Multi-source perception data fusion processing method and system based on AI large model

Abstract

This invention discloses a multi-source perception data fusion processing method and system based on an AI large-scale model, belonging to the field of data processing technology. By constructing a historical mutation case library and a dynamic time warping matching mechanism, and combining the failure probability vectors and first switching times of each data source with the forward inference output of the large-scale model, the method recursively calculates the flipped node array and the stage dominant source array. This enables early prediction or real-time identification of the dominant source switching time in millisecond-level operational change scenarios, eliminating the reliance on fixed-time window cumulative statistics in traditional methods and compressing the detection latency to a negligible level. Furthermore, the solution dynamically divides stages according to the stage dominant source and immediately terminates the current dynamic stage and triggers an immediate reallocation of fusion weights when a dominant source appears in each stage. This ensures that conflicting modal data no longer participates in fusion calculations according to the old stage weights, curbing the generation and spread of semantic pollution from the source.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for multi-source sensing data fusion processing based on AI large-scale models. Background Technology

[0002] In the field of multi-source sensing data fusion, practical industrial or IoT applications require real-time processing of data from multiple heterogeneous sensors to accurately perceive the environmental or system status. However, operating conditions often undergo rapid changes on a millisecond-level scale, and data dominance can switch instantly. This places extremely high demands on the fusion method's response speed, time-series alignment capability, and cross-stage state tracking capability.

[0003] Traditional dominant source divergence detection methods, such as the FDS divergence detection algorithm, rely on data accumulation within a fixed time window to complete statistical decisions, resulting in inherent detection latency and difficulty in achieving instantaneous response in millisecond-level sudden changes in operating conditions. When data dominance changes abruptly, the system continues to use the original fusion strategy within the detection window, causing conflicting modal data to directly participate in the fusion calculation, leading to semantic pollution. Meanwhile, mainstream recursive fusion architectures, such as Transformer, continuously encode historical data into hidden states or attention caches during operation. If erroneous information is introduced during a sudden change, semantic bias will form. Due to the lack of state backtracking and correction mechanisms, this bias cannot be cleared and will accumulate continuously with the recursive process, continuously affecting subsequent outputs. Even after the dominant source switch is completed, the system still carries historical erroneous biases, causing the fusion result to deviate from the true physical state in the long term. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for multi-source sensing data fusion processing based on a large AI model, thus solving the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Firstly, a multi-source sensing data fusion processing method based on an AI large-scale model includes the following steps: acquiring heterogeneous sensing data, parsing it using a dedicated protocol and detecting distribution offsets, and writing it into a standardized buffer; traversing the buffer to construct an autoencoder and optimize its parameters; performing feature compression based on the optimized autoencoder parameters to form a multi-source feature sequence; performing forward inference on the large model through historical mutation case library matching and dynamic time warping to generate risk assessment results and determine dynamic stages; performing weight allocation and cross-stage feature compensation on the multi-source feature sequence according to the dynamic stages to obtain a delay-corrected multi-source feature sequence; and analyzing the anomaly propagation and coupling effects of each dynamic stage based on the delay-corrected multi-source feature sequence to obtain risk score vectors for each data source in the future stage, and adjusting delay compensation according to the risk score vectors to form a cross-stage closed-loop optimization.

[0006] Preferably, heterogeneous sensing data is acquired, parsed using a dedicated protocol, and its distribution offset is detected and then standardized and written into a buffer. This includes: obtaining standardized data for data distribution offset detection from heterogeneous sensing data based on the IoT platform through MCP protocol adaptation and parsing of various dedicated protocols; analyzing the difference between the current batch data distribution and the baseline distribution based on structured data, calculating the distribution offset using the KL divergence method; if the distribution offset exceeds a preset offset threshold, automatically extracting unlabeled data from the current batch to form a real-time training set for model fine-tuning; and writing the standardized data and distribution offset into the buffer at the position corresponding to the write pointer in timestamp order to update the time range field in the buffer.

[0007] Preferably, the process involves traversing the buffer to construct and optimize the autoencoder parameters. Based on the optimized autoencoder parameters, feature compression is performed to form a multi-source feature sequence. This includes: traversing each data source within the buffer, extracting the temporal features of each data source within each sliding window, concatenating them into a feature vector, and using multiple feature vectors as a training set; constructing an autoencoder based on the training set, and using the mean squared error method and K-nearest neighbor estimation method to analyze and calculate the single-source reconstruction error and cross-source mutual information, using the weighted sum of the single-source reconstruction error and cross-source mutual information as the loss function to train and optimize the autoencoder parameters; compressing the current feature vector based on the optimized autoencoder parameters to obtain a compressed feature vector, and then integrating it to form a multi-source feature sequence.

[0008] Preferably, risk assessment results are generated by performing forward inference on a large model through historical mutation case matching and dynamic time warping. This includes: reading the historical mutation case library to obtain the dominant source switching features of historical periods, aligning each historical mutation case with the current multi-source feature sequence in time, calculating the similarity using a dynamic time warping algorithm, and obtaining a matching set between the current batch features and the historical mutation case library; converting each historical mutation case and multi-source feature sequence in the matching set into a structured text description by time step, and concatenating them to construct a structured instruction package; inputting the structured instruction package into the trained large model, and outputting the failure probability vectors and first switching times of each data source through forward inference; based on the first switching time and combined with the evolution law of the failure probability vector, recursively calculating the flip nodes and the corresponding stage dominant source labels to form a flip node array and a stage dominant source array; and summarizing the failure probability vectors of each data source, the first switching time, the flip node array, and the stage dominant source array to generate risk assessment results.

[0009] Preferably, a dynamic stage is determined, and based on the dynamic stage, weight allocation and cross-stage feature compensation are performed on the multi-source feature sequence to obtain a delayed-corrected multi-source feature sequence. This includes: determining the dynamic stage based on the flipped node array and the stage dominant source array; assigning fusion weights to each dynamic stage based on the failure probability vector, wherein the dominant source fusion weights are calculated according to the reliability ratio to obtain the fusion weight vector for each dynamic stage; combining the fusion weight vector with the compressed feature vectors of each data source within the current sliding window to perform weighted integration and generate the fused multi-source feature sequence; and performing feature delay quantization on the fused multi-source feature sequence through the dynamic stage and predicting the feature offset of future stages to complete the cross-stage feature compensation.

[0010] Preferably, through dynamic stages, feature delay quantization is performed on the fused multi-source feature sequence, and feature shifts in future stages are predicted to complete cross-stage feature compensation. This includes: extracting the fused multi-source feature sequence in each dynamic stage and analyzing the feature response time differences of each data source in different dynamic stages to quantify the feature response delay magnitude and obtain a cross-stage feature delay matrix; based on the cross-stage feature delay matrix, by analyzing the delay pattern and the evolution trend of each feature in the current dynamic stage, the shift direction and magnitude of features in future stages are predicted to obtain a delay correction prediction vector; based on the delay correction prediction vector and the fused multi-source feature sequence, the feature values ​​of each data source are adjusted by weighting to achieve cross-dynamic stage feature compensation and obtain a delay-corrected multi-source feature sequence.

[0011] Preferably, based on the delayed-corrected multi-source feature sequence, the anomaly propagation and coupling effects of each dynamic stage are analyzed to obtain the risk score vector of each data source in the future stage. The delay compensation is then adjusted according to the risk score vector to form a cross-stage closed-loop optimization, including: based on the delayed-corrected multi-source feature sequence, the reconstruction error and mutual information mutation in each dynamic stage are analyzed to obtain an anomaly impact matrix summarized by stage and time step; using the dynamic stage as a time reference and combined with the anomaly impact matrix, the anomaly evolution law is analyzed using a recursive simulation method to obtain the risk score vector of each dynamic stage; when the risk score of each data source in the risk score vector falls within the preset risk threshold range, the corresponding fusion weight in the original fusion weight vector is maintained; otherwise, the cross-stage optimization mechanism is executed.

[0012] Preferably, the process of implementing the cross-stage optimization mechanism includes: initiating the cross-stage optimization mechanism, taking data sources whose risk scores do not fall within the preset risk threshold range as the objects to be adjusted; performing linear scaling mapping on the objects to be adjusted individually to obtain mapping coefficients, and multiplying the mapping coefficients by the corresponding fusion weights in the original fusion weight vector to obtain the corrected fusion weight vector; based on the corrected fusion weight vector, repeatedly performing cross-stage feature compensation until the risk scores of each data source within the risk score vector fall within the preset risk threshold range, stopping the adjustment of the fusion weight vector, and realizing cross-stage closed-loop optimization.

[0013] Secondly, the multi-source sensing data fusion processing system based on an AI large model includes: a data acquisition module for acquiring heterogeneous sensing data, parsing it using a dedicated protocol, and detecting distribution offsets to standardize and write it into a buffer; a data processing module for traversing the buffer to construct an autoencoder and optimize its parameters, performing feature compression based on the optimized autoencoder parameters to form a multi-source feature sequence; a predictive-driven module for performing forward inference on the large model through historical mutation case matching and dynamic time warping to generate risk assessment results and determine dynamic stages, performing weight allocation and cross-stage feature compensation on the multi-source feature sequence according to the dynamic stages to obtain a delay-corrected multi-source feature sequence; and a fusion optimization module for analyzing the anomaly propagation and coupling effects of each dynamic stage based on the delay-corrected multi-source feature sequence, obtaining the future stage risk score vector of each data source, and adjusting the delay compensation according to the risk score vector to form a cross-stage closed-loop optimization.

[0014] The above-mentioned solution of the present invention includes at least the following beneficial effects: By constructing a historical mutation case library and a dynamic time warping matching mechanism, and combining the failure probability vectors and first switching times of each data source output by the forward inference of a large model, the flipped node array and the stage dominant source array are recursively calculated. This enables early prediction or real-time identification of the dominant source switching time in millisecond-level operational mutation scenarios, eliminating the reliance of traditional methods on fixed-time-window cumulative statistics and compressing the detection delay to a negligible level. Furthermore, the solution dynamically divides stages according to the stage dominant source and immediately terminates the current dynamic stage and triggers real-time redistribution of fusion weights when a stage dominant source appears. This ensures that conflicting modal data no longer participates in fusion calculations according to the old stage weights, thus curbing the generation and spread of semantic pollution from the source.

[0015] By constructing a cross-stage feature delay matrix, the feature response delay magnitude of each data source relative to the dominant source in each dynamic stage is quantified. The delay correction prediction vector is then used to predict the offset direction and offset in future stages. In turn, weighted translation and interpolation adjustment are performed on the fused multi-source feature sequence to achieve active correction of historical error information rather than passive accumulation. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention.

[0017] Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, embodiments of the present invention provide a multi-source sensing data fusion processing method based on an AI large model, including the following: S100: Acquire heterogeneous sensing data, and write it to a buffer for standardization through dedicated protocol parsing and distribution offset detection; S200: Traverse the buffer to construct an autoencoder and optimize the autoencoder parameters. Based on the optimized autoencoder parameters, perform feature compression to form a multi-source feature sequence, providing basic data for dominant source identification; S300: Perform forward inference on the large model through historical mutation case library matching and dynamic time warping to generate risk assessment results and determine dynamic stages. According to the dynamic stages, perform weight allocation and cross-stage feature compensation on the multi-source feature sequence to obtain a delay-corrected multi-source feature sequence; S400: Based on the delay-corrected multi-source feature sequence, analyze the anomaly propagation and coupling effects of each dynamic stage to obtain the future stage risk score vector of each data source, and adjust the delay compensation according to the risk score vector to form a cross-stage closed-loop optimization.

[0020] In this embodiment of the invention, during the S100 stage, the system uses the MCP protocol for unified relay and combines the parsing of various dedicated protocols to obtain standardized data and calculate the distribution offset in real time. It can automatically identify data distribution drift and use unlabeled data to form a real-time training set to quickly fine-tune the model, so that the model can still accurately adapt after data mutation or module iteration. For example, in industrial sensing scenarios, when temperature sensors drift, the model parameters can be adjusted in time without manual labeling.

[0021] The S200 extracts temporal features by traversing the buffer and trains an autoencoder. It performs weighted optimization on single-source reconstruction error and cross-source mutual information, achieving feature compression and integration of multi-source feature sequences. It retains core features while removing redundant information, providing high-precision input for large-scale model inference and risk assessment.

[0022] The S300 stage utilizes a historical mutation case library and a dynamic time warping algorithm to generate structured instruction packages. It then calculates the failure probability vector, flip nodes, and stage dominant source array through forward inference, dynamically allocates fusion weights, and combines feature delay analysis to complete cross-stage feature compensation. This ensures timely switching of dominant sources, feature time alignment, and avoids prediction errors caused by delays.

[0023] In the S400 stage, based on the delayed-corrected multi-source feature sequences, reconstruction errors and mutual information mutations are analyzed. An anomaly impact matrix is ​​constructed and, combined with the dominant source-driven weight matrix, a risk scoring vector is recursively generated. A cross-stage optimization mechanism is initiated to dynamically adjust the fusion weights, ensuring that the anomaly risks of each data source are controllable in future stages, forming a closed-loop optimization. For example, on complex industrial production lines, when pressure sensors suddenly malfunction and the dominant source switches frequently, the system can adjust weights and feature alignment in advance to avoid abnormal data affecting overall decision-making, achieving stability and accuracy of the fusion results. Simultaneously, it provides complete historical traceability, offering a reference for subsequent analysis and model iteration.

[0024] In a preferred embodiment of the present invention, S100 includes: heterogeneous sensing data based on the Internet of Things platform, through MCP protocol adaptation and parsing of each dedicated protocol, to obtain standardized data for data distribution offset detection. The standardized data structure includes source ID, data type, original value, local timestamp, and device calibration parameter pointer fields.

[0025] Specifically, the MCP protocol is first used as a unified and compatible relay layer. Then, the dedicated parsers for the Modbus, RTSP, Onvif, and TCP protocols are called respectively to unpack and decrypt the raw data streams of different types and formats one by one to extract the valid fields.

[0026] Based on structured data, the difference between the current batch data distribution and the baseline distribution is analyzed. The distribution offset is calculated using the KL divergence method, which quantifies the difference between the current distribution and the baseline distribution. If the distribution offset exceeds a preset offset threshold, it may indicate that the model iteration has caused a serious data distribution offset, resulting in a decrease in the model's current adaptability. In this case, unlabeled data from the current batch must be automatically extracted to form a real-time training set to reflect the true distribution after the current offset. Subsequently, the current weight parameters are loaded onto the model to be fine-tuned, and the real-time training set is input into the model. The output result is obtained through forward computation, and the loss function, such as the data distribution fitting loss, is calculated according to the task objective. Using the backpropagation algorithm, the gradient of the trainable parameters of each layer of the model is calculated based on the loss, and the weight parameters are updated according to the preset learning rate to achieve incremental adjustment of the model parameters, thereby allowing the model to quickly adapt to the new distribution and avoid errors in subsequent inference and detection. Conversely, if the distribution offset is within an acceptable range, the structured data can be directly output.

[0027] It should be noted that since labeled data requires manual annotation to determine whether there is any offset and the reason for the offset, which is time-consuming and costly and cannot be obtained in real time, unlabeled data is selected. The fine-tuned model refers to the core neural network in the large model or multi-source fusion model used for feature extraction, prediction or anomaly detection.

[0028] A baseline distribution is used to measure whether the current batch of data deviates from the normal or expected distribution. It is usually a statistical distribution generated from stable data accumulated over a long period of system operation, such as the mean, variance, or probability density function formed by each sensing data source over a long period of operation.

[0029] The current batch of data refers to a set of data within a continuous time window extracted from the buffer in a single processing loop, such as 200 time step data points. This is used for batch analysis of data distribution and to avoid misjudgments caused by fluctuations in single-frame data.

[0030] The standardized data and distribution offsets are written into the buffer at the positions corresponding to the write pointers in timestamp order. This is to ensure the integrity and traceability of the data, and to facilitate subsequent steps in querying and retrieving data, so as to update the time range field in the buffer.

[0031] Specifically, the write pointer is updated by incrementing it by 1. It is then checked whether the write pointer exceeds the maximum length of the buffer. If it does, the write pointer is reset to 0, overwriting the oldest data. This is to achieve buffer reuse and avoid buffer overflow. At the same time, the time range field of the buffer is updated to record the time range field currently stored in the buffer, i.e., the earliest and latest times. This is to facilitate subsequent backtracking window truncation and quick location of the required historical data.

[0032] The intermediate states of the current fusion system are read, including the hidden states of recursive units, the previous fusion weights, the previous detection results, and the previous dominant source markers. These intermediate states are written into the state storage area of ​​the buffer and mapped to the current timestamp. The purpose is to ensure that the intermediate states are synchronized with the data in time and to provide historical intermediate state references for subsequent backtracking and correction.

[0033] A buffer is a ring-shaped storage structure with fixed capacity that stores data in chronological order. It can cyclically overwrite the oldest data and is used to temporarily store recent data, such as 10 seconds of sensing data and intermediate system states, for subsequent backtracking, reasoning, and verification.

[0034] The embodiments of the present invention can unify the processing of data with different protocols and formats into structured information, and use KL divergence to measure the deviation between the current batch of data and the benchmark distribution, thereby automatically judging the model fit and generating an unlabeled training set for fine-tuning when the offset exceeds the limit, achieving rapid dynamic adaptation without relying on manual annotation, saving time and costs.

[0035] Meanwhile, the use of a circular buffer to store standardized data and intermediate states in chronological order not only ensures data integrity and traceability, but also supports historical backtracking and state correction. For example, in industrial IoT scenarios, when a temperature sensor experiences a short-term drift, the system can obtain continuous data from the past 10 seconds through the buffer, calculate the distribution offset, and fine-tune the model in a timely manner, enabling the fusion system to quickly restore accuracy and avoid subsequent prediction errors caused by single-frame anomalies or sudden changes. Overall, this improves the data reliability and inference stability of the multi-source sensing system in dynamic environments.

[0036] In a preferred embodiment of the present invention, S200 includes: Since the physical quantities of each data source contain a large amount of redundant information and random noise, and the data dimension is high and the temporal correlation is scattered, directly using them for cross-source mutual information calculation will interfere with the quantification accuracy of the correlation between different data sources, resulting in mutual information estimation deviation; when used for single-source reconstruction error calculation, it will greatly increase the reconstruction computation load, and the redundant information will make it difficult for the reconstruction error to accurately reflect the distortion degree of the core features, making it difficult to provide effective guidance for compression. Therefore, it is necessary to traverse each data source in the buffer and extract the temporal features of each data source in each sliding window. Here, the mean of multi-source sensing data is used, such as the mean of temperature and humidity in a set time window; it is used to flexibly capture the short-term fluctuation trend of data, avoid the analysis limitations brought by fixed sampling nodes, eliminate the feature judgment deviation caused by single-point abnormal noise, and splice them into feature vectors, and use multiple sets of feature vectors as training sets.

[0037] An autoencoder is constructed based on the training set, and the mean squared error method and K-nearest neighbor estimation method are used to analyze and calculate the single-source reconstruction error and cross-source mutual information. The weighted sum of the single-source reconstruction error and cross-source mutual information is used as the loss function to train the autoencoder and optimize its parameters. This step optimizes the autoencoder through training, enabling it to accurately reconstruct the features of each individual data source while retaining the relevant information between different data sources. This provides a reliable and compact input representation for subsequent multi-source feature sequence construction and dynamic stage analysis.

[0038] Specifically, the single-source reconstruction error is the difference between the original feature vector and the reconstructed feature vector of each individual data source during the compression and decompression process of the autoencoder. The mean square error method is used to calculate the squared error of each feature dimension element by element and then average it to quantify the autoencoder's ability to retain single-source information.

[0039] Cross-source mutual information is a compressed feature representation between different data sources. The mutual information value between features from different sources is estimated by the K-nearest neighbor estimation method to reflect the potential correlation and redundant information between data sources. The higher the mutual information, the stronger the correlation or shared information between features from different sources.

[0040] The loss function is a weighted sum of single-source reconstruction error and cross-source mutual information, which is used to guide the update of autoencoder parameters. This minimizes the reconstruction error while preserving cross-source correlation, thereby achieving effective compression and representation of multi-source features.

[0041] The main parameters of an autoencoder include the weight matrix and bias vector of the encoder and decoder. These parameters determine how the input features are mapped in the low-dimensional latent space and the reconstruction accuracy.

[0042] The current feature vector is compressed based on the optimized autoencoder parameters to obtain a compressed feature vector. This compressed feature vector is then integrated, with all compressed feature vectors combined according to the data source and sliding window dimensions to form a multi-source feature sequence. Rows in the matrix correspond to time steps, and columns correspond to features from different data sources, used for forward inference of the large model and comparison with historical cases. This step ensures that the multi-source features retain core information while eliminating redundancy, facilitating subsequent model inference and risk assessment.

[0043] Data source refers to different acquisition devices.

[0044] In this embodiment of the invention, in stage S200, by extracting temporal features from each data source in the buffer within a sliding window and constructing an autoencoder, effective compression of high-dimensional, multi-source data and extraction of core information are achieved, thereby improving the accuracy and efficiency of multi-source data fusion and risk assessment.

[0045] By training the autoencoder by calculating the weighted loss of single-source reconstruction error and cross-source mutual information, the system can eliminate redundant information and random noise while retaining key features, avoiding the problems of mutual information bias and excessive computation caused by directly using the original high-dimensional data.

[0046] In a preferred embodiment of the present invention, S300 includes: reading a historical mutation case library to obtain dominant source switching features of historical periods. Each dominant source switching feature includes a corresponding feature sequence, the time of the first switching, the dominant source after the first switching, whether there is a stage flip, the time of the flip node, and the dominant source label of each stage. Each historical mutation case is time-aligned according to the current multi-source feature sequence, and a similarity is calculated using a dynamic time warping algorithm to obtain a matching set between the current batch features and the historical mutation case library, including a matching case index and a similarity score, to provide accurate input for the subsequent generation of structured instruction packages.

[0047] The historical mutation case library can be obtained through long-term accumulation and annotation analysis of past multi-source sensing system operation data. The specific construction method is as follows: First, collect various types of multi-source data in different time periods, including the feature sequences of each sensor or module and their corresponding event tags or anomaly records; then, divide these data into stages and analyze the dominant source to determine the time of each dynamic stage dominant source switch, the dominant source after the switch, and whether a stage reversal occurs and the time of the reversal node; at the same time, record the dominant source tags and corresponding feature sequences of each data source in each stage; finally, store this information in a structured manner to form the historical mutation case library, which can be retrieved and compared in subsequent reasoning.

[0048] The dynamic time warping algorithm compares the matching degree between the current batch of multi-source feature sequences and the multi-source feature sequences of historical cases on the time axis, comprehensively considering the feature change patterns of each data source and the dominant source switching pattern. The similarity score comprehensively considers the field sequence, time step alignment, dominant source label change pattern, and the occurrence time of the flip node. When the similarity score exceeds the preset similarity threshold, the corresponding historical mutation case is included in the matching set. This matching process is used to accurately locate potential dominant source switching nodes and flip nodes, providing reliable input for generating structured instruction packages.

[0049] Each historical mutation case and multi-source feature sequence in the matching set is converted into a structured text description by time step, and then concatenated to construct a structured instruction package. Each instruction package contains at least a time step index, a dominant source label, a stage reversal node, and cross-source mutual information, along with compressed feature vectors and reconstruction errors. After generation, the instruction package forms a dataset that can be directly used for forward inference of large models, ensuring that the input information is complete and traceable, and providing a standardized context for dominant source switching prediction.

[0050] In multi-source sensing data analysis tasks, deep learning models trained using historical multi-source feature sequences and labeled dominant source switching information are time-series prediction models based on Transformer, LSTM, or other sequence modeling structures. Their input is a structured instruction package, and their output is the state prediction results for each data source. The training objective is to enable the model to learn to identify potential dominant source switching points and abnormal risk distribution patterns at different dynamic stages.

[0051] The structured instruction package is input into the trained large model, and the failure probability vectors of each data source and the first switching time are output through forward inference.

[0052] In the specific implementation process, forward inference first decomposes the structured instruction package into time series inputs. Each time step corresponds to the current multi-source feature compression vector and the feature information of historical matching cases. The encoder maps the input features to a high-dimensional representation space, enabling the model to simultaneously capture the change patterns of each data source in the time series and cross-source correlation features.

[0053] Subsequently, at each time step, the model processes the time series through an internal self-attention mechanism, combining historical information with current features to gradually generate a predicted distribution for the state at future time steps. By mapping the softmax activation function of the model's output layer to obtain the failure probability of each data source, continuous predicted values ​​can be normalized into a probability range, reflecting the likelihood of the data source being abnormal or failing at each time step.

[0054] Meanwhile, the model will mark the time step in the output sequence when the failure probability first exceeds the preset probability threshold, as the predicted value of the first dominant source switching time; the output of the entire forward inference, including the failure probability vector of each data source at each time step and the predicted first switching time point, will serve as the basis data for subsequent stage division and weight allocation.

[0055] After obtaining the first switchover time and the failure probability vectors of each data source, a recursive analysis method is needed to further determine the flip-over nodes and the dominant source labels for each stage in the dynamic phase. First, the first switchover time is defined as the starting point of the first dynamic phase. From this point onward, the model progressively tracks and analyzes the change in the failure probability of each data source over time, determining whether the dominant source is likely to flip by calculating the probability gradient, rate of change, and relative reliability index. For example, if at a certain time step, the failure probability of the current dominant source increases significantly, while the failure probability of another data source simultaneously decreases and falls below the probability of the dominant source, then this time step may become a flip-over node.

[0056] After determining the flipped node, the algorithm recursively uses it as the starting point of the next dynamic stage, re-evaluates the dominant source label of the stage, and selects the data source with the lowest failure probability as the dominant source. This dominant source not only represents the reference reliability of the data in the stage, but also plays a key role in cross-stage feature compensation and weight allocation.

[0057] During the recursive process, the model continuously monitors the evolution trend of failure probability at each time step, and combines the phased flip rules to gradually generate a flip node array, that is, to record all the time points when the dominant source switch occurs, and at the same time generate a phased dominant source array to identify the relatively reliable data source in each dynamic phase.

[0058] Finally, the array of flipped nodes and the number of dominant sources in each stage were determined to provide key basis for cross-stage feature compensation, weight allocation and closed-loop optimization, and to ensure timely, accurate and quantifiable response to changes in dominant sources in a multi-source sensing environment.

[0059] The dominant source array for each stage corresponds to each time period before and after the flip node, and marks the dominant source for that stage.

[0060] The failure probability vectors of each data source, the first switchover time, the flip node array, and the phase-dominant source array are summarized to generate a risk assessment result.

[0061] This embodiment uses a historical mutation case library and a dynamic time warping algorithm to achieve fine alignment of multi-source feature sequences and prediction of dominant source switching, thereby constructing a standardized structured instruction package input model, which facilitates proactive, dynamic and traceable risk assessment in multi-source environments.

[0062] Logically, the dominant source switching features in each historical mutation case are first time-aligned with the current batch of multi-source feature sequences, and the similarity is calculated using a dynamic time warping algorithm to form a high-fidelity matching set. When constructing the structured instruction package, this set not only retains the compressed feature vector and reconstruction error, but also integrates cross-source mutual information and dominant source flipping patterns, providing rich contextual information for the forward inference of the large model. Subsequently, the large model outputs the failure probability vector and the first switching time, and by recursively calculating the flipping node and the stage dominant source array, it realizes continuous tracking and risk quantification of dynamic stage evolution.

[0063] By using historical case studies to guide the judgment of current dominant source switching, predictions no longer rely solely on single instantaneous data, but rather incorporate time series patterns, phase reversal rules, and cross-source dependencies for proactive correction. For example, in smart grid monitoring, when a critical sensing line experiences a transient anomaly, the system can identify potential dominant source switching nodes in advance based on similar historical abrupt changes, switching the control strategy from high-load substations to backup lines. This dynamic adjustment is completed before a real fault is triggered, further improving system stability and decision-making accuracy, while avoiding misoperations caused by single-source noise or short-term anomalies.

[0064] Based on the array of flipped nodes and the array of dominant sources for each stage, dynamic stages are determined. Specifically, a dynamic stage is formed between every two consecutive flipped nodes. Within each dynamic stage, the dominant source is relatively stable, and the fusion weights of each data source are allocated according to the reliability ratio of the current dominant source. When a new flipped node appears, the dynamic stage ends, and the next stage begins, where the fusion weights are recalculated.

[0065] The dynamic phase is determined dynamically by taking the previous flip node as the starting point and combining the subsequent failure probability trend and possible new flip events. It relies on the real-time analysis of the failure probability sequence by a recursive algorithm, rather than static division. This approach ensures that the multi-source data fusion processing method can adapt to the dominant source switching characteristics in a multi-source sensing environment.

[0066] Based on the failure probability vector, a fusion weight is assigned to each dynamic stage. The fusion weight of the dominant source is calculated according to the reliability ratio to obtain the fusion weight vector of each dynamic stage, which includes the fusion weights of the dominant source and non-dominant source. This provides preset fusion parameters for subsequent cross-stage feature compensation, delay correction and dominant source switching, so that the integration process of the entire multi-source feature sequence can smoothly transition and dynamically adapt to the reliability changes of the data source at each stage, thereby achieving closed-loop optimization.

[0067] Specifically, the failure probability vector is converted into a reliability index, for example, reliability equals 1 minus the failure probability. Then, the fusion weight of the dominant source is calculated based on the normalized proportion of reliability at the current stage, which reflects the contribution proportion of the most reliable data source to the final fusion result in the current dynamic stage. For example, if the reliability of the dominant source is significantly higher than that of other sources, its fusion weight will be assigned a larger proportion to ensure that the final fused features are closer to the feature performance of the dominant source.

[0068] Next, a residual reliability ratio allocation strategy is adopted, that is, after determining the weight of the dominant source, the remaining fusion weights are allocated according to the reliability ratio of each non-dominant source, so that the weight of each non-dominant source is proportional to its reliability.

[0069] The fusion weight vector is combined with the compressed feature vectors from each data source within the current sliding window to generate a fused multi-source feature sequence. This sequence is used for subsequent model inference or control decisions, ensuring that the multi-source information at each sampling point or time step comprehensively considers the dominant source state and risk assessment, avoiding prediction instability or misjudgment caused by directly using single-source features or unfused multi-source features. The fusion mechanism also updates the buffer state storage area, recording the fusion weight vector, dominant source labels, and stage reversal nodes to the corresponding time steps to maintain historical traceability and provide a data foundation for backtracking analysis.

[0070] Through dynamic stages, feature delay quantization is performed on the fused multi-source feature sequences, and feature offsets in future stages are predicted to complete cross-stage feature compensation. This includes the following steps: extracting the fused multi-source feature sequences in each dynamic stage, and analyzing the differences in feature response time of each data source in different dynamic stages to quantify the feature response delay magnitude and obtain the cross-stage feature delay matrix.

[0071] In practice, the fused multi-source feature sequences are standardized according to time steps and data source dimensions to eliminate dimensional differences and unify scales. Then, based on the flipped node array and the stage dominant source array, the fused multi-source feature sequences are divided into corresponding stage segments. Delay correlation analysis is performed on the feature curve of each data source in each stage and the feature curve of the dominant source in that stage. For example, the time offset of the feature response of the data source relative to the dominant source feature is calculated by using a cross-correlation function or a delay mutual information metric. Subsequently, the average delay values ​​of each data source in each stage are summarized to construct a cross-stage feature delay matrix. The rows of the matrix correspond to the data sources, the columns correspond to the dynamic stages, and the matrix elements represent the feature response delay magnitude of each data source relative to the dominant source in each stage, i.e., the average delay value. This is used to quantify the response lag of each source across stages, supporting subsequent feature delay compensation and inter-stage weighted adjustment strategies, thereby optimizing the consistency of features across dynamic stages and the real-time performance of model prediction.

[0072] The cross-correlation function is used to measure the similarity between two time series at different time offsets. The maximum value of the cross-correlation function is taken, and the corresponding time offset is used as the delay of the feature sequence of the corresponding data source relative to the dominant source sequence, i.e., the time offset.

[0073] Based on the cross-stage feature delay matrix, by analyzing the delay pattern and the evolution trend of each feature in the current dynamic stage, the offset direction and offset magnitude of features in the future stage are predicted, and the delay correction prediction vector is obtained.

[0074] In practice, firstly, based on the cross-stage feature delay matrix, the delay change trend of each feature in consecutive stages is analyzed. For example, the delay increase / decrease rate and direction are obtained through linear fitting, moving average, or gradient analysis, forming a feature delay evolution curve. Then, the real-time evolution trend of each feature in the current dynamic stage is compared with historical delay patterns. The possible delay changes of each feature in the next stage are calculated using an autoregressive model (AR) or LSTM, that is, predicting its trend of responding earlier or later than the dominant source.

[0075] The delay pattern describes the trend, magnitude, and stability of feature values ​​in historical and current stage data, whether they are ahead or behind. It is achieved by statistically analyzing the cross-stage feature delay matrix according to feature dimensions and time steps, calculating the mean, variance, and gradient rate of change of each feature delay, and then extracting typical patterns of delay through clustering or principal component analysis, such as certain features always lagging, certain features reacting ahead, and the trend of delay magnitude changing with stage.

[0076] The system analyzes the delay correlation between different features, identifies the dependency or coupling of delays across stages, such as the lag of feature A may affect the response of feature B. Finally, the statistical and analytical results are organized into a structured representation, marking the typical delay direction (advance or lag) and delay amplitude range of each feature in each dynamic stage. This structure is the delay pattern, which provides input for the delay correction prediction vector for predicting feature offsets in future stages.

[0077] The predicted delay changes of all features are integrated to form a delay-corrected prediction vector, where each element represents the predicted offset of the corresponding feature relative to the current dominant source in the future stage, including the offset direction (negative for early, positive for late) and the offset magnitude (predicted time unit length). This vector serves to provide a correction basis for the next stage of multi-source feature fusion, enabling the fusion mechanism to adjust the time alignment of each feature in advance. This predicts and mitigates errors caused by cross-stage delays, ensuring that the actual response values ​​of each feature in the future dynamic stage are as close as possible to the ideal temporal distribution, thereby improving the accuracy and stability of multi-source fusion inference.

[0078] Based on the delayed-corrected prediction vector and the fused multi-source feature sequence, feature compensation across dynamic stages is achieved by weighting and adjusting the feature values ​​of each data source, resulting in the delayed-corrected multi-source feature sequence.

[0079] In practice, the delayed correction prediction vector is aligned with the fused multi-source feature sequence of the current dynamic stage, with each vector element corresponding to a feature time series from a data source.

[0080] For each feature, the feature value at each time step in the sequence is adjusted by weighted shift or interpolation based on its predicted offset. For example, features that are shifted ahead are adjusted forward by interpolation or time shift at the beginning of the sequence, while features that are shifted behind are adjusted by supplementing the predicted value at the end of the sequence or shifting backward. The weights can be dynamically allocated according to the reliability of the feature to ensure that the compensation process does not introduce abnormal oscillations or overcorrection.

[0081] Specifically, the reliability of the features is used as a weighting factor. For each feature sequence, the corrected position at each time step is obtained by subtracting the product of the predicted offset and the feature's reliability from the actual time at each time step. Features with higher offsets are adjusted more significantly, while features with lower reliability are adjusted less significantly.

[0082] The shifted time points are then mapped back to the original discrete time axis, and the compensation value for each time step is calculated using interpolation methods to ensure the continuity and trend of the sequence are maintained. For gaps at the beginning caused by lag, extrapolation or repetition of the first value is used to fill them; for gaps at the end caused by lag, extrapolation or repetition of the last value is used to fill them.

[0083] After performing this operation, a multi-source feature sequence with delay correction is generated. Each feature value at each time step has been time-aligned and magnitude-adjusted according to the cross-stage delay prediction, so that the time response of each feature across dynamic stages is more synchronized, eliminating information misalignment caused by feature delay, and providing accurate input for subsequent dominant source switching judgment, fusion weight allocation and risk assessment, thereby improving the prediction accuracy and stability of the multi-source fusion system in the dynamic stage evolution.

[0084] In this embodiment of the invention, a dynamic stage division and delay correction mechanism effectively solves the problems of lag in cross-stage feature response, untimely switching of dominant sources, and accumulation of prediction errors caused by abnormal interference from single-source features in traditional multi-source fusion methods. Dynamic stages are divided based on a flipped node array and a stage dominant source array, ensuring relative stability of the dominant source within each stage. Fusion weights are precisely allocated according to reliability proportions based on the failure probability vector, effectively transforming the prediction results into an executable fusion strategy. Furthermore, by constructing a cross-stage feature delay matrix and analyzing historical delay patterns, a delay correction prediction vector is formed. Weighted shifting or interpolation adjustments are made to the features of each data source, achieving cross-stage feature time alignment and amplitude compensation, thus solving the information misalignment problem caused by feature lag or pre-action. For example, in a smart power distribution monitoring scenario, the temperature sensor's response lags while the pressure sensor responds prematurely. Direct fusion may misjudge the dominant state of the equipment, triggering incorrect switching actions.

[0085] This invention adjusts temperature characteristics in advance through delayed prediction, ensuring that the fused multi-source feature sequences synchronize with the changes in the dominant source over time, thus guaranteeing a high degree of consistency between the predicted dominant source switching time and actual load mutations. Overall, this invention not only improves the prediction accuracy and stability of multi-source fusion systems during dynamic evolution stages, but also provides a reliable basis for subsequent risk assessment and control strategies through a traceable, structured fusion mechanism, making decision-making in complex, multi-stage systems more intelligent and precise.

[0086] In a preferred embodiment of the present invention, S400 includes: based on the delayed-corrected multi-source feature sequence, analyzing the reconstruction error and mutual information mutation in each dynamic stage to obtain an anomaly impact matrix summarized by stage and time step; the matrix elements contain the reconstruction error and mutual information mutation value of the corresponding feature in each stage, which are used to analyze the possible abnormal evolution law of each feature in future stages.

[0087] Specifically, within each dynamic stage, the delayed-corrected multi-source feature sequence is first divided according to time steps and data sources. The feature vectors of each data source at each time step are input into the autoencoder model to obtain the compressed and reconstructed feature sequence. Then, the element-wise difference between the reconstructed sequence and the delayed-corrected multi-source feature sequence is calculated. The reconstruction error of each time step and each data source is obtained through the mean square error method, forming the reconstruction error matrix within the stage.

[0088] The reconstruction error is used to measure the degree of feature distortion. Since the autoencoder reconstructs normal patterns well, but reconstructs abnormal features poorly, it produces a large error. If the reconstruction error is large, it means that the features at that time step deviate from the historical pattern, which may indicate an anomaly or the risk of data source failure.

[0089] Meanwhile, for feature vectors from different data sources within the same time step, the mutual information change magnitude within adjacent sliding windows is calculated using the mutual information estimation method to obtain the mutual information mutation value, which is used to measure the abnormal fluctuations in the feature dependency relationship between different data sources.

[0090] Using the dynamic stage as a time reference, that is, the state of each stage after the current time step that has not yet been fully observed, and combined with the anomaly influence matrix, the recursive simulation method is used to analyze the anomaly evolution law and obtain the risk score vector of each dynamic stage.

[0091] In practice, the phase-dominant source array is first divided into time series segments. The anomaly impact matrix is ​​split into several phase sub-matrices according to the time boundaries defined by each flip node. Each phase sub-matrice is then bound to the corresponding phase-dominant source label to form a phase-source coupling mapping relationship.

[0092] Based on this mapping relationship, for the abnormal distribution of each data source in each stage, the single data source reconstruction error and cross-source mutual information mutation value are extracted sequentially. Then, these indicators are linearly fused according to preset weight coefficients to form a state anomaly vector within the stage. The anomaly propagation direction is then adjusted based on the dominant source of the stage, so that the anomaly is recursively diffused and updated between different data sources according to the dominant source-driven weight matrix, resulting in the stage state evolution vector. Some of the weight coefficients in this method can be obtained by referring to the analytic hierarchy process.

[0093] Based on this, a unified weighted calculation is performed on the abnormal state change, cross-stage growth rate, and inter-source coupling enhancement term for each data source in the future stage to form a risk score vector. Each dimension of the risk score vector represents the risk score of the corresponding data source affected by the anomaly propagation in the future stage. This vector is used to characterize the relative risk distribution of each data source under the anomaly recursive propagation path in the future stage. The abnormal state change is calculated by averaging the anomaly comprehensive value, describing the overall trend of the anomaly level of the data source in the future stage, i.e., the magnitude or rate of the anomaly increase or decrease of the data source. The cross-stage growth rate is the result of comparing the average anomaly comprehensive value of the previous stage with the average anomaly comprehensive value of the future stage, reflecting the diffusion trend and growth intensity of the anomaly from the previous stage to the future stage, and is an important indicator of cross-stage impact. The inter-source coupling enhancement term is the result of multiplying the change difference of the anomaly comprehensive value of the corresponding data source with the corresponding weight in the dominant source driving weight matrix, and is used to quantify the potential impact of other sources on the future anomalies of the data source, i.e., the anomaly propagation or coupling effect.

[0094] Among them, the state anomaly vector within a stage is used to represent the anomaly structure within each dynamic stage. Its intrinsic parameter is the comprehensive anomaly value of each data source at different time steps within the corresponding dynamic stage, which is obtained by weighted summation of reconstruction error and cross-source mutual information mutation value.

[0095] The stage state evolution vector represents the state result after the propagation adjustment of anomalies within the same stage or across data sources under the control of the dominant source. Its specific construction process is as follows: In the corresponding dynamic stage, the state anomaly vector within the stage is used as input. According to the dominant source corresponding to the current stage, the corresponding dominant source driving weight matrix is ​​called to redistribute the anomaly comprehensive value of each data source. At the same time, a bias term is superimposed for structural correction. The anomaly evolution index of the stage is obtained through nonlinear transformation. This result is used to reflect the distribution and diffusion state of anomalies after being influenced by the dominant source in the current stage, that is, the comprehensive influence level of each data source on anomaly propagation.

[0096] The dominant source-driven weight matrix, under the dominant source type corresponding to the corresponding stage, represents the anomaly propagation relationship and intensity distribution among various data sources. It is used to control the diffusion pattern and intensity of anomalies across different data sources. Its specific construction steps are as follows: First, a set of inter-source correlation metrics is constructed based on historical multi-source data, including mutual information, time offset, and the co-occurrence frequency of anomaly events from each data source. Second, the parameters within the inter-source correlation metrics set are fused according to preset weights to obtain the basic propagation matrix, which reflects the global anomaly propagation structure under conditions without dominant constraints. Further, the stage dominant source is introduced as a conditional variable, and conditional filtering and weight redistribution are performed on the basic propagation matrix. That is, for the data source corresponding to the current stage dominant source, a Sigmoid gating function or a Softmax attention recalibration function is constructed, using the dominant source identifier as input, to dynamically weight the propagation matrix, thereby strengthening the dominant source path and suppressing the non-dominant source path. This forms a conditional propagation matrix matching the current dominant source, which is then normalized to ensure that the input weights of each data source satisfy stability constraints, such as a weight sum of 1, resulting in the final dominant source-driven weight matrix.

[0097] The bias term is a systematic offset correction amount applied to the overall abnormal distribution under the current dominant source condition. It is obtained through historical statistical offset methods. For example, when a certain source becomes the dominant source in historical data, the average offset of the abnormality of each data source is used as the bias term.

[0098] The above steps can accurately capture abnormal patterns and coupling effects of different data sources in dynamic stages, thereby enabling early identification and quantification of potential risks. For example, in an Industrial Internet of Things (IIoT) scenario, a slight drift in a temperature sensor may not immediately affect the overall system, but by using the stage state evolution vector and the dominant source-driven weight matrix, the system can identify the chain of anomalies that this drift may trigger. This allows for fine-tuning of the weights of other related sensors in the next stage, achieving preventative intervention. Compared to traditional methods that rely solely on single-source or fixed threshold judgments, this method improves the visualization and prediction capabilities of cross-source anomaly propagation, reducing the probability of false alarms and missed alarms.

[0099] When the risk scores of each data source within the risk scoring vector fall within the preset risk threshold range, the corresponding fusion weight in the original fusion weight vector is maintained; otherwise, it indicates that the corresponding data source has a higher or lower risk, and its proportion in the fusion weight needs to be reduced or increased, and a cross-stage optimization mechanism is executed.

[0100] A cross-stage optimization mechanism is initiated, and data sources whose risk scores do not fall within the preset risk threshold range are selected as the objects to be adjusted. The objects to be adjusted are individually subjected to linear scaling mapping to obtain mapping coefficients, and the mapping coefficients are multiplied by the corresponding fusion weights in the original fusion weight vector to obtain the corrected fusion weight vector.

[0101] The mapping coefficient is determined by the risk score and the sensitivity coefficient. The sensitivity coefficient is used to control the impact of the risk score on the adjustment of the fusion weight. It is obtained through simulation optimization and its value ranges from 0.1 to 10. The larger the value, the more obvious the inhibitory effect of risk on the weight.

[0102] In this invention, all parameters are dimensionless by using dimensionless processing technology to remove their dimensions, and all thresholds in this invention can be obtained by the mean-standard deviation method.

[0103] Based on the corrected fusion weight vector, cross-stage feature compensation is repeatedly performed until the risk scores of each data source within the risk scoring vector fall within the preset risk threshold range. At this point, the adjustment of the fusion weight vector stops, achieving cross-stage closed-loop optimization. This closed-loop feedback design enables the system to gradually clean up historical biases after the dominant source is switched, ensuring that the fusion result closely approximates the real physical state in the long term. This further overcomes the fundamental defects of irreversible semantic shifts and difficulty in self-correction in traditional recursive architectures.

[0104] In this embodiment of the invention, an anomaly influence matrix and a dominant source driving weight matrix are established through step S400. This enables refined quantification and closed-loop optimization of the anomaly evolution of multi-source data in dynamic stages, solving the problem of lack of systematic control over anomaly propagation paths, cross-stage risks, and feature distortion in traditional multi-source fusion. This invention calculates the reconstruction error by inputting the delay-corrected multi-source feature sequence into an autoencoder, simultaneously quantifying cross-source mutual information mutations to form a stage-specific state anomaly vector. Combined with the stage's dominant source, the anomaly propagation path is recursively adjusted through the dominant source driving weight matrix, and a risk score vector is generated for cross-stage fusion weight correction, thereby achieving anomaly control and risk closed-loop optimization. For example, in intelligent manufacturing monitoring, when anomaly values ​​from a temperature sensor increase and may trigger anomalies in pressure and vibration data, this invention can identify the temperature sensor as a high-risk source, dynamically adjust its weight in the fused feature sequence, and simultaneously time-align other sensor features according to cross-stage delay compensation to prevent anomalies from amplifying in the system. This maintains the predictive stability and real-time performance of the entire multi-source fusion system, achieving refined anomaly management and risk control, and ensuring reliable decision output.

[0105] like Figure 2As shown, embodiments of the present invention also provide a multi-source sensing data fusion processing system based on an AI large model, comprising: a data acquisition module for acquiring heterogeneous sensing data, and writing it into a standardized buffer through dedicated protocol parsing and distribution offset detection; a data processing module for traversing the buffer to construct an autoencoder and optimize its parameters, and performing feature compression based on the optimized autoencoder parameters to form a multi-source feature sequence; a predictive-driven module for performing forward inference on the large model through historical mutation case library matching and dynamic time warping, generating risk assessment results and determining dynamic stages, and performing weight allocation and cross-stage feature compensation on the multi-source feature sequence according to the dynamic stages to obtain a delay-corrected multi-source feature sequence; and a fusion optimization module for analyzing the anomaly propagation and coupling effects of each dynamic stage based on the delay-corrected multi-source feature sequence, obtaining the future stage risk score vector of each data source, and adjusting the delay compensation according to the risk score vector to form a cross-stage closed-loop optimization.

[0106] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0107] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for multi-source sensing data fusion processing based on AI large-scale models, characterized in that, The method includes: Heterogeneous sensing data is acquired, parsed using a dedicated protocol, and its distribution offset is detected, then written to a standardized buffer. The buffer is traversed to construct and optimize the autoencoder parameters. Based on the optimized autoencoder parameters, feature compression is performed to form a multi-source feature sequence. By matching historical mutation case database and dynamic time warping, forward inference is performed on large model to generate risk assessment results and determine dynamic stages. Based on the dynamic stages, weight allocation and cross-stage feature compensation are performed on multi-source feature sequences to obtain delayed-corrected multi-source feature sequences. Based on the delayed-corrected multi-source feature sequences, the abnormal propagation and coupling effects of each dynamic stage are analyzed to obtain the risk score vectors of each data source in the future stage. The delay compensation is then adjusted according to the risk score vectors to form a cross-stage closed-loop optimization.

2. The multi-source sensing data fusion processing method based on AI large model according to claim 1, characterized in that, Acquire heterogeneous sensing data, parse it using a dedicated protocol and detect distribution offsets, and then write it to a standardized buffer, including: Based on the heterogeneous sensing data of the Internet of Things platform, standardized data for data distribution offset detection is obtained through MCP protocol adaptation and parsing of various dedicated protocols. Based on structured data, the difference between the current batch data distribution and the baseline distribution is analyzed, and the distribution offset is calculated using the KL divergence method. If the distribution offset exceeds the preset offset threshold, the unlabeled data of the current batch is automatically extracted to form a real-time training set for model fine-tuning. The standardized data and distribution offsets are written to the positions corresponding to the write pointers in the buffer in timestamp order to update the time range field in the buffer.

3. The multi-source sensing data fusion processing method based on AI large model according to claim 1, characterized in that, The buffer is traversed to construct and optimize the autoencoder parameters. Based on the optimized autoencoder parameters, feature compression is performed to form a multi-source feature sequence, including: Iterate through each data source in the buffer, extract the temporal features of each data source in each sliding window, concatenate them into a feature vector, and use multiple sets of feature vectors as a training set; An autoencoder is constructed based on the training set, and the mean squared error method and the K-nearest neighbor estimation method are used to analyze and calculate the single-source reconstruction error and cross-source mutual information. The weighted sum of the single-source reconstruction error and cross-source mutual information is used as the loss function to train the autoencoder and optimize the autoencoder parameters. The current feature vector is compressed based on the optimized autoencoder parameters to obtain a compressed feature vector, which is then integrated to form a multi-source feature sequence.

4. The multi-source sensing data fusion processing method based on AI large model according to claim 1, characterized in that, By matching historical mutation case databases and performing dynamic time warping, forward inference is performed on large models to generate risk assessment results, including: Read the historical mutation case library to obtain the dominant source switching features in historical periods, and align each historical mutation case with the current multi-source feature sequence in time. Use the dynamic time warping algorithm to calculate the similarity and obtain the matching set between the current batch features and the historical mutation case library. Each historical mutation case and multi-source feature sequence in the matching set is converted into a structured text description by time step, and then concatenated to construct a structured instruction package; The structured instruction package is input into the trained large model, and the failure probability vector of each data source and the first switching time are output through forward inference. Based on the first switching time and combined with the evolution law of the failure probability vector, the flip node and the corresponding stage dominant source label are recursively calculated to form a flip node array and a stage dominant source array. The failure probability vectors of each data source, the first switchover time, the flip node array, and the phase-dominant source array are summarized to generate a risk assessment result.

5. The multi-source sensing data fusion processing method based on AI large model according to claim 1, characterized in that, Determine the dynamic stage, and based on the dynamic stage, perform weight allocation and cross-stage feature compensation on the multi-source feature sequence to obtain the delayed-corrected multi-source feature sequence, including: Dynamic stages are determined based on the flipped node array and the stage-dominant source array. The fusion weights are assigned to each dynamic stage based on the failure probability vector, where the fusion weights of the dominant sources are calculated according to the reliability ratio, thus obtaining the fusion weight vectors for each dynamic stage. The fusion weight vector is combined with the compressed feature vectors of each data source within the current sliding window, and weighted integration is performed to generate a fused multi-source feature sequence. Through dynamic stages, feature delay quantization is performed on the fused multi-source feature sequences, and feature shifts in future stages are predicted to complete cross-stage feature compensation.

6. The multi-source sensing data fusion processing method based on AI large model according to claim 1, characterized in that, Through a dynamic phase, feature delay quantization is performed on the fused multi-source feature sequence, and feature shifts in future phases are predicted to complete cross-phase feature compensation, including: Extract the fused multi-source feature sequences within each dynamic stage, and analyze the differences in feature response time of each data source within different dynamic stages to quantify the feature response delay magnitude and obtain the cross-stage feature delay matrix. Based on the cross-stage feature delay matrix, by analyzing the delay pattern and the evolution trend of each feature in the current dynamic stage, the offset direction and offset magnitude of the features in the future stage are predicted, and the delay correction prediction vector is obtained. Based on the delayed-corrected prediction vector and the fused multi-source feature sequence, feature compensation across dynamic stages is achieved by weighting and adjusting the feature values ​​of each data source, resulting in the delayed-corrected multi-source feature sequence.

7. The multi-source sensing data fusion processing method based on AI large model according to claim 1, characterized in that, Based on the delayed-corrected multi-source feature sequences, the anomaly propagation and coupling effects at each dynamic stage are analyzed to obtain the risk score vectors for each data source in the future stage. Delay compensation is then adjusted according to these risk score vectors to form a cross-stage closed-loop optimization, including: Based on the delayed-corrected multi-source feature sequences, the reconstruction error and mutual information mutation in each dynamic stage are analyzed to obtain the anomaly impact matrix summarized by stage and time step. Using the dynamic phase as a time reference and combining it with the anomaly impact matrix, the recursive simulation method is used to analyze the anomaly evolution pattern and obtain the risk score vector for each dynamic phase. When the risk scores of each data source within the risk scoring vector fall within the preset risk threshold range, the corresponding fusion weights in the original fusion weight vector are maintained; otherwise, a cross-stage optimization mechanism is executed.

8. The multi-source sensing data fusion processing method based on AI large model according to claim 1, characterized in that, The process of implementing a cross-stage optimization mechanism includes: Initiate a cross-stage optimization mechanism, and designate data sources whose risk scores do not fall within the preset risk threshold range as the objects to be adjusted; Perform a linear scaling mapping on the object to be adjusted individually to obtain the mapping coefficients, and multiply the mapping coefficients by the corresponding fusion weights in the original fusion weight vector to obtain the corrected fusion weight vector. Based on the corrected fusion weight vector, cross-stage feature compensation is repeatedly performed until the risk scores of each data source in the risk score vector fall within the preset risk threshold range. Then, the adjustment of the fusion weight vector is stopped, thus achieving cross-stage closed-loop optimization.

9. A multi-source sensing data fusion processing system based on an AI large-scale model, used to implement the multi-source sensing data fusion processing method based on an AI large-scale model as described in any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire heterogeneous sensing data, and writes it into the buffer in a standardized manner through dedicated protocol parsing and distribution offset detection. The data processing module is used to traverse the buffer to build an autoencoder and optimize its parameters. Based on the optimized autoencoder parameters, feature compression is performed to form a multi-source feature sequence. The predictive-driven module is used to perform forward inference on the large model by matching historical mutation case library and dynamic time warping, generate risk assessment results and determine dynamic stages. Based on the dynamic stages, weight allocation and cross-stage feature compensation are performed on the multi-source feature sequences to obtain the delayed-corrected multi-source feature sequences. The fusion optimization module is used to analyze the anomaly propagation and coupling effects of each dynamic stage based on the multi-source feature sequence after delay correction, obtain the risk score vector of each data source in the future stage, and adjust the delay compensation according to the risk score vector to form a cross-stage closed-loop optimization.

Images