Deep learning-based water digital employee task self-driving implementation method

Abstract

The present application relates to a kind of water affair digital employee task self-driving implementation method based on deep learning, belong to water affair industry information technology field.The technical scheme: using pre-constructed task identification model learning and reproducing human in water supply operation scene in task identification and decision logic, self-determination task to be executed and generates identification result, task is distributed to corresponding water affair digital employee execution, after water affair digital employee disposal process ends, self-driving system receives the structured task completion record of feedback of water affair digital employee, structured task completion record is added to training set as new training sample, the task identification model is retrained and dynamically iterated updates.The present application simulates human decision-making thinking by deep learning, realizes the self-driving mode of digital employee task, effectively solves the problem of task identification quality not up to standard, improves automation level, reduces labor cost.

Description

Technical Field

[0001] This invention relates to a deep learning-based method for the self-driven implementation of tasks by digital employees in the water industry, belonging to the field of information technology in the water industry. Background Technology

[0002] With the development of large-scale modeling technology, the application of artificial intelligence (AI) is becoming increasingly widespread across various industries. Digital employees, as a new application form, have been widely used in multiple fields. Digital employees, also known as digital workforce or virtual employees, are virtual labor forces built based on technologies such as Robotic Process Automation (RPA), artificial intelligence, and big data analytics. They can simulate human employees to complete routine business processes. In the digital transformation of the water industry, traditional business models are gradually shifting towards intelligence and automation. In existing technologies, AI-based digital employees can, to some extent, replace manual labor in completing specific business tasks, such as customer service inquiries and zonal leakage investigation, thereby improving the efficiency and service level of water services. However, the following problems are encountered in actual use:

[0003] (1) Tasks are usually initiated by users, and human intervention may not be able to detect anomalies in a timely manner, resulting in a certain degree of lag.

[0004] (2) In complex water affairs scenarios, business problems involve multi-dimensional data and complex business logic. Rules cannot handle complex anomalies, making identification and processing difficult, relying heavily on manual judgment, and resulting in low work efficiency. Summary of the Invention

[0005] This invention proposes a deep learning-based method for the self-driven implementation of tasks for water management digital employees. By simulating human decision-making through deep learning, it realizes a self-driven task mode for water management digital employees. Furthermore, deep learning effectively solves the problem of insufficient task recognition quality in the self-driven process of water management digital employees, improves the level of automation, reduces labor costs, and solves the aforementioned technical problems existing in existing technologies.

[0006] The technical solution of this invention is:

[0007] A deep learning-based method for self-driven task implementation by digital water employees is proposed. This method acquires and cleans employee operation and maintenance log data, transforming it into structured log data. Historical water supply business data is then obtained from this structured log data and converted into training samples with task labels. A task recognition dataset is constructed, and a task recognition model is trained based on a deep neural network. The self-driven system acquires real-time water supply business data from water-related subsystems to monitor the operational status of the water supply network and equipment. Using a pre-built task recognition model, it learns and reproduces human task recognition and decision-making logic in water supply operation and maintenance scenarios. The system autonomously determines tasks to be executed, generates recognition results, and assigns tasks to corresponding digital water employees. After the digital water employee's processing is completed, the self-driven system receives structured task completion records from the digital water employee. These records are added to the training set as new training samples, and the task recognition model is retrained and dynamically iterated, achieving automated closed-loop optimization of the self-driven processing flow of digital water employees.

[0008] The real-time water supply business data refers to multi-dimensional data collected in real time by the water supply system to reflect the operational status of the water supply business, including but not limited to reporting rate, total water supply, total water consumption, and alarm data.

[0009] The specific steps of this invention are as follows:

[0010] S1. Obtain employee operation and maintenance log data and clean the employee operation and maintenance log data; use a large language model to perform structured processing on the employee operation and maintenance log data, complete the extraction of key information and event extraction, obtain structured employee operation and maintenance log data, and obtain historical water supply business data through structured employee operation and maintenance log data.

[0011] S2. Convert historical water supply business data into training samples with task labels, construct a task recognition dataset, and train the task recognition model based on a deep neural network. The self-driven system connects to the API interface of water supply-related business subsystems to obtain real-time water supply business data as input. It uses the preset task recognition model to monitor the operation status of the water supply network and equipment in real time. When an abnormal event is encountered, the built-in trained task recognition model is triggered to automatically identify the abnormal label, automatically map the abnormal label to an executable task instruction, and output the task recognition result. This realizes the automatic identification and result output of the task, thus forming a self-driven task recognition mechanism oriented towards business scenarios.

[0012] S3. The task identification results output in step S2 are circulated and distributed. The task identification results are automatically mapped to the corresponding water digital employees through the preset task type-water digital employee mapping table. The water digital employees automatically take over the handling process of abnormal events, realizing a zero-human intervention transition from data monitoring to task assignment. The water digital employees perform corresponding standardized business handling operations. After the entire handling process is completed, the self-driven system automatically extracts the execution actions, processing time, task flow nodes, and pipeline indicator recovery status of the water digital employees, and generates a structured task completion record.

[0013] S4. The structured task completion record generated in step S3 is used as verification feedback data and transmitted back to the self-driving system in real time along the optimization link. The self-driving system extracts the structured fields in the task completion record regarding the executed actions, processing time, task flow nodes, and pipeline indicator recovery status. The structured fields are used as incremental learning sample feedback signals to dynamically iterate and update the weight parameters of the task recognition model and the matching strategy of the business rule engine, thereby realizing the automated closed-loop optimization of the water supply self-driving process.

[0014] Furthermore, in step S1, in order for the self-driving system to be able to directly and accurately identify anomalies in real-time water supply business data without relying on manually recorded employee operation and maintenance logs, the employee operation and maintenance log data must first be cleaned and processed. Step S1 specifically includes:

[0015] S11. First, obtain the employee operation and maintenance log data for each historical time period; use a large language model for structured processing and cleaning, extract key information and events from the employee operation and maintenance log data, and transform it into structured employee operation and maintenance log data; obtain historical water supply business data through the structured employee operation and maintenance log data.

[0016] S12. After acquiring the structured employee operation and maintenance log data, the self-driven system constructs multi-dimensional mapping rules and adopts a funnel-shaped fault-tolerant matching mechanism to accurately align the structured employee operation and maintenance log data with historical water supply business data. After mapping, the initial business data—the structured employee operation and maintenance log dataset—is obtained.

[0017] S13. Based on the historical water supply business data obtained after mapping the structured employee operation and maintenance log data, extract multimodal features. The multimodal features include numerical features and text features. The numerical features are derived from the indicator data in the water supply-related business subsystems; the text features are derived from the text data in the water supply-related business subsystems; and feature selection processing is performed on the numerical features.

[0018] S14. Based on the business data—structured employee operation and maintenance log dataset, construct a task identification type label system to form a business data—task identification dataset;

[0019] S15. Process and clean the business data—task identification dataset, specifically including removing or filling missing values, normalization, and vectorization, to form the final task identification dataset.

[0020] Furthermore, the multi-dimensional mapping rules in step S12 include two layers: the first layer is strong rule precise matching: priority is given to extracting the work order number, abnormal situation occurrence time, customer information, and meter number from the employee operation and maintenance log data, and performing cross-table primary key matching with the water supply-related business subsystem database;

[0021] The second layer is spatiotemporal slicing fuzzy matching: if the first layer does not match, the occurrence time extracted from the employee operation and maintenance log data is used as the benchmark, a time offset window is set, and a slice is constructed by combining the customer information entity in the employee operation and maintenance log data; within this slice range, historical water supply business data with abnormal fluctuations are retrieved and bound.

[0022] Furthermore, the feature selection process for numerical features in step S13 specifically includes: calculating the linear correlation coefficient between any two numerical features, where the two numerical features are respectively... and When numerical features With numerical features When the absolute value of the correlation coefficient reaches or exceeds the preset threshold, it is determined that there is a high linear correlation between the two features, and one of the features is deleted according to the preset rules to reduce feature redundancy.

[0023] The formula is as follows:

[0024]

[0025] in, Representation of features With features The linear correlation coefficient between them;

[0026] and They represent the first , The feature in the first The values ​​for each sample;

[0027] and These represent the average values ​​of the corresponding features within the statistical period;

[0028] Indicates the number of samples.

[0029] Furthermore, step S14 specifically includes: using the successfully mapped historical water supply business data as the input feature variable X for model training; and simultaneously, parsing the behavioral patterns in the corresponding structured employee operation and maintenance log data and converting them into task type labels Y, thereby forming the business data-task identification dataset.

[0030] Furthermore, step S15 processes and cleans the business data—task identification dataset, specifically including:

[0031] For numerical data processing: remove duplicate data records, fill in or remove missing values, and standardize key business indicators to eliminate the impact of differences in units on model training.

[0032] The standardization process uses the Z-score, for any original data point. Its corresponding Z-score is defined as:

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039]

[0040]

[0041] in, Represents the original numerical data. This represents the result of standardizing the original numerical data. The data is normalized by the mean and standard deviation to make the data follow a standard normal distribution with a mean of 0 and a standard deviation of 1. The average value of the sample data is calculated by summing all sample values ​​and dividing by the number of samples. The standard deviation indicates the dispersion of sample data and measures how much the data fluctuates relative to the mean. The larger the standard deviation, the more dispersed the data distribution. This represents the total number of data samples involved in the statistical calculation;

[0042] For text data processing: a BERT-based method is used to extract deep semantic feature vectors from the text; specifically, for text data in water supply-related business subsystems, the text is input into a pre-trained BERT model, and its bidirectional encoder is used to obtain context-related high-dimensional semantic representations.

[0043] Mapped to vector representation via the BERT model:

[0044]

[0045] Concatenate numerical features with textual features:

[0046] ,in The numerical feature vector is constructed based on the mapped historical water supply business data. The numerical features are derived from structured indicator data in the water supply-related business subsystems, such as flow rate, pressure, water supply volume and reporting rate, and are obtained after normalization. These are text-based feature vectors obtained by semantically encoding business text data using a pre-trained language model; they are then concatenated to form a unified multimodal feature representation, which is used as input for subsequent task recognition models.

[0047] Furthermore, the training of the task recognition model in step S2 specifically includes:

[0048] S21. The unified multimodal feature representation formed by concatenation in step S15 is used as a high-dimensional input tensor and input to the neural network mapping function. middle;

[0049] A mapping function representing the deep neural network model itself; The input data includes numerical features and text features; The set of parameters representing a neural network includes: the weight matrix for each layer. Bias vector ;

[0050] S22. The deep neural network model has two hidden layers, and the hidden layers use... The activation function performs a non-linear mapping on the input numerical and textual features; the output layer uses... The activation function outputs the probability of each class; the inverse of the log-likelihood function is used as the loss function. , used to optimize model parameters;

[0051] S23. To address the overfitting problem during model training, an L2 regularization mechanism is introduced into the training process of deep neural networks. Specifically, when calculating the overall loss function of the model, the self-driven system adds a regularization penalty term proportional to the sum of the squares of all weight parameters in the network, on top of the original empirical error. During backpropagation and gradient update, this mechanism applies continuous decay pressure to larger weight parameters, forcing the network to tend to learn a smaller and more uniformly distributed weight matrix. This effectively limits the overall complexity of the model, reduces the model's over-reliance on local features or noise in the training set, and significantly improves the model's generalization ability on unknown data.

[0052] Furthermore, in step S3, the process of transferring and distributing the task identification results specifically includes: the task identification model uses a variety of preset task identification results as output categories, with each task identification result corresponding to a type of water supply business scenario. The self-driving system is pre-configured with a task identification result-water digital employee mapping table, which maps the above-mentioned multiple task identification results to multiple water digital employees one-to-one, as shown in Table 2.

[0053] When the task recognition model outputs a task recognition result, and the probability of the task recognition result is not lower than a preset threshold, the self-driving system queries the task recognition result-water digital employee mapping table, directly determines the water digital employee corresponding to the task recognition result, and transmits the real-time water supply business data that triggered the recognition at the same time as the execution context to the water digital employee to complete the task distribution; when the probability of the output result is lower than the preset threshold, it is determined that the current water supply business does not require the processing of water digital employees, and no water digital employees are triggered.

[0054] For example, when the task identification model outputs the task identification result as "leakage analysis" and its probability is not lower than the preset threshold, the self-driving system queries the mapping table to determine the corresponding water digital employee as the leakage analyst, and then passes the real-time water supply business data at that moment as the execution context to the leakage analyst to complete the task distribution.

[0055] The main innovative points of this invention include the following:

[0056] 1. Addressing the reliance on manual instructions and preset rules in traditional systems, and the lag in anomaly identification, this application innovatively introduces a deep learning model to uncover the implicit nonlinear relationships between data. This enables the system to proactively and flexibly identify task scenarios, autonomously generate and drive task execution, significantly improving the system's self-driving capability and intelligent decision-making level. 2. Addressing the issue of diverse data dimensions and complex structures in water supply scenarios, this application innovatively introduces a large language model to structure complex multidimensional data and designs a multi-layer data mapping mechanism, solving the problem of difficulty in unifying and aligning multidimensional complex data structures. 3. The system transmits the task execution results of the water affairs digital employee back to the task recognition model in real time. The task recognition model adds this real result to the training samples and dynamically iterates and updates the task recognition model parameters, enabling the task recognition model to continuously optimize the task judgment logic based on the execution results, thereby improving the accuracy of task recognition.

[0057] Compared with existing technologies, the present invention has the following advantages: by simulating human decision-making thinking through deep learning, a self-driven mode for tasks of water affairs digital employees is realized, and the problem of insufficient task recognition quality of water affairs digital employees in the self-driven process is effectively solved through deep learning, thereby improving the level of automation and reducing labor costs. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the process of an embodiment of the present invention;

[0059] Figure 2 This is a flowchart illustrating the implementation of the self-driving system in an embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0061] This invention discloses a deep learning-based method for self-driven task implementation in the water sector's digital employee platform. Existing water sector digital employees mostly rely on rule engines or preset processes, lacking the ability to autonomously judge and make decisions regarding complex tasks. By training a task recognition model using deep learning and applying it to the digital platform, the system can autonomously judge and execute tasks, improving response speed. Through offline training and online learning of the model, the self-driven system can continuously optimize task recognition performance as business and environmental conditions change.

[0062] First, the self-driven system acquires employee operation and maintenance log data within a historical time period through offline batch processing and API calls. Addressing the issues of unstructured, conversational, and inconsistent formatting of employee operation and maintenance log data, this embodiment innovatively introduces a Large Language Model (LLM) for deep semantic parsing. Specifically, the self-driven system inputs preset prompt words and the original log text into the LLM, driving the LLM to perform named entity recognition and relation extraction tasks, accurately extracting key business entities from complex free text. These key business entities include, but are not limited to: event occurrence time, corresponding water supply subsystem (meter reading, zone, etc.), customer information (such as customer name, address), meter type (such as NB water meter, flow meter), description of abnormal phenomena (such as sudden drop in water pressure, turbid water quality), and task type (such as investigation and reporting rate, leakage analysis). The LLM transforms the extracted entities into a unified structured data format, serving as a standard index dictionary for subsequent extraction and matching of related business data.

[0063] Specifically, a prompt word engineering template for log information extraction is constructed, which includes: a role definition module, a task description module, a field constraint module, an output format constraint module, and an example module.

[0064] In one embodiment, the prompt word engineering template may take the following form:

[0065] # Role

[0066] You are a water utilities operations and maintenance log parsing assistant, responsible for parsing unstructured logs.

[0067] Extract key business entities.

[0068] # Task

[0069] Perform named entity recognition and event extraction on the input operation and maintenance logs, and identify and...

[0070] Output the following seven fields: Event Time, Water Supply Subsystem, Customer Name,

[0071] Address, device type, anomaly, and task type.

[0072] # Field constraints

[0073] Each field must conform to the preset value specifications, among which "water supply subsystem" and

[0074] "Task type" is taken from a predefined enumeration set; time field is formatted uniformly;

[0075] Unrecognized fields output null.

[0076] # Output Format

[0077] The output is in JSON format and no additional explanation is provided.

[0078] # Example

[0079] Input: "On March 15, 2025, xx Water Company responded to feedback from residents of xx residential area."

[0080] Abnormal water pressure was detected, and on-site investigation revealed an abnormally high regional water supply flow rate.

[0081] The leakage analysis task has been issued.

[0082] Output:

[0083] {

[0084] Event Time: 2025-03-15 15:23:12

[0085] Water Supply Subsystem: Meter Reading System

[0086] Customer Name: "xx Water Supply",

[0087] Address: "xx Residential Area"

[0088] Device Type: "NB Water Meter"

[0089] "Abnormal Phenomenon": "Abnormal water pressure, abnormally high water flow rate",

[0090] Task Type: "Leakage Analysis"

[0091] }

[0092] # Now please process the following logs:

[0093] Input: {text}

[0094] Output:

[0095] Those skilled in the art can make adaptive adjustments to the field constraints, enumeration value ranges, and examples in the above-mentioned prompt word template according to actual business scenarios, all of which fall within the protection scope of this invention.

[0096] After structuring the employee operation and maintenance log data, the system employs a funnel-shaped fault-tolerant matching mechanism to accurately align the structured employee operation and maintenance log data with historical water supply business data. Specifically, this includes:

[0097] The first layer (strong rule exact matching): Prioritizes extracting strong unique identifiers from structured employee operation and maintenance log data, such as work order numbers, precise occurrence times, customer information (e.g., unique customer IDs), and meter numbers. The self-driven system uses these unique identifiers to directly perform exact matching between primary keys or foreign keys across tables in the water supply-related business subsystem databases (e.g., meter reading systems, revenue systems, customer service systems). If a match is successful, the first-level data mapping is established.

[0098] The second layer (spatiotemporal slice fuzzy matching): If the first layer fails to match due to missing log records or inconsistent IDs, the self-driven system activates the spatiotemporal slice association mechanism. Using the "occurrence time" extracted from the structured employee operation and maintenance log data as a reference point, a preset time offset window is set (e.g., 2 hours before and after). Simultaneously, a two-dimensional spatiotemporal data slice is constructed by combining the "customer information entity" (such as the customer's water supply area or geographical coordinates) extracted from the structured employee operation and maintenance log data. Within the spatiotemporal slice, the self-driven system performs a range search on the historical water supply business database, extracts historical water supply business data records showing abnormal fluctuations (such as sudden changes in flow or pressure drop) within the slice, and binds them with the current employee operation and maintenance log data to establish a second-level data mapping. Through these two layers of mapping, the matching of historical water supply business data and structured employee operation and maintenance log data is achieved.

[0099] Based on structured employee operation and maintenance log data, task types are summarized and defined, and a task identification type tagging system is constructed. This tagging system includes several preset task categories, such as investigation and reporting rate, abnormal water usage analysis, leakage analysis, and system alarm handling.

[0100] Furthermore, a combination of rule matching and a large language model is used to label the structured employee operation and maintenance log data with task types. Rule matching is used to identify standardized task types with clear characteristics, while the large language model is used to identify complex or implicit task types, thereby improving the accuracy and coverage of task identification. The keyword rules are used to identify preset key fields in the employee operation and maintenance log data, such as "loss," "payment reminder," and "inspection"; the large language model is used to identify more complex scenarios in the employee operation and maintenance log data.

[0101] After completing the task type labeling, historical water supply business data is associated with the corresponding task identification labels to form a business data-task identification dataset. Each data record in the business data-task identification dataset includes business data feature information and its corresponding task type label, which are used for the subsequent training and optimization of the task identification model.

[0102] The data obtained after the above mapping and fusion constitutes the business data—task identification dataset. Specifically, this historical water supply business data not only includes the operational monitoring indicators of each water supply-related business subsystem, but also integrates external environmental data. The monitoring indicators mainly include: flow rate, pressure, equipment status, etc. in the zoning system; water fee collection rate, reporting rate, and number of meters in the meter reading system for the past seven days; water fee collection rate, bill anomalies, and overdue payment records in the revenue system; complaint content and processing time in the customer service system; and water supply pressure and water supply volume in the dispatching system. In addition, external related data such as weather information (such as rainfall, temperature, and humidity), holidays, and special event information are also introduced to enrich the dataset's accurate reflection of the actual complex business environment. The aforementioned equipment status data includes, but is not limited to, valve opening and closing status, water meter communication status, and equipment battery level.

[0103] Because the fused business data—the task recognition dataset—typically exhibits high-dimensional features, the system performs feature selection processing on the numerical data to avoid model overfitting and the "curse of dimensionality." Specifically, the system denotes the extracted numerical data items as follows: The system uses statistical methods to calculate the feature items of each numerical data type. and correlation score between The calculation formula is as follows: .

[0104] Furthermore, during the screening of numerical data items, the correlation index between each numerical data item is calculated. When the correlation between any two feature items is higher than a preset threshold, it is determined that the two feature items have a collinear relationship; and one of the features is deleted to reduce feature redundancy.

[0105] Furthermore, the preset threshold is 0.7-0.95, which can be adjusted according to historical data distribution, model training effect, or business scenario requirements.

[0106] After obtaining the retained core feature subset, the self-driven system needs to perform data normalization and multimodal feature fusion to eliminate dimensional differences and extract deep business semantics. Specifically, for numerical features (such as pipeline pressure, flow rate, and reporting rate), the system uses the Z-score normalization method to normalize the data, ensuring that each indicator follows a standard normal distribution with a mean of 0 and a standard deviation of 1, thereby obtaining a dimensionless numerical feature vector. Meanwhile, for text-based features (such as work order descriptions and customer complaint content) in water supply-related business subsystems, the self-driven system inputs them into a pre-trained BERT model, utilizes the bidirectional encoder of the task recognition model to obtain context-related high-dimensional semantic representations, and extracts text-based feature vectors. Subsequently, the self-driven system performs multimodal feature fusion, concatenating the two features: This forms a unified multimodal feature representation that includes cross-modal business information, which serves as the input tensor for subsequent deep neural network models.

[0107] In the construction and training phase of the task recognition model, the self-driven system performs nonlinear mapping training based on the aforementioned unified multimodal feature representation. In this embodiment, the deep neural network architecture is designed with two hidden layers, with 64 and 32 neurons respectively. This design effectively extracts the high-dimensional feature combination patterns while avoiding an overly bloated task recognition model. The hidden layers employ... After a non-linear mapping by the activation function, the data features are passed to the output layer, which consists of 16 neurons. The 16 neurons in the output layer precisely correspond to 16 pre-defined standardized task types in a water affairs operation and maintenance scenario. The trainable parameters of the deep neural network include the weight matrix W and bias vector b of each hidden layer and the output layer, which are iteratively updated based on the loss function using a backpropagation algorithm during training. The output layer... The activation function calculates the probability of each class, and the inverse of the log-likelihood function (i.e., the cross-entropy loss function) is used as the loss function to optimize the model parameters. Specifically, to address the overfitting problem caused by complex water resources data, the self-driven system introduces an L2 regularization mechanism during training. When calculating the overall loss function, a penalty term proportional to the sum of the squares of all weight parameters is added; during backpropagation to update the gradient, this mechanism applies decay pressure to larger weight parameters.

[0108] Once the task identification model is deployed and online, the self-driven system monitors water supply business data in real time. Once the task is identified and the result is output, the automated task dispatch mechanism is triggered.

[0109] Based on the core feature subset and the corresponding target output, the self-driven system completes the final iterative training of the task recognition model and deploys it into the self-driven system.

[0110] After the task identification model is deployed, the self-driven system enters the real-time self-driven operation phase. During the actual operation of the water affairs digital employee platform, the self-driven system acquires real-time water supply business data corresponding to the core feature subset, processes the data, and then inputs it into the task identification model for real-time abnormal task identification.

[0111] The table below shows the business data acquired by the system monitoring at a certain moment. The data is processed and then input into the self-driving system.

[0112] Table 1. Water supply business data acquired by the system monitoring at a certain moment.

[0113]

[0114] The task recognition model calculates the real-time feature vector of the current input and outputs a probability distribution sequence containing multiple preset task types and their corresponding predicted probabilities. For example, the output results are: [Leakage analysis: 0.71, Investigation and reporting rate: 0.02, Water demand prediction: 0.00, Overdue payment collection: 0.00, Equipment inspection: 0.02, Water quality early warning: 0.00, Pump station optimization: 0.01, Pipeline inspection: 0.03, Application review: 0.00, Water meter calibration: 0.05, Valve anomaly detection: 0.01, Pipeline maintenance: 0.03, Pressure boosting scheduling: 0.01, Complaint attribution: 0.04, Illegal water use investigation: 0.07, Energy consumption analysis: 0.00].

[0115] Furthermore, to ensure the accuracy of automated task dispatching, the self-driven system introduces a confidence assessment mechanism for the output prediction results. The self-driven system extracts the maximum predicted probability value from the probability distribution sequence and compares it with a preset confidence threshold. The confidence threshold ranges from 0.5 to 0.9 and can be calibrated based on the precision and recall performance of the task identification model on the validation set. In this embodiment, it is set to 0.6, but can be dynamically adjusted according to business needs in actual deployment. If the maximum predicted probability value is greater than or equal to 0.6 (e.g., "leakage analysis: 0.71" in the example above), the self-driven system determines that the confidence level of the identification result meets the standard, directly uses it as the final task type label, and triggers downstream water digital employees to execute the corresponding automated task instructions. If the maximum predicted probability value is less than 0.6, it indicates that the task identification model lacks sufficient understanding of the current complex working conditions, and no water digital employees are triggered. In this case, the self-driven system marks the current data record as "awaiting manual confirmation" and pushes it to manual intervention.

[0116] After outputting the task identification results, the self-driven system, based on a preset one-to-one mapping table of "task type - water digital employee," categorizes the task identification results output by the task identification model into the corresponding task type in the mapping table, and then automatically maps them to the corresponding water digital employee. The 16 preset task types in the mapping table correspond to 13 pre-deployed water digital employees; for example, "leakage analysis" corresponds to a leakage analysis digital employee, and "investigation and reporting rate" corresponds to a meter reading and inspection digital employee, forming a stable task dispatch channel. When the task identification model outputs a prediction result and passes the confidence level screening, the self-driven system automatically calls the corresponding water digital employee based on the mapping relationship, and transmits the real-time water supply business data that triggered the identification at the same time (i.e., the input data used by the task identification model for this identification) as the execution context to the water digital employee, which then takes over the handling process of abnormal events, achieving zero human intervention throughout the entire process from data monitoring to task dispatch.

[0117] Table 2 Task Type - Water Utility Digital Employee Mapping Table

[0118]

[0119] The platform currently has 13 digital water employees to support business personnel in handling daily tasks. For example... Figure 1 As shown, after the water digital employee completes the handling of an anomaly, the self-driven system automatically collects and organizes the information of the entire handling process, generating a structured task completion record. The record contains four core fields: execution action, which is the standardized business operation sequence actually performed by the water digital employee in this task, such as initiating work orders, pushing alarms, calling interfaces, etc.; processing time, which is the total process time from task dispatch to handling completion; task flow nodes, which is the flow path and node status of the task between different water digital employees or manual intervention links; and pipeline indicator recovery status, which is the recovery status of the key business indicators that triggered this task after the handling is completed, such as flow, pressure, reporting rate, non-revenue water volume, etc., relative to the time of the anomaly.

[0120] After the task completion records and manual intervention results are sent back to the self-driven system, they need to go through two stages: sample screening and training triggering, before they can be used for incremental updates of the task recognition model.

[0121] In the sample screening stage, the self-driven system performs the following screening rules on the returned samples: First, field integrity verification, that is, removing records with missing or abnormal four types of core fields; Second, treatment effectiveness verification, that is, judging whether the treatment was successful based on the "pipeline index recovery status" field, only retaining samples whose key pipeline indicators have recovered to the normal threshold range as positive samples, and marking samples with failed treatment or indicators not recovered as negative samples and storing them separately for subsequent misidentification analysis; Third, manual annotation priority, that is, assigning higher training weights to samples that have been manually corrected to strengthen the model's learning of manual correction logic.

[0122] In the training triggering phase, the self-driven system employs a dual triggering mechanism of "sample size threshold + timed detection": Incremental learning is triggered when the accumulated number of qualified incremental samples reaches a preset threshold (500 in this embodiment, adjustable from 100-2000 depending on the business scenario), or when a preset time period (7 days in this embodiment, adjustable based on the frequency of business changes) has elapsed since the last model update. Incremental learning uses a fine-tuning approach based on the currently deployed model. Based on the original network parameters, a limited number of backpropagation training rounds are performed on the incremental samples with a small learning rate. After completion, the performance of the updated task recognition model is evaluated using a validation set. If the accuracy of the new model on the validation set is not lower than that of the currently deployed model, it is replaced and deployed; otherwise, the original model is retained and the incremental samples are archived for future training.

[0123] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based method for self-driven implementation of tasks for digital water employees, characterized in that: The system acquires and cleans employee operation and maintenance log data, transforming it into structured employee operation and maintenance log data. Historical water supply business data is then obtained from this structured log data and converted into training samples with task labels. A task recognition dataset is constructed, and a task recognition model is trained based on a deep neural network. The self-driven system acquires real-time water supply business data from water-related subsystems to monitor the operational status of the water supply network and equipment. Using a pre-built task recognition model, it learns and reproduces human task recognition and decision-making logic in water supply operation and maintenance scenarios, autonomously determining tasks to be executed and generating recognition results. Tasks are then assigned to corresponding water management digital employees for execution. After the water management digital employee's processing is completed, the self-driven system receives structured task completion records from the water management digital employee. These records are added to the training set as new training samples, and the task recognition model is retrained and dynamically iterated, achieving automated closed-loop optimization of the water management digital employee self-driven processing flow.

2. The method for self-driven implementation of water affairs digital employee tasks based on deep learning according to claim 1, characterized in that... The specific steps are as follows: S1. Obtain employee operation and maintenance log data and clean the employee operation and maintenance log data; use a large language model to perform structured processing on the employee operation and maintenance log data, complete the extraction of key information and event extraction, obtain structured employee operation and maintenance log data, and obtain historical water supply business data through structured employee operation and maintenance log data. S2. Convert historical water supply business data into training samples with task labels, construct a task recognition dataset, and train the task recognition model based on a deep neural network. The self-driven system connects to the API interface of water supply-related business subsystems to obtain real-time water supply business data as input. It uses the preset task recognition model to monitor the operation status of the water supply network and equipment in real time. When an abnormal event is encountered, the built-in trained task recognition model is triggered to automatically identify the abnormal label, automatically map the abnormal label to an executable task instruction, and output the task recognition result. This realizes the automatic identification and result output of the task, thus forming a self-driven task recognition mechanism oriented towards business scenarios. S3. The task identification results output in step S2 are circulated and distributed. The task identification results are automatically mapped to the corresponding water digital employees through the preset task type-water digital employee mapping table. The water digital employees automatically take over the handling process of abnormal events, realizing a zero-human intervention transition from data monitoring to task assignment. The water digital employees perform corresponding standardized business handling operations. After the entire handling process is completed, the self-driven system automatically extracts the execution actions, processing time, task flow nodes, and pipeline indicator recovery status of the water digital employees, and generates a structured task completion record. S4. The structured task completion record generated in step S3 is used as verification feedback data and transmitted back to the self-driving system in real time along the optimization link. The self-driving system extracts the structured fields in the task completion record regarding the executed actions, processing time, task flow nodes, and pipeline indicator recovery status. The structured fields are used as incremental learning sample feedback signals to dynamically iterate and update the weight parameters of the task recognition model and the matching strategy of the business rule engine, thereby realizing the automated closed-loop optimization of the water supply self-driving process.

3. The method for self-driven implementation of water affairs digital employee tasks based on deep learning according to claim 2, characterized in that: Step S1 specifically includes: S11. First, obtain employee operation and maintenance log data within the historical time period; use a large language model for structured processing and cleaning, extract key information and events from the employee operation and maintenance log data, and transform it into structured employee operation and maintenance log data; obtain historical water supply business data through the structured employee operation and maintenance log data. S12. After acquiring the structured employee operation and maintenance log data, the self-driven system constructs multi-dimensional mapping rules and adopts a funnel-shaped fault-tolerant matching mechanism to accurately align the structured employee operation and maintenance log data with historical water supply business data. After mapping, the initial business data—the structured employee operation and maintenance log dataset—is obtained. S13. Based on the historical water supply business data obtained after mapping the structured employee operation and maintenance log data, extract multimodal features. The multimodal features include numerical features and text features. The numerical features are derived from the indicator data in the water supply-related business subsystems; the text features are derived from the text data in the water supply-related business subsystems; and feature selection processing is performed on the numerical features. S14. Based on the business data—structured employee operation and maintenance log dataset, construct a task identification type label system to form a business data—task identification dataset; S15. Process and clean the business data—task identification dataset, specifically including removing or filling missing values, normalization, and vectorization, to form the final task identification dataset.

4. The method for self-driven implementation of water affairs digital employee tasks based on deep learning according to claim 3, characterized in that... The multi-dimensional mapping rule in step S12 includes two layers: The first layer is strong rule exact matching: priority is given to extracting work order number, time of occurrence of abnormal situation, customer information and meter number from employee operation and maintenance log data, and performing cross-table primary key matching with the water supply related business subsystem database; The second layer is spatiotemporal slicing fuzzy matching: if the first layer does not match, the occurrence time extracted from the employee operation and maintenance log data is used as the benchmark, a time offset window is set, and a slice is constructed by combining the customer information entity in the employee operation and maintenance log data; within this slice range, historical water supply business data with abnormal fluctuations are retrieved and bound.

5. The method for self-driven implementation of water affairs digital employee tasks based on deep learning according to claim 3, characterized in that... The feature selection process for numerical features in step S13 specifically includes: calculating the linear correlation coefficient between any two numerical features, where the two numerical features are respectively... and When numerical features With numerical features When the absolute value of the correlation coefficient reaches or exceeds the preset threshold, it is determined that there is a high linear correlation between the two features, and one of the features is deleted according to the preset rules to reduce feature redundancy. The formula is as follows: in, Representation of features With features The linear correlation coefficient between them; and They represent the first , The feature in the first The values ​​for each sample; and These represent the average values ​​of the corresponding features within the statistical period; Indicates the number of samples.

6. The method for self-driven implementation of water affairs digital employee tasks based on deep learning according to claim 3, characterized in that: Step S14 specifically includes: using the successfully mapped historical water supply business data as the input feature variable X for model training; and simultaneously, parsing the behavioral patterns in the corresponding structured employee operation and maintenance log data and converting them into task type labels Y, thereby forming the business data-task identification dataset.

7. The method for self-driven implementation of water affairs digital employee tasks based on deep learning according to claim 3, characterized in that: Step S15 involves processing and cleaning the business data—task identification dataset—specifically including: For numerical data processing: remove duplicate data records, fill in or remove missing values, and standardize key business indicators to eliminate the impact of differences in units on model training. The standardization process uses the Z-score, for any original data point. Its corresponding Z-score is defined as: ； ； ； ； ； ； ； ； in, Represents the original numerical data. This represents the result of standardizing the original numerical data. The data is normalized by the mean and standard deviation to make the data follow a standard normal distribution with a mean of 0 and a standard deviation of 1. The average value of the sample data is calculated by summing all sample values ​​and dividing by the number of samples. The standard deviation indicates the dispersion of sample data and measures how much the data fluctuates relative to the mean. The larger the standard deviation, the more dispersed the data distribution. This represents the total number of data samples involved in the statistical calculation; For text data processing: a BERT-based method is used to extract deep semantic feature vectors from the text; specifically, for text data in water supply-related business subsystems, the text is input into a pre-trained BERT model, and its bidirectional encoder is used to obtain context-related high-dimensional semantic representations. Mapped to vector representation via the BERT model: ； Concatenate numerical features with textual features: ,in The numerical feature vector is constructed based on the mapped historical water supply business data. The numerical features are derived from the structured indicator data in the water supply-related business subsystem, including flow rate, pressure, water supply volume and reporting rate, and are obtained after normalization. These are text-based feature vectors obtained by semantically encoding business text data using a pre-trained language model; they are then concatenated to form a unified multimodal feature representation, which is used as input for subsequent task recognition models.

8. The method for self-driven implementation of water affairs digital employee tasks based on deep learning according to claim 7, characterized in that, The training of the task recognition model in step S2 specifically includes: S21. The unified multimodal feature representation formed by concatenation in step S15 is used as a high-dimensional input tensor and input to the neural network mapping function. middle; A mapping function representing the deep neural network model itself; The input data includes numerical features and text features; The set of parameters representing a neural network includes: the weight matrix for each layer. Bias vector ; S22. The deep neural network model has two hidden layers, and the hidden layers use... The activation function performs a non-linear mapping on the input numerical and textual features; the output layer uses... The activation function outputs the probability of each class; the inverse of the log-likelihood function is used as the loss function. , used to optimize model parameters; S23. An L2 regularization mechanism is introduced during the training process of deep neural networks. Specifically, when the self-driven system calculates the overall loss function of the model, it adds a regularization penalty term that is proportional to the sum of the squares of all weight parameters of the network on the basis of the original empirical error.

9. The method for self-driven implementation of water affairs digital employee tasks based on deep learning according to claim 2, characterized in that: In step S3, the process of transferring and distributing the task identification results specifically includes: the task identification model uses a variety of preset task identification results as output categories, and each task identification result corresponds to a water supply business scenario; the self-driving system is pre-configured with a task identification result-water digital employee mapping table, which maps the above-mentioned multiple task identification results to multiple water digital employees one-to-one. When the task recognition model outputs a task recognition result, and the probability of the task recognition result is not lower than a preset threshold, the self-driving system queries the task recognition result-water digital employee mapping table, directly determines the water digital employee corresponding to the task recognition result, and transmits the real-time water supply business data that triggered the recognition at the same time as the execution context to the water digital employee to complete the task distribution; when the probability of the output task recognition result is lower than the preset threshold, it is determined that the current water supply business does not require the processing of water digital employees, and no water digital employees are triggered.

Images