An industrial time series prediction-oriented time sequence and multi-process semantic enhancement large language model method

Abstract

The application belongs to the technical field of industrial data processing and intelligent prediction, and discloses a time sequence and multi-process semantic enhancement large language model method for industrial time series prediction. A semantic method based on function prototype matching is used to extract the process semantic features in the industrial time sequence data, convert the one-dimensional time sequence into a multi-dimensional semantic feature sequence, and realize industrial time sequence sample enhancement. A material flow information-time two-dimensional feature alignment method is constructed in combination with the time constraint relationship of the upstream and downstream multi-process material flow on the current time sequence process to form a standardized data sample slice structure. Finally, an encoder-decoder based on context prior and attention guidance is constructed to inject multi-process context semantics and complete time sequence token reconstruction. The reconstructed time sequence token is input into the large language model for prediction. The method improves the industrial time series prediction accuracy and provides decision support for production process optimization scheduling.

Description

Technical Field

[0001] This invention belongs to the field of information technology and relates to a large language model method for time series and multi-process semantic enhancement for industrial time series prediction. Background Technology

[0002] Process industries typically feature long-term continuous operation, multi-process synergy, and strong correlation between upstream and downstream processes. They also accumulate a wealth of industrial time-series data on equipment status, energy production and consumption, and key process variables. This data reflects the production process, energy transfer relationships, and changing operating conditions. Accurate prediction of key process variables and energy production and consumption time-series data helps to anticipate system operating status and trends, providing a basis for production scheduling, energy optimization, and operational decisions, thereby reducing energy consumption, operating costs, and carbon emissions.

[0003] To improve the effectiveness of industrial time series forecasting, existing research mainly relies on historical operational data for modeling and focuses on the changing patterns of target variables and the correlation between multiple variables. For example, Yang et al. proposed a transferable deep slow feature network to enhance the representation of key features for the problem of few-shot industrial time-series prediction (D. Yang, X. Peng, C. Jiang, X. Wu, SX Ding, and W. Zhong, “Transferable deep slow feature network with target feature attention for few-shot time-series prediction,” IEEE Transactions on Industrial Informatics, vol. 20, no. 5, pp. 7292-7302, May 2024); Liu et al. proposed a hierarchical Transformer network to fuse multi-frequency time-series features to address the problem of inconsistent sampling frequencies in industrial processes (D. Liu, Y. Wang, C. Liu, X. Yuan, and C. Yang, “Multirate-Former: An efficient Transformer-based hierarchical network for multistep prediction of multirate industrial processes,” IEEE Transactions on Instrumentation and Measurement, vol. 73, pp. 1-13, 2024). The methods described above can uncover dynamic patterns in numerical time series data, but they mainly rely on statistical correlation and do not make sufficient use of industrial knowledge such as process mechanisms, expert experience, and production plans.

[0004] To compensate for the shortcomings of purely data-driven prediction methods, existing research has begun to incorporate external knowledge and mechanistic constraints into industrial prediction models. For example, Wang et al. integrated production process characteristics and planning constraints into an industrial energy prediction model, improving energy prediction performance (T. Wang, T. Wang, J. Song, J. Zhao, H. Leung, and W. Wang, “An industrial energy prediction method integrating planning information and process correlation characteristics,” IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 23430-23442, 2025); Zhang et al. used domain knowledge as a priori causal structure to improve the stability and interpretability of soft sensing models (X. Zhang, C. Song, B. Huang, and J. Zhao, “Bayesian-based causal structure inference with a domain knowledge prior for stable and interpretable soft sensing,” IEEE Transactions on Cybernetics, vol. 54, no. 10, pp. 6081-6094, Oct. 2024). However, the above methods still mainly rely on structured priors or explicit mechanistic equations, while a large amount of operational knowledge in industrial settings exists in the form of unstructured text such as operation logs, production plans, and abnormal failure reports. Traditional models have difficulty effectively parsing the process semantics in these texts and establishing logical connections between them and numerical time series data.

[0005] In recent years, large language models and time series basic models have provided new technical paths for industrial time series forecasting. For example, Timer alleviates the heterogeneity of time-series data by unifying sequence representations (Y. Liu, H. Zhang, C. Li, X. Huang, J. Wang, and M. Long, “Timer: Generative pre-trained transformers are large time series models,” ICML, pp. 32369-32399, 2024); Time-LLM achieves prediction by reprogramming time series to a large language model label space (M. Jin, S. Wang, L. Ma, et al., “Time-LLM: Time series forecasting by reprogramming large language models,” ICLR, 2024); and ChatTS achieves cross-modal alignment by synthesizing time-series-text data (Z. Xie, Z. Li, X. He, et al., “ChatTS: Aligning time series with LLMs via synthetic data for enhanced understanding and reasoning,” Proceedings of the VLDB Endowment, vol. 18, no.8, pp. 2385-2398). (2025). The methods described above can improve time series prediction performance by leveraging the semantic understanding capabilities of large language models, but they are mostly geared towards general time series scenarios, typically assuming that numerical sequences and textual information directly correspond within the same time frame. In process industries, influenced by material transfer, process response delays, and upstream and downstream process collaboration, the semantic information of different processes and process variables often have temporal relationships, and these semantic relationships dynamically change with process stages, equipment status, and production context. Therefore, how to achieve temporal alignment and dynamic contextual fusion of industrial time series data and multi-process textual semantic information is a key issue in improving the accuracy of time series prediction in process industries. Summary of the Invention

[0006] This invention proposes a large language model method for time series and multi-process semantic enhancement for industrial time series prediction. This prediction method has high accuracy and can provide more reliable guidance for industrial energy optimization scheduling.

[0007] The technical solution of the present invention is as follows:

[0008] A large language model method for time series and multi-process semantic enhancement for industrial time series forecasting, comprising the following steps:

[0009] Step 1, Industrial Temporal Semantic Modeling:

[0010] Based on the waveform characteristics of industrial time series data, semantic modeling of intermittent and continuous industrial time series data is carried out using the prototype of double Sigmoid function and the prototype of normalized Beta function, respectively. Semantic parameters of each production semantic process are extracted by sliding matching. Industrial time series data includes shape, location and area. Semantic parameters are jointly encoded with upstream and downstream process information and task instructions to construct a semantic hybrid input for prediction tasks.

[0011] Step 1.1, Function prototype construction:

[0012] Based on the waveform characteristics of industrial time series data, corresponding function prototypes are selected to perform semantic modeling of industrial time series data.

[0013] For industrial time series data with intermittent fluctuation characteristics, a function prototype consisting of two sigmoid functions is used to describe the fluctuation segments:

[0014] (1)

[0015] in, This is the set of parameters for the function prototype; This represents the pulse amplitude, indicating the maximum value of a waveform during a production process. Indicates the pulse start position, used to determine the start time of the production process; This indicates the effective width of the high-level maintenance phase, reflecting the duration of a single production process; and Control the rate of change of the rising edge and the falling edge separately; The standard sigmoid activation function; t represents the discrete sampling time in the industrial time series data;

[0016] For industrial time series data with continuous fluctuation characteristics, in order to characterize the local fluctuation features in the industrial time series data, the normalized Beta function is selected as the function prototype, and the formula is as follows:

[0017] (2)

[0018] (3)

[0019] in, s and l represent the start time and duration of the production process, respectively; and The waveform skewness and peak position are controlled respectively; k represents the baseline offset; c represents the scale stretching parameter, which is used to adjust the intensity of local fluctuation characteristics; u=(ts) / l represents the normalized local time variable, which is used to map the interval [s,s+l] to the interval [0,1];

[0020] Step 1.2, Function prototype matching:

[0021] Further, the function prototype and Unified as Then combine it with industrial time series data Performing a point-by-point sliding matching operation along the time dimension yields the matching degree matrix:

[0022] (4)

[0023] in, To calculate the matching degree using vector dot product, T represents the length of the industrial time series data;

[0024] Since industrial time series data contains multiple production semantic processes corresponding to function prototypes, the corresponding window semantic parameters are obtained by searching for local maxima points in the matching degree matrix; the search process for local maxima points is as follows:

[0025] (5)

[0026] (6)

[0027] (7)

[0028] in, ρ represents the degree of matching between the function prototype and the industrial time series data; ρ represents the parameter set of the function prototype excluding... The parameter vector other than A, w, λ, μ}; Indicates in When fixed, make the matching degree The combination of parameters of the function prototype that achieves the maximum value; This represents the one-dimensional matching degree sequence obtained by projecting the matching degree matrix along the parameter vector ρ of the candidate prototype to its maximum value. This indicates the initial position of the nth local maximum point; Indicates Centered on, with The local search region is defined by the neighborhood radius; This indicates the time position parameter where the matching degree reaches its maximum value within the local search region;

[0029] In chronological order, each local maxima in the one-dimensional matching degree sequence corresponds to a production semantic process. Therefore, the optimal matching window determined by the nth local maximum corresponds to the nth production semantic process. Substitution The optimal shape parameter vector corresponding to the nth production semantic process is obtained:

[0030] (8)

[0031] in, and These represent the optimal pulse amplitude and optimal effective width for the nth production semantic process, respectively. and These represent the optimal parameters for controlling the rate of change of its rising and falling edges, respectively.

[0032] Therefore, in industrial time series data with intermittent fluctuation characteristics, the parameter set of the optimal function prototype corresponding to the nth production semantic process is determined by its temporal location parameters. With the optimal shape parameter vector Common characteristics are Furthermore, since the prototype function is integrated over the interval corresponding to the continuous waveform during the semantic production process, the area semantic feature is obtained as follows: Therefore, the semantic parameter set of industrial time series data with intermittent fluctuation characteristics is as follows: Similarly, in industrial time series data with continuous fluctuation characteristics, the parameter set of the optimal function prototype corresponding to the nth production semantic process is: Its corresponding area semantic feature is The final set of semantic parameters for industrial time series data with continuous fluctuation characteristics is: ;

[0033] Step 1.3, Construction with mixed inputs:

[0034] The original input of the i-th sample consists of upstream and downstream process information. and historical industrial timeline Composition, output as a future time window The Real Future Timeline Therefore, the original prediction task is represented as ;

[0035] Based on the number of local maxima in the matching degree matrix, historical industrial time series Represented as the current temporal semantics, consisting of a set of semantic parameters corresponding to multiple production semantic processes. , where N i This represents the number of semantic segments; subsequently, the current temporal semantics are... The semantic parameters corresponding to each production semantic process in the process are processed by a temporal semantic encoder. Mapped to temporal semantic embeddings consistent with the dimensions of the LLM hidden layers. Meanwhile, information on upstream and downstream processes. Text features are obtained through the native embedding layer of LLM. To make the LLM clearly define the original prediction task, the original prediction task instruction prompt word is constructed based on the prediction objective: "Predict next". Windows: And it is encoded into a task feature sequence through an LLM native embedding layer. Finally, the semantic mixed input of LLM is represented as:

[0036] (9)

[0037] Training phase, real future time series Similarly, the target-side prototype semantic supervision signal is obtained through function prototype matching. Therefore, the mapping relationship of LLM learning is: In the inference phase, LLM first generates semantic predictions of future function prototypes, and then reconstructs the prediction results into the final future time series prediction results.

[0038] Step 2, Alignment of two-dimensional features between multi-process semantics and time series data:

[0039] Using the current temporal semantic time window as a reference, the temporal correlation of upstream and downstream process information is calculated based on Gaussian distribution, candidate texts are selected and coarse-grained time windows are divided; then, the temporal offset network jointly evaluates the degree of temporal offset and semantic matching, selects the optimal matching segment, and achieves two-dimensional feature alignment between multi-process semantics and time series data through cross-attention and fusion with the current temporal semantics.

[0040] Step 2.1, Coarse-grained time window partitioning method based on Gaussian distribution:

[0041] With the current temporal semantics of the input Using a time window as a reference, Gaussian distribution is used to model the semantics of upstream and downstream processes relative to the current time sequence. The time offset is used to construct a candidate context set. Let the current temporal semantics be... The central time is The center time of the j-th candidate text is The data source is denoted as Define upstream and downstream process information relative to the current time sequence semantics. The time lag is The time correlation probability formula is defined for different data sources as follows:

[0042] (10)

[0043] The above formula uses a Gaussian distribution to measure the relationship between the j-th candidate text and the current temporal semantics. The degree of matching in time, among which, The larger the value, the stronger the temporal correlation between the two. and These represent upstream and downstream data sources, respectively. Relative to the current temporal semantics The average time delay and its fluctuation range are initialized by process experience and used as learnable parameters in training.

[0044] Based on the time correlation probability of each upstream and downstream process information We select the K upstream and downstream process information with the highest time correlation probability as candidate texts and form a candidate context set:

[0045] (11)

[0046] The above formula represents selecting K candidate texts from high to low probabilities of temporal relevance; where, Represented as current temporal semantics The m-th candidate text was selected;

[0047] Furthermore, based on the time offset distribution of the data source to which the candidate texts belong, a coarse-grained time window is assigned to the m-th candidate text:

[0048] (12)

[0049] in, This represents the m-th candidate text selected based on the probability of temporal relevance. This indicates the upstream and downstream data sources to which the candidate text belongs; and These represent the average time offset and the degree of fluctuation of the time offset of the data source of the upstream and downstream processes relative to the current time sequence semantics, respectively; This represents the range adjustment coefficient for the coarse-grained time window, used to control the expansion range of the time window on both sides of the average time offset; the above formula uses... Centered on the time window, with Define a coarse-grained time range corresponding to the m-th candidate text, using half the time window width. Then, organize the text information within this coarse-grained time range that belongs to the same upstream or downstream process as the candidate text in chronological order to form corresponding upstream and downstream process semantic fragments. ;

[0050] Step 2.2, Fine-grained semantic sample matching and fusion based on time window offset network:

[0051] From the candidate context set The algorithm identifies the semantic information of the process most relevant to the current temporal semantics, and constructs a time window offset network in the LLM native embedding layer. It adopts a lightweight MLP dual-head matching structure for jointly modeling coarse-grained time windows. Semantic fragments of upstream and downstream processes and current temporal sequence The semantic relevance and temporal offset between them are considered, and the semantics of the current time series are selected. The most relevant semantic fragments of upstream and downstream processes;

[0052] First, let's define the current temporal semantics. The semantic fragments of upstream and downstream processes in the coarse-grained time window are mapped to the LLM embedding space, introducing temporal position encoding. Process-aware semantic encoding To obtain the current temporal semantics The token sequence is represented as semantic fragments of the m-th upstream and downstream processes The token sequence is represented as ;in, , L s and L m These are the current temporal semantics. Semantic fragments related to the m-th upstream and downstream processes The number of tokens, where d is the dimension of the LLM hidden layer; For the m-th upstream and downstream process semantic fragment The initial embedding representation of the LLM; This represents the center moment of the i-th temporal semantics. This represents the center moment of the semantic segment of the m-th upstream and downstream process. The m-th upstream and downstream semantic segment represents the process semantic identifier; K represents the coarse-grained time window. The number of text fragments contained;

[0053] Subsequently, in order to achieve fine-grained fusion of semantic fragments from upstream and downstream processes with the current temporal semantics, , Discrete process information, including process type code, time location code, and production plan. Common input time window offset network; design time offset prediction for the time window offset network. and semantic matching score Output the expected time offset and semantic matching degree of the semantic fragments of the upstream and downstream processes respectively:

[0054] (13)

[0055] in, This indicates average pooling, used to obtain a fragment-level representation of the token sequence, and Subsequently, combined with the observation time offset Desired time offset Deviation in construction time consistency score and semantic matching score The fusion yields a comprehensive matching score. :

[0056] (14)

[0057] in, for exist The value of _t_ represents the historical industrial time series of the m-th retained candidate text relative to the i-th sample. Time offset; This refers to the time offset scale parameter; for and The balance coefficient; during the training phase, the comprehensive matching score, after softmax normalization, is used as the fusion weight for the candidate context. In the reasoning phase, select the candidate segment with the highest score. As the optimal matching segment:

[0058] (15)

[0059] in, The softmax temperature coefficient;

[0060] In obtaining the optimal matching fragment Then, the token sequence of the corresponding upstream and downstream process semantic segments is recorded as follows: To model the current temporal semantics token sequence and The local semantic correspondence between them is determined by cross-attention to fuse the current temporal semantics with the semantics of the matched upstream and downstream processes. The token sequence of the matched upstream and downstream process semantic fragments is adaptively aggregated by attention weights to obtain the semantically enhanced temporal representation.

[0061] (16)

[0062] in, , , Representing the l-th current temporal semantics respectively The token represents The token representation of the semantic fragment of the k-th upstream and downstream process Attention scores and relative positional bias between them; , and Let represent the learnable mapping matrices used to generate the query representation, key representation, and value representation, respectively; d represents the feature dimensions of the query representation and key representation, and the scaling factor. This is used to avoid excessive concentration of attention weights due to an excessively large dot product of the two; L m Indicates the length of the semantic fragment token sequence for the optimal matching upstream and downstream processes;

[0063] Step 3, Context-Aware Token Reconstruction for Multi-Process Operations:

[0064] After gating and fusing the time factors and process status factors in the multi-process context, the token sequence of the current temporal semantics is injected with cross-attention to obtain a context representation reflecting the upstream and downstream working conditions. This context representation is then mapped to scaling and bias parameters, and the token sequence is reconstructed with context awareness to form a context-aware representation that integrates the semantics of multiple processes and is used for subsequent prediction.

[0065] Step 3.1, Construct a context-prior and attention-guided encoder-decoder:

[0066] After step 2, the current temporal semantics of the i-th sample The token sequence is represented as ,exist The semantic fragments of upstream and downstream processes within the process construct a multi-process context token sequence. Based on this, a context-prior and attention-guided encoder-decoder is constructed; the encoder is used to apply coarse-grained time windows. Multi-process context token sequence The different working conditions are distinguished in the middle, among which Including time factor Process status factors Further construct a context-prior guided encoding process; in order to adaptively measure the time factor. Process status factors To determine the degree of influence of the current timing sequence, a gating mechanism is constructed in the encoder:

[0067] (17)

[0068] (18)

[0069] in, For gating weights, This represents element-wise multiplication. Representation layer normalization; To integrate the status information of upstream and downstream processes;

[0070] In order to obtain the current temporal semantics of the i-th sample token sequence representation Integrating upstream and downstream process status information decoder with For query, Use key-value pairs and generate contextual responses through cross-attention. In order to preserve the current temporal semantics The semantic structure of the token sequence, by... and A fusion mapping is performed to obtain a contextual representation reflecting different upstream and downstream operating conditions. :

[0071] (19)

[0072] in, , , ; , , It is a learnable linear mapping matrix, which is used for end-to-end training together with LN weights; , and They represent respectively by and The resulting query vector, key vector, and value vector are mapped.

[0073] Step 3.2, Context-Aware Token-Level Semantic Reconstruction:

[0074] After obtaining the contextual semantic representation oriented towards the current time sequence, in order to enhance the semantic contextual representation Further impact on current temporal semantics To update the token sequence, a context-aware token-level semantic reconstruction method is proposed;

[0075] The decoder further utilizes contextual representations For the current temporal semantics of the i-th sample token sequence representation Perform context-aware reconstruction; first, represent the context. Mapping to act on Scaling and bias parameters for each token:

[0076] (20)

[0077] The above equation represents the context through a linear mapping. Convert to refactoring parameters and Among them, W c and b c These represent the learnable weight matrix and bias vector of the context parameter mapping layer, respectively. Used for adjustment The scaling factor represented by each token in the table. Used to introduce the corresponding context bias information;

[0078] Subsequently, the temporal semantic representation is performed using the aforementioned scaling and bias parameters. Perform context-aware reconstruction:

[0079] (twenty one)

[0080] The above formula uses feature scaling, bias injection, and layer normalization operations to represent the context reflecting different upstream and downstream operating conditions. Information is incorporated into temporal semantic representation to obtain context-aware representation. ;

[0081] Furthermore, for In Its context-aware representation is written as:

[0082] (twenty two)

[0083] The above formula gives the token-level refactoring process, where, and These represent the scaling and bias parameters applied to the l-th token, respectively. This represents the number of tokens in the i-th sample.

[0084] The beneficial effects of this invention are as follows: This invention employs a function prototype matching method to semantically model industrial time series data, transforming the original continuous numerical sequence into a semantic parameter sequence containing fluctuation amplitude, fluctuation position, duration, waveform morphology, and drift trend, thereby improving data interpretability. It constructs a two-dimensional time series feature alignment method between material flow information and time, defining the time range and slicing samples of upstream and downstream process text information with the target time series window, enabling multi-process semantic information to form a unified data sample structure with industrial time series data. Through a gating fusion mechanism of time factors and process state factors, it obtains contextual prior features that integrate upstream and downstream process state information. Simultaneously, it adopts a context-attention-guided encoder-decoder structure to inject multi-process contextual semantics into the time series token reconstruction process, enhancing the large language model's understanding of production semantics, material flow correlation, and dynamic operating condition changes in industrial time series data, thus improving the stability, applicability, and prediction accuracy of the industrial time series prediction model. Attached Figure Description

[0085] Figure 1 This is a flowchart illustrating the application of the present invention.

[0086] Figure 2 The graphs are intermittent time series curves, where (a) represents intermittent industrial time series data, and (b) represents the prototype of the Sigmoid function and its generated semantic features.

[0087] Figure 3 The graphs are continuous time series curves, where (a) is continuous industrial time series data and (b) is the prototype of the Beta function and the semantic features generated.

[0088] Figure 4 A framework diagram for aligning semantic-temporal two-dimensional temporal features across multiple processes.

[0089] Figure 5 This is a diagram of the time-shifted network mechanism.

[0090] Figure 6 Diagram of context-aware modeling and token reconstruction mechanism.

[0091] Figure 7 The following is a comparison chart of the prediction results of intermittent time-series data (LDG converter gas recovery amount) in the example comparative experiment. (a) is a comparison of LDG converter gas recovery amount curves, and (b) is a comparison chart of RMSE for each sample.

[0092] Figure 8 The following is a comparison chart of the prediction results of intermittent time-series data (BFG blast furnace gas consumption) in the example comparative experiment. (a) is a comparison of BFG blast furnace gas consumption curves, and (b) is a comparison chart of RMSE for each sample.

[0093] Figure 9 The process diagram for reconstructing the time series based on the predicted semantic parameters of the function prototype is shown in (a) and (b) respectively. (a) is a comparison between the LDG converter gas recovery curve and the reconstructed curve, and (b) is a comparison between the LDG blast furnace gas consumption curve and the reconstructed curve. Detailed Implementation

[0094] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0095] Experiments were conducted using real production data from a large domestic steel enterprise, with validation performed using LDG recovery and BFG consumption datasets. Time-series data was collected from the SCADA system from November 2024 to October 2025, with a sampling interval of 1 minute. Text data was collected from the ERP system during the same period, covering upstream and downstream processes such as blast furnace tapping, torpedo ladle conveying, molten iron pretreatment, converter steelmaking, and continuous casting. Content included operation logs, production plans, and anomaly / fault reports. To achieve unified modeling across processes and time periods, the relevant records within each process's corresponding time period were aggregated and standardized using a "process-time period" organizational granularity, thus constructing an LLM input consisting of industrial time-series and upstream / downstream semantic text. The dataset was divided in a 7:1:2 ratio, with 70% used for training, 10% for validation, and 20% for testing, and no overlapping samples were found. The experiments were conducted on a platform equipped with an Intel Xeon Platinum 8352V processor and four NVIDIA A100 80GB GPUs. The model was implemented using PyTorch 2.5.1. To meet the computational requirements of training and inference for large language models, LoRA was used for efficient parameter fine-tuning.

[0096] First, the LDG recovery volume, BFG consumption volume, and other relevant process variables collected by the SCADA system are aligned according to a unified timestamp. For null values ​​caused by short-term communication interruptions or missing sampling, interpolation between adjacent valid sampling points or forward imputation are used. For segments with consecutive missing durations exceeding a preset threshold, they are removed from the training samples. For abnormal values ​​that clearly exceed the equipment's operating boundaries, they are screened out or corrected based on the change rate of adjacent sampling points, production logs, and equipment operating status.

[0097] Secondly, the text records in the ERP system and operation logs were standardized. Standardization included: unifying process names, equipment numbers, furnace numbers, steel grade numbers, shift numbers, time formats, and unit expressions; deleting duplicate records, invalid symbols, and templated fields irrelevant to the production process; and retaining key information characterizing the upstream and downstream material flow status, such as tapping time, molten iron arrival time, molten iron pretreatment start time, converter blowing time, converter tapping time, continuous casting batch information, production plan change information, and abnormal / fault descriptions.

[0098] Then, text records of the same process within the same or adjacent time periods are aggregated into process semantic fragments. Each process semantic fragment includes at least the process type, equipment number, start time, end time, event description, and status identifier. Through the above processing, the original numerical time series and unstructured text records are transformed into industrial time-series-text joint samples with a unified time index.

[0099] A large language model method for time series and multi-process semantic enhancement for industrial time series forecasting includes the following steps:

[0100] Step 1, Industrial Temporal Semantic Modeling:

[0101] Step 1.1, Function prototype construction:

[0102] For LDG recovery data with intermittent fluctuation characteristics, this embodiment uses a function prototype consisting of two Sigmoid functions to describe a single recovery process. The function prototype is constructed according to equation (1).

[0103] This function prototype is used to characterize the pulse amplitude, start position, duration, rise rate, and fall rate in the LDG recovery curve. For LDG recovery data, one pulse segment corresponds to the gas recovery stage in a single converter steelmaking process, and the low-value or zero-value segment between two adjacent pulses corresponds to the production interval stage.

[0104] For BFG consumption data with continuous fluctuation characteristics, this embodiment uses the normalized Beta function as the function prototype. The function prototype and its transformation form are constructed according to equations (2) and (3). This function prototype is used to describe the local peaks, valleys, waveform skew, baseline offset, and scale changes during the hot blast stove replacement process. For BFG consumption data, a local fluctuation window corresponds to a consumption change stage under the influence of a hot blast stove replacement or an adjacent replacement.

[0105] Step 1.2, Function prototype matching:

[0106] For LDG retrieval data, the function prototype shown in Equation (1) is shifted and scaled on the time axis to obtain candidate function prototypes for different starting positions and durations. Subsequently, the candidate function prototypes are matched with the historical time series of LDG, and the matching degree matrix is ​​obtained according to Equation (4). The elements in the matching degree matrix reflect the degree of similarity between the candidate function prototypes and the corresponding time segments.

[0107] After obtaining the matching degree matrix, local maxima are searched according to equations (5)-(7). Each local maximum corresponds to a production process that best matches the LDG recovery pulse. For a certain recovery process, the extracted semantic parameters include pulse amplitude, pulse start position, effective duration, rising edge variation parameters, falling edge variation parameters, and area semantic features. Among them, the area semantic features are obtained by integrating the function prototype within the corresponding interval and are used to represent the cumulative recovery intensity of this recovery process.

[0108] For BFG consumption data, the prototypes of the Beta function shown in equations (2) and (3) are matched with the local fluctuation windows in the historical time series of BFG to obtain the optimal semantic parameters corresponding to each local fluctuation process. For a certain BFG fluctuation process, the extracted semantic parameters include the start time and duration of the production process, the degree of waveform skewness and peak position, baseline offset, scale stretching parameter and area semantic features.

[0109] After function prototype matching, the original minute-level one-dimensional time series is transformed into a semantic parameter sequence composed of multiple production process semantic units. This semantic parameter sequence no longer directly represents time series segments in the form of equal-length patches, but instead represents industrial time series data as fluctuation units with production meaning, thus forming an enhanced industrial time series sample result.

[0110] Step 1.3, Construction with mixed inputs:

[0111] Based on the semantic parameters obtained from function prototype matching, a semantic hybrid input is constructed. For the i-th sample, the original input includes upstream and downstream process information (text-side information) and historical industrial time series, and the output is the true future time series within the future prediction window. The sample structure is consistent with the original prediction task organization method in the invention content, and the hybrid input is constructed according to equation (9), wherein the sample structure is based on... organize.

[0112] In practice, historical industrial time series are represented as temporal semantics after function prototype matching; textual information is represented as upstream and downstream process information after text cleaning, process identification, and time identification. Subsequently, the temporal semantics are input into the temporal semantic encoder and mapped to a temporal semantic embedding consistent with the hidden layer dimension of the large language model; the upstream and downstream process information is input into the native embedding layer of the large language model to obtain text features. .

[0113] Step 2, Alignment of two-dimensional features between multi-process semantics and time series data:

[0114] Step 2.1, Coarse-grained time window partitioning method based on Gaussian distribution:

[0115] Using the current time sequence semantics as a reference, the time range of upstream and downstream process information is defined. Specifically, the current time sequence semantics time window is assumed to have a center time and a window length. For data sources from different processes, the corresponding average time lag parameter and fluctuation range parameter are set respectively. For different processes such as blast furnace tapping, torpedo ladle conveying, molten iron pretreatment, converter steelmaking and continuous casting, since there are different sequential relationships between them and the current time sequence, the time correlation relationship of upstream and downstream process information is calculated according to Equations (10)-(12) respectively, and the coarse-grained time window division is completed. Candidate process semantic segments related to the current time sequence semantics time window are selected.

[0116] In the LDG recovery task, the candidate context set includes records such as hot metal pretreatment before and after the current converter recovery stage, converter steelmaking operation records, tapping plans, furnace switching records, and anomaly records. In the BFG consumption task, the candidate context set includes hot blast stove switching records, blast furnace tapping rhythm, gas pipeline scheduling records, energy medium balance records, and equipment anomaly information. Through this step, upstream and downstream process information scattered across different systems and time ranges can be initially categorized into the candidate context set of the current time-series semantic time window.

[0117] Step 2.2, Fine-grained semantic sample matching and fusion based on time window offset network:

[0118] Based on the candidate context set, a time window offset network is further employed for fine-grained matching. Specifically, the current temporal semantics are... The token sequence representation and the semantic fragment of the m-th upstream and downstream process. The token sequence representation, process type encoding, time location encoding, and discrete process information of the production plan are jointly input into the time window offset network. The time window offset network adopts a lightweight MLP dual-head matching structure, including a time offset prediction head and a semantic matching scoring head.

[0119] The time offset prediction head is used to output the expected time offset of the semantic segments of upstream and downstream processes relative to the time window of the current temporal semantics; the semantic matching scoring head is used to output the semantic correlation between the semantic segments of upstream and downstream processes and the time window of the current temporal semantics. Subsequently, the time offset prediction and semantic matching score are obtained according to Equation (13), and the time consistency score and semantic matching score are fused according to Equation (14) to form a comprehensive matching score. During the training phase, the comprehensive matching score is normalized according to Equation (15) to obtain the fusion weight of the semantic segments of upstream and downstream processes; during the inference phase, the candidate segment with the highest comprehensive matching score is selected. As the optimal matching fragment.

[0120] Through the above processing, the same time window no longer mechanically matches text within the same time period, but matches semantic fragments of upstream and downstream processes with actual production connections based on the material flow transmission relationship, time offset relationship and semantic correlation relationship of different processes.

[0121] After obtaining the optimal matching segment, the token of the current temporal semantics is used as the query term, and the tokens of the matched upstream and downstream process semantic segments are used as the key and value. The current temporal semantics and the upstream and downstream process semantic segments are fused through the cross-attention mechanism. The fusion process is performed according to Equation (16).

[0122] In practical implementation, cross-attention not only calculates the correlation between semantic tokens but also introduces relative position bias to preserve the temporal sequence relationship between upstream and downstream process semantic fragments and the current temporal semantic time window. For LDG recovery tasks, this mechanism can correlate upstream and downstream events such as molten iron arrival, converter blowing start, recovery start, and recovery end with LDG recovery pulses; for BFG consumption tasks, this mechanism can correlate hot blast stove changing records, BFG network status, and blast furnace production rhythm with BFG consumption fluctuations.

[0123] After cross-attention fusion, a semantically enhanced temporal representation is obtained. This temporal representation simultaneously contains the functional prototype semantic features of the current industrial time series, as well as semantic fragments of upstream and downstream processes that have temporal constraints and material flow relationships with it.

[0124] Step 3, Context-Aware Token Reconstruction for Multi-Process Operations:

[0125] Step 3.1, Construct a context-prior and attention-guided encoder-decoder:

[0126] After aligning the two-dimensional features between the multi-process semantics and time-series data, a multi-process context token sequence corresponding to the current sample is constructed based on the upstream and downstream process semantic fragments obtained from coarse-grained segmentation. Specifically, for any token in the multi-process context fragment token sequence, it includes a time factor and a process status factor. The time factor is used to characterize the time interval between the upstream and downstream processes and the target process, the material flow transfer time, or the process response delay; the process status factor is used to characterize the operating status, material conditions, equipment status, or quality parameters of the corresponding process. Taking the converter steelmaking process as an example, the time factor corresponds to the time interval information between the processes such as blast furnace tapping, torpedo ladle conveying, and molten iron pretreatment and the converter steelmaking process; the process status factor corresponds to the operating status of each process, molten iron temperature, molten iron composition, equipment status, and abnormal records.

[0127] To integrate time and process status factors, a context-prior-guided encoding process is constructed. For the i-th sample, the time factor representation is obtained. and process status factors In order to adaptively measure the influence of time factors and process state factors on the current temporal semantics, a gating mechanism is set in the encoder. The calculation process of the gating weight is shown in Equation (17).

[0128] Based on gating weights Representing the time factor and process status factors Weighted fusion is performed to obtain the upstream and downstream process status information fused by the fusion of time factor and process status factor. The calculation process is shown in equation (18). Through this step, the weights of time factors and state factors in the contextual semantic representation can be adaptively adjusted according to the process status and time delay of different samples.

[0129] After obtaining the status information of the upstream and downstream processes. Then, it is injected into the current temporal semantics. Specifically, it is represented by the token sequence of the current temporal semantics obtained after function prototype matching and two-dimensional feature alignment. As a decoder query, it integrates the status information of upstream and downstream processes. As keys and values, contextual response representations are generated through cross-attention. The calculation process is shown in equation (19).

[0130] To preserve the original semantic structure of the token sequence in the current temporal semantics, the context response representation is... The token sequence with current temporal semantics A fusion mapping is performed, and a contextual representation reflecting different upstream and downstream operating conditions is obtained through layer normalization. The calculation process is shown in equation (19). This representation simultaneously includes the semantic features of the current target time sequence, the time factors of upstream and downstream processes, and the process status factors.

[0131] Step 3.2, Context-Aware Token-Level Semantic Reconstruction:

[0132] After obtaining the context fusion representation, it is mapped to the reconstruction parameters of the token acting on the current temporal semantics. The reconstruction parameters are obtained according to Equation (20) and include scaling and bias terms. Subsequently, the current temporal token is updated in a context-aware manner according to Equation (21), and the expanded update form of the l-th token is obtained according to Equation (22).

[0133] Specifically, for time series segments with similar forms but different operating conditions, context-aware token-level semantic reconstruction can distinguish their token representations based on the semantic segments of upstream and downstream processes. For example, two LDG recovery pulses may be similar in amplitude and duration, but if their corresponding molten iron arrival time, converter start-up time, or anomaly records are different, their subsequent recovery rhythms may differ; two BFG consumption fluctuations may be similar in form, but if their corresponding hot blast stove changing sequence, gas pipeline status, or blast furnace production rhythm are different, their subsequent consumption trends may also differ. Through token-level semantic reconstruction, the model can inject information reflecting different upstream and downstream operating conditions into the current time series semantics.

[0134] After token-level reconstruction, the final input to the backbone of the large language model is not the original temporal embedding, but a context-aware embedding that integrates the semantic information from multiple processes. The large language model is trained and inferred based on this embedding, and outputs the semantic parameter prediction results within the future prediction window.

[0135] To verify the effectiveness of the proposed method, several representative recent temporal prediction models were selected for comparison, including deep learning-based models DLinear and PatchTST, traditional machine learning methods LS-SVM, and large language model-based methods LLM-Lora, Time-LLM, and CALF. These methods cover typical technical routes such as linear modeling, patch-based temporal modeling, traditional deep learning, lightweight large model adaptation, and cross-modal temporal prediction. The method is evaluated from two perspectives: prediction accuracy and inference effectiveness. For prediction accuracy, root mean square error (RMSE) and mean absolute error (MAE) are used to assess the model's temporal fitting ability. For inference effectiveness, textual contribution gain (TCG) is introduced from the perspective of textual semantic effectiveness; acceptable interval gain (AIG) is introduced to measure the improvement in the acceptability of textual information for estimating key state variables. Let Y... i For the i-th sample, and The above metrics represent the prediction results of the full model and the text-free variant, respectively, and are defined as follows:

[0136] (twenty three)

[0137] (twenty four)

[0138] in, For the RMSE function; The key state variable extraction function is used, and the experiment uses the mean and temporal semantic feature parameters; The acceptable relative error threshold was set at 10% in the experiment. The larger the TCG, the more significant the improvement in prediction error by textual information; the larger the AIG, the more significant the increase in the proportion of key state quantity estimates falling within the acceptable range due to textual information.

[0139] Figure 7 Visualizations of the small and large models for the LDG recycling task are presented. Figure 7 As shown in (a), while the small model can fit the pulse pattern of LDG recovery, it generally exhibits errors at peak boundaries and the switching positions between recovery and intermittent phases. DLinear tends to lag at abrupt changes, and while PatchTST can recover local peak shapes, its predictive stability remains insufficient when peak width and interpeak intervals vary significantly. In contrast, the large model baseline performs better in determining the presence or absence of pulses, but it is still not accurate enough in terms of peak start and end positions and local duration. Time-LLM is more prone to rhythm shifts, while CALF is generally more stable, but it still has deviations in characterizing local boundaries and details. This embodiment can more accurately fit the start and end boundaries, peak width, and intermittent intervals of the real curve on the test samples. Figure 7 (b) further shows that the RMSE of this embodiment on the 2# LDG test sample is 72.90, and the RMSE on 1#, 2# and 3# LDGs is significantly better than that of the comparison models. Table 1 shows that this embodiment achieves the best RMSE and MAE on 1#, 2# and 3# LDGs.

[0140] Regarding BFG data, Figure 8 As shown in (a), PatchTST performs well in the small model, indicating that the block modeling approach helps capture recurring local variations, but errors still exist in the depth of valleys and the location of inflection points. While DLinear and LS-SVM can follow the overall changes well, their characterization of local valleys and periodic transitions is not accurate enough. The baseline of the large model generally reflects the periodic fluctuations, with CALF being relatively more stable, indicating that the introduction of textual and historical information helps to recover periodic variations, but deviations still exist in the amplitude of local peaks and valleys and the transitions between adjacent periods. In contrast, this embodiment not only maintains the overall periodic shape better but also more accurately matches the locations of local peaks and valleys and disturbances. Figure 8 (b) further illustrates that this embodiment exhibits superior RMSE on the test set. Table 2 shows that the performance gap between models in the BFG task is smaller than that in the LDG task, but this embodiment achieves the best MAE on all three sets of BFG data, and obtains the best RMSE on BFG 2# and BFG 3#. This demonstrates that the method proposed in this embodiment can maintain good accuracy on data generated by continuously running equipment.

[0141] Table 3 shows that this embodiment achieves the highest TCG on all six datasets, significantly outperforming the LLM-based model. This indicates that it can more fully utilize the semantics of upstream and downstream processes on the text side to reduce prediction errors. Regarding the AIG metric, except for a slight difference compared to Time-LLM in 1#BFG, it achieves the best results on all other datasets, indicating that the prediction results are more likely to fall within the acceptable error range. A possible reason is that this embodiment does not use text as a simple auxiliary representation, but rather embeds upstream and downstream semantic information into the temporal modeling process through semantic-temporal alignment and temporal token reconstruction, making the prediction results subject to both process semantics and numerical constraints. In contrast, LLM-Lora and Time-LLM primarily transfer language model capabilities to temporal prediction, making it difficult to translate stable process semantics into effective constraints on temporal representations. While CALF possesses cross-modal fusion capabilities, it focuses more on inter-modal feature fusion and is insufficient in modeling the temporal sequence relationships and uncertainties between multiple processes. Therefore, this embodiment demonstrates more stable semantic inference effectiveness.

[0142] Table 1: Statistics of LDG recovery prediction indicators in comparative experiments

[0143]

[0144] Table 2: Statistics of BFG Consumption Prediction Indicators in Comparative Experiments

[0145]

[0146] Table 3: Comparison of Semantic Reasoning Performance Evaluation Metrics Based on Large Language Models (LLM)

[0147]

Claims

1. A large language model method for time series and multi-process semantic enhancement for industrial time series forecasting, characterized in that, The steps are as follows: Step 1, Industrial Temporal Semantic Modeling: Based on the waveform characteristics of industrial time series data, semantic modeling of intermittent and continuous industrial time series data is carried out using the prototype of double Sigmoid function and the prototype of normalized Beta function, respectively. Semantic parameters of each production semantic process are extracted by sliding matching. Industrial time series data includes shape, location and area. Semantic parameters are jointly encoded with upstream and downstream process information and task instructions to construct a semantic hybrid input for prediction tasks. Step 2, Alignment of two-dimensional features between multi-process semantics and time series data: Using the current temporal semantic time window as a reference, the temporal correlation of upstream and downstream process information is calculated based on Gaussian distribution, candidate texts are selected and coarse-grained time windows are divided; then, the temporal offset network jointly evaluates the degree of temporal offset and semantic matching, selects the optimal matching segment, and achieves two-dimensional feature alignment between multi-process semantics and time series data through cross-attention and fusion with the current temporal semantics. Step 3, Context-Aware Token Reconstruction for Multi-Process Operations: After gating and fusing the time factors and process status factors in the multi-process context, the token sequence of the current temporal semantics is injected with cross-attention to obtain a context representation reflecting the upstream and downstream working conditions. This context representation is then mapped to scaling and bias parameters, and the token sequence is reconstructed with context awareness to form a context-aware representation that integrates the semantics of multiple processes and is used for subsequent prediction.

2. The large language model method for time series and multi-process semantic enhancement for industrial time series prediction according to claim 1, characterized in that, The specific implementation process of step 1 is as follows: Step 1.1, Function prototype construction: Based on the waveform characteristics of industrial time series data, corresponding function prototypes are selected to perform semantic modeling of industrial time series data. For industrial time series data with intermittent fluctuation characteristics, a function prototype consisting of two sigmoid functions is used to describe the fluctuation segments: in, This is the set of parameters for the function prototype; This represents the pulse amplitude, indicating the maximum value of a waveform during a production process. Indicates the pulse start position, used to determine the start time of the production process; This indicates the effective width of the high-level maintenance phase, reflecting the duration of a single production process; and Control the rate of change of the rising edge and the falling edge separately; The standard sigmoid activation function; t represents the discrete sampling time in the industrial time series data; For industrial time series data with continuous fluctuation characteristics, in order to characterize the local fluctuation features in the industrial time series data, the normalized Beta function is selected as the function prototype, and the formula is as follows: in, s and l represent the start time and duration of the production process, respectively; and The waveform skewness and peak position are controlled respectively; k represents the baseline offset; c represents the scale stretching parameter, which is used to adjust the intensity of local fluctuation characteristics; u=(ts) / l represents the normalized local time variable, which is used to map the interval [s,s+l] to the interval [0,1]; Step 1.2, Function prototype matching: Further, the function prototype and Unified as Then combine it with industrial time series data Performing a point-by-point sliding matching operation along the time dimension yields the matching degree matrix: in, To calculate the matching degree using vector dot product, T represents the length of the industrial time series data; Since industrial time series data contains multiple production semantic processes corresponding to function prototypes, the corresponding window semantic parameters are obtained by searching for local maxima points in the matching degree matrix; the search process for local maxima points is as follows: in, The function represents the degree of matching between the function prototype and the industrial time series data; ρ represents the parameter set of the function prototype excluding... The parameter vector other than A, w, λ, μ}; Indicates in When fixed, make the matching degree The combination of parameters of the function prototype that achieves the maximum value; This represents the one-dimensional matching degree sequence obtained by projecting the matching degree matrix along the parameter vector ρ of the candidate prototype to its maximum value. This indicates the initial position of the nth local maximum point; Indicated by Centered on, with The local search region is defined by the neighborhood radius; This indicates the time position parameter where the matching degree reaches its maximum value within the local search region; In chronological order, each local maxima in the one-dimensional matching degree sequence corresponds to a production semantic process. Therefore, the optimal matching window determined by the nth local maximum corresponds to the nth production semantic process. Substitution The optimal shape parameter vector corresponding to the nth production semantic process is obtained: in, and These represent the optimal pulse amplitude and optimal effective width for the nth production semantic process, respectively. and These represent the optimal parameters for controlling the rate of change of its rising and falling edges, respectively. Therefore, in industrial time series data with intermittent fluctuation characteristics, the parameter set of the optimal function prototype corresponding to the nth production semantic process is determined by its temporal location parameters. With the optimal shape parameter vector Common characteristics are Furthermore, since the prototype function is integrated over the interval corresponding to the continuous waveform during the semantic production process, the area semantic feature is obtained as follows: Therefore, the semantic parameter set of industrial time series data with intermittent fluctuation characteristics is as follows: Similarly, in industrial time series data with continuous fluctuation characteristics, the parameter set of the optimal function prototype corresponding to the nth production semantic process is: Its corresponding area semantic feature is The final set of semantic parameters for industrial time series data with continuous fluctuation characteristics is: ; Step 1.3, Mixed Input Construction: The original input of the i-th sample consists of upstream and downstream process information. and historical industrial timeline Composition, output as a future time window Real Future Timeline Therefore, the original prediction task is represented as ; Based on the number of local maxima in the matching degree matrix, historical industrial time series Represented as the current temporal semantics, consisting of a set of semantic parameters corresponding to multiple production semantic processes. , where N i This represents the number of semantic segments; subsequently, the current temporal semantics are... The semantic parameters corresponding to each production semantic process in the process are processed by a temporal semantic encoder. Mapped to temporal semantic embeddings consistent with the dimensions of the LLM hidden layers. Meanwhile, information on upstream and downstream processes. Text features are obtained through the native embedding layer of LLM. To make the LLM clearly define the original prediction task, a prompt word for the original prediction task is constructed based on the prediction objective: "Predict next". Windows: And it is encoded into a task feature sequence through an LLM native embedding layer. Finally, the semantic mixed input of LLM is represented as: Training phase, real future time series Similarly, the target-side prototype semantic supervision signal is obtained through function prototype matching. Therefore, the mapping relationship of LLM learning is: In the inference phase, LLM first generates a prototype semantic prediction of the future function, and then reconstructs the prediction results into the final future time series prediction results.

3. The large language model method for time series and multi-process semantic enhancement for industrial time series prediction according to claim 2, characterized in that, The specific implementation process of step 2 is as follows: Step 2.1, Coarse-grained time window partitioning method based on Gaussian distribution: With the current temporal semantics of the input Using a time window as a reference, Gaussian distribution is used to model the semantics of upstream and downstream processes relative to the current time sequence. The time offset is used to construct a candidate context set. Let the current temporal semantics be... The central time is The center time of the j-th candidate text is The data source is denoted as Define upstream and downstream process information relative to the current time sequence semantics. The time lag is The time correlation probability formula is defined for different data sources as follows: The above formula uses a Gaussian distribution to measure the relationship between the j-th candidate text and the current temporal semantics. The degree of matching in time, among which, The larger the value, the stronger the temporal correlation between the two. and These represent upstream and downstream data sources, respectively. Relative to the current temporal semantics The average time delay and its fluctuation range are initialized by process experience and used as learnable parameters in training. Based on the time correlation probability of each upstream and downstream process information We select the K upstream and downstream process information with the highest time correlation probability as candidate texts and form a candidate context set: The above formula represents selecting K candidate texts from high to low probabilities of temporal relevance; where, Represented as current temporal semantics The m-th candidate text was selected; Furthermore, based on the time offset distribution of the data source to which the candidate texts belong, a coarse-grained time window is assigned to the m-th candidate text: in, This represents the m-th candidate text selected based on the probability of temporal relevance. This indicates the upstream and downstream data sources to which the candidate text belongs; and These represent the average time offset and the degree of fluctuation of the time offset of the data source of the upstream and downstream processes relative to the current time sequence semantics, respectively; This represents the range adjustment coefficient for the coarse-grained time window, used to control the expansion range of the time window on both sides of the average time offset; the above formula uses... Centered on the time window, with Define a coarse-grained time range corresponding to the m-th candidate text, using half the time window width. Then, organize the text information within this coarse-grained time range that belongs to the same upstream or downstream process as the candidate text in chronological order to form corresponding upstream and downstream process semantic fragments. ; Step 2.2, Fine-grained semantic sample matching and fusion based on time window offset network: From the candidate context set The algorithm identifies the semantic information of the process most relevant to the current temporal semantics, and constructs a time window offset network in the LLM native embedding layer. It adopts a lightweight MLP dual-head matching structure for jointly modeling coarse-grained time windows. Semantic fragments of upstream and downstream processes and current temporal sequence The semantic relevance and temporal offset between them are considered, and the semantics of the current time series are selected. The most relevant semantic fragments of upstream and downstream processes; First, let's define the current temporal semantics. The semantic fragments of upstream and downstream processes in the coarse-grained time window are mapped to the LLM embedding space, introducing temporal position encoding. Process-aware semantic encoding To obtain the current temporal semantics The token sequence is represented as semantic fragments of the m-th upstream and downstream processes The token sequence is represented as ;in, , L s and L m These are the current temporal semantics. Semantic fragments related to the m-th upstream and downstream processes The number of tokens, where d is the dimension of the LLM hidden layer; For the m-th upstream and downstream process semantic fragment The initial embedding representation of the LLM; This represents the center moment of the i-th temporal semantics. This represents the center moment of the semantic segment of the m-th upstream and downstream process. The m-th upstream and downstream process semantic segment represents the process semantic identifier; K represents the coarse-grained time window. The number of text fragments contained; Subsequently, in order to achieve fine-grained fusion of semantic fragments from upstream and downstream processes with the current temporal semantics, , Discrete process information, including process type code, time location code, and production plan. Common input time window offset network; design time offset prediction for the time window offset network. and semantic matching score Output the expected time offset and semantic matching degree of the semantic fragments of the upstream and downstream processes respectively: in, This indicates average pooling, used to obtain a fragment-level representation of the token sequence, and Subsequently, combined with the observation time offset Desired time offset Deviation Construction Time Consistency Score and semantic matching score The fusion yields a comprehensive matching score. : in, for exist The value of _t_ represents the historical industrial time series of the m-th retained candidate text relative to the i-th sample. Time offset; This refers to the time offset scale parameter; for and The balance coefficient; during the training phase, the comprehensive matching score, after softmax normalization, is used as the fusion weight for the candidate context. In the reasoning phase, select the candidate segment with the highest score. As the optimal matching segment: in, The softmax temperature coefficient; In obtaining the optimal matching fragment Then, the token sequence of the corresponding upstream and downstream process semantic segments is recorded as follows: To model the current temporal semantics token sequence and The local semantic correspondence between them is determined by cross-attention to fuse the current temporal semantics with the semantics of the matched upstream and downstream processes. The token sequence of the matched upstream and downstream process semantic fragments is adaptively aggregated by attention weights to obtain the semantically enhanced temporal representation. in, , , Representing the l-th current temporal semantics respectively The token represents The token representation of the semantic fragment of the k-th upstream and downstream process Attention scores and relative positional bias between them; , and Let represent the learnable mapping matrices used to generate the query representation, key representation, and value representation, respectively; d represents the feature dimensions of the query representation and key representation, and the scaling factor. This is used to avoid excessive concentration of attention weights due to an excessively large dot product of the two; L m This indicates the length of the semantic fragment token sequence for the optimal matching upstream and downstream processes.

4. The large language model method for time series and multi-process semantic enhancement for industrial time series prediction according to claim 3, characterized in that, The specific implementation process of step 3 is as follows: Step 3.1, Construct a context-prior and attention-guided encoder-decoder: After step 2, the current temporal semantics of the i-th sample The token sequence is represented as ,exist The semantic fragments of upstream and downstream processes within the process construct a multi-process context token sequence. Based on this, a context-prior and attention-guided encoder-decoder is constructed; the encoder is used to apply coarse-grained time windows. Multi-process context token sequence The different working conditions are distinguished in the middle, among which Including time factor Process status factors Further construct a context-prior guided encoding process; in order to adaptively measure the time factor. Process status factors To determine the degree of influence of the current timing sequence, a gating mechanism is constructed in the encoder: in, For gating weights, This represents element-wise multiplication. Representation layer normalization; To integrate the status information of upstream and downstream processes; In order to obtain the current temporal semantics of the i-th sample token sequence representation Integrating upstream and downstream process status information decoder with For query, Use key-value pairs and generate contextual responses through cross-attention. In order to preserve the current temporal semantics The semantic structure of the token sequence, by... and A fusion mapping is performed to obtain a contextual representation reflecting different upstream and downstream operating conditions. : in, , , ; , , It is a learnable linear mapping matrix, which is used for end-to-end training together with LN weights; , and They represent respectively by and The resulting query vector, key vector, and value vector are mapped. Step 3.2, Context-Aware Token-Level Semantic Reconstruction: After obtaining the contextual semantic representation oriented towards the current time sequence, in order to enhance the semantic contextual representation Further impact on current temporal semantics To update the token sequence, a context-aware token-level semantic reconstruction method is proposed; The decoder further utilizes contextual representations For the current temporal semantics of the i-th sample token sequence representation Perform context-aware reconstruction; first, represent the context... Mapping to act on Scaling and bias parameters for each token: The above equation represents the context through a linear mapping. Convert to refactoring parameters and Among them, W c and b c These represent the learnable weight matrix and bias vector of the context parameter mapping layer, respectively; Used for adjustment The scaling factor represented by each token in the table. Used to introduce the corresponding context bias information; Subsequently, the temporal semantic representation is performed using the aforementioned scaling and bias parameters. Perform context-aware reconstruction: The above formula uses feature scaling, bias injection, and layer normalization operations to represent the context reflecting different upstream and downstream operating conditions. Information is incorporated into temporal semantic representation to obtain context-aware representation. ; Furthermore, for In Its context-aware representation is written as: The above formula gives the token-level refactoring process, where, and These represent the scaling and bias parameters applied to the l-th token, respectively. This represents the number of tokens in the i-th sample.

Images