Time series data semantic method and system, electronic device and computer storage medium

Abstract

The application provides a time series data semantic method and system, electronic equipment and computer storage medium, comprising: extracting initial time / frequency domain features of time series data; mapping the initial time / frequency domain features to time / frequency domain continuous features consistent with the dimension of the word embedding space of the large language model; configuring the time / frequency domain prompt in the embedding layer of the large language model; embedding the time / frequency domain continuous features and the time / frequency domain prompt into the text instruction, and performing full splicing in the sequence dimension to obtain an input sequence; decoding the input sequence to obtain a semantic diagnosis text. By obtaining time / frequency domain continuous features consistent with the dimension of the word embedding space of the large language model, and configuring the time / frequency domain prompt in the embedding layer of the large language model, the input sequence can retain the micro amplitude changes in the time series data, improving the accuracy of feature extraction and the reasoning completeness and timeliness of the large language model, and solving the problems of low precision and high delay of existing time series data semantic schemes.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and signal processing technology, and in particular to a method and system for semanticizing time-series data, an electronic device, and a computer storage medium. Background Technology

[0002] With the rapid development of fields such as Industrial Internet of Things (IIoT), real-time medical monitoring, and embodied intelligence, introducing Large Language Models (LLM) into high-frequency continuous sensor data processing to achieve automatic semanticization and real-time reasoning of physical time-series signals has become a significant technological trend in the industry. Existing typical technical solutions usually rely on the standard paradigm of "signal acquisition—feature discretization / dimensionality reduction and compression—single-stream large language model encoding—text generation." However, when facing complex real-world business scenarios, existing technical solutions have significant limitations in their underlying architecture and physical feature mapping mechanisms, specifically manifested in the following three main technical defects: 1. Semantic Entanglement Caused by Single-Stream Passive Encoding and Global Attention: Temporal signals in the physical world naturally contain two distinct information dimensions: transient energy transitions occurring at specific time points (representing actions / processes, temporal features), and steady-state energy distributions exhibited by the signal within a certain window period (representing states / properties, frequency domain features). Existing temporal large-scale language models generally employ single-stream encoders and unconstrained global attention mechanisms (Self-Attention), forcibly mixing and superimposing these two types of time-frequency physical features, which should be orthogonal, in the same sequence. Due to the lack of underlying isolation of physical boundaries in the existing model architecture, the continuous high-frequency steady-state background energy easily overwhelms weak transient changes in global computation, causing the model to be unable to decouple "sudden actions" from "background states." In complex real-world scenarios, the system often falls into severe semantic entanglement, not only easily misjudging stable high-frequency environmental states (such as continuous processing noise) as violent impact actions, but also potentially filtering out real early transient weak faults as background noise, causing the early warning system to lose diagnostic robustness.

[0003] 2. Loss of Feature Information and "Invisibility" of Early Faults Due to Discretization and Quantization Mechanisms: High-value industrial or medical early warning systems often rely on capturing extremely small changes (e.g., a vibration amplitude of only 0.001g in a high-precision machine tool). Existing technologies, to align with the vocabulary mechanisms used by large language models for processing discrete text, commonly employ discretization methods such as Vector Quantization Variational Autoencoders (VQ-VAE) to quantize continuous, high-precision floating-point sensor data into finite discrete token IDs. However, this lossy quantization mechanism, which maps high-precision continuous space to low-resolution discrete code, inevitably introduces quantization errors at the source, leading to the loss of information on weak feature amplitudes. This obscures the microscopic physical evolution trajectory, making it difficult for the system to capture early, weak anomalous changes. Threshold alarms are typically triggered only when serious equipment damage occurs, severely reducing the practical value of predictive maintenance.

[0004] 3. Information decay and decision lag caused by traditional feature compression mechanisms: To address the problem of ultra-long temporal inputs caused by high-frequency sampling, existing large temporal models typically employ dimensionality reduction and compression mechanisms such as pooling or strided convolution to significantly compress extremely long sequences into a few feature vectors before inputting them into the large language model. However, this "compress first, then infer" architecture causes significant feature representation bottlenecks and information decay during the feature extraction stage. This results in a large number of subtle but logically relevant low-level temporal details being filtered out or missed before entering the large model for deep logical inference, thus limiting the model's ability to capture complex related features and severely affecting the timeliness and accuracy of inference decisions when facing high-frequency streaming data.

[0005] In summary, existing technologies suffer from technical obstacles when processing physical time-series signals, such as the inability of the underlying architecture to decouple time-frequency physical attributes, lossy feature quantization, and loss of details during long sequence compression. A novel network architecture and feature mapping mechanism are urgently needed to overcome these shortcomings. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for semanticizing time-series data, an electronic device, and a computer storage medium, so as to at least solve the problems of low accuracy and high latency in existing semanticizing time-series data schemes.

[0007] To address the aforementioned technical problems, this invention provides a method for semanticizing time-series data, comprising: Acquire time series data; Extract the initial time-domain and initial frequency-domain features of the time-series data; The initial temporal and frequency domain features are mapped to temporal and frequency domain continuous features consistent with the word embedding space dimension of the large language model. Configure time-domain prompts and frequency-domain prompts in the embedding layer of the large language model; The text instructions are embedded with continuous features in the time domain, continuous features in the frequency domain, time domain prompts, and frequency domain prompts, and then fully concatenated in the sequence dimension to obtain the input sequence. The input sequence is decoded using a large language model to obtain semantic diagnostic text.

[0008] Optionally, in the aforementioned time-series data semanticization method, the method for extracting the initial time-domain features and initial frequency-domain features of the time-series data includes: Based on the set sliding window, the time series data is segmented and extracted to obtain the initial time domain features and the initial frequency domain features.

[0009] Optionally, in the aforementioned time-series data semanticization method, the method for extracting the initial time-domain features and initial frequency-domain features of the time-series data includes: Initial temporal features are extracted using a local feature aggregation network; The initial frequency domain features are extracted using a neural network.

[0010] Optionally, in the aforementioned temporal data semanticization method, the method for mapping initial temporal features and initial frequency domain features to temporal continuous features and frequency domain continuous features consistent with the word embedding space dimension of a large language model includes: Constructing a cross-modal projection network based on a multilayer perceptron; A cross-modal projection network is used to perform high-dimensional continuous mapping on the initial time-domain features and initial frequency-domain features to obtain continuous features in both the time and frequency domains.

[0011] Optionally, in the aforementioned temporal data semanticization method, the method for configuring temporal and frequency domain cues in the embedding layer of the large language model includes: During the joint training phase, deep orthogonal penalty loss function constraints are applied to the parameters of the time-domain and frequency-domain cues and the implicit states of each intermediate layer of the decoder of the large language model.

[0012] Optionally, in the aforementioned temporal data semanticization method, the deep orthogonal penalty loss function is expressed as:

[0013] in, Indicates the time-domain prompt, This represents the frequency domain prompt, and M represents the total number of decoder layers in the large language model. This represents the implicit temporal state of the l-th layer. This represents the implicit frequency domain state of the l-th layer. Denotes the Frobenius norm of a matrix. This represents the weighting coefficient.

[0014] Optionally, in the aforementioned temporal data semanticization method, the method of decoding the input sequence using a large language model to obtain semantic diagnostic text includes: When performing self-attention computation in a large language model, a domain isolation attention mask matrix is ​​applied to block cross-domain feature extraction and extract transient action features and steady-state features. When decoding the input sequence, autoregressive decoding is performed based on transient action features and steady-state features to output semantic diagnostic text.

[0015] Optionally, in the aforementioned temporal data semanticization method, the configuration logic for using the application domain isolation attention mask matrix to block cross-domain feature extraction and extract transient action features and steady-state features includes: When the query vector belongs to the time-domain prompt and the key vector belongs to the frequency-domain continuous feature, or when the query vector belongs to the frequency-domain prompt and the key vector belongs to the time-domain continuous feature, the mask value at the corresponding position is set to negative infinity, so that the cross-domain attention weight after Softmax activation is forced to zero.

[0016] Optionally, in the aforementioned temporal data semanticization method, the configuration logic for using the application domain isolation attention mask matrix to block cross-domain feature extraction and extract transient action features and steady-state features further includes: When the query vector belongs to the time domain prompt and the key vector belongs to the time domain continuous feature, or when the query vector belongs to the frequency domain prompt and the key vector belongs to the frequency domain continuous feature, set the mask value of the corresponding position to 0 in order to extract transient action features and steady-state features.

[0017] Optionally, in the aforementioned temporal data semanticization method, the configuration logic for performing autoregressive decoding based on transient action features and steady-state features to output semantic diagnostic text includes: For the query vector of the text segmentation to be generated, the attention mask value of all concatenated elements in the input sequence is set to 0, so as to allow the large language model to globally access and fuse transient action features and steady-state features when generating semantic diagnostic text.

[0018] To address the aforementioned technical problems, the present invention also provides a time-series data semanticization system for implementing the time-series data semanticization method as described in any of the preceding claims, wherein the time-series data semanticization system comprises: The data acquisition module is used to acquire time-series data; The feature extraction and mapping module is used to extract the initial time-domain features and initial frequency-domain features of time-series data, and also to map the initial time-domain features and initial frequency-domain features to time-domain continuous features and frequency-domain continuous features consistent with the word embedding space dimension of the large language model. The model configuration module is used to configure time-domain and frequency-domain prompts in the embedding layer of a large language model; The large language model module includes a large language model, which is used to embed temporal continuous features, frequency continuous features, temporal prompts and frequency prompts into text instructions, and to perform full concatenation in the sequence dimension to obtain the input sequence; it is also used to decode the input sequence to obtain semantic diagnostic text.

[0019] Optionally, in the aforementioned temporal data semanticization system, the feature extraction and mapping module includes: The sliding window segmentation unit is used to segment time series data based on a set sliding window to obtain fragment data; The temporal feature extraction and mapping unit is used to extract initial temporal features from the temporal branches of fragment data and map the initial temporal features to temporal continuous features consistent with the word embedding space dimension of the large language model. The frequency domain feature extraction and mapping unit is used to extract initial frequency domain features from the frequency domain branch of the fragment data and map the initial frequency domain features to continuous frequency domain features consistent with the word embedding space dimension of the large language model.

[0020] Optionally, in the aforementioned temporal data semanticization system, the large language model includes: The sequence construction unit is used to embed time-domain continuous features, frequency-domain continuous features, time-domain prompts, and frequency-domain prompts into text instructions, and then perform full concatenation in the sequence dimension to obtain the input sequence; Domain isolation masking units are used to block cross-domain feature extraction by applying a domain isolation attention masking matrix when performing self-attention computation in a large language model, and to extract transient action features and steady-state features. The semantic diagnosis generation unit is used to perform autoregressive decoding based on transient action features and steady-state features to output semantic diagnostic text.

[0021] To address the aforementioned technical problems, the present invention also provides an electronic device, including a memory, a processor, and an executable program stored in the memory and capable of being run by the processor; when the processor runs the executable program, it executes the timing data semanticization method as described in any of the preceding claims.

[0022] To address the aforementioned technical problems, the present invention also provides a computer storage medium storing an executable program; when the executable program is executed, it implements the time-series data semanticization method as described in any of the preceding claims.

[0023] The present invention provides a method and system for semanticizing time-series data, an electronic device, and a computer storage medium, comprising: acquiring time-series data; extracting initial time-domain features and initial frequency-domain features of the time-series data; mapping the initial time-domain features and initial frequency-domain features to continuous time-domain features and continuous frequency-domain features consistent with the word embedding space dimension of a large language model; configuring time-domain prompts and frequency-domain prompts in the embedding layer of the large language model; embedding the continuous time-domain features, continuous frequency-domain features, time-domain prompts, and frequency-domain prompts into text instructions, and performing full concatenation in the sequence dimension to obtain an input sequence; and decoding the input sequence using the large language model to obtain semantically coded diagnostic text. By mapping the initial temporal and frequency domain features to continuous temporal and frequency domain features consistent with the word embedding space dimension of the large language model, and configuring temporal and frequency domain cues in the embedding layer of the large language model, the input sequence obtained after full concatenation in the sequence dimension can retain the micro-amplitude changes in the original time-series data. This improves the accuracy of feature extraction and enhances the inference completeness and timeliness of the large language model decoding, thus solving the problems of low accuracy and high latency in existing time-series data semanticization schemes. Attached Figure Description

[0024] Figure 1 This is a flowchart of the time-series data semanticization method provided in this embodiment; Figure 2 This is a schematic diagram of the internal calculation constraint logic of the domain isolation attention mask matrix provided in this embodiment; Figure 3 This is a schematic diagram of the structure of the temporal data semanticization system provided in this embodiment; Figure 4 This is a schematic diagram of the feature extraction and mapping module provided in this embodiment; Figure 5 This is a schematic diagram of the structure of the large language model provided in this embodiment. Detailed Implementation

[0025] The semantic representation method and system for time-series data, electronic devices, and computer storage media proposed in this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the drawings are all in a very simplified form and use non-precise scales, and are only used to facilitate and clarify the illustration of the embodiments of this invention. Furthermore, the structures shown in the drawings are often part of the actual structures. In particular, different figures may emphasize different aspects and sometimes use different scales.

[0026] It should be noted that the terms "first," "second," etc., used in the specification, claims, and drawings of this invention are used to distinguish similar objects in order to describe embodiments of the invention, and are not used to describe a specific order or sequence. It should be understood that such uses of terminology are interchangeable where appropriate. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] This embodiment provides a method for semanticizing time-series data, such as... Figure 1 As shown, it includes: S1, acquire time series data; S2, extract the initial time-domain features and initial frequency-domain features of the time-series data; S3 maps the initial temporal and frequency domain features to temporal and frequency domain continuous features consistent with the word embedding space dimension of the large language model. S4, configure time-domain prompts and frequency-domain prompts in the embedding layer of the large language model; S5 embeds the time-domain continuous features, frequency-domain continuous features, time-domain prompts, and frequency-domain prompts into the text instructions, and performs full concatenation in the sequence dimension to obtain the input sequence; S6 uses a large language model to decode the input sequence to obtain semantic diagnostic text.

[0028] The temporal data semanticization method provided in this embodiment maps initial temporal and frequency domain features to temporal and frequency domain continuous features consistent with the word embedding space dimension of the large language model. It also configures temporal and frequency domain prompts in the embedding layer of the large language model. This allows the input sequence obtained after full concatenation in the sequence dimension to retain the micro-amplitude changes in the original temporal data, improves the accuracy of feature extraction, and enhances the inference integrity and timeliness of the large language model decoding. This solves the problems of low accuracy and high latency in existing temporal data semanticization schemes.

[0029] Specifically, in this embodiment, step S1 involves acquiring timing data.

[0030] In practical applications, time-series data can be acquired using signal acquisition devices, such as sensors. The acquired time-series data is typically continuous streaming time-series data.

[0031] Furthermore, in this embodiment, step S2 involves extracting the initial time-domain features and initial frequency-domain features of the time-series data.

[0032] To facilitate the subsequent extraction of initial time-domain and initial frequency-domain features, in this embodiment, the time-series data is first segmented based on a set sliding window (time step size and overlap rate), thereby dividing the long-series streaming time-series data into segments of fixed-length local time windows. Then, feature extraction is performed to obtain the initial time-domain and initial frequency-domain features.

[0033] Specifically, in this embodiment, a local feature aggregation network is used to extract initial time-domain features; and a neural network is used to extract initial frequency-domain features. The specific implementation methods of using the local feature aggregation network and the neural network to extract features are well known to those skilled in the art, and will not be elaborated upon here.

[0034] In one specific embodiment, the input continuous time-series data is assumed to be a matrix. Where L represents the sequence length and C represents the number of channels of the signal acquisition device.

[0035] For the extraction of initial temporal features, a one-dimensional convolutional neural network (1D-CNN) with residual connections can be used to perform local feature aggregation while maintaining the temporal topology, and extract temporal feature maps that represent transient transitions to obtain initial temporal features.

[0036] To extract the initial frequency domain features, the one-dimensional time-series signal can be converted into a two-dimensional time-frequency graph using the short-time Fourier transform. Then, a two-dimensional convolutional neural network (2D-CNN) can be used to extract the steady-state energy distribution features in the two-dimensional time-frequency graph to obtain the initial frequency domain features.

[0037] Furthermore, in this embodiment, step S3 maps the initial temporal features and initial frequency domain features to temporal continuous features and frequency continuous features consistent with the word embedding space dimension of the large language model.

[0038] Specifically, in this embodiment, firstly, a cross-modal projection network is constructed based on a multilayer perceptron (MLP); then, the cross-modal projection network is used to perform high-dimensional continuous mapping on the initial time-domain features and initial frequency-domain features to obtain continuous time-domain features and continuous frequency-domain features. The continuous time-domain features represent transient transitions, while the continuous frequency-domain features represent steady-state distributions.

[0039] Using the example above, for continuous features in the time domain, a multilayer perceptron can be used to map the time-domain feature map representing transient transitions into continuous features in the time domain, denoted as . ,in, Let represent the number of time steps after downsampling, and D represent the embedding dimension of the large language model. For continuous features in the frequency domain, a multilayer perceptron can be used to map the steady-state energy distribution features to continuous features in the frequency domain, denoted as . ,in, Indicates the number of frequency segments.

[0040] Thus, throughout the entire process, data flows in the form of a high-precision floating-point matrix, without accessing any quantizer modules (such as vector quantization variational autoencoders (VQ-VAE) or other discretization mapping modules) that map floating-point vectors to finite word indexes (token IDs). This effectively avoids quantization errors generated during the mapping process from continuous floating-point data to discrete encoding. Simultaneously, this embodiment achieves high-fidelity preservation of the microscopic physical evolution trajectory of the original time-series signal by constructing a cross-modal continuous high-dimensional feature mapping channel. This effectively solves the technical problem of losing weak anomalous features (such as small amplitude jumps) during data preprocessing, significantly improving the system's ability to identify minute physical changes.

[0041] Of course, in other embodiments, a continuous wavelet transform (CWT) can be used to directly obtain a time-frequency continuous feature map with both high temporal and frequency resolution; subsequently, a continuous feature projector based on a visual transformer (ViT) or a state-space model (such as the Mamba architecture) is used to obtain continuous features in the time and frequency domains. This scheme also does not involve discretization and quantization operations throughout the entire process, and can achieve high-fidelity, lossless cross-modal mapping into the embedding space of a large language model.

[0042] Furthermore, in this embodiment, step S4 involves configuring time-domain prompts and frequency-domain prompts in the embedding layer of the large language model.

[0043] Typically, a large language model includes an embedding layer and a processing layer, the specific architecture of which will not be elaborated in this application. In this embodiment, learnable time-domain cues and frequency-domain cues are initialized in the embedding layer of the large language model.

[0044] To ensure strict decoupling between time-domain and frequency-domain prompts during decoding, this embodiment applies deep orthogonal penalty loss function constraints to the parameters of the time-domain and frequency-domain prompts, as well as the implicit states of each intermediate layer of the decoder in the large language model, during the joint training phase. These deep orthogonal penalty loss function constraints force the inner product of the time-domain and frequency-domain prompts to approach zero, thus achieving orthogonality of the vector space.

[0045] This embodiment presents a deep orthogonal penalty loss function, expressed as follows:

[0046] in, Indicates the time-domain prompt, , Indicates the frequency domain prompt. K represents the prompt length; M represents the total number of decoder layers in the large language model. This represents the implicit state in the time domain of the l-th layer. This represents the implicit frequency domain state of the l-th layer. Denotes the Frobenius norm of a matrix. This represents the weighting coefficient.

[0047] Of course, in other embodiments, contrastive learning loss functions (such as InfoNCE loss) or cosine distance penalty loss can also be used as deep orthogonal penalty loss functions. That is, by maximizing the similarity of features in the same domain and minimizing the similarity of features across domains (such as between time-domain cues and frequency-domain hidden states), the large language model is forced to maintain the feature independence and absolute decoupling of time-domain cues and frequency-domain cues in multi-layer decoding.

[0048] This embodiment configures temporal and frequency domain prompts in the embedding layer of a large language model, enabling the use of continuous vectors as query anchors to guide feature extraction. Simultaneously, by applying deep orthogonal penalty loss function constraints to the parameters of the temporal and frequency domain prompts and the implicit states of each intermediate layer of the large language model's decoder, it forces these two sets of prompts to maintain strict feature decoupling during the multi-layer decoding process of the large language model. This ensures that the temporal prompts are dedicated to extracting transient features, and the frequency domain prompts are dedicated to extracting steady-state features.

[0049] In practical applications, the deep orthogonal penalty loss function can be jointly optimized with the cross-entropy loss in the total loss function to improve the robustness of the system.

[0050] Specifically, the total loss function is achieved by a weighted combination of cross-entropy loss and orthogonality loss, and the total loss function is expressed as:

[0051] in, This indicates the preceding context, and T represents the sequence length. Indicates hyperparameters, This represents the input sequence.

[0052] Furthermore, in this embodiment, step S5 involves embedding the time-domain continuous features, frequency-domain continuous features, time-domain prompts, and frequency-domain prompts into the text instruction, and then performing full concatenation in the sequence dimension to obtain the input sequence.

[0053] Specifically, in this embodiment, text instructions are used. Time domain prompt Frequency domain prompt Continuous characteristics in the time domain Continuous characteristics in the frequency domain Performing full concatenation along the sequence dimension, the resulting input sequence is represented as follows: .

[0054] This embodiment completely avoids the information attenuation caused by the traditional "compression first, reasoning later" architecture by splicing all data, ensuring that the large language model has absolutely complete basic physical clues when performing deep logical reasoning.

[0055] Furthermore, in this embodiment, step S6 involves using a large language model to decode the input sequence to obtain semantic diagnostic text.

[0056] Specifically, when a large language model performs self-attention calculation, the calculation logic for its score matrix is ​​as follows:

[0057] Where M represents the structured domain isolation attention mask matrix.

[0058] In this embodiment, a domain isolation attention mask matrix is ​​applied to block cross-domain feature extraction, and transient action features and steady-state features are extracted. For example, Figure 2 As shown, the elements in matrix M The specific assignment action relationship configuration logic includes: Mask blocking action: when query vector Belongs to the time-domain prompt and key vector When it belongs to the continuous features in the frequency domain, or when the query vector Belongs to the frequency domain prompt and key vector When the feature is continuous in the time domain, let the mask value at the corresponding position be negative infinity. This forces the cross-domain attention weights after Softmax activation to zero, physically cutting off the reading path of cross-domain features.

[0059] And, same-domain extraction action: when query vector Belongs to the time-domain prompt and key vector When it belongs to the continuous features in the time domain, or when the query vector Belongs to the frequency domain prompt and key vector When the feature is continuous in the frequency domain, the mask value at the corresponding position is set to 0 to extract transient action features and steady-state features normally.

[0060] This embodiment forcibly blocks features from different physical domains at the underlying matrix calculation level, preventing cross-domain feature confusion during global attention calculation in large language models. It also solves the problem of continuous high-frequency steady-state background energy overshadowing weak transient changes during calculation, enabling the model to accurately decouple independent time-frequency features in complex noisy environments and significantly improving the accuracy of classification reasoning. Simultaneously, the introduced cue points act as "retrieval anchors" for specific physical features in global attention calculation. They dynamically increase the model's attention weight for diagnostically necessary content (such as extremely weak transient jumps) within a massive, fully concatenated set of continuous features. This effectively guides the model to focus on core features while preserving all information without loss, thus improving overall performance.

[0061] Of course, in other embodiments, at the underlying feature isolation mechanism, multi-head attention channel binding or hard gating can be used instead of the domain isolation attention mask matrix to block cross-domain feature extraction. Specifically, the attention heads of the large language model can be statically and physically divided, specifying that some attention heads can only receive and compute time-domain input, while others can only receive and compute frequency-domain input; or, before the inner product calculation of the query vector and key vector, a binary gating network layer with an output of 0 or 1 can be concatenated to physically cut off the cross-domain data flow. This alternative can also achieve absolute blocking of cross-domain features at the underlying matrix level.

[0062] Meanwhile, in this embodiment, when decoding the input sequence, autoregressive decoding is performed based on transient action features and steady-state features to output semantic diagnostic text. For example, Figure 2 As shown, the elements in matrix M The specific assignment action relationship configuration logic includes: Global Fusion Actions: For the query vector of the text segmentation to be generated, the attention mask value of all concatenated elements in the input sequence is set to 0, that is, the mask value of all key vectors in the input sequence is set to 0, so as to allow the large language model to globally access and fuse transient action features and steady-state features when generating semantic diagnostic text.

[0063] The temporal data semanticization method provided in this embodiment directly parses high-frequency continuous physical sensor signals into high-precision expert-level diagnostic text through lossless mapping of dual-domain features, orthogonal continuous prompts, and a domain-isolated attention masking mechanism. This addresses the semantic entanglement problem between transient actions and steady-state states in complex environments. To address the issue of time-frequency feature aliasing caused by the global attention mechanism in existing single-stream coding architectures when processing physical time-series signals containing both transient actions in the time domain and steady-state states in the frequency domain, this embodiment provides a time-series data semanticization method. By orthogonally constraining and isolating time-domain and frequency-domain features in the underlying computing architecture, this method prevents continuous high-frequency steady-state features from interfering with transient abrupt features, improves the accuracy of the model in extracting independent physical features under complex background noise, and solves the technical problem of low feature extraction accuracy caused by time-frequency feature aliasing.

[0064] Furthermore, addressing the shortcomings of existing technologies that employ discretization quantization methods (such as vector quantization variational autoencoders) to process continuous floating-point sensor data, including quantization errors and loss of weak physical features, this embodiment provides a temporal data semanticization method. This method circumvents discretization segmentation operations through cross-modal continuous high-dimensional feature mapping, enabling direct projection of temporal continuous signals into the embedding space of a large language model. It preserves the micro-amplitude variations of the original signal, improves the recognition rate of weak abnormal features, and solves the technical problem of loss of weak temporal features caused by the discretization quantization mechanism.

[0065] Furthermore, addressing the bottlenecks in feature representation and the shortcomings in filtering underlying information that exist in existing dimensionality reduction and compression mechanisms (such as pooling operations) when processing high-frequency long sequence data, the temporal data semanticization method provided in this embodiment introduces orthogonal continuous prompts to guide the input of all dual-domain physical features through an input architecture based on full feature concatenation and a specific attention mechanism. This avoids the loss of clues caused by excessive dimensionality reduction in the early stage, improves the completeness and timeliness of deep inference of the model, and solves the technical problems of loss of underlying details and inference delay caused by long sequence dimensionality reduction and compression.

[0066] The semanticization method for time-series data provided in this embodiment can be widely applied to business areas that require low-latency and highly robust decision-making on streaming time-series data, such as the Industrial Internet of Things (e.g., predictive maintenance of high-end equipment), embodied intelligence (e.g., real-time status perception of robots), and medical and health monitoring (e.g., emergency warning of dynamic physiological signals).

[0067] Specifically, in terms of productivity and system reliability, the domain isolation attention mechanism separates transient and steady-state characteristics, effectively reducing the false alarm rate caused by noise interference in complex environments. This improves the accuracy of abnormal state diagnosis for automated equipment (such as industrial machine tools and intelligent robots), reduces the frequency of unplanned downtime, and thus enhances the overall uptime and production continuity of the equipment.

[0068] Furthermore, in terms of operation and maintenance costs, the loss of weak fault characteristics during the quantization stage is avoided by retaining the continuous features of the dual domains without damage. This enables the system to identify minute mechanical deformations or early abnormal physiological signals, realize early detection and warning of faults, and reduce the cost of component replacement and downtime maintenance caused by severe damage to equipment in the later stages.

[0069] Furthermore, in terms of computational energy consumption and inference efficiency, a cross-modal projection network and a local feature aggregation module are adopted to replace the traditional discretization encoding and decoding modules with high computational complexity. This simplifies the front-end data preprocessing link, reduces the computing power overhead and operating power consumption of edge computing devices, and shortens the end-to-end data processing latency to meet the real-time requirements of high-frequency streaming data processing.

[0070] This embodiment also provides a time-series data semanticization system for implementing the time-series data semanticization method described above, such as... Figure 3 As shown, the time-series data semanticization system includes: The data acquisition module is used to acquire time-series data; The feature extraction and mapping module is used to extract the initial time-domain features and initial frequency-domain features of time-series data, and also to map the initial time-domain features and initial frequency-domain features to time-domain continuous features and frequency-domain continuous features consistent with the word embedding space dimension of the large language model. The model configuration module is used to configure time-domain and frequency-domain prompts in the embedding layer of a large language model; The large language model module includes a large language model, which is used to embed temporal continuous features, frequency continuous features, temporal prompts and frequency prompts into text instructions, and to perform full concatenation in the sequence dimension to obtain the input sequence; it is also used to decode the input sequence to obtain semantic diagnostic text.

[0071] The temporal data semanticization system provided in this embodiment maps initial temporal and frequency domain features to continuous temporal and frequency domain features consistent with the word embedding space dimension of the large language model through a feature extraction and mapping module. Furthermore, a model configuration module configures temporal and frequency domain prompts in the embedding layer of the large language model. This allows the input sequence obtained after the large language model module performs full concatenation in the sequence dimension to retain the micro-amplitude variations in the original temporal data, improving the accuracy of feature extraction and enhancing the inference completeness and timeliness of the large language model's decoding. This solves the problems of low accuracy and high latency in existing temporal data semanticization schemes.

[0072] Specifically, in this embodiment, such as Figure 4 As shown, the feature extraction mapping module includes: The sliding window segmentation unit is used to segment time series data based on a set sliding window to obtain fragment data; The temporal feature extraction and mapping unit is used to extract initial temporal features from the temporal branches of fragment data and map the initial temporal features to temporal continuous features consistent with the word embedding space dimension of the large language model. The frequency domain feature extraction and mapping unit is used to extract initial frequency domain features from the frequency domain branch of the fragment data and map the initial frequency domain features to continuous frequency domain features consistent with the word embedding space dimension of the large language model.

[0073] In one specific embodiment, the temporal feature extraction and mapping unit first employs a one-dimensional convolutional neural network (1D-CNN) with residual connections to perform local feature aggregation while preserving the temporal topology, extracting temporal feature maps representing transient transitions to obtain initial temporal features. Then, a temporal cross-modal projection network (based on a multilayer perceptron MLP) is used to map the temporal feature maps representing transient transitions into continuous temporal features. Similarly, the frequency domain feature extraction and mapping unit first uses a short-time Fourier transform to convert the one-dimensional time-series signal into a two-dimensional time-frequency map. Next, a two-dimensional convolutional neural network (2D-CNN) is used to extract the steady-state energy distribution features from this two-dimensional time-frequency map to obtain initial frequency domain features. Finally, a frequency domain cross-modal projection network (based on a multilayer perceptron MLP) is used to map the steady-state energy distribution features into continuous frequency domain features.

[0074] Furthermore, in this embodiment, as Figure 5 As shown, the large language model (specifically, the processing layer of the large language model) includes: The sequence construction unit is used to embed time-domain continuous features, frequency-domain continuous features, time-domain prompts, and frequency-domain prompts into text instructions, and then perform full concatenation in the sequence dimension to obtain the input sequence; Domain isolation masking units are used to block cross-domain feature extraction by applying a domain isolation attention masking matrix when performing self-attention computation in a large language model, and to extract transient action features and steady-state features. The semantic diagnosis generation unit is used to perform autoregressive decoding based on transient action features and steady-state features to output semantic diagnostic text.

[0075] The temporal data semanticization system provided in this embodiment directly parses high-frequency continuous physical sensor signals into high-precision expert-level diagnostic text through lossless mapping of dual-domain features, orthogonal continuous prompts, and a domain-isolated attention masking mechanism. This solves the semantic entanglement problem between transient actions and steady-state states in complex environments. The temporal data semanticization system provided in this embodiment can be widely applied in business areas requiring low-latency, highly robust decision-making from streaming time-series data, such as the Industrial Internet of Things, embodied intelligence, and medical health monitoring.

[0076] Furthermore, this embodiment also provides an electronic device, including a memory, a processor, and an executable program stored in the memory and capable of being run by the processor; when the processor runs the executable program, it performs the timing data semanticization method as described above.

[0077] Furthermore, this embodiment also provides a computer storage medium storing an executable program; when the executable program is executed, it implements the time-series data semanticization method as described above.

[0078] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to mutually. In addition, different parts between embodiments can also be combined with each other, and this invention does not limit this.

[0079] The semanticization method and system for time-series data, electronic device, and computer storage medium provided in this embodiment include: acquiring time-series data; extracting initial time-domain features and initial frequency-domain features from the time-series data; mapping the initial time-domain features and initial frequency-domain features to continuous time-domain features and continuous frequency-domain features consistent with the word embedding space dimension of a large language model; configuring time-domain prompts and frequency-domain prompts in the embedding layer of the large language model; embedding the continuous time-domain features, continuous frequency-domain features, time-domain prompts, and frequency-domain prompts into text instructions, and performing full concatenation in the sequence dimension to obtain an input sequence; and decoding the input sequence using the large language model to obtain semantic diagnostic text. By mapping the initial temporal and frequency domain features to continuous temporal and frequency domain features consistent with the word embedding space dimension of the large language model, and configuring temporal and frequency domain cues in the embedding layer of the large language model, the input sequence obtained after full concatenation in the sequence dimension can retain the micro-amplitude changes in the original time-series data. This improves the accuracy of feature extraction and enhances the inference completeness and timeliness of the large language model decoding, thus solving the problems of low accuracy and high latency in existing time-series data semanticization schemes.

[0080] The above description is merely a description of preferred embodiments of the present invention and is not intended to limit the scope of the present invention in any way. Any changes or modifications made by those skilled in the art based on the above disclosure shall fall within the protection scope of the claims.

Claims

1. A method for semanticizing time-series data, characterized in that, include: Acquire time series data; Extract the initial time-domain and initial frequency-domain features of the time-series data; The initial temporal and frequency domain features are mapped to temporal and frequency domain continuous features consistent with the word embedding space dimension of the large language model. Configure time-domain prompts and frequency-domain prompts in the embedding layer of the large language model; The text instructions are embedded with continuous features in the time domain, continuous features in the frequency domain, time domain prompts, and frequency domain prompts, and then fully concatenated in the sequence dimension to obtain the input sequence. The input sequence is decoded using a large language model to obtain semantic diagnostic text.

2. The semantic data representation method for time series data according to claim 1, characterized in that, The method for extracting the initial time-domain features and initial frequency-domain features of time-series data includes: Based on the set sliding window, the time series data is segmented and extracted to obtain the initial time domain features and the initial frequency domain features.

3. The semanticization method for time-series data according to any one of claims 1 or 2, characterized in that, The method for extracting the initial time-domain features and initial frequency-domain features of time-series data includes: Initial temporal features are extracted using a local feature aggregation network; The initial frequency domain features are extracted using a neural network.

4. The semantic data representation method for time series data according to claim 1, characterized in that, The method for mapping initial temporal and initial frequency domain features to temporal and frequency domain continuous features consistent with the word embedding space dimension of a large language model includes: Constructing a cross-modal projection network based on a multilayer perceptron; A cross-modal projection network is used to perform high-dimensional continuous mapping on the initial time-domain features and initial frequency-domain features to obtain continuous features in both the time and frequency domains.

5. The semantic data representation method for time series data according to claim 1, characterized in that, The method for configuring time-domain and frequency-domain prompts at the embedding layer of a large language model includes: During the joint training phase, deep orthogonal penalty loss function constraints are applied to the parameters of the time-domain and frequency-domain cues and the implicit states of each intermediate layer of the decoder of the large language model.

6. The semanticization method for time-series data according to claim 5, characterized in that, The deep orthogonal penalty loss function is expressed as follows: in, Indicates the time-domain prompt, This represents the frequency domain prompt, and M represents the total number of decoder layers in the large language model. This represents the implicit state in the time domain of the l-th layer. This represents the implicit frequency domain state of the l-th layer. Denotes the Frobenius norm of a matrix. This represents the weighting coefficient.

7. The semantic data representation method for time series data according to claim 1, characterized in that, The method of decoding the input sequence using a large language model to obtain semantic diagnostic text includes: When performing self-attention computation in a large language model, a domain isolation attention mask matrix is ​​applied to block cross-domain feature extraction and extract transient action features and steady-state features. When decoding the input sequence, autoregressive decoding is performed based on transient action features and steady-state features to output semantic diagnostic text.

8. The semantic data representation method for time series data according to claim 7, characterized in that, The configuration logic for using the application domain isolation attention mask matrix to block cross-domain feature extraction and extract transient action features and steady-state features includes: When the query vector belongs to the time-domain prompt and the key vector belongs to the frequency-domain continuous feature, or when the query vector belongs to the frequency-domain prompt and the key vector belongs to the time-domain continuous feature, the mask value at the corresponding position is set to negative infinity, so that the cross-domain attention weight after Softmax activation is forced to zero.

9. The semanticization method for time-series data according to claim 7, characterized in that, The configuration logic for using the application domain isolation attention mask matrix to block cross-domain feature extraction and extract transient action features and steady-state features also includes: When the query vector belongs to the time domain prompt and the key vector belongs to the time domain continuous feature, or when the query vector belongs to the frequency domain prompt and the key vector belongs to the frequency domain continuous feature, set the mask value of the corresponding position to 0 in order to extract transient action features and steady-state features.

10. The semantic representation method for time-series data according to claim 7, characterized in that, The configuration logic for performing autoregressive decoding based on transient action features and steady-state features to output semantic diagnostic text includes: For the query vector of the text segmentation to be generated, the attention mask value of all concatenated elements in the input sequence is set to 0, so as to allow the large language model to globally access and fuse transient action features and steady-state features when generating semantic diagnostic text.

11. A semantic data system for time series data, used to implement the semantic data system for time series data as described in any one of claims 1 to 10, characterized in that, The time-series data semanticization system includes: The data acquisition module is used to acquire time-series data; The feature extraction and mapping module is used to extract the initial time-domain features and initial frequency-domain features of time-series data, and also to map the initial time-domain features and initial frequency-domain features to time-domain continuous features and frequency-domain continuous features consistent with the word embedding space dimension of the large language model. The model configuration module is used to configure time-domain and frequency-domain prompts in the embedding layer of a large language model; The large language model module includes a large language model, which is used to embed temporal continuous features, frequency continuous features, temporal prompts and frequency prompts into text instructions, and to perform full concatenation in the sequence dimension to obtain the input sequence; it is also used to decode the input sequence to obtain semantic diagnostic text.

12. The temporal data semanticization system according to claim 11, characterized in that, The feature extraction mapping module includes: The sliding window segmentation unit is used to segment time series data based on a set sliding window to obtain fragment data; The temporal feature extraction and mapping unit is used to extract initial temporal features from the temporal branches of fragment data and map the initial temporal features to temporal continuous features consistent with the word embedding space dimension of the large language model. The frequency domain feature extraction and mapping unit is used to extract initial frequency domain features from the frequency domain branch of the fragment data and map the initial frequency domain features to continuous frequency domain features consistent with the word embedding space dimension of the large language model.

13. The temporal data semanticization system according to claim 11, characterized in that, The large language model includes: The sequence construction unit is used to embed time-domain continuous features, frequency-domain continuous features, time-domain prompts, and frequency-domain prompts into text instructions, and then perform full concatenation in the sequence dimension to obtain the input sequence; Domain isolation masking units are used to block cross-domain feature extraction by applying a domain isolation attention masking matrix when performing self-attention computation in a large language model, and to extract transient action features and steady-state features. The semantic diagnosis generation unit is used to perform autoregressive decoding based on transient action features and steady-state features to output semantic diagnostic text.

14. An electronic device, characterized in that, It includes a memory, a processor, and an executable program stored in the memory and capable of being run by the processor; when the processor runs the executable program, it performs the timing data semanticization method as described in any one of claims 1 to 10.

15. A computer storage medium, characterized in that, The computer storage medium stores an executable program; when the executable program is executed, it implements the time-series data semanticization method as described in any one of claims 1 to 10.

Images