[0047] Based on the principle of information entropy, calculate the evaluation index. In the The proportion of each sample is used to calculate the entropy value of the evaluation index using equation (8): (8) (9) Finally, the evaluation indicators were calculated. Weight values: (10) Given two sources of evidence, to balance expert experience and objective data characteristics, this embodiment uses an improved DS evidence theory to eliminate the subjective expert weights of FAHP and the objective weights of the entropy weight method, ultimately obtaining a comprehensive weight for the cable index. First, basic probability allocation functions (BPAs) are constructed using the weight vectors calculated by FAHP and the entropy weight method, respectively: (11) In the formula, and These are the cable index parameters obtained by fuzzy hierarchical analysis and entropy weight method, respectively.

[0048] According to Dempster's fusion rules, the two sources of evidence are fused: (12) In the formula, The conflict coefficient is calculated using the following formula: .

[0049] Therefore, the final comprehensive weighting of the cable indicators is: (13) In the formula, The number of parameters in the indicator layer is 12, and the overall weight satisfies the following conditions: .

[0050] 1.2 Fuzzy Comprehensive Evaluation of the Operating Status of Mining Cables This embodiment uses a small amount of 10kV cable data collected in practice for fuzzy evaluation to obtain its health label. First, the fuzzy evaluation method uses four evaluation criteria, represented as a set of evaluation criteria. These correspond to four health states: normal, attention, abnormal, and severe. The specific rules for classifying these four levels are as follows: Normal The rating range is within [90, 100]. The scoring range is [75, 89], the abnormal scoring range is [60, 74], and the severe scoring range is [0, 59].

[0051] Next, before performing fuzzy evaluation, the 12 cable indicators need to be standardized, and the values ​​of the 12 indicators are uniformly mapped to [0,1]. In order to ensure that the larger the value after normalization, the more normal the parameter indicator is, the parameter standardization in this embodiment is divided into three categories, including positive indicators, negative indicators, and moderate indicators. The calculation formula for positive indicators is Equation (14), and the indicators include the maximum value of three-phase voltage, the minimum value of three-phase voltage, and the magnitude of insulation resistance.

[0052] (14) The calculation of the reverse index is as shown in Equation (15), including three-phase voltage imbalance, maximum three-phase current, three-phase current imbalance, zero-sequence current, high-frequency leakage current, and inspection cycle.

[0053] (15) The calculation of the moderate index is as shown in Equation (16), which includes the cable surface temperature, the humidity around the cable, and the concentration of flammable and explosive gases.

[0054] (16) in, These are real-time parameter values ​​measured during the operation of mining cables; These are normalized values ​​for the parameters of mining cables. , These are the minimum and maximum values ​​of each state variable parameter during normal operation; The optimal operating critical value for the parameter.

[0055] Based on selected threshold levels, a piecewise linear membership function is used to process the cable index data. The membership vector of the index to the comment set is obtained by calculating the membership function. To form a membership matrix The parameters of the operating conditions are based on the fuzzy membership matrix formed by the hierarchical structure. Parameters of preventive testing Parameters of environmental factors The secondary evaluation matrix is ​​calculated using the indicator weights and membership matrices of the indicator layer. The calculation formula is as follows: (17) In the formula, Indicates the first The weight of each indicator in the indicator layer This represents a two-level evaluation matrix.

[0056] Second-level fuzzy relation matrix Depend on The composition and calculation formula are as follows: (18) The membership degree of the final target layer is calculated using the second-level evaluation matrix, and the final comprehensive evaluation vector is obtained based on fuzzy evaluation theory. .

[0057] (19) In the formula, These correspond to the weight values ​​of operating conditions, preventive testing, and environmental factors at the criterion level, respectively. The membership degree of the cable sample to the four levels of the evaluation set.

[0058] Based on the principle of maximum membership, the largest comprehensive evaluation vector is achieved. Determining the health status of the cable samples yields an evaluation result for the 10kV mining cable data samples. This result is then used as the initial training data for the graph neural network.

[0059] 2. Evaluation Model of Deep Q-Network Based on Uncertainty Perception To achieve accurate and intelligent assessment of cable operating status, this embodiment proposes a cable assessment model based on Deep Reinforcement Learning (DRL). The cable assessment task is abstracted into a sequential decision problem, and a Deep Q-Network (DQN) agent is designed to learn the optimal assessment strategy through interaction with the simulated environment.

[0060] 2.1 Cable Data Augmentation Based on Graph Neural Networks Raw cable monitoring data often suffers from imbalanced samples and insufficient total data volume, which severely affects the learning performance and generalization ability of intelligent algorithm models. To address this issue, this embodiment employs a data augmentation method based on Graph Neural Networks (GNNs) to construct a large-scale, class-balanced training dataset.

[0061] First, each 12-dimensional cable feature vector in the original dataset is viewed as a graph structure. One of the nodes, where Represents a set of nodes. For each cable, an operational sample record is created, with each record representing an observation of the cable's state during a specific time period or operating condition. To capture the inherent correlations between data samples, edges between nodes are constructed using the following method. : (1) Similarity measurement: Calculate the cosine similarity between all samples to obtain the similarity matrix. , For the sample The formula for calculating cosine similarity is as follows: (20) (2) Determining adjacency relationships: using K-Nearest Neighbors (KNN) identifies the most similar nodes in the feature space for each node. We have 10 neighbors, thus establishing an initial adjacency matrix. .

[0062] (3) Weighted graph construction: Combine similarity and adjacency relationship to generate a weighted adjacency matrix. , among which the side weight It's the cosine similarity between them. .

[0063] Based on the constructed graph structure, a two-layer Graph Convolutional Network (GCN) encoder was designed to aggregate and update node features. GCN updates node representations by aggregating information from neighboring nodes layer by layer. Node states consider both their own information and neighbor features during propagation at each layer, achieving deep modeling of cable feature distribution. The layer update rules are as follows: (twenty one) in, It is an adjacency matrix with self-loops added; yes The degree matrix is ​​used to train the parameter matrix; It is the first The node feature matrix of the layer, The original feature matrix ; It is the trainable weight matrix of this layer. It is a ReLU activation function. Using the encoder designed in this embodiment, each raw cable data sample is mapped to a low-dimensional latent representation. This representation not only encapsulates the features of the sample itself, but also integrates the domain structure information between different samples.

[0064] After obtaining high-quality latent representations of the data, a generator network is constructed. To augment the raw data and construct new samples of cable operating status, a conditional generation strategy is employed. The generated cable data flow is as follows: Figure 4 As shown. In this embodiment, a new cable sample is generated. The process includes the following steps: 1) The set of latent vectors generated from the GCN encoder Randomly select a vector ; 2) Generate a... Gaussian noise vectors of the same dimension This is used to increase the diversity of generated samples; 3) Generate category labels based on the target and create a corresponding one-hot encoded vector. ; Concatenate a unified input vector ; 4) Input Generator Network After passing through a fully connected layer and activation function transformation, a new 12-dimensional feature vector is output. .

[0065] The core idea of ​​this method is to use GNNs to learn the inherent distribution and topological structure of the original data, and then use this structure as a basis to generate new, high-fidelity synthetic data under the guidance of specified category conditions.

[0066] 2.2 Cable Evaluation Modeling Based on Markov Decision Process The core idea of ​​reinforcement learning is to learn the optimal policy through the interaction between the agent and the environment. To apply reinforcement learning, the cable evaluation problem is first constructed as a Markov Decision Process (MDP), whose core elements are defined as follows: (1) State Space State: The state is the basis for the agent's decision-making. In this embodiment, state Defined as a 12-dimensional real vector , representing in The vector represents the characteristic data of the cable at any given time. It reflects the key information of the cable's operating condition monitoring characteristics in this embodiment. To eliminate the influence of different dimensions between characteristics, all state data has been standardized before being input into the model.

[0067] (2) Action Space Actions are the operations that an agent can perform based on its current state. In this task, the action space is a discrete set containing four predefined cable health states. The agent at each time step Choose an action That is, the currently input cable characteristic data Provide a state assessment result.

[0068] (3) Reward Function The reward function evaluates the quality of an agent's performance of a specific action in a given state and serves as a direct signal guiding the agent's learning. To guide the agent in making accurate classifications, a tiered reward mechanism is designed as follows. Let... For the true label of the sample, reward The design is as follows: (twenty two) (twenty three) (twenty four) in, The calculation process is as follows: the probability distribution is obtained by calculating the Q value. Then calculate its normalized entropy. First, the action Q value is obtained. Then calculate the standard entropy. in .

[0069] (4) State Transition: In the MDP model for this task, each sample is considered an independent time step. When the agent changes the current state... After an action is completed, the environment will present the next sample to be evaluated as the new state. During training, the order in which samples appear is random.

[0070] 2.3 Deep Q-Network Structure Based on Uncertainty Awareness To achieve accurate assessment of cable health status and combine it with traditional empirical assessment methods, this embodiment proposes an Uncertainty-Aware Deep Q-Network (U-DQN) model. This method refactors the traditional multi-class classification problem into a decision-making task within a reinforcement learning framework. By designing a unique network structure, reward mechanism, and composite loss function, and using cable data augmented with GNN algorithms for training, the agent learns to assess the confidence level of its own predictions while simultaneously learning classification strategies.

[0071] (1) Network structure The Q-network designed in this embodiment is a fully connected neural network, and its specific structure is as follows: Input layer: Contains 12 neurons, corresponding to a 12-dimensional cable state feature vector. ; Hidden layers: Contains three hidden layers with 128, 64, and 32 neurons respectively. Each layer is followed by a correction unit (RuLu) as the activation function to introduce non-linearity; Output layer: The output of the network backbone is not directly connected to the final Q-value output layer, but instead branches into two parallel fully connected "heads" used to predict the mean and standard deviation of the Q-value distribution, respectively: ① Mean head: A linear layer that outputs a 4-dimensional vector. , representing the model's expected Q-value estimate for each action (cable state). ② Standard deviation head: Another independent linear layer, outputting a 4-dimensional vector. To ensure that the predicted standard deviation is always positive, an exponential function is applied to it to obtain the final standard deviation vector. .

[0072] By using this output layer approach, the network no longer outputs a single, deterministic Q-value, but instead outputs a Gaussian-distributed parameter for each action. ,in It directly quantifies the model's confidence in the Q-value.

[0073] To further capture the cognitive uncertainty of the model, namely the uncertainty of model parameters caused by limited training data, this embodiment integrates a Monte Carlo Dropout layer after each hidden layer. Unlike the standard practice of disabling Dropout during inference, this layer remains active during both training and inference phases. When performing model prediction, this embodiment uses the same cable data sample as input. conduct The random forward propagation, due to Dropout randomly "dropping" different neurons in each propagation, will produce... A slightly different Q-value vector This sample can be considered an approximation of the posterior distribution of the Q-value. The statistical properties of this sample set, especially its standard deviation, are... This provides a powerful, non-parametric measure of cognitive uncertainty in models.

[0074] (2) Adaptive probability calibration mechanism This embodiment introduces a learnable temperature parameter T and couples it with uncertainty dynamics. To transform the network output Q-value into a well-calibrated probability distribution that reflects the true confidence level, the final action selection probability... Calculated using the following formula: (25) (26) in, and These are the mean and standard deviation of the Q-value distribution obtained through MC Dropout sampling, respectively.

[0075] To enable the uncertainty-aware deep Q-network to collaboratively learn value assessment, uncertainty quantification, and policy regularization, this embodiment designs and implements a composite loss function. The agent is trained based on an experience replay mechanism, where a batch of experience data is sampled from the experience replay pool at each learning step. Then, the parameters of the main network are updated by minimizing the following composite loss function: (27) in, The core issue is the Q-value loss. , and These are uncertainty loss, entropy regularization loss, and temperature regularization loss, respectively. , and These are the corresponding weight hyperparameters.

[0076] a) Temporal Differential Q-value Loss

[0077] (28) in, It is the main network's view of the current state. Next action Q-value prediction, This is a temporal difference (TD) objective, whose calculation borrows from the concept of Double DQN to alleviate the problem of overestimation of the Q-value. The objective value is calculated as follows: (29) in, It is in state Next action The reward received later; It is a discount factor used to weigh the importance of future rewards; It is an indicator that is 1 if the round is over, otherwise it is 0; In the next state The optimal action is selected based on the probability distribution predicted by the target network and calibrated for uncertainty. It is the Q-value evaluation of the target network for the next state-action pair. By decoupling action selection and value evaluation in the target network, the stability of training is further enhanced.

[0078] b) Supervised uncertainty loss

[0079] Uncertainty loss is key to this model's "self-awareness," enabling it to accurately predict its own prediction errors.

[0080] (30) in, This is the standard deviation of the main network's Q-value predictions for the selected action. The target value is the absolute value of the TD error. .

[0081] c) Entropy regularization loss

[0082] To prevent the policy from prematurely converging to an overconfident probability distribution and to encourage more thorough exploration in the early stages of training, entropy regularization loss is introduced. It achieves this by penalizing the difference between the current policy's output entropy and a pre-defined target entropy. (31) in, It is the average information value of the probability distribution obtained after the Q value output by the main network is transformed by the Softmax function; It is a target entropy value, set to .

[0083] d) Temperature regularization loss

[0084] In U-DQN, temperature parameters This is a learnable parameter that directly affects the smoothness of the final decision probability. To prevent this parameter from increasing or decreasing indefinitely during training, which could lead to training instability, a simple regularization term is added to constrain it within a reasonable range: (32) in, It is the currently learnable temperature parameter in the network. It is a preset target temperature value. This loss term stabilizes the temperature parameter around a suitable benchmark value, while allowing it to be fine-tuned according to the agent's learning needs.

[0085] Figure 5 This is a flowchart of another embodiment of the mining cable condition assessment method based on the fusion of knowledge and deep Q-networks shown in this invention; see also... Figure 5 First, this embodiment employs the fuzzy multi-level hierarchical analysis method and entropy weight method from traditional empirical knowledge algorithms to calculate the subjective and objective weights of cable indicators, respectively. Then, it uses an improved DS evidence theory to perform heterogeneous weight fusion for fuzzy comprehensive evaluation of the cable. Under a data-driven intelligent algorithm, this embodiment utilizes a graph neural network to construct a graph structure for cable data, deeply mining historical cable data features and constructing a large amount of reliable cable data to enhance the cable data. In terms of evaluation decision-making, the cable evaluation problem in this embodiment is modeled as a Markov decision process. Using a graph neural network-enhanced dataset, a cable evaluation agent based on an uncertainty-aware deep Q-network is trained. Through continuous trial and error and learning in the environment, it ultimately masters the optimal evaluation strategy, achieving accurate evaluation of the cable's operating status.

[0086] To verify the reliability and effectiveness of the mining cable condition assessment method based on the fusion of knowledge and deep Q-network proposed in this embodiment for the assessment of 10kV mining cables, a series of experiments were set up and implemented in this embodiment. The experimental data used 10kV cable data generated by the GNN network. The software used was PyCharm, based on the PyTorch framework, and training was performed on a local computer equipped with an NVIDIA 4060Ti to accelerate the training process using CUDA. The main training parameter settings are shown in Table 2. These parameters were determined after debugging and optimization in the initial experiments to obtain the best training results.

[0087] Table 2 Training Parameters

[0088] After training on the local computer, the main performance curves of the model during the training process are recorded, as shown in Figure 6, which comprehensively demonstrates the performance of the agent during the training process. Starting from a random policy, the agent gradually masters the cable state assessment task through training.

[0089] Figure 6(a) illustrates the change in evaluation accuracy of the agent during training. In the early stages of training, the evaluation accuracy is low because the strategy is still being explored. As the agent receives reward and penalty signals through interaction with the environment, its strategy is continuously optimized, and the evaluation accuracy steadily improves, eventually converging and stabilizing at approximately 98%, indicating that the model's decision-making has achieved efficient evaluation capabilities.

[0090] Figure 6(b) shows the training curve of the compound loss function designed in this embodiment. In the initial stage of training, the loss value drops rapidly, reflecting a significant adjustment in the network parameter weights. After approximately 2000 training rounds, the loss curve transitions to a slow and steady decline. This indicates that the model has successfully captured the main features of the data and is moving towards more refined parameter tuning. The continuous decrease in the loss value is highly consistent with the continuous increase in accuracy shown in Figure 6(a), jointly verifying the effectiveness and stability of the model training process. Finally, the loss function converges to a very low stable value, proving that the network training has reached convergence.

[0091] Figure 6(d) shows that the average cumulative reward is one of the core metrics for evaluating the convergence of a reinforcement learning task. The curve illustrates the change in the average reward received by the agent in each round. In the early stages of training, the agent frequently executes actions that lead to misclassification, thus receiving a large number of penalties, resulting in a negative average reward. As training progresses, the agent gradually learns to distinguish and select actions that bring positive rewards, while avoiding actions that lead to penalties. Therefore, the average reward curve shows a clear monotonically increasing trend. In the later stages of training, the average reward stabilizes at a high positive level, indicating that the agent has successfully learned the optimal or near-optimal policy for this evaluation task, verifying that the model proposed in this embodiment has achieved the state evaluation objective based on the reward function.

[0092] Figure 6(c) depicts the agent's average estimate of state-action value during training. In the early stages of training, the agent's exploratory actions may occasionally yield high instantaneous rewards, leading to a temporary and immature overestimation of the value of specific actions, thus forming the first peak of the curve. As the scope of exploration expands, the agent gradually recognizes that many actions have low long-term value or may result in penalties, and its perception of the environment becomes more accurate. Therefore, the average Q-value undergoes a correction process, showing a downward trend and reaching a trough. In the later stages of training, the agent gradually grasps the inherent patterns of the task and can stably execute action sequences that bring high long-term rewards. At this point, the average Q-value steadily recovers and continues to grow. This trend is consistent with the growth trend of average reward in Figure 6(d), indicating that the agent's value estimation of its learned optimization strategy is converging, and its confidence is continuously improving.

[0093] Based on the above four indicators, the U-DQN model exhibits excellent convergence characteristics during training. The continuously increasing accuracy and average reward, the steadily decreasing loss function, and the Q-value evolution curve conforming to learning principles demonstrate that the model has successfully mastered the intelligent assessment task from multi-dimensional monitoring data to cable health status. The model is effective, well-trained, and possesses reliable capabilities for practical assessment.

[0094] To verify the model's evaluation capability, six sets of test data samples were collected as input to the U-DQN model, as shown in Table 3. The model's evaluation results for the test data, i.e., the probabilities of the four states of the cable, are shown in Table 4.

[0095] Table 3 Test Data Sample

[0096] In the test samples, the surface temperature of the cable in sample 1 exceeded the maximum surface temperature under operating conditions. If the cable continues to operate in a continuously high-temperature environment, it will accelerate the thermal aging of the cable. In sample 2, the ambient humidity and concentration of flammable and explosive gases were relatively high, but its operating parameters were all within the normal range. Samples 3 and 6 showed varying degrees of degradation in multiple indicators, including insulation resistance, three-phase current and voltage imbalance, and temperature. Meanwhile, the three-phase current imbalance and zero-sequence current in sample 4 were also at a high level. The insulation resistance of sample 5 was only 400 MΩ, and the high-frequency leakage current was also abnormal compared to other test samples.

[0097] Table 4. Prediction results of the U-DQN model

[0098] As can be seen from the model prediction probabilities in Table 4, the assessment of the operating status of the test samples is reasonable and accurate. Referring to the analysis above, Sample 1, due to its abnormal surface temperature leading to thermal aging, is classified as "abnormal" and timely maintenance and replacement are recommended. The assessment model "decouples" the environmental risks and cable operating parameters of Sample 2. Although environmental parameters fluctuate abnormally, the cable's operating parameters and insulation level remain normal, and the model classifies it as "normal." Samples 3 and 6 share common characteristics, both exhibiting deterioration in cable data indicators. The assessment model proposed in this embodiment shows high consistency in its evaluation results when assessing these indicators with varying degrees of deterioration, classifying these two similar fault modes as "caution." Sample 4 shows abnormal values ​​in both the cable's own condition and the surrounding environmental humidity. Its degree of deterioration falls between Sample 2, where only environmental factors are abnormal, and Samples 3 and 6, where both their own condition and environmental factors are abnormal. This sample is classified as "caution." Sample 5 has very obvious defects, with the highest high-frequency leakage current, which is highly likely to cause "water treeing." Simultaneously, its insulation impedance has degraded, and the model accurately classifies it as "severe," recommending immediate shutdown for maintenance.

[0099] This embodiment proposes a cable evaluation model that integrates empirical knowledge algorithms with reinforcement learning based on deep Q-learning. This model assesses the health status of 10kV mining cables. New cable data generated by a GNN using labeled cable data from traditional empirical evaluation is used as input for training to obtain the optimal cable evaluation strategy. After verification through practical examples, the following conclusions are drawn: (1) In the U-DQN model proposed in this embodiment, Q-learning reinforcement learning is combined with neural network. This embodiment introduces in detail the basic elements of the proposed method, including agent, environment, action, reward function, etc., so that the agent can interact with the environment and learn through the reward and punishment function designed in this embodiment. Experimental results show that the proposed deep reinforcement learning model can learn autonomously from the environment. By fine-tuning the parameters, the model can achieve the assessment of the health status of the cable. (2) Reinforcement learning-based models are often considered "black boxes". Future research can combine interpretable AI technology to analyze the specific reasons why the agent makes corresponding evaluation decisions, enhance the credibility and transparency of the evaluation model, and provide clearer guidance for subsequent fault diagnosis and maintenance.

[0100] The above combination Figure 1 The method for assessing the condition of mining cables based on the fusion of knowledge and deep Q-networks provided in the embodiments of the present invention has been described in detail. Next, the system for assessing the condition of mining cables based on the fusion of knowledge and deep Q-networks provided in the embodiments of the present invention will be described in conjunction with the accompanying drawings.

[0101] Figure 7This is a schematic diagram of the structure of a mining cable condition assessment system based on the fusion of knowledge and deep Q-networks, as shown in an embodiment of the present invention. (Refer to...) Figure 7 The system described in this invention includes: The weight determination module is configured to: acquire mining cable index data, establish a multi-level, multi-source data evaluation system, and determine the comprehensive weight of mining cable indexes; The comprehensive evaluation module is configured to: divide the mining cable index data into different evaluation levels and set thresholds for different levels; construct a membership matrix using a piecewise linear membership function; and determine the comprehensive evaluation vector set of mining cable indexes based on the comprehensive weights of the mining cable indexes and the membership matrix. The graph neural network module is configured to: construct a graph structure for mining cables based on a comprehensive evaluation vector set of mining cable indicators; obtain a depth feature set for mining cables using an encoder based on the graph structure; and generate new samples of mining cables using a generator network based on the depth feature set. The decision evaluation module is configured to: construct a mining cable evaluation dataset based on the comprehensive evaluation vector set of mining cable indicators and new mining cable samples; model the mining cable evaluation problem as a Markov decision process; and train a mining cable evaluation agent based on an uncertainty-aware deep Q-network based on the mining cable evaluation dataset. After training, the agent learns the optimal evaluation strategy and outputs the mining cable status evaluation results.

[0102] In some embodiments, the weight determination module is specifically configured to: use fuzzy hierarchical analysis to introduce an expert evaluation matrix to determine the subjective weights of the mining cable indicators; use entropy weight method to introduce a mining cable indicator data evaluation matrix to determine the objective weights of the mining cable indicators; and integrate the subjective and objective weights of the mining cable indicators to obtain the comprehensive weights of the mining cable indicators.

[0103] In some embodiments, the comprehensive evaluation module is specifically configured to: divide the mining cable indicator data into four evaluation levels: normal, attention, abnormal, and severe, and perform standardization processing; set thresholds for different levels, and use a piecewise linear membership function to calculate the membership vectors of the standardized mining cable indicator data belonging to different evaluation levels, forming a membership matrix; calculate a secondary evaluation matrix based on the comprehensive weight of the mining cable indicators and the membership matrix; determine a secondary fuzzy relation matrix based on the secondary evaluation matrix; and calculate a comprehensive evaluation vector set of the mining cable indicators based on the comprehensive weight of the mining cable indicators and the secondary fuzzy relation matrix.

[0104] In some embodiments, the graph neural network module is specifically configured to: treat the vectors in the comprehensive evaluation vector set of mining cable indicators as nodes of the mining cable graph structure; calculate the cosine similarity between all nodes to obtain a similarity matrix; and employ... Nearest neighbors: Find the most similar nodes in the feature space for each node. For each neighbor, an adjacency matrix is ​​constructed. Based on the similarity matrix and the adjacency matrix, a weighted adjacency matrix is ​​generated, where the edge weight is the cosine similarity between nodes. Based on the mining cable graph structure, a GCN encoder is used to aggregate neighbor node information layer by layer to update the node representation. The node state considers both its own information and neighbor features during propagation at each layer, achieving deep modeling of the cable feature distribution and obtaining a deep feature set for the mining cable. A vector is randomly extracted from the deep feature set of the mining cable. Generate a vector Gaussian noise vectors of the same dimension Generate category labels based on the target and create a corresponding one-hot encoded vector. splicing vectors ; to vector The input to the generator network, after passing through a fully connected layer and activation function transformation, generates new samples of mining cables. .

[0105] In some embodiments, a mining cable evaluation agent based on an uncertainty-aware deep Q-network includes an input layer, a hidden layer, and an output layer, with an MC Dropout layer integrated after each hidden layer to maintain activation during both training and inference phases. During the process of continuous trial and error and learning in the environment, a learnable temperature parameter T is introduced, which transforms the Q-value output by the network into the action selection probability. :

[0106]

[0107] in, and These are the mean and standard deviation of the Q-value distribution obtained through MC Dropout sampling, respectively. It is a preset target temperature value.

[0108] In some embodiments, a composite loss function is designed during the training and learning process:

[0109] in, L total Represents the composite loss function. The core issue is the Q-value loss. , and These are uncertainty loss, entropy regularization loss, and temperature regularization loss, respectively. , and These are the corresponding weight hyperparameters.

[0110] The mine cable condition assessment system based on knowledge and deep Q-network fusion according to embodiments of the present invention can correspond to the execution of the method described in the embodiments of the present invention, and the above and other operations and / or functions of each module of the mine cable condition assessment system based on knowledge and deep Q-network fusion are respectively for implementing Figure 1 For the sake of brevity, the corresponding processes of each method in the code will not be elaborated here.

[0111] See Figure 8 The diagram shows the structure of a computer device, which includes a processor, a communication interface, and a computer-readable storage medium. The processor, communication interface, and computer-readable storage medium are connected via a bus or other means. The communication interface is used to receive and send data. The computer-readable storage medium can be stored in the computer device's memory. The computer-readable storage medium stores computer programs, including program instructions, and the processor executes the program instructions stored in the computer-readable storage medium. The processor (or CPU, Central Processing Unit) is the computing and control core of the computer device, adapted to implement one or more instructions, specifically adapted to load and execute one or more instructions to achieve the corresponding steps in the embodiment of the mining cable condition assessment method based on knowledge and deep Q-network fusion.

[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the state of a mine cable based on the fusion of knowledge and deep Q network, characterized in that, The method comprises the following steps: Obtain the index data of the mining cable, establish a multi-level multi-source data evaluation system, and determine the comprehensive weight of the index of the mining cable; The index data of the mining cable is divided into different evaluation grades, and different thresholds of different grades are set, and a membership matrix is constructed by using a segmented linear membership function; Based on the comprehensive weight of the index of the mining cable and the membership matrix, the comprehensive evaluation vector set of the index of the mining cable is determined; Based on the comprehensive evaluation vector set of the index of the mining cable, the mining cable graph structure is constructed; Based on the mining cable graph structure, an encoder is used to obtain a deep feature set of the mining cable; Based on the deep feature set of the mining cable, a generator network is used to generate new samples of the mining cable; Based on the comprehensive evaluation vector set of the index of the mining cable and the new samples of the mining cable, a mining cable evaluation dataset is constructed; The mining cable evaluation problem is modeled as a Markov decision process, and a mining cable evaluation agent based on a deep Q network with uncertainty perception is trained based on the mining cable evaluation dataset. After training and learning, the optimal evaluation strategy is mastered, and the state evaluation result of the mining cable is output.

2. The method according to claim 1, wherein, Obtain the index data of the mining cable, establish a multi-level multi-source data evaluation system, and determine the comprehensive weight of the index of the mining cable; the method comprises: The subjective weight of the index of the mining cable is determined by using the fuzzy analytic hierarchy process and introducing an expert evaluation matrix. The objective weight of the index of the mining cable is determined by using the entropy weight method and introducing the index data evaluation matrix of the mining cable. The subjective weight and the objective weight of the index of the mining cable are fused to obtain the comprehensive weight of the index of the mining cable. 3.The method according to claim 1, characterized in that, The index data of the mining cable is divided into different evaluation grades, and different thresholds of different grades are set, and a membership matrix is constructed by using a segmented linear membership function; Based on the comprehensive weight of the index of the mining cable and the membership matrix, the comprehensive evaluation vector set of the index of the mining cable is determined; the method comprises: The index data of the mining cable is divided into four evaluation grades of normal, attention, abnormal and serious, and is standardized; Different thresholds of different grades are set, and the standardized index data of the mining cable is calculated by using a segmented linear membership function to obtain a membership vector of the index belonging to different evaluations, which constitutes a membership matrix; According to the comprehensive weight of the index of the mining cable and the membership matrix, a two-level evaluation matrix is calculated; According to the two-level evaluation matrix, a two-level fuzzy relationship matrix is determined; According to the comprehensive weight of the index of the mining cable and the two-level fuzzy relationship matrix, the comprehensive evaluation vector set of the index of the mining cable is calculated.

4. The method according to claim 1, wherein, Based on the comprehensive evaluation vector set of the index of the mining cable, the mining cable graph structure is constructed; the method comprises: The vectors in the comprehensive evaluation vector set of the index of the mining cable are regarded as nodes of the mining cable graph structure; Calculate the cosine similarity between all nodes to obtain a similarity matrix; adopt The nearest neighbors, find the most similar neighbors of each node in the feature space, construct an adjacency matrix; based on the similarity matrix and the adjacency matrix, generate a weighted adjacency matrix; wherein the weight of the edge is the cosine similarity between the nodes.

5. The method according to claim 1, wherein, Based on the mining cable graph structure, an encoder is used to obtain a deep feature set of the mining cable; the method comprises: based on the mining cable graph structure, a GCN encoder is used to aggregate the information of neighbor nodes layer by layer to update the node representation. The node state considers its own information and neighbor features when propagating in each layer, which realizes deep modeling of the feature distribution of the cable and obtains a deep feature set of the mining cable.

6. The method according to claim 1, wherein, Based on the deep feature set of the mining cable, a generator network is used to generate new samples of the mining cable; the method comprises: A vector is randomly selected from the depth feature set of mining cables. Generate a vector Gaussian noise vectors of the same dimension ; According to the target, a category label is generated, and a corresponding one-hot encoding vector is created , the concatenated vector ; The vector is input into the generator network, which is transformed through a fully connected layer and an activation function to generate a new sample of the mining cable .

7. The method according to claim 1, wherein, The mine cable evaluation agent based on the deep Q network of uncertainty perception comprises an input layer, a hidden layer and an output layer, and an MC Dropout layer is integrated after each hidden layer to keep the activation state during the training and inference stages. In the process of constantly trying and learning in the environment, a learnable temperature parameter T is introduced to convert the Q value output by the network into an action selection probability : wherein, and are the mean and standard deviation of the Q-value distribution obtained by MC Dropout sampling, respectively, is a preset target temperature value. 8.The method for evaluating the state of the mine cable based on the knowledge and deep Q network fusion according to claim 1, characterized in that, During the training learning process, a composite loss function is designed: in, L total Represents the composite loss function. The core issue is the Q-value loss. , and These are uncertainty loss, entropy regularization loss, and temperature regularization loss, respectively. , and These are the corresponding weight hyperparameters. It is the main network's view of the current state. Next action Q-value prediction, It is a time-difference objective. It is in state Next action The reward received later; It is a discount factor. It is an indicator that is 1 if the round is over, otherwise it is 0; In the next state The optimal action is selected based on the probability distribution predicted by the target network and calibrated for uncertainty. It is the Q-value evaluation of the target network for the next state-action pair. It is the standard deviation of the main network's prediction of the Q-value for the selected action. It is the average information value of the probability distribution obtained after the Q value output by the main network is transformed by the Softmax function; It is a target entropy value. It is the currently learnable temperature parameter in the network. It is a preset target temperature value.

9. The mine cable condition assessment system based on the fusion of knowledge and deep Q network, characterized in that, The method comprises the following steps: The weight determination module is configured to: acquire mine cable index data, establish a multi-level multi-source data evaluation system, and determine the comprehensive weight of the mine cable index; The comprehensive evaluation module is configured to: divide the mine cable index data into different evaluation grades, set threshold values for different grades, and construct a membership matrix through a segmented linear membership function; Based on the comprehensive weight of the mine cable index and the membership matrix, a set of comprehensive evaluation vectors of the mine cable index is determined; The graph neural network module is configured to: based on the set of comprehensive evaluation vectors of the mine cable index, construct a mine cable graph structure; Based on the mine cable graph structure, an encoder is used to obtain a set of deep features of the mine cable; Based on the set of deep features of the mine cable, a generator network is used to generate new samples of the mine cable; The decision evaluation module is configured to: based on the set of comprehensive evaluation vectors of the mine cable index and the new samples of the mine cable, construct a mine cable evaluation dataset; The mine cable evaluation problem is modeled as a Markov decision process, and a mine cable evaluation agent based on the deep Q network of uncertainty perception is trained based on the mine cable evaluation dataset. After training and learning, the optimal evaluation strategy is mastered, and the state evaluation result of the mine cable is output.

10. A computer device, characterized in that, a processor adapted to execute a computer program; a computer readable storage medium having stored therein a computer program, the computer program being executed by the processor to implement the steps of the method for mine cable state evaluation based on the fusion of knowledge and deep Q network according to any one of claims 1-8.

Images