An industrial equipment health prediction method based on reinforcement density clustering and hidden Markov model

Abstract

The application belongs to the technical field of industrial equipment fault diagnosis, and solves the problems of parameter solidification and limited multi-source data fusion capability of traditional methods. A health prediction method of industrial equipment based on reinforcement density clustering and hidden Markov model is proposed: multi-source sensing data is acquired, and a comprehensive feature vector is obtained by weighted fusion using an attention mechanism; an Actor-Critic reinforcement learning intelligent agent is constructed, and an optimized neighborhood radius and minimum sample number are output through a Markov decision process; density clustering is performed accordingly to generate three types of health state labels: normal, warning and fault; a hidden Markov model is constructed, and the future state of the equipment is predicted using a transition probability matrix. The application uses reinforcement learning dynamic parameter adjustment to improve the working condition adaptability, uses the attention mechanism to realize efficient fusion of cross-modal data, and combines the hidden Markov model to complete unsupervised state prediction and fault risk assessment.

Description

Technical Field

[0001] This invention belongs to the field of industrial equipment fault diagnosis technology, specifically relating to an industrial equipment health prediction method based on reinforced density clustering and hidden Markov models. Background Technology

[0002] Existing industrial equipment fault diagnosis methods mainly rely on traditional density clustering algorithms (such as DBSCAN) or supervised learning models, but they have the following significant drawbacks: Insufficient parameter fixation and dynamic adaptability: Traditional density clustering requires manual pre-setting of neighborhood radius and minimum sample size, and parameter selection is highly dependent on expert experience. For example, when equipment aging or environmental interference causes data distribution shifts, the clustering results under fixed parameters deteriorate significantly. Although the invention with publication number CN119026736A improves load prediction accuracy through neural networks, its model update relies on offline retraining and cannot respond to dynamic operating conditions in real time.

[0003] Limited multi-source data fusion capabilities: Industrial equipment sensor data has high-dimensional and multimodal characteristics (such as vibration spectrum, temperature time series, and current harmonics), and existing methods lack effective feature fusion mechanisms. The invention with publication number CN119182113A uses a CNN-LSTM hybrid model to extract spatiotemporal features, but its feature weights are fixed and it is not adapted to unlabeled clustering tasks, resulting in high computational complexity and difficulty in edge deployment.

[0004] Lack of predictive capabilities in unsupervised scenarios: Existing technologies are mostly limited to diagnosing the current state and lack the ability to quantitatively predict future health risks. Although the invention with publication number CN119514859A involves model updates, its predictive model relies on supervisory signals (such as classification labels) and cannot be directly applied to unlabeled scenarios. Traditional threshold alarm methods (such as vibration amplitude exceeding limits) have a high false alarm rate and cannot capture early, subtle anomalies.

[0005] Reward sparsity and inefficient policy optimization: Under unlabeled conditions, existing reinforcement learning methods struggle to design effective reward functions to guide parameter optimization. The invention published in CN114819425A employs adaptive training window selection, but its reward mechanism still relies on supervised loss functions, failing to address the reward sparsity problem in clustering tasks. Summary of the Invention

[0006] In order to solve at least one of the above-mentioned technical problems in the prior art, the present invention provides an industrial equipment health prediction method based on reinforced density clustering and hidden Markov model.

[0007] This invention employs the following technical solution: a method for predicting the health of industrial equipment based on reinforced density clustering and hidden Markov models, comprising the following steps: Acquire multi-source sensor data during the operation of industrial equipment, and use an attention mechanism to perform weighted fusion of the multi-source sensor data to obtain a comprehensive feature vector; A reinforcement learning agent based on the Actor-Critic framework is constructed. The comprehensive feature vector is input into the reinforcement learning agent, and the optimized neighborhood radius and minimum number of samples are output using the Markov decision process. Density clustering is performed based on the neighborhood radius and minimum number of samples to generate health status labels for industrial equipment. The health status labels include normal status, warning status and fault status. A hidden Markov model is constructed based on the density clustering results, and the future state of the industrial equipment is predicted by the transition probability matrix and the health status label of the industrial equipment.

[0008] Preferably, the multi-source sensor data is weighted and fused using an attention mechanism, including: Feature vectors are determined based on the multi-source sensor data, and a learnable weight matrix is ​​constructed based on the feature vectors. The multi-source sensor data is weighted and fused based on the learnable weight matrix and the preset bias term to obtain a comprehensive feature vector.

[0009] Preferably, before using the Markov decision process to output the optimized neighborhood radius and minimum sample size, the process further includes: Define the state-space feature vector of a Markov decision process, which includes the number of clusters, silhouette coefficient, noise point ratio, and intra-class distance variance. Define the action space of a Markov decision process; Define a reward function relating density clustering quality to historical failures.

[0010] Preferably, the reward function is: In the formula, Indicates the first dynamic weighting coefficient; Represents the profile coefficient; This represents the second dynamic weighting coefficient; This represents a measure of the current density clustering distribution. Distribution of historical normal state The degree of deviation.

[0011] Preferably, it further includes: The first dynamic weighting coefficient is adjusted according to the preset attenuation factor; The second dynamic weight coefficient is dynamically adjusted based on the adjustment results.

[0012] Preferably, the calculation formula for predicting the future state of industrial equipment using the transition probability matrix and the industrial equipment health status label is as follows: In the formula, Indicates the future The probability of the state of industrial equipment at any given time; Indicates the future The potential state at any given moment; Represents the transition probability matrix; This means that all observation data are known. At any time The probability of.

[0013] Preferably, it further includes: Set the warning signal trigger threshold according to the type of industrial equipment; The cumulative fault risk index is calculated based on the transition probability matrix and the health status label under the fault state. The judgment is made based on the cumulative fault risk index and the warning signal trigger threshold. When the cumulative fault risk index is greater than the warning signal trigger threshold, the warning signal is triggered.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This application utilizes a reinforcement learning agent to dynamically adjust density clustering parameters, replacing traditional manual trial-and-error parameter tuning and significantly reducing reliance on domain expert experience. Under dynamic conditions, the agent can autonomously optimize parameters based on real-time clustering quality feedback, overcoming the performance degradation problem caused by data distribution shifts in traditional fixed-parameter clustering methods.

[0015] This application introduces an attention mechanism to dynamically weight and fuse features from multi-source sensor data, addressing the rigidity of feature fusion in traditional methods. Through the synergistic effect of reinforcement learning and density clustering, cross-modal correlation features are effectively extracted, enhancing the ability to identify complex faults.

[0016] This application constructs a Hidden Markov Model based on optimized density clustering results, abstracting the operating state of industrial equipment into health state labels, and predicting the future state of industrial equipment through transition probability matrices. Compared with traditional black-box prediction models, this method provides an intuitive state evolution path, supporting maintenance personnel to quickly locate the root causes of potential faults.

[0017] This application utilizes Markov decision process modeling parameter optimization to enable a reinforcement learning agent to perceive changes in the operating status of industrial equipment and adjust clustering strategies in real time. Compared to static models, this method exhibits stronger adaptability in scenarios with continuous data input, reducing the risk of misjudgment due to changes in operating conditions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the process of an industrial equipment health prediction method based on reinforced density clustering and hidden Markov model provided by an embodiment of the present invention. Figure 2 This is a heat map of real-time monitoring of unit health status provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the health status transfer and prediction path of Unit 1 provided in an embodiment of the present invention; Figure 4 This is the fault risk index trend monitoring of Unit 1 provided in this embodiment of the invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be noted that the structures, proportions, sizes, etc., shown in the accompanying drawings of this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportional relationships, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should fall within the scope of the technical content disclosed in the present invention. It should be noted that in this specification, relational terms such as "first" and "second" are only used to distinguish one entity from several other entities, and do not necessarily require or imply any actual relationship or order between these entities.

[0022] This invention provides an embodiment: This invention is applied to the health status prediction of a 2.5MW wind turbine gearbox, and is used to solve problems such as load fluctuations caused by sudden wind speed changes, compound faults caused by salt spray environment (such as bearing wear and tooth surface pitting), and difficulty in adaptive clustering parameters under varying operating conditions.

[0023] like Figure 1As shown in the figure, this embodiment of the invention provides a flowchart of an industrial equipment health prediction method based on reinforced density clustering and hidden Markov models, including the following steps: S1: Acquire multi-source sensor data during the operation of industrial equipment, and use an attention mechanism to perform weighted fusion of the multi-source sensor data to obtain a comprehensive feature vector.

[0024] In this embodiment, current sensors are deployed at key circuit points in the wind turbine gearbox, triaxial vibration sensors are installed at the high-speed shaft end to collect radial and axial vibration signals, acoustic emission sensors are set at the planetary carrier to capture stress waves, and temperature sensors are arranged in the low-speed shaft bearing housing. Wavelet thresholding is used to denoise the raw multi-source sensor data to eliminate wind load turbulence interference, and a time-series feature set is constructed through Z-score normalization. Attention weights are dynamically calculated based on the real-time load rate; vibration feature weights are increased when wind speed is high, and acoustic emission feature weights are increased when wind speed is low. Finally, a comprehensive feature vector is generated through a weighted fusion formula, providing a unified input for subsequent analysis.

[0025] Optionally, the attention mechanism is used to perform weighted fusion of the multi-source sensor data, including: determining feature vectors based on the multi-source sensor data, and constructing a learnable weight matrix based on the feature vectors; and performing weighted fusion of the multi-source sensor data based on the learnable weight matrix and a preset bias term to obtain a comprehensive feature vector.

[0026] In this embodiment, the differentiated feature extraction strategies corresponding to different sensor data types are as follows: (1) Vibration signal Calculate the root mean square value in the time domain As a characteristic: in, No. A sampled value of vibration acceleration or velocity; This represents the total number of sampling points within a time window.

[0027] Calculate the dominant frequency energy in the frequency domain As a characteristic: in, The spectrum is obtained by Fourier transforming the vibration signal; and The upper and lower limits of the frequency band are preset based on the fault characteristic frequencies of the target equipment components (such as bearings and gears).

[0028] (2) Temperature signal Calculate the average temperature rise rate within the time window As a characteristic: In the formula, Indicates the current time Temperature value; express Temperature value from a given time unit ago; The length of the time window representing the computation rate.

[0029] Calculate the steady-state fluctuation variance As a characteristic: in, This is the temperature sampling sequence within the steady-state window; This is the average temperature within that window. This represents the number of sampling points within the window.

[0030] (3) Current signal Calculate harmonic distortion rate As a characteristic: in, This represents the effective value of the fundamental frequency of the current. For the first The effective value of the subharmonic; Indicates calculation The highest harmonic order considered is usually set according to industry standards or equipment characteristics.

[0031] Calculate the trend of power factor change As a characteristic: This feature is achieved by analyzing the power factor sequence within the most recent time window. Perform linear regression and use the slope of the regression line as a quantifiable value of the trend. The calculation formula is: in, This represents the number of data points within the time window.

[0032] In this embodiment, multi-source sensor data from vibration, temperature, and current sensors are collected during device operation. This multi-source sensor data is then standardized and noise-reduced to construct a multi-dimensional temporal feature set. ,in, For time The eigenvectors at that time. For the eigenvectors... For each feature value, a raw attention score is calculated using a learnable parameter matrix. for: In the formula, Indicates time The eigenvector of the first One feature; This represents a learnable weight vector; This indicates the preset bias term.

[0033] The softmax function is used to normalize the scores of all features to obtain the weight coefficients. Ensure that the sum of all weights is 1: Finally, the original feature vectors are multiplied by their corresponding weights and then summed to obtain the fused features: In the formula, Represents time Weighted aggregation of multi-source sensor data; The dimension of the feature vector represents the total number of features. Indicates time The eigenvector of the first One feature; It is the first The normalized attention weights corresponding to each feature.

[0034] In this embodiment, after aligning and standardizing the multi-source sensor data by timestamp, a sliding window (window length) is used. seconds, sliding step Generate time series feature matrix (seconds) , which serves as the joint input for density clustering and reinforcement learning in reinforcement learning agents.

[0035] S2: Construct a reinforcement learning agent based on the Actor-Critic framework, input the comprehensive feature vector into the reinforcement learning agent, and use the Markov decision process to output the optimized neighborhood radius and the minimum number of samples.

[0036] Optionally, before using the Markov decision process to output the optimized neighborhood radius and minimum number of samples, the method further includes: defining the state space feature vector of the Markov decision process, wherein the state space feature vector includes the number of clusters, the silhouette coefficient, the proportion of noise points, and the variance of intra-cluster distance; defining the action space of the Markov decision process; and setting a reward function for the correlation between density clustering quality and historical failures.

[0037] In this embodiment, the core elements of the Markov decision process are defined as follows: State-space feature vector: Extract a four-dimensional clustering quality feature vector, including the number of clusters (reflecting the complexity of data distribution), silhouette coefficient (quantifying intra-cluster density and inter-cluster separation), noise point ratio (proportion of abnormal data), and intra-cluster distance variance (intra-cluster consistency index); Action space: Design discrete action combinations, with the neighborhood radius adaptively adjusted stepwise according to the data standard deviation, and the minimum sample size varying with integer step sizes; Reward function: Integrate the silhouette coefficient and KL divergence penalty term, with dynamic weights increasing as the operating conditions deteriorate, ensuring that the optimization direction conforms to the health evolution law of the equipment.

[0038] In this embodiment, the adjustment process of the neighborhood radius and minimum number of samples in density clustering is modeled as a Markov decision process.

[0039] The state-space feature vector of the current density clustering result Specifically, it includes: Number of clusters : Reflects the complexity of the data distribution pattern; Profile coefficient The calculation method is as follows: ,(in, Indicates sample The average distance to other points within the same cluster. For the sample The minimum average distance to all points in every other cluster is used to measure the relationship between how closely a sample point is clustered with its own cluster and how well it is separated from other clusters. Noise point percentage : Indicates the proportion of abnormal data; Intra-class distance variance : Measures the consistency of sample distribution within the same cluster, calculated as the variance of the distances from all samples to their cluster center.

[0040] In this embodiment, it is defined as a parameter adjustment action: in, , representing the neighborhood radius Adjustment step size, , indicating dataset The degree of dispersion, The standard deviation of the dataset; This is called the minimum sample size adjustment step size.

[0041] In this embodiment, the reward function is: In the formula, Indicates the first dynamic weighting coefficient; Represents the profile coefficient; This represents the second dynamic weighting coefficient; This represents a measure of the current density clustering distribution. Distribution of historical normal state The degree of deviation. Among them, and The eigenvectors of the current window and the historical normal data window are approximated by Gaussian distributions, respectively.

[0042] Optionally, it further includes: adjusting the first dynamic weight coefficient according to a preset attenuation factor; and dynamically adjusting the second dynamic weight coefficient according to the adjustment result.

[0043] In this embodiment, the first dynamic weighting coefficient Second dynamic weight coefficient The adjustment strategy is as follows: This adjustment strategy implements a core idea: as time t progresses, the degree of trust in historical experience, i.e., the first dynamic weighting coefficient... The second dynamic weighting coefficient will exhibit exponential decay, while the degree of trust in the real-time clustering quality will decrease. This will increase accordingly, ensuring that the model initially relies on historical stable states and can adapt to new operating conditions more quickly in the later stages. In the formula, This represents the attenuation factor, used to balance the importance of real-time data with historical experience; This represents the first dynamic weighting coefficient at time t; This represents the second dynamic weighting coefficient at time t; This represents the first dynamic weighting coefficient at time t-1.

[0044] In this embodiment, a reinforcement learning agent based on the Actor-Critic framework is constructed. The Actor network generates continuously parameter-tuned actions and optimizes parameters through policy gradient update rules. The Critic network evaluates state-action value, updates network weights using temporal difference error, and employs a priority-based experience replay mechanism to accelerate convergence in sparse reward scenarios. When wind speed changes, the reinforcement learning agent automatically adjusts its neighborhood radius unit to accurately capture transient abnormal vibration clusters.

[0045] In this embodiment, the Deep Deterministic Policy Gradient (DDPG) algorithm is used to train and optimize reinforcement learning parameters, constructing a reinforcement learning agent within the Actor-Critic framework. Parameter optimization is achieved through the following sub-steps: (1) Collect and store experience The agent selects an action based on the current policy: (in, (To explore noise), the environment returns a reward after execution. and the next state The obtained environmental interaction experience tuples Store it in the experience replay pool D.

[0046] (2) Sampling from the experience playback pool According to a priority sampling strategy, a batch of experience is drawn from the experience replay pool D for training. The batch size is N, and the sampling probability is... for: in, For the first Instant reward for each sample; This is the reward threshold, used to determine whether the reward is sparse; For the first The proportion of noise points corresponding to the state of each sample; this strategy prioritizes sampling with sparse rewards. Or sudden abnormality The sample.

[0047] (3) Update the Critic network Critic Network Updates are performed by minimizing the time difference error (TD-error), and the loss function is: Among them, the target value Calculated by the target network: in, This is a discount factor, with a value range of (0,1), which measures the importance of future rewards; Describes the target Critic network, whose first parameter is ; Describes the target Actor network, whose second parameter is .

[0048] (4) Update the Actor network Actor Network The policy gradient is updated by maximizing the expected return evaluated by the Critic network, and is as follows: in, The gradient of the Critic network for the action; This represents the gradient of the Actor network with respect to its own parameters.

[0049] (5) Soft update target network To stabilize training, a soft update strategy is used to slowly track the main network parameters: In the formula, This represents the soft update coefficient, which is usually very small and used to control the update speed. Indicates the preset parameters of the main Q network; This represents the pre-approval parameters of the main policy network.

[0050] S3: Perform density clustering based on the neighborhood radius and minimum number of samples to generate industrial equipment health status labels, which include normal status, warning status and fault status.

[0051] In this embodiment, three health states are divided based on the optimized clustering parameters: Normal state: high-density core clusters, and the intra-class distance variance and the KL divergence of the historical normal pattern are both lower than the set threshold; Warning state: isolated abnormal feature points (boundary points) appear, or the intra-class distance variance exceeds the normal threshold but still does not reach the fault level; Fault state: low-density separated clusters (abnormal clusters) are formed or the proportion of noise points exceeds the limit.

[0052] In this embodiment, based on the operating mode and fault characteristics of the wind turbine gearbox, the overall health status of the equipment is divided into multiple sub-state modules (such as normal operating condition, slight wear, bearing imbalance, and motor overheating). Based on the sensor data distribution characteristics, each sub-module is further subdivided into several dynamic operating condition types (such as no-load operation, full-load operation, and start-stop transient). Key sub-state modules (such as gear meshing modules) are selected as the core optimization benchmark.

[0053] In this embodiment, density clustering is performed on the wind turbine gearbox data based on the optimized neighborhood radius and minimum sample number to generate three types of state labels: Normal state : Marked as a high-density core cluster; intra-class distance variance And it conforms to the historical normal pattern. divergence ; Alert status : Marked as a core cluster but does not meet all normal conditions (e.g. ), or there are isolated anomalous feature points (boundary points); Fault status : Marked as low-density separated clusters (abnormal clusters), or the proportion of noise points .

[0054] The threshold relationship satisfies threshold It is determined by statistical analysis of historical normal operation data. In practical applications, the 95th or 99th percentile of the historical normal operation data is used.

[0055] S4: Construct a hidden Markov model based on the density clustering results, and predict the future state of the industrial equipment through the transition probability matrix and the health status label of the industrial equipment.

[0056] In this embodiment, the input features of the Hidden Markov Model include: real-time features, including vibrational dominant frequency energy, temperature fluctuation variance, and current harmonic distortion rate; and historical statistical features, including the mean intra-class distance over the past hour. Moving average of noise points Operating condition labels include equipment load level (0-100%) and operating ambient temperature.

[0057] In this embodiment, a statistical state transition law of Hidden Markov Model (HMM) is constructed to quantify the state evolution law, and based on the calculated cumulative failure risk index... Make proactive early warning decisions.

[0058] In this embodiment, an Hidden Markov Model (HMM) is constructed to predict future health status: (1) Statistically count the frequency of historical state transitions and calculate the transition probability matrix. The calculation formula is: (2) Calculate the condition of all current observation sequences O using the forward-backward algorithm. Always in a state of readiness The probability of: in, For the observation sequence, These are HMM parameters.

[0059] (3) Based on the current state probability distribution and transition matrix, predict the most likely state at time t+1: in, Indicates the future The probability of the state of industrial equipment at any given time; Indicates the future The potential state at any given moment; Represents the transition probability matrix; This means that all observation data are known. At any given moment, in a state The probability of that time.

[0060] Optionally, it further includes: setting a warning signal trigger threshold according to the type of industrial equipment; calculating a cumulative fault risk index based on the transition probability matrix and the health status label under the fault state, and making a judgment based on the cumulative fault risk index and the warning signal trigger threshold; and triggering a warning signal when the cumulative fault risk index is greater than the warning signal trigger threshold.

[0061] In this embodiment, the future The formula for calculating the cumulative failure risk index within a step is: In the formula, It is the transition probability matrix Power, meaning The state distribution after the step; its elements Indicates from state After k steps, the system transitions to a fault state. The probability of. It is the current moment. In state The probability of.

[0062] when When a warning signal is triggered, among which, Dynamically set according to the type of industrial equipment.

[0063] In this embodiment, the visualization output of the health status prediction results includes: a real-time clustering distribution heatmap, showing the spatiotemporal distribution of normal, warning, and fault clusters, as shown in the example below. Figure 2 It displays the state distribution heatmap of units 1 through 5 from 10:00 to 15:00; and the state transition path diagram, showing the state distribution based on... Matrix Generates the Future The most likely transition sequence, for example, Figure 3 It displays the paths and probabilities of state transitions between the three states, and highlights the predicted state path of Unit 1 at 10:00; the fault risk index line chart is dynamically displayed. Real-time change curves, for example, Figure 4 The data shows the historical trend of the cumulative fault risk index of Unit 1 at 15:00. It can be seen that the fault risk index at this time has seriously exceeded the warning threshold and is in a high-risk state, so a warning can be issued.

[0064] This invention constructs a predictive maintenance decision-making closed loop, which uses Hidden Markov Models (HMMs) to deduce the health status transition paths and probabilities of equipment and calculates the cumulative failure risk index. ,when When the set threshold is exceeded, maintenance instructions are triggered in advance to reduce unplanned downtime losses, lower the total life cycle maintenance cost, and provide the wind power industry with a highly adaptable, highly accurate predictive, and low-latency intelligent operation and maintenance paradigm.

[0065] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the health of industrial equipment based on reinforced density clustering and hidden Markov models, characterized in that, Includes the following steps: Acquire multi-source sensor data during the operation of industrial equipment, and use an attention mechanism to perform weighted fusion of the multi-source sensor data to obtain a comprehensive feature vector; A reinforcement learning agent based on the Actor-Critic framework is constructed. The comprehensive feature vector is input into the reinforcement learning agent, and the optimized neighborhood radius and minimum number of samples are output using the Markov decision process. Density clustering is performed based on the neighborhood radius and minimum number of samples to generate health status labels for industrial equipment. The health status labels include normal status, warning status and fault status. A hidden Markov model is constructed based on the density clustering results, and the future state of the industrial equipment is predicted by the transition probability matrix and the health status label of the industrial equipment.

2. The industrial equipment health prediction method based on reinforced density clustering and hidden Markov model according to claim 1, characterized in that, The multi-source sensor data is weighted and fused using an attention mechanism, including: Feature vectors are determined based on the multi-source sensor data, and a learnable weight matrix is ​​constructed based on the feature vectors. The multi-source sensor data is weighted and fused based on the learnable weight matrix and the preset bias term to obtain a comprehensive feature vector.

3. The industrial equipment health prediction method based on reinforced density clustering and hidden Markov model according to claim 1, characterized in that, Before using Markov decision processes to output the optimized neighborhood radius and minimum sample size, the following steps are also included: Define the state-space feature vector of a Markov decision process, which includes the number of clusters, silhouette coefficient, noise point ratio, and intra-class distance variance. Define the action space of a Markov decision process; Define a reward function relating density clustering quality to historical failures.

4. The industrial equipment health prediction method based on reinforced density clustering and hidden Markov model according to claim 3, characterized in that, The reward function is: In the formula, Indicates the first dynamic weighting coefficient; Represents the profile coefficient; This represents the second dynamic weighting coefficient; This represents a measure of the current density clustering distribution. Distribution of historical normal state The degree of deviation.

5. The industrial equipment health prediction method based on reinforced density clustering and hidden Markov model according to claim 4, characterized in that, Also includes: The first dynamic weighting coefficient is adjusted according to the preset attenuation factor; The second dynamic weight coefficient is dynamically adjusted based on the adjustment results.

6. The industrial equipment health prediction method based on reinforced density clustering and hidden Markov model according to claim 1, characterized in that, The formula for predicting the future state of industrial equipment using the transition probability matrix and the industrial equipment health status label is as follows: In the formula, Indicates the future The probability of the state of industrial equipment at any given time; Indicates the future The potential state at any given moment; Represents the transition probability matrix; This means that all observation data are known. At any given moment, in a state The probability of that time.

7. The industrial equipment health prediction method based on reinforced density clustering and hidden Markov model according to claim 6, characterized in that, Also includes: Set the warning signal trigger threshold according to the type of industrial equipment; The cumulative fault risk index is calculated based on the transition probability matrix and the health status label under the fault state. The judgment is made based on the cumulative fault risk index and the warning signal trigger threshold. When the cumulative fault risk index is greater than the warning signal trigger threshold, the warning signal is triggered.

Images