A sewage treatment plant operation health condition digital evaluation and preventive maintenance scheduling method

Abstract

The present application relates to the technical field of intelligent operation and maintenance and decision optimization of industrial facilities, and in particular to a method for digital evaluation and preventive maintenance scheduling of the operation health status of a sewage treatment facility, wherein standardized high-dimensional feature sequences are generated through multi-source operation data collection and preprocessing, a process-equipment coupled directed graph and a time-varying causal adjacency matrix are constructed, and original time series data are converted into node space features that fuse physical topology and dynamic causal logic; a mechanism regularization term is introduced to extract health latent variables, and a dynamic health state transition matrix of each component and a full-plant-level health evolution trajectory are obtained through decoding; a deep policy network embedded in a differentiable process constraint manifold is constructed to generate an optimal maintenance resource allocation matrix and a multi-period scheduling window. The present application realizes precise evaluation of the health status of a sewage treatment facility and intelligent scheduling of preventive maintenance, effectively reduces the risk of equipment failure, ensures stable effluent water quality, and improves the level of operation and maintenance management and comprehensive benefits.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and decision optimization technology for industrial facilities, specifically a method for digital evaluation of the operational health status and preventive maintenance scheduling of wastewater treatment facilities. Background Technology

[0002] In recent years, with the development of the Industrial Internet of Things (IIoT) and big data technologies, some wastewater treatment plants have begun to deploy digital operation and maintenance systems. However, most existing systems remain at the stage of data collection and visualization, lacking the ability to deeply mine and analyze multi-source operational data. In terms of data preprocessing, multi-source sensor data suffers from problems such as differences in sampling frequency, transmission phase delay, and data loss. Existing methods struggle to achieve high-quality data alignment and reconstruction, leading to insufficient reliability of subsequent analysis results.

[0003] In terms of health status assessment, existing technologies mostly employ purely data-driven machine learning models, lacking the integration of knowledge about the mechanisms of wastewater treatment processes. These models are prone to physical mismatch noise, and the prediction results lack interpretability. Furthermore, most assessment methods focus only on the health status of individual equipment, ignoring the coupling relationship between process units and equipment. This makes it impossible to achieve a comprehensive health assessment from individual components to the entire plant system, and thus difficult to accurately assess the overall operational risk of the system.

[0004] In preventative maintenance scheduling, existing methods are mostly based on static rules or simple optimization algorithms, which are insufficient to handle complex scheduling problems with multiple objectives and constraints. Traditional reinforcement learning methods have limitations in handling discrete hard constraints, and the generated maintenance strategies may violate process safety requirements, such as causing the interruption of critical biological treatment processes or exceeding effluent quality standards. In addition, existing scheduling systems lack effective closed-loop iterative mechanisms, and the effects of maintenance execution cannot be fed back to the model for parameter updates. This makes it difficult for the model to adapt to dynamic changes in operating conditions such as fluctuations in influent water quality, equipment aging, and process adjustments, and the prediction accuracy and decision quality gradually decline over time.

[0005] In summary, existing wastewater treatment facility operation and maintenance technologies have significant shortcomings in areas such as data quality assurance, health status evaluation, maintenance scheduling optimization, and closed-loop iteration capabilities, failing to meet the demands of modern wastewater treatment plants for refined and intelligent operation and maintenance. Therefore, there is an urgent need to develop a digital method capable of high-quality processing of multi-source data, fusion of mechanism data, accurate evaluation of health status, and constraint-aware maintenance scheduling, thereby comprehensively improving the operational reliability and operation and maintenance management level of wastewater treatment facilities. Summary of the Invention

[0006] The purpose of this invention is to provide a method for digital evaluation of the operational health status of wastewater treatment facilities and for preventive maintenance scheduling, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for digitally evaluating the operational health status and scheduling preventative maintenance of wastewater treatment facilities includes the following steps:

[0009] S1: Real-time acquisition of multi-source operation data of wastewater treatment facilities, covering process parameters, equipment vibration and electrical characteristics, influent water quality fluctuations and environmental meteorological conditions; dynamic time warping and alignment and missing value reconstruction based on time-series attention mechanism are performed on the multi-source operation data to generate standardized high-dimensional feature sequences;

[0010] S2: Based on the process unit topology and material energy flow direction of the wastewater treatment plant, construct a process-equipment coupled directed graph structure; by calculating the conditional mutual information and Granger causality index of the node feature sequence, quantify the dynamic influence intensity between nodes, generate a time-varying causal adjacency matrix, and map the standardized high-dimensional feature sequence to the node space of the time-varying causal adjacency matrix.

[0011] S3: Input the mapped node space features into the graph structure variational autoencoder to extract the health latent variable sequence reflecting equipment degradation and process efficiency decline; introduce the mechanism regularization term containing Monod biochemical reaction kinetic constraints and hydraulic residence time balance equation to perform hidden Markov state decoding on the health latent variable sequence, and output the dynamic health state probability transition matrix of each facility component and the plant-wide health evolution trajectory.

[0012] S4: Using the dynamic health state probability transition matrix, the plant-wide health evolution trajectory, and the current available maintenance resource pool as the state space, a preventive maintenance Markov decision process is constructed; a deep policy network embedded with a differentiable process constraint manifold is deployed. The differentiable process constraint manifold maps the risk of effluent water quality exceeding standards, the uninterrupted constraints of key biochemical sections, and the mutual exclusion logic of equipment maintenance into a continuously differentiable penalty gradient field through the projection operator. This gradient field is then jointly optimized with the action value function of the deep policy network to generate the optimal maintenance resource allocation matrix and a multi-cycle scheduling time window.

[0013] S5: Issue and execute digital maintenance work orders based on the optimal maintenance resource allocation matrix; collect operational condition recovery data and water quality compliance rate after maintenance execution, and calculate the Wasserstein distance between the actual health status distribution and the predicted distribution of the dynamic health status probability transition matrix; when the Wasserstein distance exceeds the adaptive threshold, trigger online collaborative fine-tuning of the hidden Markov state decoder transition probability and the deep policy network weights, update the causal strength weights of the time-varying causal adjacency matrix, and form a closed-loop iteration of evaluation and scheduling.

[0014] As can be seen from the technical solution provided by the present invention above, the beneficial effects of the digital evaluation and preventive maintenance scheduling method for the operational health status of wastewater treatment facilities provided by the present invention are:

[0015] By using dynamic time warping alignment and missing value reconstruction based on a time-series attention mechanism, the problems of transmission phase delay and data loss caused by sampling frequency differences in multi-source sensor data are effectively solved. The generated standardized high-dimensional feature sequences have time synchronization and integrity, providing a reliable data foundation for subsequent analysis. The constructed process-equipment coupled directed graph structure and time-varying causal adjacency matrix break through the limitations of traditional static topology modeling, and can accurately quantify the dynamic influence intensity between nodes under different operating conditions. It transforms the original time-series data into node spatial features that integrate physical topological associations and dynamic causal logic, significantly improving the system's ability to represent complex processes.

[0016] By extracting low-dimensional health representations from high-dimensional features using a graph-structured variational autoencoder, computational complexity is effectively reduced. The introduction of Monod biochemical reaction kinetic constraints and hydraulic residence time balance equations as mechanism regularization terms removes physical mismatch noise in the pure data-driven model, giving health latent variables clear physical meaning. State decoding based on a hidden Markov model can accurately output the dynamic health state probability transition matrix of each facility component and the plant-wide health evolution trajectory, realizing a comprehensive and accurate evaluation from single-point equipment health to plant-wide system health, providing a scientific basis for preventive maintenance decisions.

[0017] The uninterrupted constraints of the critical biochemical section at the risk of exceeding effluent quality standards and the mutually exclusive logic of equipment maintenance are transformed into a continuously differentiable penalty gradient field. This field is then used for joint gradient optimization with the action value function of a deep policy network. This solves the problem that traditional reinforcement learning methods struggle to handle discrete hard constraints, ensuring that the generated maintenance strategy always meets process safety requirements. The multi-objective collaborative reward evaluation model can balance multiple optimization objectives, such as effluent quality safety, equipment life extension, maintenance cost control, and operational interruption risk. The generated optimal maintenance resource allocation matrix and multi-cycle scheduling time window achieve efficient utilization of limited maintenance resources and rational planning of maintenance tasks.

[0018] By quantifying the difference between the actual health status distribution and the theoretically predicted distribution, model operating condition drift can be detected in a timely manner; the triggered online collaborative fine-tuning mechanism can synchronously update the weights of the hidden Markov state decoder transition probability deep policy network and the causal strength weights of the time-varying causal adjacency matrix, enabling the system to continuously adapt to complex operating condition changes such as fluctuations in influent water quality, equipment aging, and process adjustments; as the number of closed-loop iterations increases, the model's prediction accuracy and decision quality are continuously improved, realizing continuous optimization of the operation and maintenance of wastewater treatment facilities;

[0019] Through precise health status assessment and proactive preventative maintenance scheduling, the probability of sudden equipment failures is effectively reduced, unplanned downtime is minimized, and stable compliance of effluent quality is ensured. Simultaneously, optimized allocation of maintenance resources reduces unnecessary maintenance costs, extends equipment lifespan, and improves the economic and environmental benefits of the wastewater treatment plant. Digitalized maintenance work order issuance and full-process monitoring achieve standardization and intelligent operation and maintenance management, reducing the workload of maintenance personnel and improving operational efficiency. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the steps of a method for digitally evaluating the operational health status and scheduling preventive maintenance of a wastewater treatment facility according to the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0023] like Figure 1 As shown, this embodiment of the invention provides a method for digital evaluation of the operational health status and preventive maintenance scheduling of wastewater treatment facilities, including the following steps:

[0024] S1: Real-time acquisition of multi-source operation data of wastewater treatment facilities, covering process parameters, equipment vibration and electrical characteristics, influent water quality fluctuations and environmental meteorological conditions; dynamic time warping and alignment and missing value reconstruction based on time-series attention mechanism are performed on the multi-source operation data to generate standardized high-dimensional feature sequences;

[0025] S2: Based on the process unit topology and material energy flow direction of the wastewater treatment plant, construct a process-equipment coupled directed graph structure; by calculating the conditional mutual information and Granger causality index of the node feature sequence, quantify the dynamic influence intensity between nodes, generate a time-varying causal adjacency matrix, and map the standardized high-dimensional feature sequence to the node space of the time-varying causal adjacency matrix.

[0026] S3: Input the mapped node space features into the graph structure variational autoencoder to extract the health latent variable sequence reflecting equipment degradation and process efficiency decline; introduce the mechanism regularization term containing Monod biochemical reaction kinetic constraints and hydraulic residence time balance equation to perform hidden Markov state decoding on the health latent variable sequence, and output the dynamic health state probability transition matrix of each facility component and the plant-wide health evolution trajectory.

[0027] S4: Using the dynamic health state probability transition matrix, the plant-wide health evolution trajectory, and the current available maintenance resource pool as the state space, a preventive maintenance Markov decision process is constructed; a deep policy network embedded with a differentiable process constraint manifold is deployed. The differentiable process constraint manifold maps the risk of effluent water quality exceeding standards, the uninterrupted constraints of key biochemical sections, and the mutual exclusion logic of equipment maintenance into a continuously differentiable penalty gradient field through the projection operator. This gradient field is then jointly optimized with the action value function of the deep policy network to generate the optimal maintenance resource allocation matrix and a multi-cycle scheduling time window.

[0028] S5: Issue and execute digital maintenance work orders based on the optimal maintenance resource allocation matrix; collect operational condition recovery data and water quality compliance rate after maintenance execution, and calculate the Wasserstein distance between the actual health status distribution and the predicted distribution of the dynamic health status probability transition matrix; when the Wasserstein distance exceeds the adaptive threshold, trigger online collaborative fine-tuning of the hidden Markov state decoder transition probability and the deep policy network weights, update the causal strength weights of the time-varying causal adjacency matrix, and form a closed-loop iteration of evaluation and scheduling.

[0029] In this embodiment, the core function of step S1 is to comprehensively acquire multi-dimensional status information during the operation of the wastewater treatment facility. Through data preprocessing and feature engineering, the time asynchrony, missing data, and dimensional differences of multi-source data are eliminated, generating a standardized high-dimensional feature sequence that meets the requirements of subsequent graph structure modeling. This provides a high-quality data foundation for process equipment coupling relationship analysis and health status assessment. The detailed steps are as follows:

[0030] Step S1-1: Acquisition of multi-source raw signal streams and filtering of abnormal segments:

[0031] Four types of raw signal streams are collected in parallel at a preset sampling frequency through a distributed sensor network and an industrial protocol gateway. The first type is process parameters, including dissolved oxygen concentration, mixed liquor suspended solids concentration, sludge settling ratio, reagent dosage, and reflux ratio in each biochemical treatment section. The second type is equipment vibration electrical characteristics, including three-phase current and voltage, vibration acceleration, temperature, and operating power of pump and fan motors. The third type is influent water quality fluctuation data, including influent chemical oxygen demand, biochemical oxygen demand, total nitrogen, total phosphorus, and pH value. The fourth type is environmental meteorological data, including ambient temperature, humidity, air pressure, and rainfall.

[0032] The acquired raw signal streams are uniformly calibrated with timestamps, using the system clock of the industrial control server as a reference to correct the clock deviation of each sensor node; transient interference removal is performed, and a sliding window statistical method is used to identify abnormal data points exceeding three times the standard deviation, filtering out abnormal segments caused by hardware electromagnetic interference and communication packet loss; the processed signals are classified and stored according to data type and acquisition node to generate a multi-source raw time series dataset.

[0033] Step S1-2: Dynamic time warping and alignment of multi-source data:

[0034] Using the dissolved oxygen concentration time series of the aerobic biological treatment tank in the multi-source raw time series dataset as the baseline reference sequence, denoted as... , length is The remaining source data stream sequences are respectively denoted as , length is ,in Number the data source; input the baseline reference sequence and each sequence to be aligned into the dynamic time warping alignment module;

[0035] Construct the cumulative distance cost matrix , dimension ; matrix elements Indicates the reference sequence number The nth point and the nth sequence to be aligned The cumulative distance between points; the formula for calculating the cumulative distance is: , in, For the reference sequence number The nth point and the nth sequence to be aligned Euclidean distance between points For the reference sequence number The value of each sampling point For the first The nth sequence to be aligned The values ​​of each sampling point;

[0036] From the bottom right element of the matrix The process begins by searching for the optimal nonlinear warping path in reverse, where each point on the path corresponds to a matching pair between the reference sequence and the sequence to be aligned. The sequence to be aligned is then mapped onto the time axis of the reference sequence using the optimal path, compensating for the sampling frequency differences and transmission phase delays between the multi-source data. After performing the above operations on all sequences to be aligned, a multi-source time-warped aligned sequence with strictly synchronized time axes is output.

[0037] Step S1-3: Reconstruction of missing values ​​based on temporal attention mechanism:

[0038] The multi-source time-normalized and aligned sequences are arranged according to a preset window length. With step size Perform sliding window segmentation to generate multiple consecutive data segments; calculate the temporal self-attention weight distribution for each data segment; let a certain data segment be... , dimension ,in For feature dimensions;

[0039] Temporal self-attention weight matrix The calculation formula is: ,in, To query the vector weight matrix, The key vector weight matrix, Let be the dimension of the key vector. This is a scaling factor used to prevent the dot product result from becoming too large;

[0040] Data breakpoints and missing intervals are located based on temporal self-attention weight distribution, and positions with weight values ​​below a preset threshold are identified as missing data regions. Local feature vectors of K effective windows before and after the missing interval are extracted, and these local feature vectors are aggregated in a context-aware manner according to attention weights. The aggregated feature vectors are used to fill the missing interval with temporal coherence, ensuring that the filled data maintains a consistent trend with the effective data before and after. After performing the above operations on all data segments, they are spliced ​​together to generate a complete and continuous temporal sequence.

[0041] Steps S1-4: Cross-modal joint normalization and high-dimensional feature sequence generation:

[0042] Cross-modal joint normalization is performed on complete continuous time series, and an adaptive scaling strategy based on extreme value truncation and dynamic quantile mapping is adopted. First, extreme value truncation is performed on each feature dimension, and values ​​below the 1st percentile and above the 99th percentile are truncated to the corresponding quantile values ​​to eliminate the influence of extreme outliers.

[0043] Then, dynamic quantile mapping standardization is performed, calculated using the following formula: ,in, These are the original eigenvalues. This is the 25th percentile value for this feature dimension. This is the 50th percentile value. This is the 75th percentile value;

[0044] The above operations eliminate numerical distribution offsets and scale differences between different physical dimensions; the standardized feature data are reorganized according to the process topology hierarchy and equipment physical boundaries using tensor dimensions, integrating all feature data of the same process unit or equipment into a single node feature vector; and all node feature vectors are concatenated according to the node order of the process-equipment coupled directed graph structure, generating a dimension of... The standardized high-dimensional feature sequence, in which For time step, For the number of nodes, The feature dimension for each node; this sequence serves as the graph node feature input data for step S2.

[0045] In this embodiment, the core function of step S2 is to construct a structured graph representation based on the physical topology and operational logic of the wastewater treatment plant. By quantifying the dynamic causal relationships between nodes to replace static topological connections, the original temporal characteristics are transformed into node spatial characteristics that integrate physical constraints and causal logic, providing structured input for subsequent extraction of latent health variables and state evolution modeling. The detailed steps are as follows:

[0046] Step S2-1: Construction of the directed graph structure for process-equipment coupling:

[0047] Analyze the process drawings and operating procedures of the wastewater treatment facility, extract the physical location and functional attributes of the sedimentation tank sludge treatment unit and auxiliary electromechanical equipment of each biological treatment section, and generate a set of process unit nodes and a set of equipment physical nodes.

[0048] Based on the main sewage flow direction, return branch direction, reagent dosing path, and power distribution network, the material transfer direction and energy supply direction between nodes are determined. The process unit node set is taken as the macro flow node, and the equipment physical node set is taken as the micro execution node. Directed connection edges with clear direction are established to form the initial topological skeleton.

[0049] Configure spatial hierarchy identifiers and running role labels for each node in the initial topology skeleton, bind transmission medium type identifiers and energy conversion attribute identifiers for each connecting edge, and preset edge weight storage units to carry the causal strength data for subsequent dynamic calculation;

[0050] Perform graph connectivity verification and topology redundancy pruning on the initial topology skeleton, remove invalid isolated branches and merge parallel redundant pipeline associations, and construct a process-equipment coupled directed graph structure with hierarchical progressive characteristics and complete material and energy conduction links;

[0051] Let the directed graph of process-equipment coupling be: , From the set of nodes With the set of directed edges Composition; among which, Includes all process unit nodes and equipment physical nodes. Includes directed edges connecting all nodes; the total number of nodes is The number of nodes is consistent with the number of nodes in the standardized high-dimensional feature sequence in step S1;

[0052] Step S2-2: Establishing the periodic slicing of node feature sequences and the baseline for causal computation:

[0053] Using the node connection relationships in the process-equipment coupled directed graph structure as the calculation basis, the standardized high-dimensional feature sequence generated in step S1 is processed according to a preset sliding time window length. With sliding step size Perform periodic slicing;

[0054] For each time slice Extract the feature evolution sequence of each node within this time segment, denoted as , dimension ;in This is the time slice index, with a value range of [value range missing]. to , The total number of time slices is calculated using the following formula: ,in, To standardize the total time step of the high-dimensional feature sequence, This is the floor function;

[0055] For all pairs of nodes in a graph structure that are directly connected by an edge, establish a causal relationship computation pair set; for any directed edge from a node... Pointing to node node pairs Extract nodes respectively With nodes The feature sequences within each time slice are used as input data for subsequent causal strength quantification.

[0056] Step S2-3: Calculation and fusion of conditional mutual information and Granger causality index:

[0057] For each time slice node pairs The conditional mutual information value is calculated to remove indirect correlation interference transmitted through intermediate nodes and to lock the direct information coupling strength between nodes; the formula for calculating conditional mutual information is: ,in, Time slice internal nodes With nodes In the set of intermediate nodes The mutual information value below, For nodes In time slices The set of feature values ​​within, For nodes In time slices The set of feature values ​​within, For nodes With nodes All intermediate nodes in the time slice The set of feature values ​​within, It is the probability density function;

[0058] A multi-step time-lag prediction model is constructed simultaneously to solve for the Granger causality index, evaluating the contribution of upstream node state fluctuations to the time-series prediction of downstream node characteristic evolution. The formula for calculating the Granger causality index is as follows: ,in, Time slice internal nodes For nodes Granger causality index, Let be the mean square error function. For nodes In time slices The predicted target sequence within, For nodes forward The historical characteristic sequence of steps For nodes forward The historical characteristic sequence of steps The maximum time lag order;

[0059] The conditional mutual information value and the Granger causality index are adaptively normalized so that their values ​​are mapped to the interval between 0 and 1; then, they are weighted and fused according to preset weight coefficients to generate a comprehensive dynamic influence intensity index. The calculation formula is: ,in, This is the normalized conditional mutual information value. The normalized Granger causality index. This is the weighting coefficient, with a value ranging from 0 to 1, and a default value of 0.5.

[0060] Step S2-4: Generation of time-varying causal adjacency matrix and spatial mapping of feature sequence nodes:

[0061] Based on each time slice Comprehensive dynamic influence intensity index obtained from internal calculation The preset edge weight storage unit in the process-equipment coupled directed graph structure is updated in a rolling manner; for node pairs that do not have a direct connection edge, their edge weight is set to 0.

[0062] Generate a time-varying causal adjacency matrix representing the dynamic evolution of node crossover strength under different operating conditions and time periods. , dimension ; matrix elements Representing time slices internal nodes For nodes The intensity of the comprehensive dynamic impact;

[0063] Based on the weight distribution of the time-varying causal adjacency matrix, graph topological space transformation and neighborhood feature aggregation operations are performed on the standardized high-dimensional feature sequence; for each time step ,node post-aggregation features The calculation formula is: ,in, For nodes The set of all adjacent nodes, For nodes At time step The original feature vector, The feature transformation matrix is ​​a learnable feature. For bias vectors, It is a non-linear activation function. For time step The corresponding time-varying causal adjacency matrix;

[0064] The original multidimensional time-series data is projected onto the node embedding space corresponding to the time-varying causal adjacency matrix, generating mapped node space features that integrate physical topological associations and dynamic causal logic; the dimension of the mapped node space features is... ,in, This is the transformed feature dimension; this feature serves as the input source for subsequent extraction of latent health variables.

[0065] In this embodiment, the core function of step S3 is to extract low-dimensional health characteristics that reflect equipment degradation and process efficiency decline from the node space features that integrate causal logic. By introducing mechanistic knowledge of the wastewater treatment process to constrain the learning space of the pure data-driven model, the physical interpretability and predictive accuracy of the health status assessment are improved. Finally, the dynamic health status probability transition matrix of each facility component and the plant-wide health evolution trajectory are output, providing a basis for subsequent preventive maintenance decisions. The detailed steps are as follows:

[0066] Step S3-1: Graph structure variational autoencoder encoding and initial health latent variable extraction:

[0067] The mapped node spatial features generated in step S2 are input into the encoding mapping layer of the graph structure variational autoencoder; the encoding mapping layer is composed of multiple layers of graph convolutional neural networks stacked together, which aggregates the spatiotemporal dependency representations of each node and its neighboring nodes through the graph topology message passing mechanism.

[0068] For the Graph convolutional layer, nodes Output characteristics The calculation formula is: ,in, This is the normalized time-varying causal adjacency matrix. For the first Layer nodes The output characteristics, For the first The learnable weight matrix of the layer, For the first The layer's bias vector, It is a non-linear activation function;

[0069] After multi-layer graph convolutional encoding, the output features are input into the mean network and variance network respectively to generate the mean vector of the latent space distribution. With log-variance vector Perform variational reparameterization to sample nodes from the distribution. Initial health latent variables The calculation formula is: ,in, Random noise sampled from a standard normal distribution is used to achieve backpropagation of the gradient;

[0070] Concatenate the initial health latent variables of all nodes in chronological order to generate an initial health latent variable sequence. , dimension ;in The dimension of the health latent variables is much smaller than the dimension of the original node space features. The initial health latent variable sequence can reflect the combined state of accumulated physical wear and tear on equipment and fluctuations in the biochemical efficiency of process units.

[0071] Step S3-2: Mechanism Regularization Constraints and Health Latent Variable Optimization:

[0072] The initial health latent variable sequence is input into the mechanism regularization constraint module, and the mechanism regularization term containing the Monod biochemical reaction kinetic constraint and the hydraulic residence time balance equation is loaded simultaneously; first, the Monod biochemical reaction kinetic constraint penalty term is calculated; the Monod equation describes the relationship between microbial growth rate and substrate concentration, and its expression is: ,in, The specific growth rate of microorganisms This represents the maximum specific growth rate of microorganisms. Substrate concentration, It is the half-saturation constant;

[0073] Based on the substrate degradation rate and microbial growth curve implied by the initial health latent variable sequence, the deviation from the standard Monod kinetic model is calculated; for each time step With each biochemical processing unit Deviation penalty The calculation formula is: ,in, Microbial specific growth rate obtained from health latent variables The substrate concentration obtained from the inversion is... and These are the preset kinetic parameters for the biochemical treatment unit;

[0074] Then, the constraint penalty term of the hydraulic residence time equilibrium equation is calculated; the hydraulic residence time equilibrium equation is: ,in, The volume of the reaction tank. The concentration of the substrate in the reaction tank. This refers to the inlet water flow rate. The concentration of substrate in the influent. For water flow rate, The concentration of substrate in the effluent. The substrate degradation rate;

[0075] Verify the mass conservation residuals between the volumetric loading distribution corresponding to the initial health drop variable sequence and the actual hydraulic residence time; for each time step With each reaction tank Quality conservation residual penalty The calculation formula is: ,in, The substrate concentration in the reaction tank obtained by inversion, For time step, and The actual measured influent flow rate and influent substrate concentration are... This refers to the actual measured outflow rate. The concentration of substrate in the effluent obtained by inversion, The substrate degradation rate obtained by inversion, The time step obtained by inversion from the initial health latent variable sequence Time The substrate concentration in each reaction tank For the first The volume of each reaction tank;

[0076] Summing the penalty terms of all time steps and all units generates the total mechanism regularization loss. : ,in, This represents the total number of biochemical treatment units and reaction tanks. To standardize the total time step of the high-dimensional feature sequence, and These are weighting coefficients used to balance the contributions of the two constraints;

[0077] The mechanism regularization loss is combined with the original reconstruction loss of the graph structure variational autoencoder to construct a joint loss function; the mechanism compliance gradient signal is used to perform backpropagation joint optimization on the graph structure variational autoencoder to remove the physical mismatch noise in the pure data-driven mode and output the mechanism-constrained health latent variable sequence.

[0078] Step S3-3: Hidden Markov State Decoding and Dynamic Health State Probability Transition Matrix Generation:

[0079] The sequence of mechanistic-constrained latent health variables is input into a Hidden Markov State Decoder; a set of discrete health levels is pre-defined and... Among them, the level Indicates optimal health status, level This indicates the worst health condition and that a malfunction is imminent.

[0080] Construct an observation probability emission mapping between the latent variable feature space and the discrete health level set; assume that the latent variable corresponding to each health level follows a multivariate Gaussian distribution, for each health level... Its observation probability density function is: ,in, and Health levels The mean vector and covariance matrix of the corresponding latent variable distributions, It follows a multivariate Gaussian distribution;

[0081] The forward-backward probabilistic smoothing algorithm is used to analyze the hidden state transition logic of adjacent time segments; forward probability Indicates at time At a healthy level And observed the front The probability of each latent variable is calculated using the following formula: ,in, To be based on health level Transfer to health level The state transition probability, For a moment Health potential variables For a moment At a healthy level And observed the front Forward probabilities of one health latent variable;

[0082] Backward probability Indicates at time At a healthy level And observed from arrive The probability of the latent variable is calculated using the following formula: ,in, To preset the total number of discrete health levels, To be based on health level Transfer to health level The state transition probability, For a moment Mechanistic constraints on latent health variables For a moment At a healthy level And observed from arrive The backward probabilities of all health latent variables;

[0083] Calculate time based on forward and backward probabilities. At a healthy level posterior probability : ;

[0084] The Baum-Welch algorithm is used to iteratively estimate the parameters of the Hidden Markov Model, including the initial state probability distribution, the state transition probability matrix, and the observation probability distribution parameters. After multiple iterations and convergence, a dynamic health state probability transition matrix characterizing the degradation rate and functional recovery potential of each facility component is generated. , dimension ; matrix elements Indicates at time From health level Transfer to health level The probability of;

[0085] Step S3-4: Plant-wide health evolution trajectory simulation:

[0086] Based on the dynamic health status probability transition matrix of each facility component, and relying on the hierarchical transmission topology of the process-equipment coupled directed graph structure, cross-component health status risk weighted fusion and plant-wide performance extrapolation are performed.

[0087] First, for each node Assigning health risk weights The weights are pre-set based on the importance of each node in the process flow; key process units and core equipment have higher weights, while auxiliary equipment and secondary pipelines have lower weights; the sum of the weights of all nodes is 1.

[0088] For each time step Calculate the overall health of the entire plant. : ,in, For a moment node At a healthy level The posterior probability;

[0089] The state transition probabilities of single-point components are cascaded and aggregated along the main network of material flow and energy flow to assess the transmission impact of the deterioration of the health status of upstream components on downstream components; the Monte Carlo simulation method is used to predict the distribution of the health status of each component and the overall plant-wide health status change at multiple future time steps based on the dynamic health status probability transition matrix.

[0090] The output represents the spatiotemporal evolution pattern of the overall operational health of the wastewater treatment plant, which includes historical health curves and future predicted health curves. The plant-wide health evolution trajectory serves as the state space construction benchmark for the preventive maintenance Markov decision process in step S4.

[0091] In this embodiment, the core function of step S4 is to transform the facility health status evolution information and maintenance resource constraints into a standardized decision problem. A multi-objective optimization maintenance scheduling strategy is solved using a deep reinforcement learning method embedding process constraints. Under the premise of ensuring effluent water quality safety and continuous process operation, the optimal allocation of maintenance resources and multi-cycle time-series planning are achieved, providing a decision-making basis for subsequent digital maintenance work order issuance. The detailed steps are as follows:

[0092] Step S4-1: Preventive maintenance Markov decision process state space construction:

[0093] The dynamic health status probability transition matrix output in step S3 is fused with the plant-wide health evolution trajectory through time-series slicing to extract a local health status snapshot for the current decision-making cycle. This local health status snapshot includes the posterior probability distribution of the health level of each facility component at the current moment, and its future... The probability distribution of health status prediction for each decision cycle and the changing trend of overall plant-wide health status;

[0094] The local health status snapshot is concatenated with the resource attributes in the current available maintenance resource pool using a multi-dimensional tensor. The available maintenance resource pool attributes include the number and skill level of maintenance personnel, the types and quantities of spare parts in stock, the availability of specialized tools and equipment, and the remaining maintenance budget. The construction dimension is... Multidimensional state representation vector ,in Index for the current decision-making cycle;

[0095] Multidimensional state representation vector As the state input for the preventive maintenance Markov decision process, it fully represents the real-time operating status and maintenance support capabilities of the system, providing comprehensive state information for subsequent strategy reasoning.

[0096] Step S4-2: Maintaining the action space definition and constructing the multi-objective collaborative reward evaluation model:

[0097] Based on the node hierarchy distribution of the process-equipment coupled directed graph structure and the physical boundaries of facility components, a discretized maintenance action space is defined; each maintenance action corresponds to an intervention operation performed on a specific facility component, including four basic types: preventive maintenance, component replacement, condition monitoring, and resource standby.

[0098] The available resources in the current maintenance resource pool are matched with the potential intervention behaviors of each component for feasibility analysis; based on the skill requirements, spare parts requirements, and tool requirements for component maintenance, infeasible action combinations are filtered out, and a candidate set of action combinations covering all legal maintenance operations is generated. Each element in the action combination candidate set This represents a complete maintenance scheme that includes multiple parallel or serial maintenance operations.

[0099] Construct a multi-objective collaborative reward evaluation model, with core optimization indicators including the safety margin for effluent quality compliance, the expected gain of remaining equipment service life, the resource consumption cost of a single maintenance, and the risk of process operation interruption; the total reward function... The calculation formula is: ,in, In order to provide a safety margin bonus for effluent water quality to meet standards, This is a bonus based on the expected gain for the remaining useful life of the equipment. The cost of resources consumed per maintenance session. Penalty for the risk of process interruption. , , , Let be the weight coefficients for each objective, satisfying... ;

[0100] A dynamic time discount factor is set based on the efficiency decay slope of the plant-wide health evolution trajectory. It is used to quantify the current value of future rewards; the dynamic time discount factor decreases as the health decay slope increases, which can prioritize short-term safety gains when the system's health deteriorates;

[0101] The dynamic health state probability transition matrix is ​​conditionally coupled with the action combination candidate set to generate the action conditional state transition kernel. It is used to simulate the transition probability distribution of the health status of each facility component after a specific maintenance strategy is implemented;

[0102] The multidimensional state representation vector, action combination candidate set, multi-objective collaborative reward evaluation model and action condition state transition kernel are structured and encapsulated to form a preventive maintenance Markov decision process containing a complete state space, action space, reward mechanism and state transition logic.

[0103] Step S4-3: Construction of differentiable process-constrained manifolds and generation of continuously differentiable penalized gradient fields:

[0104] Define the geometric topological boundary of the differentiable process-constrained manifold, and transform three types of core process constraints into continuously differentiable penalty functions; the first type is the effluent quality exceeding risk constraint; the effluent quality exceeding risk is transformed into a continuous cost mapping of water quality indicators deviating from the safety threshold; for the... The effluent water quality indicators and their penalties The calculation formula is: ,in, To perform maintenance actions The predicted first The effluent water quality index values, For the first Safety thresholds for various water quality indicators This is the penalty intensity coefficient;

[0105] The second category is the constraint that the critical biochemical segment cannot be interrupted; this constraint is converted into a time-series overlap penalty term between the maintenance operation time window and the continuous period of the biochemical reaction; for the first... A key biochemical segment, its penalty item The calculation formula is: ,in, To perform maintenance actions The next Each biochemical segment at time The value of the operation interruption indicator. For a moment Importance coefficient of biochemical reaction To maintain the length of the scheduling time window;

[0106] The third category is equipment maintenance mutual exclusion logic constraints; this abstracts the equipment maintenance mutual exclusion logic into a Boolean logic continuous relaxation expression for resource occupancy conflicts; for a set of devices sharing the same maintenance resource, the penalty term... The calculation formula is: ,in, To share the first A collection of devices that maintain resources. For equipment At any moment The maintenance execution instruction value ranges from 0 to 1;

[0107] The orthogonal projection operator is used to smoothly project the three types of discrete hard constraints—the risk of effluent quality exceeding standards, the uninterrupted constraint of key biochemical stages, and the mutual exclusion logic of equipment maintenance—to a high-dimensional parameter space, generating a continuously differentiable penalty gradient field representing the constraint violation sensitivity and directional derivative. Total constraint penalty The weighted sum of all constraint penalty terms: ,in, , , These are the weighting coefficients for each constraint penalty term;

[0108] Step S4-4: Deep Policy Network Training and Optimal Maintenance Scheduling Policy Generation:

[0109] A deep policy network is constructed, which includes a spatiotemporal feature encoding layer, a policy probability output layer, and an action value evaluation layer. The spatiotemporal feature encoding layer is composed of multiple layers of temporal convolutional neural networks and graph attention networks, and is used to extract spatiotemporal dependent features from multidimensional state representation vectors.

[0110] Preventively maintain the multidimensional state representation vector of the Markov decision process The input spatiotemporal feature encoding layer extracts the latent state representations and outputs them to the policy probability output layer and the action value evaluation layer; the policy probability output layer generates the initial maintenance action probability distribution. The action value assessment layer outputs the corresponding baseline action value score. ;

[0111] By introducing a continuously differentiable penalty gradient field into the action-value function optimization path of a deep policy network, a joint loss objective of policy reward and constraint penalty is constructed. : ,in, For policy gradient loss, For the value function loss, and These are the loss weighting coefficients;

[0112] The policy gradient and manifold constraint gradient are calculated synchronously through backpropagation mechanism. Gradient vector synthesis and optimization step size are adaptively adjusted between the upward direction of maximizing action value and the projection direction of the feasible region of the constraint manifold. This drives the parameters of the deep policy network to iterate towards the convergence frontier that takes into account both long-term maintenance benefits and process safety boundaries.

[0113] After parameter convergence, the deep policy network is subjected to multi-cycle action sequence sampling and deterministic policy decoding; the output continuous action probability distribution is discretized into specific maintenance execution instructions, and resource quota secondary verification and time window conflict resolution are performed based on process material flow direction and energy transmission link.

[0114] Generate an optimal maintenance resource allocation matrix and multi-cycle scheduling time windows covering all plant facility components; the dimension of the optimal maintenance resource allocation matrix is... ,in, For the number of facility components, To maintain the number of resource types, matrix elements represent the number of resources allocated to the corresponding components; the multi-cycle scheduling time window includes the start time, end time, and execution order of each maintenance task; and the digital output of the preventive maintenance scheduling strategy is completed, providing a basis for the issuance of maintenance work orders in step S5.

[0115] In this embodiment, the core function of step S5 is to transform the optimal maintenance scheduling strategy into executable digital maintenance tasks. By monitoring the maintenance execution process and its effects throughout the entire process, quantifying model prediction deviations and triggering adaptive parameter updates, a closed-loop iteration of health status evaluation and preventive maintenance scheduling is achieved, continuously improving the system's adaptability to complex operating conditions. The detailed steps are as follows:

[0116] Step S5-1: Digital maintenance work order generation and targeted distribution:

[0117] The optimal maintenance resource allocation matrix generated in step S4 and the multi-cycle scheduling time window are structured and parsed to extract the target component identifier, resource quota details, execution time window and pre-security verification rules for each maintenance task; the above information is encapsulated in a unified data format to generate a standardized digital maintenance work order data package.

[0118] Digital maintenance work order data packages are pushed to the corresponding mobile operation and maintenance terminals and field control units through industrial IoT edge gateways; the digital twin interactive view and standardized operation guidance procedures bound to the target components are retrieved simultaneously, and the three-dimensional structure of the components, historical maintenance records and key operation points are displayed in real time on the mobile operation and maintenance terminals to guide on-site personnel to perform preventive maintenance, component replacement or condition detection operations in sequence.

[0119] Step S5-2: Monitoring the maintenance execution process and collecting data on operational condition recovery:

[0120] During maintenance, IoT sensors deployed at key nodes collect data in real time on equipment disassembly and assembly status, lubrication volume, electrical parameter retest values, and short-term bypass switching records of process sections. All collected data are accompanied by precise time and space stamps to form a complete maintenance process trajectory data stream, which is then uploaded to the cloud data platform.

[0121] After maintenance is completed, a recovery command is issued to the process control system to gradually restore the normal operation of each process unit. Operating condition recovery data and effluent compliance rate are continuously collected within 24 hours after maintenance, including core indicators such as dissolved oxygen concentration, mixed liquor suspended solids concentration, effluent chemical oxygen demand, total nitrogen, and total phosphorus in each biological treatment section. Abnormal fluctuations are filtered and steady-state features are extracted from the collected operating condition recovery data to remove transient fluctuation data during system recovery, generating a post-maintenance performance benchmark dataset. The post-maintenance performance benchmark dataset and the maintenance process trajectory data stream are used together as feedback input sources for subsequent model validation and closed-loop optimization.

[0122] Step S5-3: Quantification of Health Status Distribution Deviation and Adaptive Threshold Calculation:

[0123] Based on the post-maintenance performance benchmark dataset, the actual operational health level percentage of each facility component is reconstructed through a health status inverse mapping rule. The actual health level percentages of each component are then arranged chronologically to construct an empirical distribution of the actual health status after maintenance. ;

[0124] Extract the predicted output of the dynamic health status probability transition matrix from step S3 within the corresponding maintenance cycle, calculate the theoretical health level probability distribution of each component after maintenance, and construct the theoretical predicted health status distribution. ;

[0125] An optimal transmission distance metric algorithm is employed to quantify the cumulative cost of the empirical distribution of actual health status versus the theoretically predicted distribution of health status during the probabilistic quality transfer process, generating a Wasserstein distance value that characterizes the degree of model prediction bias. The formula for calculating the Wasserstein distance is as follows: ,in, Let be the set of all joint distributions, and let their marginal distributions be respectively and , For mathematical expectation, The distance is Euclidean.

[0126] By combining the historical periodic distance fluctuation average and the seasonal disturbance coefficient of influent water quality, the adaptive deviation tolerance threshold is dynamically calculated; the formula for calculating the adaptive threshold is: ,in, This is the average Wasserstein distance over the past 30 maintenance cycles. The standard deviation of the Wasserstein distance over the past 30 maintenance cycles. This is the coefficient of variation, with a default value of 2. This represents the seasonal disturbance coefficient of the current influent water quality. This is the seasonal disturbance weighting coefficient, with a default value of 0.5;

[0127] Step S5-4: Online collaborative fine-tuning of the model and update of causal strength weights:

[0128] When the Wasserstein distance value exceeds the adaptive bias tolerance threshold, it is determined that the current health evolution model and scheduling strategy have experienced operational drift, triggering an online collaborative fine-tuning mechanism;

[0129] First, the residual gradients of the actual empirical distribution of health status and the theoretically predicted distribution of health status are used to perform a Bayesian prior update on the state transition probabilities of the Hidden Markov State Decoder; the updated state transition probabilities are calculated using the following formula: ,in, For the updated health level Transfer to health level The probability, For historical observations of health levels Transfer to health level Number of times, The health level observed after this maintenance Transfer to health level The number of increments;

[0130] Meanwhile, the actual water discharge compliance rate and resource consumption feedback are used as reinforcement learning reward signals to perform small-step gradient backtracking fine-tuning on the action value evaluation layer of the deep policy network. During the fine-tuning process, a learning rate decay strategy is adopted. The initial learning rate is set to 1 / 100 of the original training learning rate, and the learning rate is decayed to 0.9 of the original rate every 10 iterations to ensure that the fine-tuning process does not destroy the effective knowledge already learned by the model.

[0131] The fine-tuned hidden state transition logic is fed back to the causal quantization step S2. Based on the actual coupling changes in the evolution of node features before and after maintenance, the fusion weights of conditional mutual information and Granger causality index are recalibrated. The causal strength weights of the time-varying causal adjacency matrix are updated on a rolling basis so that the graph structure can accurately reflect the actual operating characteristics of the system after maintenance.

[0132] Step S5-5: Closed-loop iteration and parameter synchronization for the next decision cycle:

[0133] The updated time-varying causal adjacency matrix parameters, hidden Markov state decoder transition probability matrix, and deep policy network weight parameters are synchronously saved to the model parameter library; at the beginning of the next decision cycle, the latest model parameters are automatically loaded for state space construction and policy inference.

[0134] The system continuously monitors its operational status and maintenance effectiveness, repeating the complete process from steps S1 to S5 to achieve adaptive closed-loop iteration of digital health status assessment and preventive maintenance scheduling. As the number of iterations increases, the model's understanding of the operational patterns of wastewater treatment facilities deepens, and the accuracy of health status prediction and the quality of maintenance scheduling decisions are continuously improved.

[0135] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for digitally evaluating the operational health status and scheduling preventive maintenance of wastewater treatment facilities, characterized in that: Includes the following steps: S1: Real-time acquisition of multi-source operation data of wastewater treatment facilities, covering process parameters, equipment vibration and electrical characteristics, influent water quality fluctuations and environmental meteorological conditions; dynamic time warping and alignment and missing value reconstruction based on time-series attention mechanism are performed on the multi-source operation data to generate standardized high-dimensional feature sequences; S2: Based on the process unit topology and material energy flow direction of the wastewater treatment plant, construct a process-equipment coupled directed graph structure; by calculating the conditional mutual information and Granger causality index of the node feature sequence, quantify the dynamic influence intensity between nodes, generate a time-varying causal adjacency matrix, and map the standardized high-dimensional feature sequence to the node space of the time-varying causal adjacency matrix. S3: Input the mapped node space features into the graph structure variational autoencoder to extract the health latent variable sequence reflecting equipment degradation and process efficiency decline; introduce the mechanism regularization term containing Monod biochemical reaction kinetic constraints and hydraulic residence time balance equation to perform hidden Markov state decoding on the health latent variable sequence, and output the dynamic health state probability transition matrix of each facility component and the plant-wide health evolution trajectory. S4: Construct a preventive maintenance Markov decision process using the dynamic health state probability transition matrix, the plant-wide health evolution trajectory, and the currently available maintenance resource pool as the state space; Deploy a deep policy network embedded with a differentiable process constraint manifold. The differentiable process constraint manifold maps the risk of effluent water quality exceeding standards, the uninterrupted constraints of key biochemical stages, and the mutual exclusion logic of equipment maintenance into a continuously differentiable penalty gradient field through the projection operator. This field is then used in conjunction with the action value function of the deep policy network for gradient optimization to generate the optimal maintenance resource allocation matrix and multi-cycle scheduling time window. S5: Issue and execute digital maintenance work orders based on the optimal maintenance resource allocation matrix; collect operational condition recovery data and water quality compliance rate after maintenance execution, and calculate the Wasserstein distance between the actual health status distribution and the predicted distribution of the dynamic health status probability transition matrix; when the Wasserstein distance exceeds the adaptive threshold, trigger online collaborative fine-tuning of the hidden Markov state decoder transition probability and the deep policy network weights, update the causal strength weights of the time-varying causal adjacency matrix, and form a closed-loop iteration of evaluation and scheduling.

2. The method for digital evaluation of the operational health status and preventive maintenance scheduling of a wastewater treatment facility according to claim 1, characterized in that: Real-time collection of multi-source operational data of wastewater treatment facilities, which includes process parameters, equipment vibration and electrical characteristics, influent water quality fluctuations and environmental meteorological conditions; Dynamic time warping and alignment and missing value reconstruction based on temporal attention mechanism are performed on multi-source running data to generate standardized high-dimensional feature sequences; Specifically, it includes: Through a distributed sensor network and an industrial protocol gateway, raw signal streams containing process parameters, equipment vibration and electrical characteristics, influent water quality fluctuations and environmental meteorological conditions are collected in parallel at a preset sampling frequency; the raw signal streams are subjected to unified timestamp calibration and transient interference removal, and abnormal segments caused by hardware noise and communication packet loss are filtered out to generate multi-source raw time series datasets. Using the core process unit's operating cycle time sequence in the multi-source raw time series data as the benchmark reference sequence, the other source data stream sequences and the benchmark reference sequence are input into the dynamic time warping and alignment module; by constructing a cumulative distance cost matrix and searching for the optimal nonlinear warping path, the sampling frequency difference and transmission phase delay between the multi-source data are compensated, and a multi-source time warping and alignment sequence with strict time axis synchronization is output. The multi-source time-normalized and aligned sequence is divided into sliding windows with a preset step size. The temporal self-attention weight distribution is calculated for the data segments in each window. Based on the weight distribution, the data breakpoints and missing intervals are located. The local feature vectors of adjacent valid windows are extracted and subjected to context-aware weighted aggregation. Temporal coherence filling is performed on the missing intervals to generate a complete and continuous time series sequence. Cross-modal joint standardization is performed on the complete continuous time series. An adaptive scaling strategy based on extreme value truncation and dynamic quantile mapping is adopted to eliminate numerical distribution offsets and scale differences between different physical dimensions. The standardized feature data is reorganized into tensor dimensions and concatenated into channels according to the process topology level and equipment physical boundary to generate a standardized high-dimensional feature sequence, which is used as the graph node feature input data in step S2.

3. The method for digital evaluation of the operational health status and preventive maintenance scheduling of a wastewater treatment facility according to claim 1, characterized in that: Based on the process unit topology and material energy flow direction of the wastewater treatment plant, a process-equipment coupled directed graph structure is constructed; specifically including: Analyze the process drawings and operating procedures of the wastewater treatment facility, extract the physical location and functional attributes of each biochemical treatment section, sedimentation tank, sludge treatment unit and auxiliary electromechanical equipment, and generate a set of process unit nodes and a set of equipment physical nodes. Based on the main flow direction of sewage, the direction of return branch, the dosing path of reagents and the power distribution network, the material transfer direction and energy supply direction between nodes are determined. The process unit node set is the macro flow node, and the equipment physical node set is the micro execution node. Directed connection edges with clear direction are established to form the initial topological skeleton. Configure spatial hierarchy identifiers and running role labels for each node in the initial topology skeleton, bind transmission medium type identifiers and energy conversion attribute identifiers for each connection edge, and preset edge weight storage units to carry the causal strength data for subsequent dynamic calculation; Perform graph connectivity verification and topology redundancy pruning on the initial topology skeleton, remove invalid isolated branches and merge parallel redundant pipeline associations, and construct a process-equipment coupled directed graph structure with hierarchical progressive characteristics and complete material and energy conduction links.

4. The method for digital evaluation of the operational health status and preventive maintenance scheduling of a wastewater treatment facility according to claim 3, characterized in that: By calculating the conditional mutual information and Granger causality index of the node feature sequences, the dynamic influence strength between nodes is quantified, a time-varying causal adjacency matrix is ​​generated, and the standardized high-dimensional feature sequences are mapped to the node space of the time-varying causal adjacency matrix. Specifically, it includes: Using the node connection relationship in the process-equipment coupled directed graph structure as the calculation basis, the standardized high-dimensional feature sequence is periodically sliced ​​according to a preset sliding time window to extract the feature evolution sequence of each node in a continuous time segment; For node pairs with directly connected edges in a graph structure, the conditional mutual information value is calculated to remove indirect association interference transmitted through intermediate nodes and lock the direct information coupling strength between nodes; simultaneously, a multi-step time-lag prediction model is constructed to solve the Granger causality index and evaluate the contribution rate of upstream node state fluctuations to the time-series prediction of downstream node feature evolution. The conditional mutual information value and Granger causality index are adaptively normalized and weighted to generate a comprehensive dynamic influence intensity index. Based on this index, the preset edge weight storage unit in the process-equipment coupled directed graph structure is updated in a rolling manner, and a time-varying causal adjacency matrix representing the dynamic evolution of the interaction strength of nodes under different operating conditions is output. Based on the weight distribution pattern of the time-varying causal adjacency matrix, graph topology space transformation and neighborhood feature aggregation operations are performed on the standardized high-dimensional feature sequence. The original multi-dimensional time series data is projected onto the node embedding space corresponding to the time-varying causal adjacency matrix, generating mapped node space features that integrate physical topological association and dynamic causal logic, which serve as the input source for subsequent health latent variable extraction.

5. The method for digital evaluation of the operational health status and preventive maintenance scheduling of a wastewater treatment facility according to claim 1, characterized in that: The mapped node spatial features are input into a graph-structured variational autoencoder to extract a sequence of health latent variables reflecting equipment degradation and process efficiency decline. A mechanism regularization term incorporating Monod biochemical reaction kinetic constraints and hydraulic residence time balance equations is introduced to perform hidden Markov state decoding on the health latent variable sequence, outputting the dynamic health state probability transition matrix of each facility component and the plant-wide health evolution trajectory. Specifically, this includes: The mapped node spatial features are input into the encoding mapping layer of the graph structure variational autoencoder. The spatiotemporal dependency representations of neighboring nodes are aggregated through the graph topology message passing mechanism. After variational reparameterization processing, they are projected onto a low-dimensional continuous latent space to generate an initial health latent variable sequence that reflects the accumulation of physical wear and tear of equipment and the fluctuation of biochemical efficiency of process units. The initial health latent variable sequence is input into the mechanism regularization constraint module, and the mechanism regularization term containing Monod biochemical reaction kinetic constraints and hydraulic residence time balance equation is loaded simultaneously. By calculating the deviation penalty of the substrate degradation rate and microbial growth curve implied by the initial health latent variable sequence from the standard kinetic model, and verifying the mass conservation residual between the volumetric loading distribution and the actual hydraulic residence time corresponding to the initial health latent variable sequence, a mechanism compliance gradient signal is generated. The mechanism compliance gradient signal is used to perform backpropagation joint optimization on the graph structure variational autoencoder to remove the physical mismatch noise in the pure data-driven mode and output the mechanism-constrained health latent variable sequence. The sequence of mechanism-constrained latent health variables is input into the Hidden Markov State Decoder to construct the observation probability emission mapping between the latent variable feature space and the preset discrete health level set. The hidden state transition logic of adjacent time segments is analyzed by dynamic programming path search and forward and backward probability smoothing algorithm, the temporal conditional probability between each health level is calculated, and a dynamic health state probability transition matrix characterizing the degradation rate and functional recovery potential of each facility component is generated. Based on the dynamic health status probability transition matrix of each facility component, and relying on the hierarchical transmission topology of the process-equipment coupled directed graph structure, cross-component health status risk weighted fusion and plant-wide efficiency extrapolation are performed. The state transition probabilities of single-point components are cascaded and aggregated along the material flow and energy flow backbone network to output the plant-wide health evolution trajectory that characterizes the spatiotemporal evolution law of the overall operational health of the wastewater treatment plant, and serves as the state space construction benchmark for the preventive maintenance Markov decision process in step S4.

6. The method for digital evaluation of the operational health status and preventive maintenance scheduling of a wastewater treatment facility according to claim 1, characterized in that: Using the dynamic health state probability transition matrix, the plant-wide health evolution trajectory, and the currently available maintenance resource pool as the state space, a preventative maintenance Markov decision process is constructed; specifically, it includes: The dynamic health status probability transition matrix is ​​fused with the plant-wide health evolution trajectory through time-series slicing. A snapshot of the local health status in the current decision-making cycle is extracted and spliced ​​with the resource type, quantity and skill level attributes in the current available maintenance resource pool through multi-dimensional tensor splicing to construct a multi-dimensional state representation vector that represents the real-time operating status and maintenance support capability of the system. Based on the node hierarchy distribution of the process-equipment coupled directed graph structure and the physical boundaries of facility components, a discretized maintenance action space is defined. The available resources in the current available maintenance resource pool are matched with the potential intervention behaviors of each component to generate a candidate set of action combinations covering preventive maintenance, component replacement, condition detection and resource standby. A multi-objective collaborative reward evaluation model is constructed, with the safety margin for effluent quality compliance, the expected gain of the remaining service life of the equipment, the resource consumption cost of a single maintenance, and the risk of process operation interruption as the core optimization indicators. A dynamic time discount factor is set in combination with the efficiency decay slope of the plant-wide health evolution trajectory to quantify the comprehensive net contribution value of different action combinations to the long-term health benefits of the wastewater treatment facility. The dynamic health state probability transition matrix is ​​conditionally coupled with the action combination candidate set to generate an action condition state transition kernel. This simulates the transition probability distribution of the health state of each facility component after executing a specific maintenance strategy. The multi-dimensional state representation vector, action combination candidate set, multi-objective collaborative reward evaluation model, and action condition state transition kernel are then structurally encapsulated to form a preventive maintenance Markov decision process that includes a complete state space, action space, reward mechanism, and state transition logic. This provides a standardized decision base for subsequent policy network training.

7. The method for digital evaluation of the operational health status and preventive maintenance scheduling of a wastewater treatment facility according to claim 6, characterized in that: A deep policy network embedded with a differentiable process-constrained manifold is deployed. This manifold, through a projection operator, maps the risk of effluent quality exceeding standards, the uninterrupted constraints of critical biochemical stages, and the mutual exclusion logic of equipment maintenance into a continuously differentiable penalty gradient field. This field is then jointly gradient-optimized with the action value function of the deep policy network to generate the optimal maintenance resource allocation matrix and a multi-cycle scheduling time window. Specifically, this includes: A deep policy network is constructed, which includes a spatiotemporal feature encoding layer, a policy probability output layer, and an action value evaluation layer. The multidimensional state representation vector of the preventive maintenance Markov decision process is input into the spatiotemporal feature encoding layer to extract the hidden state representation and output the initial maintenance action probability distribution and the corresponding benchmark action value score. Define the geometric topological boundary of the differentiable process constraint manifold, transform the risk of effluent water quality exceeding the standard into a continuous cost mapping of water quality indicators deviating from the safety threshold, transform the uninterrupted constraint of the key biochemical stage into a time-series overlap penalty term of maintenance operation time window and biochemical reaction continuous period, abstract the mutual exclusion logic of equipment maintenance into a Boolean logic continuous relaxation expression of resource occupation conflict, and use the orthogonal projection operator to smoothly project the above three types of discrete hard constraints, namely the risk of effluent water quality exceeding the standard, the uninterrupted constraint of the key biochemical stage and the mutual exclusion logic of equipment maintenance, into a high-dimensional parameter space to generate a continuous differentiable penalty gradient field characterizing the constraint violation sensitivity and directional derivative; By introducing a continuously differentiable penalty gradient field into the action value function optimization path of a deep policy network, a joint loss objective of policy gain and constraint penalty is constructed. The policy gradient and manifold constraint gradient are calculated simultaneously through the backpropagation mechanism. Gradient vector synthesis and adaptive adjustment of the optimization step size are performed between the upward direction of maximizing action value and the projection direction of the feasible region of the constraint manifold. This drives the parameters of the deep policy network to iterate towards the convergence frontier that takes into account both long-term maintenance benefits and process safety boundaries. After parameter convergence, the deep policy network is subjected to multi-cycle action sequence sampling and deterministic policy decoding. The output continuous action probability distribution is discretized into specific maintenance execution instructions. Based on the process material flow and energy transmission link, resource quota secondary verification and time window conflict resolution are performed to generate the optimal maintenance resource allocation matrix and multi-cycle scheduling time sequence window covering all plant facilities and components, thus completing the digital output of the preventive maintenance scheduling strategy.

8. The method for digital evaluation of the operational health status and preventive maintenance scheduling of a wastewater treatment facility according to claim 1, characterized in that: Digital maintenance work orders are issued and executed based on the optimal maintenance resource allocation matrix; specifically including: The optimal maintenance resource allocation matrix and multi-cycle scheduling time window are structured and parsed to extract the target component identifier, resource quota details, execution time window and pre-security verification rules of each maintenance task, and generate a standardized digital maintenance work order data package. Digital maintenance work order data packages are pushed to the corresponding mobile operation and maintenance terminals and field control units through industrial IoT edge gateways. Simultaneously, digital twin interactive views and standardized operation instructions bound to the target components are retrieved to guide on-site personnel to perform preventive maintenance, component replacement or condition detection operations in sequence. During maintenance, IoT sensors deployed at key nodes collect data in real time on equipment disassembly and assembly status, lubrication volume, electrical parameter remeasurement values, and short-term bypass switching records of process sections, forming a maintenance process trajectory data stream with time and space stamps. After the maintenance operation is completed, a recovery command is issued to the process control system. The operation recovery data and effluent compliance rate after the maintenance are collected. Abnormal fluctuations are filtered and steady-state features are extracted from the operation recovery data to generate a post-maintenance performance benchmark dataset. The post-maintenance performance benchmark dataset and the maintenance process trajectory data stream are used as feedback input sources for subsequent model verification and closed-loop optimization.

9. The method for digital evaluation of the operational health status and preventive maintenance scheduling of a wastewater treatment facility according to claim 8, characterized in that: Collect operational condition recovery data and effluent compliance rate after maintenance execution, calculate Wasserstein distance between actual health status distribution and predicted distribution of dynamic health status probability transition matrix; when Wasserstein distance exceeds adaptive threshold, trigger online collaborative fine-tuning of hidden Markov state decoder transition probability and deep policy network weight, update causal strength weight of time-varying causal adjacency matrix, and form evaluation and scheduling closed-loop iteration. Specifically, it includes: Based on the post-maintenance performance benchmark dataset, the actual operational health level ratio of each facility component is restored through the health status inverse mapping rule, and the empirical distribution of the actual health status after maintenance is constructed; the predicted output of the dynamic health status probability transition matrix output in step S3 within the corresponding maintenance cycle is extracted, and the theoretical predicted health status distribution is constructed. The optimal transmission distance metric algorithm is used to quantify the cumulative cost of the actual health status empirical distribution and the theoretical predicted health status distribution in the probabilistic quality migration process, and generate Wasserstein distance values ​​that characterize the degree of model prediction bias; combined with the historical periodic distance fluctuation mean and the seasonal disturbance coefficient of influent water quality, the adaptive bias tolerance threshold is dynamically calculated. When the Wasserstein distance value exceeds the adaptive deviation tolerance threshold, it is determined that the current health evolution model and scheduling strategy have experienced operational condition drift, triggering an online collaborative fine-tuning mechanism. The residual gradient of the actual health state empirical distribution and the theoretically predicted health state distribution is used to update the state transition probability of the Hidden Markov State Decoder using Bayesian prior. At the same time, the actual water discharge compliance rate and resource consumption feedback are used as reinforcement learning reward signals to perform small-step gradient backtracking fine-tuning on the action value evaluation layer of the deep policy network. The fine-tuned hidden state transition logic is fed back to the causal quantification step S2. Based on the actual coupling changes in the evolution of node features before and after maintenance, the fusion weights of conditional mutual information and Granger causality index are recalibrated, and the causal strength weights of the time-varying causal adjacency matrix are updated on a rolling basis. The updated graph structure parameters and optimized policy network parameters are synchronously loaded into the state space construction and policy reasoning process of the next decision cycle, realizing the adaptive closed-loop iteration of digital health status evaluation and preventive maintenance scheduling.

Images