CN-122022077-A - Maintenance fund management platform budget intelligent measuring and calculating method and system based on machine learning

CN122022077ACN 122022077 ACN122022077 ACN 122022077ACN-122022077-A

Abstract

The application discloses a machine learning-based intelligent calculation method and system for a maintenance fund management platform budget, wherein the method comprises the steps of respectively obtaining on-site investigation images, description texts and archive data of maintenance projects, inputting the obtained on-site investigation images of the maintenance projects into a pre-constructed convolutional neural network engineering quantity verification model to generate standardized engineering quantity data, calculating budget prices of the maintenance projects according to the standardized engineering quantity data and the archive data, simultaneously predicting maintenance fund expenditure in a future period, carrying out semantic analysis on the description texts through a natural language processing technology, matching policy terms and calculating compliance risks, transversely comparing the prices in the description texts by utilizing an anomaly detection algorithm, and outputting reasonable budget results, and simultaneously carrying out optimal allocation of different period limit deposits on the total funds of a private user by combining future maintenance fund expenditure prediction in preset time. The application improves the refinement level of maintenance fund management.

Inventors

BAO MINGMING
CHEN XIUFANG
DUAN WEIQI
LU LIFENG
SHI BO
HAN CHAO
LI TIANNING
HUANG HONGRUI
ZHANG YUNHAO
ZHAO JIAYU
HAN YUANYUAN
WAN YUNLIANG

Assignees

山东嘉友互联软件股份有限公司

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. The intelligent calculation method for the budget of the maintenance fund management platform based on machine learning is characterized by comprising the following steps of: Respectively acquiring on-site investigation images, description texts and archive data of maintenance projects; Inputting the acquired on-site investigation image of the maintenance project into a pre-constructed convolutional neural network engineering quantity verification model to generate standardized engineering quantity data; Calculating the budget price of the maintenance project according to the standardized engineering quantity data and the archive data, and simultaneously predicting the maintenance fund expenditure in a future preset period; semantic analysis is carried out on the description text of the maintenance project through a natural language processing technology, policy clauses are matched, the compliance risk is calculated, and meanwhile, the price in the description text is transversely compared through an anomaly detection algorithm; and outputting a reasonable budget result after the comparison is successful, and simultaneously carrying out optimal allocation of different period limit deposits on the sum of the funds of the private users by combining the maintenance funds expenditure prediction of the future preset time.
2. The intelligent calculation method of the maintenance fund management platform budget based on machine learning according to claim 1, wherein the convolutional neural network engineering quantity verification model comprises a trunk feature extraction network, an input end of which is used for receiving a preprocessed on-site investigation image, an output end of which is connected with an input end of a regional advice network, an output end of which is connected with an input end of an ROI alignment layer, an output end of which is respectively connected with input ends of a classification branch, a bounding box regression branch, a mask segmentation branch and an OCR digital extraction sub-network, and output ends of the classification branch, the bounding box regression branch, the mask segmentation branch and the OCR digital extraction sub-network are connected with an output layer, and a calculation formula of a loss function adopted when the convolutional neural network engineering quantity verification model is trained is as follows: Wherein, the As a function of the total loss, For the weight coefficient of the classification loss, In order to classify the loss of the device, The weight coefficients for the bounding box regression loss, For the bounding box regression loss, The weight coefficients lost for the mask segmentation, For the mask segmentation loss to be sufficient, The lost weighting coefficients are identified for OCR numbers, For the purpose of OCR digital recognition loss, Is the number of positive samples that are to be taken, As a total number of candidate regions, For the index of the candidate region, As the class weight factor is used, Model predictive first The probability that the candidate regions belong to the category t, In order to be able to take the focus parameter as such, For the cross-correlation of the predicted frame with the real frame, To predict the euclidean distance between the frame center point and the true frame center point, In order to predict the parameters of the frame, As a parameter of the real frame, For the diagonal length of the smallest bounding rectangle of the predicted and real frames, As a measure of the uniformity of the aspect ratio, And The height and width of the mask image respectively, For the index of the pixel(s), Model predictive first The probability that an individual pixel belongs to a damaged area, Is the first The true label of the individual pixels is that, In order to smooth the term(s), In order for the coefficient of balance to be present, For the pixel-level loss, As the total number of pixels, As a class balancing factor, For the length of the feature sequence, Outputting characters for the model at the t-th time step Is a function of the probability of (1), Mapping to target tag sequences by merging repeated characters and removing blanks Is provided with a set of paths, For one of the paths, the character sequence output at each time step is represented.
3. The machine learning-based intelligent calculation method for a budget for a maintenance fund management platform according to claim 1, wherein inputting the acquired on-site survey image of the maintenance project into a pre-constructed convolutional neural network engineering quantity verification model to generate standardized engineering quantity data comprises: performing size normalization, color space conversion and illumination balance processing on the obtained on-site investigation image, estimating a scale factor of image pixels and actual physical length based on image EXIF information, and obtaining a preprocessed image tensor; Inputting the preprocessed image tensor into a trunk feature extraction network of a convolutional neural network engineering quantity verification model to perform multi-scale damage feature extraction, and generating a multi-scale feature pyramid; generating a dense anchor point frame on each level of the feature pyramid through a regional suggestion network, and screening candidate regions by adopting non-maximum suppression after foreground or background classification and bounding box regression; Mapping each candidate region to a corresponding level of a feature pyramid through ROI alignment layer operation, and uniformly sampling to a fixed size to obtain an aligned region feature map; The aligned regional feature images are simultaneously input into a classification branch, a boundary box regression branch, a mask segmentation branch and an OCR digital extraction sub-network, and damage types and grades, accurate boundary box coordinates of a damaged region, pixel-level binary masks of the damaged region and size labels in the images are respectively output; the method comprises the steps of performing pixel counting on a pixel-level binary mask output by a mask dividing branch, calculating a damaged area by combining a scale coefficient, calculating a damaged length by a skeleton extraction algorithm, and estimating a damaged width or depth based on a minimum circumscribed rectangle of the mask, wherein the calculation formulas are respectively as follows: Wherein, the In order to damage the area of the substrate, And The height and width of the mask image respectively, Is the first The mask value of the individual pixels is set, To indicate the function when When 1 is taken, otherwise 0 is taken, For the physical width of the image sensor, In order to take a picture of the distance, Is the focal length of the lens and, For the pixel width of the image, In order to damage the length of the cable, In order to obtain a single-pixel wide central line pixel set after skeleton extraction of the mask, To damage width or depth; And fusing the damage type and grade, the accurate boundary frame coordinates of the damage area, the damage length, the damage width and the size marks in the image to obtain standardized engineering quantity data.
4. The intelligent computing method for budget of maintenance fund management platform based on machine learning according to claim 1, wherein computing the budget price of the maintenance project according to the standardized engineering quantity data in combination with the archive data while predicting maintenance fund expenditure in a future preset period comprises: Extracting a base period rated unit price, a material market real-time unit price and a site measure fee matched with a maintenance project from the archive data; According to the standardized engineering quantity data, calculating the reference cost of the current maintenance project by combining the base period rated unit price, the material market real-time unit price and the site measure cost, wherein the calculation formula is as follows: Wherein, the As a benchmark fee for the current maintenance project, To base each sub-project amount in the standardized project amount data, Is the standard rated unit price of the utility model, Is the first The weight coefficient of the seed material in the total cost, As an index of the current price it is, As the price index of the reference period, In order to obtain the number of kinds of materials, In order to divide the number of projects, Is the site measure fee; Based on the maintenance type feature vector, correcting the reference cost through an experience correction coefficient to obtain a final budget price, wherein the calculation formula is as follows: Wherein, the For the final budget price to be reached, In order to maintain the type of feature vector, For a non-linear adjustment term based on engineering quantities, In order to adjust the coefficient of the coefficient, Is the first The weight of the class experience is that, As a function of the rules of the rule, Is an empirical gradient term; And simultaneously, predicting maintenance fund expenditure in a future preset period.
5. The intelligent computing method for budget of a maintenance fund management platform based on machine learning according to claim 4, wherein said predicting maintenance fund expenditure in a future preset period comprises: Based on house characteristic vectors and historical actual expenditure values, maintenance fund expenditure in a future preset period is generated through a time sequence prediction model, and a calculation formula is as follows: Wherein, the Is the predicted first The value of the capital expenditure for the period maintenance, For the first empirical gradient term, reflecting the trend of degradation based on house characteristics, First, the The weighting coefficients of the individual empirical gradient terms, The representation is based on house characteristics Is the first of (2) A degradation-like trend or a maintenance requirement change gradient, As the house characteristic vector of the t-th period, Is the first The coefficient of the order auto-regressive, Is the first The historical actual maintenance capital expenditure values for the period, Is the first The order of the moving average coefficient is such that, Is the first A history of the residual terms of the period, The weight coefficients for the non-linear feature map term, For a non-linear feature mapping term based on house features, And The autoregressive order and the moving average order, respectively.
6. The intelligent computing method of the maintenance fund management platform budget based on the machine learning according to claim 1, wherein the semantic parsing of the descriptive text of the maintenance project by the natural language processing technology, the matching of policy clauses and the calculation of compliance risk comprise: performing word segmentation, part-of-speech tagging and dependency syntactic analysis on the description text of the maintenance project to generate a preprocessed text sequence, and simultaneously calculating the weight of each word by adopting a TF-IDF algorithm; Inputting the text sequence into an entity recognition model of a professional dictionary and a rule base which are integrated with the maintenance fund field, recognizing and extracting maintenance objects, maintenance ranges, material specifications and monetary digital entities in the text, and generating an entity set; Encoding each identified entity text to obtain an entity embedded vector, and simultaneously calculating the weight of each entity by combining the vocabulary weight of the corresponding entity, wherein the calculation formula is as follows: Wherein, the Is the first The integrated weight of the individual entities is used, Output for entity recognition model The confidence score of the individual entity, Is the first The number of words contained in the individual entity text, The sum of TF-IDF weights of all words in the entity text; constructing a weighted query vector according to the weight of each entity and the entity embedding vector, wherein the calculation formula is as follows: Wherein, the In order to weight the query vector, For the total number of entities identified from the text, Is the first Embedding vectors for the individual entities; Each clause text in the policy clause library is coded in advance, a clause embedded vector is generated, a nearest neighbor index is established, the weighted query vector is input into a matching engine, matching clauses are executed, and a compliance risk is calculated.
7. The machine learning based maintenance fund management platform budget intelligent measuring and calculating method according to claim 6, wherein inputting the weighted query vector into a matching engine, performing matching terms and calculating compliance risk comprises: And (3) carrying out accurate matching on the entity text in the entity set and forbidden keywords in the policy clause library, if keyword matching is not triggered in the matching process, searching a preset clause which is most similar to the weighted query vector in the index, calculating weighted cosine similarity of the weighted query vector and each candidate clause embedded vector, outputting the clause with the highest similarity score and the maximum semantic similarity value, wherein the calculation formula is as follows: Wherein, the To weight the query vector and the first The policy terms embed cosine similarity of vectors, In order to weight the query vector, Is the first The embedded vector of the policy terms, Is the L2 norm of the vector; according to the maximum semantic similarity value, combining the semantic distance between the entity type and the clause applicable type, calculating a compliance risk index, wherein a calculation formula is as follows: Wherein, the In order to be a risk index of compliance, For the maximum semantic similarity value, In order for the attenuation coefficient to be a factor, For the semantic distance between the entity type and the matched clause applicable type, E is the entity set, Numbering policy terms most similar to the weighted query vector; If matching exists in the matching process, the compliance risk index is directly output, and the matched clause numbers are recorded.
8. The intelligent computing method for the budget of the maintenance fund management platform based on machine learning according to claim 1, wherein the step of using an anomaly detection algorithm to transversely compare prices in descriptive texts comprises the steps of: Extracting a declaration price from a description text of a maintenance project, and constructing an input feature vector by combining calculated budget price, standardized engineering quantity data and historical similar project price information in archive data; Inputting the input characteristic vector into a pre-trained self-encoder, obtaining hidden layer representation through the encoder, and reconstructing through a decoder to obtain an output vector, wherein the calculation formula is as follows: Wherein, the For the hidden layer representation of the encoder output, In order to input the feature vector(s), For the encoder function, input feature vectors Mapping to the hidden-layer space is performed, In order to activate the function, For the weight matrix of the encoder, For the offset vector of the encoder, For the reconstructed output vector of the decoder, In order for the decoder function to be a function, For the weight matrix of the decoder, Is a bias vector for the decoder; and calculating a weighted mean square reconstruction error according to the output vector, wherein the calculation formula is as follows: Wherein, the In order to weight the mean square reconstruction error, As a dimension of the features, For inputting feature vectors Is the first of (2) The number of components of the composition, Reconstructing a vector Is the first of (2) The number of components of the composition, Is normal sample at the first The average value over the individual features is used, In the first place for abnormal samples The average value over the individual features is used, Is normal sample at the first The standard deviation in the individual features is used, In the first place for abnormal samples Standard deviation on individual features; comparing the weighted mean square reconstruction error with a preset threshold, judging that the declaration price is normal when the weighted mean square reconstruction error is smaller than the preset threshold, and outputting a comparison passing mark; and when the weighted mean square reconstruction error is larger than a preset threshold value, judging that the declared price is abnormal, and outputting an abnormal alarm and an abnormal type.
9. The intelligent computing method of the maintenance fund management platform budget based on machine learning according to claim 1, wherein the computing method is characterized in that after the comparison is successful, a reasonable budget result is output, and meanwhile, the optimal configuration of different period limit deposits for the total amount of the funds of the private user is carried out by combining the maintenance fund expenditure prediction of the future preset time, and the computing method comprises the following steps: constructing an objective function for maximizing total income and punishing liquidity risks according to the reasonable budget result, the maintenance fund expenditure value of the future preset time, the current sum of funds of a special user and the annual interest rate of the regular deposit of each term, wherein the calculation formula is as follows: Simultaneously setting a fund total constraint and a non-negative integer constraint, wherein the calculation formula is as follows: Wherein, the For the value of the objective function, Is the first The annual interest rate of the regular deposit, Is the first The amount of the allocation of the regular deposit, For the liquidity risk penalty factor, For the purpose of the year index, For the predicted payout amount of the t-th year, To the first Total amount of funds due by the end of the year, The sum of funds for the current private household; And solving by an integer programming solver to obtain the optimal configuration.
10. The utility model provides a maintenance fund management platform budget intelligence measurement and calculation system based on machine learning which characterized in that includes: The acquisition module is used for acquiring on-site investigation images, description texts and archive data of maintenance projects; The engineering quantity intelligent verification module is used for inputting the acquired on-site investigation image of the maintenance project into a pre-constructed convolutional neural network engineering quantity verification model to generate standardized engineering quantity data; the cost prediction module is used for calculating the budget price of the maintenance project according to the standardized engineering quantity data and the archive data, and predicting the maintenance fund expenditure in a future preset period; The matching analysis module is used for carrying out semantic analysis on the description text of the maintenance project through a natural language processing technology, matching policy clauses and calculating compliance risks, and simultaneously, carrying out transverse comparison on the prices in the description text by utilizing an anomaly detection algorithm; and the configuration module outputs a reasonable budget result after successful comparison, and simultaneously optimally configures different period limit deposits for the sum of the funds of the private users by combining the maintenance funds expenditure prediction of the future preset time.

Description

Maintenance fund management platform budget intelligent measuring and calculating method and system based on machine learning Technical Field The invention relates to the technical field of maintenance fund management, in particular to an intelligent budget measuring and calculating method and system for a maintenance fund management platform based on machine learning. Background The maintenance funds are used as core funds for updating and reforming, and the accuracy and the compliance of budget measurement and the rationality of funds allocation are directly related to the funds use efficiency. The budget measurement work of the current maintenance fund management platform is still mainly carried out in a traditional manual mode, and obvious technical short plates exist in the links of engineering quantity verification, budget calculation, compliance verification, fund configuration and the like, so that the modern fine management requirements are difficult to adapt. The traditional maintenance project engineering quantity verification highly depends on-site manual investigation, measurement and statistics of workers, so that a great amount of manpower and material resources are consumed, the investigation efficiency is low, the engineering quantity verification is easily influenced by factors such as manual subjective judgment, measuring tool precision, on-site environment and the like, the engineering quantity data error is large, the standardization degree is low, and hidden data distortion hidden troubles are measured and calculated for subsequent budget. Meanwhile, the manually recorded investigation result is mostly reserved in the form of unstructured text or paper documents, and is difficult to realize data linkage with a budget measuring and calculating system, so that the overall working efficiency is further reduced. In the budget price calculation link, the existing mode directly uses fixed quota standard, dynamic adaptation of material market price fluctuation, maintenance type differentiation requirement and historical maintenance experience is lacked, characteristics such as building year and structure type of an unbonded house are subjected to personalized correction, deviation of the budget price and actual construction cost is large, and the problems that funds are wasted due to overhigh budget and normal construction cannot be supported due to overlow budget are easily caused. In addition, for the prediction of future maintenance fund expenditure, a simple trend extrapolation method is mostly adopted, multidimensional information such as house degradation characteristics, historical expenditure data and the like are not fully fused, the accuracy of a prediction result is not enough, and effective data support cannot be provided for fund long-term planning. In addition, for the rationality verification of the declared price of the maintenance project, an effective transverse comparison mechanism is lacking, and the abnormal conditions such as false alarm, high-priced price and the like cannot be rapidly identified only by means of manual experience judgment, so that the normalization and rationality of fund use are difficult to be effectively ensured. Disclosure of Invention In order to solve the technical problems, the application provides the following technical scheme: In a first aspect, an embodiment of the present application provides a machine learning-based intelligent measurement and calculation method for a budget of a maintenance fund management platform, including: Respectively acquiring on-site investigation images, description texts and archive data of maintenance projects; Inputting the acquired on-site investigation image of the maintenance project into a pre-constructed convolutional neural network engineering quantity verification model to generate standardized engineering quantity data; Calculating the budget price of the maintenance project according to the standardized engineering quantity data and the archive data, and simultaneously predicting the maintenance fund expenditure in a future preset period; semantic analysis is carried out on the description text of the maintenance project through a natural language processing technology, policy clauses are matched, the compliance risk is calculated, and meanwhile, the price in the description text is transversely compared through an anomaly detection algorithm; and outputting a reasonable budget result after the comparison is successful, and simultaneously carrying out optimal allocation of different period limit deposits on the sum of the funds of the private users by combining the maintenance funds expenditure prediction of the future preset time. In one possible implementation manner, the convolutional neural network engineering quantity verification model comprises a trunk feature extraction network, wherein the input end of the trunk feature extraction network is used for receiving a preprocessed on-site investigat