Method and system for low-carbon denitrification and efficient biofilm formation in industrial circulating water aquaculture

Abstract

The application relates to the technical field of water treatment, and discloses a method and system for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture. The method comprises the following steps: collecting biochemical bin environmental parameters, combining a flow field three-dimensional grid to construct a space-time state attribute graph tensor; configuring a time domain convolution kernel to extract biological activity evolution characteristics; inputting the characteristics into a pre-trained prediction network to output degradation rate and biological membrane thickness prediction values; generating a virtual impact load tensor after the membrane thickness meets the standard, obtaining a degradation rate fluctuation curve and calculating system anti-recovery time; when the recovery time meets the standard, taking low carbon and short cycle as the target, generating a dosing and disturbance control benchmark vector through reinforcement learning; outputting a safety instruction execution control through expert rule limiting amplitude, continuously checking the degradation rate residual error stability, and determining that the biofilm formation is completed and the waterway is switched after the biofilm formation is completed. The application solves the problems of long biofilm formation period, high carbon source consumption and difficult prediction of system stability, shortens the biofilm formation period, reduces carbon consumption and improves the impact resistance and operation stability of the biochemical system.

Description

Technical Field

[0001] This application relates to the field of water treatment technology, and in particular to a method and system for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture. Background Technology

[0002] Recirculating aquaculture systems (RAS) utilize biofilm reactors for water purification, a core technological approach to ensuring nitrogen balance in aquaculture water. The biofilm formation process—the attachment of microorganisms to the substrate surface and the formation of a biodegradation-enabled microbial community—directly determines the system's stable operation. However, current biofilm formation technologies generally face limitations such as long cultivation cycles, high carbon source consumption, and unpredictable system stability, making it difficult to meet the demands of modern, high-efficiency aquaculture.

[0003] Existing biofilm formation methods largely rely on empirical parameters to control aeration rate and feed concentration, failing to accurately grasp the spatiotemporal coupling relationship between the internal flow field distribution and microbial growth status within the reactor. For example, in reactors employing aeration and tumbling flow fields, significant differences exist in dissolved oxygen concentration and substrate transfer rates at different spatial locations, resulting in highly uneven microbial attachment speeds on the packing surface. When the flow velocity in the corner areas of the reactor is too low, nutrient supply in these areas is severely insufficient, leading to stagnation of microbial reproduction; meanwhile, the excessive shear force generated near the aeration ports continuously washes away the nascent, unstable biofilm. This spatial heterogeneity severely hinders the overall biofilm formation process due to localized weak points; even if most of the packing material has formed effective bacterial communities, the system still needs to wait for the lagging areas to reach the required standards before it can be put into operation. Furthermore, existing control strategies lack precise matching between microbial reproductive cycles and metabolic patterns. The generation time of nitrifying bacteria is approximately ten hours; if the aeration interval is improperly set, newly attached bacteria will detach due to lack of oxygen; an imbalance in the carbon-to-nitrogen ratio leads to excessive proliferation of heterotrophic bacteria, thereby inhibiting autotrophic nitrifying bacteria. In practice, the rapid biofilm formation is often accelerated by blindly increasing the amount of carbon source added. This not only significantly increases operating costs but also causes large areas of biofilm to detach after the system is fully operational due to sudden high-load shocks. For example, when a sudden increase in stocking density causes a surge in ammonia nitrogen concentration, a biofilm system that has not undergone shock resistance assessment may completely lose its degradation capacity within hours, leading to a rapid deterioration of water quality. Summary of the Invention

[0004] To address the above problems, this application provides a method and system for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture, thereby solving the problems existing in the background art.

[0005] In a first aspect, this application provides a method for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture systems, comprising: S1. Obtain the state environment parameter sequence collected by the sensor array in the isolation biochemical chamber, determine the spatial adjacency weight of each sensor based on the three-dimensional mesh of the aeration tumbling flow field, and construct the spatiotemporal state attribute graph tensor in combination with the state environment parameter sequence. S2. Obtain the reproductive cycle parameters of the target microorganism and configure the temporal convolution kernel. Use the configured temporal convolution kernel to perform graph convolution and temporal convolution operations on the spatiotemporal state attribute graph tensor to extract the biological activity evolution feature sequence. S3. Input the bioactivity evolution feature sequence into the pre-trained prediction network and output the predicted value of the target substance degradation rate and the predicted value of the biofilm thickness at future time steps. The loss function of the prediction network includes a penalty term constructed based on the standard biochemical kinetic equation and the carbon-nitrogen ratio limiting metabolic boundary parameter. S4. Determine whether the predicted value of the biofilm thickness is greater than or equal to the preset thickness threshold. If so, generate a virtual impact load tensor of the simulated substance release curve, and obtain the fluctuation curve of the target substance degradation rate based on the virtual impact load tensor. S5. Based on the fluctuation curve, determine the resilience recovery time of the isolation biochemical chamber. When the resilience recovery time is less than the preset time threshold, take the minimum material consumption and the shortest biofilm formation cycle as the optimization objectives, and generate a benchmark control vector containing the material addition pump flow rate and the opening degree of the fluid disturbance valve through reinforcement learning operators. S6. Based on the preset expert rule base, the baseline control vector is subjected to boundary limiting to generate a safety control command. The dosing pump and fluid disturbance device are driven according to the safety control command. The residual between the actual degradation rate and the predicted value of the target substance degradation rate is calculated. If the residual is less than the preset residual threshold for N consecutive monitoring cycles, the biofilm formation is determined to be successful and a valve switching command to connect the target waterway is output.

[0006] Secondly, this application provides a system for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture, used to implement the aforementioned method for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture, comprising: The tensor construction module is used to obtain the state environment parameter sequence collected by the sensor array in the isolation biochemical chamber, determine the spatial adjacency weight of each sensor based on the three-dimensional mesh of the aeration tumbling flow field, and construct the spatiotemporal state attribute graph tensor in combination with the state environment parameter sequence. The feature extraction module is used to obtain the reproductive cycle parameters of the target microorganism and configure the temporal convolution kernel. The configured temporal convolution kernel is used to perform graph convolution and temporal convolution operations on the spatiotemporal state attribute graph tensor to extract the biological activity evolution feature sequence. The prediction output module is used to input the bioactivity evolution feature sequence into a pre-trained prediction network and output the predicted values ​​of the target substance degradation rate and biofilm thickness at future time steps. The loss function of the prediction network includes a penalty term constructed based on the standard biochemical kinetic equation and the carbon-nitrogen ratio limiting metabolic boundary parameter. The impact test module is used to determine whether the predicted value of the biofilm thickness is greater than or equal to a preset thickness threshold. If so, a virtual impact load tensor of the simulated substance release curve is generated, and the fluctuation curve of the target substance degradation rate is obtained based on the virtual impact load tensor. The benchmark generation module is used to determine the resilience recovery time of the isolation biochemical chamber based on the fluctuation curve. When the resilience recovery time is less than a preset time threshold, the module generates a benchmark control vector containing the material addition pump flow rate and the opening degree of the fluid disturbance valve by using reinforcement learning operators with the optimization objectives of minimizing material consumption and minimizing biofilm formation period. The safety execution module is used to generate safety control instructions by limiting the baseline control vector according to the preset expert rule base, drive the dosing pump and fluid disturbance device according to the safety control instructions, calculate the residual between the actual degradation rate and the predicted value of the target substance degradation rate, and if the residual is less than the preset residual threshold for N consecutive monitoring cycles, the biofilm is determined to be successful and a valve switching instruction to connect the target water channel is output.

[0007] Thirdly, this application provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the aforementioned method for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture.

[0008] The beneficial effects of this application are as follows: This application proposes a method and system for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture. Compared with existing technologies, this application obtains environmental parameters through a sensor array, determines the spatial adjacency weights of each sensor based on a three-dimensional mesh of the aeration and tumbling flow field, and constructs a spatiotemporal state attribute graph tensor. This explicitly incorporates the spatial heterogeneity information inside the reactor into the modeling, overcoming the shortcomings of traditional methods that ignore the influence of the spatial distribution of the flow field on biofilm formation. Subsequently, graph convolution and temporal convolution operations are performed using temporal convolution kernels that match the reproductive cycle of the target microorganisms to extract the biological activity evolution feature sequence containing spatial mass transfer information and temporal evolution law, enabling the subsequent prediction network to accurately grasp the growth rhythm and metabolic kinetics of the microorganisms. A penalty term based on the standard biochemical kinetic equation and the carbon-nitrogen ratio limiting metabolic boundary parameter is introduced into the prediction network to constrain the prediction results to conform to the inherent laws of biochemical reactions and avoid invalid predictions that deviate from the actual kinetic boundary. Based on this, the resilience and recovery capacity of the biofilm are assessed by virtual impact loads to ensure that the system after biofilm formation can withstand the impact of actual working conditions such as sudden changes in aquaculture density. With material consumption and biofilm formation cycle as optimization objectives, control vectors are generated through reinforcement learning, and then the execution safety is ensured by boundary limiting of expert rule base, forming a complete closed-loop control link of "perception-prediction-evaluation-optimization-execution-verification". Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of a method for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture according to this application; Figure 2 This is a graph showing the comparison of cumulative carbon source consumption in this application; Figure 3 This is a graph showing the comparison of biofilm formation cycles in 20 independent experiments conducted in this application; Figure 4 This is a schematic diagram of the structure of a low-carbon, nitrogen-removing, and highly efficient biofilm formation system for factory-scale recirculating aquaculture in this application. Detailed Implementation

[0011] This application provides a method and system for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0012] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of a method for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture systems, as described in this application, includes: S1. Obtain the state environment parameter sequence collected by the sensor array in the isolation biochemical chamber, determine the spatial adjacency weight of each sensor based on the three-dimensional mesh of the aeration tumbling flow field, and construct the spatiotemporal state attribute graph tensor by combining the state environment parameter sequence.

[0013] In one specific embodiment, performing S1 includes the following steps: The sequence of state and environmental parameters collected by the sensor array in the isolation biochemical chamber is obtained. The velocity field distribution is obtained by performing fluid dynamics simulation on the aeration and tumbling flow field in the isolation biochemical chamber. The physical position of each sensor is mapped to a three-dimensional mesh to obtain the corresponding node of the sensor. Based on the velocity field distribution, the dominant streamline direction of each sensor corresponding node is extracted. Based on the angle between the line connecting each pair of corresponding sensor nodes and the dominant streamline direction and the spatial distance between corresponding nodes, the spatial adjacency weight between each sensor is calculated, and a spatial adjacency weight matrix is ​​generated. Using the monitoring locations of each sensor as graph nodes, the multidimensional environmental parameters corresponding to each time step in the state-environment parameter sequence as node attributes, and the corresponding elements of the spatial adjacency weight matrix as edge weights between graph nodes, a spatiotemporal state attribute graph tensor is constructed.

[0014] Specifically, the isolation biochemical tank refers to a biochemical tank that operates independently from the circulating water system of the aquaculture pond during the biofilm formation process. Inside the tank, soaked or crushed bass feed and a special carbon source are added according to the biofilm formation operation procedures for microbial cultivation, and the water temperature inside the tank must be maintained above 26 degrees Celsius. The sensor array refers to a collection of various water quality sensors spatially distributed within the isolation biochemical tank, including water temperature sensors, dissolved oxygen sensors, pH sensors, ammonia nitrogen sensors, nitrite nitrogen sensors, and nitrate nitrogen sensors. These sensors are layered along the length, width, and depth of the biochemical tank, ensuring coverage of both the high-speed tumbling zone near the aerator and the low-speed zone in the corners of the tank. The environmental parameter sequence refers to the sequence formed by arranging water quality data collected continuously by each sensor according to a fixed sampling period in chronological order. The multidimensional environmental parameters are six types of parameters at the same sampling time: water temperature, dissolved oxygen concentration, pH value, ammonia nitrogen concentration, nitrite nitrogen concentration, and nitrate nitrogen concentration. The unit of water temperature is degrees Celsius, and the units of dissolved oxygen concentration, ammonia nitrogen concentration, nitrite nitrogen concentration, and nitrate nitrogen concentration are milligrams per liter. The pH value is dimensionless. For example, the sampling period is set to 10 minutes, and the collected data is handled according to three levels of status: normal values ​​are directly stored in the database; if a parameter exceeds the upper limit of the range or communication is interrupted at a certain moment, instead of using the arithmetic average of the preceding and following moments to fill in the missing data, the point is marked as missing, and the measurement values ​​of the effective nodes in the neighborhood are weighted and filled in by taking advantage of the strong correlation between water quality in adjacent spatial locations in the aeration tumbling flow field and using spatial adjacency weights. At the same time, it is explicitly prohibited to average and smooth out high values ​​exceeding the range to avoid losing the true signals of impact or local accumulation; if the same sensor is missing for three consecutive moments or more than one-third of the nodes are abnormal at the same moment, an alarm is triggered and the control process is suspended until the data link is manually confirmed to be restored before continuing. In this way, the integrity and reliability of the sequence are guaranteed while retaining abnormal information.

[0015] The aeration tumbling flow field refers to the three-dimensional flow pattern formed when air is introduced into the bottom of the isolation biological chamber by an aeration device, causing the air bubbles to rise, tumbling the water, and driving the biochemical ball packing material to tumble up and down. Computational fluid dynamics simulation is performed on this flow field: a geometry is established according to the actual geometric dimensions of the isolation biological chamber, and the geometry is discretized into a three-dimensional mesh. The three-dimensional mesh is a computational mesh composed of several hexahedral elements; for example, the side length of the hexahedral elements is taken to be 0.05 meters to 0.1 meters. The water in the chamber is considered as an incompressible fluid, and the continuity equation and Reynolds-mean momentum equation are solved in steady state. The continuity equation is... , The Hamiltonian operator represents the partial derivative with respect to spatial coordinates. U is the velocity vector field, i.e., the velocity field distribution output by the simulation, in meters per second. The velocity vector at the corresponding sensor node is extracted subsequently. That is, the value of the velocity vector field U at this node; the physical meaning of the continuity equation is that the mass of water flowing in and out of any three-dimensional grid cell is equal; the Reynolds-averaged momentum equation is obtained by averaging the transient Navier-Stokes equation over time, which is... Where ρ is the density of water, in kilograms per cubic meter. This is the convection term, characterizing the transport of momentum in water during flow. P represents pressure, measured in Pascals. For the pressure gradient term, The dynamic viscosity of water is expressed in Pascals per second. The viscosity is turbulent eddy viscosity, measured in Pascals per second. For the Laplace operator, The term represents the viscous diffusion term, where g is the gravitational acceleration vector in meters per second squared. The Reynolds stress introduced in the equation is closed according to the Boussinesq eddy viscosity assumption and expressed as turbulent eddy viscosity. Characterization: all terms in the equations are in Newtons per cubic meter; the standard k-ε model is used for turbulence, and the turbulent eddy viscosity is calculated according to... Calculation, where The model's empirical constant is set to 0.09, k is the turbulent kinetic energy in square meters per second squared, and ε is the turbulent kinetic energy dissipation rate in square meters per second cubicd. The area where the aerator is located is set as the velocity inlet boundary, with the inlet velocity obtained by dividing the aeration rate by the cross-sectional area of ​​the aeration region. The water surface inside the tank is set as the pressure outlet boundary, and the tank walls and fixed structural components inside the tank are set as non-slip wall boundaries. The bio-ball packing area is not treated with wall surface treatment; instead, the entire packing accumulation area is regarded as a porous medium region, and a resistance source term is added to the momentum equation. Describes the flow hindrance effect when fluid passes through a packing layer, where K is the permeability coefficient. The inertial drag coefficient is determined by the packing density and experimental pressure drop data, ensuring that the simulated velocity field distribution reflects the actual flow pattern of water flowing through the gaps in the packing material as bubbles rise, rather than the erroneous assumption of fluid flowing around a fixed rigid body. The biofilm attachment and detachment process on the packing surface does not depend on the wall boundary conditions, but is indirectly transmitted through the spatial adjacency weights derived from the flow field velocity distribution in subsequent graph convolution operations to convey the influence of flow field shear force and mass exchange intensity on the biofilm formation process. Convergence is determined when the residuals of all equations are below 0.0001, and the velocity field distribution is output. The velocity field distribution is the velocity vector at the center of each three-dimensional grid cell, which contains velocity components along the three coordinate axes, in meters per second. Mapping the physical position of each sensor to the three-dimensional grid means: reading the three-dimensional coordinates of each sensor's installation position, searching for the grid cell closest to that coordinate in the three-dimensional grid, and marking the center of that grid cell as the corresponding node of the sensor, so that each sensor has a unique node in the flow field.

[0016] The dominant streamline direction is defined as the velocity vector at node i. The direction of the vector represents the main direction of water flow and material transport at node i. For any two sensors corresponding to nodes i and j, the spatial coordinate vector of node i represents the direction of the main flow and material transport at node j. Spatial coordinate vector of node j Subtract the values ​​to obtain the line vectors. The magnitude of the connecting vector This represents the spatial distance between the two nodes, in meters; the velocity vector at node i is taken. As the dominant streamline direction from i to j, the connecting vector Angle with the direction of the dominant streamline cosine value according to The calculation, in physical terms, indicates whether the mainstream fluid flow at node i directly points towards node j, reflecting the smoothness of material transport from i to j. This represents the dot product of two vectors. Represents the velocity along the line connecting nodes i and j; spatial adjacency weight. Calculation, where This is the distance attenuation coefficient, in meters, and is taken as half of the effective water depth of the isolation biological chamber. For example, when the effective water depth is 2.5 meters... Take 1.25 meters, spatial adjacency weight Let be a dimensionless number ranging from 0 to 1. Iterate through all pairwise node combinations and calculate the ... Fill in the i-th row and j-th column of the matrix, and set the diagonal elements to 1 to indicate that the node is completely associated with itself, to generate an N-order spatial adjacency weight matrix, where N is the number of sensors. For example, when 12 sensors are arranged in the isolation biochemical chamber, N is 12, and a 12-row, 12-column spatial adjacency weight matrix is ​​generated.

[0017] Since the six types of environmental parameters have different dimensions and numerical ranges, directly using them as node attributes would cause parameters with larger numerical ranges to dominate in subsequent convolution operations. Therefore, max-min standardization is performed on each type of parameter in the state environmental parameter sequence. That is, a sample value is subtracted from the minimum value of that type of parameter in the sequence, and then divided by the difference between the maximum and minimum values ​​of that type of parameter in the sequence, so that each type of parameter is converted into a dimensionless number of 0-1. The spatiotemporal state attribute graph tensor is composed of a node attribute tensor and a spatial adjacency weight matrix: the monitoring position of each sensor is taken as the graph node, and the monitoring positions of N sensors mean that the graph node set contains N nodes in total; the 6-dimensional environmental parameter vector of each sampling time after standardization is taken as the node attribute of the corresponding graph node at that time. The node attribute is the feature vector describing the state of the graph node. The node attributes of all sampling times are stacked in chronological order to form a node attribute tensor with dimensions T×N×6, where T is the total number of sampling times; the corresponding elements in the spatial adjacency weight matrix are taken as the edge weight between graph nodes. The edge weight is the value of the edge connecting two graph nodes, which represents the strength of the material transfer correlation between the spatial positions monitored by the two sensors. The data correspondence is as follows: the six types of water quality data output by the six types of sensors correspond to the standardized six-dimensional node attributes; the T sampling times correspond to the T time slices of the node attribute tensor; the N sensors correspond to the N graph nodes; the flow field correlation between each pair of sensors corresponds to the N×N edge weights in the spatial adjacency weight matrix; the spatiotemporal state attribute graph tensor constructed in this way simultaneously carries the temporal evolution information of water quality parameters and the spatial distribution information of the aeration tumbling flow field during the biofilm formation process, serving as the input for subsequent graph convolution and temporal convolution operations.

[0018] S2. Obtain the reproductive cycle parameters of the target microorganism and configure the temporal convolution kernel. Use the configured temporal convolution kernel to perform graph convolution and temporal convolution operations on the spatiotemporal state attribute graph tensor to extract the biological activity evolution feature sequence.

[0019] In one specific embodiment, performing S2 includes the following steps: Obtain the reproductive cycle parameters of the target microorganism that dominates denitrification during biofilm formation, and configure the receptive field size of the temporal convolution kernel based on the reproductive cycle parameters; Perform graph convolution operation on the spatiotemporal state attribute graph tensor, aggregate the neighborhood attributes of each node according to the edge weights of the graph structure, extract the mass transfer gradient between nodes, and obtain the node feature sequence containing spatial mass transfer information. The configured temporal convolution kernel is used to perform temporal convolution operation on the node feature sequence to extract the evolutionary pattern of biological activity in the time dimension and generate a biological activity evolution feature sequence.

[0020] Specifically, the target microorganisms refer to the microbial community that attaches to and grows on the surface of the biofilm packing material during the biofilm formation process and dominates the denitrification function. These include ammonia-oxidizing bacteria that oxidize ammonia nitrogen to nitrite nitrogen under aerobic conditions, nitrite-oxidizing bacteria that oxidize nitrite nitrogen to nitrate nitrogen in one step, and denitrifying bacteria that reduce nitrate nitrogen to nitrogen gas under special carbon source addition conditions. The reproductive cycle parameter refers to the generation time of the target microorganisms, i.e., the time required for the bacterial population to complete one generation of proliferation. Since the generation time of ammonia-oxidizing bacteria is significantly longer than that of denitrifying bacteria, the nitrification process constitutes a limiting factor in the biofilm formation cycle. Therefore, the generation time of ammonia-oxidizing bacteria is taken as the reproductive cycle parameter τ, in hours. The temporal convolution kernel refers to the set of convolution weights that slide along the time dimension, and the receptive field refers to the number of consecutive sampling times covered by one operation of the temporal convolution kernel, denoted as R. The calculation method for configuring the receptive field size based on the reproductive cycle parameters is as follows: R is the integer obtained by rounding up the ratio of τ to the sampling period Δt, where Δt is the sampling period of the state environment parameter sequence in S1. After unifying τ and Δt to hours, the ratio becomes dimensionless. Rounding up ensures that the receptive field completely covers one generation time. For example, the generation time of ammonia-oxidizing bacteria is about ten hours. τ is 12 hours, Δt is 10 minutes (1 / 6 of an hour), and R is 12 divided by 1 / 6 and rounded up to 72. This means that one operation of the time-domain convolution kernel covers 72 consecutive sampling moments, corresponding to all observation data within one generation time. The configuration logic is as follows: when the receptive field is smaller than the generation time, the convolution window cannot cover the complete activity change process of the bacteria from attachment to proliferation, and the extracted patterns are incomplete; when the receptive field is larger than the generation time, the convolution window mixes in early historical data unrelated to the current metabolic state, weakening the correspondence between features and biological activity.

[0021] Graph convolution refers to the weighted aggregation operation performed on node attributes according to the connection relationships and edge weights between graph nodes in the graph structure. Let X be the node attribute tensor in the spatiotemporal state attribute graph tensor, with dimensions T×N×6, where T is the total number of sampling times, and N is the number of sensors (i.e., the number of graph nodes). Each sampling time t corresponds to a node attribute matrix of dimension N×6. Let A be the spatial adjacency weight matrix, and let the element in the i-th row and j-th column of A be the spatial adjacency weight calculated by S1. Before performing graph convolution, apply self-loops to A and normalize it: add A to the N-order identity matrix I, which is a square matrix with diagonal elements of 1 and all other elements of 0. Adding ensures that each node retains its own attributes during aggregation; calculate the degree matrix D, which is a diagonal matrix where the i-th diagonal element is equal to the sum of all elements in the i-th row of matrix A+I; then proceed according to... Calculate the normalized adjacency matrix This ensures that the range of aggregated values ​​does not become unbalanced with the number of node connections, guaranteeing the numerical stability of multi-layer convolution operations. The node attribute matrix for each sampling time t... Perform a 2-layer graph convolution, with the first layer outputting... , second layer output ,in For dimension The trainable weight matrix, For dimension The trainable weight matrix is ​​used, with ReLU as the activation function. The output is 0 when the input is less than 0, and the output is the input itself otherwise. The first layer aggregates the attributes of each node's directly adjacent nodes. The second layer further aggregates the results, ensuring each node covers a two-hop neighborhood with which it has an indirect material transfer relationship. For example, Take 16, Taking 8, that is, the first layer maps the 6-dimensional attributes of each node to 16-dimensional intermediate features, and the second layer compresses it into 8-dimensional features. The i-th row vector This is the feature representation of node i after aggregating neighborhood attributes at time t. Since the adjacency weight is derived from the velocity field distribution of the aeration tumbling flow field, the aggregation result carries spatial mass transfer information.

[0022] For node i and its neighboring node j, according to Calculate the mass transfer gradient between the two at time t, where Let be the spatial distance between two nodes in S1, in meters. The eigenvector is a dimensionless number, and the mass transfer gradient is in units of one-meter. To ensure dimension consistency when concatenating the mass transfer gradient with the dimensionless eigenvector, the mass transfer gradient is multiplied by the distance attenuation coefficient in S1. Transform into a dimensionless form. The unit is meters, and the value is consistent with the calculation of the S1 spatial adjacency weight, that is, half of the effective water depth of the isolation biological chamber. For example, when the effective water depth is 2.5 meters... Take a value of 1.25 meters. For node i, transfer the dimensionless mass gradient between it and all its adjacent nodes according to the corresponding edge weights. Weighted summation yields the mass transfer gradient vector at node i. A larger edge weight indicates a more direct exchange of matter between two nodes along the flow direction, and a higher proportion of the gradient in the summation. Node pairs whose connecting lines are nearly perpendicular to the dominant streamline direction are suppressed because their edge weights approach zero. The eigenvector of each node at time t is... With the gradient vector of matter transfer Concatenating them along the feature dimensions, since each component of both is a dimensionless number with consistent dimensions, results in a dimension of 2. The node feature vectors are exemplarily 16-dimensional. The node feature vectors of all nodes and all sampling times are arranged in chronological order to form a node feature sequence with a dimension of T×N×16. This sequence simultaneously contains the node's own state, neighborhood aggregation state, and spatial mass transfer information.

[0023] Temporal convolution refers to a one-dimensional convolution operation performed along the time dimension of the node feature sequence using a configured temporal convolution kernel. Graph convolution is performed independently on each sampling time slice, while temporal convolution is performed independently on the time sequence of each graph node. Spatial and temporal operations are concatenated sequentially. Let the 16-dimensional vector of node i at time t in the node feature sequence be denoted as . Perform convolution on the time series of each node: Where Ψ is the trainable weight tensor of the temporal convolution kernel, m is the kernel index, ν is the time offset, and f is the input feature dimension index. Let be the bias parameters of the m-th convolutional kernel, and let the input features be dimensionless numbers, and the output be... The convolution kernel is dimensionless, and its sliding step along the time dimension is one sampling time. Causal padding is used in the convolution, i.e., R−1 zero vectors are padded at the beginning of the time series, ensuring that the output at time t is calculated only from the node features of time t and the preceding 71 times, preventing data leakage from future times to the current operation. An example of 32 convolution kernels is used, meaning each node outputs 32-dimensional features at each time step. All outputs are arranged chronologically to form a biological activity evolution feature sequence with dimensions T×N×32. The evolutionary pattern of biological activity refers to the coordinated change pattern of water quality parameters over a generation timescale: from day 18 to day 21 of biofilm culture, the concentrations of nitrite nitrogen and nitrate nitrogen rise to their highest values, and then gradually fall back to 0 to 0.2 mg / L and 0 to 50 mg / L respectively under the action of denitrification catalyst and special carbon source. Within a generation time window, this process is characterized by a coordinated pattern of accelerated decrease in ammonia nitrogen concentration, peak of nitrite nitrogen concentration, and accumulation and subsequent decline of nitrate nitrogen concentration. Matching the receptive field with the generation time allows the temporal convolution kernel to completely capture this pattern within a window.

[0024] S3. Input the bioactivity evolution feature sequence into the pre-trained prediction network, and output the predicted values ​​of the target substance degradation rate and biofilm thickness at future time steps. The loss function of the prediction network includes a penalty term constructed based on the standard biochemical kinetic equation and the carbon-nitrogen ratio limiting metabolic boundary parameter.

[0025] In one specific embodiment, performing S3 includes the following steps: Obtain the standard biochemical kinetic equation and the carbon-nitrogen ratio limiting metabolic boundary parameter, and construct a penalty term based on the standard biochemical kinetic equation and the carbon-nitrogen ratio limiting metabolic boundary parameter. The penalty term is added to the loss function of the prediction network, and the network parameters of the prediction network are jointly optimized by backpropagation algorithm to obtain the pre-trained prediction network. The evolutionary sequence of biological activity is input into a pre-trained prediction network, and the predicted values ​​of the target material degradation rate and biofilm thickness at future time steps are output through forward propagation.

[0026] Specifically, the target substance refers to ammonia nitrogen co-degraded by ammonia-oxidizing bacteria and nitrite-oxidizing bacteria during biofilm culture. The target substance degradation rate is the ratio of the decrease in ammonia nitrogen concentration in the isolation biochemical chamber within the predicted time span to the initial ammonia nitrogen concentration, a dimensionless number between 0 and 1. The predicted biofilm thickness refers to the average thickness of the microbial biofilm layer attached to the surface of the biochemical ball packing along the normal direction of the packing surface, in micrometers. The input to the prediction network is a feature segment extracted from the biological activity evolution feature sequence output by S2, starting from the current time. The segment length is taken as the receptive field R of the temporal convolution kernel, i.e., 72 sampling times, corresponding to one generation of ammonia-oxidizing bacteria. The segment dimension is 72×N×32, where N is the number of graph nodes and 32 is the feature dimension of the temporal convolution output. Node-dimensional pooling is performed before inputting into the network: for each sampling time, the 32-dimensional feature vectors of all N graph nodes are arithmetically averaged by their components to obtain a 32-dimensional global feature vector z. The m-th component of z is then... Calculation, where For the m-th feature of node i at time t in the bioactivity evolution feature sequence, pooling compresses the spatial state of all monitoring locations under the coverage of the aeration and tumbling flow field into a bioactivity characterization of the entire isolated biochemical chamber. The pooled feature matrix, with a dimension of 72×32, is input into the prediction network in sequence. The prediction network consists of two layers of Long Short-Term Memory (LSTM) networks and two fully connected layers connected in series. In the LSM network, each layer has a hidden dimension of 64. The activation functions for the input, forget, and output gates are all sigmoid functions σ. Candidate states are generated using the hyperbolic tangent function. The forget gate bias is initialized to 1 to maintain gradient propagation over long sequences. The 64-dimensional hidden state at the end of the second LSM network is mapped to 32 dimensions by the first fully connected layer and activated by ReLU. The second fully connected layer maps the 32-dimensional feature to a 2-dimensional output vector. The first component of the output vector is activated by sigmoid to obtain the predicted degradation rate of the target substance at future time steps. The second component is activated by Softplus to obtain the predicted value of biofilm thickness. Softplus activation ensures that the thickness output is a non-negative number.

[0027] The training dataset is constructed from the operational records of historical biofilm formation batches in the isolated biochemical chamber. Training samples are generated using a sliding window method: each sampling time is taken as the endpoint, and the pooled feature matrix of the preceding 72 sampling times is used as input. The actual degradation rate and actual biofilm thickness within a time interval Δφ after that time are used as labels, where Δφ is the duration of the future time step, exemplarily taken as 6 hours, corresponding to 36 sampling periods. The actual degradation rate label is... Calculation, where Let t be the ammonia nitrogen concentration measured by the ammonia nitrogen sensor. The measured ammonia nitrogen concentration at time t+Δφ is expressed in milligrams per liter, and the ratio is dimensionless. The actual biofilm thickness label is obtained weekly by randomly retrieving biospheres from the isolated biochemical chamber, staining them, and measuring the membrane thickness along the cross-section under an optical microscope. The arithmetic mean of five biosphere measurements is taken each time, and labels between adjacent measurement dates are padded using linear interpolation. For example, records from 12 biofilm formation batches, each with approximately 4300 sampling times, are accumulated. After removing 72 sampling times from the initial biofilm formation stage (less than a complete window), approximately 50,000 training samples are generated, divided into training and validation sets at an 8:2 ratio by batch. Since the degradation rate label is dimensionless while the thickness label is in micrometers, the thickness label and predicted thickness are max-min standardized before being used in loss calculation. During inference, the standardized thickness output by the network is de-standardized to restore the micrometer unit before output.

[0028] The loss function L of the prediction network is a weighted superposition of the basic loss, the kinetic penalty term, and the carbon-nitrogen ratio boundary penalty term, i.e. Loss weight Take 0.5, Take 0.3. Base loss. The mean squared error between the predicted degradation rate and the degradation rate label on a batch of samples is added to the mean squared error between the predicted thickness and the standardized thickness label, characterizing the goodness of fit between the prediction and the measured data. Kinetic penalty term. Derived from the standard biochemical kinetic equation, specifically the Monod equation, which characterizes the combined limitation of substrate concentration and dissolved oxygen concentration on the nitrification degradation rate, the ammonia nitrogen degradation rate v is calculated according to... The calculation is performed, where S is the inversely standardized value of the measured ammonia nitrogen concentration at the corresponding time point, and O is the measured dissolved oxygen concentration at the same time point, both in milligrams per liter. The ammonia nitrogen half-saturation constant is taken as 1.0 mg / L. The dissolved oxygen half-saturation constant is taken as 0.5 mg / L. The maximum degradation rate at water temperature W is obtained by temperature correction from a reference value of 20 degrees Celsius. ,in, This represents a reference value of 20 degrees Celsius, where κ is a temperature correction factor, taken as 1.072, for example. Using 2.4 mg / L / h, the dimension of v is mg / L / h. The theoretical degradation rate over the predicted time span is derived from the degradation rate. ,in The dimension of is milligrams per liter, consistent with the dimension of S, and the ratio is a dimensionless number. The kinetic penalty term is calculated according to... The calculation uses M as the batch size and δ as the tolerance, set to 0.05. This means that a gradient is only generated when the predicted degradation rate deviates from the theoretical value of the Monod equation by more than 0.05, thus avoiding excessive constraints on the network due to errors in kinetic parameter measurement. A carbon-to-nitrogen ratio boundary penalty term is also included. Derived from the carbon-nitrogen ratio limiting metabolic boundary parameter, the carbon-nitrogen ratio is calculated as follows: The calculation is as follows: G is the concentration of biodegradable organic carbon in the storage area, calculated from the special carbon source addition record, in milligrams per liter; χ is a dimensionless number representing the upper limit of the carbon-to-nitrogen ratio in the limiting metabolic boundary parameters. For example, take 6, when χ exceeds Heterotrophic bacteria have an advantage in the competition for dissolved oxygen and the attachment space of the packing material. The metabolic upper limit of autotrophic nitrifying bacteria is reduced proportionally, with the reduction factor being... Calculations were performed to derive the limiting degradation rate from the reduced limiting degradation rate. The carbon-to-nitrogen ratio boundary penalty term is calculated as follows: The calculations apply only when the predicted degradation rate exceeds the metabolic limit of the carbon-nitrogen ratio constraint.

[0029] During training, the graph convolutional layers and temporal convolutional kernels in S2, together with the prediction network (LSTM and fully connected layers) in S3, constitute a complete differentiable computational graph. The training set is input in batches of 32. The Adam optimizer, using backpropagation, performs end-to-end joint gradient optimization on all trainable parameters, including the graph convolutional layers, temporal convolutional layers, LSTM network, and fully connected layers. The gradient of the loss function is propagated back from the prediction network output layer to the graph convolutional input layer, ensuring that the feature extraction parameters in S2 are optimized synchronously with the downstream prediction task. The learning rate is set to 0.001, the first-order momentum decay coefficient to 0.9, the second-order momentum decay coefficient to 0.999, the gradient is clipped according to the global norm with a clipping threshold of 1.0, and the training epochs are set to 200. Training is stopped early when the validation set loss no longer decreases after 20 consecutive epochs, and the network parameters are rolled back to the values ​​at the lowest validation loss, resulting in the pre-trained prediction network. During the inference phase, the bioactivity evolution feature sequence of the previous 72 sampling times is pooled at the node dimension and input into the pre-trained prediction network. After forward propagation, the pooled features pass through a two-layer long short-term memory network at each time step. The hidden state at the last time step is mapped through a two-layer fully connected layer, and the predicted values ​​of the target substance degradation rate and biofilm thickness for the next 6 hours are output. The thickness prediction value is denormalized to recover the micrometer unit, and the degradation rate prediction value can be converted into a percentage for reading. For example, when the nitrite nitrogen concentration rises to the peak on the 19th day of biofilm culture, the prediction network outputs a predicted ammonia nitrogen degradation rate of 0.62 and a predicted biofilm thickness of 180 micrometers for the next 6 hours.

[0030] S4. Determine whether the predicted value of biofilm thickness is greater than or equal to the preset thickness threshold. If so, generate a virtual impact load tensor for the simulated substance release curve, and obtain the fluctuation curve of the target substance degradation rate based on the virtual impact load tensor.

[0031] In one specific embodiment, performing S4 includes the following steps: Determine whether the predicted value of biofilm thickness is greater than or equal to the preset thickness threshold. If so, generate a virtual impact load tensor containing the time-series perturbation data of the target substance concentration according to the simulated substance release curve. The virtual impact load tensor is superimposed onto the state environment parameter sequence to generate the disturbance state environment parameter sequence. Based on the perturbation state environmental parameter sequence, a perturbation spatiotemporal state attribute graph tensor is constructed, and the perturbation biological activity evolution feature sequence is extracted through graph convolution and temporal convolution. The perturbation bioactivity evolution feature sequence is input into a pre-trained prediction network to obtain the predicted values ​​of the target substance degradation rate for multiple consecutive future time steps, and the values ​​are arranged in chronological order to generate the fluctuation curve of the target substance degradation rate.

[0032] Specifically, the preset thickness threshold is the limit for determining whether a biofilm meets the conditions for undergoing impact testing, denoted as . The unit is micrometers, and the value is based on the biofilm thickness level on the packing surface corresponding to the one-time feed supplementation test after the water quality indicators have stabilized in the biofilm formation operation procedure. For example, 150 micrometers is used. When the biofilm thickness prediction value output by S3... Less than When the microbial community on the surface of the bio-ball packing material has not yet formed a stable film layer and lacks the structural basis to withstand impact, no disturbance is generated. The biofilm cultivation continues under the original conditions, and a new thickness prediction value is obtained in the next monitoring cycle for reassessment. Greater than or equal to At this point, the virtual impact testing process begins. The simulated substance release curve refers to the curve showing the change in concentration over time of nitrogenous organic matter in the feed releasing ammonia nitrogen into the water through ammonification when 200 ppm bass feed is soaked and then added to the isolation biochemical tank in accordance with the biofilm testing procedure. The nitrogen content of the bass feed, converted based on protein content, is approximately 7.2%, and the ammonification conversion rate is taken as 0.7. The peak increase in ammonia nitrogen concentration corresponding to a feed dosage of 200 mg / L is shown. The equivalent concentration is approximately 10 mg / L. The release process is described by a single-peak curve: Where n is the perturbation time step number, The sampling period for the state environment parameter sequence in S1 is 1 / 6 hour. This is the release rate constant, expressed in parts per hour; for example, it is taken as 0.05. The unit is milligrams per liter, and the curve reaches its peak at the 20th hour of the disturbance. The concentration then decays exponentially, consistent with the actual release pattern of ammonia nitrogen concentration rising and then falling after feed addition. The virtual impact load tensor, denoted as J, is obtained by mapping the release curve to all graph nodes according to the flow field spatial distribution. Its dimension is E×N, where E is the total number of perturbation time steps (288 corresponds to a 48-hour perturbation duration), and N is the number of graph nodes. The feed addition point is mapped to the corresponding graph nodes in the three-dimensional mesh. The element in the nth row and ith column of J is sorted by Calculation, where For normalized spatial adjacency weights, i.e. The first element of the spatial adjacency weight matrix A generated directly from S1 is taken directly from the first element of the spatial adjacency weight matrix A. The value in the i-th row is 0 to 1 and is dimensionless. The physical meaning of this mapping method is as follows: the ammonia nitrogen released at the dosing point is transported to each monitoring location along the dominant flow direction of the aeration tumbling flow field. The nodes with a stronger hydraulic correlation with the dosing point in the flow direction receive a higher load share, while nodes located in the flow dead zone or backflow dead angle receive only a lower load even if they are spatially close.

[0033] The process of superimposing the virtual impact load tensor onto the state environment parameter sequence is as follows: the measured state environment parameter sequence of the previous 71 sampling times is taken as the history segment, and the 6-dimensional environmental parameter values ​​measured by each sensor at the current time are taken as the steady-state baseline. The steady-state baseline is copied moment by moment along the future E time steps to obtain the extension segment. The total length of the history segment and the extension segment after splicing them in chronological order is 71+E. Then, the ammonia nitrogen increment of each graph node in the nth row of the virtual impact load tensor is superimposed onto the ammonia nitrogen concentration channel of the corresponding sensor at the nth time step of the extension segment according to the one-to-one correspondence between graph nodes and sensors. The five parameter channels of water temperature, dissolved oxygen, pH, nitrite nitrogen, and nitrate nitrogen maintain the steady-state baseline values ​​within the extension segment to generate the disturbed state environment parameter sequence. The data correspondence is as follows: the E values ​​of the release curve correspond to the E time steps of the extension segment, the spatial attenuation coefficients of the N graph nodes correspond to the ammonia nitrogen channels of the N sensors, and the 6 types of environmental parameters correspond to the 6 parameter channels of the sequence. Among them, only the ammonia nitrogen channel is disturbed and rewritten, while the other channels remain in a steady state. In this way, a feed shock is transformed into a sensor sequence form that isomorphic to the measured data, so that the subsequent processing link can be directly reused without modification.

[0034] The process of constructing a perturbation spatiotemporal state attribute graph tensor based on the perturbation state environment parameter sequence is as follows: Following the max-min normalization method in S1, the maximum and minimum values ​​used for normalization are taken from the statistical values ​​of historical membrane-forming data during the training phase and kept unchanged. After normalization of the six parameter channels, they are used as node attributes in dimensionless form. The N-order spatial adjacency weight matrix A generated in S1 is used as the edge weights between graph nodes, forming a perturbation node attribute tensor with dimensions (71+E)×N×6. For each time step of the perturbation node attribute tensor... The same two-layer graph convolution operation as S2 is performed. The graph convolution weight matrix takes the values ​​obtained after training in S2. After aggregating the neighborhood attributes layer by layer, it is concatenated with the dimensionless mass transfer gradient vector to obtain the 16-dimensional node features of each graph node at each sampling time. Then, using the temporal convolution kernel with a receptive field R of 72, a number of convolution kernels of 32, and causal padding from S2, convolution is performed along the time dimension. The temporal convolution weights also take the values ​​obtained after training in S2. The output is a perturbation biological activity evolution feature sequence with a dimension of (71+E)×N×32. Since the standardized parameters, adjacency weight matrix, and convolution weights are all consistent with the processing link under normal working conditions, the perturbation sequence and the normal sequence are mapped to the same feature space. The difference in the response of the prediction network to the perturbation input comes only from the load change of the ammonia nitrogen channel, thus ensuring that the fluctuation curve reflects the impact itself rather than the deviation of the processing link.

[0035] The process of inputting the perturbed biological activity evolution feature sequence into the pre-trained prediction network employs a rolling inference approach: for each future time step n after perturbation, a feature segment of 72 sampling times with that time as the endpoint is extracted from the perturbed biological activity evolution feature sequence. This segment is then processed by node-dimensional pooling to obtain a 72×32 feature matrix, which is input into the pre-trained prediction network. After forward propagation through a 2-layer Long Short-Term Memory network and a 2-layer fully connected layer, the output for the time step n is obtained. That is, the predicted degradation rate of the target substance over 6 hours. The process rolls from n=1 to n=E, obtaining a total of E predicted degradation rate values. These predicted values ​​are then arranged in chronological order according to their corresponding time steps to generate a fluctuation curve of the target substance's degradation rate. The horizontal axis of the curve represents the time step after the disturbance, and the vertical axis represents the predicted degradation rate value, thus comprehensively depicting the biofilm's response process under feed shock, from inhibited degradation rate, fluctuation with peak ammonia nitrogen concentration, to gradual recovery to steady state. For example, the fluctuation curve drops to a trough of 0.41 near the 8th hour of the disturbance, then rises above 0.6 and tends to stabilize at the 30th hour, indicating that the membrane layer has the ability to absorb shocks. The fluctuation curve is output to S5 to identify the starting point of the degradation rate deviating from steady state and the ending point of recovery to steady state, and to calculate the stress recovery time.

[0036] S5. Based on the fluctuation curve, determine the stress recovery time of the isolation biochemical chamber. When the stress recovery time is less than the preset time threshold, take the minimum material consumption and the shortest biofilm formation cycle as the optimization objectives, and generate a benchmark control vector containing the material addition pump flow rate and the opening degree of the fluid disturbance valve through reinforcement learning operators.

[0037] In one specific embodiment, performing S5 includes the following steps: Based on the fluctuation curve, the starting moment when the degradation rate of the target substance deviates from the steady state and the ending moment when it returns to the steady state are identified. The time difference between the starting moment and the ending moment is calculated to determine the stress recovery time of the isolation biochemical chamber. If the stress recovery time is less than the preset time threshold, then the optimization objectives are to minimize material consumption and shorten the biofilm formation period. A baseline control vector containing the material addition pump flow rate and the opening degree of the fluid disturbance valve is generated by reinforcement learning operators.

[0038] Specifically, the steady-state degradation rate refers to the baseline degradation rate of the target substance when the biofilm in the isolation biochemical chamber is in normal operating condition before the virtual shock occurs, denoted as The arithmetic mean of the predicted degradation rate values ​​for the 12 consecutive sampling times preceding the starting point of the fluctuation curve generated by S4 is taken as a dimensionless number between 0 and 1. The starting point of deviation from steady state is identified according to the following rule: the predicted degradation rate values ​​are compared point by point starting from the first time step of the fluctuation curve. and The difference will be the first one to satisfy | - The time steps exceeding the deviation limit are denoted as... The deviation limit is a dimensionless number, exemplarily taken as 0.05; the termination time of the return to steady state is identified according to the following rules: from Then, a search was performed to find that all 12 consecutive time steps met the condition. - For segments whose deviation from the limit is less than or equal to the limit, the first time step of that segment is recorded as... A 2-hour stability confirmation window is established for 12 consecutive time steps to avoid misjudgments caused by single-point fluctuations. The resilience and recovery time is calculated according to... The calculation shows that Δt is the sampling period, which is 1 / 6 of an hour, and the calculation result is in hours. For example, the fluctuation curve breaks through the steady-state zone at the 48th time step of the disturbance and returns to the steady-state zone at the 192nd time step, with a recovery time of 24 hours. The stress recovery time is compared with a preset time threshold. The preset time threshold is strictly set based on the experience of water quality recovery after a 200ppm feed shock in the biofilm operation procedure, with 24 hours as an example. When the stress recovery time is greater than or equal to the preset time threshold, the virtual stress resistance of the membrane is determined to be insufficient, no control vector is generated, the original culture conditions are maintained, and the prediction and shock test are re-executed in the next monitoring cycle. When the virtual stress recovery time is less than the preset time threshold, it indicates that the membrane structure has passed the initial screening of virtual shock, triggering a small-dose physical disturbance test: 50ppm of bass feed simulation solution is added to the isolation biochemical chamber at one time, and sensor data is collected continuously for 48 hours after the disturbance. The measured stress recovery time is calculated according to the same steady-state deviation identification rule in S5. If the measured stress recovery time is still less than the preset time threshold, the membrane stress resistance is officially determined to be qualified, and the reinforcement learning control vector generation stage is entered. If the measured recovery time is unqualified, the current batch is abandoned, and the biofilm culture and prediction verification are re-executed in S1. The virtual test result should not be used as the basis for jumping.

[0039] The reinforcement learning operator selected is the deep deterministic policy gradient algorithm, consisting of a policy network and a value network, suitable for control problems where both pump flow rate and valve opening are continuous variables. The state vector, denoted as x, has a dimension of 34 and is composed of a 32-dimensional global feature vector obtained by pooling the current bioactivity evolution feature sequence through node dimension, the predicted degradation rate of the target substance at the current time, and the predicted biofilm thickness. The action vector, denoted as y, has a dimension of 2, with the first component being... The pumping flow rate for material addition, i.e., the delivery flow rate of the special carbon source mother liquor metering pump, is expressed in liters per hour, with a value ranging from 0 to 50. (Second component) This represents the valve opening degree affected by fluid disturbance, expressed as a percentage, ranging from 0 to 100. The reward value is denoted as b, and is calculated according to... Calculate: Consumption Cost Item ,in This refers to the concentration of the carbon source mother liquor, expressed in milligrams per liter. The mass of carbon source added in a single step is expressed in milligrams. For reference consumption, the maximum allowable daily dosage is used, in milligrams. A dimensionless number; periodic cost term Λ represents the estimated remaining biofilm attachment time in hours, calculated from the current predicted biofilm thickness reaching a preset thickness threshold. The estimated time required is as follows: Dimensionless number; penalty term for exceeding the limit The value is set to 1 if the predicted degradation rate is below 0.3 or the predicted biofilm thickness decreases after the action is performed; otherwise, it is set to 0. Take 0.6, A value of 0.4 reflects the prioritization of low-carbon goals over periodic goals. The data correspondences are as follows: the 32-dimensional pooling feature corresponds to the 34 components of the state vector in the 2-dimensional prediction value; the control quantities of the two actuators correspond to the two components of the action vector; and the material consumption and biofilm formation cycle, the two optimization goals, correspond to the reward values. and Two items.

[0040] The structure and training process of the policy network and value network are as follows: The policy network consists of three fully connected layers. The input layer is 34-dimensional, and each of the two hidden layers contains 128 neurons and is activated by ReLU. The output layer is 2-dimensional and linearly mapped to the action value range after hyperbolic tangent activation. The value network takes a 36-dimensional vector as input, which is a concatenation of the state vector and the action vector. After passing through two hidden layers, each containing 128 neurons, it outputs a scalar state-action value Q(x,y). Training uses a pre-trained prediction network as the environment model. The entire training process is performed in an offline virtual environment without interaction with the real isolated biochemical chamber. Each action is executed, and the action remains constant within that step. The state-environment parameter sequence is advanced according to the following rule: the carbon source concentration increment is converted from the pump flow rate. The material balance state variable G, accumulated within the environmental model, does not participate in the construction of node attributes of the spatiotemporal state attribute graph tensor. It only serves as an independent scalar in determining the carbon-nitrogen ratio constraint and calculating the reward function. The initial value of G is determined by the baseline value in the silo before the first addition, and is dynamically updated at each subsequent time step according to the pumping increment and microbial consumption, thereby avoiding direct dependence on the organic carbon sensor. The isolation biochemical chamber has a volume of 105,000 liters; the dissolved oxygen increment is calculated based on the valve opening degree. The degradation rate is superimposed on the dissolved oxygen channel; the ammonia nitrogen channel decays according to the degradation rate prediction value output by the prediction network, that is, the ammonia nitrogen concentration at the end of this step is the product of the current ammonia nitrogen concentration and (1-η), where η is the degradation rate prediction value of the current step. The ammonia nitrogen concentration at each sampling time within the step transitions from the current value to the end value according to an exponential law; the nitrite nitrogen channel and the nitrate nitrogen channel are converted according to the nitrogen conservation relationship: the net accumulation of the ζ share in the ammonia nitrogen degradation in this step is the nitrite nitrogen increment, where ζ is the dimensionless net accumulation coefficient of nitrite nitrogen, which is exemplarily taken as 0.2; the carbon source driven denitrification removal is taken as μ×ΔG, where μ is the dimensionless denitrification stoichiometric coefficient, which is exemplarily taken as 0.3; the nitrate nitrogen increment is the difference between the (1-ζ) share of ammonia nitrogen degradation and the denitrification removal, and the difference is taken as 0 when it is negative; the water temperature channel and the pH channel maintain the current measured baseline value during the advancement, which is consistent with the treatment method of the extension section in S4. After updating the state environment parameter sequence, the features are reconstructed and input into the prediction network to obtain the predicted degradation rate and film thickness for the next time step. Based on this, the reward value is calculated and the state is advanced. Each step corresponds to Δφ, which is a virtual advancement of 6 hours. Training adopts an experience replay mechanism with a replay buffer capacity of 100,000 records, a batch size of 256, a policy network learning rate of 0.0003, a value network learning rate of 0.001, a discount factor of 0.95, and a target network that tracks the online network using a soft update method with a soft update coefficient of 0.005. The buffer is filled with random actions for the first 5,000 steps. During training, truncated Gaussian exploration noise with a standard deviation of 0.1 is superimposed on the output of the policy network. The number of training rounds is 5,000, with 30 steps per round. Training stops when the average reward fluctuation of the most recent 100 rounds is less than 5%.

[0041] During the inference phase, the current state vector is input into the trained policy network, and a 2D action vector is output through forward propagation. After mapping, a baseline control vector containing the material addition pump flow rate and the valve opening for fluid disturbance is obtained. For example, after passing the virtual impact test on the 22nd day of biofilm formation, the policy network outputs a pump flow rate of 12.5 liters per hour and a valve opening of 65%.

[0042] S6. Based on the preset expert rule base, the baseline control vector is boundary-limited to generate a safety control command. The dosing pump and fluid disturbance device are driven according to the safety control command. The residual between the actual degradation rate and the predicted value of the target substance degradation rate is calculated. If the residual is less than the preset residual threshold for N consecutive monitoring cycles, the biofilm is determined to be successful and a valve switching command to connect the target waterway is output.

[0043] In one specific embodiment, performing S6 includes the following steps: The preset expert rule base is invoked to perform boundary limiting processing on the material dosing pump flow rate and fluid disturbance valve opening parameters in the baseline control vector, and generate safety control commands. Drive the dosing pump and fluid disturbance device according to the safety control command, collect the actual environmental parameters in the isolation biochemical chamber, and calculate the actual degradation rate based on the actual environmental parameters. Calculate the residual between the actual degradation rate and the predicted degradation rate of the target substance, and determine whether the residual is less than the preset residual threshold for N consecutive monitoring cycles. If so, the biofilm formation is considered successful, and a valve switching command to connect the target water path is output, where N is a positive integer greater than 3.

[0044] Specifically, the expert rule base refers to a set of rules solidified from biofilm formation operation procedures and biochemical reaction safety boundaries. Rules are stored in pairs of conditions and limits. The condition portion references the water temperature, dissolved oxygen concentration, and ammonia nitrogen concentration measured by the sensor array within the isolated biochemical chamber, as well as the degradable organic carbon concentration calculated from the special carbon source addition record. The limit portion specifies the upper and lower limits allowed for the material addition pump flow rate and the fluid disturbance valve opening under the current operating conditions. Boundary limiting refers to the calculation of compressing each component of the baseline control vector to the corresponding upper and lower limit intervals. For the material addition pump flow rate, the expert rule base infers the upper limit based on the carbon-to-nitrogen ratio limiting metabolic boundary parameter: the current carbon-to-nitrogen ratio is calculated by the degradable organic carbon concentration G and the measured ammonia nitrogen concentration S within the chamber. To prevent heterotrophic bacteria from suppressing autotrophic nitrifying bacteria in the competition for dissolved oxygen and the space for attachment to the packing material, the carbon-to-nitrogen ratio after this feeding cycle must not exceed the upper limit of the carbon-to-nitrogen ratio. That is, 6, the upper limit of pumping flow rate is derived from the allowable additional organic carbon mass. ,in The isolation biochemical chamber has a volume of 105,000 liters. For special carbon source mother liquor concentration, For a monitoring cycle of 6 hours, molecules The allowable additional organic carbon mass is expressed in milligrams, with the denominator being... The unit of measurement is milligrams per liter multiplied by hours, and the unit of measurement of the quotient is liters per hour, consistent with the pump flow rate; lower limit of pump flow rate. A value of 0 indicates that when water quality indicators are close to the target, it is permissible to suspend feed to reduce carbon consumption. Regarding the valve opening for fluid disturbances, the rule is based on the dissolved oxygen safety limit: to suppress excessive shear force near the aeration port from eroding the still-developing initial biofilm, the dissolved oxygen concentration in the tank must not exceed the dissolved oxygen safety limit. That is, 8 mg / L. Using a conversion factor of 0.04 mg / L per percentage point between valve opening and dissolved oxygen increment, the upper limit of valve opening can be calculated. Molecular weight is expressed in milligrams per liter, and the quotient is expressed in percentage points; lower limit of aperture. A 20% threshold is used to maintain the minimum opening degree required for the bio-ball packing material to tumble up and down with the water flow. The amplitude limit calculation is based on... Execution, valve opening component Similarly, the rule base also includes rate of change limits, meaning that the change in pumping flow rate between two adjacent monitoring cycles must not exceed 5 liters per hour, and the change in valve opening must not exceed 10 percentage points. If these limits are exceeded, the limit value is used to prevent sudden shear changes in the water body caused by abrupt actions of the actuator. The data correspondence is as follows: the two components of the baseline control vector correspond to two limiting channels; the measured water temperature, dissolved oxygen, ammonia nitrogen, and organic carbon data correspond to the conditional inputs of the rules; and the output is a safety control command with the same dimension as the baseline control vector. .

[0045] Safety control commands are issued to the dosing pump and fluid disturbance device for execution. The dosing pump is a metering pump that delivers the special carbon source mother liquor. Operating at a flow rate of [number] liters per hour, the fluid disturbance device, i.e., the regulating valve group on the aeration branch, is [operated according to...]. The aeration intensity was adjusted by a percentage point. During operation, the sensor array continuously collected actual environmental parameters within the isolated biological chamber at a 10-minute sampling cycle, including six types of measured data: water temperature, dissolved oxygen concentration, pH value, ammonia nitrogen concentration, nitrite nitrogen concentration, and nitrate nitrogen concentration. At the end of each monitoring cycle, the measured ammonia nitrogen concentration at the start of the cycle was recorded. Measured concentration at the end of the cycle ,according to Calculate the actual degradation rate The numerator and denominator are both in milligrams per liter, and the ratio is a dimensionless number ranging from 0 to 1. Within the same monitoring period, the environmental parameter sequence is synchronously updated with the bioactivity evolution feature sequence via graph convolution and temporal convolution, and then input into a pre-trained prediction network to obtain the predicted degradation rate of the target substance corresponding to that monitoring period. The actual degradation rate and the predicted value are strictly aligned on the time axis, that is, they correspond to the same start and end time and the same predicted time span, ensuring that the physical meaning of subsequent residual comparisons is consistent.

[0046] Residual refers to the absolute measure of the deviation of the actual degradation rate from the predicted value. The residual for the t-th monitoring period is calculated as follows: The calculation yields a dimensionless number. The residual is then compared to a preset residual threshold. For comparison, a value of 0.08 is chosen, which is greater than the allowable deviation of 0.05 for the kinetic penalty term, to accommodate the flow error of the metering pump and the measurement noise of the ammonia nitrogen sensor; when Less than When this occurs, it indicates that the prediction network's characterization of the effects of drug administration and perturbation is still reliable, and the continuous achievement count is incremented by 1; when Greater than or equal to When the consecutive compliance count is zeroed, the current safety control command continues to operate and the system enters the next monitoring cycle for recalibration. When the residual is less than [a certain value] for N consecutive monitoring cycles... Successful biofilm formation is determined by the timing of N, a positive integer greater than 3, exemplified as 4, corresponding to a stable verification window of 24 consecutive hours. This aligns with the experience in the biofilm formation operation procedure where water quality returns to normal within 2 to 3 days after a single addition of 200 ppm bass feed. Upon successful biofilm formation, a valve switching command is output to connect to the target waterway. The target waterway refers to the circulating waterway between the isolated biological tank and the aquaculture pond. The command drives the electric valve assembly to close the independent inlet bypass during the biofilm formation stage and open the circulating inlet and outlet valves from the isolated biological tank to the aquaculture pond. This integrates the biofilm-formed biological tank into the main recirculating aquaculture loop, corresponding to the operation in the biofilm formation operation procedure where ammonia nitrogen, nitrite nitrogen, and nitrate nitrogen levels recover and are then recirculated into the aquaculture pond.

[0047] Please see Figure 2 , Figure 2 The graph shows the comparison of cumulative carbon source consumption. The horizontal axis represents biofilm formation time, and the vertical axis represents cumulative carbon source consumption. The blue curve represents the method of this application, and the red curve represents the traditional method. The graph shows that during the entire 30-day biofilm formation cycle, the traditional method's carbon source consumption consistently increases and fluctuates greatly, accumulating to nearly 60g at 30 days. In contrast, the carbon source consumption of the method of this application increases gradually, with a peak of only about 32g. The total amount of carbon source added is significantly reduced for the same biofilm formation time. This graph verifies the advantages of the method of this application in low-carbon regulation by optimizing the carbon source pumping flow rate through reinforcement learning and matching the actual metabolic needs of microorganisms. It solves the pain point of blindly overdosing carbon sources in traditional processes, reducing aquaculture operating costs and water organic load.

[0048] Please see Figure 3 , Figure 3 The figure shows the comparison of biofilm formation cycles in 20 independent experiments. The horizontal axis is divided into two groups: the method of this application and the traditional method, and the vertical axis represents the biofilm formation completion cycle. The figure shows that the biofilm formation cycle of the method of this application is concentrated in the range of 18 to 22 days, with small data dispersion and stable and controllable cycle. The biofilm formation cycle of the traditional method is distributed in a wide range of 24 to 40 days, with a median of more than 30 days and some extremely long batches. This figure illustrates that the traditional method ignores the spatial heterogeneity of the flow field in the biochemical chamber and cannot predict the growth state of microorganisms, resulting in a long biofilm formation cycle and large batch differences. In contrast, the method of this application captures the spatiotemporal coupling law between the flow field and microorganisms through spatiotemporal graph convolution and dynamically predicts the biofilm maturation progress, thereby compressing the overall biofilm formation time and reducing batch fluctuations.

[0049] Please see Figure 4 The following describes a system for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture systems, as described in the embodiments of this application. This system includes: The tensor construction module is used to obtain the state environment parameter sequence collected by the sensor array in the isolation biochemical chamber, determine the spatial adjacency weight of each sensor based on the three-dimensional mesh of the aeration tumbling flow field, and construct the spatiotemporal state attribute graph tensor by combining the state environment parameter sequence. The feature extraction module is used to obtain the reproductive cycle parameters of the target microorganism and configure the temporal convolution kernel. The configured temporal convolution kernel is used to perform graph convolution and temporal convolution operations on the spatiotemporal state attribute graph tensor to extract the biological activity evolution feature sequence. The prediction output module is used to input the bioactivity evolution feature sequence into the pre-trained prediction network and output the predicted values ​​of the target substance degradation rate and biofilm thickness at future time steps. The loss function of the prediction network includes a penalty term constructed based on the standard biochemical kinetic equation and the carbon-nitrogen ratio limiting metabolic boundary parameter. The impact test module is used to determine whether the predicted value of biofilm thickness is greater than or equal to the preset thickness threshold. If so, it generates a virtual impact load tensor of the simulated substance release curve and obtains the fluctuation curve of the target substance degradation rate based on the virtual impact load tensor. The baseline generation module is used to determine the resilience recovery time of the isolation biochemical chamber based on the fluctuation curve. When the resilience recovery time is less than the preset time threshold, the module takes the minimum material consumption and the shortest biofilm formation cycle as the optimization objectives and generates a baseline control vector containing the material addition pump flow rate and the opening of the fluid disturbance valve through reinforcement learning operators. The safety execution module is used to generate safety control commands by limiting the baseline control vector according to the preset expert rule base. The module drives the dosing pump and fluid disturbance device according to the safety control commands, calculates the residual between the actual degradation rate and the predicted value of the target substance degradation rate, and if the residual is less than the preset residual threshold for N consecutive monitoring cycles, the module determines that biofilm formation is successful and outputs a valve switching command to connect the target water circuit.

[0050] Through the collaborative efforts of the aforementioned components, the system achieves integrated collaborative control of the entire biofilm attachment process, including intelligent sensing, dynamic simulation, resilience verification, optimization decision-making, and safe execution. The tensor construction module is responsible for organically integrating the multidimensional water quality parameters collected by the sensor array with the three-dimensional spatial distribution of the aeration tumbling flow field. It extracts the mass transfer correlations between monitoring nodes through fluid dynamics simulation, constructing a graph tensor carrying spatiotemporal coupling information, providing an input basis with both physical consistency and spatial resolution for subsequent feature extraction. The feature extraction module configures temporal convolution kernels based on the intrinsic reproductive rhythm of the target microorganisms. Through cascaded operations of graph convolution and temporal convolution, it captures the evolutionary patterns of biochemical reactions on a generational timescale while aggregating mass transfer information from the spatial neighborhood, elevating the original monitoring data into a bioactivity evolution feature sequence containing the dynamics of microbial activity. The prediction output module receives the feature sequence and propagates it forward through a prediction network embedded with prior knowledge of biochemical dynamics, outputting the predicted degradation rate and film thickness for future time steps. The kinetic and carbon-nitrogen ratio boundary penalty terms introduced in its loss function ensure that the predicted trajectory does not deviate from the intrinsic physical limits of the biochemical reactions, providing a scientifically reliable quantitative basis for subsequent evaluation and decision-making. The impact testing module actively generates simulated high-load impacts after the predicted film thickness meets the standard. A virtual perturbation tensor drives the same prediction link to continuously extrapolate the degradation rate response fluctuation curve under impact, quantitatively characterizing the stress recovery time of the biofilm system. This virtual online testing method avoids actual impact risks and fills the technical blind spot of unpredictable stability during biofilm formation. After the stress resistance capacity passes the preset threshold, the benchmark generation module uses material consumption and biofilm formation cycle as dual optimization objectives. It uses deep reinforcement learning operators to search for the optimal control strategy in the continuous action space, generating a benchmark vector that matches the pump flow rate and perturbation opening degree to match the current biochemical state. This achieves a fine match between carbon source addition and aeration intensity, reducing redundant carbon consumption and shortening biofilm formation time from the source. The safety execution module applies physical boundary limits to the benchmark vector based on the expert rule base, generating safe and executable control commands to drive the dosing pump and fluid perturbation device. It verifies the control effect by calculating the continuous residual between the actual degradation rate and the predicted value. When the residual continuously meets the standard, it indicates that the system is operating within the predicted trajectory. Finally, it determines that biofilm formation is successful and outputs a water path switching command, thus connecting the intelligent decision-making to on-site execution and the execution effect to system evaluation at the end link.

[0051] This application also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the method for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture.

[0052] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the methods and systems described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0053] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture, characterized in that, include: S1. Obtain the state environment parameter sequence collected by the sensor array in the isolation biochemical chamber, determine the spatial adjacency weight of each sensor based on the three-dimensional mesh of the aeration tumbling flow field, and construct the spatiotemporal state attribute graph tensor in combination with the state environment parameter sequence. S2. Obtain the reproductive cycle parameters of the target microorganism and configure the temporal convolution kernel. Use the configured temporal convolution kernel to perform graph convolution and temporal convolution operations on the spatiotemporal state attribute graph tensor to extract the biological activity evolution feature sequence. S3. Input the bioactivity evolution feature sequence into a pre-trained prediction network and output the predicted value of the target substance degradation rate and the predicted value of the biofilm thickness at future time steps. The loss function of the prediction network includes a penalty term constructed based on the standard biochemical kinetic equation and the carbon-nitrogen ratio limiting metabolic boundary parameter. S4. Determine whether the predicted value of the biofilm thickness is greater than or equal to the preset thickness threshold. If so, generate a virtual impact load tensor of the simulated substance release curve, and obtain the fluctuation curve of the target substance degradation rate based on the virtual impact load tensor. S5. Based on the fluctuation curve, determine the resilience recovery time of the isolation biochemical chamber. When the resilience recovery time is less than the preset time threshold, take the minimum material consumption and the shortest biofilm formation cycle as the optimization objectives, and generate a benchmark control vector containing the material addition pump flow rate and the opening degree of the fluid disturbance valve through reinforcement learning operators. S6. Based on the preset expert rule base, the baseline control vector is subjected to boundary limiting to generate a safety control command. The dosing pump and fluid disturbance device are driven according to the safety control command. The residual between the actual degradation rate and the predicted value of the target substance degradation rate is calculated. If the residual is less than the preset residual threshold for N consecutive monitoring cycles, the biofilm formation is determined to be successful and a valve switching command to connect the target waterway is output.

2. The method according to claim 1, characterized in that, S1 includes: The sequence of state and environmental parameters collected by the sensor array in the isolation biochemical chamber is obtained. The velocity field distribution is obtained by performing fluid dynamics simulation on the aeration and tumbling flow field in the isolation biochemical chamber. The physical position of each sensor is mapped to a three-dimensional mesh to obtain the corresponding node of the sensor. Based on the velocity field distribution, the dominant streamline direction of each sensor corresponding node is extracted. Based on the angle between the line connecting each pair of corresponding sensor nodes and the dominant streamline direction and the spatial distance between the corresponding nodes, the spatial adjacency weight between each sensor is calculated, and a spatial adjacency weight matrix is ​​generated. Using the monitoring locations of each sensor as graph nodes, the multidimensional environmental parameters corresponding to each time step in the state-environment parameter sequence as node attributes, and the corresponding elements of the spatial adjacency weight matrix as edge weights between graph nodes, a spatiotemporal state attribute graph tensor is constructed.

3. The method according to claim 1, characterized in that, S2 includes: Obtain the reproductive cycle parameters of the target microorganism that dominates denitrification during biofilm formation, and configure the receptive field size of the temporal convolution kernel according to the reproductive cycle parameters; A graph convolution operation is performed on the spatiotemporal state attribute graph tensor. The neighborhood attributes of each node are aggregated according to the edge weights of the graph structure. The mass transfer gradient between nodes is extracted to obtain a node feature sequence containing spatial mass transfer information. The node feature sequence is subjected to temporal convolution operation using the configured temporal convolution kernel to extract the evolutionary pattern of biological activity over time and generate a biological activity evolution feature sequence.

4. The method according to claim 1, characterized in that, S3 includes: Obtain the standard biochemical kinetic equation and the carbon-nitrogen ratio limiting metabolic boundary parameter, and construct a penalty term based on the standard biochemical kinetic equation and the carbon-nitrogen ratio limiting metabolic boundary parameter; The penalty term is superimposed on the loss function of the prediction network, and the network parameters of the prediction network are jointly optimized by backpropagation algorithm to obtain the pre-trained prediction network. The bioactivity evolution feature sequence is input into a pre-trained prediction network, and the predicted values ​​of the target material degradation rate and biofilm thickness at future time steps are output through forward propagation.

5. The method according to claim 1, characterized in that, S4 includes: Determine whether the predicted value of the biofilm thickness is greater than or equal to a preset thickness threshold. If so, generate a virtual impact load tensor containing the time-series perturbation data of the target substance concentration according to the simulated substance release curve. The virtual impact load tensor is superimposed onto the state environment parameter sequence to generate a disturbance state environment parameter sequence. Based on the perturbation state environment parameter sequence, a perturbation spatiotemporal state attribute graph tensor is constructed, and the perturbation biological activity evolution feature sequence is extracted by graph convolution and temporal convolution. The perturbation bioactivity evolution feature sequence is input into a pre-trained prediction network to obtain the predicted values ​​of the target substance degradation rate for multiple consecutive future time steps, and the predicted values ​​are arranged in chronological order to generate a fluctuation curve of the target substance degradation rate.

6. The method according to claim 1, characterized in that, S5 includes: Based on the fluctuation curve, the starting moment when the degradation rate of the target substance deviates from the steady state and the ending moment when it returns to the steady state are identified. The time difference between the starting moment and the ending moment is calculated to determine the stress recovery time of the isolation biochemical chamber. If the stress recovery time is less than a preset time threshold, then the optimization objectives are to minimize material consumption and shorten the biofilm formation period. A baseline control vector containing the material addition pump flow rate and the opening degree of the fluid disturbance valve is generated by reinforcement learning operators.

7. The method according to claim 1, characterized in that, S6 includes: The preset expert rule base is invoked to perform boundary limiting processing on the material dosing pump flow rate and fluid disturbance valve opening parameters in the benchmark control vector, and generate safety control commands. The dosing pump and fluid disturbance device are driven to operate according to the safety control command, and the actual environmental parameters in the isolation biochemical chamber are collected. The actual degradation rate is calculated based on the actual environmental parameters. Calculate the residual between the actual degradation rate and the predicted degradation rate of the target substance, and determine whether the residual is less than the preset residual threshold for N consecutive monitoring cycles. If so, determine that biofilm formation is successful and output a valve switching command to connect the target waterway, where N is a positive integer greater than 3.

8. A system for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture, used to implement the method for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture as described in any one of claims 1 to 7, characterized in that, include: The tensor construction module is used to obtain the state environment parameter sequence collected by the sensor array in the isolation biochemical chamber, determine the spatial adjacency weight of each sensor based on the three-dimensional mesh of the aeration tumbling flow field, and construct the spatiotemporal state attribute graph tensor in combination with the state environment parameter sequence. The feature extraction module is used to obtain the reproductive cycle parameters of the target microorganism and configure the temporal convolution kernel. The configured temporal convolution kernel is used to perform graph convolution and temporal convolution operations on the spatiotemporal state attribute graph tensor to extract the biological activity evolution feature sequence. The prediction output module is used to input the bioactivity evolution feature sequence into a pre-trained prediction network and output the predicted values ​​of the target substance degradation rate and biofilm thickness at future time steps. The loss function of the prediction network includes a penalty term constructed based on the standard biochemical kinetic equation and the carbon-nitrogen ratio limiting metabolic boundary parameter. The impact test module is used to determine whether the predicted value of the biofilm thickness is greater than or equal to a preset thickness threshold. If so, a virtual impact load tensor of the simulated substance release curve is generated, and the fluctuation curve of the target substance degradation rate is obtained based on the virtual impact load tensor. The benchmark generation module is used to determine the resilience recovery time of the isolation biochemical chamber based on the fluctuation curve. When the resilience recovery time is less than a preset time threshold, the module generates a benchmark control vector containing the material addition pump flow rate and the opening degree of the fluid disturbance valve by using reinforcement learning operators with the optimization objectives of minimizing material consumption and minimizing biofilm formation period. The safety execution module is used to generate safety control instructions by limiting the baseline control vector according to the preset expert rule base, drive the dosing pump and fluid disturbance device according to the safety control instructions, calculate the residual between the actual degradation rate and the predicted value of the target substance degradation rate, and if the residual is less than the preset residual threshold for N consecutive monitoring cycles, the biofilm is determined to be successful and a valve switching instruction to connect the target water channel is output.

9. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by the processor, it implements a method for low-carbon denitrification and efficient biofilm formation in factory-scale recirculating aquaculture as described in any one of claims 1 to 7.

Images