Real-time monitoring and early warning method for microbial pollution risk of primary pulp production line

Abstract

The invention provides a puree production line microbial pollution risk real-time monitoring and early warning method, which comprises the following steps: deploying multiple types of sensors, an industrial camera and an operation log interface, collecting environmental temperature and humidity, pH value, dissolved oxygen, image and operation behavior multi-modal data in a fermentation tank in real time, and carrying out edge calculation and standardized preprocessing to obtain the microbial pollution risk of a puree production line. Data missing, abnormity and time sequence difference are eliminated; multi-modal features and a dynamic knowledge graph constructed based on a production process and a pollution event are fused, and a deep learning model and a graph neural network are utilized to realize microbial pollution risk probability intelligent prediction; risk grading and a dynamic weighting algorithm are introduced, grading early warning signals and disposal suggestions are automatically generated, and a production control system is linked to execute response measures; according to the subsequent production state, continuous feedback is carried out, the model and the early warning threshold are adjusted in a self-adaptive mode, and the accuracy, response efficiency and production safety of pollution early warning are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of microbial contamination risk monitoring technology in pulp production process, and in particular to a method for real-time monitoring and early warning of microbial contamination risk in pulp production line. Background Technology

[0002] The field of microbial contamination risk monitoring and early warning technology in pulp production has gradually evolved from traditional manual detection and single-sensor monitoring to intelligent and diversified approaches. With the application of emerging technologies such as the Industrial Internet of Things, machine vision, and artificial intelligence, industrial fermentation and pulp production enterprises have generally added online environmental sensors, automated equipment, and information management systems to achieve automatic collection of production process parameters and preliminary risk identification. Some leading companies have introduced data-driven methods based on deep learning, using time-series data of environmental physical quantities such as temperature, humidity, and pH values, as well as fermentation process images, to assist in determining microbial contamination events. However, these methods are limited by their insufficient ability to express prior knowledge of complex production processes, and the generalization and interpretability of the models urgently need improvement.

[0003] From a technological perspective, mainstream solutions focus on two directions: First, automated systems primarily for environmental measurement and data visualization monitoring, which can achieve real-time acquisition of multi-channel parameters, but are limited by single-point threshold warnings and struggle to handle nonlinear anomaly detection caused by multi-source data fusion; Second, big data analysis platforms based on deep learning models, including convolutional neural networks and recurrent neural networks, which can be used for environmental anomaly identification and pollution trend prediction, but they are highly dependent on data-driven approaches and lack logical constraints from process flow and microbiological knowledge, resulting in insufficient model interpretability, a high false alarm rate, and limited responsiveness to new production models or rare pollution events.

[0004] Some existing research and patents have attempted to incorporate knowledge bases, process standard operating procedures (SOPs), and empirical rules into risk monitoring. However, these primarily rely on static rules or fixed graphs, offering limited adaptability to dynamic evolution of production processes and multimodal fusion prediction. For typical application scenarios, such as pollution risk monitoring in pulp fermentation processes, current technology cannot yet perform in-depth correlation analysis between environmental data and operational behaviors. Knowledge structure updates lag behind on-site anomalies, leading to delayed early warning system responses and frequent false alarms or missed detections. This is particularly pronounced in large-scale, continuous production lines, impacting enterprise cost control and product quality and safety. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, the present invention provides a method for real-time monitoring and early warning of microbial contamination risk in pulp production lines.

[0006] The technical solution of this invention is implemented as follows: A method for real-time monitoring and early warning of microbial contamination risk in a pulp production line, comprising: S1: Collect multimodal real-time data on the temperature, humidity, pH value, dissolved oxygen concentration, image data, and operation logs of the fermentation tank in the pulp production line, and record the data collection timestamp and equipment identification information; S2: Perform missing value imputation, outlier removal and time alignment preprocessing on the collected multimodal data to eliminate timing deviations caused by differences in sampling frequencies of different sensors; S3: Construct a dynamic knowledge graph covering the pulp production process, microbial growth characteristics, and historical contamination events, and design an incremental update mechanism based on graph neural networks to achieve continuous optimization and expansion of the knowledge graph; S4: Input the preprocessed multimodal data into the multimodal fusion deep learning model, and combine it with the structured semantic information provided by the knowledge graph to generate fusion feature representations to enhance the model's ability to explain pollution risks; S5: Based on the fusion of feature representation and prior rules in the knowledge graph, calculate the probability of microbial contamination risk under the current working conditions, and feed it back to the knowledge graph through a two-way interaction mechanism to correct potential erroneous associations; S6: Based on the set pollution risk level threshold, determine whether the current risk probability exceeds the preset threshold. If it does, generate a pollution warning signal along with the risk level and suggested treatment measures. S7: Transmits the pollution warning signal to the control system and executes warning response actions in conjunction with it, including but not limited to suspending feeding, starting the sterilization device, and recording the abnormal log. S8: After the early warning response is executed, continuously monitor the production line status data and dynamically adjust the early warning threshold and model parameters according to subsequent data changes to improve the accuracy and adaptability of subsequent predictions.

[0007] The present invention provides a method for real-time monitoring and early warning of microbial contamination risk in a pulp production line, which has the following beneficial effects: (1) This invention innovatively integrates dynamically updated knowledge graphs and multimodal deep learning reasoning, achieving efficient collaboration between structured knowledge and data-driven prediction. By introducing a dynamic knowledge graph update mechanism based on graph neural networks, the graph structure and relationships can be automatically supplemented and corrected based on multi-source data such as temperature and humidity, pH value, dissolved oxygen, images, and operation logs collected in real time at the production site, significantly enhancing the timeliness and adaptability of knowledge modeling. Compared with traditional static knowledge bases or rule systems, this mechanism can quickly adapt to new pollution patterns and complex process changes, avoid misjudgments caused by knowledge lag, and ensure that risk judgment is always based on the latest and most relevant information; (2) This invention fully leverages the complementary advantages of multi-source heterogeneous data through unified preprocessing and integration of multimodal data. By employing techniques such as missing value imputation, outlier removal, time synchronization, and normalization, it effectively solves problems such as inconsistent sensor frequencies and varying data quality, thereby improving the consistency of model input and data reliability. High dynamic range image processing enhances the visibility of microbial products, and combined with text log structuring and encoding, it greatly enriches the observable features of pollution states, providing solid support for subsequent feature learning and risk identification of deep neural networks. (3) This invention introduces knowledge-aware multimodal deep fusion modeling technology, employs graph attention mechanism and gated recurrent units, and dynamically injects the knowledge graph structure weights into the model feature space, enabling the model to automatically focus on the process nodes and microbial associations that contribute the most to pollution risk prediction. This approach not only significantly enhances the model's generalization ability and contextual semantic interpretability, but also effectively alleviates the "black box" problem of traditional deep learning models, making the model's decision path more interpretable; (4) This invention features a unique two-way interaction mechanism between the model and the knowledge graph, which can dynamically adjust the graph weights in real time based on inference errors and confidence levels. It utilizes a model feedback loop to enhance the knowledge graph's representation and reasoning capabilities, continuously optimizes the structure, and strengthens critical paths, overcoming the problems of "blind spot accumulation" in existing inference models and "rigidity" in knowledge graphs. The feedback enhancement and incremental reasoning methods effectively reduce the false positive and false negative rates, and significantly improve overall robustness and adaptability. Attached Figure Description

[0008] Figure 1 This is a flowchart of a method for real-time monitoring and early warning of microbial contamination risk in a pulp production line according to the present invention; Figure 2 This is a sub-flowchart of a method for real-time monitoring and early warning of microbial contamination risk in a pulp production line according to the present invention; Figure 3 This is another sub-flowchart of the method for real-time monitoring and early warning of microbial contamination risk in a pulp production line according to the present invention. Detailed Implementation

[0009] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0010] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0011] like Figure 1 As shown, this application provides a method for real-time monitoring and early warning of microbial contamination risk in a pulp production line, specifically including: S1: Collect multimodal real-time data on the temperature, humidity, pH value, dissolved oxygen concentration, image data, and operation logs of the fermentation tank in the pulp production line, and record the data collection timestamp and equipment identification information; S2: Perform missing value imputation, outlier removal and time alignment preprocessing on the collected multimodal data to eliminate timing deviations caused by differences in sampling frequencies of different sensors; S3: Construct a dynamic knowledge graph covering the pulp production process, microbial growth characteristics, and historical contamination events, and design an incremental update mechanism based on graph neural networks to achieve continuous optimization and expansion of the knowledge graph; S4: Input the preprocessed multimodal data into the multimodal fusion deep learning model, and combine it with the structured semantic information provided by the knowledge graph to generate fusion feature representations to enhance the model's ability to explain pollution risks; S5: Based on the fusion of feature representation and prior rules in the knowledge graph, calculate the probability of microbial contamination risk under the current working conditions, and feed it back to the knowledge graph through a two-way interaction mechanism to correct potential erroneous associations; S6: Based on the set pollution risk level threshold, determine whether the current risk probability exceeds the preset threshold. If it does, generate a pollution warning signal along with the risk level and suggested treatment measures. S7: Transmits the pollution warning signal to the control system and executes warning response actions in conjunction with it, including but not limited to suspending feeding, starting the sterilization device, and recording the abnormal log. S8: After the early warning response is executed, continuously monitor the production line status data and dynamically adjust the early warning threshold and model parameters according to subsequent data changes to improve the accuracy and adaptability of subsequent predictions.

[0012] Step S1: Collect multimodal real-time data on the temperature, humidity, pH value, dissolved oxygen concentration, image data, and operation logs of the fermentation tank in the pulp production line, and record the data acquisition timestamp and equipment identification information. Specifically, this includes: S1.1: Based on the industrial Internet of Things architecture, deploy multiple types of sensors to continuously sample the temperature, humidity, pH value, and dissolved oxygen concentration in the fermenter to obtain multi-channel time series data, providing a data foundation for building a complete state perception of the fermentation process; To address the need for multi-channel environmental status acquisition within the fermenter, an Industrial Internet of Things (IIoT) architecture is adopted to physically connect temperature and humidity sensors (temperature measurement resolution of 0.1℃, humidity measurement resolution of 0.1%RH), pH electrochemical sensors (voltage signal sampling rate of 1000Hz), and dissolved oxygen optical probes (excitation wavelength of 470nm, sampling frequency of 1Hz) to edge acquisition nodes, thereby enabling continuous sampling of various physical quantities. Furthermore, through a steady-state sampling control algorithm (parameter: sampling interval) Minutes, sampling batch size The temperature and humidity sensor acquisition channels are polled at a constant speed to ensure the integrity of time series data and cached in real time to the memory of edge nodes; Furthermore, by using the potential difference measurement method (parameter: the reference electrode potential stabilizes at...), The pH sensor is driven to continuously output a voltage signal with a sampling rate of 1000Hz. The voltage value is converted into a pH value using an analog-to-digital converter, and a time stamp identifier is added by a time stamp generator. Furthermore, the fluorescence lifetime signal returned by the dissolved oxygen optical probe was analyzed using phase modulation scattering, and a curve fitting algorithm was employed (parameter: fitting order). The dissolved oxygen concentration values ​​were obtained, and independent time series entries were generated. The modal synchronization control algorithm is used to align the time series data of the three sensors according to a unified time reference, ensuring the temporal consistency of different physical quantities in the subsequent multimodal data fusion process, and realizing the data foundation for continuous perception of the fermentation process status. For example, in a certain fermentation batch of pulp production, the temperature and humidity sensor is set to a sampling interval of [missing information]. Minutes, temperature measurement resolution is ℃ and humidity measurement resolution is %RH; pH sensor uses reference electrode potential Voltage signal sampling rate Hz; Dissolved oxygen probe setting excitation wavelength nm, sampling frequency Hz. During data acquisition, the temperature channel receives data per hour. Continuous samples, pH channel output per minute One measurement value, dissolved oxygen channel output per hour The lifetime resolution value is calculated. The three channels are aligned using a modal synchronization control algorithm based on the principle of minimizing timestamp differences, generating a length of [value missing]. A multi-channel time series matrix was generated, containing four columns of data: temperature, humidity, pH, and dissolved oxygen. In subsequent feature processing, it was verified that this acquisition strategy resulted in alignment deviations of the temperature and humidity data below a certain threshold. Seconds, the deviation between pH and dissolved oxygen is less than This ensures the input accuracy of multimodal fusion prediction within seconds; S1.2: The state of materials inside the fermenter is captured in timed images using an industrial camera and image acquisition card. High dynamic range imaging technology is used to acquire image data to enhance the visibility and recognition accuracy of microbial metabolites in the images. After receiving the time series data of the multi-channel environmental sensor inside the fermenter from the output of step S1.1 as background conditions, in order to meet the requirement of acquiring the visual state of the material inside the fermenter, an image acquisition unit composed of an industrial camera and an image acquisition card (parameters: photosensitive element resolution ≥5MP, signal-to-noise ratio >45dB, frame rate ≥30fps) is used to realize the timed acquisition function of the image inside the fermenter. High dynamic range imaging (HDR) technology (parameters: exposure times ≥ 3, exposure time series is short-medium-long exposure combination, dynamic range enhancement ratio ≥ 60dB) is used to acquire multiple exposure images of the same shooting position, and the multiple exposure sequences are combined into a high dynamic range single frame image through HDR fusion algorithm to obtain visual data with balanced brightness and shadow details; Furthermore, gamma correction and color balance algorithms are used (parameter: gamma value set to...). With a color temperature adjustment range of 4500K–6500K, the image brightness and color are standardized and adjusted to obtain image data suitable for subsequent microbial metabolite feature extraction. Furthermore, a combination of spatial domain filtering and frequency domain enhancement is used (the spatial domain employs a median filter kernel size). × The frequency domain uses a high-frequency component enhancement factor. This process achieves image noise suppression and texture detail enhancement, and generates processing results that salience microbial metabolic features; Furthermore, through morphological gradient operations (structural element size) × Detect the edge contours of metabolic products on the surface of materials and calculate the visibility index of metabolic products. ,in The number of pixels representing the area of ​​the region where metabolites are identified. The total number of pixels in the image is used to quantify the identifiability of metabolites. Through the above multi-stage image processing algorithm chain, the timed acquisition of multi-exposure image sequences in the previous step is transformed into structured visual data with high dynamic range, noise suppression, detail enhancement and quantified recognizability, thereby improving the visibility and recognition accuracy of microbial metabolites. For example, in a fermentation workshop producing 100 tons of raw pulp per day, an industrial camera using a Sony IMX264 image sensor with a resolution of 2448×2048 pixels, set to short exposure time in HDR shooting mode, is used. / Seconds, medium exposure / Seconds, long exposure / The dynamic range of the synthesized image is increased to 75dB, and after gamma correction and color temperature adjustment (5000K), the noise level is reduced to the original level. Texture detail sharpening was improved by 20%, and the consistency rate between the outlines of metabolites detected by morphological gradients and manually annotated samples reached 94%, with a visibility index of [missing information]. ≈0.024, ensuring that the subsequent model inference module can accurately identify contaminated areas and achieve input quality assurance for workshop pollution risk early warning; S1.3: Based on the Manufacturing Execution System (MES) interface, acquire the operation log data of operators during the fermentation process, including feeding time, stirring rate setting, and sterilization operation records, to form an operation behavior dataset that is strongly correlated with the process flow; S1.4: During the data acquisition process, the acquisition timestamp and device identification information of each type of data are recorded synchronously. A unified time base is used to align the timestamps of multi-source data to ensure the consistency of the timing of subsequent multimodal data fusion. S1.5: The raw multimodal data collected is initially encapsulated and standardized in format by edge computing nodes to generate structured data packets, which are then transmitted to cloud or local processing modules for in-depth preprocessing and analysis.

[0013] Step S2 involves preprocessing the collected multimodal data, including missing value imputation, outlier removal, and time alignment, to eliminate timing deviations caused by differences in sampling frequencies from different sensors. Specifically, this includes: S2.1: Perform missing value identification processing on the temperature and humidity sensor data, pH sensor data, dissolved oxygen concentration sensor data, image acquisition equipment data and operation log text data collected from the fermenter, and fill in the missing samples based on the linear interpolation algorithm and the preceding and following timestamp data to obtain a complete data sequence; For temperature and humidity sensor data, pH sensor data, dissolved oxygen concentration sensor data, image acquisition equipment data, and operation log text data collected from the fermenter, a missing value marking method (parameters: sampling interval Δt, threshold setting β) is used to automatically detect and mark missing sampling points in the data sequence. Furthermore, by using the timestamp adjacent difference calculation method (parameters: timestamp sequence T, expected sampling interval Δt), the missing points are accurately located, and a list of missing data location indices is obtained; Furthermore, a linear interpolation algorithm (parameter: interpolation window size w) is used to estimate the missing samples. The interpolation calculation formula is as follows: in, These are the interpolated missing data values. The value of the previous valid sampling point. The value of the next valid sampling point. and These are the timestamps of the valid sampling points before and after. The timestamp of the missing point; Furthermore, the accuracy of the interpolation is verified by using a residual test method for the interpolation results (parameter: residual threshold γ), and a corrected complete data sequence is generated. By using the above linear interpolation algorithm and residual verification processing method, the results of the previous step are transformed into stable input data that are complete in time sequence and can be used for subsequent anomaly detection and multimodal fusion processing. The expected technical effect is to reduce the inconsistency of multi-source data and improve the model inference accuracy. For example, in a fermentation monitoring scenario of a pulp production line, the temperature and humidity sensor has a sampling frequency of 1Hz, the pH sensor has a sampling frequency of 0.5Hz, and the sampling interval Δt is set to 1s and 2s respectively. The missing value detection threshold β is set to any interval in the sequence that exceeds Δt × 1.5. In actual operation, a single missing point appears at the 300th second in a certain temperature and humidity data sequence. The timestamps before and after the missing point are 299s and 301s respectively, corresponding to temperatures of 35.2℃ and 35.6℃. The interpolation calculation is as follows: The calculation result is The residual test threshold γ was set to 0.05℃. After passing the test, this value was used to replace the missing points, resulting in a continuous and complete temperature data sequence. In the image acquisition and operation log data, missing frames and records were also located by comparing timestamp differences, and then interpolated with adjacent samples in the same time period to ensure the consistency of data time and the completeness of input in the multimodal fusion stage. S2.2: Anomaly detection is performed on the data streams of each sensor based on the sliding window statistical method. The mean and standard deviation of the data within the window are calculated, data points exceeding three times the standard deviation are identified, and the median replacement strategy is used to correct the anomalies in order to obtain a sensor data stream with enhanced stability. S2.3: Time synchronization processing of multimodal data streams is performed based on the timestamp alignment mechanism. The minimum timestamp difference matching algorithm is used to align sensor data with image frames and operation log entries at different sampling frequencies in time sequence, so as to eliminate the time sequence misalignment problem caused by sampling frequency differences. S2.4: Perform data format standardization conversion on the time-aligned multimodal data, normalize the numerical sensor data to the [0,1] interval, convert the image data to a uniform size and color space, and extract and encode the structured fields of the operation log text to obtain a preprocessed dataset in a uniform format; S2.5: Construct a multimodal feature vector matrix based on the preprocessed unified format dataset, and concatenate and fuse numerical features, image feature vectors and text embedding vectors to obtain a structured multimodal feature representation that can be used as input for deep learning models.

[0014] Step S3: Construct a dynamic knowledge graph covering the pulp production process, microbial growth characteristics, and historical contamination events, and design an incremental update mechanism based on graph neural networks to achieve continuous optimization and expansion of the knowledge graph. Figure 2 As shown, it specifically includes: S3.1: Perform text parsing and entity recognition on the original pulp production process flow documents, literature on microbial growth patterns, and records of historical contamination events to extract key entity nodes and their attribute information; S3.2: Based on the extracted key entity nodes and their attribute information, the semantic relationships between entities are identified using a relation extraction algorithm, and a triple representation of the initial knowledge graph is constructed to form a structured knowledge storage; In the input conditions, the key entity nodes and their corresponding attribute information extracted from S3.1, including the pulp production process flow nodes, microbial classification and their growth parameters, historical pollution event nodes and associated process parameter datasets, are used as the input objects for this sub-step relationship extraction process. A relation extraction algorithm based on dependency parsing (parameter: dependency tree construction rule = production process category first, microbiology category second) is adopted to parse the syntactic dependency relationship between key entities in the same text paragraph and search for specific dependency patterns in the parse tree to lock in potential semantic associations; Furthermore, a pattern matching algorithm based on open domain information extraction (parameter: syntax pattern library = includes process step-equipment relationship, microorganism-environmental factor relationship, pollution event-time node relationship) is used to achieve semantic mapping of dependency parsing results and obtain entity pairs and their relationship type annotation results; Furthermore, by using a statistical weighting evaluation method, the credibility of the extracted entity relations is quantified, and a confidence value for each relation triple is generated. Furthermore, a rule-based conflict resolution algorithm (parameter: prioritize retaining high-confidence relations and delete contradictory and low-confidence relations) is used to filter relation triples with semantic conflicts or redundancy, ensuring the accuracy and structural consistency of the resulting dataset. The triplet construction engine (parameter: triple format=(entity1, relation, entity2)) transforms the relation results filtered in the previous step into an initial knowledge graph storage structure and writes them into the knowledge graph database, thus realizing the establishment of the initial structured knowledge storage. For example, in a processing task, the input includes a set of entities {“Fermenter A”, “Temperature Monitor”, “Lactic Acid Bacteria”, “Contamination Event #045”} from a production process SOP document, along with attribute information (Fermenter A location = Workshop 1, Temperature Monitor accuracy = ±0.1℃, Optimal growth temperature for lactic acid bacteria = 35℃, Contamination event occurrence time = 2023-05-21). The relation extraction algorithm first uses dependency parsing to identify the relationships “Fermenter A”-“Installation”-“Temperature Monitor”, “Lactic Acid Bacteria”-“Growth Temperature”-“35℃”, and “Contamination Event #045”-“Occurrence”-“2023-05-21”. In the pattern matching step, these relationships are mapped to process equipment relationships, microbial characteristic relationships, and event time relationships, and weights are calculated. Assuming the frequencies of these relationships are 3, 5, and 2, the total number of texts is 10, and the context consistency coefficients are 0.9, 0.95, and 0.85, the weights of the three relationships are calculated according to the formula. , , The conflict resolution step retains the first two relationships and removes the third low-weight relationship. The final triplet is constructed as (fermenter A, installation, temperature monitor) and (lactic acid bacteria, growth temperature, 35℃), and stored in the Neo4j knowledge graph database. Subsequent steps can directly call this structured knowledge to support graph neural network embedding learning, which significantly improves the prior support of subsequent reasoning steps. S3.3: Input the constructed initial knowledge graph triples into the graph neural network model for embedding learning to obtain low-dimensional vector representations of nodes and relations, which are used to support subsequent graph reasoning and update operations; The input condition is the initial knowledge graph triple dataset that has been constructed, which contains node and relationship information between the pulp production process, microbial growth characteristics and historical contamination events; A graph embedding initialization algorithm (parameters: number of nodes N, number of relation types R) is adopted to initialize the vector representation space of each node and relation in the knowledge graph, ensuring that the input data matches the dimension of the graph neural network embedding layer; Furthermore, by employing a feature propagation algorithm based on Graph Convolutional Networks (GCN) (parameters: number of convolutional layers L, adjacency matrix A, feature matrix X), multi-order neighborhood propagation and aggregation of node features in the graph structure are achieved, resulting in semantically enhanced node embedding vectors. ; Furthermore, through the normalized linear transformation method (parameter: transformation matrix) Activation function This implements a non-linear mapping of node embedding vectors and generates a low-dimensional representation. This is to compress feature dimensions and improve model inference efficiency; Furthermore, by using the TransE-based relation encoding algorithm (parameters: embedding dimension d, learning rate α), relation triples are encoded. Mapping to a vector space and training is completed by minimizing the following energy function: in The value is the energy function value. , , These are vector representations of the head entity, relation, and tail entity, respectively. For vector norm operations; By jointly optimizing and training node and relation vectors, a low-dimensional embedding representation with structural fidelity and semantic interpretability is generated, providing a high-quality vector foundation for subsequent knowledge graph reasoning and incremental updates. For example, in a knowledge graph of pulp production processes, let the number of nodes N=500, the number of relation types R=20, the embedding dimension d=64, the number of GCN convolutional layers L=2, the learning rate α=0.01, the adjacency matrix A defined by the process dependencies between nodes, and the initial feature matrix X composed of process parameters and production data. GCN is used for computation. , then calculate The final node embedding matrix is ​​obtained. Using the TransE training process, the head entity "Fermentation Tank 1", the relation "Contamination Occurrence", and the tail entity "Excessive Lactic Acid Bacteria" are selected, with the optimization objective being to minimize... The L2 norm is iteratively updated until the loss function converges. Finally, on the validation set, the node classification accuracy is improved to over 92%, the recall rate for predicting pollution events reaches 89%, and the embedding vectors present a clear process flow organization structure in the visualization analysis. S3.4: Based on real-time collected multimodal data and model reasoning result feedback information, the node relationships in the knowledge graph are dynamically evaluated and corrected to achieve incremental updates and topology optimization of the knowledge graph; S3.5: Perform graph structure optimization and semantic consistency verification on the updated knowledge graph to improve the reasoning efficiency and semantic accuracy of the graph, and provide high-quality prior knowledge support for model fusion reasoning.

[0015] Step S4: The preprocessed multimodal data is input into a multimodal fusion deep learning model. Combined with the structured semantic information provided by the knowledge graph, a fusion feature representation is generated to enhance the model's ability to explain pollution risks. For example... Figure 3 As shown, it specifically includes: S4.1: Standardize and normalize the structured sensor data such as temperature, humidity, pH value, and dissolved oxygen concentration in the preprocessed multimodal data to eliminate the influence of different physical dimensions on model training and obtain a structured input feature vector with unified dimensions. S4.2: Perform ResNet-50-based feature extraction on the image data to extract the high-level semantic feature map of the image and generate deep feature vectors of the image modalities as visual input for multimodal fusion; S4.3: Perform BERT-based embedding encoding on the operation log text data to generate semantic feature vectors of the text modality in order to capture the process operation and abnormal behavior information contained in the log; S4.4: Based on the relationship between process nodes and microbial growth in the knowledge graph, a structured semantic feature matrix is ​​constructed, and a knowledge-aware weight allocation vector is generated through a graph attention mechanism to guide the model to focus on key feature modalities; Based on the updated knowledge graph nodes and relationship data, a knowledge embedding mapping method (parameters: node low-dimensional vector dimension is 128, relationship weight is initially set to mean distribution) is used to transform the association between process nodes and microbial growth into a structured semantic feature matrix. Furthermore, by using the relationship influence coefficient calculation method (parameter: relationship type influence coefficient range [0.1, 0.9]), basic weight values ​​are assigned to each feature dimension in the feature matrix, and a weighted structured feature matrix is ​​obtained; Furthermore, a graph attention mechanism is employed (parameters: attention coefficient α initially valued at 0.2, normalization method using Softmax) to amplify the weights of feature nodes with high correlation between the process flow and microbial growth, generating a node-level weighted coefficient vector. The calculation formula is as follows: in, For nodes For nodes Attention coefficient For nodes Embedded vector, For nodes Embedded vector, It is a linear mapping matrix. For attention parameter vectors, Vector concatenation operator; Furthermore, a node weight normalization processing method (parameter: L2 norm constraint value is 1) is adopted to uniformly normalize the magnitude of the weighted coefficient vector and obtain the knowledge perception weight allocation vector. By using graph attention mechanism, the structured feature matrix from the previous step is transformed into weighted data that integrates prior knowledge of the domain, thereby achieving the expected technical effect of guiding the model to automatically focus on key feature modalities in pollution risk prediction. For example, in one specific implementation, the input knowledge graph includes process flow nodes such as fermentation temperature, pH control point, and stirring frequency, as well as microbial growth-related nodes such as lactic acid bacteria and yeast, totaling 200 nodes. Each node's embedding vector dimension is set to 128, and the initial relation weights are assigned a mean weight of 0.5. Influence coefficients are calculated for the node relationships: the relationship between fermentation temperature and lactic acid bacteria growth is assigned a weight of 0.8, the relationship between stirring frequency and yeast growth is assigned a weight of 0.7, and the remaining relationships are weighted according to default values. In the graph attention mechanism, the initial value of α is set to 0.2. Softmax is used to amplify the attention of the fermentation temperature-lactic acid bacteria related feature vector, resulting in a weighting coefficient of 0.92; the pH control point-yeast relationship yields a weighting coefficient of 0.87; and other relationships are processed using LeakyReLU, maintaining coefficients around 0.5. After L2 normalization of all node weights, the highest weight node has a modulus of 1, and the weights of the remaining nodes are scaled proportionally. The final weight allocation vector has a length of 200. The node with the largest weight is given priority by the model in the pollution risk prediction results, and the prediction accuracy is improved to 97%. The interpretability test shows that the model’s attention distribution on key nodes is consistent with the domain experts by 0.94. S4.5: Multimodal fusion of structured sensor feature vectors, image semantic feature vectors, text semantic feature vectors, and knowledge graph semantic weight vectors is performed. Gated recurrent units and cross-attention mechanisms are used for feature alignment and semantic enhancement to generate fused feature representations, which serve as inputs for the subsequent pollution risk inference module.

[0016] Step S5: Based on the fusion of feature representation and prior rules in the knowledge graph, calculate the probability of microbial contamination risk under the current working conditions, and feed it back to the knowledge graph through a two-way interaction mechanism to correct potential erroneous associations. Specifically, this includes: S5.1: Perform semantic mapping processing on the fusion feature representation output by the multimodal fusion deep learning model to obtain an embedding vector representation consistent with the semantic space of entity nodes in the knowledge graph, thereby achieving semantic alignment between the model inference results and the knowledge graph structure; S5.2: Based on the semantically aligned embedded vector representation and prior rules in the knowledge graph, propagation reasoning based on graph neural networks is performed to generate a pollution risk confidence score that integrates domain knowledge, serving as a preliminary result for microbial pollution risk prediction. S5.3: Perform similarity measurement calculations on the preliminary pollution risk prediction results and the risk levels under similar historical operating conditions to identify whether the current operating condition belongs to a known pollution pattern, thereby determining whether it is necessary to introduce extended rules from the knowledge graph for secondary reasoning correction. S5.4: Based on the prediction error and confidence changes from model feedback, perform weight update operations on relevant nodes and edges in the knowledge graph to correct potential erroneous associations, and enhance the semantic expression of key contamination paths through graph attention mechanism; Based on the prediction error and confidence change data vectors from model feedback, a weighted difference update method is used (parameter: node weights). Edge weight Learning rate This enables dynamic adjustment of node and edge weights in the knowledge graph to respond to real-time changes in model inference results; Furthermore, through the error propagation algorithm (parameter: prediction error vector) Current weight matrix This allows the error to be mapped from the model output layer embedding space to the knowledge graph structure domain, resulting in a node weight correction matrix. With edge weight correction matrix ; Furthermore, the confidence level change rate is calculated using the confidence level change rate calculation method. The formula is: in, The current prediction confidence level, The confidence level from the previous time step is used to normalize the weight update magnitude. Furthermore, through graph attention mechanism (parameter: attention coefficient matrix) eigenvector set This enables semantic enhancement of key information paths and guides weight allocation towards edges and nodes with higher pollution risk correlation during node update. By using an incremental update mechanism, the knowledge graph structure after weight correction is transformed into new node embedding and edge embedding representations, thereby improving the accuracy of path parsing and semantic expression in pollution risk reasoning. For example, in the batch data of the pulp production line operation, the prediction error vector is: In the formula for calculating the rate of change of confidence, the confidence level at the previous time point... = Current confidence level = The rate of change was obtained. = In weighted difference update, set = The resulting matrix of node weight corrections and edge weight corrections is input into the graph attention mechanism training module. The path with the highest correlation coefficient in the attention coefficient matrix is ​​weighted. The verification results show that the updated knowledge graph improves the accuracy of target contamination pattern matching by approximately %. %, Explanatory scores improved This meets the accuracy and response speed requirements of real-time early warning systems; S5.5: Perform a rule-based reasoning-based feedback enhancement operation on the updated knowledge graph to generate a corrected pollution risk probability distribution, which serves as the final microbial pollution risk prediction output for use by the subsequent early warning judgment module; The updated knowledge graph undergoes feedback enhancement processing based on rule-based reasoning, employing a multi-layer rule matching algorithm (parameter: rule priority weight set). Node embedding matrix Relationship weight matrix This enables secondary semantic correction and risk probability reassessment of the model's inference results. Furthermore, through fuzzy set reasoning methods (parameter: membership function) (Gaussian type selected), fuzzy threshold (Set to 0.65)) to achieve a fuzzy expression of the probability interval of pollution risk and obtain the difference matrix ΔP of the probabilities before and after correction, which is used to evaluate the correction magnitude introduced by the rule; Furthermore, through a normalized weighted cumulative algorithm (parameter: weight coefficients) Normalized benchmark This involves weighted fusion of the rule-based reasoning results with the original risk probabilities output by the fusion features to generate an intermediate corrected probability vector. ; Furthermore, a logical consistency check algorithm is used (parameter: rule conflict detection threshold). (Set to 0.1), Semantic consistency metric function This involves verifying the rule-based suitability of the modified probability vector P_inter, and performing the final mapping transformation of the probability vector if the verification passes, to obtain the modified pollution risk probability distribution. ; By using the above-mentioned feedback enhancement processing method based on rule reasoning, the prediction results of the previous step are transformed into the final pollution risk probability distribution that combines domain knowledge constraints, thereby improving the semantic accuracy of risk assessment and the reliability of decision-making. For example, in a pulp production environment, a set of rule priority weights is set. = {0.35, 0.25, 0.20, 0.20}, Node embedding matrix The dimension is 128×512, and the relation weight matrix is... The dimension is 512×512. When using fuzzy set reasoning, the membership function... A Gaussian fuzzy type was selected, with a standard deviation σ set to 0.15 and a fuzzy threshold. Set it to 0.65. The original risk probability vector is... After fuzzification, the difference matrix ΔP is obtained as follows: The normalized weighted cumulative algorithm uses weight coefficients. = {0.6, 0.4}, normalized baseline = 1.0 Combines the results from the two sources mentioned above to generate for Logical consistency checks are performed by setting a threshold for rule conflict detection. = 0.1, semantic consistency measure function Based on cosine similarity calculation, when When ≥0.95, it maps to the final probability distribution. for This distribution, in field verification, reduced the false alarm rate of early warnings by 15% and improved the response accuracy to 92%.

[0017] Step S6: Based on the set pollution risk level threshold, determine whether the current risk probability exceeds the preset threshold. If it does, generate a pollution warning signal along with the risk level and suggested handling measures. Specifically, this includes: S6.1: Based on the pollution risk probability jointly output by knowledge graph and deep learning model, combined with the distribution characteristics of historical early warning events, a dynamic weighted algorithm is used to calculate the current risk level coefficient in order to quantify the degree of pollution risk under the current working conditions; S6.2: Normalize the calculated risk level coefficients to obtain standardized pollution risk scores, which are mapped to a preset five-level pollution risk level range to achieve automatic risk level classification. S6.3: Based on the preset multi-level risk threshold strategy, a threshold comparison algorithm is used to judge the current pollution risk score. If the score exceeds the threshold of the corresponding level, a risk level identification signal is generated as the basis for triggering the early warning. S6.4: Based on the generated risk level identification signal, call the predefined disposal suggestion knowledge base, execute the rule engine-based disposal measure recommendation, and generate an early warning disposal suggestion text that matches the current risk level; The risk level indicator signal data is parsed, and a rule engine-based condition matching method (parameter: the rule set comes from the disposal suggestion knowledge base) is used to implement the processing strategy invocation for signals of different risk levels; Furthermore, by using a conditional matching algorithm (parameters: risk level, pollution type, production section information), the algorithm can accurately retrieve matching entries from the disposal suggestion knowledge base and obtain the corresponding disposal rule ID set. Furthermore, a rule decoding method (parameter: set of rule IDs) is adopted to perform semantic parsing and logical judgment on multiple instructions in the disposal rule, and generate a structured execution sequence of disposal measures; Furthermore, a multi-dimensional priority ranking algorithm (parameters: execution time constraint, resource consumption, safety factor) is used to optimize the priority of the disposal measure sequence and generate a sorted list of disposal measures. Furthermore, a text generation template method (parameters: list of disposal measures, risk level description, and implementation precautions) is adopted to realize the textualization and formatting of the list of disposal measures and generate a complete warning disposal suggestion text; By using rule engine calls and template generation, the risk level identification signal from the previous step is transformed into structured early warning and handling suggestion text, enabling the description of response instructions that can be directly executed by the control system. For example, in the microbial contamination risk prediction of the fermentation process in the pulp production line, when the system outputs a risk level indicator signal of level three, corresponding to lactic acid bacteria contamination, and the production section information is two-stage fermentation, the rule engine calls the disposal rules in the disposal suggestion knowledge base. The matching conditions are risk level = 3, contamination type = lactic acid bacteria, and section = two-stage fermentation, retrieving the disposal rule ID set {R52, R87}. The rule decoding method parses R52 as "reduce fermentation temperature by 2℃, adjust stirring rate -15%", and R87 as "pause feeding for 30 minutes, start low-temperature sterilization mode". The multi-dimensional priority ranking algorithm ranks the measures according to execution time constraints (urgent tasks take priority), resource consumption (ranked by available equipment in the section), and safety factor (prioritize reducing bacterial activity in the environment). The ranking results are: 1. Pause feeding for 30 minutes; 2. Start low-temperature sterilization mode; 3. Reduce fermentation temperature by 2℃; 4. Adjust stirring rate -15%. The text generation template method converts the list into disposal suggestion text: "It is recommended to immediately suspend feeding for 30 minutes and simultaneously start the low-temperature sterilization mode; subsequently, reduce the fermentation temperature by 2°C and appropriately adjust the stirring rate to inhibit lactic acid bacteria activity. The above measures should be implemented under on-site monitoring conditions to ensure stable process parameters." The generated suggestion text is encapsulated and enters the subsequent transmission process, achieving the technical effect of providing clear and actionable instructions before the response. S6.5: Encapsulate the risk level identification signal and the corresponding disposal suggestion text to generate a structured pollution early warning signal data packet, which is then transmitted to the control system and triggers the corresponding early warning response mechanism.

[0018] Step S7: Transmit the pollution warning signal to the control system and trigger a warning response, including but not limited to suspending feeding, starting the sterilization device, and recording the abnormality log. Specifically, this includes: S7.1: Based on the pollution warning signal output by S6, the risk level, warning timestamp and suggested handling measures are transmitted to the PLC controller through the industrial Ethernet communication protocol to realize real-time synchronization of warning information and command issuance; Based on the structured pollution early warning signal data packets output by S6, the industrial Ethernet communication protocol of the IEEE 802.3 standard (parameters: data link layer frame format, physical layer rate 1000 Mbps) is used to realize the encapsulation and transmission functions of risk level, early warning timestamp and recommended handling measures. Furthermore, through the Modbus TCP data mapping mechanism (parameters: register address range 40001-40020, function code 0x10), the risk level is mapped to the control register field, the warning timestamp is encoded into the time register in Unix time format, and the rule ID corresponding to the suggested handling measures is encoded into the operation register, so as to realize the parsing of information and instruction caching at the PLC end; Furthermore, the CRC-32 check algorithm (parameters: polynomial 0x04C11DB7, initial value 0xFFFFFFFF) is used to perform redundancy check on the payload of the industrial Ethernet data frame, generate check codes to verify the integrity of data transmission, and reduce the probability of false triggering caused by signal noise or link packet loss. Furthermore, by utilizing a real-time priority queue scheduling mechanism (parameters: the highest priority is the pollution warning signal queue, and the delay threshold is ≤5 ms), pollution warning signal data packets are prioritized for forwarding in the industrial Ethernet switch, ensuring that the PLC controller receives the warning data within the millisecond delay range; By using the above-mentioned industrial Ethernet protocol encapsulation, Modbus TCP data mapping, CRC check and priority scheduling processing, the structured pollution early warning signal generated in the previous step is converted into control data that the PLC can parse, achieving millisecond-level synchronization of early warning information and high reliability of command issuance. For example, in a microbial contamination risk prediction system for a pulp production line, the PLC controller is a Siemens S7-1500 with an industrial Ethernet port speed of 1000 Mbps. The industrial switch supports IEEE 802.1Q VLAN and priority scheduling. In the contamination warning signal data packet, the risk level field occupies 2 bytes, with a value range of 1~5, encoded to address 40001 of Modbus TCP; the warning timestamp field occupies 8 bytes, representing the number of seconds since 00:00 on 1970-01-01, encoded to addresses 40002~40003; the suggested treatment measure field occupies 4 bytes as the rule ID, encoded to address 40004. A 4-byte checksum generated using CRC-32 is appended to the end of the data frame. During signal transmission, the average link delay is 3.8 ms, the CRC error rate is 0, and the PLC immediately triggers the corresponding control strategy template upon receiving the signal. Verification results show that all warning instructions are issued within 6 ms, meeting the system's design requirements for low latency and high reliability. S7.2: In the PLC controller, the corresponding control strategy template is executed based on the risk level to adjust the production line control parameters using PID control, so as to dynamically adjust the feeding rate or suspend the feeding operation to prevent the spread of pollution. S7.3: Trigger the sterilization device control relay based on the warning signal, and start the CIP cleaning system based on the preset sterilization time and temperature curve to perform in-situ cleaning and sterilization of the fermenter and pipelines; S7.4: Record abnormal log entries in the SCADA system, including warning time, risk level, triggering conditions, executed actions and equipment response status, to generate complete event tracing data; S7.5: Feedback the early warning response execution results to the model inference module as the basis for subsequent online optimization of model parameters and dynamic adjustment of early warning thresholds, so as to improve the closed-loop adaptability of the system.

[0019] Step S8: After the early warning response is executed, continuously monitor the production line status data and dynamically adjust the early warning threshold and model parameters based on subsequent data changes to improve the accuracy and adaptability of subsequent predictions. Specifically, this includes: S8.1: Continuously collect production line status data after the early warning response is executed, including the temperature and humidity, pH value, dissolved oxygen concentration, image data and operation logs inside the fermenter, in order to obtain real-time input data streams for model feedback optimization; S8.2: Perform missing value imputation, outlier removal and time alignment on the collected feedback data stream to eliminate sensor acquisition timing deviations and improve data consistency, and obtain a cleaned feedback input dataset; S8.3: Based on the cleaned feedback input dataset, combined with the structured semantic information in the knowledge graph, perform multimodal feature fusion calculation to generate feedback fusion feature vectors, thereby enhancing the model's semantic understanding of the current production state; S8.4: Based on feedback fusion feature vectors and historical pollution event data, an online incremental learning algorithm is used to dynamically update the parameters of the deep learning model in order to adapt to new pollution patterns and improve the model's generalization ability, thereby obtaining an updated pollution risk prediction model. S8.5: Based on the prediction error and risk probability distribution output by the updated pollution risk prediction model, and combined with the prior rules and risk level mapping relationship defined in the knowledge graph, the pollution early warning threshold parameters are dynamically adjusted to achieve adaptive optimization of the early warning mechanism.

[0020] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0021] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for real-time monitoring and early warning of microbial contamination risk in a pulp production line, characterized in that, Includes the following steps: S1: Collect multimodal real-time data from the fermentation tank in the pulp production line and record the data collection timestamp and equipment identification information; S2: Perform missing value imputation, outlier removal and time alignment preprocessing on the multimodal real-time data to obtain preprocessed multimodal data; S3: Construct a dynamic knowledge graph covering the pulp production process, microbial growth characteristics, and historical contamination events, and design an incremental update mechanism based on graph neural networks; S4: Input the preprocessed multimodal data into the multimodal fusion deep learning model, and generate fusion feature representations by combining the structured semantic information provided by the knowledge graph; S5: Based on the fused feature representation and the prior rules in the knowledge graph, calculate the probability of microbial contamination risk under the current working conditions, and feed it back to the knowledge graph through a two-way interaction mechanism to correct potential erroneous associations; S6: Based on the set pollution risk level threshold, determine whether the current risk probability exceeds the preset threshold. If it does, generate a pollution warning signal along with the risk level and suggested handling measures. S7: Transmit the pollution warning signal to the control system and execute the warning response action accordingly; S8: After the early warning response is executed, continuously monitor the production line status data and dynamically adjust the early warning threshold and model parameters according to subsequent data changes.

2. The method for real-time monitoring and early warning of microbial contamination risk in a pulp production line according to claim 1, characterized in that, Step S1 specifically includes: Based on the industrial Internet of Things architecture, multiple types of sensors are deployed to continuously sample the temperature, humidity, pH value, and dissolved oxygen concentration in the fermenter to obtain multi-channel time series data. The fermenter uses an industrial camera and image acquisition card to periodically acquire images of the material state inside the fermenter, and high dynamic range imaging technology is used to acquire image data. Based on the production execution system interface, the operation log data of the operators during the fermentation process is obtained, including feeding time, stirring rate setting, sterilization operation record, forming an operation behavior dataset that is strongly correlated with the process flow; During the data acquisition process, the acquisition timestamp and device identification information of each type of data are recorded synchronously, and a unified time base is used to align the timestamps of multi-source data. The raw multimodal data collected is initially encapsulated and standardized by edge computing nodes to generate structured data packets.

3. The method for real-time monitoring and early warning of microbial contamination risk in a pulp production line according to claim 2, characterized in that, The various types of sensors include a temperature and humidity sensor, a pH electrochemical sensor, and a dissolved oxygen optical probe, all of which are physically connected to the edge acquisition node. The temperature and humidity sensor has a sampling interval of 1 / 60 minute, a temperature measurement resolution of 0.1℃, a humidity measurement resolution of 0.1%RH, a pH sensor reference electrode potential that is stable at 0.2 volts, a voltage signal sampling rate of 1000 Hz, and a dissolved oxygen probe with an excitation wavelength of 470 nm and a sampling frequency of 1 Hz.

4. The method for real-time monitoring and early warning of microbial contamination risk in a pulp production line according to claim 1, characterized in that, Step S2 specifically includes: Missing value identification processing was performed on temperature and humidity sensor data, pH sensor data, dissolved oxygen concentration sensor data, image acquisition equipment data, and operation log text data collected from the fermenter. Missing samples were filled in based on linear interpolation algorithm and previous and subsequent timestamp data to obtain a complete data sequence. The sliding window statistical method is used to detect outliers in each sensor data stream. The mean and standard deviation of the data within the window are calculated, data points exceeding three times the standard deviation are identified, and the median replacement strategy is used to correct outliers, thereby obtaining a sensor data stream with enhanced stability. The time synchronization of multimodal data streams is performed based on the timestamp alignment mechanism. The minimum timestamp difference matching algorithm is used to align sensor data with image frames and operation log entries at different sampling frequencies in time. Data format standardization conversion is performed on the time-aligned multimodal data to normalize numerical sensor data to a unified range, convert image data to a unified size and color space, and extract and encode structured fields from operation log text to obtain a preprocessed dataset in a unified format. A multimodal feature vector matrix is ​​constructed based on the preprocessed dataset. Numerical features, image feature vectors and text embedding vectors are concatenated and fused to obtain a structured multimodal feature representation.

5. The method for real-time monitoring and early warning of microbial contamination risk in a pulp production line according to claim 1, characterized in that, Step S3 specifically includes: Text parsing and entity recognition were performed on raw pulp production process documents, literature on microbial growth patterns, and records of historical pollution events to extract key entity nodes and their attribute information. Based on the key entity nodes and their attribute information, a relation extraction algorithm is used to identify the semantic relationships between entities and construct a triple representation of the initial knowledge graph. The initial knowledge graph triples are input into a graph neural network model for embedding learning to obtain low-dimensional vector representations of nodes and relationships; Based on real-time collected multimodal data and model reasoning feedback information, the relationships between nodes in the knowledge graph are dynamically evaluated and corrected. For the updated knowledge graph, perform graph structure optimization and semantic consistency verification.

6. The method for real-time monitoring and early warning of microbial contamination risk in a pulp production line according to claim 5, characterized in that, In step S3, the construction of the dynamic knowledge graph includes text parsing to obtain key processes, microorganisms and pollution event entities, using dependency syntax and open domain information extraction to identify entity relationships, storing triple structures in the knowledge graph database, initializing the number of nodes and relationship types through graph neural network embedding, and using TransE to encode node relationships in low-dimensional vectors.

7. The method for real-time monitoring and early warning of microbial contamination risk in a pulp production line according to claim 1, characterized in that, Step S4 specifically includes: The structured sensor data in the preprocessed multimodal data is standardized and normalized to obtain a structured input feature vector with uniform dimensions; Perform ResNet-50-based feature extraction on image data to extract high-level semantic feature maps of the image and generate image semantic feature vectors of image modalities; The operation log text data is embedded and encoded using the BERT model to generate text semantic feature vectors of the text modality. Based on the relationship between process nodes and microbial growth in the knowledge graph, a structured semantic feature matrix is ​​constructed, and a knowledge graph semantic weight vector with knowledge perception is generated through a graph attention mechanism. The structured input feature vector, the image semantic feature vector, the text semantic feature vector, and the knowledge graph semantic weight vector are fused in a multimodal manner. Gated recurrent units and cross-attention mechanisms are used for feature alignment and semantic enhancement to generate a fused feature representation.

8. The method for real-time monitoring and early warning of microbial contamination risk in a pulp production line according to claim 1, characterized in that, Step S5 specifically includes: Semantic mapping is performed on the fusion feature representation output by the multimodal fusion deep learning model to obtain an embedding vector representation consistent with the semantic space of entity nodes in the knowledge graph; Based on the embedded vector representation and prior rules in the knowledge graph, propagation reasoning based on graph neural networks is performed to generate a pollution risk confidence score that integrates domain knowledge, which serves as a preliminary prediction result of microbial pollution risk. The preliminary pollution risk prediction results are compared with the risk levels under similar historical conditions to calculate similarity measurement, identify whether the current condition belongs to a known pollution pattern, and thus determine whether it is necessary to introduce extended rules from the knowledge graph for secondary reasoning correction. Based on the prediction error and confidence changes from model feedback, perform weight update operations on relevant nodes and edges in the knowledge graph; Perform rule-based reasoning-based feedback enhancement operations on the updated knowledge graph to generate a revised pollution risk probability distribution.

9. The method for real-time monitoring and early warning of microbial contamination risk in a pulp production line according to claim 8, characterized in that, In step S5, the final pollution risk probability distribution is corrected by multi-level rule matching and fuzzy set reasoning method. The membership function is Gaussian type, the fuzzy threshold is set to 0.65, the rule priority weight set and the weighted cumulative algorithm normalization benchmark are used to generate the final probability distribution, and the logical conflict test threshold is set to 0.

1.

10. The method for real-time monitoring and early warning of microbial contamination risk in a pulp production line according to claim 1, characterized in that, In step S6, after the pollution risk score is standardized, it is automatically divided according to the preset five-level risk level range. The current risk level coefficient is corrected by using a dynamic weighting algorithm and the distribution of historical early warning events. The risk level, pollution type, and process section are matched by the rule engine and the disposal suggestion knowledge base to generate sorted disposal measure suggestion text.

Images