Social governance work order dispatching method based on emergency degree assessment, medium and equipment

Abstract

The application provides a social governance work order dispatching method based on emergency degree evaluation, a medium and equipment, and belongs to the technical field of social governance information. Text data and image data in a social governance work order are collected, a text encoder and an image encoder are called, and the text data and the image data are respectively converted into a text feature vector and an image feature vector; consistency detection is performed on the text feature vector and the image feature vector of the same work order; the emotion intensity is calculated based on the text feature vector, and a corresponding work order event danger degree and time coding are combined to splice and generate a vector; the spliced and generated vector is input into a trained neural network model to generate a work order emergency index, the work order emergency index is compared with a preset threshold value to divide the emergency degree of the work order; and the work order is dispatched according to the emergency degree of the work order. By introducing a multi-layer perceptron network of cross features, residual connection and Swish activation, the disposal timeliness of the social governance work order is significantly improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of social governance information, and particularly relates to a social governance work order dispatching method based on emergency degree evaluation, a medium and equipment. BACKGROUND

[0002] Government service hotline and city operation monitoring center need to handle a large number of social governance work orders every day. These work orders usually contain multi-modal information such as text and images, and involve multiple fields such as public safety, traffic order, municipal facilities and environmental sanitation. At present, the work order grading and dispatching process mainly relies on manual experience combined with static rules, and determines the emergency degree through a keyword or business prompt word dictionary, which is difficult to analyze the risks implied in colloquial expressions and images. In addition, manual review and telephone transfer and other links, the average time of a single work order preliminary examination reaches 8-10 minutes, and a large number of backlog may occur during peak periods, resulting in delayed response and low disposal efficiency. The existing technology is difficult to adapt to high concurrency scenarios and sudden emergencies, and there is no method to simultaneously fuse emotion features, risk degree features and time information in a unified vector space to realize real-time emergency index generation and dispatching of social governance work orders. SUMMARY

[0003] The present application provides a social governance work order dispatching method based on emergency degree evaluation, a medium and equipment, which significantly improves the disposal efficiency of social governance work orders by introducing a multi-layer perceptron network with cross features, residual connection and Swish activation.

[0004] To achieve the above purpose, the application adopts the following technical solutions:

[0005] In a first aspect, the application provides a social governance work order dispatching method based on emergency degree evaluation, comprising the following steps:

[0006] Step 1: Collecting text data and image data in social governance work orders, calling a text encoder and an image encoder, and converting the text data and image data into text feature vectors and image feature vectors respectively;

[0007] Step 2: Consistency detection of the text feature vectors and image feature vectors of the same work order, and dispatching the work orders that do not pass the consistency detection to a manual review queue;

[0008] Step 3: For the work orders that pass the consistency detection, calculating the emotion intensity based on the text feature vectors, and combining the corresponding work order event risk degree and time coding to splice and generate a vector;

[0009] Step 4: Construct and train a neural network model including a multilayer perceptron with residual connections. Input the concatenated vector into the trained neural network model to generate a work order urgency index. By comparing the work order urgency index with a preset threshold, the urgency level of the work order is classified.

[0010] Step 5: Dispatch work orders according to their urgency.

[0011] Optionally, in step 1, the text encoder includes an embedding layer, a bidirectional LSTM, a global average pooling layer, and a fully connected layer connected in sequence, with a ReLU activation function after the fully connected layer; the text encoder finally outputs a text feature vector with a vector dimension of 256.

[0012] Optionally, in step 1, the image encoder includes a series of sequentially connected layers: a 32-channel 3*3 convolutional layer, a 2*2 max pooling layer, a 64-channel 3*3 convolutional layer, a 2*2 max pooling layer, a 128-channel 3*3 convolutional layer, a 2*2 max pooling layer, a feature aggregation layer, an average pooling layer, and a fully connected layer. Each 2*2 max pooling layer is followed by a ReLU activation function. The image encoder finally outputs an image feature vector with a dimension of 256.

[0013] Optionally, in step 2, the consistency detection specifically includes:

[0014] Calculate the cosine similarity between text feature vectors and image feature vectors. Cosine similarity With set threshold If a comparison is made, If the consistency check is passed, then it is considered to have passed. If so, it is considered that the consistency test has not been passed;

[0015] Among them, through Slide to update subscript Indicates the time step. and They represent and time step Threshold and cosine similarity at time, The parameters are set to allow the threshold to gradually converge to the optimal discrimination point.

[0016] Optionally, in step 3, the calculation process of the emotion intensity is as follows: input the text feature vector into the Senta sentiment analysis model, output the probability vectors of different emotions, and form the emotion intensity vector. ,in, express The probability vector corresponding to each emotion A scalar representing the intensity of emotion. represents the dimension of ;

[0017] The calculation process of the risk degree is: creating a one-hot vector, writing the weight of the predefined ticket event type in the corresponding dimension of the one-hot vector, setting the remaining dimensions to 0, and generating a risk degree vector wherein represents a weight corresponding to the event type;

[0018] The calculation process of the time encoding is: dividing 24h into equal time periods, and performing One-Hot encoding according to the ticket event occurrence time to obtain a dimensional vector, then multiplying each dimension according to the predefined weight of each time period, and compressing through a Sigmoid function to obtain a time encoding vector ;

[0019] Finally, the vector is spliced.

[0020] Optionally, in step 4, the process of generating the ticket emergency index by the neural network model is:

[0021] Step 4.1: normalizing the vector to obtain a normalized feature vector ; generating attention weights , re-weighting by using to obtain a feature vector ;

[0022] ;

[0023]

[0024] In the formula, represents the Hadamard product, is the weight corresponding to the th feature component in , is an intermediate vector obtained after a linear transformation is performed on , , , and are respectively a learning weight matrix and a bias term, and the subscript is a dimension index, and are respectively the th and the th exp represents exponential function;

[0025] Step 4.2: Perform latent vector decomposition on the feature vector to obtain cross vector :

[0026]

[0027] wherein, is a low-rank representation of the second-order interaction feature, is a second-order cross weight obtained by inner product of latent vectors in the factorization machine, representing the interaction intensity between the th feature and the th feature, and are the th and the th feature components of the feature vector ;

[0028] concatenate and to generate vector , representing horizontal concatenation;

[0029] Step 4.3: Input to a multi-layer perceptron with residual connection, the multi-layer perceptron contains hidden layers, each hidden layer is followed by a Swish activation function, and the network structure is represented as:

[0030] ;

[0031] wherein, , is the output vector of the th hidden layer, and are the weight vector and the bias term of the th layer, respectively;

[0032] The multi-layer perceptron finally outputs the ticket urgency index , wherein is a Sigmoid function, and are the weight matrix and the bias vector of the output layer.

[0033] Optionally, in step 4, the neural network model constructs a joint loss function with ticket label and ticket handling duration , and trains the model using a gradient optimizer, and the joint loss function is represented as:

[0034] ;

[0035] wherein, the model training batch size is , denotes the ticket urgency index prediction value output by the model for the i th ticket sample, denotes the ticket urgency index label value of the i th ticket sample, , denotes the actual response time length of the i th ticket sample; , denotes the actual response time length of the i th ticket sample; denotes the handling time length of the i th ticket sample, which is the difference between the ticket creation time and the ticket handling completion time, denotes the handling time length prediction value output by the model for the i th ticket sample. In step 5, the assigning according to the urgency of the ticket is specifically: According to the urgency of the ticket, the priority of the ticket is divided, and the tickets are sequentially assigned according to the priority of the ticket.

[0036] In the second aspect, the present application provides a computer readable storage medium, which stores a computer program, wherein the computer program enables a computer to execute the social governance ticket assigning method based on urgency evaluation as described in the first aspect.

[0037] In the third aspect, the present application provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the social governance ticket assigning method based on urgency evaluation as described in the first aspect is realized.

[0038] The present application has the following beneficial effects:

[0039] (1) In the neural network model for generating the urgency index, the present application adopts the combination design of factorization machine, residual connection and Swish MLP, which effectively solves the problems of large cross-feature dimension, label sparsity and gradient vanishing in the social governance ticket.

[0040] Swish activation function retains non-zero gradient in the negative interval, which, in combination with residual connection, can significantly alleviate gradient attenuation, speed up convergence and smooth the prediction curve; the residual path also retains first-order semantics after splicing low-rank cross-features and original features, reducing the overfitting risk in small sample scenarios.

[0041]

[0042] ​​(2) The joint loss function fusing the emergency index and the disposal time length is adopted, on one hand, the weight coefficient can be flexibly adjusted according to business requirements, the actual response time length is introduced into the optimization target, the model can learn the emergency degree and the work order disposal load at the same time, the deficiency of only relying on single emergency index regression and being difficult to measure the disposal cost is made up, and the explainability is improved; on the other hand, the disposal time length prediction provides an additional gradient signal for the network, and the overfitting caused by label sparsity can be relieved. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 It is a structural schematic diagram of a text encoder in the application.

[0044] Figure 2 It is a structural schematic diagram of an image encoder in the application.

[0045] Figure 3 It is a flow chart of work order emergency index generation in the application.

[0046] Figure 4 It is a network structure schematic diagram of a multilayer perceptron introducing residual connection and Swish activation function in the application.

[0047] Figure 5 It is a work order emergency index generation, order dispatching and parameter updating process in the application.

[0048] Figure 6 It is a social governance work order processing system interface diagram in the application.

[0049] Figure 7 It is a social governance work order database system interface diagram in the application. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application.

[0051] In an embodiment, the application provides a social governance work order dispatching method based on emergency degree evaluation, including the following steps:

[0052] Step 1: collect text data and image data in the social governance work order, call the text encoder and the image encoder, and convert the text data and the image data into a text feature vector and an image feature vector respectively.

[0053] The social governance work order data mainly includes two types: the text and image uploaded by citizens through the complaint platform, or the text and image uploaded by community grid members, which constitutes the available modal data of the social governance work order.

[0054] For subsequent model training, this embodiment collected 11,376 historical social governance work orders. The fields of the historical work orders include work order ID, event occurrence time, event occurrence location, event description, event image, event type (16 categories), work order dispatch time, work order handling department, handling completion time, and user satisfaction. Each work order was converted into a minimal modal combination of text data and image data.

[0055] The raw modal data is converted into feature vectors of uniform dimension, and the text data is input into the text encoder. Its basic structure is as follows Figure 1 As shown in the figure, h represents the overall text representation obtained by average pooling. t This indicates that the hidden state vector and text feature vector are output at time t. Input image data into the image encoder. Its basic structure is as follows Figure 2 As shown, the output image feature vector .

[0056] Step 2: Perform consistency checks on the text feature vectors and image feature vectors of the same work order. Work orders that fail the consistency check are assigned to the manual review queue.

[0057] To detect the consistency between text modal data and image modal data, the text feature vectors of the same work order are analyzed. With image feature vectors Positive samples for modal data consistency detection are labeled as 1; simultaneously, work order IDs are randomly matched to construct negative samples (image and text mismatch), which are labeled as 0; the ratio of positive to negative samples is set to 1:1; the network structure uses two fully connected layers and ReLU activation, and cosine similarity is calculated after output; binary cross-entropy loss is used as the loss function, the dataset is divided into training and validation sets in an 8:2 ratio, the optimizer is Adam, the initial learning rate is set to 0.001, and early stopping is achieved if the validation set loss does not decrease after 10 epochs.

[0058] Cosine similarity Initial threshold setting It is 0.65; if If the consistency check is passed, it is determined that the multimodal content is consistent, and vector C is constructed accordingly; if The work order is marked as "image and text mismatch" and sent for manual review. It is automatically assigned to the manual review queue. Based on the review result (1 = consistent, 0 = inconsistent), it is approved. Slide to update subscript Indicates the time step, typically 1 day or 1 week, set. This allows the threshold to gradually converge to the optimal distinction point as the work order business data accumulates.

[0059] Step 3: For the work order detected by consistency, calculate the emotional intensity based on the text feature vector, and splice the generated vector combined with the danger level and time coding of the corresponding work order event.

[0060] In this embodiment, first, the Calculate the emotional intensity E based on the Baidu sentiment analysis model Senta fine-tuning, output the probability vector of different emotions (such as anger, disgust, fear, sadness, joy, surprise, anxiety, calm), output The sum of the first 8 dimensions of different emotion probabilities is 1, and the 9th dimension is the emotion intensity scalar .

[0061] Then, according to the selected event type of the work order, calculate the event danger level according to the pre-defined weight of 16 types of events. The weight of the event type is: 0.95 for fighting with weapons, 0.70 for crowd gathering, 0.60 for vehicle parking violation, 0.70 for vehicle rear-end collision, 0.90 for fire, 0.95 for gas safety, 0.45 for road occupation, 0.80 for building falling, 0.95 for wall falling, 0.65 for cable exposure, 0.55 for manhole cover missing, 0.85 for road collapse, 0.65 for water pipe burst, 0.45 for garbage accumulation, and 0.4 for other abnormalities. According to the selected event type, create a one-hot vector and write the weight in the corresponding dimension , and set the remaining dimensions to 0 to generate the event danger level vector .

[0062] Next, calculate the time coding , divide 24h into 8 equal time periods according to the event occurrence time , and do One-Hot coding to get an 8-dimensional base vector; then multiply each dimension according to the time period emergency weight table (night safety segment weight is high, daytime weight is relatively low), and compress it through the Sigmoid function to get .

[0063] Finally, output the vector .

[0064] Step 4: Build and train a neural network model including a multi-layer perceptron (MLP) with residual connection, input the spliced generated vector into the trained neural network model, generate the work order emergency index, and divide the emergency degree of the work order by comparing the work order emergency index with the preset threshold. The structure of the neural network model is shown in Figure 3 and Figure 4 .

[0065] First, generate the attention weight , use to process the normalized features The reweighting is performed to obtain :

[0066] ;

[0067] wherein, and are attention weight generation layer learning weight matrices and bias terms; a normalized weight vector is output by a function, representing the importance of each feature dimension, ensuring that each component ∈ (0, 1) and the sum is 1, so that the network focuses on key feature dimensions, and further, the weight of the th feature component of can be calculated as follows:

[0068]

[0069] wherein, is an intermediate vector obtained by performing a linear transformation on the input feature vector , the subscript is a dimension index, and exp represents an exponential function;

[0070] to obtain the feature vector :

[0071] ;

[0072] wherein, is multiplied by the original feature vector by Hadamard product to achieve feature reweighting, and parameters and are updated by backpropagation.

[0073] Then, the latent vector decomposition is performed on by using the factorization machine, the weighted features are input, and the cross vector is obtained:

[0074]

[0075] wherein, is a low-rank representation of the second-order interaction feature, is a second-order cross weight obtained by the inner product of the latent vector in the factorization machine, representing the interaction intensity of the th dimension and the th dimension feature, wherein ; The first and the second components of the weighted feature vector . . .

[0076] Then, the input is fed into a multi-layer perceptron with residual connections, containing at least 3 hidden layers, with Swish as the activation function, and the number of hidden units set to 256→128→64. The network structure can be represented as:

[0077] ;

[0078] wherein, y is the output vector of the i-th hidden layer, W and b are the weight vector and bias term of the i-th layer, respectively, with Swish activation function , , , wherein , , ,

[0079] In the training process of the neural network model, the embodiment constructs a joint loss function with the work order label and the work order handling duration , adopts a gradient optimizer to train the model, optimizes the network parameters, and the model loss function can be expressed as:

[0080] ;

[0081] wherein, the loss function includes two parts, the emergency index loss and the handling delay loss, and the weight can be set to 0.5 to balance the two parts of error of the emergency index and the actual handling delay. The above formula can be further represented as the following formula:

[0082] ;

[0083] wherein, the model training batch size is , denotes the work order emergency index prediction value output by the model for the i-th work order sample, is the work order emergency index label value (calculated by the actual response duration) of the i-th work order sample; ​​​​​​​​​​​The processing duration of the work order sample is the difference between the work order creation time and the work order processing completion time, is the processing duration prediction value output by the first work order sample model.

[0084] The processing duration prediction value can be calculated by the following formula:

[0085]

[0086] wherein, is the actual response duration of the first work order sample, which is the difference between the work order creation time and the work order dispatch time (i.e., the dispatch time), in minutes. The faster the response, the closer the label to 1. The model training uses the Adam optimizer, the training and verification ratio is 8:2, the initial learning rate is 0.001, the verification set loss does not decrease for 10 epochs, and the optimal weight is saved. The data set used for training is the historical social governance work order collected in step 1.

[0087] When , it automatically enters the "highest priority" work order queue; when , it enters the "high priority" work order queue; when , it enters the "ordinary priority" work order queue; when , it enters the "low priority" work order queue.

[0088] Step 5: Dispatch the work order according to the priority from "highest-high-ordinary-low" in turn.

[0089] The following takes a work order as an example to explain the generation of the emergency index, work order dispatching and parameter updating in this embodiment.

[0090] First, work order data collection and preprocessing. After the data is entered by the platform, it is stored in the database in json data format. The work order example is a gas pipeline leakage work order in a certain district:

[0091] {"reporter": "Mr. Wang (137****1234)", "report time": "2023-10-15 16:23:45", "event description": "There is a strong gas smell in the corridor of Building 23 in a certain community. Suspected pipeline leakage! Please handle as soon as possible!", "event type": "gas safety", "event address": "58, Xilu, Xianxi Street, Xianxi District, Xianxi City", "attachments": ["gas_leak_20231015.jpg"]}.

[0092] Second, input the event description text ["There is a strong gas smell in the corridor of Building 23 in a certain community. Suspected pipeline leakage! Please handle as soon as possible!"] into the text encoder ​, get the text feature vector .

[0093] Third step, input the image ["gas_leak_20231015162345.jpg"] to the image encoder get the image feature vector .

[0094] Fourth step, detect the consistency of the two modal feature vectors, calculate the cosine similarity , , greater than the threshold value 0.65, this work order passes the modal consistency detection and enters the subsequent C vector construction process.

[0095] Fifth step, based on the text feature vector calculate the emotion intensity vector , =[0.02,0.01,0.03,0.02,0.0.3,0.04,0.83,0.02,0.92] , where 0.83 represents the probability component of anxiety emotion, and 0.92 represents the emotion intensity.

[0096] Calculate the event risk level vector , create the event one-hot vector [0,0,0,0,0,1,…,0], multiply it with the event type weight to generate =[0, 0, 0, 0, 0, 0.95,…,0] , where 0.95 corresponds to the gas safety event.

[0097] Calculate the time encoding vector , the event occurrence time t=16:23 falls in the 6th segment (15:00-17:59), its one-hot encoding is =[0,0,0,0.67,0,0,0,0], query the time period weight table to get the weight 0.7, first multiply by the weight vector, then compress it by Sigmoid, get =[0,0,0,0.67,0,0,0,0] .

[0098] Sixth step, concatenate to get the interaction vector .

[0099] Seventh step, calculate the work order emergency index according to the vector , the output of this work order example is:

[0100] ​{"Work Order ID": "XXGL202310151623458", "Emergency Index": 0.94, "Emergency Level": Highest Priority, "Evaluation Basis": ["Event Type: Gas Safety", "Danger Level: Extremely High", "Time Encoding: Late Peak Hours", "Emotional Intensity: Extremely Anxious", "Responsible Department: Emergency Management"]}.

[0101] In the eighth step, according to the defined threshold rule, the work order is marked as "highest priority" according to the emergency index =0.94, the work order template is matched, the mandatory fields are automatically filled, and manual review is necessary.

[0102] In the ninth step, according to the work order emergency index, the work order is automatically assigned to the responsible department, and the model predicted disposal time is marked as the expected disposal time, which is 30 minutes.

[0103] In the tenth step, after the responsible department completes the on-site disposal, the disposal result is fed back, the actual disposal time of the work order is recorded as 42 minutes, the work order satisfaction is obtained through user evaluation as 5 points, and the model parameters are updated using the work order disposal time and user satisfaction.

[0104] Based on the method flow proposed in this embodiment, the work order time generated by relying on manual experience and static rules is about 8.7 minutes, and after applying this method, the work order preliminary review time is <1 min, the work order processing efficiency is improved by 8-10 times, the emergency work order recognition accuracy is increased by about 60%, and when the work order volume increases significantly, the efficiency can be further improved, and the potential loss and impact caused by high-risk events can be reduced.

[0105] Figure 6 and Figure 7 are respectively the social governance work order processing system and the database system interface diagram for implementing the method of this embodiment.

[0106] Compared with the existing method using only ReLU-MLP, the residual connection MLP introduced in this embodiment uses Swish as the activation function, has the characteristics of faster convergence and smoother prediction, avoids gradient disappearance, and outputs smooth and continuous. In addition, the low-rank cross features generated by the factorization machine are spliced with the original features to increase the dimension, and the residual connection retains more original feature information than other models, preventing overfitting in the case of small data volume. By comparing the baseline model, it can be seen that the residual MLP model designed for emergency index generation in this embodiment has good comprehensive performance in evaluation indicators. Compared with the traditional ReLU-MLP, the F1 value is improved to 0.872, the mean absolute error is reduced to 0.053, and the parameter amount is reduced by about 25% on the work order dataset, showing faster training convergence, more stable prediction output, and better comprehensive performance. As shown in Table 1.

[0107] Table 1 Performance comparison of the method of the embodiment and the traditional ReLU-MLP

[0108]

[0109] In another embodiment, the present application provides a computer readable storage medium, storing a computer program, the computer program causing a computer to execute the social governance work order dispatching method based on emergency degree assessment of the foregoing embodiment.

[0110] In another embodiment, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the social governance work order dispatching method based on emergency degree assessment of the foregoing embodiment when executing the computer program.

[0111] In the embodiments disclosed in the present application, the computer storage medium can be a tangible medium, which can contain or store programs for use by or in connection with an instruction execution system, apparatus or device. The computer storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses or devices, or any suitable combination of the above. More specific examples of computer storage media can include one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CDROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0112] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed in the present application can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those of ordinary skill in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0113] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for those of ordinary skill in the art, some improvements and refinements without departing from the principles of the present application shall be considered within the protection scope of the present application.

Claims

1. A social governance work order dispatching method based on emergency level assessment, characterized in that, The method comprises the following steps: Step 1: Collecting text data and image data in a social governance work order, calling a text encoder and an image encoder, and converting the text data and the image data into a text feature vector and an image feature vector respectively; Step 2: Performing consistency detection on the text feature vector and the image feature vector of the same work order, and assigning the work order that fails the consistency detection to a manual review queue; Step 3: For the work order that passes the consistency detection, calculating the emotional intensity based on the text feature vector, and combining the danger level and the time code of the corresponding work order event to splice a generated vector; Step 4: Constructing and training a neural network model comprising a multilayer perceptron with residual connection, inputting the spliced generated vector into the trained neural network model to generate a work order emergency index, and dividing the emergency degree of the work order by comparing the work order emergency index with a preset threshold; Step 5: Assigning the work order according to the emergency degree of the work order.

2. The social governance work order dispatching method based on emergency level assessment according to claim 1, characterized in that: In step 1, the text encoder comprises an embedding layer, a bidirectional LSTM, a global average pooling layer and a full connection layer connected in sequence, and a Relu activation function is arranged after the full connection layer; and the text encoder finally outputs a text feature vector with a vector dimension of 256.

3. The social governance work order dispatching method based on emergency level assessment of claim 1, wherein: In step 1, the image encoder comprises a 32-channel 3*3 convolution layer, a 2*2 max pooling layer, a 64-channel 3*3 convolution layer, a 2*2 max pooling layer, a 128-channel 3*3 convolution layer, a 2*2 max pooling layer, a feature aggregation layer, an average pooling layer and a full connection layer connected in sequence, wherein a Relu activation function is arranged after each 2*2 max pooling layer; and the image encoder finally outputs an image feature vector with a vector dimension of 256.

4. The social governance work order dispatching method based on emergency level assessment of claim 1, wherein: In step 2, the consistency detection is specifically: calculating a cosine similarity of the text feature vector and the image feature vector comparing the cosine similarity with a set threshold value , if , it is considered that the consistency test is passed, if , it is considered that the consistency test is not passed; wherein, by Sliding update , subscript represents the time step, and respectively represent and the threshold value and the cosine similarity at the time step , is a set parameter for making the threshold value gradually converge to the optimal discrimination point.

5. The social governance work order dispatching method based on emergency level assessment of claim 1, wherein: In step 3, the calculation process of the emotion intensity is as follows: inputting the text feature vector into the sentiment analysis model Senta, outputting a probability vector of different emotions to form an emotion intensity vector wherein, represents a probability vector corresponding to an emotion, represents an emotion intensity scalar, represents a dimension of ; The calculation process of the risk degree is: creating a one-hot vector, writing the weight of the predefined work order event type in the corresponding dimension of the one-hot vector, and setting the remaining dimensions to 0 to generate a risk degree vector wherein represents a weight corresponding to an event type The calculation process of the time encoding is: dividing 24 hours into equal time periods, according to the time of occurrence of the work order event to obtain a one-hot encoding dimension vector, then multiplying each dimension according to a predefined weight of each time period, and compressing through a Sigmoid function to obtain a time encoding vector ; Final stitching to generate vector .

6. The social governance work order dispatching method based on emergency level assessment of claim 5, wherein: In step 4, the process of generating the work order emergency index by the neural network model is: Step 4.1: Normalization of the vector The normalized feature vector is obtained by normalizing the vector ; generating attention weights , re-weighting the vector to obtain the feature vector ; ； wherein denotes the Hadamard product, is corresponding to the weight of the th feature component, is an intermediate vector obtained after a linear transformation of , , and are the learning weight matrix and bias term, respectively, with subscript being the dimension index, and are the th and th feature component of , respectively, exp denotes the exponential function; Step 4.2: Factorizing the eigenvectors using a factorization machine Performing latent vector decomposition to obtain cross-vectors : wherein is the low-rank representation of the second-order interaction feature, is the second-order cross weight obtained by the inner product of the hidden vectors in the factorization machine, indicating the interaction intensity between the th and the th features, and are the th and the th feature components of the feature vector , respectively. are concatenated to form a vector with concatenated to form a vector , denotes a horizontal concatenation Step 4.3: The input is fed into a multi-layer perceptron with residual connections, the multi-layer perceptron comprising one hidden layer, each hidden layer followed by a Swish activation function, the network structure represented as: one hidden layer, each hidden layer followed by a Swish activation function, the network structure represented as: ； wherein, , is the output vector of the th hidden layer, and are the weight vector and bias term of the th layer, respectively. Multilayer perceptron final output ticket urgency index wherein is a sigmoid function, and is a weight matrix and bias vector of the output layer.

7. The social governance work order dispatching method based on emergency level assessment of claim 6, wherein: In step 4, the neural network model takes the ticket label and ticket handling duration A joint loss function is constructed, and the model is trained using a gradient optimizer, and the joint loss function is represented as: ； In the formula, the model training batch size is , Refers to the first The predicted urgency index of a work order, output by a sample work order model. For the first The work order urgency index label value of a sample of work orders. , For the first The actual response time of each work order sample; For the first The processing time for each work order sample is the difference between the work order creation time and the work order processing completion time. For the first The predicted processing time output by the sample work order model.

8. The social governance work order dispatching method based on emergency level assessment of claim 1, wherein: In step 5, the assigning of the work order according to the emergency degree of the work order is specifically: According to the emergency degree of the work order, the priority of the work order is divided, and the work order is assigned in sequence according to the priority of the work order.

9. A computer readable storage medium storing a computer program, characterized in that, The computer program enables the computer to execute the social governance work order assigning method based on emergency degree evaluation as claimed in any one of claims 1-8.

10. An electronic device, comprising: It comprises: a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the social governance work order assigning method based on emergency degree evaluation as claimed in any one of claims 1-8 when executing the computer program.

Images