A large model-based fermentation process data analysis method and system

Abstract

The application relates to the technical field of intelligent control of industrial fermentation processes, and discloses a fermentation process data analysis method and system based on a large model, which aligns sensor time series data, process text logs and image data according to a unified timestamp, carries out abnormal value cleaning, missing value interpolation and standardization processing, and obtains preprocessed data; based on the preprocessed data, each modal feature is extracted through a time series neural network, a language model and a visual model respectively, and a fusion representation sequence is generated; the fusion representation sequence and a current process context are input into a fermentation field large model, and an abnormality detection, root cause diagnosis, trend prediction or optimization recommendation task is executed; a structured analysis report is output, and operator feedback is collected for model iteration; and the application improves the fusion analysis efficiency of multi-source data and the explainability of process optimization suggestions.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for industrial fermentation processes, specifically to a data analysis method and system for fermentation processes based on a large model. Background Technology

[0002] Modern bio-fermentation processes involve the real-time acquisition of multi-source heterogeneous data, including temperature, dissolved oxygen, pH, stirring speed, exhaust gas composition, cell morphology, and substrate concentration. This data is characterized by high dimensionality, strong temporal sequence, and complex coupling relationships between variables. Traditional fermentation data analysis primarily relies on Statistical Process Control (SPC) based on mechanistic models or shallow machine learning methods, which have the following limitations: First, they struggle to simultaneously process structured sensor data, unstructured text process records, and microscopic information about cell morphology from images; second, they lack sufficient feature extraction and knowledge generalization capabilities when dealing with massive amounts of historical batch data; and third, anomaly diagnosis and process optimization suggestions heavily depend on expert experience, resulting in low efficiency in knowledge accumulation and reuse. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a data analysis method and system for fermentation processes based on a large model.

[0004] The first aspect of this invention provides a method for analyzing fermentation process data based on a large model, the method comprising the following steps: The sensor time series data, process text logs and image data are aligned according to a unified timestamp, and outlier cleaning, missing value imputation and standardization are performed to obtain preprocessed data. Based on the preprocessed data, modal features are extracted through temporal neural networks, language models and visual models respectively. In the feature fusion stage, the numerical time series and visualization curves are aligned at the fragment granularity. The shared semantics and unique semantics of the modalities are separated through discretization and decoupling interaction. Then, the relevant signals are dynamically focused on to generate a fused representation sequence. By inputting the fused characterization sequence and the current process context into a large model in the fermentation field, tasks such as anomaly detection, root cause diagnosis, trend prediction, or optimization recommendation can be performed. Output a structured analysis report and collect feedback from operators for model iteration. The structured analysis report includes the ranking of key influencing factors, confidence assessment, and control recommendations.

[0005] Optionally, in a first implementation of the first aspect of the present invention, the step of aligning sensor time-series data, process text logs, and image data according to a unified timestamp, and performing outlier cleaning, missing value imputation, and standardization to obtain preprocessed data includes: Collect sensor timing data, process text logs, and image data, and align them according to a unified timestamp; Outliers are identified using a sliding window-based statistical method. If all adjacent data points before and after an outlier are normal, the current outlier is identified as an isolated outlier and removed. If multiple consecutive data points are marked as outliers, these data points are marked as interruption or process mutation events. The fermentation stage to which the current data point belongs is identified. Linear interpolation is used to fill in the gaps during the lag phase and stationary phase, while local similarity-based interpolation is used to fill in the gaps during the logarithmic phase. The inflection point of the curve is detected based on the changes in the slope of the cell growth curve and the changes in the product synthesis rate. The intervals between the inflection points are labeled as the lag phase, logarithmic phase, stationary phase, and decline phase, respectively. The data after outlier cleaning and missing value imputation are standardized to obtain preprocessed data.

[0006] Optionally, in a second implementation of the first aspect of the present invention, based on the preprocessed data, features of each modality are extracted through a temporal neural network, a language model, and a visual model, respectively. In the feature fusion stage, the numerical time series and the visualization curve are aligned at the fragment granularity. The shared semantics and unique semantics of the modality are separated through discretization and decoupling interaction. Then, the system dynamically focuses on query-related signals to generate a fused representation sequence, including: Based on the preprocessed data, temporal feature vectors, semantic feature vectors, and visual feature vectors are extracted using temporal neural networks, language models, and visual models, respectively. The continuous sensor time-series data in the preprocessed data is segmented into semantically complete time-series segments, which include the numerical sequence of all sensor variables within the time interval and the corresponding visualization curve segments. Using time series segments as the basic unit, fine-grained cross-modal alignment of time series feature vectors, semantic feature vectors, and visual feature vectors is performed between the numerical time series domain and the visualization curve domain. The aligned temporal feature vector, semantic feature vector, and visual feature vector are decomposed into modality-shared semantics and modality-complementary semantics. Based on the current analysis and query task, signal segments in the time sequence are dynamically identified through an attention mechanism, and then weighted and summed before being spliced ​​together to generate a comprehensive representation sequence that includes time sequence information, textual semantic information, and visual information.

[0007] Optionally, in a third implementation of the first aspect of the present invention, the step of extracting temporal feature vectors, semantic feature vectors, and visual feature vectors based on preprocessed data through a temporal neural network, a language model, and a visual model, respectively, includes: The sensor time series data in the preprocessed data is input into a time series neural network based on the Transformer architecture. The input data is divided into multiple fixed-length subsequences. After position encoding of each subsequence, it is input into a multi-head self-attention network to capture the long-range dependencies between different time steps and output the time series feature vector corresponding to each subsequence. The preprocessed process text logs are input into a BERT-based language model to extract the contextual semantic representation of the text and output the semantic feature vector corresponding to each text record. The image data from the preprocessed data is input into a vision model based on the ViT architecture. The input image is divided into a sequence of image blocks, the visual features of each image block are extracted, and the visual feature vector corresponding to the whole image is output.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, before inputting the fused characterization sequence and the current process context into a large fermentation model to perform anomaly detection, root cause diagnosis, trend prediction, or optimization recommendation tasks, the method further includes: Collect fermentation process literature, standard operating procedure documents, historical batch reports, and experience texts to build a domain-pre-trained corpus; Incremental pre-training of the large language model on a pre-training corpus enables the model to learn professional terminology, process principles, and operating procedures in the fermentation field, thus forming a large model in the fermentation field.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, the step of inputting the fused characterization sequence and the current process context into a large model in the fermentation field to perform anomaly detection, root cause diagnosis, trend prediction, or optimization recommendation tasks includes: Load a large model in the fermentation field and concatenate the fused characterization sequence with the current process context into the model input sequence; The current fermentation stage is identified and a stage-aware location code is generated and superimposed on the input sequence. A sparse gating network is used to route and filter historical batch fragments according to the current fermentation stage, activating only the candidate fragments with the highest relevance to the current stage. The current time series data and candidate segments are decomposed into trends at multiple scales. After matching similarity at each scale, the weighted sum is calculated to retrieve the top-ranked similar historical segments. In each Transformer layer of the model, a cross-attention sublayer is added to use the retrieved historical fragments as key vectors and fuse them with the current sequence representation layer by layer in depth. Starting with the observed anomalous variables, we backtrack backward level by level along the dynamic causal graph carrying time delay parameters to locate root cause candidates where there are no upstream nodes of anomalous variables, and obtain rigid causal backtracking results. In the process of generating analysis conclusions from large models in the fermentation field, entity relationships are monitored in real time, and candidate relationships are temporarily stored in a queue; The generated reasoning paths of the large model in the fermentation field are input into the knowledge graph to retrieve relevant constraints, which are then encoded as soft constraint loss signals to adjust the reasoning direction of subsequent generation. Based on the analysis task type specified by the user, the corresponding analysis results are extracted.

[0010] Optionally, in the sixth implementation of the first aspect of the present invention, before the step of real-time monitoring of entity relationships and temporary storage of candidate relationships in the queue during the generation of analysis conclusions from the large model in the fermentation field, the method further includes: The user-specified task type is converted into an instruction prefix vector and injected into the starting position of the decoder of the large fermentation domain model; Based on the task type, observation facts are extracted from the fused representation sequence and organized into structured observation text blocks; The observed text blocks and retrieved historical fragments are compared item by item to identify the differences, similarities, or gaps between the current pattern and the historical pattern. The rigid causal backtracking results are injected into the generation context as a skeleton, and the fermentation domain big model transforms them into a causal narrative and supplements the explanation of the validity of each link. Output analysis conclusions based on task type, including anomaly detection, root cause diagnosis, trend prediction, or optimization recommendation tasks.

[0011] A second aspect of the present invention provides a data analysis system for fermentation processes based on a large model, the system comprising: The data processing module is used to align sensor time-series data, process text logs and image data according to a unified timestamp, perform outlier cleaning, missing value imputation and standardization processing to obtain preprocessed data; The multimodal data fusion module is used to extract modal features based on preprocessed data through temporal neural networks, language models and visual models respectively. In the feature fusion stage, the numerical time series and visualization curves are aligned at the fragment granularity. The shared semantics and unique semantics of the modalities are separated through discretization and decoupling interaction. Then, the module dynamically focuses on query-related signals to generate a fused representation sequence. The model processing module is used to input the fused characterization sequence and the current process context into a large model in the fermentation field to perform tasks such as anomaly detection, root cause diagnosis, trend prediction, or optimization recommendation. The results output module is used to output structured analysis reports and collect feedback from operators for model iteration. The structured analysis reports include the ranking of key influencing factors, confidence assessment, and control recommendations.

[0012] A third aspect of the present invention provides a large-scale model-based fermentation process data analysis device, the large-scale model-based fermentation process data analysis device comprising a memory and at least one processor, the memory storing instructions; the at least one processor calling the instructions in the memory to cause the large-scale model-based fermentation process data analysis device to perform various steps of the large-scale model-based fermentation process data analysis method as described in any of the preceding claims.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the large-model-based fermentation process data analysis method as described in any of the preceding claims.

[0014] The technical solution provided by this invention aligns sensor time-series data, process text logs, and image data according to a unified timestamp, performs outlier cleaning, missing value imputation, and standardization to obtain preprocessed data. Based on the preprocessed data, features of each modality are extracted using a time-series neural network, a language model, and a visual model, respectively. In the feature fusion stage, numerical time series and visualization curves are aligned at the fragment granularity. Through discretization and decoupling interaction, shared semantics and unique semantics of the modalities are separated, and then the focus is dynamically shifted to query relevant signals to generate a fused representation sequence. The fused representation sequence and the current process context are input into a large-scale model in the fermentation field to perform anomaly detection, root cause diagnosis, trend prediction, or optimization recommendation tasks. A structured analysis report is output, and operator feedback is collected for model iteration. This invention is the first to technically introduce a large-scale model into fermentation process analysis, breaking through the limitation of traditional methods that can only handle a single data modality, and realizing sensing... This system enables unified modeling and collaborative analysis of machine data, textual knowledge, and visual information. Through domain-adapted large-scale models, it not only identifies data patterns but also understands the biological mechanisms and engineering logic behind fermentation processes, providing in-depth analytical conclusions rather than superficial statistical correlations. Combining time-series modeling and causal inference, it effectively identifies key control nodes and true causal paths in the fermentation process, providing robust evidence for process optimization and reducing decision-making risks caused by spurious correlations. It significantly reduces the data analysis skills required of operators, enabling frontline process engineers to easily access intelligent analysis support and accelerating the transformation from experience-driven fermentation production to data-driven intelligence. Through knowledge graphs and online learning mechanisms, analytical capabilities continuously improve with the accumulation of production data, enhancing the efficiency of multi-source data fusion analysis and the interpretability of process optimization suggestions. This forms enterprise-specific fermentation process knowledge assets, realizing the long-term value accumulation of analytical models. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 A flowchart illustrating the fermentation process data analysis method based on a large model provided in this embodiment of the invention; Figure 2 A schematic diagram of the structure of a fermentation process data analysis system based on a large model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a fermentation process data analysis device based on a large model, provided in an embodiment of the present invention. Detailed Implementation

[0017] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular 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 orders other than those 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, apparatus, product, or device 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 devices.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 The flowchart of the fermentation process data analysis method based on a large model provided in this embodiment of the invention includes the following steps: Step 101: Align sensor time series data, process text logs and image data according to a unified timestamp, perform outlier cleaning, missing value imputation and standardization processing to obtain preprocessed data; In this embodiment, three types of raw data are acquired from sensor data acquisition, process log recording, and image acquisition equipment. For sensor time-series data, each record carries a precise timestamp; for process text logs, timestamps are added based on the log recording time; and for image data, timestamps are added based on the capture time. Using a unified reference clock as a benchmark, the timestamps of all data are converted to the same time coordinate system; a fixed sampling interval is set, dividing the time axis into equal-length time slots. For each time slot, all measured values ​​within that time slot are extracted from the sensor data; all text records generated within that time slot are extracted from the process text logs; and all images acquired within that time slot are extracted from the image data. If a certain type of data is missing within a time slot, a missing data marker is recorded.

[0019] In this embodiment, a sliding window-based statistical method is used to identify outliers. Specifically, for each sensor variable, a fixed-length sliding window is set, and the mean and standard deviation of the data within the window are calculated. Data points deviating from the mean by more than a preset multiple of the standard deviation are marked as candidate outliers. Further neighborhood verification is performed on these candidate outliers. If a data point is marked as outlier but its immediate neighbors are all normal, that point is confirmed as an isolated outlier and removed. If multiple consecutive data points are marked as outliers, this data segment is marked as a data interruption or process mutation event and handed over to the subsequent analysis module for processing.

[0020] In this embodiment, a staged imputation method based on fermentation stage characteristics is employed. First, the fermentation stage to which the current data point belongs is identified: lag phase, logarithmic phase, stationary phase, or decay phase. Then, an imputation strategy is selected based on the biological characteristics of different stages: in stages with gradual changes, such as lag and stationary phases, a linear imputation method is used, filling in the missing data using a weighted average of known data points before and after the missing point; in stages with drastic changes, such as the logarithmic phase, a local similarity-based imputation method is used, searching for the most similar segment to the sequence before and after the missing point from historical normal batches, and filling in the missing data with the corresponding value of that segment. For continuous large missing segments that cannot be effectively imputed, the data segment is marked as unusable.

[0021] In this embodiment, each sensor variable is standardized to map variables of different dimensions to the same numerical range, eliminating the impact of dimensional differences on subsequent analysis. Based on the statistical patterns of historical batch data, the fermentation stage boundary of the current batch is automatically identified. Based on the changes in the slope of the cell growth curve and the changes in the product synthesis rate, the inflection point of the curve is detected, and the intervals between the inflection points are labeled as the lag phase, logarithmic phase, stationary phase, and decay phase, respectively.

[0022] Step 102: Based on the preprocessed data, extract the features of each modality through temporal neural network, language model and visual model respectively. In the feature fusion stage, align the numerical time series and visualization curve at the fragment granularity. Separate the shared semantics and unique semantics of the modality through discretization and decoupling interaction. Then dynamically focus on the query related signals to generate the fused representation sequence. In this embodiment, the sensor time-series data in the preprocessed data is input into a Transformer-based temporal encoder for feature extraction. The encoder first divides the continuous time series into multiple equal-length subsequences according to a preset fixed time window. Adjacent subsequences may overlap partially to maintain the continuity of temporal information. For each subsequence, the encoder first performs temporal position encoding, generating a position vector reflecting its temporal order for each time step in the subsequence. This position vector is then added to the data value vector of the corresponding time step, enabling the subsequent attention mechanism to perceive the sequential relationship between data points. After position encoding, the subsequence is fed into a multi-head self-attention network. This network uses multiple parallel attention heads to calculate the attention weight distribution between different time steps in the sequence. Each attention head focuses on different temporal dependency patterns; for example, some heads focus on short-range rapid changes, while others focus on long-range slow trends. The outputs of all attention heads are concatenated and linearly transformed, and then nonlinearly mapped through a feedforward network. Finally, each subsequence is mapped to a fixed-dimensional temporal feature vector, which fully represents the comprehensive dynamic characteristics of all sensor variables and their intervariate coupling relationships within that time period.

[0023] In this embodiment, the process text logs in the preprocessed data are input into a pre-trained language model based on the BERT architecture for semantic encoding. First, multiple text records within the same time window are concatenated into a continuous text sequence in chronological order, with a specific classification marker added at the beginning of the sequence. Separator markers are inserted between each record to maintain boundary information. The concatenated text sequence is then converted into a corresponding marker index sequence after word segmentation, and along with the positional and paragraph codes of each marker, it is fed into the multi-layer bidirectional Transformer network of the BERT model. In each layer of the model, the network uses a bidirectional self-attention mechanism to allow each marker to simultaneously pay attention to all other markers to its left and right, thereby capturing the bidirectional semantic dependencies of the context. After multi-layered progressive abstraction, the model extracts the deep semantic vector corresponding to the classification marker position at the output as the overall semantic representation of the text records within that time window. This vector integrates the complete semantic information of all process operation records, anomaly notes, and observation descriptions within that time period.

[0024] In this embodiment, the microscopic images of bacterial cells and the color images of fermentation broth in the preprocessed data are input into a vision model based on the ViT architecture for visual feature extraction. The model first scales the input image to a uniform, fixed size, then cuts the entire image into multiple image patches of the same size according to a fixed grid division method. Each image patch covers a local spatial region of the original image. After each image patch is flattened into a one-dimensional pixel sequence, it is mapped to an image patch embedding vector through a linear projection layer. Simultaneously, a spatial location encoding reflecting the spatial position is added to each image patch embedding to preserve the relative layout relationship of each image patch in the original image. All image patch embedding vectors, along with the location encoding, are input into a multi-layer multi-head self-attention network of the vision model. In each layer, the network calculates attention weights between each image patch and all other image patches in the entire image, thereby capturing the visual dependencies between local patches and the global context. After multi-layer self-attention processing, the model performs global average pooling to aggregate the output vectors corresponding to all image patches, or extracts the output vectors corresponding to the classification label positions, and finally forms a fixed-dimensional visual feature vector for the entire image. This vector condenses key visual information such as cell morphology distribution, hyphal density, and color changes in fermentation broth.

[0025] In this embodiment, continuous sensor time-series data is segmented into semantically complete time-series segments. The segment length is dynamically adjusted based on the biological stages of the fermentation process: shorter segment lengths are used during rapidly changing phases such as the lag and logarithmic phases to preserve detailed information; longer segment lengths are used during periods of slower change such as the stationary and decay phases to improve computational efficiency. Each time-series segment corresponds to a time interval with complete semantics, containing the numerical sequence of all sensor variables within that interval and the corresponding visualization curve segment. Using generated time series segments as basic units, fine-grained cross-modal alignment is performed between the numerical time series domain and the visualization curve domain. For each time series segment, its numerical sequence and corresponding visualization curve are mapped to a unified latent space through different encoding networks, obtaining the segment's representation in the numerical domain and its representation in the visualization domain. Positive sample pairs are constructed, where the representations of the same time series segment in the numerical domain and the representations in the visualization domain constitute a positive sample pair; simultaneously, negative sample pairs are constructed, where different time series segments are combined across domains to form a negative sample pair. By optimizing the objective function, the distance between positive sample pairs in the latent space is forced to continuously decrease, while the distance between negative sample pairs is continuously increased, ultimately achieving semantic alignment between the numerical time series and the visualization curve at the segment granularity.

[0026] In this embodiment, the aligned multimodal features are further decomposed into two parts: modality-shared semantics and modality-specific semantics. Specifically, the features of each modality are input into a decoupled network, which encodes the features into a discrete set of latent variables. One set of latent variables represents semantic information shared across multiple modalities, such as the common expression of the semantic "fermentation temperature increases" across different modalities. The other set of latent variables represents complementary information specific to that modality, such as the morphological details of the cells observed in an image that cannot be obtained from sensor data. A constraint mechanism ensures that shared semantics remain consistent across different modalities, while allowing modality-specific semantics to freely express the unique information of each modality, thereby achieving adaptive synergy between shared and complementary information across modalities.

[0027] In this embodiment, based on the current analysis query task, the most informative key signal segments in the time series are dynamically identified. Specifically, the relevance score of each time series segment to the current query task is calculated; segments with higher relevance scores are considered more critical. Higher attention weights are assigned to highly relevant segments, allowing the model to focus on these key segments in subsequent processing; lower attention weights are assigned to less relevant segments. This mechanism enables the model to quickly locate the time interval most relevant to the current query within hundreds of hours of fermentation data. Aligned, decoupled, and weighted multimodal features are then spliced ​​and integrated in a unified latent space to generate a comprehensive representation sequence containing temporal information, textual semantic information, and visual information.

[0028] Step 103: Input the fused characterization sequence and the current process context into the large model of the fermentation field, and perform anomaly detection, root cause diagnosis, trend prediction or optimization recommendation tasks; In this embodiment, a large-scale model specifically for the fermentation field is constructed based on a general large-scale language model, employing a two-stage strategy for domain adaptation. A massive amount of fermentation process literature, standard operating procedure documents, historical batch reports, and expert experience texts are collected to construct a domain-specific pre-training corpus. The general large-scale language model is incrementally pre-trained on this corpus, enabling the model to learn fermentation field terminology, process principles, and operating specifications, forming a large-scale fermentation model. A fine-tuning dataset covering typical tasks such as fermentation anomaly diagnosis, process parameter optimization, yield prediction, and quality traceability is designed. Each instruction sample includes a task description, input context, and standard output. Dynamic rank adaptive LoRA technology is used for efficient fine-tuning: a learnable rank allocation mechanism is introduced on top of standard LoRA, simultaneously training LoRA modules with multiple rank ranges. During training, the representations learned by the adapter modules under different ranks are sorted, allowing the model to dynamically select the optimal rank based on the complexity of different tasks, achieving efficient parameterization and task-adaptive domain adaptation.

[0029] In this embodiment, the specific fermentation stage of the current batch is identified based on the labeled fermentation stage information and the dynamic characteristics of real-time data. A stage-aware routing mechanism is adopted, using temporal position encoding to determine the stage category of the current time point, and outputting the retrieval preference weight corresponding to the stage through a sparse gating structure to guide the subsequent retrieval process to focus on the historical knowledge most relevant to that stage.

[0030] In this embodiment, a multi-scale time series knowledge base is constructed based on the Kolmogorov-Arnold network. The complete time series of each sensor variable in the historical batch is input into the Kolmogorov-Arnold network. Based on the Kolmogorov-Arnold representation theorem, this network decomposes the multivariate time series dependencies into trend components at multiple scales. Each trend component at each scale is parameterized by an independent spline function. The decomposed trend components at each scale are indexed and stored in vector form, forming the multi-scale time series knowledge base. During retrieval, the current time series segment is also decomposed at multiple scales, and the knowledge base is searched for the historical segments most similar to the current components at each scale. Retrieved historical similar fragments are used as contextual information and input into the large-scale fermentation model along with the current fused representation sequence. The retrieval results not only serve as prompts injected into the model's input context but also undergo layer-by-layer deep fusion with the current hidden state of the large-scale model through a cross-attention mechanism. In each Transformer layer of the model, the retrieved knowledge vectors interact with the representation of the current sequence through additional cross-attention sublayers, enabling the model to fully utilize historical experience at every stage of inference. The model integrates current data with historical knowledge to execute user-specified analysis tasks.

[0031] In this embodiment, the user-specified task type, such as anomaly detection, root cause diagnosis, trend prediction, or optimization recommendation, is converted into a specific instruction prefix vector, which is then inserted into the starting position of the large model decoder. This prefix vector sets the generation target for the entire subsequent inference chain of the model. Different task types will activate the neural pathways in the model parameters that are most relevant to that task, so that the model focuses on the task target when generating content at each step. Depending on the task type, the model selectively extracts key observational facts from the fused representation sequence. For anomaly detection tasks, the model focuses on extracting the deviation of each sensor variable from historical statistical thresholds, the duration of the deviation, and the co-variable deviation characteristics. For trend prediction tasks, the model focuses on extracting the slope, acceleration, and precursor features of each key variable in the current time period. For root cause diagnosis and optimization recommendation tasks, the model focuses on extracting the current state and temporal evolution trajectory of all variables related to anomalous events or yield targets. The model organizes these extracted observational facts into structured observational text blocks. The model compares the extracted observations with patterns from retrieved historical similar segments item by item. For anomaly detection, the model identifies the differences and degrees of difference between the current pattern and historical normal batch patterns, while also identifying the similarity to historically known failure patterns. For trend prediction, the model finds multiple subsequent development trajectories from historical similar segments that best match the current previous curve, and identifies the differences in process conditions corresponding to different trajectories. For root cause diagnosis, the model identifies the common features between the current abnormal pattern and historical similar abnormal patterns. For optimization and recommendation, the model identifies the gaps between the current pattern and historical high-yield or high-quality batch patterns. When the model generates causal inference content, the root cause candidate results and causal backtracking paths that have already been completed by the temporal structure causal inference module are injected as additional hard inputs into the model's current generation context. Using this hard input as a framework, the model transforms the rigid causal chain of output (variable A → variable B → variable C) into a flexible causal narrative that conforms to natural language logic. The narrative further explains why each link in the chain holds true; for example, it explains why an anomaly in variable A is transmitted to variable B through metabolic pathways. For trend prediction tasks, the model infers the dynamic coupling relationship between current variables at this stage based on the subsequent changes in historical similar trajectories. For optimization and recommendation tasks, the model infers the chain reaction paths that may be triggered by adjusting specific process parameters at this stage. The model integrates the observation text, pattern text, and causal text generated in the first three stages, and locks the final conclusion output according to the task type. In the anomaly detection task, the model calculates anomaly scores based on the degree of observation bias and pattern deviation, compares them with preset anomaly judgment rules, and outputs specific anomaly type labels and anomaly severity levels. In the root cause diagnosis task, the model embeds the returned root cause candidate variable names into the conclusion sentence and sorts the root causes according to the effect strength of each link in the causal chain. In the trend prediction task, the model makes a comprehensive inference about the future trend of the current batch based on the subsequent development curves of retrieved historical similar trajectories, and outputs the predicted inflection point time and predicted value range. In the optimization recommendation task, the model generates a list of optimization measures containing specific regulatory variables, adjustment directions, adjustment magnitudes, and expected effects based on the gap items determined in the pattern recognition stage and the chain reaction paths predicted in the causal inference stage. The optimization measures are then sorted from high to low according to expected effects and locked as the final conclusion.

[0032] In this embodiment, for anomaly detection and root cause diagnosis tasks, a temporal structure causal reasoning module is activated. This module constructs a dynamic causal graph with fermentation time as the axis. Each node in the graph represents a key variable, such as temperature, dissolved oxygen, pH, substrate concentration, and product concentration. The directed edges between nodes represent the causal dependencies between variables, and each edge carries a time delay parameter, indicating how long it takes for the effect to be observed in the outcome variable after a change in the causal variable. The construction of the dynamic causal graph combines causal discovery algorithms with domain knowledge constraints: first, statistical dependencies between variables are learned from historical data using a data-driven approach to discover candidate causal edges; then, prior knowledge from the fermentation domain is used to prune and verify the candidate edges. During inference, for observed anomalous states, the influence path is traced back along the causal graph, starting from the anomalous variable and tracing back level by level along the reverse direction of the causal edges. During the backtracking process, the propagation delays between different variables are aligned, ultimately locating the most likely causal source variable.

[0033] In this embodiment, entity relationships are continuously extracted from the analysis conclusions and reasoning paths generated by the model during the analysis process. A delayed relationship extraction strategy is adopted: instead of immediately extracting relationships for each analysis conclusion, the relationship extraction task is postponed to a subsequent retrieval stage. Relationship extraction and graph updates are performed only when a piece of knowledge is actually retrieved and used, thereby reducing the construction cost of the knowledge graph. The extracted entity relationships are stored in the knowledge graph in the form of triples. Knowledge in the knowledge graph is stored in the form of vector indexes, supporting fast retrieval based on semantic similarity.

[0034] In this embodiment, during the model's generation of analytical conclusions, the knowledge graph provides structured reasoning constraints to the large model. The causal chains and logical rules in the knowledge graph are encoded as soft-constraint loss functions. As the model decodes and generates each step of the reasoning content, it checks in real-time whether the currently generated reasoning path is consistent with the known causal relationships and logical rules in the knowledge graph. If an inconsistency is detected, the model is implicitly guided by adjusting the generation probability, constraining the reasoning path to converge in a direction consistent with domain knowledge, thereby improving the rigor of the analytical conclusions.

[0035] Step 104: Output a structured analysis report and collect feedback from operators for model iteration. The structured analysis report includes the ranking of key influencing factors, confidence assessment, and control recommendations.

[0036] In this embodiment, the importance of each variable affecting the current analysis result is assessed based on the output of the time-series causal inference module. The assessment criteria include: the length of the causal path from the variable to the target variable in the dynamic causal graph, the strength of the causal effect along the path, and the degree to which the variable's current value deviates from the normal range. The impact score of each variable is calculated by combining these factors, and the variables are sorted from highest to lowest score to generate a ranking list of key impact factors.

[0037] In this embodiment, the credibility of the analysis conclusions is evaluated from multiple dimensions. These dimensions include: current data quality, the similarity between the retrieved historical similar batches and the current batch, the certainty of the causal reasoning path, and the internal consistency of each reasoning step when the model generates the conclusions. The overall confidence level of the analysis conclusions is output by combining the scores from each dimension.

[0038] In this embodiment, specific process control recommendations are generated based on the analysis conclusions and the ranking of key influencing factors. For anomaly diagnosis scenarios, the recommendations include the process parameters that need to be adjusted, the direction of adjustment, and the adjustment range; for production forecasting scenarios, the recommendations include possible process optimization measures. Each recommendation is accompanied by a causal explanation of its expected effect.

[0039] In this embodiment, the complete reasoning process of the model is decomposed into four steps: data observation, pattern recognition, causal inference, and conclusion generation, which are then presented layer by layer in a visual manner. The data observation step displays the key observational facts extracted by the model from multimodal data; the pattern recognition step displays the data patterns and historical similar cases identified by the model; the causal inference step displays the complete path of the model's backtracking reasoning along the dynamic causal graph; and the conclusion generation step displays the final analysis conclusions and regulatory recommendations. Each step is labeled with the data and knowledge sources on which it is based, enabling operators to understand the complete derivation logic of the analysis conclusions.

[0040] In this embodiment, the ranking of key influencing factors, confidence assessment, regulatory recommendations, and multi-step inference chains are integrated into a structured analysis report, presented to operators using a combination of natural language and visual charts. A feedback collection interface is embedded in the analysis report to collect operators' evaluations of the analysis conclusions, including: whether the analysis conclusions are accurate, whether the regulatory recommendations are reasonable and feasible, and whether the inference chain is clearly and understandably presented. Simultaneously, the actual regulatory measures taken by operators and their effects are recorded. The collected operator feedback and actual regulatory results are used as new training signals to iteratively optimize the model through an efficient online parameter update mechanism. This mechanism only updates the parameters of the dynamic rank adaptive LoRA adapter, keeping all parameters of the base model unchanged, thus avoiding catastrophic forgetting while absorbing new knowledge. The updated model parameters take effect immediately, allowing analytical capabilities to continuously improve with the accumulation of production batches.

[0041] Please see Figure 2 A schematic diagram of the structure of a fermentation process data analysis system based on a large model provided in this embodiment of the invention. The system includes: The data processing module is used to align sensor time-series data, process text logs and image data according to a unified timestamp, perform outlier cleaning, missing value imputation and standardization processing to obtain preprocessed data; The multimodal data fusion module is used to extract modal features based on preprocessed data through temporal neural networks, language models and visual models respectively. In the feature fusion stage, the numerical time series and visualization curves are aligned at the fragment granularity. The shared semantics and unique semantics of the modalities are separated through discretization and decoupling interaction. Then, the module dynamically focuses on query-related signals to generate a fused representation sequence. The model processing module is used to input the fused characterization sequence and the current process context into a large model in the fermentation field to perform tasks such as anomaly detection, root cause diagnosis, trend prediction, or optimization recommendation. The results output module is used to output structured analysis reports and collect feedback from operators for model iteration. The structured analysis reports include the ranking of key influencing factors, confidence assessment, and control recommendations.

[0042] Figure 3 This is a schematic diagram of a large-scale model-based fermentation process data analysis device 300 provided in an embodiment of the present invention. The large-scale model-based fermentation process data analysis device 300 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the large-scale model-based fermentation process data analysis device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the large-scale model-based fermentation process data analysis device 300 to implement the method provided in the above embodiment.

[0043] The fermentation process data analysis device 300 based on a large model may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The structure of the fermentation process data analysis device based on a large model shown does not constitute a limitation on the computer device provided by the present invention. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0044] The present invention 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 executed on a computer, cause the computer to perform the various steps of the fermentation process data analysis method based on a large model provided in the above embodiments.

[0045] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described equipment or apparatus / unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0046] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0047] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended technical solutions and their equivalents.

Claims

1. A data analysis method for fermentation processes based on a large model, characterized in that, The method includes the following steps: The sensor time series data, process text logs and image data are aligned according to a unified timestamp, and outlier cleaning, missing value imputation and standardization are performed to obtain preprocessed data. Based on the preprocessed data, modal features are extracted through temporal neural networks, language models and visual models respectively. In the feature fusion stage, the numerical time series and visualization curves are aligned at the fragment granularity. The shared semantics and unique semantics of the modalities are separated through discretization and decoupling interaction. Then, the relevant signals are dynamically focused on to generate a fused representation sequence. By inputting the fused characterization sequence and the current process context into a large model in the fermentation field, tasks such as anomaly detection, root cause diagnosis, trend prediction, or optimization recommendation can be performed. Output a structured analysis report and collect feedback from operators for model iteration. The structured analysis report includes the ranking of key influencing factors, confidence assessment, and control recommendations.

2. The data analysis method for fermentation processes based on a large model as described in claim 1, characterized in that, The process involves aligning sensor time-series data, process text logs, and image data according to a unified timestamp, performing outlier cleaning, missing value imputation, and standardization to obtain preprocessed data, including: Collect sensor timing data, process text logs, and image data, and align them according to a unified timestamp; Outliers are identified using a sliding window-based statistical method. If all adjacent data points before and after an outlier are normal, the current outlier is identified as an isolated outlier and removed. If multiple consecutive data points are marked as outliers, these data points are marked as interruption or process mutation events. The fermentation stage to which the current data point belongs is identified. Linear interpolation is used to fill in the gaps during the lag phase and stationary phase, while local similarity-based interpolation is used to fill in the gaps during the logarithmic phase. The inflection point of the curve is detected based on the changes in the slope of the cell growth curve and the changes in the product synthesis rate. The intervals between the inflection points are labeled as the lag phase, logarithmic phase, stationary phase, and decline phase, respectively. The data after outlier cleaning and missing value imputation are standardized to obtain preprocessed data.

3. The data analysis method for fermentation processes based on a large model as described in claim 1, characterized in that, Based on the preprocessed data, modal features are extracted using temporal neural networks, language models, and visual models, respectively. In the feature fusion stage, numerical time series and visualization curves are aligned at the fragment granularity. Modal shared semantics and unique semantics are separated through discretization and decoupling interaction. Then, the system dynamically focuses on query-related signals to generate a fused representation sequence, including: Based on the preprocessed data, temporal feature vectors, semantic feature vectors, and visual feature vectors are extracted using temporal neural networks, language models, and visual models, respectively. The continuous sensor time-series data in the preprocessed data is segmented into semantically complete time-series segments, which include the numerical sequence of all sensor variables within the time interval and the corresponding visualization curve segments. Using time series segments as the basic unit, fine-grained cross-modal alignment of time series feature vectors, semantic feature vectors, and visual feature vectors is performed between the numerical time series domain and the visualization curve domain. The aligned temporal feature vector, semantic feature vector, and visual feature vector are decomposed into modality-shared semantics and modality-complementary semantics. Based on the current analysis and query task, signal segments in the time sequence are dynamically identified through an attention mechanism, and then weighted and summed before being spliced ​​together to generate a comprehensive representation sequence that includes time sequence information, textual semantic information, and visual information.

4. The data analysis method for fermentation processes based on a large model as described in claim 3, characterized in that, Based on the preprocessed data, temporal feature vectors, semantic feature vectors, and visual feature vectors are extracted respectively through temporal neural networks, language models, and visual models, including: The sensor time series data in the preprocessed data is input into a time series neural network based on the Transformer architecture. The input data is divided into multiple fixed-length subsequences. After position encoding of each subsequence, it is input into a multi-head self-attention network to capture the long-range dependencies between different time steps and output the time series feature vector corresponding to each subsequence. The preprocessed process text logs are input into a BERT-based language model to extract the contextual semantic representation of the text and output the semantic feature vector corresponding to each text record. The image data from the preprocessed data is input into a vision model based on the ViT architecture. The input image is divided into a sequence of image blocks, the visual features of each image block are extracted, and the visual feature vector corresponding to the whole image is output.

5. The data analysis method for fermentation processes based on a large model as described in claim 1, characterized in that, Before inputting the fused characterization sequence and the current process context into a large fermentation model to perform tasks such as anomaly detection, root cause diagnosis, trend prediction, or optimization recommendation, the following steps are also included: Collect fermentation process literature, standard operating procedure documents, historical batch reports, and experience texts to build a domain-pre-trained corpus; Incremental pre-training of the large language model on a pre-training corpus enables the model to learn professional terminology, process principles, and operating procedures in the fermentation field, thus forming a large model in the fermentation field.

6. The data analysis method for fermentation processes based on a large model as described in claim 1, characterized in that, The process of inputting the fused characterization sequence and the current process context into a large-scale fermentation model to perform tasks such as anomaly detection, root cause diagnosis, trend prediction, or optimization recommendation includes: Load a large model in the fermentation field and concatenate the fused characterization sequence with the current process context into the model input sequence; The current fermentation stage is identified and a stage-aware location code is generated and superimposed on the input sequence. A sparse gating network is used to route and filter historical batch fragments according to the current fermentation stage, activating only the candidate fragments with the highest relevance to the current stage. The current time series data and candidate segments are decomposed into trends at multiple scales. After matching similarity at each scale, the weighted sum is calculated to retrieve the top-ranked similar historical segments. In each Transformer layer of the model, a cross-attention sublayer is added to use the retrieved historical fragments as key vectors and fuse them with the current sequence representation layer by layer in depth. Starting with the observed anomalous variables, we backtrack backward level by level along the dynamic causal graph carrying time delay parameters to locate root cause candidates where there are no upstream nodes of anomalous variables, and obtain rigid causal backtracking results. In the process of generating analysis conclusions from large models in the fermentation field, entity relationships are monitored in real time, and candidate relationships are temporarily stored in a queue; The generated reasoning paths of the large model in the fermentation field are input into the knowledge graph to retrieve relevant constraints, which are then encoded as soft constraint loss signals to adjust the reasoning direction of subsequent generation. Based on the analysis task type specified by the user, the corresponding analysis results are extracted.

7. The data analysis method for fermentation processes based on a large model as described in claim 6, characterized in that, Before real-time monitoring of entity relationships and temporary storage of candidate relationships in a queue during the generation of analytical conclusions from a large-scale model in the fermentation field, the method further includes: The user-specified task type is converted into an instruction prefix vector and injected into the starting position of the decoder of the large fermentation domain model; Based on the task type, observation facts are extracted from the fused representation sequence and organized into structured observation text blocks; The observed text blocks and retrieved historical fragments are compared item by item to identify the differences, similarities, or gaps between the current pattern and the historical pattern. The rigid causal backtracking results are injected into the generation context as a skeleton, and the fermentation domain big model transforms them into a causal narrative and supplements the explanation of the validity of each link. Output analysis conclusions based on task type, including anomaly detection, root cause diagnosis, trend prediction, or optimization recommendation tasks.

8. A data analysis system for fermentation processes based on a large model, characterized in that, The system includes: The data processing module is used to align sensor time-series data, process text logs and image data according to a unified timestamp, perform outlier cleaning, missing value imputation and standardization processing to obtain preprocessed data; The multimodal data fusion module is used to extract modal features based on preprocessed data through temporal neural networks, language models and visual models respectively. In the feature fusion stage, the numerical time series and visualization curves are aligned at the fragment granularity. The shared semantics and unique semantics of the modalities are separated through discretization and decoupling interaction. Then, the module dynamically focuses on query-related signals to generate a fused representation sequence. The model processing module is used to input the fused characterization sequence and the current process context into a large model in the fermentation field to perform tasks such as anomaly detection, root cause diagnosis, trend prediction, or optimization recommendation. The results output module is used to output structured analysis reports and collect feedback from operators for model iteration. The structured analysis reports include the ranking of key influencing factors, confidence assessment, and control recommendations.

9. A data analysis device for fermentation processes based on a large model, characterized in that, The large-scale model-based fermentation process data analysis device includes a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the large-scale model-based fermentation process data analysis device to perform each step of the large-scale model-based fermentation process data analysis method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the fermentation process data analysis method based on a large model as described in any one of claims 1-7.

Images