Machine learning-based maintenance fund management platform budget intelligent calculation method and system

Abstract

The application discloses a kind of based on machine learning's maintenance fund management platform budget intelligent measurement method and system, the method includes: the field reconnaissance image of maintenance project, description text and archive data are respectively acquired;The field reconnaissance image of the maintenance project obtained is input into the pre-constructed convolutional neural network engineering quantity verification model, to generate standardization engineering quantity data;According to standardization engineering quantity data, the budget price of maintenance project is calculated in combination with archive data, while the maintenance fund expenditure in future period is predicted;The semantic analysis of description text is carried out by natural language processing technology, and the compliance risk is calculated by matching policy provisions and outputting reasonable budget result after the price in description text is compared horizontally using anomaly detection algorithm, while the optimal configuration of different term deposits of total amount of special account funds is carried out in combination with the maintenance fund expenditure prediction of future preset time.The application improves the fine level of maintenance fund management.

Description

Technical Field

[0001] This invention relates to the field of maintenance fund management technology, specifically to a machine learning-based intelligent budget calculation method and system for maintenance fund management platforms. Background Technology

[0002] As the core funding for renovation and upgrading, the accuracy, compliance, and rationality of the budget calculation and allocation of maintenance funds directly affect the efficiency of fund utilization. Currently, the budget calculation work of maintenance fund management platforms still relies mainly on traditional manual methods, which have significant technical shortcomings in areas such as project quantity verification, budget calculation, compliance review, and fund allocation, making it difficult to adapt to the needs of modern refined management.

[0003] Traditional maintenance project quantity verification relies heavily on manual on-site surveys, measurements, and statistics. This not only consumes significant manpower and resources and is inefficient, but is also susceptible to subjective judgment, the accuracy of measuring tools, and on-site environmental factors, leading to large errors and low standardization in quantity data. This creates potential for data distortion in subsequent budget calculations. Furthermore, manually recorded survey results are often stored in unstructured text or paper documents, making it difficult to integrate with budget calculation systems and further reducing overall work efficiency.

[0004] In the budget calculation stage, existing methods often directly apply fixed quota standards, lacking dynamic adaptation to fluctuations in material market prices, differentiated needs for different types of repairs, and historical repair experience. They also fail to incorporate personalized adjustments based on characteristics such as the building's age and structural type, leading to significant discrepancies between budgeted prices and actual construction costs. This can result in over-budgeting leading to wasted funds or under-budgeting failing to support normal construction. Furthermore, forecasting future maintenance expenditures often relies on simple trend extrapolation methods, failing to fully integrate multi-dimensional information such as building degradation characteristics and historical expenditure data. This results in insufficient accuracy of forecasts and an inability to provide effective data support for long-term financial planning.

[0005] Furthermore, there is a lack of effective cross-comparison mechanisms for verifying the reasonableness of repair project price declarations. Relying solely on manual experience makes it difficult to quickly identify abnormal situations such as false or inflated prices, and the standardization and rationality of fund use cannot be effectively guaranteed. Summary of the Invention

[0006] In order to solve the above-mentioned technical problems, this application proposes the following technical solution:

[0007] In a first aspect, embodiments of this application provide a machine learning-based intelligent budget calculation method for a maintenance fund management platform, including:

[0008] The site survey images, descriptive texts, and archival data for the maintenance project were acquired respectively.

[0009] The acquired on-site survey images of the maintenance project are input into a pre-built convolutional neural network engineering quantity verification model to generate standardized engineering quantity data;

[0010] The budget price of the maintenance project is calculated based on the standardized engineering quantity data and the archive data, and the maintenance fund expenditure in the future preset period is predicted.

[0011] Natural language processing technology is used to semantically parse the description text of the repair project, match it with policy clauses and calculate compliance risks. At the same time, anomaly detection algorithms are used to make horizontal comparisons of prices in the description text.

[0012] After successful comparison, a reasonable budget result is output. At the same time, the total amount of funds in the special account is optimally allocated to deposits of different terms based on the future maintenance fund expenditure forecast at a preset time.

[0013] In one possible implementation, the convolutional neural network engineering quantity verification model includes: a backbone feature extraction network, the input of which is used to receive preprocessed on-site survey images; the output of which is connected to the input of a region proposal network; the output of which is connected to the input of an ROI Align layer; the output of which is connected to the inputs of a classification branch, a bounding box regression branch, a mask segmentation branch, and an OCR digital extraction sub-network; and the outputs of which are connected to an output layer. The loss function used during training of the convolutional neural network engineering quantity verification model is calculated using the following formula:

[0014]

[0015]

[0016]

[0017]

[0018]

[0019]

[0020] in, For the total loss function, These are the weighting coefficients for the classification loss. For classifying losses, These are the weighting coefficients for the bounding box regression loss. For bounding box regression loss, These are the weighting coefficients for the mask segmentation loss. For mask segmentation loss, These are the weighting coefficients for the OCR digit recognition loss. For OCR digital recognition loss, The number of positive samples. The total number of candidate regions. For the index of the candidate region, For category weighting factors, The model predicts the first The probability that each candidate region belongs to category t. To focus parameters, The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. To predict the Euclidean distance between the center point of the bounding box and the center point of the ground truth bounding box, For the parameters of the prediction box, The parameters are for the actual bounding box. Let be the diagonal length of the minimum bounding rectangle between the predicted bounding box and the ground truth bounding box. As a measure of aspect ratio consistency, and These are the height and width of the mask image, respectively. For pixel index, The model predicts the first The probability that a pixel belongs to a damaged area. For the first The real label of each pixel For smoothing terms, For balance coefficient, For pixel-level loss, This represents the total number of pixels. As a class balance factor, The length of the feature sequence. Output the character for the model at time step t. The probability, For all sequences that can be mapped to the target label sequence by merging duplicate characters and removing whitespace. The set of paths For one of the paths, it represents the character sequence output at each time step.

[0021] In one possible implementation, the acquired on-site survey images of the maintenance project are input into a pre-built convolutional neural network engineering quantity verification model to generate standardized engineering quantity data, including:

[0022] The acquired field survey images are normalized in size, converted in color space and balanced in illumination. The scale factor of the image pixels and the actual physical length is estimated based on the image EXIF ​​information to obtain the preprocessed image tensor.

[0023] The preprocessed image tensor is input into the backbone feature extraction network of the convolutional neural network engineering quantity verification model to extract multi-scale damage features and generate a multi-scale feature pyramid.

[0024] Dense anchor boxes are generated at each level of the feature pyramid through a region proposal network. After foreground or background classification and bounding box regression, candidate regions are selected by non-maximum suppression.

[0025] Each candidate region is mapped to the corresponding level of the feature pyramid through the ROI Align layer operation, and then uniformly sampled to a fixed size to obtain the aligned region feature map;

[0026] The aligned region feature maps are simultaneously input into the classification branch, bounding box regression branch, mask segmentation branch and OCR digital extraction subnetwork, and output the damage type and level, the precise bounding box coordinates of the damaged area, the pixel-level binary mask of the damaged area and the size annotation in the image, respectively.

[0027] Pixel counts are performed on the pixel-level binary mask output from the mask segmentation branch. The damaged area is calculated using a scale factor, the damaged length is calculated using a skeleton extraction algorithm, and the damaged width or depth is estimated based on the minimum bounding rectangle of the mask. The calculation formulas are as follows:

[0028]

[0029]

[0030]

[0031] in, For the damaged area, and These are the height and width of the mask image, respectively. For the first The mask value of each pixel. For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. The physical width of the image sensor, For shooting distance, For the lens focal length, The pixel width of the image. For the length of damage, This refers to the set of single-pixel wide centerline pixels obtained after skeleton extraction from the mask. Damage to width or depth;

[0032] The damage type and level, the precise bounding box coordinates of the damaged area, the damaged area, the damaged length, the damaged width, and the dimension annotations in the image are fused to obtain standardized engineering quantity data.

[0033] In one possible implementation, the budget price of the maintenance project is calculated based on the standardized engineering quantity data and the archival data, while the maintenance fund expenditure within a preset future period is predicted, including:

[0034] Extract the base period quota unit price, real-time material market unit price, and on-site measures fee that match the maintenance project from the archive data;

[0035] Based on the standardized engineering quantity data, combined with the base period quota unit price, real-time material market unit price, and on-site measures fee, the benchmark cost of the current maintenance project is calculated using the following formula:

[0036]

[0037] in, This is the base cost for the current repair project. Based on the quantities of each sub-item of the standardized engineering quantity data, The benchmark unit price, For the first The weighting coefficient of each material in the total cost The current price index, The price index for the base period, For the number of material types, For the quantity of sub-items, For on-site measures costs;

[0038] Based on the repair type feature vector, the baseline cost is adjusted using an empirical correction coefficient to obtain the final budget price. The calculation formula is as follows:

[0039]

[0040] in, For the final budget price, For the maintenance type feature vector, This is a non-linear adjustment term based on the quantity of work. To adjust the coefficient, For the first Weight of class experience, For regular functions, For empirical gradient terms;

[0041] At the same time, the expenditure on maintenance funds in the future is predicted within a predetermined period.

[0042] In one possible implementation, the prediction of maintenance fund expenditures over a predetermined future period includes:

[0043] Based on the building feature vector and historical actual expenditure values, a time-series prediction model is used to generate maintenance fund expenditures for a preset future period. The calculation formula is as follows:

[0044]

[0045] in, For the predicted first Maintenance fund expenditure value, Let be the i-th empirical gradient term, reflecting the degradation trend based on house characteristics. No. The weight coefficients of each empirical gradient term, Indicated based on house characteristics The Degradation trend or gradient of maintenance needs Let be the house feature vector in period t. For the first Autoregressive coefficients of order 1 For the first Historical actual maintenance expenditure value during the period For the first Moving average coefficient, For the first Historical residuals of the period These are the weighting coefficients of the nonlinear feature mapping term. For nonlinear feature mapping terms based on house features, and These are the autoregression order and the moving average order, respectively.

[0046] In one possible implementation, the step of semantically parsing the description text of the maintenance project using natural language processing technology, matching policy clauses, and calculating compliance risks includes:

[0047] The description text of the maintenance project is segmented, part-of-speech tagging and dependency parsing are performed to generate a preprocessed text sequence. At the same time, the weight of each word is calculated using the TF-IDF algorithm.

[0048] The text sequence is input into an entity recognition model that integrates a professional dictionary and rule base in the field of maintenance funds to identify and extract maintenance objects, maintenance scope, material specifications, and monetary entities from the text, and generate an entity set.

[0049] Each identified entity text is encoded to obtain an entity embedding vector. Simultaneously, the weight of each entity is calculated by combining the corresponding word weights. The calculation formula is as follows:

[0050]

[0051] in, For the first The overall weight of each entity The first output of the entity recognition model The confidence score of each entity. For the first The number of words contained in an entity text. This is the sum of the TF-IDF weights of all words in the entity text;

[0052] A weighted query vector is constructed based on the weight of each entity and its embedding vector. The calculation formula is as follows:

[0053]

[0054] in, For weighted query vectors, This represents the total number of entities identified from the text. For the first Embedding vectors of each entity;

[0055] Each clause text in the policy clause library is pre-encoded to generate a clause embedding vector, and a nearest neighbor index is built. The weighted query vector is then input into the matching engine to perform clause matching and calculate compliance risks.

[0056] In one possible implementation, the weighted query vector is input into the matching engine to execute matching terms and calculate compliance risks, including:

[0057] The entity text in the entity set is precisely matched with the prohibited keywords in the policy clause library. If no keyword matching is triggered during the matching process, a preset number of clauses most similar to the weighted query vector are retrieved from the index. The weighted cosine similarity between the weighted query vector and the embedding vector of each candidate clause is calculated, and the clause with the highest similarity score and the maximum semantic similarity value are output. The calculation formula is as follows:

[0058]

[0059] in, For weighted query vector and the first The cosine similarity of the embedding vectors of the policy clauses. For weighted query vectors, For the first The embedding vector of each policy clause Let L2 be the norm of the vector;

[0060] Based on the maximum semantic similarity value, and combined with the semantic distance between the entity type and the applicable clause type, the compliance risk index is calculated using the following formula:

[0061]

[0062] in, This is a compliance risk index. The maximum semantic similarity value. The attenuation coefficient is... Let E be the semantic distance between the entity type and the applicable type of the matched terms, and let E be the set of entities. The policy clause number that is most similar to the weighted query vector;

[0063] If a match is found during the matching process, the compliance risk index will be output directly, and the matching clause number will be recorded.

[0064] In one possible implementation, the step of using an anomaly detection algorithm to perform a horizontal comparison of prices in the descriptive text includes:

[0065] Extract the declared price from the description text of the maintenance project, and combine it with the calculated budget price, standardized engineering quantity data, and historical price information of similar projects in the archive data to construct an input feature vector;

[0066] The input feature vector is fed into a pre-trained autoencoder to obtain the hidden layer representation, which is then reconstructed by the decoder to obtain the output vector. The calculation formula is as follows:

[0067]

[0068]

[0069] in, This is the hidden layer representation output by the encoder. For the input feature vector, For encoder functions, the input feature vector is... Mapped to the hidden space, For activation function, Here is the weight matrix of the encoder. This is the bias vector of the encoder. The output vector reconstructed by the decoder, For decoder functions, Here is the weight matrix of the decoder. This is the bias vector for the decoder;

[0070] The weighted mean square reconstruction error is calculated based on the output vector, using the following formula:

[0071]

[0072] in, For weighted mean square reconstruction error, For feature dimension, Input feature vector The One portion, For reconstructing vectors The One portion, For normal samples in the first The mean of each feature For the abnormal sample in the first The mean of each feature For normal samples in the first Standard deviation of each feature For the abnormal sample in the first Standard deviation over each feature;

[0073] The weighted mean square reconstruction error is compared with a preset threshold. When the weighted mean square reconstruction error is less than the preset threshold, the declared price is determined to be normal, and a comparison pass flag is output.

[0074] When the weighted mean square reconstruction error is greater than a preset threshold, the declared price is determined to be abnormal, and an abnormal alarm and abnormal type are output.

[0075] In one possible implementation, after a successful comparison, a reasonable budget result is output. Simultaneously, based on future maintenance fund expenditure forecasts over a predetermined period, the total amount of funds in the special account is optimally allocated to deposits of different maturities, including:

[0076] Based on the aforementioned reasonable budget results, the future maintenance fund expenditure value at a predetermined time, the current total amount of funds in the special account, and the annualized interest rates of fixed deposits of various terms, an objective function is constructed to maximize total return while penalizing liquidity risk. The calculation formula is as follows:

[0077]

[0078] Simultaneously, constraints are set for the total amount of funds and non-negative integers. The calculation formula is as follows:

[0079]

[0080] in, The objective function value, For the first The annualized interest rate of fixed deposits, For the first The amount allocated to fixed-term deposits, This is a liquidity risk penalty coefficient. For year indexing, Let be the projected expenditure amount for year t. To reach the first The total amount of deposits that matured before the end of the year, This represents the current total amount of funds in the special account;

[0081] The optimal configuration is obtained by solving the problem using an integer programming solver.

[0082] Secondly, embodiments of this application provide a machine learning-based intelligent budget calculation system for a maintenance fund management platform, comprising:

[0083] The acquisition module is used to acquire on-site survey images, descriptive texts, and archival data for maintenance projects;

[0084] The intelligent quantity verification module is used to input the acquired on-site survey images of the maintenance project into a pre-built convolutional neural network quantity verification model to generate standardized quantity data.

[0085] The cost prediction module is used to calculate the budget price of the maintenance project based on the standardized engineering quantity data and the archive data, and to predict the maintenance fund expenditure within a preset period in the future.

[0086] The matching analysis module is used to perform semantic parsing of the description text of maintenance projects using natural language processing technology, match policy clauses and calculate compliance risks. At the same time, it uses anomaly detection algorithms to perform horizontal comparison of prices in the description text.

[0087] The configuration module outputs a reasonable budget result after successful comparison, and optimizes the allocation of the total amount of special account funds for deposits of different terms by combining the future maintenance fund expenditure forecasts for a preset time.

[0088] Compared with the prior art, the beneficial effects of this application are as follows:

[0089] This application addresses the problems of traditional maintenance fund budget calculations relying on manual surveys and large errors in engineering quantity statistics. By using a convolutional neural network engineering quantity verification model, it achieves intelligent analysis of on-site survey images, accurately extracts engineering quantity data of damaged parts, has a high degree of standardization, effectively reduces human subjective errors, and improves the accuracy and efficiency of engineering quantity verification. At the same time, this application combines price indices and maintenance type feature vectors from archival data to correct the budget, making the budget price more in line with the actual market situation. Furthermore, it uses a time series forecasting model to achieve accurate medium- and long-term forecasts of maintenance fund expenditures, providing data support for financial planning.

[0090] This application utilizes natural language processing technology to perform semantic analysis of the descriptive text and intelligent matching of policy clauses, automatically calculating the compliance risk index to achieve intelligent and automated compliance review, significantly improving review efficiency and reducing compliance risks. It also uses an autoencoder anomaly detection algorithm combined with historical data from similar projects to conduct multi-dimensional comparisons of declared prices, accurately identifying price anomalies, effectively regulating maintenance fund application behavior, and ensuring the rationality of fund use.

[0091] This application constructs an objective function that maximizes total return and penalizes liquidity risk, and combines it with future expenditure forecasts to achieve optimal allocation of deposits with different maturities for special account funds. Under the premise of ensuring the liquidity of maintenance fund payments, it improves the investment returns of special account funds and achieves refined fund management. Attached Figure Description

[0092] Figure 1 A flowchart illustrating a machine learning-based intelligent budget calculation method for a maintenance fund management platform, as provided in this application embodiment;

[0093] Figure 2 A flowchart for generating standardized engineering quantity data using a convolutional neural network engineering quantity verification model provided in this application embodiment;

[0094] Figure 3 This is a structural diagram of the intelligent budget calculation system for the maintenance fund management platform based on machine learning, provided in an embodiment of this application. Detailed Implementation

[0095] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0096] Figure 1 A flowchart illustrating a machine learning-based intelligent budget calculation method for a maintenance fund management platform, as provided in this application embodiment, is shown below. Figure 1 This embodiment of a machine learning-based intelligent budget calculation method for a maintenance fund management platform includes:

[0097] S101, respectively acquires on-site survey images, descriptive texts and archival data for maintenance projects.

[0098] In this embodiment, a high-definition camera module of a mobile acquisition terminal captures multi-angle images of the damaged area at the repair project site. During acquisition, the EXIF ​​information of the images is recorded simultaneously, including image sensor parameters, shooting distance, lens focal length, shooting time, and geographical location information. The acquired images are structured and stored according to the naming rule of project number-damaged area-shooting angle, and uploaded to the image database of the maintenance fund management platform via an encrypted transmission protocol. The text input module or speech-to-text module of the mobile acquisition terminal collects the text describing the repair project submitted by the repair applicant, including information such as the repair object, repair scope, repair content, declared price, material specifications, construction process, and repair period. The text generated by speech-to-text is grammatically corrected and semantically completed. Field validation is performed on the entered text to ensure that no key information is missing before storing it in the text database of the maintenance fund management platform in a structured text format. From the archive database of the maintenance fund management platform, based on key identifiers such as project number, building number, building year, and apartment type, the platform extracts the corresponding basic archive data, historical maintenance archive data, quota unit price archive data, material price archive data, and special account fund archive data. Among them, the basic archive data includes the building structure type, building area, number of floors, and supporting facilities; the historical maintenance archive data includes the engineering quantity, budget price, actual expenditure amount, and construction period of similar maintenance projects; the quota unit price archive data includes the base period quota unit price of each sub-item; the material price archive data includes the base period price, current market price, and price index of various maintenance materials; and the special account fund archive data includes the total amount of special account funds, fund deposit and withdrawal records, and deposit interest rates for each term.

[0099] S102, the acquired on-site survey images of the maintenance project are input into the pre-built convolutional neural network engineering quantity verification model to generate standardized engineering quantity data.

[0100] In this embodiment, the convolutional neural network engineering quantity verification model includes: a backbone feature extraction network, which adopts ResNet101 with fused feature pyramids. Its input is used to receive preprocessed on-site survey images, and its output is connected to the input of the region proposal network. The region proposal network is connected to each level of the feature pyramid. Its input receives feature maps from each level of the feature pyramid, and its output generates candidate regions and their corresponding foreground or background classification scores and bounding box regression offsets. The output of the region proposal network is connected to the input of the ROI Align layer. The output of the ROI Align layer is connected to the input of the classification branch, bounding box regression branch, mask segmentation branch, and OCR digital extraction sub-network, respectively. The outputs of the classification branch, bounding box regression branch, mask segmentation branch, and OCR digital extraction sub-network are connected to the output layer. In addition, a pyramid coordinate channel attention module is introduced in the mask segmentation branch. By simultaneously modeling the inter-channel dependencies and spatial location importance, the attention to the target region is enhanced and background interference is suppressed.

[0101] The formula for calculating the loss function used when training the convolutional neural network engineering quantity verification model is as follows:

[0102]

[0103]

[0104]

[0105]

[0106]

[0107]

[0108] in, For the total loss function, These are the weighting coefficients for the classification loss. For classifying losses, These are the weighting coefficients for the bounding box regression loss. For bounding box regression loss, These are the weighting coefficients for the mask segmentation loss. For mask segmentation loss, These are the weighting coefficients for the OCR digit recognition loss. For OCR digital recognition loss, The number of positive samples. The total number of candidate regions. For the index of the candidate region, For category weighting factors, The model predicts the first The probability that each candidate region belongs to category t. To focus parameters, The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. To predict the Euclidean distance between the center point of the bounding box and the center point of the ground truth bounding box, For the parameters of the prediction box, The parameters are for the actual bounding box. Let be the diagonal length of the minimum bounding rectangle between the predicted bounding box and the ground truth bounding box. As a measure of aspect ratio consistency, and These are the height and width of the mask image, respectively. For pixel index, The model predicts the first The probability that a pixel belongs to a damaged area. For the first The real label of each pixel For smoothing terms, For balance coefficient, For pixel-level loss, This represents the total number of pixels. As a class balance factor, The length of the feature sequence. Output the character for the model at time step t. The probability, For all sequences that can be mapped to the target label sequence by merging duplicate characters and removing whitespace. The set of paths For one of the paths, it represents the character sequence output at each time step.

[0109] See Figure 2In this embodiment, the acquired on-site survey images are subjected to size normalization, color space conversion, and illumination equalization. Based on the image EXIF ​​information, the scale factor between image pixels and actual physical length is estimated to obtain a preprocessed image tensor. This preprocessed image tensor is then input into the backbone feature extraction network of the trained convolutional neural network engineering quantity verification model for multi-scale damage feature extraction, generating a multi-scale feature pyramid. At each level of the feature pyramid, a region proposal network generates dense anchor boxes. After foreground or background classification and bounding box regression, non-maximum suppression is used to filter candidate regions. Each candidate region is then processed using ROI (Region of Interest). The alignment layer operation maps to the corresponding level of the feature pyramid and samples them uniformly to a fixed size, resulting in aligned region feature maps. These aligned region feature maps are simultaneously input into the classification branch, bounding box regression branch, mask segmentation branch, and OCR digit extraction subnetwork, respectively outputting the damage type and level, the precise bounding box coordinates of the damaged area, the pixel-level binary mask of the damaged area, and the size annotations in the image. The pixel-level binary mask output by the mask segmentation branch is used for pixel counting, and the damaged area is calculated using a scale factor. The damaged length is calculated using a skeleton extraction algorithm, and the damaged width or depth is estimated based on the minimum bounding rectangle of the mask. The calculation formulas are as follows:

[0110]

[0111]

[0112]

[0113] in, For the damaged area, and These are the height and width of the mask image, respectively. For the first The mask value of each pixel. For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. The physical width of the image sensor, For shooting distance, For the lens focal length, The pixel width of the image. For the length of damage, This refers to the set of single-pixel wide centerline pixels obtained after skeleton extraction from the mask. To determine the width or depth of damage, the damage type and level, the precise bounding box coordinates of the damaged area, the damaged area, the damaged length, the damaged width, and the dimension annotations in the image are merged to obtain standardized engineering quantity data.

[0114] S103 calculates the budget price of maintenance projects based on standardized engineering quantity data and archival data, and forecasts maintenance fund expenditures within a preset future period.

[0115] In this embodiment, the base period quota unit price, real-time material market unit price, and on-site measures fee matching the maintenance project are extracted from the archive data. Based on the standardized engineering quantity data, combined with the base period quota unit price, real-time material market unit price, and on-site measures fee, the benchmark cost of the current maintenance project is calculated. The calculation formula is as follows:

[0116]

[0117] in, This is the base cost for the current repair project. Based on the quantities of each sub-item of the standardized engineering quantity data, The benchmark unit price, For the first The weighting coefficient of each material in the total cost The current price index, The price index for the base period, For the number of material types, For the quantity of sub-items, Cost of on-site measures.

[0118] Because real-world repair processes involve occasional repair markups, emergency repair surcharges, and bulk purchase discounts, a non-linear correction is needed to the aforementioned baseline architecture. In this embodiment, based on the repair type feature vector, the baseline cost is corrected using an empirical correction coefficient to obtain the final budget price. The calculation formula is as follows:

[0119]

[0120]

[0121] in, For the final budget price, For the maintenance type feature vector, This is a non-linear adjustment term based on the quantity of work. To adjust the coefficient, For the first Weight of class experience, For regular functions, For the empirical gradient term, For specific feature values ​​extracted from the feature vector, such as a repair urgency score, and These are linear coefficients.

[0122] Simultaneously, based on the building feature vector and historical actual expenditure values, a time-series prediction model is used to generate future maintenance fund expenditures for a preset period. The calculation formula is as follows:

[0123]

[0124] in, For the predicted first Maintenance fund expenditure value, Let be the i-th empirical gradient term, reflecting the degradation trend based on house characteristics. No. The weight coefficients of each empirical gradient term, Indicated based on house characteristics The Degradation trend or gradient of maintenance needs Let be the house feature vector in period t. For the first Autoregressive coefficients of order 1 For the first Historical actual maintenance expenditure value during the period For the first Moving average coefficient, For the first Historical residuals of the period These are the weighting coefficients of the nonlinear feature mapping term. For nonlinear feature mapping terms based on house features, and These are the autoregression order and the moving average order, respectively.

[0125] S104 uses natural language processing technology to semantically analyze the description text of the maintenance project, match policy clauses and calculate compliance risks. At the same time, it uses anomaly detection algorithms to perform horizontal comparison of prices in the description text.

[0126] In this embodiment, the description text of the maintenance project is segmented, part-of-speech tagging is performed, and dependency parsing is conducted to generate a preprocessed text sequence. Simultaneously, the TF-IDF algorithm is used to calculate the weight of each word. The text sequence is then input into an entity recognition model that integrates a professional dictionary and rule base for the maintenance funds domain. This model identifies and extracts the maintenance object, maintenance scope, material specifications, and monetary numerical entities from the text, generating an entity set. Each identified entity text is encoded to obtain an entity embedding vector. Furthermore, the weight of each entity is calculated by combining the corresponding word weights. The calculation formula is as follows:

[0127]

[0128] in, For the first The overall weight of each entity The first output of the entity recognition model The confidence score of each entity. For the first The number of words contained in an entity text. This is the sum of the TF-IDF weights of all words in the entity text. A weighted query vector is constructed based on the weight of each entity and its entity embedding vector, calculated using the following formula:

[0129]

[0130] in, For weighted query vectors, This represents the total number of entities identified from the text. For the first The embedding vector of each entity is generated by pre-encoding the text of each clause in the policy clause library, creating a clause embedding vector, and building a nearest neighbor index. The weighted query vector is then input into the matching engine to execute the matching clause and calculate compliance risks.

[0131] In this embodiment, the entity text in the entity set is precisely matched with the prohibited keywords in the policy clause library. If no keyword matching is triggered during the matching process, a preset number of clauses most similar to the weighted query vector are retrieved from the index. The weighted cosine similarity between the weighted query vector and the embedding vector of each candidate clause is calculated, and the clause with the highest similarity score and the maximum semantic similarity value are output. The calculation formula is as follows:

[0132]

[0133] in, For weighted query vector and the first The cosine similarity of the embedding vectors of the policy clauses. For weighted query vectors, For the first The embedding vector of each policy clause Let L2 be the norm of the vector;

[0134] Based on the maximum semantic similarity value and the semantic distance between the entity type and the applicable clause type, the compliance risk index is calculated using the following formula:

[0135]

[0136] in, This is a compliance risk index. The maximum semantic similarity value. The attenuation coefficient is... Let E be the semantic distance between the entity type and the applicable type of the matched terms, and let E be the set of entities. The policy clause number that is most similar to the weighted query vector is used. If a match is found during the matching process, the compliance risk index is directly output and the matched clause number is recorded.

[0137] Simultaneously, an anomaly detection algorithm is used to perform horizontal comparison of prices in the description text. In this embodiment, the declared price is extracted from the description text of the maintenance project, and combined with the calculated budget price, standardized engineering quantity data, and historical price information of similar projects in the archive data, an input feature vector is constructed. The input feature vector is then input into a pre-trained autoencoder to obtain the hidden layer representation, and finally reconstructed by the decoder to obtain the output vector. The calculation formula is as follows:

[0138]

[0139]

[0140] in, This is the hidden layer representation output by the encoder. For the input feature vector, For encoder functions, the input feature vector is... Mapped to the hidden space, For activation function, Here is the weight matrix of the encoder. This is the bias vector of the encoder. The output vector reconstructed by the decoder, For decoder functions, Here is the weight matrix of the decoder. Given the decoder's bias vector, the weighted mean square reconstruction error is calculated based on the output vector using the following formula:

[0141]

[0142] in, For weighted mean square reconstruction error, For feature dimension, Input feature vector The One portion, For reconstructing vectors The One portion, For normal samples in the first The mean of each feature For the abnormal sample in the first The mean of each feature For normal samples in the first Standard deviation of each feature For the abnormal sample in the first The standard deviation of each feature is used to compare the weighted mean square reconstruction error with a preset threshold. When the weighted mean square reconstruction error is less than the preset threshold, the declared price is determined to be normal, and a comparison pass flag is output. When the weighted mean square reconstruction error is greater than the preset threshold, the declared price is determined to be abnormal, and an abnormal alarm and abnormal type are output.

[0143] S105, after successful comparison, outputs reasonable budget results, and at the same time, combines the future maintenance fund expenditure forecast to optimize the allocation of the total amount of special account funds for deposits of different terms.

[0144] In this embodiment, an objective function is constructed to maximize total return and penalize liquidity risk based on the reasonable budget results, the maintenance fund expenditure value for the next 1-5 years, the current total amount of funds in the special account, and the annualized interest rate of time deposits of various terms. The calculation formula is as follows:

[0145]

[0146] Simultaneously, constraints are set for the total amount of funds and non-negative integers. The calculation formula is as follows:

[0147]

[0148] Where i=1,2,3,4 correspond to 7-day, 1-year, 3-year, and 5-year fixed deposits, respectively, representing the deposit terms. The respective durations are 0.23 months, 12 months, 36 months, and 60 months. The objective function value, For the first The annualized interest rate of fixed deposits, For the first The amount allocated to fixed-term deposits, This is a liquidity risk penalty coefficient. For year indexing, Let be the projected expenditure amount for year t. To reach the first The total amount of deposits that matured before the end of the year, Given the current total amount of funds in the special account, the optimal allocation is obtained by solving the problem using an integer programming solver.

[0149] Corresponding to the above embodiment of the intelligent budget calculation method for a maintenance fund management platform based on machine learning, this application also provides an embodiment of an intelligent budget calculation system for a maintenance fund management platform based on machine learning.

[0150] See Figure 3 This application provides a machine learning-based maintenance fund management platform budget intelligent calculation system 20, comprising:

[0151] The acquisition module 201 is used to acquire on-site survey images, descriptive texts, and archival data for maintenance projects.

[0152] The intelligent quantity verification module 202 is used to input the acquired on-site survey images of the maintenance project into a pre-built convolutional neural network quantity verification model to generate standardized quantity data.

[0153] The cost forecasting module 203 is used to calculate the budget price of maintenance projects based on standardized engineering quantity data and archival data, and to forecast maintenance fund expenditures within a preset period.

[0154] The matching analysis module 204 is used to perform semantic parsing on the description text of the maintenance project using natural language processing technology, match policy clauses and calculate compliance risks, and at the same time, use anomaly detection algorithms to perform horizontal comparison of prices in the description text.

[0155] Configuration module 205 outputs a reasonable budget result after successful comparison, and optimizes the allocation of the total amount of special account funds for deposits of different terms by combining the future maintenance fund expenditure forecasts for a preset time.

[0156] In this embodiment, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0157] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0158] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A machine learning-based intelligent budget calculation method for a maintenance fund management platform, characterized in that, include: The site survey images, descriptive texts, and archival data for the maintenance project were acquired respectively. The acquired on-site survey images of the maintenance project are input into a pre-built convolutional neural network engineering quantity verification model, which outputs the damage type and level, the precise bounding box coordinates of the damaged area, the pixel-level binary mask of the damaged area, and the size annotations in the image. The damage type and level, the precise bounding box coordinates of the damaged area, the damaged area, the damaged length, the damaged width, and the size annotations in the image are fused to obtain standardized engineering quantity data. The budget price of the maintenance project is calculated based on the standardized engineering quantity data and the archive data, and the maintenance fund expenditure in the future preset period is predicted. Natural language processing technology is used to semantically analyze the description text of maintenance projects, match it with policy clauses, and calculate compliance risks, including: The description text of the maintenance project is segmented, part-of-speech tagging and dependency parsing are performed to generate a preprocessed text sequence. At the same time, the weight of each word is calculated using the TF-IDF algorithm. The text sequence is input into an entity recognition model that integrates a professional dictionary and rule base in the field of maintenance funds to identify and extract maintenance objects, maintenance scope, material specifications, and monetary entities from the text, and generate an entity set. Each identified entity text is encoded to obtain an entity embedding vector. Simultaneously, the weight of each entity is calculated by combining the corresponding word weights. The calculation formula is as follows: in, For the first The overall weight of each entity The first output of the entity recognition model The confidence score of each entity. For the first The number of words contained in an entity text. This is the sum of the TF-IDF weights of all words in the entity text; A weighted query vector is constructed based on the weight of each entity and its embedding vector. The calculation formula is as follows: in, For weighted query vectors, This represents the total number of entities identified from the text. For the first Embedding vectors of each entity; Each clause text in the policy clause library is pre-encoded to generate a clause embedding vector, and a nearest neighbor index is built. The weighted query vector is then input into the matching engine to perform clause matching and calculate compliance risks. Simultaneously, anomaly detection algorithms are used to perform horizontal comparisons of prices in the descriptive text, including: Extract the declared price from the description text of the maintenance project, and combine it with the calculated budget price, standardized engineering quantity data, and historical price information of similar projects in the archive data to construct an input feature vector; The input feature vector is fed into a pre-trained autoencoder to obtain the hidden layer representation, which is then reconstructed by the decoder to obtain the output vector. The calculation formula is as follows: in, This is the hidden layer representation output by the encoder. For the input feature vector, For encoder functions, the input feature vector is... Mapped to the hidden space, For activation function, Here is the weight matrix of the encoder. This is the bias vector of the encoder. The output vector reconstructed by the decoder, For decoder functions, Here is the weight matrix of the decoder. This is the bias vector for the decoder; The weighted mean square reconstruction error is calculated based on the output vector, using the following formula: in, For weighted mean square reconstruction error, For feature dimension, Input feature vector The One portion, For reconstructing vectors The One portion, For normal samples in the first The mean of each feature For the abnormal sample in the first The mean of each feature For normal samples in the first Standard deviation of each feature For the abnormal sample in the first Standard deviation over each feature; The weighted mean square reconstruction error is compared with a preset threshold. When the weighted mean square reconstruction error is less than the preset threshold, the declared price is determined to be normal, and a comparison pass flag is output. When the weighted mean square reconstruction error is greater than a preset threshold, the declared price is determined to be abnormal, and an abnormal alarm and abnormal type are output. After successful comparison, a reasonable budget result is output. At the same time, the total amount of funds in the special account is optimally allocated to deposits of different terms based on the future maintenance fund expenditure forecast at a preset time.

2. The intelligent budget calculation method for a maintenance fund management platform based on machine learning as described in claim 1, characterized in that, The convolutional neural network engineering quantity verification model includes: a backbone feature extraction network, the input of which receives preprocessed on-site survey images; the output of which is connected to the input of a region proposal network; the output of which is connected to the input of an ROI Align layer; the output of which is connected to the inputs of a classification branch, a bounding box regression branch, a mask segmentation branch, and an OCR digital extraction sub-network; and the outputs of which are connected to an output layer. The loss function used during training of the convolutional neural network engineering quantity verification model is calculated using the following formula: in, For the total loss function, These are the weighting coefficients for the classification loss. For classifying losses, These are the weighting coefficients for the bounding box regression loss. For bounding box regression loss, These are the weighting coefficients for the mask segmentation loss. For mask segmentation loss, These are the weighting coefficients for the OCR digit recognition loss. For OCR digital recognition loss, The number of positive samples. The total number of candidate regions. For the index of the candidate region, For category weighting factors, The model predicts the first The probability that each candidate region belongs to category t. To focus parameters, The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. To predict the Euclidean distance between the center point of the bounding box and the center point of the ground truth bounding box, For the parameters of the prediction box, The parameters are for the actual bounding box. Let be the diagonal length of the minimum bounding rectangle between the predicted bounding box and the ground truth bounding box. As a measure of aspect ratio consistency, and These are the height and width of the mask image, respectively. For pixel index, The model predicts the first The probability that a pixel belongs to a damaged area. For the first The real label of each pixel For smoothing terms, For balance coefficient, For pixel-level loss, This represents the total number of pixels. As a class balance factor, The length of the feature sequence. Output the character for the model at time step t. The probability, For all sequences that can be mapped to the target label sequence by merging duplicate characters and removing whitespace. The set of paths For one of the paths, it represents the character sequence output at each time step.

3. The intelligent budget calculation method for a maintenance fund management platform based on machine learning as described in claim 1, characterized in that, The acquired on-site survey images of the repair project are input into a pre-constructed convolutional neural network engineering quantity verification model. The model outputs the damage type and level, the precise bounding box coordinates of the damaged area, the pixel-level binary mask of the damaged area, and the dimension annotations in the image. The damage type and level, the precise bounding box coordinates of the damaged area, the damaged area, the damaged length, the damaged width, and the dimension annotations in the image are then fused to obtain standardized engineering quantity data, including: The acquired field survey images are normalized in size, converted in color space and balanced in illumination. The scale factor of the image pixels and the actual physical length is estimated based on the image EXIF ​​information to obtain the preprocessed image tensor. The preprocessed image tensor is input into the backbone feature extraction network of the convolutional neural network engineering quantity verification model to extract multi-scale damage features and generate a multi-scale feature pyramid. Dense anchor boxes are generated at each level of the feature pyramid through a region proposal network. After foreground or background classification and bounding box regression, candidate regions are selected by non-maximum suppression. Each candidate region is mapped to the corresponding level of the feature pyramid through the ROI Align layer operation, and then uniformly sampled to a fixed size to obtain the aligned region feature map; The aligned region feature maps are simultaneously input into the classification branch, bounding box regression branch, mask segmentation branch and OCR digital extraction subnetwork, and output the damage type and level, the precise bounding box coordinates of the damaged area, the pixel-level binary mask of the damaged area and the size annotation in the image, respectively. Pixel counts are performed on the pixel-level binary mask output from the mask segmentation branch. The damaged area is calculated using a scale factor, the damaged length is calculated using a skeleton extraction algorithm, and the damaged width or depth is estimated based on the minimum bounding rectangle of the mask. The calculation formulas are as follows: in, For the damaged area, and These are the height and width of the mask image, respectively. For the first The mask value of each pixel. For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. The physical width of the image sensor, For shooting distance, For the lens focal length, The pixel width of the image. For the length of damage, This refers to the set of single-pixel wide centerline pixels obtained after skeleton extraction from the mask. Damage to width or depth; The damage type and level, the precise bounding box coordinates of the damaged area, the damaged area, the damaged length, the damaged width, and the dimension annotations in the image are fused to obtain standardized engineering quantity data.

4. The intelligent budget calculation method for the maintenance fund management platform based on machine learning according to claim 1, characterized in that, The budget price for the maintenance project is calculated based on the standardized engineering quantity data and the archival data. Simultaneously, maintenance expenditures within a predetermined future period are predicted, including: Extract the base period quota unit price, real-time material market unit price, and on-site measures fee that match the maintenance project from the archive data; Based on the standardized engineering quantity data, combined with the base period quota unit price, real-time material market unit price, and on-site measures fee, the benchmark cost of the current maintenance project is calculated using the following formula: in, This is the base cost for the current repair project. Based on the quantities of each sub-item of the standardized engineering quantity data, The benchmark unit price, For the first The weighting coefficient of each material in the total cost The current price index, The price index for the base period, For the number of material types, For the quantity of sub-items, For on-site measures costs; Based on the repair type feature vector, the baseline cost is adjusted using an empirical correction coefficient to obtain the final budget price. The calculation formula is as follows: in, For the final budget price, For the maintenance type feature vector, This is a non-linear adjustment term based on the quantity of work. To adjust the coefficient, For the first Weight of class experience, For regular functions, For empirical gradient terms; At the same time, the expenditure on maintenance funds in the future is predicted within a predetermined period.

5. The intelligent budget calculation method for the maintenance fund management platform based on machine learning according to claim 4, characterized in that, The forecasting of maintenance fund expenditures within a predetermined future period includes: Based on the building feature vector and historical actual expenditure values, a time-series prediction model is used to generate maintenance fund expenditures for a preset future period. The calculation formula is as follows: in, For the predicted first Maintenance fund expenditure value, Let be the i-th empirical gradient term, reflecting the degradation trend based on house characteristics. No. The weight coefficients of each empirical gradient term, Indicated based on house characteristics The Degradation trend or gradient of maintenance needs Let be the house feature vector in period t. For the first Autoregressive coefficients of order 1 For the first Historical actual maintenance expenditure value during the period For the first Moving average coefficient, For the first Historical residuals of the period These are the weighting coefficients of the nonlinear feature mapping term. For nonlinear feature mapping terms based on house features, and These are the autoregression order and the moving average order, respectively.

6. The intelligent budget calculation method for a maintenance fund management platform based on machine learning according to claim 1, characterized in that, The weighted query vector is input into the matching engine to execute matching terms and calculate compliance risks, including: The entity text in the entity set is precisely matched with the prohibited keywords in the policy clause library. If no keyword matching is triggered during the matching process, a preset number of clauses most similar to the weighted query vector are retrieved from the index. The weighted cosine similarity between the weighted query vector and the embedding vector of each candidate clause is calculated, and the clause with the highest similarity score and the maximum semantic similarity value are output. The calculation formula is as follows: in, For weighted query vector and the first The cosine similarity of the embedding vectors of the policy clauses. For weighted query vectors, For the first The embedding vector of each policy clause Let L2 be the norm of the vector; Based on the maximum semantic similarity value, and combined with the semantic distance between the entity type and the applicable clause type, the compliance risk index is calculated using the following formula: in, This is a compliance risk index. The maximum semantic similarity value. The attenuation coefficient is... Let E be the semantic distance between the entity type and the applicable type of the matched terms, and let E be the set of entities. The policy clause number that is most similar to the weighted query vector; If a match is found during the matching process, the compliance risk index will be output directly, and the matching clause number will be recorded.

7. The intelligent budget calculation method for a maintenance fund management platform based on machine learning according to claim 1, characterized in that, After successful comparison, a reasonable budget result is output. Simultaneously, based on future maintenance fund expenditure forecasts over a preset time period, the optimal allocation of funds in the special account for deposits of different maturities is determined, including: Based on the aforementioned reasonable budget results, the future maintenance fund expenditure value at a predetermined time, the current total amount of funds in the special account, and the annualized interest rates of fixed deposits of various terms, an objective function is constructed to maximize total return while penalizing liquidity risk. The calculation formula is as follows: Simultaneously, constraints are set for the total amount of funds and non-negative integers. The calculation formula is as follows: in, The objective function value, For the first The annualized interest rate of fixed deposits, For the first The amount allocated to fixed-term deposits, This is a liquidity risk penalty coefficient. For year indexing, Let be the projected expenditure amount for year t. To reach the first The total amount of deposits that matured before the end of the year, This represents the current total amount of funds in the special account; The optimal configuration is obtained by solving the problem using an integer programming solver.

8. A machine learning-based intelligent budget calculation system for a maintenance fund management platform, implemented using the machine learning-based intelligent budget calculation method for a maintenance fund management platform as described in any one of claims 1-7, characterized in that, include: The acquisition module is used to acquire on-site survey images, descriptive texts, and archival data for maintenance projects; The intelligent engineering quantity verification module is used to input the acquired on-site survey images of the maintenance project into a pre-built convolutional neural network engineering quantity verification model, and output the damage type and level, the precise bounding box coordinates of the damaged area, the pixel-level binary mask of the damaged area, and the size annotations in the image. The damage type and level, the precise bounding box coordinates of the damaged area, the damaged area, the damaged length, the damaged width, and the size annotations in the image are fused to obtain standardized engineering quantity data. The cost prediction module is used to calculate the budget price of the maintenance project based on the standardized engineering quantity data and the archive data, and to predict the maintenance fund expenditure within a preset period in the future. The matching analysis module uses natural language processing technology to semantically parse the descriptive text of maintenance projects, match it with policy clauses, and calculate compliance risks, including: The description text of the maintenance project is segmented, part-of-speech tagging and dependency parsing are performed to generate a preprocessed text sequence. At the same time, the weight of each word is calculated using the TF-IDF algorithm. The text sequence is input into an entity recognition model that integrates a professional dictionary and rule base in the field of maintenance funds to identify and extract maintenance objects, maintenance scope, material specifications, and monetary entities from the text, and generate an entity set. Each identified entity text is encoded to obtain an entity embedding vector. Simultaneously, the weight of each entity is calculated by combining the corresponding word weights. The calculation formula is as follows: in, For the first The overall weight of each entity The first output of the entity recognition model The confidence score of each entity. For the first The number of words contained in an entity text. This is the sum of the TF-IDF weights of all words in the entity text; A weighted query vector is constructed based on the weight of each entity and its embedding vector. The calculation formula is as follows: in, For weighted query vectors, This represents the total number of entities identified from the text. For the first Embedding vectors of each entity; Each clause text in the policy clause library is pre-encoded to generate a clause embedding vector, and a nearest neighbor index is built. The weighted query vector is then input into the matching engine to perform clause matching and calculate compliance risks. Simultaneously, anomaly detection algorithms are used to perform horizontal comparisons of prices in the descriptive text, including: Extract the declared price from the description text of the maintenance project, and combine it with the calculated budget price, standardized engineering quantity data, and historical price information of similar projects in the archive data to construct an input feature vector; The input feature vector is fed into a pre-trained autoencoder to obtain the hidden layer representation, which is then reconstructed by the decoder to obtain the output vector. The calculation formula is as follows: in, This is the hidden layer representation output by the encoder. For the input feature vector, For encoder functions, the input feature vector is... Mapped to the hidden space, For activation function, Here is the weight matrix of the encoder. This is the bias vector of the encoder. The output vector reconstructed by the decoder, For decoder functions, Here is the weight matrix of the decoder. This is the bias vector for the decoder; The weighted mean square reconstruction error is calculated based on the output vector, using the following formula: in, For weighted mean square reconstruction error, For feature dimension, Input feature vector The One portion, For reconstructing vectors The One portion, For normal samples in the first The mean of each feature For the abnormal sample in the first The mean of each feature For normal samples in the first Standard deviation of each feature For the abnormal sample in the first Standard deviation over each feature; The weighted mean square reconstruction error is compared with a preset threshold. When the weighted mean square reconstruction error is less than the preset threshold, the declared price is determined to be normal, and a comparison pass flag is output. When the weighted mean square reconstruction error is greater than a preset threshold, the declared price is determined to be abnormal, and an abnormal alarm and abnormal type are output. The configuration module outputs a reasonable budget result after successful comparison, and optimizes the allocation of the total amount of special account funds for deposits of different terms by combining the future maintenance fund expenditure forecasts for a preset time.

Images