An industrial decision support system and method based on large models and artificial intelligence

Abstract

The application discloses an industrial decision support system and method based on a large model and artificial intelligence, and belongs to the technical field of industrial intelligent decision-making.The system comprises the following steps: collecting industrial field data streams in real time and preprocessing to generate a structured time series dataset; using an industrial state analysis large model to extract a characteristic vector sequence, combining a sliding time window division and a neural Granger causality test algorithm to generate a causality knowledge rule set; using a neural differential equation framework to predict equipment state data prediction values; performing risk benefit assessment based on the prediction values to generate a structured decision instruction; after executing the instruction, comparing the prediction values and measured value deviations by using a sliding window cumulative error statistical method to dynamically optimize the model; and through the degenerative modeling and dynamic prediction of the causality rule constraint, and the closed-loop verification of the sliding window cumulative error statistical method, high-precision prediction of the equipment state, risk tracing and model self-adaptive optimization are realized.

Description

Technical Field

[0001] This invention belongs to the field of industrial intelligent decision-making technology, and in particular, it is an industrial decision support system and method based on large models and artificial intelligence. Background Technology

[0002] Industrial decision support systems (ICS) are a core technology in the field of intelligent manufacturing. They aim to achieve real-time monitoring and optimization decisions for the production process by integrating the IoT sensing layer, the industrial internet platform layer, and the application service layer. Current mainstream systems rely on big data analytics and employ machine learning predictive models for fault warnings and energy efficiency optimization. Simultaneously, rule engines and operations research algorithms are applied to resource allocation and risk response strategy generation, forming a closed-loop architecture of "perception-analysis-decision," significantly improving equipment reliability and production scheduling efficiency.

[0003] However, existing methods still have two limitations: the predictive models lack dynamic adaptability, equipment degradation modeling relies heavily on static historical data for training, making it difficult to integrate causal drift caused by changes in operating conditions in real time, resulting in the accumulation of prediction bias as the dynamics of operating conditions intensify; the depth of causal reasoning is limited, traditional methods do not fully mine the implicit causal chains in multimodal industrial data, especially when dealing with nonlinear coupling parameters such as vibration signals and temperature fluctuations, it is difficult to construct an interpretable set of physical rules, which restricts the accuracy and traceability of decision-making. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an industrial decision support system and method based on large models and artificial intelligence, which solves the problems of insufficient dynamic adaptability of prediction models and insufficient mining of implicit causal chains in multimodal industrial data.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an industrial decision support method based on large models and artificial intelligence, comprising the following steps:

[0008] S1. Real-time acquisition of industrial field data streams and preprocessing to generate structured time-series datasets;

[0009] S2. Analyze the structured time series dataset using the large industrial state analysis model to obtain the industrial state feature vector sequence. Divide the industrial state feature vector sequence using a sliding time window and perform causal verification to generate a causal knowledge rule set.

[0010] S3. Predict the predicted values ​​of equipment status data using a neural differential equation framework;

[0011] S4. Based on the predicted values ​​of equipment status data, use the large industrial status analysis model to conduct risk and benefit assessment and generate structured decision instructions.

[0012] S5. Execute structured decision instructions, collect real-time measured values ​​of equipment status data, and perform deviation analysis and evaluation of the execution effect between predicted and measured values ​​of equipment status data, while dynamically optimizing the neural differential equation framework.

[0013] Furthermore, in S1, the industrial field data stream includes equipment fault data, equipment vibration signal data, equipment temperature readings, and equipment pressure fluctuation data;

[0014] Data cleaning and standardization are performed on industrial field data streams to generate structured time-series datasets.

[0015] Furthermore, in S2, the structured time-series dataset is analyzed using a large-scale industrial state analysis model to obtain a sequence of industrial state feature vectors. This sequence is then divided using a sliding time window, including:

[0016] A large-scale industrial state analysis model is constructed based on structured time-series datasets;

[0017] The industrial state analysis model is used to extract features from the structured time series dataset to obtain the industrial state feature vector sequence.

[0018] A sliding time window divider is used to divide the industrial state feature vector sequence into time window data segments of fixed duration.

[0019] Furthermore, a neural Granger causality test algorithm is used to generate a set of causal knowledge rules, including:

[0020] Based on the time window data segments, causal tests are performed to obtain candidate causal pairs;

[0021] The obtained candidate causal pairs are subjected to time window continuity verification to obtain a stable causal chain;

[0022] Based on stable causal chains, a set of causal knowledge rules is obtained through a rule template matching strategy.

[0023] Furthermore, S3 specifically refers to:

[0024] A neural differential equation framework is used to embed a set of causal knowledge rules as physical constraints into the differential equation structure;

[0025] Based on real-time acquired industrial field data streams and a neural differential equation framework, the time derivative of the equipment status data at the current moment is calculated, and integral calculations are performed to generate predicted values ​​of the equipment status data.

[0026] Furthermore, S4 specifically refers to:

[0027] Based on equipment status data predictions, identify potential failure risks;

[0028] Based on potential failure risks, a quantified failure risk is generated using a dynamic risk propagation algorithm based on a set of causal knowledge rules.

[0029] Through the task decision and optimization layer of the industrial condition analysis big model, decision analysis is performed on the quantified fault risks to generate structured decision instructions.

[0030] Furthermore, S5 specifically refers to:

[0031] Execute structured decision instructions and simultaneously collect measured values ​​of device status data with timestamps;

[0032] Align the timestamps of the measured equipment status data with the timestamps of the predicted equipment status data to generate a measured time series dataset;

[0033] Based on the measured time series dataset and the predicted values ​​of the status data, the root mean square error and the mean absolute percentage error of the predicted values ​​of the equipment status data and the measured values ​​of the equipment status data within the same time window are calculated to generate the prediction deviation value of the equipment status data.

[0034] Based on the predicted deviation value of the equipment status data, identify the abnormal deviation time window in the measured time series dataset, and locate the abnormal deviation source by combining the equipment topology network, and generate deviation source location data.

[0035] The causal knowledge rules are used to analyze the deviation contribution of the deviation source location data to generate a deviation contribution quantification matrix.

[0036] Based on the deviation contribution metric matrix, an incremental learning mechanism is used to update the neural differential equation framework.

[0037] Secondly, the present invention provides an industrial decision support system based on large models and artificial intelligence, including a data acquisition and preprocessing module, an industrial status analysis module, an equipment degradation prediction module, a decision generation module, and a decision execution and verification module;

[0038] The data acquisition and preprocessing module is used to acquire industrial field data streams in real time and preprocess them to generate structured time-series datasets.

[0039] The industrial state analysis module is used to analyze structured time series datasets using a large industrial state analysis model to obtain industrial state feature vector sequences. It divides the industrial state feature vector sequences using a sliding time window and performs causal tests to generate a set of causal knowledge rules.

[0040] The equipment degradation prediction module is used to predict equipment status data using a neural differential equation framework.

[0041] The decision generation module is used to generate structured decision instructions based on the predicted values ​​of equipment status data and the risk-benefit assessment using a large industrial status analysis model.

[0042] The decision execution and verification module is used to execute structured decision instructions, collect real-time measured values ​​of equipment status data, and perform deviation analysis between predicted and measured values ​​of equipment status data to evaluate the execution effect. At the same time, it dynamically optimizes the neural differential equation framework.

[0043] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the industrial decision support method based on large models and artificial intelligence as described in the first aspect of the present invention.

[0044] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the industrial decision support method based on large models and artificial intelligence as described in the first aspect of the present invention.

[0045] The beneficial effects of this invention are as follows: This invention organically combines the degradation modeling of causal rule constraints, dynamic prediction, and closed-loop verification using the sliding window cumulative error statistical method, as well as the deep coupling and local technological innovation of each module: At the degradation modeling level, causal knowledge rules are embedded as physical constraints into neural differential equations to achieve the fusion of mechanism and data; at the dynamic prediction level, prediction is driven by the "direction-intensity-delay" triplet based on the causal chain, improving the adaptability to operating condition drift; at the closed-loop verification level, through deviation source localization, causal contribution matrix, and incremental learning mechanism, the model achieves self-explanation, self-diagnosis, and self-optimization, ultimately forming an autonomous decision-making closed loop of "perception-prediction-decision-verification," comprehensively improving the reliability, decision-making efficiency, and dynamic environmental adaptability of industrial systems. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 A flowchart illustrating an industrial decision support method based on large models and artificial intelligence;

[0048] Figure 2 This is a schematic diagram of an industrial decision support system based on large models and artificial intelligence.

[0049] Figure 3 A flowchart for generating a causal knowledge rule set for a neural Granger causality test algorithm;

[0050] Figure 4 This is a flowchart of the sliding window cumulative error statistical method. Detailed Implementation

[0051] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0052] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0053] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0054] Reference Figures 1-4 This is one embodiment of the present invention, which provides an industrial decision support method based on large models and artificial intelligence, including the following steps:

[0055] S1. Real-time acquisition of industrial field data streams and preprocessing to generate structured time-series datasets;

[0056] Industrial field data streams include equipment fault data, equipment vibration signal data, equipment temperature readings, and equipment pressure fluctuation data;

[0057] In industrial environments, vibration sensors are deployed directly onto the surface of mechanical equipment to capture vibration signal data in real time and continuously output vibration signals; temperature sensors are installed in close contact with the heat source area of ​​the equipment to continuously measure the temperature readings; pressure sensors are configured to connect to the pressure port of the equipment to continuously monitor the pressure fluctuation data; and equipment fault data is obtained by recording equipment status change events.

[0058] It should be noted that this invention is applicable to intelligent operation and maintenance scenarios for rotating mechanical equipment (such as centrifugal compressors, pumps, fans, steam turbines, etc.). In rotating mechanical equipment, the changes in parameters such as vibration signals, temperature, and pressure are constrained by classical physical laws, and causal modeling and degradation prediction are achieved based on the following natural laws: equipment vibration signal data reflects the dynamic mechanical state of rotating mechanical bearings, gears, or rotors, and is one of the indicators of early failures; equipment temperature readings reflect the equipment's thermal load status and are used to monitor abnormal frequencies such as frictional overheating, cooling failure, or electrical overload; equipment pressure fluctuation data reflects the sealing performance and flow stability degradation degree of fluid systems (such as pipelines, compressor cavities); and equipment fault data serves as discrete event signals, used to verify prediction results, trigger decision commands, and participate in deviation source location.

[0059] Data cleaning and standardization are performed on industrial field data streams to generate structured time-series datasets. This includes: filling missing values ​​in equipment temperature readings or equipment pressure fluctuation data with the mean; detecting and removing outliers in equipment vibration signal data using outlier clustering; removing duplicate records in equipment fault data based on hash algorithms; and dividing the cleaned equipment vibration signal data, equipment temperature readings, and equipment pressure fluctuation data into fixed-length subsequences in chronological order using a sliding window. Each subsequence serves as a sample in the structured time-series dataset, with adjacent samples arranged consecutively by timestamps to generate the structured time-series dataset.

[0060] S2. Analyze the structured time series dataset using the large industrial state analysis model to obtain the industrial state feature vector sequence. Divide the industrial state feature vector sequence using a sliding time window and perform causal verification to generate a causal knowledge rule set.

[0061] The large-scale industrial state analysis model includes a multimodal fusion and feature extraction layer, an industrial mechanism embedding and fine-tuning layer, and a task decision and optimization layer.

[0062] The multimodal fusion and feature extraction layer generates a joint feature representation of the multimodal based on a structured time-series dataset using a feature-level fusion method;

[0063] Furthermore, the equipment vibration signal data in the structured time-series dataset is converted into a two-dimensional time-series image using a Markov transfer field, and the equipment temperature readings and equipment pressure fluctuation data are converted into one-dimensional time-series vectors using a time-domain statistical feature extraction method. Subsequently, a feature-level fusion method is applied to concatenate the feature vectors of the two-dimensional time-series image with the feature vectors of the one-dimensional equipment temperature readings and equipment pressure fluctuation data to form a multimodal joint feature representation. The multimodal joint feature representation provides the training layer with cross-modal complementary information and high-dimensional feature interaction capabilities.

[0064] The training layer employs an industrial mechanism embedding fine-tuning method to fine-tune the multimodal joint feature representation, thereby obtaining a fused feature representation;

[0065] Furthermore, industrial knowledge is transformed into industrial mechanism embedded vector features through an industrial mechanism embedding fine-tuning method; the industrial mechanism embedded vector features are subjected to linear transformation to obtain an industrial task instruction set; the industrial task instruction set is mapped into a soft cue vector through a knowledge graph instruction fine-tuning method; and then fused with a one-dimensional time-series vector and a two-dimensional time-series image through a feature connection method to generate a fused feature representation.

[0066] The task decision and optimization layer refines and aligns the fused feature representations to obtain an industrial state feature vector sequence, makes decisions based on the industrial state feature vector sequence, and finally outputs the industrial state feature vector sequence and real-time control commands.

[0067] Furthermore, a bidirectional spatiotemporal feature extraction method is used to extract primary features with physical correlation and temporal dependence from the fusion feature representation of the input. Temporal alignment and structured reconstruction are then used to obtain the industrial state feature vector sequence. A dynamic optimization algorithm is applied to the industrial state feature vector to generate real-time control commands. Finally, based on the industrial state feature vector sequence, iterative learning is used to update the large-scale industrial state analysis model, forming a self-optimizing closed loop of "perception-decision-execution-feedback" to achieve continuous optimization of the industrial system.

[0068] The industrial state analysis model is used to extract features from the structured time series dataset to obtain the industrial state feature vector sequence.

[0069] A sliding time window divider is used to divide the industrial state feature vector sequence into time window data segments of fixed duration.

[0070] Furthermore, the window size of the sliding time window divider is determined based on the time periodicity of the industrial state feature vector sequence, with the window size corresponding to a fixed time span. The window slides sequentially along the time axis of the industrial state feature vector sequence based on the fixed time span, advancing the window by a fixed time unit with each slide. There is partial overlap between adjacent windows. A circular buffer stores the current window data segment of the industrial state feature vector sequence, storing the industrial state feature vector sequences in chronological order. When a new industrial state feature vector sequence arrives, it is appended to the end of the circular buffer, while the earliest industrial state feature vector sequence exceeding the current window's start time is removed. The window sliding process strictly aligns the timestamps of the industrial state feature vector sequences, ensuring that the timestamps of the industrial state feature vector sequences within each time window data segment are continuous and evenly distributed. Finally, a fixed-length time window data segment is output, with each time window data segment containing a subset of industrial state feature vector sequences of the same length.

[0071] Based on the time window data segments, the neural Granger causality test algorithm is used to obtain candidate causal pairs;

[0072] Furthermore, based on the physical meaning of each time window data segment, a hypothetical causal pair to be tested is defined for each time window data segment (e.g., assuming that changes in equipment temperature readings are the cause of abnormal equipment vibration signal data, or that abnormal equipment pressure fluctuation data is the cause of equipment fault data). The neural Granger causality test algorithm is used to obtain the direction-intensity-delay triplet of the hypothetical causal pair for each time window data segment. Based on the direction-intensity-delay triplet of the hypothetical causal pair, the Granger causality test is used to determine the causal direction and time-varying characteristics to obtain candidate causal pairs for each time window data segment.

[0073] For the obtained candidate causal pairs, the sliding window sequence matching method is used to verify the continuity of the time window to obtain a stable causal chain;

[0074] Furthermore, the candidate causal pairs for each time window data segment are arranged in order of their start timestamps, forming a time-order indexed sequence of candidate causal pairs. Then, the sliding window sequence matching parameters are initialized. The window size of the sliding window sequence matching method is consistent with the fixed duration of the time window data segments to ensure that each window covers the same number of consecutive time window data segments. Simultaneously, the sliding step size is set to advance one time window data segment at a time, ensuring no overlap between adjacent sliding windows. Subsequently, the continuity verification of candidate causal pairs within the sliding window is performed. For each subset of consecutive time window data segments covered by each sliding window, the direction-intensity-delay triples of all candidate causal pairs are extracted, verifying whether the same candidate causal pair appears consecutively in all time window data segments within the sliding window. If a candidate causal pair passes the above verification in multiple consecutive sliding windows, a stable causal chain is generated. The stable causal chain record includes the causal pair body, the mean of the direction-intensity-delay triples in the consecutive time window data segments, and the time span covered by the stable causal chain.

[0075] Based on stable causal chains, a set of causal knowledge rules is obtained through a rule template matching strategy.

[0076] Furthermore, four descriptive fields are defined: causal subject field, direction attribute field, intensity attribute field, and time delay attribute field. Specifically, the causal pair subjects recorded in the stable causal chain are separated into cause and effect entities using semantic parsing, and filled into the causal subject field in a "cause → effect" format. Next, the direction value is extracted from the direction-intensity-time delay triplet, converted to a direction symbol using a symbolic function mapping method, and filled into the direction attribute field. Simultaneously, the intensity value is extracted from the direction-intensity-time delay triplet, preprocessed using a normalization method to obtain a normalized value, and filled into the intensity attribute field. The time delay value is extracted from the direction-intensity-time delay triplet, converted to actual duration using a time unit conversion method, and filled into the time delay attribute field. Finally, the four filled fields are combined in a fixed order using a field concatenation method to generate a single causal knowledge rule, and all single causal knowledge rules are traversed using set operations to integrate them into a structured causal knowledge rule set.

[0077] S3. Predict the predicted values ​​of equipment status data using a neural differential equation framework;

[0078] A neural differential equation framework is used to embed a set of causal knowledge rules as physical constraints into the differential equation structure;

[0079] Furthermore, the individual causal knowledge rules in the causal knowledge rule set are analyzed. The direction attribute field of each individual causal knowledge rule is mapped to the coupling direction using the signed linear coefficient method; the intensity attribute field of each individual causal knowledge rule is transformed into the action coefficient using the Sigmoid function transformation method; and the time delay attribute field of each individual causal knowledge rule is quantized into a delay operator using the peak detection method of the cross-correlation function. Based on the coupling direction, action coefficient, and delay operator, an embeddable neural differential equation is constructed to calculate the physical constraint value, expressed as:

[0080] ,

[0081] Where R is the physical constraint value, d is the action coefficient (0-1), d is the coupling direction, and z is the delay operator.

[0082] Based on physical constraints, the residuals with physical constraints are added to the neural differential equation framework using the residual constraint construction method to calculate the derivative of the state vector. The gradient of the residuals with physical constraints is calculated using the adjoint sensitivity method, and the neural differential equation framework is corrected by backpropagation using a stochastic optimization algorithm. A neural differential equation framework that conforms to the physical constraints of the causal knowledge rule set is generated.

[0083] It should be noted that the range of values ​​for the action coefficient is determined by the mathematical properties of the Sigmoid function.

[0084] The real-time industrial field data stream is used in conjunction with a neural differential equation framework to calculate the time derivative of the equipment status data at the current moment, and then perform integral calculations to generate predicted values ​​of the equipment status data.

[0085] Furthermore, the system receives real-time industrial field data streams and inputs them into a neural differential equation framework in chronological order. Based on the current equipment status data and the physical constraints of the embedded causal knowledge rule set, the neural differential equation framework calculates the time derivatives of equipment vibration signal data, equipment temperature readings, and equipment pressure fluctuation data. Based on the calculated time derivatives, an explicit Euler method is used to generate predicted values ​​for the equipment status data in the next time window, expressed as:

[0086] ;

[0087] in, For the next time port The predicted value of the equipment status data within the time window t, S(t) represents the equipment status data within the time window t. The time window length, It is the derivative in the time dimension.

[0088] S4. Based on the predicted values ​​of equipment status data, use the large industrial status analysis model to conduct risk and benefit assessment and generate structured decision instructions.

[0089] Based on the predicted values ​​of equipment status data, dynamic baseline residual analysis is used to identify potential failure risks.

[0090] Furthermore, the predicted values ​​of the current equipment status data are compared point-by-point with the real-time collected equipment status data according to timestamps to generate a dynamic baseline residual sequence. Each residual term explicitly represents the deviation between the predicted and measured values ​​of the three types of equipment status data. A sliding time window divider is used to segment the dynamic baseline residual sequence into segments of fixed duration. Within each time window, the upper and lower bounds of the confidence interval are calculated using the least squares method for the current window residual sequence. Within the sliding time window, if multiple consecutive residual terms in the dynamic baseline residual sequence (e.g., 5 consecutive sampling points) simultaneously exceed the upper or lower bound of the confidence interval, a potential fault risk indicator is triggered. The potential fault risk indicator records the specific data type of the equipment status that exceeds the limit (equipment vibration signal data, equipment temperature reading, or equipment pressure fluctuation data), the direction of deviation (positive or negative exceedance), and the time span of continuous exceedance. Based on the potential fault risk, a quantified fault risk is generated using a dynamic risk propagation algorithm based on a causal knowledge rule set. Through the task decision and optimization layer of the industrial status analysis model, the quantified fault risk is analyzed to generate structured decision instructions.

[0091] S5. Execute structured decision instructions, collect real-time measured values ​​of equipment status data, and perform deviation analysis and evaluation of the execution effect between predicted and measured values ​​of equipment status data, while dynamically optimizing the neural differential equation framework.

[0092] Execute structured decision instructions and simultaneously collect measured values ​​of device status data with timestamps;

[0093] Furthermore, the structured decision instructions are converted into standardized equipment control signals; the equipment operating state is changed according to the standardized equipment control signals; at the instant the instruction is completed, the measured values ​​of equipment vibration signal data, equipment temperature readings, and equipment pressure fluctuation data are collected synchronously, and each measured value is marked with a timestamp; thus, the measured values ​​of equipment status data with timestamps are obtained.

[0094] Align the timestamps of the measured equipment status data with the timestamps of the predicted equipment status data to generate a measured time series dataset;

[0095] Furthermore, based on the differences in timestamp range and time resolution between the measured and predicted values ​​of equipment status data, the least common multiple method is used to calculate the reference data frequency. A reference timestamp sequence with equal intervals is generated based on the reference data frequency, with each timestamp being an integer multiple of the reference data frequency, ensuring strict continuity and no overlap. For each timestamp of the measured equipment status data, the nearest neighbor principle is applied to match the nearest reference timestamp in the reference timestamp sequence. Each timestamp of the predicted equipment status data is mapped to the reference timestamp sequence using the same matching rule. If a reference timestamp has no corresponding measured or predicted value, the timestamp position is retained and marked as null. The matched measured and predicted equipment status data values ​​are arranged in the order of the reference timestamps. Three columns of data are generated to record the measured and predicted values ​​of vibration, temperature, and pressure under a unified reference timestamp, forming a measured time-series dataset.

[0096] Based on the measured time series dataset and the predicted values ​​of the equipment status data, the sliding window cumulative error statistical algorithm is used to calculate the root mean square error and the mean absolute percentage error between the predicted values ​​of the equipment status data and the measured values ​​of the equipment status data within the same time window, and to generate the prediction deviation value of the equipment status data.

[0097] Furthermore, a fixed-duration sliding time window divider is employed to align the measured and predicted equipment status data in the measured time-series dataset with a unified timestamp, generating multiple consecutive time window data segments. Each time window data segment contains a subset of the sequence of measured and predicted equipment status data of the same duration. For each time window data segment, the squared residuals of the predicted and measured equipment status data are calculated point-by-point, expressed as:

[0098] ;

[0099] Where I is the squared residual, L(a) is the predicted value of the equipment status data within time window a, and S(a) is the measured value of the equipment status data within time window a.

[0100] The mean square error is calculated by summing all the squared residuals and dividing the sum by the length of the time window data segment. The root mean square error (RMSE) of the current time window data segment is then calculated by taking the square root of the RMS error. For each device status data point within the same time window data segment, the absolute percentage error is calculated using the following expression:

[0101] ;

[0102] Where U represents the absolute percentage error;

[0103] Accumulate all absolute percentage errors within the current time window data segment and calculate the ratio with the length of the time window data segment to generate the average absolute percentage error. Use the root mean square error and average absolute percentage error calculated for each time window data segment as the prediction deviation value of the equipment status data corresponding to the time window data segment. Arrange the prediction deviation values ​​of the equipment status data corresponding to each time window data segment according to the timestamp to form a prediction deviation value sequence aligned with the timestamp of the time window data segment.

[0104] Based on the predicted deviation value sequence of equipment status data, the abnormal deviation time window in the measured time series dataset is identified through density deviation analysis. Combined with the equipment topology network, the physical source of the abnormal deviation is located, and deviation source location data is generated.

[0105] Furthermore, based on the predicted deviation value sequence of equipment status data and the measured values ​​of historical equipment status data, the kernel density estimation method is used to calculate the probability density distribution function of the predicted deviation value and the historical baseline probability density distribution function. The KL divergence method is used to calculate the density deviation index and the historical density deviation index for the probability density distribution function of the time window and the historical baseline probability density distribution function, respectively. When the density deviation index continuously exceeds three times the standard deviation of the historical density deviation index mean, the current time window is determined to be an abnormal deviation time window. The start and end timestamps of the abnormal deviation time window are recorded. Logical connection information between devices is actively collected through a standardized network protocol. Subsequently, a structured topology is constructed using graph theory algorithms, and device identifiers are manually entered to construct the device topology network. Based on the device topology network... Based on the physical location of the equipment, a minimum spanning tree algorithm is used to construct the equipment monitoring network structure diagram. The connection weight of each node in the equipment topology network is calculated through excessive centrality analysis. Within the abnormal deviation time window, the predicted deviation value sequence of equipment status data of all nodes in the equipment topology network is extracted. The predicted deviation values ​​of adjacent nodes are weighted and aggregated through node connection weights to generate node-level aggregated deviation values. The equipment topology network node with the largest node-level aggregated deviation value is located as the physical source of the abnormal deviation. The equipment identifier and topological coordinates of the located node are recorded. The start and end timestamps of the abnormal deviation time window are associated and integrated with the equipment identifier and topological coordinates of the physical source of the abnormal deviation to generate triplet structure data containing time dimension, spatial dimension and deviation intensity dimension, forming deviation source location data.

[0106] The random forest algorithm is used to analyze the deviation contribution of the deviation source location data and combined with causal knowledge rules to generate a deviation contribution quantification matrix.

[0107] Furthermore, the triplet structure data of the deviation source location data is used as input features. The time dimension corresponds to the start and end timestamps of the abnormal deviation time window, the spatial dimension corresponds to the device identifier and topological coordinates of the physical source of the abnormal deviation, and the deviation intensity dimension corresponds to the node-level aggregated deviation value. A random forest algorithm is trained based on the input features. Multiple decision trees are constructed to fit the deviation source location data. Each decision tree selects a split node based on the Gini coefficient, recursively splitting the triplet structure data until the leaf node purity reaches the standard or the maximum depth is reached. After training, the average impurity reduction method is used to calculate the performance of each feature in the random forest algorithm. The total reduction in impurity caused by split nodes in all decision trees is normalized to obtain the initial contribution scores of each feature in the deviation source localization data. Simultaneously, the causal knowledge rule set is analyzed, and the symbol mapping results of the direction attribute field and the normalized values ​​of the intensity attribute field in a single causal knowledge rule are extracted. The symbol mapping results of the direction attribute field are converted into influence direction weight coefficients between features using the symbolic function mapping method, and the normalized values ​​of the intensity attribute field are used as the influence strength coefficients between features. Combining the initial contribution scores, influence direction weight coefficients, and influence strength coefficients, the final deviation contribution score is calculated using a weighted product method, expressed as:

[0108] ,

[0109] in, The final deviation contribution score is given by u, where u is the initial contribution score. β is the influence direction weighting coefficient (value range: 0-1), and β is the influence intensity coefficient (value range: 0-1).

[0110] A deviation contribution quantification matrix is ​​generated by traversing all features. The rows of the matrix represent the data features that locate the deviation source, the columns represent the causal knowledge rule identifiers, and the matrix elements are the final deviation contribution scores of the corresponding features under a specific causal rule.

[0111] Based on the bias contribution metric matrix, an incremental learning mechanism is used to update the neural differential equation framework.

[0112] Furthermore, when the final deviation contribution score of the deviation contribution metric matrix in a continuous time window exceeds three times the standard deviation of the historical density deviation index mean, the measured time-series dataset of the corresponding time window is extracted as an incremental sample. A lightweight adaptation layer is inserted into the neural differential equation framework using a parameter-efficient fine-tuning method, preserving the original feature propagation path. The incremental samples are input into the neural differential equation framework through a gradient descent optimizer to calculate the derivative of the state vector and simultaneously calculate the gradient value of the residual term. The adaptation layer parameters are updated using backpropagation, while parameter regularization constrains the change amplitude of the core parameters of the neural differential equation framework. During the iteration process, a time decay strategy is used to adjust the step size of the gradient descent optimizer to ensure monotonically convergent loss function. The state prediction accuracy of the neural differential equation framework after the adaptation layer update is evaluated on the incremental samples. If the verification loss does not decrease for several consecutive rounds, the update is terminated, and finally, the optimized neural differential equation framework is generated.

[0113] This embodiment also provides an industrial decision support system based on large models and artificial intelligence, including:

[0114] Data acquisition and preprocessing module, industrial status analysis module, equipment degradation prediction module, decision generation module, and decision execution and verification module;

[0115] The data acquisition and preprocessing module is used to acquire industrial field data streams in real time and preprocess them to generate structured time-series datasets.

[0116] The industrial state analysis module is used to analyze structured time series datasets using a large industrial state analysis model to obtain industrial state feature vector sequences. It divides the industrial state feature vector sequences using a sliding time window and performs causal tests to generate a set of causal knowledge rules.

[0117] The equipment degradation prediction module is used to predict equipment status data using a neural differential equation framework.

[0118] The decision generation module is used to generate structured decision instructions based on the predicted values ​​of equipment status data and the risk-benefit assessment using a large industrial status analysis model.

[0119] The decision execution and verification module is used to execute structured decision instructions, collect real-time measured values ​​of equipment status data, and perform deviation analysis between predicted and measured values ​​of equipment status data to evaluate the execution effect. At the same time, it dynamically optimizes the neural differential equation framework.

[0120] This embodiment also provides a computer device applicable to industrial decision support methods based on large models and artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the industrial decision support method based on large models and artificial intelligence proposed in the above embodiment.

[0121] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0122] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the industrial decision support method based on large models and artificial intelligence as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0123] In summary, this invention achieves high-precision prediction of equipment status, risk tracing, and adaptive model optimization by combining degradation modeling and dynamic prediction based on causal rule constraints with closed-loop verification using the sliding window cumulative error statistical method. Ultimately, it forms an autonomous decision-making closed loop of "perception-prediction-decision-verification," comprehensively improving the reliability, decision-making efficiency, and dynamic environmental adaptability of industrial systems.

[0124] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An industrial decision support method based on large models and artificial intelligence, characterized in that, Includes the following steps: S1. Real-time acquisition of industrial field data streams and preprocessing to generate structured time-series datasets; S2. Analyze the structured time-series dataset using a large-scale industrial state analysis model to obtain industrial state feature vector sequences. Divide these sequences using a sliding time window and perform causal checks to generate a set of causal knowledge rules, including: Based on the time window data segments, causal tests are performed to obtain candidate causal pairs; The obtained candidate causal pairs are subjected to time window continuity verification to obtain a stable causal chain; Based on stable causal chains, a set of causal knowledge rules is obtained through a rule template matching strategy. S3. Predict the predicted values ​​of equipment status data using a neural differential equation framework; S4. Based on the predicted values ​​of equipment status data, use the large industrial status analysis model to conduct risk and benefit assessment and generate structured decision instructions. S5. Execute structured decision instructions, collect real-time measured values ​​of equipment status data, and perform deviation analysis between predicted and measured values ​​of equipment status data to evaluate the execution effect. Simultaneously, dynamically optimize the neural differential equation framework, specifically: Execute structured decision instructions and simultaneously collect measured values ​​of device status data with timestamps; Align the timestamps of the measured equipment status data with the timestamps of the predicted equipment status data to generate a measured time series dataset; Based on the measured time series dataset and the predicted values ​​of the status data, the root mean square error and the mean absolute percentage error of the predicted values ​​of the equipment status data and the measured values ​​of the equipment status data within the same time window are calculated to generate the prediction deviation value of the equipment status data. Based on the predicted deviation value of the equipment status data, identify the abnormal deviation time window in the measured time series dataset, and locate the abnormal deviation source by combining the equipment topology network, and generate deviation source location data. The causal knowledge rules are used to analyze the deviation contribution of the deviation source location data to generate a deviation contribution quantification matrix. Based on the deviation contribution metric matrix, an incremental learning mechanism is used to update the neural differential equation framework.

2. The industrial decision support method based on large models and artificial intelligence according to claim 1, characterized in that, In S1, the industrial field data stream includes equipment fault data, equipment vibration signal data, equipment temperature readings, and equipment pressure fluctuation data; Data cleaning and standardization are performed on industrial field data streams to generate structured time-series datasets.

3. The industrial decision support method based on large models and artificial intelligence according to claim 2, characterized in that, In S2, a large-scale industrial state analysis model is used to analyze the structured time-series dataset to obtain a sequence of industrial state feature vectors. A sliding time window is then used to divide this sequence of industrial state feature vectors, including: A large-scale industrial state analysis model is constructed based on structured time-series datasets; The industrial state analysis model is used to extract features from the structured time series dataset to obtain the industrial state feature vector sequence. A sliding time window divider is used to divide the industrial state feature vector sequence into time window data segments of fixed duration.

4. The industrial decision support method based on large models and artificial intelligence according to claim 1, characterized in that, S3 specifically refers to: A neural differential equation framework is used to embed a set of causal knowledge rules as physical constraints into the differential equation structure; Based on real-time acquired industrial field data streams and a neural differential equation framework, the time derivative of the equipment status data at the current moment is calculated, and integral calculations are performed to generate predicted values ​​of the equipment status data.

5. The industrial decision support method based on large models and artificial intelligence according to claim 4, characterized in that, S4 specifically refers to: Based on equipment status data predictions, identify potential failure risks; Based on potential failure risks, a quantified failure risk is generated using a dynamic risk propagation algorithm based on a set of causal knowledge rules. Through the task decision and optimization layer of the industrial condition analysis big model, decision analysis is performed on the quantified fault risks to generate structured decision instructions.

6. An industrial decision support system based on large models and artificial intelligence, based on the industrial decision support method based on large models and artificial intelligence as described in any one of claims 1 to 5, characterized in that, include, The data acquisition and preprocessing module is used to acquire industrial field data streams in real time and preprocess them to generate structured time-series datasets. The industrial state analysis module is used to analyze structured time series datasets using a large industrial state analysis model to obtain industrial state feature vector sequences. It divides the industrial state feature vector sequences using a sliding time window and performs causal tests to generate a set of causal knowledge rules. The equipment degradation prediction module is used to predict equipment status data using a neural differential equation framework. The decision generation module is used to generate structured decision instructions based on the predicted values ​​of equipment status data and the risk-benefit assessment using a large industrial status analysis model. The decision execution and verification module is used to execute structured decision instructions, collect real-time measured values ​​of equipment status data, and perform deviation analysis between predicted and measured values ​​of equipment status data to evaluate the execution effect. At the same time, it dynamically optimizes the neural differential equation framework.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the industrial decision support method based on large models and artificial intelligence as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the industrial decision support method based on large models and artificial intelligence as described in any one of claims 1 to 5.

Images