Quality of service assessment system

Abstract

The application relates to the field of engineering consulting service quality control, and discloses a service quality evaluation system, which comprises a data acquisition module, a dynamic topology construction module, a double-channel evaluation module, an improvement suggestion generation module and an evaluation report generation module. By constructing a dynamic topology network with engineering documents as nodes and multi-dimensional correlations as edges, combining real-time compliance detection channels with deep semantic matching and deep evolution analysis channels with quality state differential equation prediction, quantitative evaluation of standard compliance and analysis of quality evolution trend are realized. Based on the abnormal propagation path detection of the topology gradient field, an industry knowledge graph is generated to optimize suggestions, and through dynamic correlation mapping technology, multi-source evaluation data is integrated into a standardized report. The application improves the structuralization, dynamization and intellectualization level of engineering consulting service quality evaluation, and has good practicability and engineering adaptation ability.

Description

Service quality assessment system Technical Field

[0001] This invention relates to the field of quality control technology for engineering consulting services, and in particular to a service quality assessment system. Background Technology

[0002] In the field of engineering consulting, service quality assessment is a core element in ensuring project implementation effectiveness, but traditional assessment methods suffer from significant technical bottlenecks. Existing technologies often combine manual review with static rule matching, which struggles to effectively handle the complex relationships between the massive amounts of heterogeneous documents in modern engineering consulting projects. This results in assessments being singular in scope and lacking dynamic adaptability. Especially when facing frequent design changes and iterative requirements, conventional systems cannot capture the topological evolution of document relationship networks in real time, leading to lags in quality assessments and insufficient operability of decision-making recommendations.

[0003] In addition, existing automated assessment tools are mostly limited to the analysis of a single data source and lack the ability to conduct collaborative analysis of multi-dimensional quality elements such as the completeness of the reference to normative clauses and the evolution trajectory of document versions, resulting in fragmented assessment results that are difficult to trace and verify.

[0004] These technical deficiencies severely restrict the accuracy and timeliness of quality control in engineering consulting services, making it difficult to meet the full-cycle management needs of new smart engineering construction. Summary of the Invention

[0005] The purpose of this invention is to provide a service quality assessment system that solves the problems of lack of cross-document correlation analysis, insufficient dynamic adaptability, and fragmented assessment results in traditional engineering consulting quality assessment methods.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The service quality assessment system includes:

[0008] The data acquisition module is configured to automatically collect structured documents from the entire lifecycle of engineering consulting projects, including contract texts, design drawings, and meeting minutes.

[0009] The dynamic topology construction module is configured to receive structured documents collected by the data acquisition module, construct a dynamic topology network with individual documents as nodes and multidimensional relationships between documents as edges, and dynamically calculate edge weights based on document semantic relevance, temporal continuity, and normative citation strength.

[0010] The dual-channel evaluation module, connected to the dynamic topology construction module, includes a real-time compliance detection channel and a deep evolution analysis channel, wherein:

[0011] The real-time compliance detection channel is configured to perform deep semantic matching between structured documents and a pre-built industry standard database to generate a standard compliance score for the structured documents.

[0012] The deep evolution analysis channel is configured to construct a quality state differential equation based on the dynamic topology network and output the service quality evolution trend prediction result.

[0013] The improvement suggestion generation module is configured to receive the specification compliance score and service quality evolution trend prediction results, identify abnormal propagation paths in the dynamic topology network by constructing the topological gradient field, and generate improvement suggestions for service quality optimization by combining industry knowledge graphs.

[0014] The assessment report generation module is configured to integrate the compliance score, service quality evolution trend prediction results, and improvement suggestions to output a structured assessment report.

[0015] Preferably, the data acquisition module includes:

[0016] The data acquisition module includes:

[0017] The document type recognition unit is configured to automatically distinguish between three types of documents: contract text, design drawings, and meeting minutes using a convolutional neural network model.

[0018] The streaming processing unit is configured to perform incremental data acquisition using a dynamic window mechanism; when the accumulated amount of acquired data reaches a set threshold, document type recognition and data standardization processing are triggered.

[0019] Preferably, the dynamic topology building module performs:

[0020] The dynamic topology building module is configured to execute:

[0021] Define a set of nodes V = {v k |v k ∈Contract terms entity, design drawings entity, meeting minutes entity};

[0022] A dynamic edge weight function is constructed to calculate the edge weight between nodes. The edge weight is determined by a weighted combination of semantic similarity between documents, proximity of document generation time, and overlap of commonly referenced normative clauses.

[0023] Preferably, the dynamic edge weight function is specifically:

[0024] ;

[0025] in, ; Document generation time interval; This represents the time decay factor, based on the total project duration. set up; and Representing documents respectively and The set of referenced normative clauses; This represents the complete set of pre-built industry standards databases; Indicates the proximity of document generation times; Indicates the percentage overlap of commonly cited normative clauses within the entire normative database set; These are the corresponding weight parameters, used to adjust the influence of semantic similarity, temporal proximity, and clause overlap on edge weights, respectively.

[0026] Preferably, the real-time compliance detection channel includes:

[0027] The deep semantic matching unit is configured to use a pre-trained language model to compute documents. With the terms and conditions Semantic similarity:

[0028] ;

[0029] in, Representing documents respectively With the terms and conditions semantic vectors, For trainable weight matrix, For the sigmoid function, For bias terms;

[0030] The clause integrity verification unit is configured to statistically analyze the coverage rate of clauses cited in structured documents, and the coverage rate is calculated using the following formula:

[0031] ;

[0032] in, Document The actual set of normative clauses referenced; This represents a predefined subset of specifications that must be followed based on the project type;

[0033] The comprehensive scoring unit is configured to calculate documents based on semantic similarity and coverage. The standard compliance score is calculated using the following formula:

[0034] ;

[0035] in, For document Standard compliance score; This represents the semantic matching weight coefficient.

[0036] Preferably, the differential equation of mass state is expressed as:

[0037] ;

[0038] in, The schedule deviation rate is expressed as (actual progress - planned progress) / planned progress. This indicates the cost overrun rate, calculated as actual cost / budgeted cost. The defect rate is expressed as the number of defective items detected divided by the total number of items detected. This represents the impact factor of design changes, and its value is positively correlated with the level of change. This represents the impact factor of demand changes, and its value is positively correlated with the frequency of changes. Represents the time-varying state transition matrix; This represents the time-varying input coupling matrix.

[0039] Preferably, the improvement suggestion generation module includes:

[0040] The mass gradient field construction unit is configured to build the mass influence matrix:

[0041] ;

[0042] in, This represents the comprehensive quality scoring function, which is calculated as the weighted average of the progress score, cost control score, and quality inspection score of the tasks associated with the document node. Represents nodes in a dynamic topology network and The edge weights; This indicates the total number of document nodes in the current network topology.

[0043] The anomaly propagation analysis unit is configured to detect key impact paths based on the following formula:

[0044] ;

[0045] in, This is a key anomaly propagation path. This indicates the document node from which an anomaly was detected. To key delivery milestones The set of all connected paths; This represents the rate of change of the edge weights over time.

[0046] It is recommended that the generation unit be configured based on the aforementioned key impact path. Based on the normative dependencies and empirical rules in the industry knowledge graph, optimization suggestions are generated for path nodes.

[0047] Preferably, the anomaly propagation analysis unit is further configured as follows:

[0048] Solving using an improved Dijkstra algorithm The path weight is defined as:

[0049] ;

[0050] in, Represents the edges in the quality gradient matrix Element; This represents the rate of change of edge weights over time; This represents a path from the starting node to the ending node.

[0051] The algorithm terminates when any of the following thresholds are met:

[0052] Path length ;in, Indicates the preset maximum path length;

[0053] Cumulative weight ;in, This represents the critical threshold for path influence.

[0054] Preferably, the evaluation report generation module is configured to execute:

[0055] Receive and integrate the compliance score and service quality evolution trend prediction results from the dual-channel evaluation module, as well as the improvement suggestions from the improvement suggestion generation module;

[0056] Based on the node connection relationships of the dynamic topology network, establish a correlation mapping between evaluation indicators;

[0057] The fusion results and correlation mappings are converted into a standardized report format that can be parsed by machines according to a preset template;

[0058] When the rate of change of any side weight in the dynamic topology network exceeds a set threshold, an incremental report update is automatically initiated.

[0059] In summary, the present invention has at least one of the following beneficial technical effects:

[0060] 1. This invention utilizes dynamic topology network modeling technology to construct a dynamic network structure with semantic, temporal, and normative associations for multi-source engineering documents such as contract texts and design drawings according to their node connection relationships, thereby supporting the establishment of association mappings between evaluation indicators. This structure overcomes the limitations of traditional evaluation methods that suffer from scattered documents and weak associations, providing a structural foundation for fusion processing and traceable report output in the evaluation report generation module, and realizing unified modeling and orderly association between evaluation elements across documents.

[0061] 2. This invention configures a real-time compliance detection channel and a deep evolutionary analysis channel in the dual-channel assessment module. The former generates a compliance score through semantic matching, while the latter predicts the service quality evolution trend based on state differential equations. The assessment report generation module integrates the two types of results and generates a standardized report accordingly. This allows the system to accurately record the current compliance status and dynamically adjust the assessment results before changes in trend indicators occur, thus providing a foundation for downstream improvement suggestion generation and report updates.

[0062] 3. In the improvement suggestion generation module, this invention generates a hierarchical suggestion chain through quality gradient field analysis and knowledge graph reasoning, and then passes this chain as input to the evaluation report generation module. This suggestion chain is not only presented in a structured manner in the main body of the report, but also achieves traceable path location through indicator-node association mapping. In this way, the system can dynamically label quality anomalies and intelligently recommend related document nodes, avoiding the problems of vagueness and lack of specificity in traditional improvement strategies.

[0063] 4. This invention achieves structural binding and version change awareness between indicators and original document nodes by configuring a dynamic association coding mechanism and an incremental update strategy in the evaluation report generation module. During report generation, local indicator recalculation is triggered based on the association matrix and edge weight change rate, outputting only the differing parts and marking the changed nodes. This effectively reduces the evaluation recalculation cost in scenarios with high-frequency document changes, ensuring the real-time nature and version consistency of the report content. Attached Figure Description

[0064] Figure 1 is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0065] The present invention will be further described in detail below with reference to the accompanying drawings.

[0066] This invention provides a service quality assessment system that integrates dynamic topology analysis and a dual-channel assessment mechanism to achieve quantitative assessment and decision support for service quality throughout the entire lifecycle of engineering consulting projects.

[0067] As shown in Figure 1, the service quality assessment system includes a data acquisition module, a dynamic topology construction module, a dual-channel assessment module, an improvement suggestion generation module, and an assessment report generation module.

[0068] The following is a detailed description of each module in the system of this invention.

[0069] In this embodiment, the data acquisition module achieves full-cycle acquisition and preprocessing of engineering documents through multimodal data processing and adaptive acquisition strategies. The acquired document types include contract texts, design drawings, and meeting minutes. This module combines deep neural networks and streaming technology to construct a complete technical chain from raw data input to standardized output, ensuring both acquisition efficiency and data quality.

[0070] In terms of document type recognition, the data acquisition module processes data through a feature extraction layer and a classification decision layer based on a convolutional neural network (CNN).

[0071] Preferably, a ResNet structure is employed to enhance the multi-scale feature fusion capability of the input document. Specifically, after the input document passes through multiple convolutional neural networks, the feature map is reduced to the category discrimination space via a global average pooling layer. During this process, geometric contour feature vectors are extracted from scanned image-format design drawings using an edge detection operator, while for text documents, an attention mechanism is used to capture semantic association patterns between paragraphs. The extracted feature data is then processed through the output layer of the classifier, using a softmax function to generate the document type probability distribution. Its decision function can be expressed as:

[0072] ;

[0073] in, This represents the high-dimensional representation vector output by the feature extraction network, where K=3 corresponds to three document categories: contract text, design drawings, and meeting minutes. This design effectively addresses the insufficient generalization ability of traditional rule-matching methods in fuzzy document classification, improving the accuracy and robustness of document classification.

[0074] In streaming data processing, the data acquisition module employs a dynamic windowing mechanism to achieve incremental data acquisition, ensuring that the system can efficiently process large amounts of data while avoiding resource waste caused by fixed time windows. Specifically, the data acquisition module sets a window trigger threshold based on information entropy. The data processing flow is automatically triggered whenever the accumulated amount of acquired data meets the following conditions:

[0075] ;

[0076] in, This represents the collection of documents within the current time window. Assuming a priori distribution of historical document types, For configurable information thresholds, This represents the Kullback-Leibler divergence, used to measure the difference between the current document type distribution and the historical document type distribution. This mechanism effectively ensures real-time performance during data acquisition and avoids invalid data collection caused by a fixed window size.

[0077] The data acquisition module ensures timely document collection and standardized preprocessing through streaming processing and intelligent triggering mechanisms, and flexibly adapts to constantly changing document types and structures during system operation. Through automatic classification using deep neural networks and an incremental data acquisition strategy, the system can efficiently and accurately acquire and process various documents in engineering projects, providing fundamental data support for subsequent quality assessment and analysis.

[0078] In this embodiment, the dynamic topology construction module is configured to receive structured documents collected by the data acquisition module and construct a dynamic topology network with individual documents as nodes and multidimensional relationships between documents as edges. This module can dynamically calculate the edge weights between document nodes based on the semantic relevance, temporal continuity, and canonical reference strength between documents, thereby establishing a topology network that reflects the complex relationships between documents.

[0079] To achieve the above functionality, the dynamic topology construction module first executes a transformation process, systematically mapping each input document and its associated information into nodes and edges conforming to a preset data structure. The specific implementation process is as follows:

[0080] Document to Node data structure conversion:

[0081] For each document processed by the data acquisition module, the system generates a unique node data structure. This process ensures that each document has a standardized digital identity within the topology network. Node generation involves filling in the following fields:

[0082] 1. Key (Unique ID): The system generates a globally unique identifier for each document. This identifier can be generated by hashing the document content or by combining the document's filename, path, and creation timestamp, ensuring its uniqueness within the network and serving as the node's "identity card."

[0083] 2. `type`: This field directly utilizes the output of the data acquisition module. The data acquisition module, using a convolutional neural network-based classifier, has identified the document as "contract text," "design drawings," or "meeting minutes," among other types. This classification result is directly assigned to the node's `type` field for subsequent classification processing and analysis.

[0084] 3. Data (Attributes): This is a collection of attributes used to store the document's core metadata and content features. Specifically, it may include: document title, creation date, author, file storage path, and document summary or keyword vectors extracted by a natural language processing model. These attributes provide rich information for subsequent calculations and queries.

[0085] 4. `rel` (List of Related Edges): This field is initially set to an empty list during node creation. As subsequent steps calculate and generate the related edges originating from this node, the unique IDs of these edges are added to this list. This allows for quick indexing of all direct relationships originating from any node.

[0086] Transformation of document-to-edge data structures and calculation of edge weights:

[0087] After all nodes in the network have been created, the system begins to calculate and generate edges representing the relationships between the nodes. The system iterates through each node pair, evaluating whether a sufficiently strong association exists between them. If so, an edge data structure is generated. The edge generation process is as follows:

[0088] 1. Determining the starting point (source) and ending point (target): For two document nodes whose relationship is being evaluated, the system assigns their respective keys (unique IDs) to the source and target fields of the edge, thus clarifying the directionality of the edge.

[0089] 2. Definition of type: The type of an edge describes the nature of the association relationship. In this embodiment, since the edge weight is a fusion of multiple relationships, the edge type can be uniformly labeled as "comprehensive association relationship". In a more refined implementation, different types can also be assigned based on the main factors that trigger edge generation (such as "strong semantic correlation", "temporal proximity", etc.).

[0090] 3. Dynamic calculation and assignment of weights: This is the core step in quantifying the multi-dimensional relationships between documents into a single numerical value to meet the requirements of the edge data structure.

[0091] Step 1: Calculate scores for each dimension. The system first independently calculates the correlation scores for the three core dimensions:

[0092] Semantic association score: By comparing the semantic vectors of the content of two documents using a deep learning model, a numerical value reflecting the similarity of the content is obtained.

[0093] Temporal continuity score: Based on the generation timestamps of two documents, the time interval is calculated and converted into a numerical value representing temporal proximity using a preset time decay function. The closer the times, the higher the score.

[0094] Normative citation strength score: By comparing the set of normative clauses cited by two documents with the complete set of industry normative databases, the overlap ratio of the norms cited by the two documents is calculated, and a numerical value representing normative consistency is obtained.

[0095] Step Two: Normalization and Weighted Summation. To merge these three scores from different sources and with different dimensions into a unified weight, the system first normalizes each score, mapping them to the same numerical range (e.g., 0 to 1). Then, based on preset weight parameters, the system performs a weighted sum of these three normalized scores.

[0096] Step 3: Assignment. The final, single, quantified total score is assigned to the weight field of the edge.

[0097] Through the above transformation process, the dynamic topology building module successfully transforms unstructured documents and their implicit relationships into a machine-readable standardized graph data structure composed of nodes and edges containing specific fields, laying the foundation for subsequent quantitative analysis.

[0098] As an example, the dynamic topology building module first defines a set of nodes. This collection contains all document types obtained through the data acquisition module. In this topology network, each document serves as a node. The multidimensional relationships between documents are connected by edges. The weights of the edges are dynamically calculated to reflect the strength of the association between documents.

[0099] The relationships between documents include, but are not limited to, the following: semantic similarity, proximity of document generation time, and overlap of common normative clauses referenced between documents.

[0100] Specifically, the edge weights between documents are calculated using a dynamic edge weight function, which quantifies the weighted combination of these multidimensional relationships. The edge weight function is expressed as follows:

[0101] ;

[0102] in:

[0103] Document and The semantic similarity between documents. This value is calculated using a deep learning model to assess the similarity of document content and reflect the degree of matching between them.

[0104] Document and The generation time interval between them and These are the timestamps of document generation. This value is used to measure the temporal proximity between documents, representing their temporal relevance.

[0105] This represents the time decay factor, which is based on the total project duration. This coefficient dynamically adjusts the impact of document generation time proximity, giving more weight to documents generated more recently when calculating edge weights.

[0106] and Representing documents respectively and A collection of cited normative clauses. This collection reflects the coverage of industry standards referenced in the document.

[0107] This represents the complete set of pre-built industry standards databases, which is a collection of all standard clauses used to calculate the overlap of standard clauses between documents.

[0108] The document generation time proximity is represented by a Gaussian function, and the influence of time proximity gradually weakens as the time interval increases.

[0109] Document and The percentage of overlap of commonly cited normative clauses in the entire set of normative databases measures the consistency among document contents in terms of normative clause citations.

[0110] These are the corresponding weight parameters, used to adjust the influence of semantic similarity, temporal proximity, and clause overlap on the calculation of edge weights.

[0111] This edge weight function can dynamically calculate the edge weight between each pair of document nodes based on the semantic content, time sequence, and reference to normative clauses of the document, thereby establishing a dynamic topological network that accurately reflects the relationships between documents.

[0112] The dynamic topology building block not only relies on the direct semantic relationships between documents, but also considers the temporal arrangement of documents and their compliance with industry standards. Through this multi-dimensional calculation method, the topology network can accurately reflect the complex relationships between documents, which is helpful for subsequent service quality assessment and optimization.

[0113] In this embodiment, the dual-channel evaluation module is connected to the dynamic topology construction module and includes two main functional channels: a real-time compliance detection channel and a deep evolutionary analysis channel. Through these two channels, the dual-channel evaluation module can achieve real-time evaluation of structured documents and long-term trend prediction of their quality status, thereby providing comprehensive support for service quality optimization.

[0114] The real-time compliance detection channel is configured to perform deep semantic matching between structured documents and a pre-built industry standard database to generate a standard compliance score for each structured document. Specifically, the structured document is first compared with the standard clauses using a semantic matching unit. The matching degree between each document and the standard clause is calculated using the following formula:

[0115] ;

[0116] in, Representing documents respectively With Terms semantic vectors, For weight parameters, For bias terms, The sigmoid activation function is used. The semantic vectors of the document and the specification clauses are generated using a deep learning model (such as pre-trained BERT or other language models) for semantic matching and comparison. This matching method accurately assesses the degree of matching between the document and the specification clauses, thereby calculating the document's compliance score.

[0117] For clause integrity verification, the clause integrity verification unit is configured to statistically analyze the coverage of referenced specification clauses in structured documents and define the subset of specifications that must be followed. Then, the coverage ratio of the actual cited clauses in the document is calculated using set operations:

[0118] ;

[0119] in, For document A set of indexed normative clauses actually referenced. Preferably, for clauses not covered. Tracing the set of associated document nodes along the dynamic topology network This generates a supplementary citation suggestion chain. This mechanism enables the tracing and source tracking of specification compliance defects.

[0120] The deep evolution analysis channel is configured to construct a quality state differential equation based on the dynamic topology network and output service quality evolution trend prediction results. In this channel, a differential equation model describing the project quality state is first constructed, which is used to dynamically track the changing trends of the project in terms of schedule, cost, and quality.

[0121] The differential equation of mass state is expressed as:

[0122] ;

[0123] Specifically, first, a three-dimensional state vector is defined. Each component represents a core dimension of quality evolution:

[0124] Schedule Deviation Rate:

[0125] ;

[0126] in, This represents the actual progress value (e.g., the amount of work completed). This represents the planned schedule value. Normalization is used to eliminate project size discrepancies, and the value range is constrained to [−1, 1]. Negative values ​​indicate schedule lag.

[0127] Cost overrun rate:

[0128] ;

[0129] in, This represents the actual cumulative cost. This represents the budgeted cost. This ratio retains its original dimensional characteristics, and a value ≥ 0 reflects the degree of deviation from cost control.

[0130] Quality defect rate:

[0131] ;

[0132] in, The number of defects detected. This represents the total number of items tested. Standardized to the [0,1] interval, it reflects the level of quality compliance.

[0133] The module extracts the time-varying system matrix from historical data using the dynamic mode decomposition method:

[0134] ;

[0135] in:

[0136] The state transition matrix has elements. Describe the dynamic coupling relationships between state variables:

[0137] diagonal elements : Characterizes the self-decay effect of each state variable (such as resource crowding caused by schedule lag);

[0138] off-diagonal elements Cross-dimensional impact coefficient (e.g., the positive promotion of quality defects by cost overruns).

[0139] The matrix is ​​extracted from historical data through Dynamic Mode Decomposition (DMD).

[0140] For the input coupling matrix, elements Characterizes the transmission strength of external inputs to state variables:

[0141] Design changes to the first The influence gain of each state variable;

[0142] : Requirements change on the first The influence gain of each state variable.

[0143] To input the change matrix, This indicates a pseudo-inverse operation.

[0144] This data-driven approach effectively captures the dynamic evolution characteristics of engineering systems.

[0145] The differential equation of mass state can be further expressed as:

[0146] ;

[0147] Where the input vector Includes factors influencing design changes and factors influencing demand changes. Preferably, The value is positively correlated with the change level. The values ​​are positively correlated with the frequency of changes, and are mapped to the standard dimension space through normalization.

[0148] This equation is solved using the fourth-order Runge-Kutta numerical integration method, enabling multi-step prediction of service quality evolution trends. Details are as follows:

[0149] The quality evolution process is described by the following initial value problem:

[0150] ;

[0151] Discretization is performed using the fourth-order Runge-Kutta numerical method:

[0152] 1. Time discretization

[0153] Time interval Divided into step size of 1 node .

[0154] 2. Iterative calculation

[0155] For each time step Calculate the slope:

[0156] ;

[0157] ;

[0158] ;

[0159] ;

[0160] Update state variables:

[0161] ;

[0162] The module employs a heterogeneous computing architecture at the implementation level. Preferably, the semantic matching engine is deployed on a GPU accelerator for real-time inference, while the differential equation solver is deployed on a CPU cluster for high-precision numerical calculations. A priority scheduler is configured on the data bus between the two channels to ensure that the initial conditions for long-term trend prediction can be dynamically corrected based on real-time detection results. This design achieves collaborative optimization between evaluation channels, improving the overall evaluation accuracy of the system.

[0163] In this embodiment, the improvement suggestion generation module is connected to the dual-channel evaluation module to receive the standard compliance score of the structured document and the service quality evolution trend prediction result derived from the topology network. Based on this, it combines the topological relationship between project documents and the industry knowledge graph to output suggestions for quality optimization.

[0164] To achieve the above functions, the improvement suggestion generation module includes a mass gradient field construction unit, an anomaly propagation analysis unit, and a suggestion generation unit. These units work together to form a complete suggestion generation logic flow.

[0165] The quality gradient field construction unit is configured to build a quality influence matrix based on a dynamic topology network, which is used to quantify the impact of changes in the association strength between document nodes on the overall service quality, and to provide gradient sensitivity basis for subsequent anomaly propagation path identification.

[0166] The first step in constructing the mass gradient field unit is to define the comprehensive quality score function. This function measures the overall performance of the tasks associated with each document node across quality dimensions. To align with typical quality assessment systems in project management, it uses three basic dimensions as its components: progress score, schedule score, and performance evaluation score. Cost control score and quality inspection scores .

[0167] Comprehensive quality scoring function The calculation uses a weighted average form, expressed as follows:

[0168] ;

[0169] in, This refers to any document node in the network topology. These are weighting coefficients used to reflect the relative importance of schedule, cost, and quality across different project phases or types. The weighting parameters can be dynamically determined using a optimized Analytic Hierarchy Process (AHP), combined with expert evaluation and historical project data.

[0170] The three scoring dimensions mentioned above can be calculated from the task data associated with the document nodes:

[0171] Progress Score It can be determined based on the degree of matching between the planned progress and the actual progress;

[0172] Cost control score It can be quantified based on the degree of deviation between budgeted costs and actual costs;

[0173] Quality inspection score This can be derived from the quality inspection results corresponding to the task at that node, such as the proportion of non-conforming items and the number of re-inspections.

[0174] In obtaining the comprehensive quality function Subsequently, the mass gradient field construction unit further constructs the mass influence matrix. This is used to reflect the impact of edge weight changes on the quality scoring function in a dynamic topology network. The first-order response relationship. This matrix is ​​defined as follows:

[0175] ;

[0176] in, Represents document nodes in a network topology and The edge weights between the two represent the strength of their semantic, temporal, or normative relationship. This represents the total number of nodes in the network topology.

[0177] To obtain each matrix element For the specific value of the mass gradient field, the mass function is differentiated using the autodiff technique. Because... Essentially, they are state variables. The weighted linear combination of the terms leads to the partial derivatives which can be further expanded as follows:

[0178] ;

[0179] In the above expression This reflects the indirect impact of changes in topological edge weights on quality indicators across various dimensions, and is modeled based on the information flow paths between nodes in a dynamic topological network. Specifically:

[0180] For schedule deviation rate This partial derivative captures how changes in the temporal order between document nodes causally affect the task completion rate;

[0181] For cost overrun rate This partial derivative term reflects the correlation between the strength of a specific relationship (such as the reference relationship between design and budget) and cost variations;

[0182] For quality defect rate This derivative measures the potential driving effect of changes in semantic consistency or canonical dependencies between documents on the final detection results.

[0183] The automatic differentiation process is performed within the backpropagation framework, enabling the acquisition of the gradient contribution of each edge weight change to the scoring function without analytical derivation, thus ensuring the traceability and scalability of the computation. Preferably, this computation is performed in a runtime environment that supports tensor computation, such as using tensor graph construction tools to improve computational efficiency.

[0184] The resulting matrix Each element The sensitivity of node edge weights to changes in service quality during dynamic adjustment was quantified. This matrix serves as input for subsequent anomaly propagation analysis units, providing accurate gradient references for critical path search and anomaly causal chain tracing.

[0185] After constructing the quality gradient field, the anomaly propagation analysis unit is configured to identify critical impact paths in the topology network, thereby locating path segments that may lead to service quality degradation. This identification task is based on the following objective function:

[0186] ;

[0187] in, Candidate paths represent the paths from the starting node where the anomaly was detected. To key delivery milestones All connected paths; The derivative of the edge weights over time reflects the degree of dynamic fluctuation in the relationship. Let be the weight response strength of the edge in the gradient field. By solving for the maximum value of this product across all paths, the path with the most significant propagation of anomalies can be identified. .

[0188] The path search space is limited to starting from the anomaly source node. To key delivery milestones The set of all connected paths. Preferably, the source node. The activation condition is that the state variable exceeds a dynamic threshold:

[0189] ;

[0190] in, The mean and standard deviation of the historical data.

[0191] The rate of change of edge weights Calculated using sliding window difference:

[0192] ;

[0193] The objective function design highlights the cumulative effect and dynamic evolution characteristics of anomaly propagation, effectively identifying the transmission links of key issues.

[0194] To improve path identification efficiency, the anomaly propagation analysis unit is further configured to use an improved Dijkstra algorithm for path search and filtering. The path evaluation function is defined as follows:

[0195] ;

[0196] This design balances the differences in sensitivity contributions across different sides through nonlinear transformations, avoiding numerical overflow. Preferably, the algorithm uses a dual termination condition:

[0197] Path length ,in, The maximum path length is represented, and the complexity of the constraint suggestion chain is within interpretable limits.

[0198] Cumulative weight To ensure the effectiveness of the recommendations, among which, Indicates the critical threshold. The statistical distribution is set based on historical data, and the value is three times the standard deviation of the normal distribution.

[0199] It is recommended that the generation unit be configured based on the aforementioned anomaly propagation path. In addition, it outputs improvement suggestions corresponding to path nodes based on the normative dependencies and empirical rules in the industry knowledge graph.

[0200] In practice, the path nodes are first mapped to entities such as engineering clauses, process nodes, and construction points in the knowledge graph, and then joint reasoning is performed based on the dependencies encoded in the graph (such as hierarchical references of standard clauses and constraint rules between items) and the expert rule base (such as rule-based reasoning of non-compliant behaviors).

[0201] The reasoning process employs the following logic:

[0202] If there are high-gradient abrupt changes in the path nodes, it is recommended to first check the consistency of the two documents connected by the edge in terms of canonical references or chronological arrangements.

[0203] If the end node of the path is a critical deliverable document and its quality score is significantly lower than the average of the preceding nodes, it is recommended to backtrack and check from its associated process nodes or input documents.

[0204] If there are segments in the path where the edge weights change drastically over time, it is recommended to check with time series graphs to see if there are frequent changes in the phased design or external intervention events.

[0205] The final generated recommendations are output in a structured format, including the recommendation location (corresponding document ID), recommendation content (text description), and the basis for association (graph node or rule matching results).

[0206] In this embodiment, the evaluation report generation module is configured to receive and integrate the specification compliance score and service quality evolution trend prediction results from the dual-channel evaluation module, as well as the improvement suggestion information from the improvement suggestion generation module. Based on the node connection relationship of the constructed dynamic topology network, it establishes the correlation mapping between evaluation indicators and finally generates a standardized evaluation report in a structured manner from the multi-source evaluation results. It also has an incremental dynamic update mechanism to deal with the evaluation offset caused by real-time changes in network status.

[0207] During the data reception and fusion phase, the evaluation report generation module first establishes a unified data input channel to receive the following input:

[0208] The compliance score from the dual-channel assessment module is denoted as This reflects the document's alignment with industry standards.

[0209] The service quality evolution trend prediction results from the deep evolution analysis channel are denoted as The dynamic trend of quality status is usually represented by a time series curve.

[0210] Improvement suggestions from the improvement suggestion generation module are denoted as This includes key impact path nodes and their recommendations.

[0211] Since the aforementioned data may originate from different modules and have different data refresh cycles, to ensure time series consistency, the evaluation report generation module introduces a time series alignment mechanism, employing an alignment constraint strategy:

[0212] ;

[0213] in, As the system's base time, The design adaptively adjusts based on the data collection frequency. This ensures temporal consistency among evaluation elements and avoids misalignment of causal relationships.

[0214] To establish a structured correspondence between indicators and document nodes in the report, the evaluation report generation module is further configured with an association mapping construction mechanism, which constructs an indicator-node association matrix based on the node connection relationships in the dynamic topology network. Its element definition is as follows:

[0215] ;

[0216] The correlation between evaluation indicators and nodes is supported by the following two types of conditions:

[0217] document nodes Metadata and metrics included The semantic similarity is higher than a preset threshold This similarity calculation is based on an embedding model or vector space mapping;

[0218] index The computational logic directly uses data from the nodes. Derived data (such as edge weights, timestamps, quality scores, etc.).

[0219] This correlation matrix provides a structural explanation for each evaluation item in the report and provides index support for subsequent difference backtracking and optimization suggestion positioning.

[0220] In the output phase, the evaluation report generation module adopts a structured output method that combines template-driven and content reasoning to generate a standardized report format that meets machine parsing requirements.

[0221] Preferably, a standardized report template is defined:

[0222] ;

[0223] in:

[0224] Used to encapsulate basic project information, affiliated unit, task stage, time range, and other metadata;

[0225] The content, including compliance scores, service quality evolution trend predictions, and improvement suggestions, is presented in a tree structure, with nested outputs based on node and indicator dimensions.

[0226] Records metadata such as evaluation version information, generated timestamps, report summary verification hash values, etc.

[0227] To ensure the standardization and compliance of the report structure, the module defines the structural constraints of the template through JSON Schema, preferably conforming to the machine-parseable report interface specifications defined by international standards such as ISO / IEC 20560, in order to support subsequent automated review and audit trail.

[0228] To enhance the system's adaptability in dynamic environments, the evaluation report generation module is equipped with an edge weight change monitoring mechanism and an incremental report update engine. This mechanism monitors the edge weight change rate in the dynamic topology network in real time, triggering a global update when the following conditions are met:

[0229] ;

[0230] This means that the average rate of change of edge weights in the network over time exceeds a set threshold. When this happens, the incremental report update process is automatically triggered.

[0231] During incremental updates, the module only recalculates the subset of evaluation metrics affected by changes in edge weights:

[0232] ;

[0233] in, This represents the set of document nodes that are directly related to changes in edge weights.

[0234] To ensure the traceability and version control of updated data, the module uses a columnar storage method to maintain the report version history, and stores each update as a Delta record, including fields such as timestamp, difference range, and structure hash. Preferably, version integrity verification is performed by constructing a Merkle tree structure to support arbitrary version rollback and audit retrieval of evaluation history.

[0235] In summary, this invention constructs a dynamic topology network with document nodes at its core and relationships as its edges, integrating a dual-channel evaluation mechanism and quality optimization path identification methods to form a service quality quantitative management framework that spans the entire project lifecycle. The system as a whole is based on the deep semantic matching capability of the real-time compliance detection channel and the state prediction capability of the deep evolutionary analysis channel to generate compliance scores and evolutionary trend results. Furthermore, it relies on the quality gradient field and industry knowledge graph to derive optimization paths and improvement suggestions. Finally, through the dynamic association mapping and incremental update mechanism based on node connections in the evaluation report generation module, it achieves the fusion output of multi-source evaluation information and the generation of structured reports, constructing a complete, traceable, and adjustable service evaluation closed loop to support intelligent monitoring and continuous improvement of engineering quality.

[0236] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A service quality assessment system, characterized in that, include: The data acquisition module is configured to automatically collect structured documents from the entire lifecycle of engineering consulting projects, including contract texts, design drawings, and meeting minutes. A dynamic topology construction module is configured to receive structured documents collected by the data acquisition module, construct a dynamic topology network with individual documents as nodes and multidimensional relationships between documents as edges, and dynamically calculate edge weights based on document semantic relevance, temporal continuity, and normative citation strength. A dual-channel evaluation module, connected to the dynamic topology construction module, includes a real-time compliance detection channel and a deep evolutionary analysis channel. The real-time compliance detection channel is configured to perform deep semantic matching of structured documents with a pre-built industry normative database to generate normative compliance scores for the structured documents. The deep evolutionary analysis channel is configured to... The system constructs a quality state differential equation using a dynamic topology network and outputs a service quality evolution trend prediction result. An improvement suggestion generation module is configured to receive the specification compliance score and the service quality evolution trend prediction result, identify abnormal propagation paths by constructing the topological gradient field of the dynamic topology network, and generate improvement suggestions for service quality optimization by combining industry knowledge graphs. An evaluation report generation module is configured to integrate the specification compliance score, service quality evolution trend prediction result, and improvement suggestions to output a structured evaluation report. The real-time compliance detection channel includes a deep semantic matching unit configured to use a pre-trained language model to calculate document... With the terms and conditions Semantic similarity: ;in, Representing documents respectively With the terms and conditions semantic vectors, For trainable weight matrix, For the sigmoid function, This is a bias term; the clause integrity verification unit is configured to statistically measure the coverage of clauses in the structured document reference specifications, and the coverage calculation formula is: ;in, Document The actual set of normative clauses referenced; This represents a predefined subset of mandatory specifications to be followed based on the project type; the comprehensive scoring unit is configured to calculate documents based on semantic similarity and coverage. The standard compliance score is calculated using the following formula: ;in, For document Standard compliance score; These are the semantic matching weight coefficients; the mass state differential equation is expressed as: ;in, The schedule deviation rate is expressed as (actual progress - planned progress) / planned progress. This indicates the cost overrun rate, calculated as actual cost / budgeted cost. The defect rate is expressed as the number of defective items detected divided by the total number of items detected. This represents the impact factor of design changes, and its value is positively correlated with the level of change. This represents the impact factor of demand changes, and its value is positively correlated with the frequency of changes. Represents the time-varying state transition matrix; This represents the time-varying input coupling matrix.

2. The service quality assessment system according to claim 1, characterized in that, The data acquisition module includes: a document type identification unit, configured to automatically distinguish between three types of documents—contract text, design drawings, and meeting minutes—using a convolutional neural network model; a streaming processing unit, configured to perform incremental data acquisition using a dynamic window mechanism; and a document type identification and data standardization processing unit, which is triggered when the accumulated data volume reaches a set threshold.

3. The service quality assessment system according to claim 1, characterized in that, The dynamic topology building module is configured to execute: Define a node set V = {v k |v k ∈Contract clause entity, design drawing entity, meeting minutes entity}; Construct a dynamic edge weight function to calculate the edge weight between nodes. The edge weight is determined by a weighted combination of semantic similarity between documents, proximity of document generation time, and overlap of commonly referenced normative clauses.

4. The service quality assessment system according to claim 3, characterized in that, The dynamic edge weight function is specifically as follows: ;in, ； Indicates the document generation time interval; This represents the time decay factor, based on the total project duration. set up; and Representing documents respectively and The set of referenced normative clauses; This represents the complete set of pre-built industry standards databases; Indicates the proximity of document generation times; Indicates the percentage overlap of commonly cited normative clauses within the entire normative database set; These are the corresponding weight parameters, used to adjust the influence of semantic similarity, temporal proximity, and clause overlap on edge weights, respectively.

5. The service quality assessment system according to claim 1, characterized in that, The improvement suggestion generation module includes: a mass gradient field construction unit, configured to construct a mass influence matrix. ;in, This represents the comprehensive quality scoring function, which is calculated as the weighted average of the progress score, cost control score, and quality inspection score of the tasks associated with the document node. Represents nodes in a dynamic topology network and The edge weights; This represents the total number of document nodes in the current network topology; the anomaly propagation analysis unit is configured to detect critical impact paths based on the following formula: ;in, This is a key anomaly propagation path. Candidate paths, This indicates the document node from which an anomaly was detected. To key delivery milestones The set of all connected paths; This represents the rate of change of edge weights over time; it suggests generating units configured based on the aforementioned key influence paths. Based on the normative dependencies and empirical rules in the industry knowledge graph, optimization suggestions are generated for path nodes.

6. The service quality assessment system according to claim 5, characterized in that, The anomaly propagation analysis unit is further configured to solve the problem using an improved Dijkstra algorithm. The path weight is defined as: ;in, Represents the edges in the quality gradient matrix Element; This represents the rate of change of edge weights over time; This represents a path from the starting node to the ending node; the algorithm terminates when any of the following thresholds is met: path length ;in, Indicates the preset maximum path length; cumulative weight ;in, This represents the critical threshold for path influence.

7. The service quality assessment system according to claim 1, characterized in that, The evaluation report generation module is configured to: receive and integrate the compliance score and service quality evolution trend prediction results from the dual-channel evaluation module, as well as the improvement suggestions from the improvement suggestion generation module; establish a correlation mapping between evaluation indicators based on the node connection relationship of the dynamic topology network; convert the fusion results and correlation mapping into a machine-readable standardized report format according to a preset template; and automatically initiate incremental report updates when the change rate of any side weight in the dynamic topology network exceeds a set threshold.

Images