Workplace-oriented data processing method and system

Abstract

The invention relates to a workplace-oriented data processing method and a workplace-oriented data processing system, in particular to the field of data processing, which can realize real-time accurate assessment and dynamic intervention of workplace staff health risks, improve the accuracy of risk prediction through multi-modal data fusion, enhance the interpretability of a model in combination with a knowledge graph, and improve the accuracy of risk prediction. A personalized intervention strategy is generated by utilizing reinforcement learning, and the system is continuously self-optimized by virtue of an online learning mechanism, so that a complete closed loop of perception, decision, action and learning is finally formed, the health risk occurrence rate is effectively reduced, and the health level and the working efficiency of employees are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, and more particularly, to a workplace-oriented data processing method and system. BACKGROUND

[0002] In recent years, with the increasing emphasis of enterprises on employee health management, digital health monitoring systems have been widely introduced in workplaces such as large-scale manufacturing enterprises and science and technology parks. Under this system, employees are usually required to submit health self-assessment questionnaires regularly through mobile applications or internal online platforms. The questionnaire content involves subjective feeling descriptions, such as fatigue level, sleep quality, emotional state, and muscle soreness. At the same time, enterprises provide employees with intelligent wearable devices, such as smart wristbands or smart badges, to continuously collect objective physiological parameters, including heart rate variability, daily steps, activity duration, and environmental noise. The internal driving factor of this practice is that enterprises aim to reduce the incidence of occupational diseases, improve attendance rate and production efficiency through proactive health management. The external driving factor may be the group health risks brought about by sudden public health events, seasonal influenza outbreaks, or extreme weather conditions. However, the questionnaire text is unstructured data, and its expression method has high diversity. There are significant differences in structure and semantics between the text and the time-series numerical data generated by sensors, which poses a great challenge to the comprehensive and accurate assessment of individual and group health status.

[0003] Currently, there are obvious limitations in the analysis technology for workplace health data. In terms of text processing, existing solutions mostly rely on rule-based keyword matching or traditional machine learning classifiers, such as using regular expressions for pattern filtering or using support vector machines for simple sentiment polarity judgment of text. Such methods stay at the surface level of language understanding and are difficult to effectively capture context-dependent relationships, semantic disambiguation, and subtle sentiment tendencies in text, resulting in insufficient understanding of employee subjective descriptions. In terms of data fusion, most systems process text questionnaire data and sensor physiological data separately, lacking effective multi-modal information alignment and collaborative analysis mechanisms, and unable to construct a unified representation of employee health status as a whole. In addition, the analysis process of existing technical solutions is usually in an offline batch processing mode, with a delay of several hours or even several days from data collection, cleaning to the generation of risk assessment reports, which seriously restricts the ability to conduct real-time early warning and dynamic intervention on acute health risks (such as sudden discomfort caused by excessive fatigue). This technical deficiency makes it difficult for enterprise health management strategies to achieve the key transition from passive response to proactive prevention, and cannot meet the urgent needs of modern workplaces for efficient and intelligent health protection systems. SUMMARY

[0004] The present application aims at the technical problems existing in the prior art, and provides a workplace-oriented data processing method and system to solve the problems raised in the background art.

[0005] The technical solution of the present application to solve the above technical problems is as follows: a workplace-oriented data processing method, specifically comprising the following steps: Step S1, real-time receiving first data from an employee terminal and second data from a sensor through a stream processing framework, and buffering and aligning the first data and the second data in a sliding time window of a preset time length; the first data is health questionnaire data containing natural language text, and the second data is time series data containing physiological parameters; Step S2, performing natural language processing on the first data in the sliding time window to extract symptom entity features and emotion vector features contained therein, and performing time domain and frequency domain feature extraction on the second data in the sliding time window to obtain a physiological feature vector; then, using a dynamic semantic alignment layer, calculating cross-modal attention weights based on the symptom entity features, emotion vector features and physiological feature vector, and generating a fusion feature vector according to the weights; Step S3, inputting the fusion feature vector into a time series prediction model, and simultaneously querying a pre-constructed professional health knowledge graph to obtain a graph subgraph vector associated with the symptom entity features from the knowledge graph; in the processing process of the time series prediction model, using the hidden state generated by the time series prediction model to fuse the graph subgraph vector and the hidden state through a cross-attention mechanism to generate a knowledge-enhanced state vector, and outputting a health risk prediction probability based on this; Step S4, obtaining current work context information of the corresponding employee from an enterprise information system, and combining the health risk prediction probability and the current work context information to form an agent state; inputting the agent state into a multi-agent reinforcement learning strategy network pre-trained through reinforcement learning, and outputting a personalized intervention action for the employee from the strategy network; executing the personalized intervention action, and after execution, collecting new first data and second data through the employee terminal and the sensor as feedback data; using the feedback data to fine-tune the parameters of the time series prediction model and the multi-agent reinforcement learning strategy network in an online learning manner; In a preferred embodiment, in step S1, the specific operation of buffering and aligning the first data and the second data in a sliding time window of a preset time length is as follows: For each piece of first data received in real time, calculate its text richness index, which is the ratio of the number of effective words in the natural language text contained in the first data to the total number of text words; For each piece of second data received in real time, a signal quality index is calculated, which is the ratio of the signal power of the physiological parameter signal represented by the piece of second data to the noise power of the preset high-frequency noise band within a preset short-time calculation window.

[0006] In a preferred embodiment, after obtaining the text richness index and the signal quality index, a dynamic fusion weight factor is further calculated for each piece of first data and each piece of second data within the sliding time window, and the calculation process includes: The text richness index of the first data is directly taken as its instant quality score, and the signal quality index of the second data is divided by a preset normalization reference value to obtain its instant quality score; A context consistency score is calculated, which is the reciprocal of the Mahalanobis distance between the feature value of the current data and the mean of the feature value set of all data of the same type within the sliding time window, and the reciprocal of the distance value is taken as the score result; A timeliness decay factor is calculated, which is the result of an exponential function with a natural constant as the base, the absolute difference between the timestamp of the current data and the center time point of the sliding time window multiplied by a preset decay coefficient as the exponent, and the negative value as the result; The instant quality score and the context consistency score are multiplied by the first preset coefficient and the second preset coefficient respectively, and then added to obtain a weighted sum; The weighted sum is input into an S-shaped growth function for nonlinear mapping to obtain an intermediate weight value; The intermediate weight value is multiplied by the timeliness decay factor to obtain a dynamic fusion weight factor with a value range between zero and one; After the above calculation, the output of the sliding time window is an enhanced aligned buffer set, which explicitly includes each first data and each second data within the window, and each data is associated with its corresponding dynamic fusion weight factor.

[0007] In a preferred embodiment, in step S2, the specific operations of performing natural language processing on the first data within the sliding time window to extract symptom entity features and sentiment vector features, and performing time domain and frequency domain feature extraction on the second data to obtain physiological feature vectors are as follows: Each piece of first data and each piece of second data is read from the enhanced aligned buffer set obtained in step S1, and for each piece of first data, a named entity recognition model is used to identify symptom entities representing health status from natural language text, and each identified symptom entity is converted into a fixed-dimensional symptom entity feature; Meanwhile, a deep learning sentiment analysis model is used to process the natural language text, and a sentiment vector feature representing multi-dimensional sentiment intensity is output. For each piece of second data, first, signal preprocessing is performed, including denoising and normalization, then time domain features and frequency domain features are extracted from the preprocessed signal, the time domain features include root mean square and approximate entropy, and the frequency domain features include the power spectrum density main frequency energy obtained through spectrum analysis; Finally, the extracted time domain features and frequency domain features are spliced into a fixed-dimensional physiological feature vector. Thus, the feature extraction of the first data and the second data is completed, and the symptom entity feature and the sentiment vector feature corresponding to each piece of first data, and the physiological feature vector corresponding to each piece of second data are obtained.

[0008] In a preferred embodiment, the specific operation of calculating cross-modal attention weights and generating a fusion feature vector based on the symptom entity feature, the sentiment vector feature and the physiological feature vector using the dynamic semantic alignment layer is as follows: First, the symptom entity feature, the sentiment vector feature and the physiological feature vector are respectively subjected to weighted average pooling operation using the dynamic fusion weight factor corresponding to each first data and second data calculated in step S1, the weighted average pooling operation refers to: multiplying the symptom entity features of all first data in the sliding window by the corresponding dynamic fusion weight factor and summing them up, and then dividing by the sum of the dynamic fusion weight factors of all first data, thereby obtaining the pooled symptom entity feature; In the same way, the sentiment vector feature is pooled to obtain the pooled sentiment vector feature, and the physiological feature vectors of all second data are pooled to obtain the pooled physiological feature vector; Then, the pooled symptom entity feature, the pooled sentiment vector feature and the pooled physiological feature vector are respectively input into three different fully connected layers for linear transformation, each fully connected layer includes a trainable weight matrix and a bias vector, the linear transformation is used to map the features of different modalities to a common semantic space with the same dimension, and the hyperbolic tangent activation function is used for nonlinear activation of the transformation result, and finally the projected symptom entity feature, the projected sentiment vector feature and the projected physiological feature vector are obtained; Next, the cross-modal attention weights are calculated, a trainable query vector is used to perform dot product operation with the projected symptom entity feature, the projected sentiment vector feature and the projected physiological feature vector respectively, to obtain three initial attention scores; Subsequently, the three initial attention scores are input into the Softmax function for normalization to obtain three final attention weight values with a value range of zero to one and a sum of one, respectively corresponding to the symptom entity feature, the sentiment vector feature and the physiological feature vector. Finally, the projected symptom entity features, the projected emotion vector features, and the projected physiological feature vectors are multiplied by the respective final attention weight values, and then the three weighted feature vectors are added to generate a final fusion feature vector.

[0009] In a preferred embodiment, the operation of inputting the fusion feature vector into the time series prediction model and simultaneously querying the pre-built professional health knowledge graph to obtain a graph subgraph vector in step S3 is as follows: First, the fusion feature vector output in step S2 is input into a time series prediction model in chronological order. The time series prediction model adopts a recurrent neural network structure, specifically a gated recurrent unit network. For each time step, the time series prediction model calculates the hidden state of the current time step based on the fusion feature vector of the current time step and the hidden state of the previous time step. The hidden state of the current time step is calculated based on the fusion feature vector of the current time step and the hidden state of the previous time step through a gated recurrent unit operation. At the same time, based on the symptom entity features extracted from the first data in step S2, the pre-built professional health knowledge graph is queried. The professional health knowledge graph is stored in a graph structure, with nodes representing health-related entities, including symptoms, diseases, and risk factors, and edges representing relationships between entities. When querying, the symptom entity features are used as keys to retrieve the entities and relationships directly connected to them in the knowledge graph, forming a local subgraph. Then, a graph attention network is used to encode the local subgraph to generate a graph subgraph vector. The graph attention network calculates the attention weight between each node and its neighbor nodes in the local subgraph through a multi-head attention mechanism, and aggregates the feature information of the neighbor nodes according to the weight. After passing through a layer of graph attention layer, a fixed-dimensional graph subgraph vector is finally generated through the output layer of the network. At this point, the hidden state of each time step and the corresponding graph subgraph vector are obtained.

[0010] In a preferred embodiment, the specific operation of using the hidden state generated by the time series prediction model and fusing the graph subgraph vector with the hidden state through a cross-attention mechanism to generate a knowledge-enhanced state vector and output a health risk prediction probability is as follows: First, the knowledge-enhanced fusion is achieved using a cross-attention mechanism, where the graph subgraph vector is used as the query source, and the hidden state output by the time series prediction model is used as the key and value. Specifically, the graph subgraph vector is linearly projected through a trainable first projection matrix to map it to a space with the same dimension as the hidden state, obtaining a query vector. Meanwhile, the hidden state is linearly projected by a trainable second projection matrix to obtain a key vector, and the hidden state is linearly projected by a trainable third projection matrix to obtain a value vector; Then, a similarity score between the query vector and the key vector is calculated, and the similarity score is realized by a dot product operation between the query vector and the key vector, and is divided by a scaling factor, which is the square root of the dimension of the hidden state; Next, the similarity score is input into a Softmax function for normalization to obtain an attention weight vector; Based on the attention weight, the value vector is weighted and summed to generate a knowledge context vector; Subsequently, the knowledge context vector is linearly transformed by a trainable fourth projection matrix, and then the transformed result is added to the original hidden state through a residual connection, and then normalized by a layer to generate a knowledge-enhanced state vector; Finally, the knowledge-enhanced state vector is input into a fully connected layer with a trainable weight matrix and bias term to linearly transform the knowledge-enhanced state vector; Then, the linearly transformed result is input into a Sigmoid activation function for nonlinear mapping, and the output value is compressed between zero and one, which is the health risk prediction probability of the current time step.

[0011] In a preferred embodiment, the step S4 of obtaining the current work context information of the corresponding employee from the enterprise information system and combining the health risk prediction probability with the current work context information to form an agent state, and inputting the agent state into the multi-agent reinforcement learning policy network to output a personalized intervention action is specifically as follows: First, the current work context information of the corresponding employee is obtained from the enterprise information system in real time, and the current work context information includes the first type of information: real-time work task load intensity, the second type of information: current work role identifier, and the third type of information: real-time environmental stress index; The health risk prediction probability output by step S3 is spliced with the first type of information, the second type of information and the third type of information to form a comprehensive agent state vector; Then, the agent state vector is input into a multi-agent reinforcement learning policy network which is pre-trained by reinforcement learning; the multi-agent reinforcement learning policy network first extracts features of the agent state vector through a fully connected layer to obtain a hidden representation of each agent; Next, the attention mechanism is used to calculate the mutual influence weight between the current agent and other adjacent agents in the workplace. The specific calculation process is as follows: the hidden representation of the current agent is spliced with the hidden representation of each adjacent agent, the spliced vector is linearly transformed through a trainable weight matrix, and then is nonlinearly mapped through an activation function to obtain the initial influence score of each adjacent agent on the current agent; then all the initial influence scores are input into the Softmax function for normalization processing, so that the sum of the influence weights of the current agent and all adjacent agents is one, and the final influence weight is obtained. According to the calculated influence weight, the hidden representations of the adjacent agents are weighted and summed to obtain the context-aware hidden representation of the current agent. Finally, the context-aware hidden representation of the current agent is passed through a fully connected layer, and the Softmax activation function is used to output the probability distribution of each possible intervention action. According to the probability distribution, the intervention action with the highest probability is selected as the personalized intervention action for the employee.

[0012] In a preferred embodiment, the specific operation of collecting feedback data after performing the personalized intervention action and using the feedback data to fine-tune the parameters of the time series prediction model and the multi-agent reinforcement learning strategy network in an online learning manner is as follows: First, the personalized intervention action output by the multi-agent reinforcement learning strategy network is executed, and a predetermined time interval is waited after execution. New first data is collected through the employee terminal and new second data is collected through the sensor, and the new first data and second data are used as feedback data. Then, using the feedback data, the trainable parameters of the time series prediction model used in step S3 and the trainable parameters of the multi-agent reinforcement learning strategy network used in step S4 are sequentially subjected to online incremental learning. For the time series prediction model, the new first data and second data in the feedback data are used to calculate the new health risk prediction probability according to the method of steps S1 to S3, and the loss function is calculated with the actual health state label, and the preliminary update parameters of the time series prediction model are generated by the gradient descent method. For the multi-agent reinforcement learning strategy network, the policy gradient method based on importance sampling is used for updating to generate the preliminary update parameters of the strategy network. Finally, a momentum update mechanism is used to soft-update the time series prediction model preliminary update parameters and the strategy network preliminary update parameters generated above, which specifically includes: obtaining the old parameters of the time series prediction model and the old parameters of the strategy network before online incremental learning, and performing weighted average calculation using the preliminary update parameters of the time series prediction model and the old parameters of the time series prediction model, and the weight is controlled by a preset update coefficient. Likewise, the preliminary updated parameters of the policy network and the old parameters of the policy network are used to calculate a weighted average, and the weight is controlled by the same update coefficient, and the parameters obtained after the weighted average are used as the final updated parameters of the time series prediction model and the policy network respectively.

[0013] The application also provides a workplace-oriented data processing system, specifically comprising: a data receiving and alignment module, a feature extraction and fusion module, a risk prediction module, and an intervention and learning module, wherein The data receiving and alignment module is configured to receive health questionnaire data from an employee terminal and physiological parameter time series data from a sensor in real time through a stream processing framework, and buffer and align the data in a preset sliding time window. The feature extraction and fusion module is configured to extract symptom entity features and emotion vector features from the questionnaire data in the window, extract physiological feature vectors from the sensor data, and calculate cross-modal attention weights using a dynamic semantic alignment layer to generate a fusion feature vector. The risk prediction module is configured to input the fusion feature vector into a time series prediction model, combine a subgraph vector obtained from a professional health knowledge graph, generate a knowledge-enhanced state vector through a cross-attention mechanism, and then output a health risk prediction probability. The intervention and learning module is configured to construct an agent state according to the health risk prediction probability and work context information obtained from an enterprise information system, output personalized intervention actions through a multi-agent reinforcement learning policy network, and perform online fine-tuning on the time series prediction model and the policy network parameters based on feedback data collected after the execution of the actions.

[0014] The application has the beneficial effects of: enabling real-time and accurate assessment and dynamic intervention of the health risks of employees in the workplace, improving the accuracy of risk prediction through multi-modal data fusion, enhancing the interpretability of the knowledge graph-enhanced model, generating personalized intervention strategies using reinforcement learning, and continuously optimizing the system itself using an online learning mechanism, ultimately forming a complete closed loop of perception, decision-making, action, and learning, effectively reducing the incidence of health risks, and improving the health level and work efficiency of employees. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The application provides a method flowchart. Figure 2 The application provides a system structure block diagram. DETAILED DESCRIPTION

[0016] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of the present application.

[0017] In the description of the present application, the terms "first", "second" are used only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0018] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or explanation". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can realize the present application without using these specific details. In other examples, well-known structures and processes will not be described in detail to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the shown embodiments, but is consistent with the broadest scope in accordance with the principles and characteristics disclosed in the present application.

[0019] Embodiment 1 The present embodiment provides a workplace-oriented data processing method as shown in Figure 1 The present embodiment provides a workplace-oriented data processing method as shown in Step S1, receiving first data from an employee terminal and second data from a sensor in real time through a stream processing framework, and buffering and aligning the first data and the second data in a sliding time window of a preset time length; the first data is health questionnaire data containing natural language text, and the second data is time series data containing physiological parameters; Step S2, performing natural language processing on the first data in the sliding time window to extract symptom entity features and sentiment vector features contained therein, and performing time domain and frequency domain feature extraction on the second data in the sliding time window to obtain a physiological feature vector; then, using a dynamic semantic alignment layer, calculating cross-modal attention weights based on the symptom entity features, the sentiment vector features and the physiological feature vector, and generating a fusion feature vector according to the weights; Step S3, input the fusion feature vector into the time series prediction model, and query the pre-constructed professional health knowledge graph at the same time, to obtain a graph subgraph vector associated with the symptom entity feature from the knowledge graph; in the processing process of the time series prediction model, the hidden state generated by the time series prediction model is used to fuse the graph subgraph vector and the hidden state through a cross-attention mechanism to generate a knowledge-enhanced state vector, and a health risk prediction probability is output based on the knowledge-enhanced state vector; Step S4, obtain the current work context information of the employee from the enterprise information system, and combine the health risk prediction probability and the current work context information to form an agent state; input the agent state into a multi-agent reinforcement learning strategy network pre-trained through reinforcement learning, and output a personalized intervention action for the employee from the strategy network; execute the personalized intervention action, and collect new first data and second data as feedback data through the employee terminal and the sensor after execution; and fine-tune the parameters of the time series prediction model and the multi-agent reinforcement learning strategy network in an online learning manner using the feedback data.

[0020] In this embodiment, it is specifically required to explain that the specific operation of buffering and aligning the first data and the second data in the step S1 is as follows: For each piece of first data received in real time, a text richness index is calculated, which is the ratio of the number of effective words in the natural language text contained in the first data (effective words refer to the remaining substantive words after excluding common stop words (such as ‘of’, ‘of’, ‘in’), which usually reflect the health status, such as ‘dizziness’, ‘fatigue’, ‘soreness’, etc.) to the total number of text words; For each piece of second data received in real time, a signal quality index is calculated, which is the ratio of the signal power of the physiological parameter signal represented by the second data to the noise power of the preset high-frequency noise band in a preset time length of the short calculation window, and the physiological parameter time series signal can specifically include heart rate, skin electricity, body movement, and the preset time length of the short calculation window is preferably 5 seconds to 30 seconds; the preset high-frequency noise band is set to a frequency band of 40Hz to 60Hz for the heart rate signal of the second data, which usually contains main noises such as power frequency interference, and the signal power and the noise power can be estimated by using the periodogram method or the Welch method; A sliding time window is established based on the preset window length (the preset sliding time window length is preferably 5 minutes to 15 minutes, for example, 10 minutes, to balance the real-time requirement and the stability of data statistics), and all the calculated text richness indexes and signal quality indexes in the window are used as the input of the subsequent weighting operation; After obtaining the text richness indicator and the signal quality indicator, a dynamic fusion weight factor is further calculated for each piece of first data and each segment of second data within the sliding time window, and the calculation process includes: The text richness indicator of the first data is directly taken as its instant quality score, and the signal quality indicator of the second data is divided by a preset normalization reference value (which can be set according to historical data statistical characteristics and can be 0.5) to obtain its instant quality score; A context consistency score is calculated, for the current data, the current data including the first data or the second data to be calculated, the context consistency score is obtained by calculating the Mahalanobis distance between the feature value of the current data and the mean value of the feature value set of all data of the same type within the sliding time window, and the reciprocal of the distance value is taken as the score result, for the first data, the feature value can be the text richness indicator or the semantic vector obtained by the word embedding model; for the second data, the feature value can be the time domain feature (such as mean value, standard deviation) or the frequency domain feature (such as power spectral density) extracted from the physiological parameter signal; the Mahalanobis distance used in the calculation of the context consistency score can effectively measure the deviation degree of the current data point from the overall distribution of the historical data in the sliding window, and the distance calculation considers the covariance structure of the data features, so it can better reflect the statistical sense of abnormality than the Euclidean distance; the technical effect lies in that it can automatically identify and suppress the outlier data points caused by instantaneous device interference or individual abnormal behavior, thereby ensuring the smoothness and logical coherence of the final fused feature vector sequence, and avoiding the dramatic fluctuations of the subsequent risk prediction model due to individual noise points; A time decay factor is calculated, which is the calculation result of an exponential function with a natural constant as the base, the absolute difference between the timestamp of the current data and the center time point of the sliding time window multiplied by a preset decay coefficient as the exponent, and the negative value as the result; The instant quality score and the context consistency score are multiplied by the first preset coefficient and the second preset coefficient respectively, and then added to obtain a weighted sum; The weighted sum is input into an S-shaped growth function for nonlinear mapping to obtain an intermediate weight value, the S-shaped growth function is a monotonically increasing function with an output value between zero and one, and the S-shaped growth function is preferably a Sigmoid function, which functions to map the value range of the weighted sum to the interval (0, 1) and provide a smooth nonlinear transformation in the interval, the first preset coefficient and the second preset coefficient are adjustable hyperparameters for balancing the relative importance of instant quality and context consistency in weight calculation, and their specific values can be determined through grid search or optimization algorithm in the system training stage, and the decay coefficient is preferably 0.1 to 0.5 to control the speed of time decay; The intermediate weight value is multiplied by the timeliness decay factor to obtain a dynamic fusion weight factor with a value range between zero and one. The dynamic fusion weight factor fuses the information of the three dimensions of real-time quality, context consistency and timeliness, and realizes fine quantization of the data value. The technical effect is that in the data fusion stage, the data with high quality, consistent with historical laws and strong timeliness can be adaptively relied on, so as to provide purer and more reliable feature input for the subsequent model. The optimization of the data level preprocessing cooperates with the subsequent complex model, and produces a synergistic effect of improving the overall performance, which is not a simple superposition of the effects of each step. After the above calculation, the output of the sliding time window is an enhanced aligned buffer set, which explicitly contains each first data and each second data in the window, and each data is associated with its corresponding dynamic fusion weight factor. The enhanced aligned buffer set can be implemented in a data structure such as a list or a dictionary in memory, for example, a record can be represented as (data unit identifier, data content, dynamic fusion weight factor), so as to ensure that each data and its dynamic fusion weight factor can be seamlessly retrieved and called by the identifier in the subsequent step.

[0021] In this embodiment, it is specifically necessary to explain that in step S2, the first data in the sliding time window is subjected to natural language processing to extract symptom entity features and sentiment vector features, and the second data is subjected to time domain and frequency domain feature extraction to obtain physiological feature vectors. Each first data and each second data in the enhanced aligned buffer set obtained from step S1 is read. For each first data, a named entity recognition model is used to identify symptoms entities representing health status from natural language text, and each identified symptom entity is converted into a fixed-dimensional symptom entity feature. The named entity recognition model is preferably based on BERT architecture and fine-tuned on biomedical text corpus, and can accurately identify symptoms entities closely related to workplace health such as "dizziness", "weakness", "shoulder and neck pain", etc. The fixed dimension is set to 100 dimensions, and the mapping of entities to vectors is realized by a pre-trained word embedding model such as Word2Vec or GloVe, which ensures the semantic richness of feature representation and the normativity of model processing. At the same time, a deep learning sentiment analysis model is used to process the natural language text, and outputs a sentiment vector feature representing multi-dimensional sentiment intensity. The sentiment dimensions at least include pleasantness and arousal. The sentiment analysis model can use a network structure based on RoBERTa, and the dimension of the output sentiment vector feature is set to 50. The introduction of multi-dimensional sentiment intensity representation instead of simple positive and negative judgment has the beneficial effect of being able to more finely capture subtle emotional changes of employees, for example, the combination of low pleasantness and high arousal may indicate an anxious state, thereby providing more discriminative feature input for subsequent health risk prediction. For each piece of second data, first, signal preprocessing is performed, including denoising and normalization, denoising can adopt wavelet transform threshold method, normalization can adopt Z-Score standardization method, to eliminate the influence of baseline drift and individual difference, then time domain features and frequency domain features are extracted from the preprocessed signal, time domain features include root mean square and approximate entropy, root mean square reflects signal energy, approximate entropy describes signal complexity, frequency domain features include power spectral density main frequency energy obtained through spectral analysis, in the calculation of approximate entropy, the pattern dimension is set to 2, and the tolerance threshold is set to 0.2 times the standard deviation of the signal; The power spectral density main frequency energy refers to the integral value of the signal in the frequency range of 0.04-0.15Hz, which corresponds to the low frequency component in heart rate variability, which can effectively reflect the sympathetic nervous activity level; Finally, the extracted time domain features and frequency domain features are spliced into a fixed dimension physiological feature vector, the fixed dimension can be set according to the number of selected features, for example, if 5 time domain features and 3 frequency domain features are selected, the physiological feature vector is 8-dimensional; This kind of multi-feature splicing method can comprehensively describe the physiological state from different angles, and its technical effect is to avoid the limitation of single feature, and to enhance the anti-interference ability and robustness of feature representation; At this point, the feature extraction of the first data and the second data is completed, the symptom entity feature and the emotion vector feature corresponding to each piece of first data are obtained, and the physiological feature vector corresponding to each piece of second data is obtained; The specific operation of calculating the cross-modal attention weight and generating the fusion feature vector based on the symptom entity feature, the emotion vector feature and the physiological feature vector by using the dynamic semantic alignment layer is as follows: First, using the dynamic fusion weight factor corresponding to each first data and second data calculated in step S1, respectively, the symptom entity feature, the emotion vector feature and the physiological feature vector are subjected to weighted average pooling operation, the weighted average pooling operation is to multiply the symptom entity features of all first data in the sliding window by the corresponding dynamic fusion weight factor and then sum them up, and then divide the sum by the sum of the dynamic fusion weight factors of all first data, so as to obtain the pooled symptom entity feature; In the same way, the emotion vector feature is pooled to obtain the pooled emotion vector feature, and the physiological feature vectors of all second data are pooled to obtain the pooled physiological feature vector, the beneficial effect of this design is that it organically integrates the dynamic fusion weight factor calculated in step S1 into the feature extraction processing stage, realizes the guiding effect of data quality on feature representation; The feature contribution of high-quality data is greater, so it occupies a dominant position in the pooled feature, effectively suppressing the negative influence of low-quality or noise data on the overall feature representation, providing a cleaner and more reliable feature basis for subsequent fusion; Then, the pooled symptom entity features, the pooled sentiment vector features and the pooled physiological feature vectors are respectively input to three different fully connected layers for linear transformation, each of which includes a trainable weight matrix and a bias vector, the linear transformation is used to map the features of different modalities to a common semantic space with the same dimension, and a hyperbolic tangent activation function is used for nonlinear activation of the transformation result, finally obtaining the projected symptom entity features, the projected sentiment vector features and the projected physiological feature vectors, the dimension of the common semantic space is a unified preset key parameter, which can be set to 64 in a preferred embodiment; using the hyperbolic tangent activation function can limit the projected feature values within the range of (-1, 1), which is conducive to the stability of training; the technical effect of this projection design is that it maps heterogeneous features of different sources and different physical meanings to a unified metric space, solving the semantic gap problem of direct comparison between modalities, and providing a fair comparison basis for subsequent attention weight calculation; Then, the cross-modal attention weights are calculated by using a trainable query vector to perform dot product operations with the projected symptom entity features, the projected sentiment vector features and the projected physiological feature vectors respectively, obtaining three initial attention scores, the query vector is a trainable parameter with the same dimension as the projected feature, for example, 64 dimensions; during model training, the query vector will learn to focus on those feature dimensions that are most important for the current health risk assessment; Subsequently, the three initial attention scores are input into a Softmax function for normalization to obtain three final attention weight values with a value range of zero to one and a sum of one, respectively corresponding to the symptom entity features, the sentiment vector features and the physiological feature vectors; Softmax normalization ensures the probabilistic interpretation of the weight values, and its function characteristics amplify the weight difference between important modalities and unimportant modalities, enabling the model to dynamically and adaptively adjust the attention degree to different information sources; for example, when the sensor data quality is very high, the weight of the physiological feature vector may automatically increase; and when the employee describes the key symptoms in the questionnaire, the weight of the symptom entity features will be significantly increased, this dynamic adjustment capability is not possessed by the simple feature splicing method, which embodies the intelligence and superiority of the present method; Finally, the projected symptom entity features, the projected sentiment vector features and the projected physiological feature vectors are respectively multiplied by the respective final attention weight values, and then the three weighted feature vectors obtained are added to generate a final fusion feature vector; this weighted sum operation realizes deep fusion of multi-modal information, and its technical effect lies in that the generated fusion feature vector not only retains the information of the original modal, but also reflects the relative importance of different modalities in the current context; the fusion feature vector as the input of step S3 can more comprehensively and accurately represent the comprehensive health status of the employee, thereby significantly improving the accuracy and reliability of the subsequent health risk prediction model. Through a series of operations of weighted pooling, projection, attention mechanism and weighted sum in this step, a synergistic effect that cannot be achieved by simply stacking each link is produced, that is, multi-modal feature fusion under data quality guidance, semantic alignment and dynamic self-adaptation is realized.

[0022] In this embodiment, it is specifically required to explain that the specific operation of inputting the fusion feature vector into the time series prediction model and simultaneously querying the pre-constructed occupational health knowledge graph to obtain the graph subgraph vector in step S3 is as follows: First, the fusion feature vector output by step S2 is input into a time series prediction model in time sequence. The time series prediction model adopts a recurrent neural network structure, specifically a gated recurrent unit network. The gated recurrent unit network effectively captures long-term dependencies in time series through update gate and reset gate mechanisms, avoiding the gradient disappearance problem in traditional recurrent neural networks. For each time step, the time series prediction model calculates the hidden state of the current time step based on the fusion feature vector of the current time step and the hidden state of the previous time step. The hidden state of the current time step is calculated based on the fusion feature vector of the current time step and the hidden state of the previous time step through a gated recurrent unit operation. The hidden state is a fixed-dimensional real vector, and the dimension is set to 64. The hidden state is used to encapsulate the time series dynamic information of the current and historical time steps, and its technical effect lies in that it can effectively learn the evolution law of the health status. At the same time, based on the symptom entity features extracted from the first data in step S2, the pre-constructed occupational health knowledge graph is queried. The knowledge graph is stored in a Neo4j graph database and efficiently retrieved through Cypher query language. The occupational health knowledge graph is stored in a graph structure, and the nodes represent health-related entities, including symptoms, diseases and risk factors. The edges represent the relationships between entities, including but not limited to "may cause", "related to" and "belong to symptoms". A rich semantic network is formed. When querying, the symptom entity features are taken as keys to retrieve entities and relationships directly connected to the keys in the knowledge graph to form a local subgraph; the local subgraph contains a central symptom node and all entities and relationships associated with the node at the first degree, ensuring the completeness of knowledge; Then, a graph attention network is used to encode the local subgraph to generate a graph subgraph vector; the graph attention network calculates the attention weight between each node in the local subgraph and its neighbor nodes through a multi-head attention mechanism, uses eight attention heads, each of which captures the relationship between nodes from different angles, aggregates the feature information of the neighbor nodes according to the weight, and finally generates a fixed-dimension graph subgraph vector through the output layer of the network after at least one graph attention layer, and the fixed dimension is set to 100, which ensures the richness of knowledge representation; At this point, the hidden state at each time step and the corresponding graph subgraph vector are obtained as the input for subsequent fusion; The specific operation of generating a knowledge-enhanced state vector and outputting a health risk prediction probability by fusing the graph subgraph vector and the hidden state generated by the time series prediction model through a cross-attention mechanism is as follows: First, the cross-attention mechanism is used to realize knowledge-enhanced fusion, in which the graph subgraph vector is taken as the query source, and the hidden state output by the time series prediction model is taken as the key and the value; specifically, the graph subgraph vector is linearly projected through a trainable first projection matrix to map it to a space with the same dimension as the hidden state to obtain a query vector; the projection operation converts the features of different modalities to the same semantic space, laying a foundation for subsequent attention calculation; At the same time, the hidden state is linearly projected through a trainable second projection matrix to obtain a key vector, and the hidden state is linearly projected through a trainable third projection matrix to obtain a value vector, and the parameters of the first, second and third projection matrices are learned during model training; the dimensions of the three projection matrices are set to 100x64, 64x64 and 64x64 respectively, and are adaptively adjusted through training; Then, the similarity score between the query vector and the key vector is calculated, and the similarity score is realized by the dot product operation of the query vector and the key vector, and is divided by a scaling factor, which is the square root of the dimension of the hidden state, i.e., divided by 8, to prevent the problem of gradient vanishing caused by too large inner product, to stabilize the training process; Next, the similarity score is input into a Softmax function for normalization to obtain an attention weight vector, and each element of the weight vector represents the relevance of the corresponding part of the graph subgraph vector to the current hidden state; the Softmax function ensures that the sum of all weights is 1, and has a clear probability interpretation; Based on the attention weight, the value vector is weighted and summed to generate a knowledge context vector, which encapsulates the most relevant knowledge information in the current timing context; the technical effect lies in dynamically screening the most important knowledge information and avoiding the interference of irrelevant knowledge; Subsequently, the knowledge context vector is linearly transformed through a trainable fourth projection matrix, and then the transformed result is added to the original hidden state through a residual connection, and then layer normalization processing is performed, the residual connection ensures that the model does not destroy the original timing information when injecting new knowledge, and the layer normalization improves the training stability, to balance the original timing information and the injected knowledge, and generate a knowledge-enhanced state vector; Finally, the knowledge-enhanced state vector is input into a fully connected layer, which has a trainable weight matrix and a bias term, for linear transformation of the knowledge-enhanced state vector; the weight matrix dimension of the fully connected layer is set to 64x32, and the bias term dimension is 32; Then the linearly transformed result is input into a Sigmoid activation function for non-linear mapping, and the output value is compressed to between zero and one, which is the health risk prediction probability of the current time step, and the numerical value is positively correlated with the health risk degree, and the closer the numerical value is to one, the higher the risk; the Sigmoid function maps continuous values to probability outputs, providing intuitive basis for subsequent decision-making.

[0023] In this embodiment, it needs to be specifically explained that in step S4, the current work context information of the corresponding employee is obtained from the enterprise information system, and the health risk prediction probability and the current work context information jointly constitute the agent state, and the agent state is input into the multi-agent reinforcement learning policy network to output the specific operation of the personalized intervention action: First, the current work context information of the corresponding employee is obtained from the enterprise information system in real time, and the current work context information includes the first type of information: real-time work task load intensity (obtained through the API interface of the enterprise project management system, quantified as the ratio of the number of current unfinished tasks to the total task capacity), the second type of information: current work role identifier (represented by one-hot encoding, such as [1, 0, 0] representing a development role and [0, 1, 0] representing a test role), and the third type of information: real-time environmental stress index (calculated by combining the number of attendance abnormalities, meeting intensity, and project urgency); The health risk prediction probability output by step S3 is spliced with the first type of information, the second type of information, and the third type of information, specifically, the health risk prediction probability, the real-time work task load intensity, the current work role identifier, and the real-time environmental stress index are combined into a column vector in a predetermined order, thereby forming a comprehensive agent state vector, the dimension of the agent state vector is four-dimensional, and the effective fusion of the health state and the working environment is realized; Then, the agent state vector is input into a multi-agent reinforcement learning strategy network pre-trained by reinforcement learning. The strategy network adopts a centralized training and distributed execution architecture, where centralized training refers to training the strategy network on the server side, and distributed execution refers to distributing the parameters of the strategy network to each employee terminal for independent decision-making. The multi-agent reinforcement learning strategy network first extracts features of the agent state vector through a fully connected layer to obtain the hidden representation of each agent. The activation function of the fully connected layer is ReLU function, and the hidden representation dimension is set to 128 dimensions. Next, the influence weight between the current agent and other adjacent agents in the workplace is calculated using an attention mechanism. The specific calculation process is as follows: the hidden representation of the current agent is spliced with the hidden representation of each adjacent agent, the spliced vector is linearly transformed through a trainable weight matrix, and then nonlinearly mapped through an activation function to obtain the initial influence score of each adjacent agent on the current agent. The activation function is LeakyReLU function, and the negative slope coefficient is set to 0.2. Then, all initial influence scores are input into the Softmax function for normalization processing, so that the sum of the influence weights of the current agent and all adjacent agents is one, and the final influence weight is obtained. The technical effect of this attention mechanism is that it can dynamically capture the mutual influence relationship in team collaboration. For example, when an adjacent employee is in a high load state, the system will automatically reduce the attention weight to it to avoid the transmission of negative emotions. According to the calculated influence weight, the hidden representations of adjacent agents are weighted and summed to obtain the context-aware hidden representation of the current agent. This step realizes the organic integration of swarm intelligence, so that the decision of each agent considers not only its own state but also the overall situation of the team. Finally, the context-aware hidden representation of the current agent is input into a fully connected layer, and the Softmax activation function is used to output the probability distribution of each possible intervention action. According to the probability distribution, the intervention action with the highest probability is selected as the personalized intervention action for the employee. Intervention actions include adjusting task allocation, suggesting rest duration, recommending health activities, etc. The system will automatically select the optimal intervention strategy according to the probability distribution. After executing the personalized intervention action, feedback data is collected and used to fine-tune the parameters of the time series prediction model and the multi-agent reinforcement learning strategy network in an online learning manner. First, the personalized intervention action output by the multi-agent reinforcement learning strategy network is executed, and after execution, a preset time interval (recommended setting is 2 hours to ensure that the intervention effect fully emerges) is waited for, new first data is collected through the employee terminal, and new second data is collected through the sensor, and the new first data and second data are used as feedback data; the setting of this time interval is verified through experiments, which can achieve the best balance between timeliness and accuracy; Then, using the feedback data, the trainable parameters of the time series prediction model used in step S3 and the trainable parameters of the multi-agent reinforcement learning strategy network used in step S4 are sequentially subjected to online incremental learning; for the time series prediction model, using the new first data and second data in the feedback data, the new health risk prediction probability is calculated according to the method of steps S1 to S3, and the loss function is calculated with the actual health state label (cross-entropy loss function is used), and the preliminary update parameters of the time series prediction model are generated by the gradient descent method (learning rate is set to 0.001); online learning enables the model to quickly adapt to changes in employee health status and improve prediction accuracy; For the multi-agent reinforcement learning strategy network, the policy gradient method based on importance sampling is used for updating, specifically: the agent state, the intervention action performed in the state, the reward value obtained after the action is performed, and the subsequent state are extracted from the feedback data; the importance weight is calculated, which is the ratio of the probability of the current strategy network selecting the intervention action to the probability of the old strategy network selecting the same intervention action, and the importance weight effectively solves the distribution mismatch problem in online learning; then, the policy gradient is calculated, which is the product of the importance weight and the advantage function with respect to the derivative of the policy network parameter, and the advantage function represents the advantage of the intervention action in the current state relative to the average action; the advantage function is calculated using the generalized advantage estimator, and the discount factor is set to 0.99; then, the update amount of the policy network parameter is calculated, which is equal to the learning rate multiplied by the expected value of the policy gradient, and the learning rate is a preset normal number less than one (recommended value is 0.0003); finally, the update amount is subtracted from the current policy network parameter to generate the preliminary update parameters of the strategy network; this process enables the strategy network to learn from the intervention effect and continuously optimize decision quality; Finally, the momentum update mechanism is used to soft-update the time series prediction model preliminary update parameters and the strategy network preliminary update parameters generated above, specifically including: obtaining the old parameters of the time series prediction model and the old parameters of the strategy network before online incremental learning, calculating the weighted average using the preliminary update parameters of the time series prediction model and the old parameters of the time series prediction model, and the weight is controlled by a preset update coefficient; the introduction of the update coefficient ensures the smoothness of parameter updating, and its technical effect lies in preventing model performance oscillation caused by new knowledge covering old knowledge; Likewise, the preliminary updated parameters of the policy network and the old parameters of the policy network are used to calculate a weighted average, and the weight is controlled by the same update coefficient. The parameters obtained after the weighted average are used as the final updated parameters of the time series prediction model and the policy network, respectively. This soft updating mechanism significantly improves the stability and convergence of the system. The update coefficient is preferably set to 0.01. Through this weighted average operation, the model parameters are smoothly transitioned, avoiding sharp jumps in parameter values and ensuring system stability. A small update coefficient ensures the progressive accumulation of system experience, absorbing new knowledge while not forgetting old experience.

[0024] Embodiment 2 The embodiment provides a workplace-oriented data processing system as shown in Figure 2 The workplace-oriented data processing system specifically comprises a data receiving and alignment module, a feature extraction and fusion module, a risk prediction module, and an intervention and learning module, wherein: The data receiving and alignment module is configured to receive health questionnaire data from an employee terminal and physiological parameter time series data from a sensor in real time through a stream processing framework, and buffer and align the data in a preset sliding time window. The feature extraction and fusion module is configured to extract symptom entity features and emotion vector features from the questionnaire data in the window, extract physiological feature vectors from the sensor data, and calculate cross-modal attention weights using a dynamic semantic alignment layer to generate a fusion feature vector. The risk prediction module is configured to input the fusion feature vector into a time series prediction model, combine a graph subgraph vector obtained from a professional health knowledge graph, and generate a knowledge-enhanced state vector through a cross-attention mechanism to output a health risk prediction probability. The intervention and learning module is configured to construct an agent state according to the health risk prediction probability and work context information obtained from an enterprise information system, output personalized intervention actions through a multi-agent reinforcement learning policy network, and perform online fine-tuning on the time series prediction model and the policy network parameters based on feedback data collected after the execution of the actions.

[0025] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0026] Those skilled in the art will appreciate that embodiments of the present application can be devised for a variety of applications. It is intended that the present application be limited only by the scope of the appended claims, and it is intended that various modifications and alterations made by those skilled in the art be considered as within the scope of the present application. The embodiments of the present application will be described with reference to the attached drawings identified below.

[0027] The present application is described in reference to the drawings using a flowchart and / or a block diagram of the method, apparatus (system) and computer program product according to embodiments of the application. It will be understood that each block of the flowchart and / or block diagram, and combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0028] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0029] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0030] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to cover all such modifications and variations as fall within the scope of the present application.

[0031] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A data processing method for a workplace, characterized by, Specifically, the following steps are included: Step S1: Receive first data from the employee terminal and second data from the sensor in real time through a stream processing framework, and place the first data and second data within a sliding time window of a preset time length for buffering and alignment; the first data is health questionnaire data containing natural language text, and the second data is time-series data containing physiological parameters. Step S2: Perform natural language processing on the first data within the sliding time window to extract the symptom entity features and emotion vector features contained therein. At the same time, perform time domain and frequency domain feature extraction on the second data within the sliding time window to obtain physiological feature vectors. Then, using a dynamic semantic alignment layer, calculate cross-modal attention weights based on the symptom entity features, emotion vector features and physiological feature vectors, and generate fusion feature vectors based on the weights. Step S3: Input the fused feature vector into the time series prediction model, and simultaneously query the pre-constructed occupational health knowledge graph to obtain the graph sub-graph vector associated with the symptom entity features from the knowledge graph; during the processing of the time series prediction model, use the hidden state generated by the time series prediction model to fuse the graph sub-graph vector with the hidden state through a cross-attention mechanism to generate a knowledge-enhanced state vector, and output the health risk prediction probability based on this. Step S4: Obtain the current work context information of the corresponding employee from the enterprise information system, and combine the health risk prediction probability with the current work context information to form an agent state; input the agent state into a multi-agent reinforcement learning policy network that has been pre-trained through reinforcement learning, and have the policy network output a personalized intervention action for the employee; execute the personalized intervention action, and after execution, collect new first and second data as feedback data through the employee terminal and sensors; use the feedback data to fine-tune the parameters of the time-series prediction model and the multi-agent reinforcement learning policy network in an online learning manner.

2. The workplace-oriented data processing method according to claim 1, characterized in that: In step S1, the specific operation of placing the first data and the second data within a sliding time window of a preset time length for buffering and alignment is as follows: For each piece of first data received in real time, calculate its text richness index, which is the ratio of the number of effective words in the natural language text contained in that piece of first data to the total number of words in the text. For each segment of second data received in real time, its signal quality index is calculated. The signal quality index is the ratio of the signal power of the physiological parameter signal represented by the segment of second data to the noise power of a preset high-frequency noise band within a short calculation window of preset duration.

3. The workplace-oriented data processing method according to claim 2, characterized in that: After obtaining the text richness index and signal quality index, a dynamic fusion weight factor is further calculated for each piece of first data and each piece of second data within the sliding time window. The calculation process includes: The text richness index of the first data is directly used as its instant quality score, and the signal quality index of the second data is divided by the preset normalized benchmark value to obtain its instant quality score. To calculate the context consistency score, for the current data, which includes the first or second data to be calculated, the context consistency score is obtained by calculating the Mahalanobis distance between the feature value of the current data and the mean of the feature value set of all data of the same type within the sliding time window, and the reciprocal of the distance value is used as the score result. The time-related decay factor is calculated using an exponential function with the natural constant as the base and the negative value of the result of multiplying the absolute difference between the current data's timestamp and the center time point of the sliding time window by a preset decay coefficient. The instant quality score and the contextual consistency score are multiplied by the first preset coefficient and the second preset coefficient, respectively, and then summed to obtain a weighted sum; The weighted sum is input into the sigmoid growth function for nonlinear mapping to obtain the intermediate weight values; Multiply the intermediate weight value by the time-related decay factor to obtain a dynamic fusion weight factor with a value between zero and one. After the above calculations, the output of the sliding time window is an enhanced alignment buffer set, which explicitly contains every first data point and every second data point within the window, and each data point is associated with its corresponding dynamic fusion weight factor.

4. A workplace-oriented data processing method according to claim 3, characterized in that: In step S2, the specific operations of performing natural language processing on the first data within the sliding time window to extract symptom entity features and emotion vector features, and simultaneously performing time-domain and frequency-domain feature extraction on the second data to obtain physiological feature vectors are as follows: Read each first data and each second data from the enhanced alignment buffer set obtained in step S1. For each first data, use the named entity recognition model to identify symptom entities representing health status from natural language text, and convert each identified symptom entity into a fixed-dimensional symptom entity feature. Meanwhile, a deep learning sentiment analysis model is used to process natural language text and output a sentiment vector feature that represents the intensity of multi-dimensional sentiment. For each segment of the second data, signal preprocessing is first performed, including denoising and normalization. Then, time-domain features and frequency-domain features are extracted from the preprocessed signal. The time-domain features include root mean square and approximate entropy, and the frequency-domain features include power spectral density and main frequency energy obtained through spectrum analysis. Finally, the extracted time-domain features and frequency-domain features are concatenated into a fixed-dimensional physiological feature vector. At this point, feature extraction of the first and second data is complete, resulting in symptom entity features and emotional vector features corresponding to each piece of the first data, as well as physiological feature vectors corresponding to each piece of the second data.

5. A workplace-oriented data processing method according to claim 4, characterized in that: The specific operation of using the dynamic semantic alignment layer to calculate cross-modal attention weights based on symptom entity features, emotional vector features, and physiological feature vectors, and to generate fused feature vectors, is as follows: First, using the dynamic fusion weight factor corresponding to each first data and second data calculated in step S1, weighted average pooling operations are performed on the symptom entity features, emotional vector features, and physiological feature vectors respectively. The weighted average pooling operation means: multiplying the symptom entity features of all first data in the sliding window by their corresponding dynamic fusion weight factors, summing them, and then dividing by the sum of the dynamic fusion weight factors of all first data to obtain the pooled symptom entity features. In the same way, the emotion vector features are pooled to obtain pooled emotion vector features, and the physiological feature vectors of all second data are pooled to obtain pooled physiological feature vectors. Then, the pooled symptom entity features, pooled emotion vector features, and pooled physiological feature vectors are input into three different fully connected layers for linear transformation. Each fully connected layer contains a trainable weight matrix and a bias vector. The linear transformation is used to map the features of different modalities to a common semantic space with the same dimension. The hyperbolic tangent activation function is used to perform non-linear activation on the transformation results, and finally the projected symptom entity features, projected emotion vector features, and projected physiological feature vectors are obtained. Next, cross-modal attention weights are calculated by using a trainable query vector and performing dot product operations with the projected symptom entity features, the projected emotion vector features, and the projected physiological feature vector, respectively, to obtain three initial attention scores. Subsequently, the three initial attention scores are input into the Softmax function for normalization, resulting in three final attention weight values ​​that range from zero to one and sum to one, corresponding to symptom entity features, emotional vector features, and physiological feature vectors, respectively. Finally, the projected symptom entity features, projected emotional vector features, and projected physiological feature vectors are multiplied by their respective final attention weight values, and then the three weighted feature vectors are added together to generate the final fused feature vector.

6. A workplace-oriented data processing method according to claim 5, characterized in that: In step S3, the specific operation of inputting the fused feature vector into the time series prediction model and simultaneously querying the pre-constructed occupational health knowledge graph to obtain the graph subgraph vector is as follows: First, the fused feature vector output from step S2 is input into a time-series prediction model in chronological order. This time-series prediction model adopts a recurrent neural network structure, specifically a gated recurrent unit network. For each time step, the time series prediction model calculates the hidden state of the current time step based on the fused feature vector of the current time step and the hidden state of the previous time step; wherein, the hidden state of the current time step is calculated by gated recurrent unit operation based on the fused feature vector of the current time step and the hidden state of the previous time step. Simultaneously, based on the symptom entity features extracted from the first data in step S2, a pre-constructed occupational health knowledge graph is queried; The occupational health knowledge graph is stored in a graph structure, where nodes represent health-related entities, including symptoms, diseases, and risk factors, and edges represent the relationships between entities. During the query, the entity features of the symptom are used as keys to retrieve the entities and relationships directly connected to them in the knowledge graph, forming a local subgraph; Then, a graph attention network is used to encode the local subgraph to generate a graph subgraph vector. The graph attention network calculates the attention weight between each node in the local subgraph and its neighboring nodes through a multi-head attention mechanism, and aggregates the feature information of the neighboring nodes based on the weight. After processing by a graph attention layer, a fixed-dimensional graph subgraph vector is finally generated through the output layer of the network. At this point, the hidden state and the corresponding graph subgraph vector for each time step are obtained.

7. A workplace-oriented data processing method according to claim 6, characterized in that: The specific operation of using the hidden state generated by the time-series prediction model and fusing the graph subgraph vector with the hidden state through the cross-attention mechanism to generate a knowledge-enhanced state vector and output the health risk prediction probability is as follows: First, a cross-attention mechanism is used to achieve knowledge enhancement fusion, in which the graph subgraph vector is used as the query source and the hidden state output by the time series prediction model is used as the key and value. Specifically, the graph subgraph vector is linearly projected through a trainable first projection matrix to map it into a space with the same dimension as the hidden state, thus obtaining the query vector; Simultaneously, the hidden state is linearly projected through a trainable second projection matrix to obtain the key vector, and the hidden state is linearly projected through a trainable third projection matrix to obtain the value vector. Then, the similarity score between the query vector and the key vector is calculated. The similarity score is achieved by the dot product of the query vector and the key vector and divided by a scaling factor, which is the square root of the hidden state dimension. Next, the similarity score is input into the Softmax function for normalization to obtain the attention weight vector; Based on attention weights, the value vectors are weighted and summed to generate a knowledge context vector; Subsequently, the knowledge context vector is linearly transformed through a trainable fourth projection matrix, and then the transformed result is added to the original hidden state through a residual connection. Finally, through layer normalization, a knowledge-enhanced state vector is generated. Finally, the knowledge-enhanced state vector is input into a fully connected layer with a trainable weight matrix and bias terms to perform a linear transformation on the knowledge-enhanced state vector. The result of the linear transformation is then input into the Sigmoid activation function for nonlinear mapping, compressing the output value to between zero and one. This output value is the health risk prediction probability at the current time step.

8. A workplace-oriented data processing method according to claim 7, characterized in that: In step S4, the specific operations of obtaining the current work context information of the corresponding employee from the enterprise information system, combining the health risk prediction probability with the current work context information to form the agent state, and inputting the agent state into the multi-agent reinforcement learning policy network to output personalized intervention actions are as follows: First, the current work context information of the corresponding employee is obtained in real time from the enterprise information system. The current work context information includes the first type of information: real-time work task load intensity, the second type of information: current work role identifier, and the third type of information: real-time environmental stress index. The health risk prediction probability output in step S3 is concatenated with the first type of information, the second type of information, and the third type of information to form a comprehensive agent state vector. Then, the agent's state vector is input into a multi-agent reinforcement learning policy network that has been pre-trained through reinforcement learning; the multi-agent reinforcement learning policy network first extracts features from the agent's state vector through a fully connected layer to obtain the hidden representation of each agent; Next, the attention mechanism is used to calculate the mutual influence weights between the current agent and other neighboring agents in the workplace. The specific calculation process is as follows: the hidden representation of the current agent is concatenated with the hidden representation of each neighboring agent, the concatenated vector is linearly transformed through a trainable weight matrix, and then nonlinearly mapped through an activation function to obtain the initial influence score of each neighboring agent on the current agent. Then, all initial influence scores are input into the Softmax function for normalization, so that the sum of the influence weights of the current agent and all neighboring agents is one, thus obtaining the final influence weight; Based on the calculated influence weights, the hidden representations of neighboring agents are weighted and summed to obtain the context-aware hidden representation of the current agent. Finally, the context-aware hidden representation of the current agent is passed through a fully connected layer, and the probability distribution of each possible intervention action is output using the Softmax activation function. Based on this probability distribution, the intervention action with the highest probability is selected as the personalized intervention action for the employee.

9. A workplace-oriented data processing method according to claim 8, characterized in that: The specific steps for collecting feedback data after performing personalized intervention actions and using this feedback data to fine-tune the parameters of the time-series prediction model and the multi-agent reinforcement learning policy network through online learning are as follows: First, the personalized intervention action output by the multi-agent reinforcement learning policy network is executed, and after execution, a preset time interval is waited. New first data is collected through the employee terminal and new second data is collected through the sensor. The new first data and the new second data are used as feedback data. Then, using the feedback data, online incremental learning is performed on the trainable parameters of the time-series prediction model used in step S3 and the trainable parameters of the multi-agent reinforcement learning policy network used in step S4. For the time-series prediction model, using the new first and second data in the feedback data, the new health risk prediction probability is calculated according to the method of steps S1 to S3, and the loss function is calculated with the actual health status label. The initial update parameters of the time-series prediction model are generated by gradient descent. For multi-agent reinforcement learning policy networks, an importance sampling-based policy gradient method is used for updating to generate preliminary update parameters for the policy network. Finally, a momentum update mechanism is used to perform soft updates on the preliminary update parameters of the generated time series prediction model and the preliminary update parameters of the policy network. Specifically, this includes: obtaining the old parameters of the time series prediction model and the old parameters of the policy network before online incremental learning; and performing a weighted average calculation using the preliminary update parameters of the time series prediction model and the old parameters of the time series prediction model, with the weights controlled by preset update coefficients. Similarly, a weighted average is calculated using the initial updated parameters of the policy network and the old parameters of the policy network, with the weights controlled by the same update coefficients. The parameters obtained after weighted averaging are used as the final updated parameters of the time series prediction model and the policy network, respectively.

10. A workplace-oriented data processing system applied to a workplace-oriented data processing method as described in any one of claims 1-9, characterized in that: Specifically, it includes: The system comprises a data receiving and alignment module, a feature extraction and fusion module, a risk prediction module, and an intervention and learning module. The data receiving and alignment module is used to receive health questionnaire data from employee terminals and physiological parameter time-series data from sensors in real time through a stream processing framework, and to buffer and align the data within a preset sliding time window. The feature extraction and fusion module is used to extract symptom entity features and emotional vector features from the questionnaire data in the window, extract physiological feature vectors from the sensor data, and use the dynamic semantic alignment layer to calculate cross-modal attention weights to generate fused feature vectors. The risk prediction module is used to input the fused feature vector into the time series prediction model, and combine it with the graph subgraph vector obtained from the occupational health knowledge graph query to generate a knowledge-enhanced state vector through a cross-attention mechanism, and then output the health risk prediction probability. The intervention and learning module is used to construct the agent state based on the predicted health risk probability and the work context information obtained from the enterprise information system, output personalized intervention actions through a multi-agent reinforcement learning policy network, and fine-tune the time-series prediction model and policy network parameters online based on the feedback data collected after the actions are executed.

Images