Tender full-process automation and decision support system based on multi-agent cooperation

Abstract

The application relates to the technical field of decision support, and discloses a bidding whole-process automation and decision support system based on multi-agent cooperation. First, the updating intensity sequence and global risk score data of the demand extraction side and the review and early warning side are collected, and the instantaneous phase and effective frequency are extracted through signal analysis. Then, phase relationship and frequency ratio analysis are carried out, the frequency mismatch and coupling strength are calculated, and whether the system is in a locked state is determined. When the locking effect is detected, the real risk signal is extracted through the unlocking residual error module to eliminate the false fluctuations caused by the model rhythm difference. Then, the residual error signal quality is evaluated by using a signal-to-noise ratio calculation module, and a document-level risk score is comprehensively output in combination with a mapping function, so that accurate identification and intelligent decision support of the bidding document risk are realized, and the accuracy and stability of risk prediction are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of decision support technology, and more specifically, to a decision support system for the entire bidding process based on multi-agent collaboration. Background Technology

[0002] In automated bid analysis and risk management, two types of intelligent systems are commonly used: a requirements analysis module and a document review module. The former extracts elements such as clauses, requirements, and technical specifications from the bidding documents; the latter performs intelligent reviews of the generated bid documents in terms of compliance, technology, and pricing. Traditional systems often operate in a sequential manner: the analysis module first extracts the content, and then the review module evaluates the results. While this unidirectional process can improve processing efficiency, it is prone to delays, information loss, and task drift when dealing with long, multi-layered documents and multi-source data. To alleviate these problems, the industry has begun to adopt a dual-system interactive working mechanism, enabling the two systems to continuously correct and update each other throughout the bidding cycle, achieving dynamic learning and prediction.

[0003] Under this interactive mechanism, the two intelligent systems need to continuously exchange data and update their models based on their respective learning rates, feedback frequencies, and processing cycles. However, practice shows that there is an inherent asynchrony in the update rhythm, internal training step size, and information transmission between the two systems. The parsing module typically relies on batch learning from document streams, resulting in a longer update cycle; while the review module mostly performs short-cycle fine-tuning based on real-time evaluation results. This rhythm difference causes the overall state of the system to shift periodically over time. When the update cycle ratio of the two modules approaches a certain stable relationship, the risk prediction results will exhibit a fixed periodic behavior, that is, the risk score will repeat a pattern over time, decoupling from changes in the actual document content. This phenomenon is not an algorithmic error, but an endogenous dynamic synchronization effect, stemming from the rhythmic coordination formed by the two systems in the time dimension, making the prediction results dominated by the update rhythm. Summary of the Invention

[0004] This invention provides a fully automated bidding process and decision support system based on multi-agent collaboration, which solves the technical problem mentioned in the background: in the process of dual-system collaborative prediction, risk scoring is periodically locked due to the system rhythm, resulting in the results being insensitive to changes in document content and failing to reflect the true risks.

[0005] This invention relates to a multi-agent collaborative bidding process automation and decision support system, comprising:

[0006] The data acquisition module collects the intensity sequence updated from the demand extraction side, the intensity sequence updated from the review and early warning side, and the global risk scoring sequence.

[0007] The data parsing module performs signal analysis on the demand extraction side update intensity sequence and the review and early warning side update intensity sequence to obtain the instantaneous phase of the demand extraction side, the instantaneous phase of the review and early warning side, and their respective effective frequencies.

[0008] The data analysis module performs phase relationship analysis based on the instantaneous phase of the demand extraction side and the instantaneous phase of the review and early warning side to obtain the frequency ratio;

[0009] The coupling strength calculation module calculates the frequency mismatch between the effective frequency of the demand extraction side and the effective frequency of the review and early warning side, as well as the coupling strength between the two update strength sequences.

[0010] The lockout state determination module determines the lockout state based on frequency mismatch, coupling strength, and frequency ratio; when in a lockout state, it locates the dominant lockout frequency in the spectrum of the global risk score sequence.

[0011] The delocking residual module constructs a locking component based on the locking dominant frequency and uses the difference between the global risk score sequence and the locking component as the delocking residual.

[0012] The signal-to-noise ratio (SNR) calculation module defines the locked phase based on the frequency ratio, sets the phase cross section to determine the set of in-phase moments, calculates the statistics of the delocking residual on the set of in-phase moments, and obtains the content SNR.

[0013] The risk scoring module processes the statistics of the delocked residuals and the content signal-to-noise ratio through a mapping function, and outputs a document-level risk score.

[0014] The beneficial effects of this invention include: by introducing a dual-agent collaborative mechanism, dynamic linkage and adaptive learning are achieved between the demand extraction side and the review and early warning side throughout the entire bidding process, effectively identifying and eliminating the periodic locking problem of risk scoring caused by differences in system rhythm. Through comprehensive analysis of frequency ratio, coupling strength, and locking status, the system can automatically separate false risk fluctuations caused by model coordination rhythm and extract risk signals that are truly related to the document content. Compared with traditional one-way or static evaluation systems, this invention significantly improves the sensitivity and accuracy of risk prediction, realizes real-time self-correction and intelligent judgment of bidding risks, thereby improving the reliability and scientific nature of automated bidding decisions. Attached Figure Description

[0015] Figure 1 This is a block diagram of the bidding process automation and decision support system based on multi-agent collaboration of the present invention. Detailed Implementation

[0016] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0017] like Figure 1 As shown, the bidding process automation and decision support system based on multi-agent collaboration includes:

[0018] The data acquisition module collects the intensity sequence updated from the demand extraction side, the intensity sequence updated from the review and early warning side, and the global risk scoring sequence.

[0019] The data parsing module performs signal analysis on the demand extraction side update intensity sequence and the review and early warning side update intensity sequence to obtain the instantaneous phase of the demand extraction side, the instantaneous phase of the review and early warning side, and their respective effective frequencies.

[0020] The data analysis module performs phase relationship analysis based on the instantaneous phase of the demand extraction side and the instantaneous phase of the review and early warning side to obtain the frequency ratio;

[0021] The coupling strength calculation module calculates the frequency mismatch between the effective frequency of the demand extraction side and the effective frequency of the review and early warning side, as well as the coupling strength between the two update strength sequences.

[0022] The lockout state determination module determines the lockout state based on frequency mismatch, coupling strength, and frequency ratio; when in a lockout state, it locates the dominant lockout frequency in the spectrum of the global risk score sequence.

[0023] The delocking residual module constructs a locking component based on the locking dominant frequency and uses the difference between the global risk score sequence and the locking component as the delocking residual.

[0024] The signal-to-noise ratio (SNR) calculation module defines the locked phase based on the frequency ratio, sets the phase cross section to determine the set of in-phase moments, calculates the statistics of the delocking residual on the set of in-phase moments, and obtains the content SNR.

[0025] The risk scoring module processes the statistics of the delocked residuals and the content signal-to-noise ratio through a mapping function, and outputs a document-level risk score.

[0026] In one embodiment of the present invention, the process of collecting the demand extraction side update intensity sequence, the review and early warning side update intensity sequence, and the global risk scoring sequence includes:

[0027] Define a set of time steps, which contains all time steps from the first time step to the Nth time step, where N is the total number of time steps, and each time step is represented by t;

[0028] At each time step t, demand extraction lateral vector evidence and review warning lateral vector evidence are obtained, and the value of the global risk score sequence at that time step t is recorded.

[0029] The value of the demand extraction side update strength sequence at time step t is calculated as follows: the Euclidean norm of the difference between the vector evidence at time step t and the vector evidence at time step t-1; where t ranges from the second time step to the Nth time step.

[0030] The value of the updated intensity sequence on the review and early warning side at time step t is calculated as follows: the Euclidean norm of the difference between the vector evidence on the review and early warning side at time step t and the vector evidence at time step t-1; where t ranges from the second time step to the Nth time step.

[0031] The time step set is a set of all time nodes covering the bidding risk monitoring cycle, divided at fixed time intervals, used to synchronously record dual-agent data and risk scores.

[0032] N is the total number of time step sets, which needs to be set according to the bidding period and data sampling frequency (e.g., if the bidding period is 7 days and sampling is performed twice a day, N is 14).

[0033] t is the identifier for a single time step, with values ​​from 1 to N (t=1 corresponds to the first sampling time, and t=N corresponds to the last sampling time), used to locate the observation data at each time node.

[0034] The demand extraction lateral vector evidence is a vector of key elements extracted from the tender documents by the demand extraction agent at time step t. (e.g., including three dimensions: project budget, number of technical indicators that meet the requirements, and delivery cycle requirements. Each dimension is represented by a value from 0 to 100, with 100 for full budget and 0 for insufficient budget, 100 for the percentage of technical indicators that meet the requirements, and 100 for compliant delivery cycle and 0 for overdue delivery cycle.)

[0035] The review and early warning lateral vector evidence is the evaluation vector of the tender documents by the review and early warning agent at time step t (such as including three dimensions: compliance rate of compliance clauses, reasonableness of price, and completeness of technical solution, each dimension is represented by a value of 0-100, compliance rate of compliance clauses ×100, price within a reasonable range is 100, exceeding is 0, completeness of technical solution is 100, and missing items are 0).

[0036] The global risk score sequence at time step t is a comprehensive score reflecting the bidding risk level at time step t, which is output collaboratively by the two agents. The range is 0-100 (0 represents no risk and 100 represents extremely high risk). It is generated by the system based on the weighted calculation of vector evidence from the two agents (e.g., each agent has a weight of 50%, and the total score = the mean of the vector evidence from the demand side × 50% + the mean of the vector evidence from the review side × 50%).

[0037] The value of the demand extraction side update intensity sequence at time step t is an indicator that quantifies the magnitude of change in the demand extraction side vector evidence between adjacent time steps; the greater the magnitude of change, the higher the update intensity. Specifically, the steps are as follows: First, identify all elements in the corresponding dimensions of the demand extraction side vector evidence at time step t and the vector evidence at time step t minus one. Second, calculate the difference between each pair of corresponding dimension elements (element at time step t minus the element at time step t minus one). Third, square each difference. Fourth, sum all the squared results to obtain a total. Fifth, take the square root of this total; the result is the value of the demand extraction side update intensity sequence at time step t.

[0038] The value of the update intensity sequence of the review and early warning side at time step t is an indicator that quantifies the magnitude of change in the vector evidence of the review and early warning side between adjacent time steps. The greater the magnitude of change, the higher the update intensity. Specifically, the first step is to find all elements of the corresponding dimension in the vector evidence of the review and early warning side at time step t and the vector evidence at time step t minus one. The second step is to calculate the difference between each pair of corresponding dimension elements (element at time step t minus element at time step t minus one). The third step is to square each difference. The fourth step is to sum all the squared results to obtain the total. The fifth step is to take the square root of this total, and the result is the value of the update intensity sequence of the review and early warning side at time step t.

[0039] It should be noted that the intelligent agent is as follows:

[0040] The demand extraction agent focuses on extracting key elements from bidding documents. Its core function is to extract key elements such as project budget, number of technical specifications met, and delivery cycle requirements from bidding documents and quantify these elements into vector evidence (e.g., quantifying the degree of compliance of each element with a value from 0 to 100, with 100 for full budget and 100 for complete technical specification compliance). The demand extraction side update intensity sequence is the quantified result of the changes in the vector evidence of the demand extraction agent.

[0041] The review and early warning agent focuses on bid document evaluation. Its core function is to evaluate the generated bid documents from dimensions such as compliance rate, price reasonableness, and technical solution completeness. The evaluation results are also quantified as vector evidence (e.g., 100% compliance with compliance clauses is assigned 100, and a price within a reasonable range is assigned 100). The updated intensity sequence of the review and early warning agent is the quantified result of changes in the vector evidence of the review and early warning agent.

[0042] Specifically, the global risk scoring sequence is the result of collaborative output between the demand extraction agent and the review and early warning agent.

[0043] The demand extraction-side intelligent agent is a text parsing algorithm system that focuses on extracting and quantifying key elements of bidding documents. It transforms unstructured / semi-structured bidding document content into structured vector evidence that can be used for risk analysis, specifically in the following form:

[0044] With Natural Language Processing (NLP) technology at its core, it includes sub-modules such as named entity recognition (extracting key values ​​such as project budget and delivery cycle), text classification (determining the type of technical indicators, such as hardware requirements and software compatibility), and information extraction (extracting mandatory and priority items from the terms), ensuring accurate location of risk-related elements in the bidding documents;

[0045] Extract elements directly related to bidding risks from the tender documents, such as project budget, technical specifications, delivery time limits, and qualification requirements.

[0046] Extract a project budget of no less than 5 million yuan and quantify it into a budget adequacy dimension (take 100 for budgets ≥ 5 million yuan, and convert proportionally for budgets < 5 million yuan, such as taking 80 for budgets of 4 million yuan).

[0047] Extract the three core technical indicators and quantify them into a technical indicator coverage dimension (100 for full coverage, 67 for coverage of two indicators, and so on).

[0048] Each time step outputs multi-dimensional quantitative vector evidence (such as three core dimensions: budget adequacy, technical indicator coverage, and delivery cycle compliance, with each dimension ranging from 0 to 100).

[0049] The review and early warning intelligent agent is a comprehensive evaluation algorithm system focusing on bid document assessment and risk warning. It evaluates and quantifies generated bid documents from the dimensions of compliance, reasonableness, and completeness, outputting vector evidence to support risk analysis. Its specific form is as follows:

[0050] The system integrates a rule engine and machine learning, with the rule engine used to match compliance clauses (such as mandatory rules like requiring official seals on tender documents) and machine learning models (such as classifiers and regression models) used to evaluate non-mandatory indicators (such as the reasonableness of the price and the feasibility of the technical solution).

[0051] The tender documents are evaluated based on several dimensions, including compliance rate, price reasonableness, completeness of technical solutions, and validity of qualification documents. For example:

[0052] Check whether the tender documents meet the 10 compliance clauses, and quantify them as compliance clause compliance rate (100 for meeting 10 clauses, 80 for meeting 8 clauses).

[0053] The model is trained based on historical bid price data to determine whether the current bid is within a reasonable range, and is quantified as a dimension of bid reasonableness (100 is taken within the range, and the deviation is converted when it is outside the range).

[0054] Each time step outputs multi-dimensional quantitative vector evidence (such as compliance rate of compliance clauses, reasonableness of quotation, and completeness of technical solution, with each dimension ranging from 0 to 100).

[0055] The demand extraction agent and the review and early warning agent do not operate independently, but rather form a closed-loop collaboration by continuously exchanging vector evidence and dynamically adjusting the evaluation strategy.

[0056] Vector evidence (such as "technical requirements") from the demand extraction side agent will be used as the evaluation benchmark for reviewing the early warning side agent (to review whether the tender documents meet the requirements).

[0057] The vector evidence reviewed by the early warning agent (such as technical solutions not meeting the indicators) will be fed back to the demand extraction agent to adjust the priority of element extraction (such as focusing on supplementary clauses that do not meet the indicators).

[0058] The vector evidence changes (i.e., the update strength sequence) of both, together with the globally risk score sequence of the collaborative output, constitute the core data chain for bid risk prediction.

[0059] In one embodiment of the present invention, signal analysis is performed on the demand extraction side update intensity sequence and the review and warning side update intensity sequence to obtain the instantaneous phase of the demand extraction side, the instantaneous phase of the review and warning side, and their respective effective frequencies, including:

[0060] Set the set of time steps and the number of time steps, and set the time step size;

[0061] Hilbert transform is applied to the demand extraction side update intensity sequence to construct the demand extraction side analytic signal, and Hilbert transform is applied to the review and early warning side update intensity sequence to construct the review and early warning side analytic signal.

[0062] The argument of the analytical signal on the demand extraction side is taken as the instantaneous phase on the demand extraction side, and the argument of the analytical signal on the review and early warning side is taken as the instantaneous phase on the review and early warning side.

[0063] Phase expansion is performed on the instantaneous phase of the demand extraction side, the phase difference between adjacent time steps is calculated, and the phase difference is divided by the product of the time step size and twice π to obtain the instantaneous frequency sequence of the demand extraction side; where the time step value is the number of time steps from the second time step.

[0064] Phase expansion is performed on the instantaneous phase of the review and early warning side, the phase difference between adjacent time steps is calculated, and the phase difference is divided by the product of the time step size and twice π to obtain the instantaneous frequency sequence of the review and early warning side; where the time step value is the number of time steps from the second time step.

[0065] The median of the instantaneous frequency sequence on the demand extraction side is taken as the effective frequency on the demand extraction side, and the median of the instantaneous frequency sequence on the review and early warning side is taken as the effective frequency on the review and early warning side.

[0066] The time step set is a set of all time nodes covering the parsing period of the dual-agent update intensity sequence signal, used to synchronize and match the update intensity data with the parsing process.

[0067] The number of time steps is the total number of time steps in the set, which is consistent with the total number of time steps N. It must match the length of the update intensity sequence to be parsed. For example, if the update intensity sequence has 13 data points, the number of time steps is 13.

[0068] The time step is the time interval between two adjacent time steps, in units such as seconds or minutes (e.g., if the update intensity is recorded every 5 minutes, the time step is 5 minutes). It needs to be set according to the data update frequency of the two agents.

[0069] The Hilbert transform is an operation that converts a real-valued update intensity sequence into an imaginary signal. Its purpose is to construct an analytic signal containing both the real part (the original sequence) and the imaginary part (the transformed result). Specifically, the first step is to determine the update intensity sequence to be transformed (e.g., a demand-side sequence containing all values ​​from time step 2 to the number of time steps). The second step is to calculate the integral of each value at each time step in the sequence with 1 divided by π and multiplied by the time difference (the time difference being the time interval between the current time step and all other time steps). The third step is to obtain the Hilbert transform result (the imaginary part signal) corresponding to the update intensity sequence.

[0070] The demand extraction side analytic signal is a complex signal, with the real part being the value of the demand extraction side update intensity sequence and the imaginary part being the result of the sequence after Hilbert transform.

[0071] The analytical signal of the review and early warning side is a complex signal, with the real part being the value of the update intensity sequence of the review and early warning side, and the imaginary part being the result of the sequence after Hilbert transformation.

[0072] The instantaneous phase on the demand extraction side is the angle corresponding to the analytical signal (complex signal) on the demand extraction side, ranging from negative π to positive π, reflecting the phase state of the signal at that time step.

[0073] The instantaneous phase of the review and early warning side is the angle corresponding to the analytical signal (complex signal) of the review and early warning side, ranging from negative π to positive π, reflecting the phase state of the signal at that time step.

[0074] Phase unwrapping is an operation that corrects the instantaneous phase folding problem (the original instantaneous phase jumps when it exceeds π or is lower than -π, and the unwrapped phase becomes continuous). Specifically, the first step is to obtain the instantaneous phase of all time steps on a certain side (such as the demand side); the second step is to calculate the difference between the instantaneous phase of the current time step and the instantaneous phase of the previous time step, starting from the second time step; the third step is to calculate the unwrapped phase of the current time step if the difference is greater than π, then the unwrapped phase of the current time step is equal to the current instantaneous phase minus 2π; if the difference is less than negative π, then the unwrapped phase of the current time step is equal to the current instantaneous phase plus 2π; if the difference is between negative π and π, the unwrapped phase is equal to the current instantaneous phase; the fourth step is to calculate the unwrapped phase of the first time step, which is equal to its instantaneous phase; and so on, to obtain the complete unwrapped phase sequence.

[0075] The phase difference between adjacent time steps on the demand extraction side is the result of subtracting the unfolded phase of the previous time step from the unfolded phase of the next time step in the unfolded demand side phase.

[0076] The phase difference between adjacent time steps on the review and early warning side is the result of subtracting the unfolded phase of the previous time step from the unfolded phase of the next time step in the unfolded review side phase.

[0077] The instantaneous frequency sequence of the demand extraction side is a sequence reflecting the frequency of the demand side update intensity signal at each time step. Specifically, the first step is to obtain the phase difference between adjacent time steps on the demand extraction side (from the second time step to the number of time steps); the second step is to calculate the product of the time step length and 2π; the third step is to divide the phase difference of each time step by the product to obtain the instantaneous frequency of the corresponding time step; the instantaneous frequencies of all time steps constitute the instantaneous frequency sequence of the demand extraction side.

[0078] The instantaneous frequency sequence of the review and early warning side is a sequence reflecting the frequency of the review and early warning side update intensity signal at each time step. Specifically, the first step is to obtain the phase difference between adjacent time steps of the review and early warning side (from the second time step to the number of time steps); the second step is to calculate the product of the time step length and 2π; the third step is to divide the phase difference of each time step by the product to obtain the instantaneous frequency of the corresponding time step; the instantaneous frequencies of all time steps constitute the instantaneous frequency sequence of the review and early warning side.

[0079] The effective frequency of demand extraction is the median of the instantaneous frequency sequence of demand, which is used to characterize the stable frequency of the demand update intensity signal (the median can suppress the influence of outliers).

[0080] The effective frequency of the review and early warning side is the median of the instantaneous frequency sequence of the review side, which is used to characterize the stable frequency of the update intensity signal of the review side (the median can suppress the influence of outliers).

[0081] In one embodiment of the present invention, a phase relationship analysis is performed based on the instantaneous phase of the demand extraction side and the instantaneous phase of the review and early warning side to obtain a frequency ratio, including:

[0082] Define the set of time steps and the number of time steps, where the number of time steps is N;

[0083] Phase expansion is performed on the instantaneous phase of the demand extraction side to obtain the expanded phase of the demand extraction side, and phase expansion is performed on the instantaneous phase of the review and early warning side to obtain the expanded phase of the review and early warning side.

[0084] The difference between the unfolded phase on the demand extraction side at the first time step and the Nth time step is calculated as the cumulative phase increment on the demand extraction side, and the difference between the unfolded phase on the review and early warning side at the first time step and the Nth time step is calculated as the cumulative phase increment on the review and early warning side.

[0085] The ratio of the cumulative phase increment on the demand extraction side to the cumulative phase increment on the review and early warning side is taken as the real number form of the frequency ratio.

[0086] Set an approximation threshold, and use the continued fraction method to determine the integer numerator and integer denominator, such that the difference between the real number form and the result of dividing the numerator by the denominator is no greater than the approximation threshold. The frequency ratio is represented by the numerator and denominator.

[0087] The time step set is a set of all time nodes covering the phase relationship analysis cycle, used to match the instantaneous phase data of the two agents.

[0088] N is the total number of time steps in the set, which is consistent with the total number of time steps N and the number of time steps, and must match the length of the instantaneous phase sequence of the two agents.

[0089] The cumulative phase increment on the demand extraction side is the total change in the demand-side expanded phase over the entire analysis period, reflecting the overall magnitude of the demand-side phase change. Specifically, the first step is to find the value corresponding to the first time step in the demand extraction side expanded phase; the second step is to find the value corresponding to the Nth time step in the demand extraction side expanded phase; the third step is to subtract the value of the first time step from the value of the Nth time step, and the result is the cumulative phase increment on the demand extraction side.

[0090] The cumulative phase increment of the review and early warning side is the total change of the expanded phase of the review side during the entire analysis period, reflecting the overall change range of the review side phase. Specifically, the first step is to find the value corresponding to the first time step in the expanded phase of the review and early warning side; the second step is to find the value corresponding to the Nth time step in the expanded phase of the review and early warning side; the third step is to subtract the value of the first time step from the value of the Nth time step, and the result is the cumulative phase increment of the review and early warning side.

[0091] The real form of the frequency ratio is a real number that characterizes the relationship between the phase change rates of the two agents, reflecting the relative magnitude of the frequencies on the demand side and the review side. Specifically, the first step is to obtain the cumulative phase increment on the demand extraction side; the second step is to obtain the cumulative phase increment on the review and early warning side; the third step is to divide the cumulative phase increment on the demand extraction side by the cumulative phase increment on the review and early warning side, and the result is the real form of the frequency ratio.

[0092] The approximation threshold controls the upper limit of the error between the integer fraction (numerator / denominator) and the real form of the frequency ratio. It is a positive decimal (such as 0.01 or 0.005). The higher the error requirement, the smaller the threshold should be.

[0093] The continued fraction method is a calculation method that converts the real form of a frequency ratio into its simplest integer fraction (numerator / denominator), ensuring that the error between the fraction and the real number is less than or equal to the approximation threshold. Specifically, the first step is to obtain the real form of the frequency ratio, denoted as R. The second step is to take the integer part of R as the first coefficient of the continued fraction and calculate the difference between R and this integer part, denoted as the remainder r1. The third step is to stop the calculation if the absolute value of r1 is less than or equal to the approximation threshold, with the numerator as the first coefficient and the denominator as 1. If r1 is greater than the approximation threshold, calculate the integer part of 1 / r1 as the second coefficient, and then calculate the remainder r2 between 1 / r1 and this integer part. The fourth step is to repeat the third step until the absolute value of the remainder in a certain step is less than or equal to the approximation threshold, and then convert all coefficients into their simplest integer fractions (numerator and denominator are coprime) according to the continued fraction rules, ensuring that |R - numerator / denominator| is less than or equal to the approximation threshold.

[0094] The integer numerator is the numerator of the simplest fraction output by the continued fraction method. It is coprime with the integer denominator and together they form the frequency ratio.

[0095] The integer denominator is the denominator of the simplest fraction output by the continued fraction method. It is coprime with the integer numerator and together they form the frequency ratio.

[0096] The frequency ratio is the simplest fraction expressed as an integer numerator / integer denominator, reflecting the ratio of the effective frequency of the demand extraction side to that of the review and early warning side (e.g., numerator=1, denominator=2, which means the frequency of the demand side is 1 / 2 of that of the review side).

[0097] In one embodiment of the present invention, calculating the frequency mismatch between the effective frequency of the demand extraction side and the effective frequency of the review and early warning side, and the coupling strength of the two update strength sequences, includes:

[0098] Calculate the difference between the effective frequency of the demand extraction side and the effective frequency of the review and early warning side, and take the absolute value of the difference as the frequency mismatch;

[0099] Define a set of discrete frequency points, which contains multiple discrete frequency points;

[0100] The frequency domain representation of the demand extraction side is obtained by performing a Discrete Fourier Transform on the demand extraction side update intensity sequence, and the frequency domain representation of the review and early warning side is obtained by performing a Discrete Fourier Transform on the review and early warning side update intensity sequence.

[0101] The square of the complex modulus of the frequency domain representation of the demand extraction side is used as the self-spectral density of the demand extraction side, and the square of the complex modulus of the frequency domain representation of the review and early warning side is used as the self-spectral density of the review and early warning side.

[0102] The product of the frequency domain representation of the demand extraction side and the frequency domain representation of the review and early warning side is used as the cross-spectral density.

[0103] For each discrete frequency point in the set of discrete frequency points, calculate the ratio of the complex modulus of the cross-spectral density to the square root of the product of the self-spectral density of the demand extraction side and the self-spectral density of the review and early warning side.

[0104] The maximum value among the above ratios corresponding to all discrete frequency points is taken as the coupling strength.

[0105] Frequency mismatch is an indicator that characterizes the degree of difference in the effective frequencies of two agents. The larger the absolute value of the difference, the greater the frequency difference. Specifically, the first step is to obtain the effective frequencies of the demand extraction side and the review and warning side; the second step is to calculate the difference between the two effective frequencies (the effective frequency of the demand side minus the effective frequency of the review side); the third step is to take the absolute value of the difference, and the result is the frequency mismatch.

[0106] The discrete frequency point set is a set of frequency values ​​used to calculate the spectral density (self-spectrum, cross-spectrum), and it needs to cover the possible range of effective frequencies of the two agents. Specifically, the first step is to determine the frequency range (take 0 to twice the maximum effective frequency of the two agents, such as 0 to 20 Hz if the maximum effective frequency is 10 Hz); the second step is to set the number of sets (take an integer power of 2, such as 256 or 512, the more the number, the higher the accuracy); the third step is to divide the frequency range into equal intervals of the corresponding number, and the endpoints of each interval are discrete frequency points, and all points form the discrete frequency point set.

[0107] The Discrete Fourier Transform (DFT) is an operation that converts the time-domain update intensity sequence into a frequency-domain representation for subsequent spectral density calculations. Specifically, the first step is to obtain the update intensity sequence of a certain side (e.g., the demand side), and denote the sequence length as M (matching the number of time steps). The second step is to calculate the sum of each value of the sequence corresponding to that frequency point multiplied by the cosine term (2π × frequency point × time step / M) minus the sine term (2π × frequency point × time step / M). The third step is to obtain a complex number (the real part is the sum of the cosine terms, and the imaginary part is the sum of the negative sine terms) from the complex numbers corresponding to all frequency points, which together form the frequency-domain representation of that side.

[0108] The frequency domain representation of the demand extraction side is a set of complex numbers output by the Discrete Fourier Transform, with each complex number corresponding to the frequency domain information (real part + imaginary part) of a discrete frequency point.

[0109] The frequency domain representation of the review and early warning side is a set of complex numbers output by the discrete Fourier transform, with each complex number corresponding to the frequency domain information (real part + imaginary part) of a discrete frequency point.

[0110] Demand decimation-side autospectral density is an index characterizing the energy distribution of the demand-side frequency domain representation. Specifically, the first step is to take a complex number at a discrete frequency point in the demand decimation-side frequency domain representation; the second step is to calculate the modulus of this complex number (the square of the real part plus the square of the imaginary part, and then taking the square root of the result); the third step is to square the modulus, and the result is the demand decimation-side autospectral density at that frequency point; the results of all frequency points constitute the demand decimation-side autospectral density set.

[0111] The self-spectral density of the review and early warning side is an index characterizing the energy distribution of the review and early warning side in the frequency domain representation. Specifically, the first step is to take a complex number at a certain discrete frequency point in the review and early warning side frequency domain representation; the second step is to calculate the modulus of the complex number (the square of the real part plus the square of the imaginary part, and then taking the square root of the result); the third step is to perform a square operation on the modulus, and the result is the self-spectral density of the review and early warning side at that frequency point; the results of all frequency points constitute the set of self-spectral densities of the review and early warning side.

[0112] Cross-spectral density is a complex number index characterizing the correlation of the frequency domain representations of two agents. Specifically, the first step is to take the demand extraction side's complex number and the review and early warning side's complex number at the same discrete frequency point. The second step is to find the complex conjugate of the review side's complex number (keeping the real part unchanged and taking the opposite of the imaginary part). The third step is to multiply the demand side's complex number with the review side's complex conjugate (multiplying the real part by the real part minus multiplying the imaginary part to get the new real part, and multiplying the real part by the imaginary part plus multiplying the imaginary part to get the new imaginary part). The resulting complex number is the cross-spectral density at that frequency point. The results for all frequency points form the cross-spectral density set.

[0113] The ratio (at discrete frequency points) is an indicator of the degree of coupling between two agents at a certain frequency point. The larger the ratio, the stronger the coupling. Specifically, the steps are as follows: First, take the complex cross-spectral density, demand-side self-spectral density, and review-side self-spectral density at a certain discrete frequency point; second, calculate the modulus of the complex cross-spectral density (square of the real part plus square of the imaginary part, then take the square root); third, calculate the product of the demand-side self-spectral density and the review-side self-spectral density, then take the square root of the product; fourth, divide the modulus of the cross-spectral density by the result of the third step, and the result is the ratio at that frequency point.

[0114] Coupling strength is an indicator that characterizes the overall coupling degree of two agents. Taking the maximum value of the ratio of all frequency points can highlight the contribution of the strongest coupling frequency.

[0115] In one embodiment of the present invention, a locking state is determined based on frequency mismatch, coupling strength, and frequency ratio; when in a locking state, the dominant locking frequency is located in the spectrum of the global risk scoring sequence, including:

[0116] The ratio of frequency mismatch to the sum of the effective frequencies on the demand extraction side and the effective frequencies on the review and early warning side is used as the normalized frequency mismatch.

[0117] Calculate the ratio of the effective frequency on the demand extraction side to the effective frequency on the review and early warning side, and the difference between the frequency ratios. Take the absolute value of this difference as the frequency ratio deviation.

[0118] Two threshold coefficients are set, both of which are positive real numbers;

[0119] When the normalized frequency mismatch is not greater than the product of the first threshold coefficient and the coupling strength, and the frequency ratio deviation is not greater than the product of the second threshold coefficient and the coupling strength, it is determined to be in a locked state.

[0120] Define a discrete frequency set, which contains multiple discrete frequency points;

[0121] The spectral amplitude is obtained by performing a discrete Fourier transform on the global risk score sequence;

[0122] The sum of the numerator of the effective frequency of the demand extraction side divided by the frequency ratio and the denominator of the effective frequency of the review and early warning side divided by the frequency ratio is used as the center estimated frequency.

[0123] Set the bandwidth, which is a positive real number;

[0124] Select discrete frequency points from the discrete frequency set that are located within an interval centered on the estimated central frequency and with a bandwidth of [missing information].

[0125] The discrete frequency corresponding to the maximum value in the spectral amplitude of the discrete frequency point within the interval is taken as the locked dominant frequency.

[0126] Normalized frequency mismatch is a mismatch index after eliminating the influence of frequency magnitude, which facilitates the determination of the locking state of dual agents with different frequency magnitudes. Specifically, the first step is to obtain the frequency mismatch, the effective frequency of the demand extraction side, and the effective frequency of the review and warning side; the second step is to calculate the sum of the effective frequency of the demand side and the effective frequency of the review side; the third step is to divide the frequency mismatch by the sum, and the result is the normalized frequency mismatch.

[0127] Frequency ratio deviation is an indicator that characterizes the difference between the actual frequency ratio and the theoretical frequency ratio (frequency ratio) of two agents. The smaller the deviation, the closer the frequency ratio is to the theoretical value. Specifically, the first step is to calculate the ratio of the effective frequency on the demand extraction side to the effective frequency on the review and warning side (effective frequency on the demand side divided by effective frequency on the review side); the second step is to obtain the frequency ratio; the third step is to calculate the difference between the result of the first step and the frequency ratio, and take the absolute value of the difference. The result is the frequency ratio deviation.

[0128] The first threshold coefficient is a positive real number that controls the normalized frequency mismatch judgment standard, with a value range of 0.1-0.3 (in the bidding scenario, it is necessary to balance the sensitivity of locked state recognition and the false judgment rate).

[0129] The second threshold coefficient is a positive real number that controls the standard for judging the deviation of the frequency ratio, with a value range of 0.1-0.3 (consistent with the value range of the first threshold coefficient to ensure consistent judgment logic).

[0130] The locked state is a state that indicates whether the frequencies of the two agents have entered a stable synchronization. It is locked when two conditions are met and not greater than 2; otherwise, it is unlocked.

[0131] Spectral amplitude is the complex modulus of the global risk scoring sequence in the frequency domain, reflecting the signal energy intensity at different frequency points. The stronger the energy, the larger the spectral amplitude.

[0132] The center estimated frequency is determined by locking the predicted position of the dominant frequency and combining the effective frequency of the two agents with the frequency ratio. Specifically, the first step is to obtain the effective frequency of the demand extraction side, the effective frequency of the review and early warning side, and the numerator and denominator of the frequency ratio. The second step is to calculate the effective frequency of the demand side divided by the numerator of the frequency ratio, and then calculate the effective frequency of the review side divided by the denominator of the frequency ratio. The third step is to calculate the sum of the two results, take half of the sum, and the result is the center estimated frequency.

[0133] The bandwidth is the search range around the center estimated frequency, used to filter possible dominant frequencies. It is set to 5%-10% of the center estimated frequency (e.g., if the center estimated frequency is 10 Hz, the bandwidth is 0.5-1 Hz).

[0134] These frequency points represent the candidate range for identifying the dominant frequency, with the interval being from the center estimated frequency minus half the bandwidth to the center estimated frequency plus half the bandwidth.

[0135] The dominant locking frequency is the main frequency in the global risk scoring sequence that is affected by the two-agent phase-locking mechanism. The largest spectral amplitude represents the strongest energy of this frequency, which is the core frequency of phase-locking.

[0136] In one embodiment of the present invention, a locking component is constructed based on the locking dominant frequency, and the difference between the global risk score sequence and the locking component is used as the delocking residual, including:

[0137] Define the time step set, the number of time steps, and the time step size. The time step set contains time steps from the first time step to the corresponding time step number.

[0138] Set the maximum positive integer for the harmonic order;

[0139] For each time step, based on the time step size of the dominant frequency harmonic order and the time step itself, construct the cosine basis function and the sine basis function respectively, where the harmonic order ranges from 1 to the maximum positive integer of the set harmonic order;

[0140] The cosine basis function and sine basis function corresponding to each time step are arranged into a vector for that row in order of harmonic order. The vectors corresponding to all time steps together form the design matrix.

[0141] The values ​​of the global risk score sequence at each time step are arranged in time step order to form a risk score vector;

[0142] The parameter vector is calculated using the least squares method. This parameter vector is obtained by multiplying the inverse of the product of the transpose of the design matrix and the design matrix, and then multiplying the transpose of the design matrix and the risk score vector.

[0143] For each time step, the coefficients of each harmonic order in the parameter vector are multiplied by the cosine basis function and sine basis function of each harmonic order for that time step, and then all the product results are summed to obtain the locked component for that time step.

[0144] Calculate the difference between the global risk score sequence value at each time step and the locked component at that time step. This difference is the unlocking residual at that time step.

[0145] The maximum positive integer of the harmonic order is the highest order of the dominant harmonic of the locked frequency, which is used to cover the harmonic components related to phase lock. The value range is 1-5 (in the bidding scenario, the energy of harmonics within the third order is already significant, and the energy of harmonics above the fifth order can be ignored to avoid redundant calculations).

[0146] The cosine basis function is a cosine-based fundamental function for fitting locked components, which varies with the time step and harmonic order. Specifically, the first step is to obtain the current time step t, the locked dominant frequency, a certain harmonic order k (k ranges from 1 to the largest positive integer), and the time step size. The second step is to calculate the result of 2 multiplied by π multiplied by the locked dominant frequency multiplied by the harmonic order k multiplied by the time step size multiplied by the current time step t. The third step is to take the cosine value of the result, which is the cosine basis function for the current time step t and harmonic order k.

[0147] The sinusoidal basis function is a sinusoidal fundamental function for fitting the locked component, which varies with the time step and harmonic order. Specifically, the first step is to obtain the current time step t, the locked dominant frequency, a certain harmonic order k (k ranges from 1 to the largest positive integer), and the time step size. The second step is to calculate the result of 2 multiplied by π multiplied by the locked dominant frequency multiplied by the harmonic order k multiplied by the time step size multiplied by the current time step t. The third step is to take the sine value of the result, which is the sinusoidal basis function for the current time step t and harmonic order k.

[0148] The design matrix is ​​a matrix used for least squares fitting, with each row corresponding to a combination of basis functions for one time step. The construction logic is as follows: for each time step, row vectors are formed in the order of harmonic order 1 cosine basis function → harmonic order 1 sine basis function → harmonic order 2 cosine basis function → harmonic order 2 sine basis function → … → maximum harmonic order cosine basis function → maximum harmonic order sine basis function. All row vectors are arranged vertically to form the design matrix.

[0149] The risk score vector is a vector form of the global risk score sequence, with elements being the score value at each time step, arranged in the order of time steps (e.g., score at time step 1 → score at time step 2 → ... → score at time step N).

[0150] The least squares method is a mathematical method for solving parameter vectors, with the aim of minimizing the error between the design matrix × parameter vector and the risk score vector.

[0151] The parameter vector is the coefficient vector of the least squares fitting, containing the coefficients of the cosine and sine basis functions of each harmonic order. Specifically, the steps are as follows: First, calculate the transpose of the design matrix (interchange rows and columns); second, calculate the product of the transposed design matrix and the original design matrix; third, find the inverse of the product matrix (if the matrix is ​​not invertible, use singular value decomposition to find the pseudo-inverse); fourth, calculate the product of the transposed design matrix and the risk scoring vector; fifth, multiply the inverse matrix (or pseudo-inverse) obtained in the third step by the product result of the fourth step, and the resulting vector is the parameter vector.

[0152] The locked component is the periodic component in the global risk score caused by phase-locking between two agents. Specifically, the first step is to obtain the parameter vector (containing cosine and sine coefficients for harmonic orders 1 to the maximum order), the cosine basis functions and sine basis functions for each harmonic order, at the current time step t. The second step is to multiply the cosine basis function of time t by the cosine coefficient corresponding to k in the parameter vector, and multiply the sine basis function of time t by the sine coefficient corresponding to k. The third step is to add the above two products for all harmonic orders, and the sum is the locked component at time step t.

[0153] The delocking residual is the remaining part of the risk score after removing the periodic component of the phase-locked loop, reflecting the real risk related to the document content. Specifically, the first step is to obtain the global risk score sequence value at the current time step t; the second step is to obtain the locking component at time step t; the third step is to subtract the locking component from the global risk score value, and the difference is the delocking residual at time step t.

[0154] In one embodiment of the present invention, the locked phase is defined based on the frequency ratio, a phase cross section is set to determine the set of in-phase times, and statistics of the delocking residual are calculated on the set of in-phase times to obtain the content signal-to-noise ratio, including:

[0155] Define a set of time steps, which contains all time steps from the first time step to the Nth time step, where N is the number of time steps;

[0156] The first product is obtained by multiplying the integer numerator of the frequency ratio by the instantaneous phase of the review and warning side, and the second product is obtained by multiplying the integer denominator of the frequency ratio by the instantaneous phase of the demand extraction side. The difference between the first product and the second product is calculated and used as the locking phase.

[0157] Set the phase cross-section angle and the phase window width, where the phase window width is a positive real number; perform phase foldback processing on the difference between the locked phase and the phase cross-section angle at each time step, and take the absolute value of the difference after processing;

[0158] Select time steps whose absolute value is not greater than the width of the phase window to form a set of in-phase moments, and count the number of elements in the set of in-phase moments;

[0159] Calculate the sum of the delocking residuals for all time steps within the in-phase time set, and divide the sum by the number of elements in the in-phase time set to obtain the mean of the delocking residuals.

[0160] Calculate the difference between the delocking residual and the mean for each time step within the in-phase time set. Square each difference and sum all the squaring results. Divide this sum by the number of elements in the in-phase time set to obtain the variance of the delocking residual.

[0161] Calculate the sum of squares of the locked components of all time steps within the in-phase time set, divide the sum of squares by the number of elements in the in-phase time set to obtain the locked energy density; set a positive real constant, calculate the ratio of the variance of the delocking residual to the sum of the positive real constant and the locked energy density, and use this ratio as the content signal-to-noise ratio.

[0162] The first product is an intermediate variable for constructing the locked phase, reflecting the relationship between the instantaneous phase on the review side and the numerator of the frequency ratio. Specifically, the first step is to obtain the integer numerator of the frequency ratio; the second step is to obtain the instantaneous phase on the review warning side at a certain time step; the third step is to multiply the integer numerator by the instantaneous phase, and the result is the first product at that time step.

[0163] The second product is an intermediate variable for constructing the locked phase, reflecting the relationship between the instantaneous phase on the demand side and the denominator of the frequency ratio. Specifically, the first step is to obtain the integer denominator of the frequency ratio; the second step is to obtain the instantaneous phase on the demand extraction side at a certain time step; the third step is to multiply the instantaneous phase by the integer denominator, and the result is the second product at that time step.

[0164] The lock phase is a phase index that characterizes the phase-locked synchronization state of two agents. The smaller the difference, the stronger the synchronization. Specifically, the first step is to obtain the first product and the second product of a certain time step. The second step is to subtract the second product from the first product, and the result is the lock phase of that time step.

[0165] The phase cross-section angle is the reference phase for screening in-phase moments, and its value ranges from negative π to positive π. It is usually set to 0 (corresponding to the reference state for phase synchronization of two agents).

[0166] The phase window width is the maximum range of deviation from the phase cross-section angle allowed, with a value of 0.1π-0.3π (e.g., 0.2π, approximately 36 degrees, balancing the number of in-phase moments with synchronization accuracy).

[0167] Phase reversal processing is an operation that adjusts the difference between the locked phase and the phase cross-section angle to within the range of negative π to positive π, avoiding abnormally large phase differences due to periodicity. Specifically, the first step is to calculate the difference between the locked phase and the phase cross-section angle at a certain time step; the second step is to subtract 2π from the difference if the difference is greater than π; if the difference is less than negative π, add 2π to the difference; if the difference is between negative π and positive π, it remains unchanged; the third step is to obtain the difference after reversal processing.

[0168] The absolute value of the difference after processing is the basis for determining whether the time step belongs to the in-phase moment. The smaller the absolute value, the closer it is to the reference phase section.

[0169] The set of in-phase moments is the set of time steps where the locked phase is close to the phase cross-section angle, reflecting the period when the synchronization between the two agents is relatively high.

[0170] The number of elements in the set of in-phase moments is the denominator for calculating the delocking residual statistic and the locking energy density, reflecting the sample size at in-phase moments.

[0171] The mean of the delocking residuals is the average level of the residuals within the same phase time step, reflecting the benchmark value of content-related risk. Specifically, the first step is to obtain the delocking residuals for all time steps within the same phase time step set; the second step is to calculate the sum of these residuals; the third step is to divide the sum by the number of elements in the same phase time step set, and the result is the mean of the delocking residuals.

[0172] The variance of the delocking residual is the degree of dispersion of the residual within the same phase time step, reflecting the fluctuation range of content-related risk. Specifically, the first step is to obtain the delocking residual and the residual mean for each time step within the set of same phase time steps; the second step is to calculate the difference between each residual and the mean, and square each difference; the third step is to calculate the sum of all the squared results; the fourth step is to divide the sum by the number of elements in the set of same phase time steps, and the result is the variance of the delocking residual.

[0173] Lock-in energy density is the average energy of the lock-in components within the in-phase time step, reflecting the strength of the phase-locked components. Specifically, the first step is to obtain the lock-in components of all time steps within the in-phase time step set; the second step is to calculate the square of each lock-in component and then sum all the squared results; the third step is to divide the sum by the number of elements in the in-phase time step set, and the result is the lock-in energy density.

[0174] Positive real constants are tiny values ​​that prevent the denominator from being zero due to the locked energy density being zero. Their values ​​range from 10 to the power of -6 to 10 to the power of -4 (e.g., 10 to the power of -5 will not affect the normal calculation results).

[0175] Content signal-to-noise ratio (SNR) is an indicator that characterizes the relative strength of the content-related residual energy and the phase-locked component energy. The larger the ratio, the more significant the content signal. Specifically, the first step is to obtain the variance, positive real constant, and lock energy density of the delocking residual; the second step is to calculate the sum of the positive real constant and the lock energy density; and the third step is to divide the variance of the delocking residual by the sum, and the result is the content signal-to-noise ratio.

[0176] In one embodiment of the present invention, the statistics of the delocking residuals and the content signal-to-noise ratio are processed by a mapping function to output a document-level risk score, including:

[0177] Take the square root of the variance of the locked residual to obtain the standard deviation of the locked residual;

[0178] Four real-valued weight parameters are defined as the first weight parameter, the second weight parameter, the third weight parameter, and the fourth weight parameter.

[0179] Calculate the product of the first weight parameter, the mean of the locked residual, and the second weight parameter; the product of the standard deviation of the locked residual and the third weight parameter; and the product of the natural logarithm of the sum of the content signal-to-noise ratio and the fourth weight parameter. Use this sum as the intermediate score.

[0180] Define a mapping function that is the sum of one divided by one raised to the power of the negative intermediate score and the natural constant; substitute the intermediate score into the mapping function to obtain the document-level risk score.

[0181] The standard deviation of the delocking residual is an indicator of the degree of dispersion of the residual within the same phase time. It is easier to understand intuitively than the variance (it is of the same order of magnitude as the residual). Specifically, the first step is to obtain the variance of the delocking residual; the second step is to perform a square root operation on the variance (if the variance is 0, the standard deviation is 0); the third step is to obtain the standard deviation of the delocking residual.

[0182] The first weighting parameter is the baseline weight of the intermediate score, which is used to set the basic level of risk scoring. The value range is 0.2-0.4 (in bidding scenarios, the baseline score needs to balance the risk-free and high-risk ranges to avoid extreme deviations).

[0183] The second weighting parameter is the weight for locking the mean of the residuals, which is used to adjust the impact of the content risk benchmark level on the final score. The value range is 0.3-0.5 (the mean reflects the core content risk, and the weight must be higher than other non-benchmark indicators).

[0184] The third weighting parameter is the weight for locking the residual standard deviation, which is used to adjust the impact of content risk fluctuations on the final score. The value range is 0.1-0.2 (fluctuation is an auxiliary indicator, and its weight is lower than the mean).

[0185] The fourth weighting parameter is a weight of the natural logarithm of the sum of the content signal-to-noise ratio and the content signal-to-noise ratio. It is used to adjust the influence of the content signal strength on the final score, and its value ranges from 0.1 to 0.2 (the signal-to-noise ratio is a reliability indicator, and the weight is equivalent to the standard deviation).

[0186] The intermediate score is a comprehensive indicator that integrates the baseline score, the mean of content risk, risk volatility, and signal reliability. Its value has no fixed range and needs to be normalized using a mapping function. Specifically, the process involves: first, calculating the mean of the delocking residuals multiplied by the second weighting parameter; second, calculating the standard deviation of the delocking residuals multiplied by the third weighting parameter; third, calculating the sum of the first weighting parameter and the content signal-to-noise ratio, taking the natural logarithm of this sum, and multiplying it by the fourth weighting parameter; and fourth, adding the first weighting parameter to the results of the first, second, and third steps to obtain the final intermediate score.

[0187] The mapping function is a function that normalizes intermediate scores to the 0-1 range (i.e., the sigmoid function), ensuring that document-level risk scoring conforms to the business convention of 0 being the lowest risk and 1 being the highest risk. Specifically, the first step is to determine the value of the natural constant (usually 2.71828, but five decimal places can be retained for simplified calculation); the second step is to calculate the negative power of the natural constant for the intermediate score (i.e., the intermediate score for which the exponent of the natural constant is negative); the third step is to calculate the sum of one and the result of this power; the fourth step is to divide one by this sum, and the result is the output value of the mapping function.

[0188] The natural constant is the core constant in the mapping function, used for exponential operations, and its usual value is 2.71828.

[0189] Document-level risk score is the final result that characterizes the overall risk level of the tender document. The value ranges from 0 to 1, with higher risk being closer to 1 and lower risk being closer to 0.

[0190] It's important to note that document-level risk scoring provides bidding teams with actionable risk assessment criteria: a score close to 1 indicates high content risk in the bid documents (such as substandard technical solutions or missing compliance clauses), requiring priority adjustment of those clauses; a score close to 0 represents a low level of risk, allowing the bidding process to proceed. Furthermore, the signal-to-noise ratio (SNR) associated with the document-level risk score further informs decision-makers about its reliability; a high SNR indicates less susceptibility to phase-locked loop interference, leading to higher decision-making credibility. In summary, document-level risk scoring is a cornerstone for automating the entire bidding process and providing decision support.

[0191] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A bidding full-process automation and decision support system based on multi-agent collaboration, characterized in that, The method comprises the following steps: Data acquisition module, collect demand extraction side update intensity sequence, review warning side update intensity sequence and global risk score sequence, including: Set a time step set, which contains all time steps from the first time step to the Nth time step, where N is the total number of time steps, and each time step is represented by t; At each time step t, obtain the demand extraction side vector evidence and the review warning side vector evidence, and record the value of the global risk score sequence at the time step t; Calculate the value of the demand extraction side update intensity sequence at time step t, specifically: the Euclidean norm of the difference between the vector evidence of the demand extraction side at time step t and the vector evidence at time step t-1; wherein t is valued from the second time step to the Nth time step; Calculate the value of the review warning side update intensity sequence at time step t, specifically: the Euclidean norm of the difference between the vector evidence of the review warning side at time step t and the vector evidence at time step t-1, wherein t is valued from the second time step to the Nth time step; Data analysis module, based on the demand extraction side instantaneous phase and the review warning side instantaneous phase, perform phase relationship analysis to obtain the frequency ratio; Coupling strength calculation module, calculate the frequency mismatch of the demand extraction side effective frequency and the review warning side effective frequency, and the coupling strength of the two update intensity sequences; Locking state determination module, determine the locking state according to the frequency mismatch, coupling strength and frequency ratio; when in the locking state, locate the locking dominant frequency in the frequency spectrum of the global risk score sequence; Unlocking residual module, based on the locking dominant frequency, construct a locking component, and take the difference between the global risk score sequence and the locking component as the unlocking residual; ​ ​ ​ ​ ​ ​ ​ The signal-to-noise ratio calculation module defines a locked phase based on a frequency ratio, sets a phase section to determine an in-phase time set, calculates a statistic of a de-locking residual on the in-phase time set, and obtains a content signal-to-noise ratio; The risk score module processes the statistic of the de-locking residual and the content signal-to-noise ratio through a mapping function, and outputs a document-level risk score.

2. The multi-agent collaboration based tender full-process automation and decision support system according to claim 1, characterized in that, The phase relationship analysis is performed based on the instantaneous phase of the demand extraction side and the instantaneous phase of the review warning side to obtain a frequency ratio, including: A time step set and a time step number are set, and the time step number is N; The instantaneous phase of the demand extraction side is phase-expanded to obtain a demand extraction side expanded phase, and the instantaneous phase of the review warning side is phase-expanded to obtain a review warning side expanded phase; The difference between the demand extraction side expanded phase at the first time step and the Nth time step is calculated as a demand extraction side cumulative phase increment, and the difference between the review warning side expanded phase at the first time step and the Nth time step is calculated as a review warning side cumulative phase increment; The ratio of the demand extraction side cumulative phase increment to the review warning side cumulative phase increment is taken as a real number form of the frequency ratio; An approximation threshold is set, and a continued fraction method is used to determine an integer numerator and an integer denominator, so that the difference between the real number form and the result of the numerator divided by the denominator is not greater than the approximation threshold, and the frequency ratio is represented by the numerator and the denominator. 3.The multi-agent collaboration based tender full-process automation and decision support system according to claim 2, characterized in that, The frequency mismatch between the demand extraction side effective frequency and the review warning side effective frequency, and the coupling strength of the two update intensity sequences, are calculated, including: The difference between the demand extraction side effective frequency and the review warning side effective frequency is calculated, and the absolute value of the difference is taken as the frequency mismatch; A discrete frequency point set is set, which includes a plurality of discrete frequency points; The discrete Fourier transform is performed on the demand extraction side update intensity sequence to obtain a demand extraction side frequency domain representation, and the discrete Fourier transform is performed on the review warning side update intensity sequence to obtain a review warning side frequency domain representation; The square of the complex modulus of the demand extraction side frequency domain representation is calculated as the demand extraction side autospectrum density, and the square of the complex modulus of the review warning side frequency domain representation is calculated as the review warning side autospectrum density; The product of the complex conjugate of the demand extraction side frequency domain representation and the review warning side frequency domain representation is calculated as the cross-spectrum density; For each discrete frequency point in the discrete frequency point set, the ratio of the complex modulus of the cross-spectrum density to the square root of the product of the demand extraction side autospectrum density and the review warning side autospectrum density is calculated; The maximum value of the above ratio corresponding to all discrete frequency points is taken as the coupling strength.

4. The multi-agent collaboration based tender full-process automation and decision support system according to claim 3, characterized in that, The locked state is determined according to the frequency mismatch, the coupling strength, and the frequency ratio; When in the locked state, the locked dominant frequency is located in the frequency spectrum of the global risk score sequence, including: The ratio of the frequency mismatch to the sum of the demand extraction side effective frequency and the review warning side effective frequency is calculated as a normalized frequency mismatch; The difference between the ratio of the demand extraction side effective frequency to the review warning side effective frequency and the frequency ratio is taken as the frequency ratio deviation; Two threshold coefficients are set, both of which are positive real numbers; When the normalized frequency mismatch is not greater than the product of the first threshold coefficient and the coupling strength, and the frequency ratio deviation is not greater than the product of the second threshold coefficient and the coupling strength, it is determined that the locked state is reached. Set a discrete frequency set, the discrete frequency set contains a plurality of discrete frequency points; Perform a discrete Fourier transform on the global risk score sequence to obtain a spectrum amplitude; Calculate the sum of the numerator of the demand extraction side effective frequency divided by the frequency ratio and the denominator of the review early warning side effective frequency, and take half of the sum as the center estimated frequency; Set a frequency bandwidth, the frequency bandwidth is a positive real number; Select a discrete frequency point in the discrete frequency set within the interval centered on the center estimated frequency and with the frequency bandwidth as the range; Take the discrete frequency point corresponding to the maximum value in the spectrum amplitude corresponding to the discrete frequency point in the interval as the locked dominant frequency. 5.The multi-agent collaboration based tender full-process automation and decision support system according to claim 4, characterized in that, Based on the locked dominant frequency, construct a locked component, and take the difference between the global risk score sequence and the locked component as the unlocked residual, including: Set a time step set and a time step length, where the time step set contains time steps from the first time step to the time step number corresponding to the time step number; Set the maximum positive integer of the harmonic order number; For each time step, construct a cosine basis function and a sine basis function based on the locked dominant frequency, the harmonic order number, the time step length, and the time step, where the harmonic order number is from 1 to the set maximum positive integer of the harmonic order number; Arrange the cosine basis function and the sine basis function corresponding to each time step in the order of the harmonic order number to form a vector of the row, and arrange the vectors corresponding to all time steps to form a design matrix; Arrange the values of the global risk score sequence at each time step in the order of the time step to form a risk score vector; Calculate the parameter vector by multiplying the inverse matrix of the product of the transpose of the design matrix and the design matrix, and then multiplying the product of the transpose of the design matrix and the risk score vector; For each time step, multiply the coefficients corresponding to each harmonic order number in the parameter vector by the cosine basis function and the sine basis function corresponding to each harmonic order number of the time step, and then sum all the products to obtain the locked component of the time step. Calculate the difference between the value of the global risk score sequence at each time step and the locked component of the time step, which is the unlocked residual of the time step. 6.The multi-agent collaboration based tender full-process automation and decision support system according to claim 5, characterized in that, Define a locked phase based on the frequency ratio, set a phase section to determine a set of in-phase time instants, calculate a statistic of the unlocked residual on the set of in-phase time instants, and obtain a content signal-to-noise ratio, including: Set a time step set, the time step set contains all time steps from the first time step to the Nth time step, where N is the number of time steps; Take the integer multiple of the numerator of the frequency ratio and the review early warning side instantaneous phase to obtain a first product, take the integer multiple of the denominator of the frequency ratio and the demand extraction side instantaneous phase to obtain a second product, calculate the difference between the first product and the second product, and take the difference as the locked phase; Set a phase section angle and a phase window width, where the phase window width is a positive real number; perform phase folding processing on the difference between the locked phase of each time step and the phase section angle, and take the absolute value of the processed difference; Select time steps with absolute values not greater than the phase window width to form a set of in-phase time instants, and count the number of elements in the set of in-phase time instants; Calculate the sum of the unlocked residuals of all time steps in the set of in-phase time instants, divide the sum by the number of elements in the set of in-phase time instants, and obtain the mean of the unlocked residuals. Calculate the difference between the de-locking residual of each time step in the in-phase time set and the mean value, square each difference, sum all the square results, divide the sum by the number of elements in the in-phase time set, and obtain the variance of the de-locking residual; Calculate the square sum of the locking component of all time steps in the in-phase time set, divide the square sum by the number of elements in the in-phase time set, and obtain the locking energy density; Set a positive real constant, calculate the ratio of the variance of the de-locking residual to the sum of the positive real constant and the locking energy density, and take the ratio as the content signal-to-noise ratio.

7. The multi-agent collaboration based tender full-process automation and decision support system according to claim 6, characterized in that, After processing the statistics of the de-locking residual and the content signal-to-noise ratio by the mapping function, the document-level risk score is output, including: Take the square root of the variance of the de-locking residual to obtain the standard deviation of the de-locking residual; Set four real weight parameters, which are the first weight parameter, the second weight parameter, the third weight parameter, and the fourth weight parameter; Calculate the product of the first weight parameter plus the mean value of the de-locking residual and the second weight parameter, plus the product of the standard deviation of the de-locking residual and the third weight parameter, plus the natural logarithm of the sum of one and the content signal-to-noise ratio, and the fourth weight parameter, and take the sum as the intermediate score; Set the mapping function, which is one divided by the sum of one and the negative intermediate score power of the natural constant; Substitute the intermediate score into the mapping function to obtain the document-level risk score.

Images