Time series-based multi-regional questionnaire data fusion and questionnaire dynamic generation method

Abstract

The application discloses a time series-based multi-territory questionnaire data fusion and questionnaire dynamic generation method, which comprises the following steps: after capturing multi-source questionnaire data from different questionnaire survey terminals, adopting a lightweight time sequence ontology model to standardize the semantics in the multi-source questionnaire data; establishing a soft alignment mechanism to perform adaptive cross-territory space-time alignment; constructing a hyperbolic space-embedded space-time semantic graph to perform cross-cultural dimension alignment; further adopting an adaptive item weaving engine to dynamically generate the content of subsequent survey questionnaires; finally, using topological sorting to identify a parallel execution set and a sequential execution set, and adopting a genetic algorithm to perform heterogeneous data resource scheduling; and based on an interpretable space-time prediction model, performing multi-territory and multi-dimensional prediction feedback. The application solves the problems of timestamp ambiguity and semantic uncertainty in the multi-territory questionnaire data fusion process, and significantly improves the final questionnaire data prediction evaluation precision and feedback accuracy.

Description

Technical Field

[0001] This invention relates to the field of questionnaire data processing technology, specifically to a method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series. Background Technology

[0002] Traditional automated questionnaire generation methods primarily rely on pre-set formatted frameworks for uniform output. Time-series surveys, however, require dynamic adjustments to subsequent questionnaire content based on preliminary survey results, as well as analysis of temporal data correlations across different regions to reveal their mutual influence and mechanisms. However, existing pre-defined, fixed questionnaire designs prevent real-time sharing of dynamically adjusted data, and online multimedia questionnaire data resides independently on different terminals. When time-series forecasting based on multi-dimensional survey data is required, the current method of using isolated data from a single terminal results in low prediction accuracy and poor reliability.

[0003] In existing questionnaire survey systems, survey data typically comes from different regions and survey channels, such as online questionnaires, face-to-face interviews, telephone interviews, and sensor data. Acquiring this questionnaire data is usually done through either "passive reception" (waiting for terminal uploads) or "manual export" (manually downloading Excel files from various platforms). However, these data vary significantly in format, structure, sampling frequency, and representation methods, resulting in poor timeliness, inconsistent formats, and difficulty in tracing historical states, making direct integration and comparison challenging. For example, some regions may use new technologies such as smart cards, GPS, or mobile phones to generate continuous data streams, leading to significant differences in data quality, including missing values, outliers, and noisy data. These issues reduce data quality and affect the accuracy of subsequent analysis. Furthermore, the timestamps of time-series data are inconsistent due to time zones, recording precision, and data latency, affecting the time alignment of cross-regional data. Summary of the Invention

[0004] The purpose of this invention is to provide a method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series, so as to solve the problems mentioned in the background art.

[0005] The specific technical solution provided by this invention is as follows: A method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series, comprising the following operational steps: Step S1: Capture multi-source questionnaire data from authorized different questionnaire survey terminals using smart adapters and incremental logs.

[0006] Preferably, capturing multi-source questionnaire data from authorized different questionnaire survey terminals includes four retrieval strategies: Fetching data is achieved through a dual-link approach of API proactive polling and Webhook. The MQTT protocol is used for data retrieval in fragments. Perform time-series database protocol conversion and downsampling retrieval; Retrieve data based on incremental logs.

[0007] Step S2: Use a lightweight temporal ontology model to standardize the semantics in the multi-source questionnaire data based on temporal ontology-driven and drift decay compensation.

[0008] Preferably, the lightweight temporal ontology model includes: defining static attributes and temporal dynamic attributes for questionnaire items; using a pre-trained domain language model to perform entity linking and sentiment sequence extraction for unstructured text data; and performing unit conversion and dimension unification for sensor numerical streams based on the ontology.

[0009] Preferably, the operation of the lightweight temporal ontology model includes: The system reads raw event stream data from the message pipeline using a sequential consumption mode, and automatically triggers the corresponding payload unpacker based on the terminal type identifier carried in the unique identifier of the data source to obtain the raw payload object carrying metadata. Load a lightweight temporal ontology model containing entity, attribute, and relation layers to perform mapping and semantic annotation on structured, semi-structured, and unstructured data; Semantic quantization is performed on the identified unstructured text segments to obtain an initial sequence of semantic events with original semantic strength and semantic vectors; Perform time-series effectiveness decay compensation based on drift amount; All types of data are uniformly encapsulated into standardized multidimensional semantic tensors for output.

[0010] Step S3: Establish a soft alignment mechanism, calculate the probabilistic optimal correspondence between two sequences on a continuous time axis, and perform adaptive cross-regional spatiotemporal alignment.

[0011] Preferably, the soft alignment mechanism includes: The output single-region semantic temporal stream is decomposed into multi-scale representations at different temporal granularities, and similarity is calculated on high-frequency details and low-frequency trends respectively. By introducing variational inference, the alignment path is modeled as a continuous probability distribution, thus achieving soft alignment; Define a unified objective function to optimize both the alignment inference network parameters and the sequence reconstruction parameters involved. Based on the learned soft alignment distribution, the aligned sequence is projected onto a unified virtual time grid, eliminating the sampling frequency differences and phase misalignments of the original sequence on the time axis.

[0012] Step S4: Construct a spatiotemporal semantic graph of hyperbolic space embedding and perform cross-cultural dimension alignment.

[0013] Preferably, cross-cultural dimension alignment includes: inputting item texts and answer distributions from different regions; using a joint encoder to encode the item texts into semantic vectors via cross-lingual BERT, and simultaneously encoding the answer distributions into distribution vectors; inputting the semantic vectors and distribution vectors into a fusion layer to obtain a fusion vector, and then projecting the fusion vectors from Euclidean space to hyperbolic space to obtain hyperbolic embeddings; then using core concept nodes in the ontology library of a lightweight temporal ontology model as anchor points, calculating the hyperbolic distance between the hyperbolic embeddings of all items and each anchor point; when the minimum distance of an item is less than a set threshold, the item is mapped to the core dimension space.

[0014] Step S5: Use an adaptive item weaving engine to dynamically generate the content of subsequent survey questionnaires.

[0015] Preferably, the dynamic generation of subsequent questionnaire content includes: for each core dimension of each region, running an online change point detection algorithm to calculate the posterior probability of the data at the current time point relative to the historical predicted distribution; when the calculated posterior probability is greater than a set threshold, triggering a questionnaire dynamic generation event; after triggering the questionnaire dynamic generation event, calling an adaptive item weaving engine to generate a new measurement kernel vector; combining the measurement kernel vector with the predefined cultural shell encoding of the target region to dynamically generate the final questionnaire item text; this item is added to the next round of questionnaires, while recording its predicted discrimination and difficulty.

[0016] Step S6: Use topological sorting to identify the sets that can be executed in parallel and the sets that can be executed sequentially, and use a genetic algorithm to schedule heterogeneous data resources.

[0017] Preferably, the use of genetic algorithms for heterogeneous data resource scheduling includes: modeling the task scheduling problem as a multi-objective optimization problem; using chromosome encoding, where each chromosome represents a scheduling scheme, and performing a set simulation process on each chromosome; and outputting the scheduling scheme corresponding to the current optimal chromosome after performing genetic operations.

[0018] Step S7: Perform multi-regional time series prediction feedback based on an interpretable spatiotemporal prediction model.

[0019] Preferably, the interpretable spatiotemporal prediction model includes: combining a spatiotemporal graph convolutional network with a neural Hawkes process for multi-regional temporal prediction.

[0020] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention enables the proactive retrieval of required data from multiple isolated terminals and performs drift quantization, improving the utilization rate of data from different dimensions across different terminals. Simultaneously, this invention introduces variational inference into the field of questionnaire spatiotemporal alignment, modeling the alignment path as a probability distribution. This makes the alignment process no longer a "geometric deformation" of the original sequence, but rather a "probabilistic information fusion," solving the hard alignment distortion problem caused by differences in sampling frequencies. It also achieves cross-cultural semantic alignment. Finally, this invention differs from existing dynamic questionnaire systems that simply jump based on preset rules. The item weaving engine employed in this invention enables the questionnaire's logical structure to autonomously decide on the addition, deletion, and order adjustment of questions under the goal of maximizing information entropy. This allows for the dynamic generation of subsequent questionnaire content based on preliminary survey results and inter-regional influence relationships. This significantly improves the accuracy of the final prediction and evaluation of questionnaire content data and the accuracy of feedback. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the steps of the multi-regional questionnaire data fusion and dynamic questionnaire generation method provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the operational logic architecture provided in an embodiment of the present invention. Detailed Implementation

[0022] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0023] Example 1: In this embodiment, the present invention mainly focuses on the questionnaire and corresponding questionnaire system designed for children. While traditional questionnaires for investigating children can reflect the relevant survey data, existing questionnaires for children cover a wide range of topics and require complex time, space, and geographical considerations. This results in survey data from multiple schools, districts, cities, provinces, and other scales. These data are independent of each terminal's respective questionnaire system, and their data formats, structures, sampling frequencies, and representation methods differ significantly, making direct integration and comparison impossible. This hinders the fusion and interconnection of different questionnaire systems. Consequently, effective evaluation cannot be obtained when further assessing and providing feedback on the children's situation based on long-term time series data.

[0024] In this embodiment, combined with Figures 1-2As shown in the figure, the method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series described in this embodiment includes the following steps: Step S1: Capture multi-source questionnaire data from authorized different questionnaire survey terminals using smart adapters and incremental logs.

[0025] In this embodiment, the present invention establishes a heterogeneous data access gateway that actively pulls data, adapts to the protocol, and supports incremental change capture. This gateway aggregates questionnaire data from different terminals into a unified time-series message pipeline in the form of raw, immutable event streams. Each data packet pulled from a terminal is forcibly encapsulated into a specified structure by the gateway, ultimately outputting an ordered queue of raw event streams carrying a unified pull envelope. This solves the physical problems of isolated multi-terminal data sources and uncontrollable latency.

[0026] For example, the present invention constructs the following four pull strategies for different types of terminals: Strategy 1: Online questionnaires and web-based platforms For this terminal type, this invention employs a dual-link approach of API active polling and Webhook for data retrieval. Specifically, this includes targeting third-party online questionnaire platforms (such as Wenjuanxing or Qualtrics API) or self-built front-end event tracking. Utilizing the platform's open API interface, the polling frequency is dynamically adjusted based on a custom field in the returned HTTP header. This custom field informs the client how many legitimate requests can be initiated within the current rate-limiting time window. In this invention, the polling interval is automatically shortened during periods of frequent data updates (such as large sample collection periods) and lengthened during periods of data calm to conserve resources. Simultaneously, for terminals supporting real-time callbacks, an HTTPS tunnel is established to receive instant push notifications. This link serves as a "fast data channel" for polling, ensuring zero-latency data entry of key answers. Each retrieval is recorded with a timestamp watermark to ensure accurate recovery from the breakpoint in case of network interruptions, preventing duplicate retrievals or data loss.

[0027] Strategy Two: Face-to-face and Telephone Interview Terminals For this terminal type, this invention performs offline audio streaming upload and ASR edge preprocessing. Since face-to-face interview terminals are often in weak network environments, HTTP large file uploads are not used; instead, MQTT fragmented file retrieval is employed. The cloud gateway, acting as an MQTTBroker, actively retrieves file fragments that have not been successfully synchronized on the terminal side. A lightweight feature extraction probe is deployed on the terminal device. During retrieval, the gateway requests not only the audio stream but also the short-time energy spectrum calculated by the probe and the interview timestamp mapping table. This invention transmits only the extracted feature vector packets, rather than the complete high-definition audio, significantly reducing the backhaul bandwidth pressure while protecting the interviewee's voice privacy.

[0028] Strategy 3: Sensor and Smart Card Data Stream For this terminal type, this invention performs time-series database protocol conversion and downsampling retrieval. It targets high-throughput time-series data such as GPS, smart card swipe records, and mobile phone accelerometer data. A pluggable adapter layer is used to convert the output of different sensor databases (such as InfluxDB, TimescaleDB, and even private binary logs) into a standard exposure format. Furthermore, this invention does not directly retrieve millisecond-level raw noise data. The gateway guides the retrieval granularity based on the set data fusion requirements. For example, if the analysis finds that the data quality in a certain region is high, the gateway increases the sampling rate to retrieve details; if it finds that the data is entirely noise, the gateway only retrieves the mean and variance statistics.

[0029] Strategy 4: Legacy databases and offline forms In this terminal type, the present invention captures incremental logs based on CDC. For past survey data stored in old SQL Server, Access, or Excel files, the present invention monitors a specified shared folder. When a new .xlsx file is added, streaming SAX parsing is triggered, avoiding memory overflow when loading large files. For relational databases, the present invention does not perform a full table scan, but instead enables database log mining and pulls the corresponding log stream. A logical timestamp is injected during log pulling to mark the drift of the record between the "business occurrence time" and the "system pull time".

[0030] This invention addresses the problems of isolated data sources, uncontrollable data collection delays, chaotic transmission formats, and difficulty in tracing historical states in existing questionnaire survey systems under multi-terminal environments. By establishing an active pull gateway and pull envelope mechanism, it for the first time injects the transmission delay of multi-source data as an explicit parameter into the data stream and assigns a unified spatiotemporal reference identity to each piece of data. This differs from the simple API integration or manual upload / export methods in existing technologies. It achieves highly robust fragmented transmission in weak network environments, edge feature compression and backhaul under bandwidth-constrained conditions, and non-intrusive incremental log capture at the database level. It constructs an ordered, complete, and tamper-proof cross-terminal original event tracing stream, providing a solid data foundation with quantifiable physical delays for subsequent semantic parsing and temporal fusion steps.

[0031] Step S2: Use a lightweight temporal ontology model to standardize the semantics in the multi-source questionnaire data based on temporal ontology-driven and drift decay compensation.

[0032] In this embodiment, the present invention employs a lightweight temporal ontology model for the social survey domain. This ontology model defines not only the static attributes of questionnaire items but also their temporal dynamic attributes. The original event stream in the output event stream queue is standardized, that is, the heterogeneity of data from different terminals in terms of format, field definition, and semantic expression is standardized. For unstructured text data (such as open-ended question answers and interview transcripts), the present invention utilizes a pre-trained domain language model for entity linking and sentiment sequence extraction. For sensor numerical streams, the present invention performs unit conversion and dimensional unification based on the ontology model.

[0033] For example, running a lightweight temporal ontology model includes: Step S21: Read the raw event stream data from the message pipeline using the sequential consumption mode. Automatically trigger the corresponding payload unpacker based on the terminal type identifier carried in the unique identifier of the data source. Calculate the initial source confidence based on the terminal type and network conditions (MQTT retransmission count, HTTP status code history). For example: online questionnaires uploaded via wired fiber optic networks. The characteristics of the face-to-face interview recording after three retransmissions under weak network conditions It outputs the raw payload object carrying metadata.

[0034] Step S22: Load the lightweight time-series ontology model, which includes entity, attribute, and relation layers. For structured data (online questionnaire JSON, sensor values), directly associate field names with ontology attributes using a pre-configured mapping rule engine. For semi-structured and unstructured data (open-ended question text, transcribed interview recordings), first perform coarse-grained segmentation (e.g., segmenting question-and-answer rounds by silent segments in the recording), then proceed to sub-stage three for fine-grained extraction, converting all local event times to UTC timestamps and unifying the precision to milliseconds. For timestamps missing millisecond portions in sensor streaming data, use linear interpolation to fill in the missing milliseconds, outputting semi-structured time-series document objects with ontology semantic tags.

[0035] Step S23: Perform refined semantic quantization on the identified unstructured text paragraphs. Specifically, based on the current survey topic and ontology attributes, this invention retrieves and assembles time-aware prompt templates from a template library, and performs restricted decoding on mask vocabulary constraints. For example, the attitude mask only allows the output of tokens from the vocabulary of "positive," "neutral," and "negative," while the intensity mask only allows the output of "weak," "moderate," and "strong," etc. After calculating the Logits distribution of the output tokens, continuous intensity values ​​are obtained through Softmax mapping. Simultaneously, when a long text contains multiple question-and-answer rounds, a sliding window segmentation is used. The window length is adaptively determined by the interview speaking speed. Each window generates a semantic snapshot vector, forming the individual semantic temporal stream of the interviewee, resulting in an initial semantic event sequence with original semantic intensity and semantic vectors.

[0036] Step S24: Read the clock drift / transmission and processing delay from the envelope. Set the half-life parameter The half-life parameter in this invention Based on the pre-defined questionnaire survey type, such as a half-life of 24 hours for a follow-up survey and 72 hours for an attitude stability survey, the decay factor is calculated. ,in The original intensity of each window is corrected using an attenuation factor, while the initial confidence level of the information source is also adjusted. Also affected by drift, the overall confidence level of this semantic event ultimately... for: ,in For smoothing coefficients, This represents the variance of sentiment fluctuation within the text. Greater fluctuations indicate unstable respondents' attitudes and a slightly lower confidence level. It is a slice representation of the original semantic intensity sequence, and outputs a valid semantic event sequence after drift decay calibration.

[0037] Step S25: The various types of data processed above are uniformly encapsulated into a standardized multidimensional semantic tensor. That is, for each region and each unified time frame, the semantic snapshots of all respondents and sensors within the window are aggregated to construct a feature vector including sentiment statistics, semantic mean, sensor aggregated features and average confidence. The metadata quintuple is finally encapsulated and output, and the processed data stream is written to the time series feature storage engine in real time. A semantic integrity tag is attached to the end of the stream for each region.

[0038] This invention utilizes a drift-aware decay function to achieve continuous quantification of data timeliness, rather than discrete truncation and discarding. Simultaneously, by explicitly modeling terminal-specific transmission delays as correction factors for semantic weights, it achieves lossless representation transformation of unstructured interviews, sensor streams, and online questionnaires within a unified temporal semantic space. The output is a standardized semantic tensor stream carrying high-precision timestamps and confidence labels, eliminating the source-level timestamp ambiguity and semantic uncertainty for subsequent cross-regional spatiotemporal alignment.

[0039] Step S3: Establish a soft alignment mechanism, calculate the probabilistic optimal correspondence between two sequences on a continuous time axis, and perform adaptive cross-regional spatiotemporal alignment.

[0040] In this embodiment, although the timestamps of the standardized data stream are normalized, the sampling frequencies of different regions and terminals still vary significantly (for example, an online questionnaire in location A is submitted once a day, while a mobile phone sensor in location B records the location once per minute). Direct resampling would lead to information loss. Therefore, this invention establishes a soft alignment mechanism, which does not force rigid consistency of timestamps.

[0041] For example, the soft alignment mechanism is specifically as follows: The output single-region semantic temporal stream is decomposed into multi-scale representations at different temporal granularities, and similarity is calculated on high-frequency details and low-frequency trends respectively.

[0042] Specifically, this invention extracts the feature matrices from the outputs of the two regions A and B to be aligned, and uses the Daubechies-4 (db4) wavelet basis function to independently perform multi-level wavelet decomposition on each feature dimension. Each sequence is decomposed into a set of approximation coefficients and detail coefficients, and then the coefficients of each level are reconstructed back to the original time length (the inverse wavelet transform only retains specific frequency bands). The resulting sequence has a decomposition level determined by the sequence length plus a frequency band filter. Finally, each region obtains a multi-scale representation set.

[0043] By introducing variational inference, the alignment path is modeled as a continuous probability distribution, thus achieving soft alignment; Specifically, traditional Dynamic Time Warp (DTW) outputs a deterministic discrete alignment path, which cannot quantify the uncertainty of alignment and is not differentiable. This invention defines the alignment path as a set of mappings from the time index of sequence A to the time index of sequence B. Using the mean field assumption, the alignment path is decomposed into an independent product of the alignment relationships at each time step. Simultaneously, a smoothness prior is introduced to encourage alignment paths to be close to the diagonal, avoiding ill-conditioned one-to-many warps. Furthermore, in this invention, the expected value of any loss function dependent on the alignment path can be calculated using Monte Carlo sampling or reparameterization techniques.

[0044] Define a unified objective function to optimize the alignment inference network parameters and subsequent sequence reconstruction parameters. Specifically, the weighted Euclidean distance expectation under soft alignment is calculated on each wavelet decomposition frequency band. The total alignment loss is the weighted sum of the losses of each frequency band, resulting in a multi-scale alignment loss. Then, a KL divergence regularization term is introduced. Based on the multi-scale alignment loss and the KL divergence regularization term, the final total objective function is obtained. The alignment inference network parameters are iteratively updated using the gradient descent method. After convergence, the optimal variational distribution is obtained, which is the optimal soft alignment relationship between the two sequences in a probabilistic sense.

[0045] Based on the learned soft alignment distribution, the aligned sequence is projected onto a unified virtual time grid to eliminate the sampling frequency differences and phase misalignments of the original sequence on the time axis; Specifically, based on the temporal resolution requirements of the questionnaire survey, this invention defines an equally spaced reference time axis. For each virtual moment, its corresponding position on the original time axis of sequence A is determined by the marginal probability of the alignment distribution. A projection weight matrix from sequence A to the virtual grid is defined, and similarly, the projection weight matrix of sequence B is calculated. The original features are then weighted and averaged using these projection weights to obtain the aligned sequence on the virtual grid. Finally, a unified spatiotemporal latent variable tensor incorporating dual-regional information is output. Simultaneously with outputting the alignment results, this invention uses a variational distribution to calculate the variance of the projection weights of each virtual grid point as a measure of alignment uncertainty, quantifying the uncertainty caused by soft alignment and providing a confidence reference for subsequent calculations.

[0046] Step S4: Construct a spatiotemporal semantic graph of hyperbolic space embedding and perform cross-cultural dimension alignment.

[0047] In this embodiment, the problem that data fields (such as questions and answers) in standardized event streams are still "isolated" strings or numbers is addressed. This is achieved by constructing a dynamically growing hyperbolic space graph that maps these fragmented semantics into a unified, comparable spatiotemporal semantic space.

[0048] For example, the operation of the hyperbolic space embedding spatiotemporal semantic graph includes: The present invention takes as input item text and answer distribution from region A, and uses a joint encoder to encode the item text into semantic vectors via cross-lingual BERT (such as XLM-RoBERTa), while simultaneously encoding the answer distribution (such as Likert scale frequency) into distribution vectors. Then, a trainable fusion layer is used to obtain a fusion vector, which is then projected from Euclidean space to hyperbolic space (Poincaré sphere model) to obtain the hyperbolic embedding.

[0049] Then, using the core concept nodes in the ontology as anchor points, the hyperbolic embeddings of all items are calculated to determine the hyperbolic distance between each anchor point. If the minimum distance is less than a threshold, the item is mapped to the core dimension space. This achieves the mapping of items from different languages ​​and expressions to a unified construct space. In addition to semantic edges, this invention also constructs temporal and geographic edges.

[0050] Temporal edge: If the semantic graphs of the same region are strongly correlated at time t and time t+1 (e.g., the two surveys target the same construct), then an edge with temporal weights is established between the corresponding nodes of the two time layers.

[0051] Geographic edge: Based on the aligned time series, calculate the temporal similarity of the same dimension in different regions. If the temporal similarity is higher than a set threshold, then establish a geographic association edge for the corresponding nodes of regions A and B in the graph.

[0052] This invention addresses the semantic problem in existing technologies where questionnaire data from different regions and cultural backgrounds are difficult to directly compare. Unlike simple translation alignment or generic word vectors, this invention introduces hyperbolic space embedding to represent hierarchical cultural dimensions and designs a joint encoder to fuse the question text and response distribution. This ensures that each node (item, dimension, region) in the spatiotemporal semantic graph has unique hyperbolic space coordinates, achieving precise anchoring and comparability of cross-cultural constructs.

[0053] Step S5: Based on the preliminary survey results (such as abnormal scores in a certain dimension in a certain region) and the influence relationships between regions, dynamically generate the content of subsequent survey questionnaires.

[0054] In this embodiment, the present invention runs an online change point detection algorithm (such as Bayesian Online Change Point Detection, BOCPD) for each core dimension of each region. It calculates the posterior probability of the data at the current time point relative to the historical predicted distribution. When the calculated posterior probability exceeds a set threshold, a questionnaire dynamic generation event is triggered.

[0055] For example, when the questionnaire dynamic generation event is triggered, this invention calls the "adaptive item weaving engine" and, based on the TRIZ local quality principle, aggregates three types of neighbor information for the nodes corresponding to the triggering event in the spatiotemporal semantic graph at specific time layers, regions, and dimensions: Historical self-attention neighbors: nodes in the same region and dimension that have been in the past L time steps; Cross-regional contemporaneous neighbors: nodes of the same dimension in other regions at the current time step t; Cross-dimensional neighbor relationships: Nodes in other dimensions within the current region and at the current time step; By using a perturbation-enhanced graph attention network, attention weights are calculated for each neighbor type using graph attention layers. After weighted aggregation, the results are normalized using softmax to output the final aggregated neighbor representation.

[0056] In particular, in response to the fact that traditional graph neural networks only output deterministic node representations, this invention introduces a learnable perturbation vector to make the output deviate from the aggregate representation within a reasonable range. That is, the information entropy of the current node is first calculated as a measure of uncertainty. The higher the information entropy, the more uncertain the cognition of that dimension is, and the more exploratory new items are needed. The higher the information entropy (the greater the uncertainty), the greater the perturbation variance. This invention generates the final new measurement kernel vector by sampling the perturbation vector from the multivariate normal distribution. Since the generated kernel vector is an abstract mental construct representation, it needs to be converted into readable item text for the target region.

[0057] This invention further utilizes a Conditional Variational Autoencoder (CVAE) as a dynamically equivalent translation engine. The Conditional Variational Autoencoder consists of an encoder, a decoder, and a latent space. The conditional variables are: Measure kernel vector; The predefined cultural shell encoding of the target region is extracted from the semantic graph and includes information such as the region's scores in multiple dimensions, common metaphors, and sentence structure preferences. Then, a reconstruction loss function, an equivalence loss function, and KL divergence are constructed. The reconstruction loss function ensures that the generated text is syntactically and semantically sound. The equivalence loss function requires that the reconstructed kernel be as close as possible to the original kernel. KL divergence is used to make the latent distribution approximate a standard normal distribution, thus facilitating the generation of diverse text. Reconstruction Loss Function Equivalent loss function and KL divergence They are represented as follows:

[0058]

[0059]

[0060] in, For expectation operator, The variational posterior probability distribution of the encoder output. These are latent features extracted from the input by the encoder, used to generate diverse question texts. The generation probability distribution of the decoder output. The newly generated question text (sequence) is the output target of the decoder. The measurement kernel vector is generated by the perturbation-enhanced graph attention network. Encode the cultural shell of a predefined target region. The reconstruction loss is equal to the negative expectation of the logarithm of the generation probability under the encoder distribution. Minimizing this loss forces the decoder to accurately reconstruct the item text from the latent variable z and the condition variable. The reconstructed kernel vector is obtained by mapping the generated title text back to the kernel space through a semantic anchoring function. For the KL divergence operator, It follows a normal distribution (Gaussian distribution). To represent the covariance matrix of the normal distribution, the KL divergence loss is equal to the KL divergence between the posterior distribution of the encoder output and the standard normal distribution. Minimizing this loss forces the distribution of the latent variable z to approximate the standard normal distribution, thus making the latent space continuous and regular, facilitating sampling and the generation of diverse outputs. Finally, the total loss is obtained. , , Let be the regularization coefficient.

[0061] Therefore, an encoder is not needed when actually generating new items. Latent variables are sampled directly. The data is then input into the decoder to generate the final item text, which is added to the next round of the questionnaire. The predicted discrimination and difficulty are recorded. Simultaneously, different z samples can be sampled multiple times to obtain multiple candidate items for manual review or automatic selection.

[0062] Step S6: Use topological sorting to identify the sets that can be executed in parallel and the sets that can be executed sequentially, and use a genetic algorithm to schedule heterogeneous data resources.

[0063] In this embodiment, the present invention constructs a directed acyclic graph where nodes are computational tasks and edges are data dependencies, for example: Task S2_Clean (Region A) depends on the output of Task S1_Pull (Region A).

[0064] Task S3_Map Update (Time Window T) depends on the completion of S2_Cleanup in all regions.

[0065] Task S4_Kernel Generation (Triggering Event) depends on the completion of S3_Graph Update.

[0066] These dependencies are constructed as a DAG. ,in For task nodes, For directed edges express Must Previously completed. This invention uses topological sorting to hierarchically divide the DAG, identifying tasks with no dependencies (i.e., no paths connecting them), which are marked as tasks that can be executed in parallel, while tasks with paths (direct or indirect dependencies) between any two nodes are marked as tasks that must be executed sequentially.

[0067] This invention is based on genetic algorithm-driven data resource scheduling optimization. It models the task scheduling problem as a multi-objective optimization problem, with the objectives being to minimize the overall completion time and maximize resource utilization. In other words, the multi-objective problem is transformed into a single-objective fitness function, and the fitness value of the scheduling scheme S is calculated. :

[0068] in, , , , These are the weight coefficients of the term "Ge". , , Each is an optimization objective. It is a very small positive number, used to prevent the denominator from being zero. To avoid penalties for violating dependencies or exceeding resource limits, this invention omits the normalization factor in the above formula for simplicity. In practical applications, each term will be normalized (divided by its respective theoretical maximum value). This fitness value... The larger the value, the better the scheduling scheme.

[0069] Then, chromosome encoding is used, with each chromosome representing a scheduling scheme, employing indirect encoding: the length is... an integer array, the first Bit representation of task The assigned resource node number The execution order is separately encoded by a dependency-based priority list, or the start time of each task is simulated using a list scheduler based on dependency constraints and resource allocation when evaluating fitness. For each chromosome, the following simulation process is performed: a. Initialize the availability schedule (idle timeline) of all resource nodes; b. Based on the priority order of chromosome B, retrieve the tasks sequentially. ; c. Check if all its predecessor tasks have been completed. If not, skip the waiting process. d. Locate the resource node Find the first sufficiently long free window (that satisfies the task) on the available timeline of that node. The estimated execution time, which is given by the historical performance model. e. Allocate this window, update the available timeline for the node, and record the start and end times of the task; i. Repeat until all task scheduling is complete, then calculate... , , Then calculate fitness. The algorithm predicts the execution time model, with parameters updated online via historical execution logs. After genetic operations, the algorithm iterates through multiple generations to continuously evolve a near-optimal scheduling scheme. The algorithm converges when the optimal fitness does not improve for several consecutive generations or when the maximum number of iterations is reached. The algorithm outputs the scheduling scheme corresponding to the current optimal chromosome.

[0070] Finally, based on the optimal scheduling scheme output by the genetic algorithm, each computing node is dynamically configured. For example, high-load graph attention inference tasks are scheduled to nodes with sufficient GPU memory, while I / O-intensive S1 fetch tasks are scheduled to nodes with high network bandwidth. The system monitors the load of each resource in real time, and if a significant deviation occurs, the genetic algorithm is warm-started for rescheduling to achieve dynamic rebalancing.

[0071] Step S7: Perform multi-regional time series prediction feedback based on an interpretable spatiotemporal prediction model.

[0072] In this embodiment, the spatiotemporal prediction model constructed by this invention uses a spatiotemporal semantic graph as its input structure. The node feature matrix contains the scores of each dimension over time, and the adjacency matrix contains geographical and temporal similarity. A spatiotemporal graph convolutional network is used to capture complex spatiotemporal dependencies. Furthermore, a neural Hawkes process is integrated to model the intensity of events (such as "significant changes in a certain dimension's score"), which depends on historical events. A neural network replaces the preset exponential decay kernel, enabling it to learn complex, nonlinear cross-regional and cross-dimensional triggering effects. Counterfactual prediction is then performed, and a backdoor criterion adjustment is executed. The value of the intervention variable is fixed as the new value. Using the trained neural Hawkes process model, the conditional probability distribution of all other variables under the intervention is recalculated to obtain the counterfactual time series. Based on the model, multi-regional and multi-dimensional predicted values ​​for multiple future time steps are output.

[0073] For example, this invention addresses the fundamental problem of existing technologies—namely, the low prediction accuracy and inability to reveal cross-regional dynamic mechanisms—based on this step. Unlike traditional time-series models (such as ARIMA and LSTM), this invention combines spatiotemporal graph convolutional networks with neural Hawkes processes, significantly improving the accuracy of multi-regional time-series prediction by utilizing graph structure information.

[0074] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0075] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series, characterized by: The following steps are included: Step S1: Capture multi-source questionnaire data from authorized questionnaire terminals using smart adapters and incremental logs; Step S2: Use a lightweight temporal ontology model to standardize the semantics in the multi-source questionnaire data based on temporal ontology-driven and drift decay compensation. Step S3: Establish a soft alignment mechanism, calculate the probabilistic optimal correspondence between two sequences on a continuous time axis, and perform adaptive cross-regional spatiotemporal alignment; Step S4: Construct a spatiotemporal semantic graph of hyperbolic space embedding and perform cross-cultural dimension alignment; Step S5: Use an adaptive item weaving engine to dynamically generate the content of subsequent survey questionnaires; Step S6: Use topological sorting to identify the sets that can be executed in parallel and the sets that can be executed sequentially, and use a genetic algorithm to schedule heterogeneous data resources; Step S7: Based on an interpretable spatiotemporal prediction model, perform multi-regional and multi-dimensional prediction feedback.

2. The method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series as described in claim 1, characterized in that: Step S1 involves capturing multi-source questionnaire data from different authorized questionnaire survey terminals, employing four retrieval strategies: Fetching data is achieved through a dual-link approach of API proactive polling and Webhook. The MQTT protocol is used for data retrieval in fragments. Perform time-series database protocol conversion and downsampling retrieval; Retrieve data based on incremental logs.

3. The method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series as described in claim 2, characterized in that: The lightweight temporal ontology model in step S2 includes: Define the static and time-series dynamic attributes of questionnaire items; For unstructured text data, a pre-trained domain language model is used to extract entity links and sentiment sequences; For sensor numerical streams, unit conversion and dimensional unification are performed based on the data itself.

4. The method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series as described in claim 3, characterized in that: The operation of a lightweight temporal ontology model includes: The system reads raw event stream data from the message pipeline using a sequential consumption mode, and automatically triggers the corresponding payload unpacker based on the terminal type identifier carried in the unique identifier of the data source to obtain the raw payload object carrying metadata. Load a lightweight temporal ontology model containing entity, attribute, and relation layers to perform mapping and semantic annotation on structured, semi-structured, and unstructured data; Semantic quantization is performed on the identified unstructured text segments to obtain an initial sequence of semantic events with original semantic strength and semantic vectors; Perform time-series effectiveness decay compensation based on drift amount; All types of data are uniformly encapsulated into standardized multidimensional semantic tensors for output.

5. The method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series as described in claim 4, characterized in that: The soft alignment mechanism in step S3 includes: The output single-region semantic temporal stream is decomposed into multi-scale representations at different temporal granularities, and similarity is calculated on high-frequency details and low-frequency trends respectively. By introducing variational inference, the alignment path is modeled as a continuous probability distribution, thus achieving soft alignment; Define a unified objective function to optimize both the alignment inference network parameters and the sequence reconstruction parameters involved. Based on the learned soft alignment distribution, the aligned sequence is projected onto a unified virtual time grid, eliminating the sampling frequency differences and phase misalignments of the original sequence on the time axis.

6. The method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series as described in claim 5, characterized in that: Step S4, cross-cultural dimension alignment, includes: Input the question texts and answer distributions from different regions; A joint encoder is used to encode the question text into a semantic vector using cross-language BERT, while the response distribution is encoded into a distribution vector. The semantic vector and distribution vector are input into the fusion layer to obtain the fusion vector. The fusion vector is then projected from Euclidean space to hyperbolic space to obtain the hyperbolic embedding. Then, using the core concept nodes in the ontology library of the lightweight temporal ontology model as anchor points, calculate the hyperbolic embedding of all items and the hyperbolic distance between each anchor point; When the minimum distance of a question item is less than a set threshold, the question item is mapped to the core dimension space.

7. The method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series as described in claim 6, characterized in that: The content of the subsequent survey questionnaire dynamically generated in step S5 includes: For each core dimension of each region, run an online change point detection algorithm to calculate the posterior probability of the data at the current time point relative to the historical predicted distribution. When the calculated posterior probability is greater than the set threshold, trigger a questionnaire dynamic generation event. When the questionnaire dynamic generation event is triggered, the adaptive item weaving engine is invoked to generate a new measurement kernel vector; By combining the measurement kernel vector with the predefined cultural shell encoding of the target region, the final questionnaire item text is dynamically generated; This item was added to the next round of the questionnaire, and its predicted discrimination and difficulty were recorded.

8. The method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series as described in claim 7, characterized in that: Step S5 also includes: a learnable perturbation vector and a conditional variational autoencoder.

9. The method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series as described in claim 8, characterized in that: Step S6, which uses a genetic algorithm for heterogeneous data resource scheduling, includes: The task scheduling problem is modeled as a multi-objective optimization problem; Chromosome encoding is used, with each chromosome representing a scheduling scheme, and a set simulation process is executed for each chromosome; After performing genetic operations, output the scheduling scheme corresponding to the current optimal chromosome.

10. The method for multi-regional questionnaire data fusion and dynamic questionnaire generation based on time series as described in claim 9, characterized in that: The interpretable spatiotemporal prediction model in step S7 includes: combining spatiotemporal graph convolutional networks with neural Hawkes processes to perform multi-regional and multi-dimensional prediction feedback.

Images