Intelligent distribution system based on deep learning and dynamic resource scheduling

Abstract

The invention discloses an intelligent distribution system based on deep learning and dynamic resource scheduling, and relates to the technical field of resource scheduling. Comprising the steps of 1, creating an intelligent distribution system, 2, integrating an NLP engine through a user request analysis module, carrying out intention recognition and semantic analysis on a user request, extracting a service type and related parameters, carrying out triple positioning by fusing GPS / base station / Wi-Fi, and obtaining user position information, and 3, constructing a service personnel digital portrait library through a dynamic resource management module, 4, calculating the geographic proximity of candidate service personnel and the user position through an intelligent matching module by adopting a Manhattan distance formula, training an LSTM neural network based on historical service data by utilizing a service quality prediction model, and obtaining the current position and the moving speed of the service personnel through the LSTM neural network; and step 5, obtaining user service evaluation data through a feedback module, triggering a service quality prediction model to perform retraining, and outputting a service quality prediction value in the confidence interval, generating a real-time comprehensive priority according to the geographic proximity, the service quality prediction value and the service emergency degree, and recommending services according to the comprehensive priority, and step 5, obtaining user service evaluation data through the feedback module, and triggering the service quality prediction model to perform retraining. And updating the service personnel digital portrait library by adopting an incremental learning strategy.

Description

Technical Field

[0001] The present invention discloses an intelligent allocation system based on deep learning and dynamic resource scheduling, and relates to the technical field of resource scheduling. Background Art

[0002] Existing online service platforms generally have the following defects:

[0003] Response delay: Manual customer service processing is inefficient, and queues are long during peak hours;

[0004] Imbalanced resource allocation: Service personnel distribution does not match user demand, resulting in idle services in some areas and overloaded services in others.

[0005] Fluctuations in service quality: There is a lack of a dynamic evaluation mechanism for service personnel’s historical performance, which makes it impossible to better maintain service quality. Summary of the Invention

[0006] In response to the problems of the prior art, the present invention provides an intelligent allocation system based on deep learning and dynamic resource scheduling, which is applied to intelligent task matching and service quality optimization in online service platforms.

[0007] The specific scheme proposed by the present invention is:

[0008] The present invention provides an intelligent allocation method based on deep learning and dynamic resource scheduling, comprising:

[0009] Step 1: Create an intelligent allocation system, which includes a user request parsing module, a dynamic resource management module, an intelligent matching module, and a feedback module.

[0010] Step 2: Integrate the NLP engine through the user request parsing module to perform intent recognition and semantic analysis on user requests, extract service types and related parameters, and integrate GPS / base station / Wi-Fi for triple positioning to obtain user location information.

[0011] Step 3: Build a digital portrait library of service personnel through the dynamic resource management module, record the service personnel's response speed, user ratings and skill tags through the digital portrait library of service personnel, and use the real-time interface to connect to the third-party map API to obtain the service personnel's current location and movement speed.

[0012] Step 4: The smart matching module uses the Manhattan distance formula to calculate the geographic proximity between the candidate service personnel and the user's location. The service quality prediction model is used to train the LSTM neural network based on historical service data, and the service quality prediction value within the confidence interval is output. The real-time comprehensive priority is generated based on the geographic proximity, service quality prediction value, and service urgency, and the service is recommended based on the comprehensive priority.

[0013] Step 5: Obtain user service evaluation data through the feedback module, trigger the retraining of the service quality prediction model, and use the incremental learning strategy to update the service personnel digital portrait library.

[0014] Furthermore, in step 2 of the intelligent allocation method based on deep learning and dynamic resource scheduling, the NLP engine is integrated through the user request parsing module, including:

[0015] For the integrated NLP engine, train and fine-tune the NLP model for the vertical domain:

[0016] Decompose the model tasks: perform intent recognition and text classification tasks; perform entity extraction and sequence labeling tasks,

[0017] Training configuration:

[0018] Pre-training weights: load the pre-training parameters of bert-base-chinese,

[0019] Configure the loss function: use cross entropy loss for intent classification and CRF loss for entity extraction.

[0020] Hyperparameters: learning rate 5e-5, batch size 16, and training epochs 3.

[0021] Furthermore, step 3 of the intelligent allocation method based on deep learning and dynamic resource scheduling is to construct a digital portrait library of service personnel, including:

[0022] Collect behavioral data, evaluation data, and qualification data related to service personnel,

[0023] The collected behavior data includes: work order timestamps and response behaviors. The work order timestamps include the user's order placement time t1, the service personnel's order acceptance time t2, the arrival time t3, and the service completion time t4. The response behaviors include the order acceptance status, cancellation records during the process, and timeout reasons.

[0024] The collected evaluation data includes: rating records, text evaluation, constraint quantity,

[0025] Rating records: user rating data of 1-5 stars for the service,

[0026] Text evaluation: data on service details described by users,

[0027] Constraint quantity: at least 50 historical reviews;

[0028] The content of the collected qualification data includes: qualification certificate data, historical work order label data,

[0029] Number of qualification certificates: professional qualification certificates, language certificates,

[0030] Historical work order tag data: automatically labeled skill types and skills with additional user notes.

[0031] Furthermore, step 4 of the intelligent allocation method based on deep learning and dynamic resource scheduling generates a real-time matching index based on geographic proximity, service quality prediction value, and service urgency, including:

[0032] Use the formula W = α × D + β × Q + γ × E to generate the comprehensive priority W.

[0033] D represents the geographical proximity between the candidate service personnel and the user, Q represents the predicted service quality value of the candidate service personnel, and E represents the urgency level of the current service task, which is a binary variable, E∈{0,1}, set 0=normal, 1=urgent, and γ is the urgency weight, which is the importance weight of each factor in the comprehensive priority, satisfying α+β+γ=1.

[0034] Furthermore, the intelligent allocation method based on deep learning and dynamic resource scheduling is characterized by triggering retraining of the service quality prediction model, including:

[0035] Determine whether the trigger conditions are met. If the trigger conditions are met, the service quality prediction model is retrained. The trigger conditions include:

[0036] Data volume threshold conditions:

[0037] The cumulative amount of newly added evaluation data is ≥ N to ensure that the model has enough new data to learn;

[0038] Data distribution change detection conditions:

[0039] Use statistical tests such as the KS test or model performance monitoring such as a sudden increase in the RMSE between the predicted and true scores to determine whether the new data is consistent with the historical data distribution;

[0040] Business rule triggering conditions: If the platform launches a new service type, such as high-end home appliance repair, the service standards will be adjusted, such as shortening the response time to 15 minutes, forcing retraining to adapt to the new rules.

[0041] The present invention also provides an intelligent allocation system based on deep learning and dynamic resource scheduling, which includes a user request parsing module, a dynamic resource management module, an intelligent matching module and a feedback module.

[0042] The user request parsing module integrates the NLP engine to perform intent recognition and semantic analysis on user requests, extract service types and related parameters, and integrate GPS / base station / Wi-Fi for triple positioning to obtain user location information.

[0043] The dynamic resource management module builds a digital portrait library of service personnel, records the service personnel's response speed, user ratings and skill tags through the digital portrait library, and uses the real-time interface to connect to the third-party map API to obtain the service personnel's current location and movement speed.

[0044] The intelligent matching module uses the Manhattan distance formula to calculate the geographic proximity between candidate service personnel and user locations, uses the service quality prediction model to train the LSTM neural network based on historical service data, outputs the service quality prediction value within the confidence interval, generates a real-time comprehensive priority based on geographic proximity, service quality prediction value, and service urgency, and recommends services based on the comprehensive priority.

[0045] The feedback module obtains user service evaluation data, triggers the retraining of the service quality prediction model, and adopts an incremental learning strategy to update the service personnel digital portrait library.

[0046] Furthermore, the user request parsing module of the intelligent allocation system based on deep learning and dynamic resource scheduling is integrated with an NLP engine, including:

[0047] For the integrated NLP engine, train and fine-tune the NLP model for the vertical domain:

[0048] Decompose the model tasks: perform intent recognition and text classification tasks; perform entity extraction and sequence labeling tasks,

[0049] Training configuration:

[0050] Pre-training weights: load the pre-training parameters of bert-base-chinese,

[0051] Configure the loss function: use cross entropy loss for intent classification and CRF loss for entity extraction.

[0052] Hyperparameters: learning rate 5e-5, batch size 16, and training epochs 3.

[0053] Furthermore, the dynamic resource management module of the intelligent allocation system based on deep learning and dynamic resource scheduling constructs a digital portrait library of service personnel, including:

[0054] Collect behavioral data, evaluation data, and qualification data related to service personnel,

[0055] The collected behavior data includes: work order timestamps and response behaviors. The work order timestamps include the user's order placement time t1, the service personnel's order acceptance time t2, the arrival time t3, and the service completion time t4. The response behaviors include the order acceptance status, cancellation records during the process, and timeout reasons.

[0056] The collected evaluation data includes: rating records, text evaluation, constraint quantity,

[0057] Rating records: user rating data of 1-5 stars for the service,

[0058] Text evaluation: data on service details described by users,

[0059] Constraint quantity: at least 50 historical reviews;

[0060] The content of the collected qualification data includes: qualification certificate data, historical work order label data,

[0061] Number of qualification certificates: professional qualification certificates, language certificates,

[0062] Historical work order tag data: automatically labeled skill types and skills with additional user notes.

[0063] Furthermore, step 4 of the intelligent allocation system based on deep learning and dynamic resource scheduling generates a real-time matching index based on geographic proximity, service quality prediction value, and service urgency, including:

[0064] Use the formula W = α × D + β × Q + γ × E to generate the comprehensive priority W.

[0065] D represents the geographical proximity between the candidate service personnel and the user, Q represents the predicted service quality value of the candidate service personnel, and E represents the urgency level of the current service task, which is a binary variable, E∈{0,1}, set 0=normal, 1=urgent, and γ is the urgency weight, which is the importance weight of each factor in the comprehensive priority, satisfying α+β+γ=1.

[0066] Furthermore, the feedback module of the intelligent allocation system based on deep learning and dynamic resource scheduling triggers the retraining of the service quality prediction model, including:

[0067] Determine whether the trigger conditions are met. If the trigger conditions are met, the service quality prediction model is retrained. The trigger conditions include:

[0068] Data volume threshold conditions:

[0069] The cumulative amount of newly added evaluation data is ≥ N to ensure that the model has enough new data to learn;

[0070] Data distribution change detection conditions:

[0071] Use statistical tests such as the KS test or model performance monitoring such as a sudden increase in the RMSE between the predicted and true scores to determine whether the new data is consistent with the historical data distribution;

[0072] Business rule triggering conditions: If the platform launches a new service type, such as high-end home appliance repair, the service standards will be adjusted, such as shortening the response time to 15 minutes, forcing retraining to adapt to the new rules.

[0073] The benefits of the present invention are:

[0074] Improved user service experience: Accurate intent recognition and demand matching, parsing user requests through the NLP engine, quickly identifying service types such as repair, housekeeping, and delivery, and key parameters such as time, location, and device model, reducing the tedious manual input of users; multi-source positioning improves location accuracy, integrating GPS, base station, and Wi-Fi positioning to solve the accuracy issues of single positioning methods, such as weak indoor GPS signals, and quickly obtain the user's accurate location;

[0075] Optimized service resource allocation: Dynamic resource management is implemented, recording service personnel's response speed, user ratings, and skill tags to achieve dynamic resource scheduling; the real-time interface connects to the third-party map API to obtain the service personnel's current location and movement speed, and dynamically adjust the scheduling strategy;

[0076] Improving dispatch efficiency through intelligent matching: Manhattan distance is used to quickly identify nearby service personnel, reducing their commuting time. Compared to Euclidean distance, Manhattan distance is more suitable for urban gridded roads, with high computational efficiency and more realistic results. Using service quality predictions avoids the "nearby but unprofessional" problem caused by relying solely on geographic proximity, balancing efficiency and quality.

[0077] Combine geographic proximity, service quality predictions, and service urgency to generate real-time comprehensive priorities for better matching services.

[0078] User service evaluation data is used to trigger incremental learning of the LSTM model and improve prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0080] Figure 1 It is a schematic flow chart of the method of the present invention. DETAILED DESCRIPTION

[0081] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0082] Example 1

[0083] The present invention provides an intelligent allocation method based on deep learning and dynamic resource scheduling, comprising:

[0084] Step 1: Create an intelligent allocation system, which includes a user request parsing module, a dynamic resource management module, an intelligent matching module, and a feedback module.

[0085] Step 2: Integrate the NLP engine through the user request parsing module to perform intent recognition and semantic analysis on user requests, extract service types and related parameters, and integrate GPS / base station / Wi-Fi for triple positioning to obtain user location information. The integration of the NLP engine through the user request parsing module includes:

[0086] For the integrated NLP engine, train and fine-tune the NLP model for the vertical domain:

[0087] Decompose the model tasks: perform intent recognition and text classification tasks; perform entity extraction and sequence labeling tasks,

[0088] Training configuration:

[0089] Pre-training weights: load the pre-training parameters of bert-base-chinese,

[0090] Configure the loss function: use cross entropy loss for intent classification and CRF loss for entity extraction.

[0091] Hyperparameters: learning rate 5e-5, batch size 16, and training epochs 3.

[0092] Step 3: Build a digital portrait library of service personnel through the dynamic resource management module, record the service personnel's response speed, user ratings and skill tags through the digital portrait library of service personnel, and use the real-time interface to connect to the third-party map API to obtain the service personnel's current location and movement speed.

[0093] The construction of a digital portrait library of service personnel includes:

[0094] Collect behavioral data, evaluation data, and qualification data related to service personnel,

[0095] The collected behavior data includes: work order timestamps and response behaviors. The work order timestamps include the user's order placement time t1, the service personnel's order acceptance time t2, the arrival time t3, and the service completion time t4. The response behaviors include the order acceptance status, cancellation records during the process, and timeout reasons.

[0096] The collected evaluation data includes: rating records, text evaluation, constraint quantity,

[0097] Rating records: user rating data of 1-5 stars for the service,

[0098] Text evaluation: data on service details described by users,

[0099] Constraint quantity: at least 50 historical reviews;

[0100] The content of the collected qualification data includes: qualification certificate data, historical work order label data,

[0101] Number of qualification certificates: professional qualification certificates, language certificates,

[0102] Historical work order tag data: automatically labeled skill types and skills with additional user notes.

[0103] Step 4: The smart matching module uses the Manhattan distance formula to calculate the geographic proximity between the candidate service personnel and the user's location. The service quality prediction model is used to train the LSTM neural network based on historical service data. The service quality prediction value within the confidence interval is output. A real-time comprehensive priority is generated based on the geographic proximity, service quality prediction value, and service urgency. Services are recommended based on the comprehensive priority. The real-time matching index generated based on geographic proximity, service quality prediction value, and service urgency includes:

[0104] Use the formula W = α × D + β × Q + γ × E to generate the comprehensive priority W.

[0105] D represents the geographical proximity between the candidate service personnel and the user, Q represents the predicted service quality value of the candidate service personnel, and E represents the urgency level of the current service task, which is a binary variable, E∈{0,1}, set 0=normal, 1=urgent, and γ is the urgency weight, which is the importance weight of each factor in the comprehensive priority, satisfying α+β+γ=1.

[0106] Step 5: Obtain user service evaluation data through the feedback module, trigger the retraining of the service quality prediction model, and use the incremental learning strategy to update the service personnel digital portrait library.

[0107] Triggering the retraining of the service quality prediction model may include:

[0108] Determine whether the trigger conditions are met, and if so, trigger the retraining of the service quality prediction model.

[0109] Trigger conditions include:

[0110] Data volume threshold conditions:

[0111] The cumulative amount of newly added evaluation data is ≥ N, for example, N = 1000, to ensure that the model has enough new data to learn;

[0112] Data distribution change detection condition:

[0113] Method: Determine whether the new data is consistent with the historical data distribution through statistical tests such as KS test or model performance monitoring such as sudden increase of RMSE of predicted score and true score.

[0114] Business rule triggering condition: If a new service type such as high-end home appliance maintenance is put on the platform, adjust the service standard such as shorten the response time to 15 minutes, and trigger retraining to adapt to the new rules.

[0115] Embodiment 2

[0116] The application also provides an intelligent allocation system based on deep learning and dynamic resource scheduling, which comprises a user request analysis module, a dynamic resource management module, an intelligent matching module and a feedback module,

[0117] The user request analysis module integrates an NLP engine, performs intent recognition and semantic analysis on user requests, extracts service types and related parameters, and performs triple positioning by combining GPS / base station / Wi-Fi to obtain user location information,

[0118] The dynamic resource management module constructs a service personnel digital portrait library, records service personnel response speed, user ratings and skill labels through the service personnel digital portrait library, and uses a real-time interface to connect a third-party map API to obtain the current location and moving speed of the service personnel,

[0119] The intelligent matching module uses the Manhattan distance formula to calculate the geographical proximity of candidate service personnel and user location, uses a service quality prediction model to train an LSTM neural network based on historical service data, outputs service quality prediction values within the confidence interval, generates real-time comprehensive priority based on geographical proximity, service quality prediction values and service urgency, and recommends services according to the comprehensive priority,

[0120] The feedback module obtains user service evaluation data, triggers service quality prediction model retraining, and uses incremental learning strategy to update the service personnel digital portrait library.

[0121] The information interaction and execution process between the modules in the above device are based on the same concept as the method embodiments of the application, and the specific content can be referred to the description in the method embodiments of the application, which will not be repeated here.

[0122] Similarly, the device of the present invention improves the user service experience: accurate intent recognition and demand matching, parsing user requests through the NLP engine, quickly identifying service types such as maintenance, housekeeping, and delivery, and key parameters such as time, location, and equipment model, reducing the tedious manual input of users; multi-source positioning improves location accuracy, integrating GPS, base station, and Wi-Fi triple positioning to solve the accuracy problems of single positioning methods, such as weak indoor GPS signals, and quickly obtain the user's accurate location;

[0123] Optimized service resource allocation: Dynamic resource management is implemented, recording service personnel's response speed, user ratings, and skill tags to achieve dynamic resource scheduling; the real-time interface connects to the third-party map API to obtain the service personnel's current location and movement speed, and dynamically adjust the scheduling strategy;

[0124] Improving dispatch efficiency through intelligent matching: Manhattan distance is used to quickly identify nearby service personnel, reducing their commuting time. Compared to Euclidean distance, Manhattan distance is more suitable for urban gridded roads, with high computational efficiency and more realistic results. Using service quality predictions avoids the "nearby but unprofessional" problem caused by relying solely on geographic proximity, balancing efficiency and quality.

[0125] Combine geographic proximity, service quality predictions, and service urgency to generate real-time comprehensive priorities for better matching services.

[0126] User service evaluation data is used to trigger incremental learning of the LSTM model and improve prediction accuracy.

[0127] It should be noted that not all steps and modules in the above-mentioned processes and device structures are required, and certain steps or modules can be omitted according to actual needs. The execution order of each step is not fixed and can be adjusted as needed. The system structure described in the above-mentioned embodiments can be a physical structure or a logical structure, that is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or may be implemented by certain components in multiple independent devices.

[0128] The above embodiments are merely preferred embodiments for the purpose of fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are within the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

Claims

1. An intelligent allocation method based on deep learning and dynamic resource scheduling, characterized by include: Step 1: Create an intelligent allocation system, which includes a user request parsing module, a dynamic resource management module, an intelligent matching module, and a feedback module. Step 2: Integrate the NLP engine through the user request parsing module to perform intent recognition and semantic analysis on user requests, extract service types and related parameters, and integrate GPS / base station / Wi-Fi for triple positioning to obtain user location information. Step 3: Build a digital portrait library of service personnel through the dynamic resource management module, record the service personnel's response speed, user ratings and skill tags through the digital portrait library of service personnel, and use the real-time interface to connect to the third-party map API to obtain the service personnel's current location and movement speed. Step 4: The smart matching module uses the Manhattan distance formula to calculate the geographic proximity between the candidate service personnel and the user's location. The service quality prediction model is used to train the LSTM neural network based on historical service data, and the service quality prediction value within the confidence interval is output. The real-time comprehensive priority is generated based on the geographic proximity, service quality prediction value, and service urgency, and the service is recommended based on the comprehensive priority. Step 5: Obtain user service evaluation data through the feedback module, trigger the retraining of the service quality prediction model, and use the incremental learning strategy to update the service personnel digital portrait library.

2. The intelligent allocation method based on deep learning and dynamic resource scheduling according to claim 1 is characterized by In step 2, the NLP engine is integrated through the user request parsing module, including: For the integrated NLP engine, train and fine-tune the NLP model for the vertical domain: Decompose the model tasks: perform intent recognition and text classification tasks; perform entity extraction and sequence labeling tasks, Training configuration: Pre-training weights: load the pre-training parameters of bert-base-chinese, Configure the loss function: use cross entropy loss for intent classification and CRF loss for entity extraction. Hyperparameters: learning rate 5e-5, batch size 16, and training epochs 3.

3. The intelligent allocation method based on deep learning and dynamic resource scheduling according to claim 1, Its characteristic is that the service personnel digital portrait library is constructed in step 3, including: Collect behavioral data, evaluation data, and qualification data related to service personnel, The collected behavior data includes: work order timestamps and response behaviors. The work order timestamps include the user's order placement time t1, the service personnel's order acceptance time t2, the arrival time t3, and the service completion time t4. The response behaviors include the order acceptance status, cancellation records during the process, and timeout reasons. The collected evaluation data includes: rating records, text evaluation, constraint quantity, Rating records: user rating data of 1-5 stars for the service, Text evaluation: data on service details described by users, Constraint quantity: at least 50 historical reviews; The content of the collected qualification data includes: qualification certificate data, historical work order label data, Number of qualification certificates: professional qualification certificates, language certificates, Historical work order tag data: automatically labeled skill types and skills with additional user notes.

4. The intelligent allocation method based on deep learning and dynamic resource scheduling according to claim 1 is characterized by: In step 4, a real-time matching index is generated based on geographic proximity, service quality prediction value, and service urgency, including: Use the formula W = α × D + β × Q + γ × E to generate the comprehensive priority W. D represents the geographical proximity between the candidate service personnel and the user, Q represents the predicted service quality value of the candidate service personnel, and E represents the urgency level of the current service task, which is a binary variable, E∈{0,1}, set 0=normal, 1=urgent, and γ is the urgency weight, which is the importance weight of each factor in the comprehensive priority, satisfying α+β+γ=1.

5. The intelligent allocation method based on deep learning and dynamic resource scheduling according to claim 1 is characterized by: Step 5 triggers retraining of the service quality prediction model, including: Determine whether the trigger conditions are met. If the trigger conditions are met, the service quality prediction model is retrained. The trigger conditions include: Data volume threshold conditions: The cumulative amount of newly added evaluation data is ≥ N to ensure that the model has enough new data to learn; Data distribution change detection conditions: Use statistical tests such as the KS test or model performance monitoring such as a sudden increase in the RMSE between the predicted and true scores to determine whether the new data is consistent with the historical data distribution; Business rule triggering conditions: If the platform launches a new service type, such as high-end home appliance repair, the service standards will be adjusted, such as shortening the response time to 15 minutes, forcing retraining to adapt to the new rules.

6. An intelligent allocation system based on deep learning and dynamic resource scheduling, characterized by The intelligent allocation system includes a user request parsing module, a dynamic resource management module, an intelligent matching module and a feedback module. The user request parsing module integrates the NLP engine to perform intent recognition and semantic analysis on user requests, extract service types and related parameters, and integrate GPS / base station / Wi-Fi for triple positioning to obtain user location information. The dynamic resource management module builds a digital portrait library of service personnel, records the service personnel's response speed, user ratings and skill tags through the digital portrait library, and uses the real-time interface to connect to the third-party map API to obtain the service personnel's current location and movement speed. The intelligent matching module uses the Manhattan distance formula to calculate the geographic proximity between candidate service personnel and user locations, uses the service quality prediction model to train the LSTM neural network based on historical service data, outputs the service quality prediction value within the confidence interval, generates a real-time comprehensive priority based on geographic proximity, service quality prediction value, and service urgency, and recommends services based on the comprehensive priority. The feedback module obtains user service evaluation data, triggers the retraining of the service quality prediction model, and adopts an incremental learning strategy to update the service personnel digital portrait library.

7. The intelligent allocation system based on deep learning and dynamic resource scheduling according to claim 6 is characterized by: The user request parsing module integrates the NLP engine, including: For the integrated NLP engine, train and fine-tune the NLP model for the vertical domain: Decompose the model tasks: perform intent recognition and text classification tasks; perform entity extraction and sequence labeling tasks, Training configuration: Pre-training weights: load the pre-training parameters of bert-base-chinese, Configure the loss function: use cross entropy loss for intent classification and CRF loss for entity extraction. Hyperparameters: learning rate 5e-5, batch size 16, and training epochs 3.

8. The intelligent allocation system based on deep learning and dynamic resource scheduling according to claim 6 is characterized by dynamic The resource management module builds a digital portrait library of service personnel, including: Collect behavioral data, evaluation data, and qualification data related to service personnel, The collected behavior data includes: work order timestamps and response behaviors. The work order timestamps include the user's order placement time t1, the service personnel's order acceptance time t2, the arrival time t3, and the service completion time t4. The response behaviors include the order acceptance status, cancellation records during the process, and timeout reasons. The collected evaluation data includes: rating records, text evaluation, constraint quantity, Rating records: user rating data of 1-5 stars for the service, Text evaluation: data on service details described by users, Constraint quantity: at least 50 historical reviews; The content of the collected qualification data includes: qualification certificate data, historical work order label data, Number of qualification certificates: professional qualification certificates, language certificates, Historical work order tag data: automatically labeled skill types and skills with additional user notes.

9. The intelligent allocation system based on deep learning and dynamic resource scheduling according to claim 6 is characterized by: In step 4, a real-time matching index is generated based on geographic proximity, service quality prediction value, and service urgency, including: Use the formula W = α × D + β × Q + γ × E to generate the comprehensive priority W. D represents the geographical proximity between the candidate service personnel and the user, Q represents the predicted service quality value of the candidate service personnel, and E represents the urgency level of the current service task, which is a binary variable, E∈{0,1}, set 0=normal, 1=urgent, and γ is the urgency weight, which is the importance weight of each factor in the comprehensive priority, satisfying α+β+γ=1.

10. The intelligent allocation system based on deep learning and dynamic resource scheduling according to claim 1 is characterized by The feedback module triggers retraining of the service quality prediction model, including: Determine whether the trigger conditions are met. If the trigger conditions are met, the service quality prediction model is retrained. The trigger conditions include: Data volume threshold conditions: The cumulative amount of newly added evaluation data is ≥ N to ensure that the model has enough new data to learn; Data distribution change detection conditions: Use statistical tests such as the KS test or model performance monitoring such as a sudden increase in the RMSE between the predicted and true scores to determine whether the new data is consistent with the historical data distribution; Business rule triggering conditions: If the platform launches a new service type, such as high-end home appliance repair, the service standards will be adjusted, such as shortening the response time to 15 minutes, forcing retraining to adapt to the new rules.

Images