Whole-process consultation project risk prediction method based on artificial intelligence

Abstract

The invention discloses a whole-process consultation project risk prediction method based on artificial intelligence, and belongs to the technical field of project management risk prediction, and the method comprises the steps: collecting a progress report, a quality detection report and multi-source heterogeneous data of a resource use condition in a project management process, and carrying out the manual marking of the collected project process data. Sample generation is carried out based on an adaptive synthetic sampling SMOTE algorithm, and project process data expansion is realized. A deep neural network classifier is used for carrying out whole-process consultation project risk prediction, the deep neural network classifier comprises a time sequence feature separation layer, and a weight updating process is optimized through enhanced global search. According to the SMOTE algorithm based on the dynamic parameter gating mechanism, through the steps of weight matrix initialization, high-dynamic feature anomaly detection, feature importance dynamic evaluation, synthetic sample generation and adjustment and the like, synthetic samples related to high-risk features are preferentially generated, the problem of sample imbalance is effectively solved, and the accuracy of the SMOTE algorithm is improved. And the sensitivity of the model to key risk characteristics is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of project management risk prediction, and particularly relates to a risk prediction method for the whole-process consulting project based on artificial intelligence. Background Art

[0002] In modern project management, the whole-process consulting project faces a complex and changeable risk environment. Traditional risk management and prediction methods are difficult to meet the needs of accurately identifying and effectively controlling the risks of the whole project process. There are many deficiencies in the existing technologies in dealing with multi-source heterogeneous data, solving the problem of sample imbalance, and optimizing the risk prediction model, resulting in the inability to effectively guarantee the accuracy and timeliness of risk prediction, thus affecting the smooth progress and final delivery of the project. In addition, with the rapid development of artificial intelligence technology, there are still many challenges in applying artificial intelligence technology to the risk prediction of the whole-process consulting project.

[0003] The existing publicly disclosed patents have the following technical problems: 1. When dealing with sample imbalance, the existing technologies usually adopt simple oversampling or undersampling methods, which cannot effectively solve the problem of fewer high-risk samples, resulting in insufficient recognition ability of the model for high-risk features. 2. The risk prediction models in the existing technologies often lack an independent processing and dynamic optimization mechanism for key features, resulting in lower sensitivity of the model to problems such as project schedule delay and cost overrun, and low prediction accuracy. 3. The models of the existing technologies lack a dynamic adjustment mechanism during the training process and cannot automatically optimize the weight update direction according to the changes in project data, resulting in poor adaptability of the model when facing complex project data. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the above-mentioned shortcomings of the existing technologies and provide a risk prediction method for the whole-process consulting project based on artificial intelligence.

[0005] The technical solution adopted to solve the above technical problem is: A risk prediction method for the whole-process consulting project based on artificial intelligence, comprising the following steps:

[0006] S1, Project process data collection and annotation

[0007] Collect multi-source heterogeneous data such as progress reports, quality inspection reports, and resource usage in the project management process, and manually annotate the collected project process data. The annotation categories include three categories: low risk, medium risk, and high risk;

[0008] S2, Project process data expansion

[0009] Generate samples based on the SMOTE algorithm of adaptive synthetic sampling to achieve project process data expansion; S3, Risk prediction of the whole-process consulting project

[0010] Use a deep neural network classifier for the risk prediction of the whole process consulting project. The deep neural network classifier includes a time series feature separation layer, and the project progress ratio is R a and the time delay is T a Allocate independent neuron channels and optimize the weight update process through enhanced global search.

[0011] Furthermore, the multi-source heterogeneous data in the above S1 includes the project progress ratio R a , the quality score Q a , the resource consumption rate C a , the personnel change frequency P a , the cost deviation B a , the time delay T a , the number of contract changes M a , the risk event occurrence rate E a , the customer satisfaction S a and the number of legal disputes L a .

[0012] Furthermore, the process of sample generation based on the SMOTE algorithm with adaptive synthetic sampling in the above S2 includes the following steps:

[0013] S201. Initialize the dynamic parameter gating mechanism: Based on the dimension D of data project management and the preset number of features F fs , randomly generate the initial weight matrix and the initial bias vector and use the feature variability threshold θ c as the constraint condition for subsequent sample generation;

[0014] The dimension D is determined by the number of original data features. When the number of features F fs = 10, set D to 0.5 times the number of features;

[0015] The initial weight matrix is initialized using the Xavier normal distribution, and its standard deviation is

[0016] S202. High-dynamic feature anomaly detection: For the resource consumption rate C a and the cost deviation B a , identify the critical point of abnormal resource allocation in the data project management process through the coefficient of variation of features CV c , where the coefficient of variation of features CV c is calculated based on the ratio of the standard deviation σ c of the feature value to the mean μ c ;

[0017] S203. Dynamic evaluation of feature importance: Use the information gain ratio Gc (f) Quantify the frequency of personnel changes P a and the incidence rate of risk events E a for relevance, and based on the information gain ratio G c (f) Dynamically adjust the interpolation weight λ c to preferentially generate synthetic samples of high-risk features;

[0018] S204, Synthetic sample generation and adjustment: According to the real-time feature distribution and the interpolation weight λ c , combined with the dynamic adjustment coefficient γ c , generate new samples through the interpolation formula where γ c is automatically adjusted according to the chaos degree of project progress data to control the generation intensity of high-risk data;

[0019] S205, Repeat the above steps until the preset number of synthetic samples n c is reached, which is used to train the risk prediction model to optimize resource allocation and schedule delay prediction.

[0020] Furthermore, the initial weight matrix in the above S201 has a dimension of D×F fs where F fs is 10 preset features;

[0021] The feature variability threshold θ c is determined by calculating the variance of the original dataset x c , where x c is a matrix containing N c samples.

[0022] Furthermore, the feature coefficient of variation CV c in the above S202 has the following calculation formula:

[0023]

[0024] where the standard deviation σ c and the mean μ c are calculated respectively by the following formulas:

[0025]

[0026] In the formula, N c is the total number of samples, x c,i is the i-th sample. When CV c exceeds θ c , mark the corresponding feature as a high-risk area.

[0027] Furthermore, the adjustment formula for the interpolation weight λ c in the above S203 is:

[0028]

[0029] Wherein, ∑ f∈F G c (f) is the sum of the information gain ratios of all features, F is the set of all features,

[0030] G c (f) is the information gain ratio of feature f;

[0031] G c (f) is defined as:

[0032]

[0033] Wherein, H(Y) is the entropy of the project progress data, H(Y|f) is the conditional entropy given feature f, and H(f) is the entropy of feature f.

[0034] Furthermore, the dynamic adjustment coefficient γ in S204 c is adjusted exponentially according to the entropy of the project progress data, and the calculation formula is:

[0035] γ c = γ 0 ·e -βH(Y)

[0036] Wherein, γ 0 is the initial adjustment coefficient, e is the natural constant, β is the attenuation factor, and the value range of the attenuation factor β is 0.1 ≤ β ≤ 0.5;

[0037] The formula for interpolating to generate a new sample is:

[0038] x c new = x c + γ c ·λ c ·(x' c - x c )

[0039] Wherein, x' c and x c are two randomly selected minority class samples, and x c new is the synthesized new sample.

[0040] Furthermore, the training process of the deep neural network classifier in S3 includes the following steps:

[0041] S301, initialize the parameters of the deep neural network classifier: construct a 6-layer fully connected deep neural network classifier, and according to the preset number of features and the characteristics of the engineering data, for the project progress ratio R with time series correlationa associated with the time delay T a Allocate independent hidden layer neurons and randomly generate the initial weight matrix and the initial bias vector to match the feature dimension;

[0042] The 6 - layer fully - connected deep neural network is: the parallel cascade layer of the time - series feature separation layer and the input layer, hidden layer 1, hidden layer 2, hidden layer 3, hidden layer 4, output layer, and the dropout rate is set to 0.2;

[0043] S302, enhanced global search optimization: By calculating the Euclidean distance between the weight matrix and the optimal weight dynamically adjust the reflection intensity γ p , and combine the hyperbolic tangent activation function to constrain the magnitude of the weight update ΔW p to optimize the model's ability to capture the inflection point of the progress delay by dynamically adjusting the weight update amount;

[0044] S303, feature - sensitivity - driven weight adjustment: Based on the feature sensitivity and the local curvature optimize the weight update direction through the composite loss function L p and preferentially correct the neuron connection weights corresponding to the highly sensitive features;

[0045] S304, iterative training until the stop condition is met: Repeat the above steps to make the model converge to the optimal parameters of risk prediction and output the early warning of project cost over - run and the probability of progress delay.

[0046] Furthermore, in the S301, the dimension of the initial weight matrix is F fs ×P fs , where F fs = 10 is the preset number of features, and P fs is the number of hidden layer neurons;

[0047] The dimension of the initial bias vector is the same as P fs ;

[0048] In the S302, the calculation formula of the reflection intensity γ p is:

[0049]

[0050] In the formula, κ p is the reflection intensity sensitivity parameter, exp is the natural exponential function, is the Euclidean distance between the weight at the t - th iteration and the optimal weight, is the weight for the current t-th iteration, is the optimal weight;

[0051] The weight update amount ΔW p is calculated as follows:

[0052]

[0053] where λ p is the learning rate of the deep neural network classifier, γ p is the reflection intensity, and tanh is the hyperbolic tangent activation function.

[0054] Furthermore, the composite loss function L in S303 p is defined as a weighted combination of cross-entropy loss and root mean square loss, and the calculation formula is:

[0055]

[0056] In the formula, is the true label of the i-th p sample, is the predicted probability of the i-th p sample by the deep neural network classifier, and N p is the number of batch samples input to the deep neural network classifier;

[0057] The feature sensitivity is calculated as follows:

[0058]

[0059] In the formula, is the input value of the i-th c feature, and ya is the predicted probability of the sample by the deep neural network classifier.

[0060] The beneficial effects of the present invention are as follows: (1) The SMOTE algorithm based on the dynamic parameter gating mechanism of the present invention preferentially generates synthetic samples related to high-risk features through steps such as initializing the weight matrix, high-dynamic feature anomaly detection, dynamic evaluation of feature importance, and generation and adjustment of synthetic samples, which can effectively solve the problem of sample imbalance and enhance the sensitivity of the model to key risk features.

[0061] (2) The deep neural network classifier containing the time series feature separation layer in the present invention improves the model's ability to identify risk features and prediction accuracy by independently processing key time series features such as project progress ratio and time delay, and combining enhanced global search optimization and weight adjustment driven by feature sensitivity.

[0062] (3) In the present invention, a composite loss function combining cross-entropy and root mean square loss is used, and the weight update direction is dynamically adjusted based on feature sensitivity and local curvature, which can preferentially optimize the features that have a greater impact on the prediction results and further improve the performance of the model. Description of the Drawings

[0063] Figure 1 It is a comparison chart of F1 scores in different algorithms of the present invention and a comparison chart of ROC-AUC in different algorithms.

[0064] Figure 2 It is a comparison chart of the accuracy rates of different models in project risk prediction.

[0065] Figure 3 A comparison chart of F1 scores of different models in schedule delay prediction. Detailed Implementation Manner

[0066] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0067] The method for predicting the risks of the whole-process consulting project based on artificial intelligence in this embodiment includes the following steps:

[0068] S1, Collection and Annotation of Project Process Data

[0069] In the process of project management, multi-source heterogeneous data such as progress reports, quality inspection reports, and resource usage conditions are collected, and the collected project process data is manually annotated. The annotation categories include three categories: low risk, medium risk, and high risk.

[0070] Progress report: Extract time series data such as the project progress ratio R a and time delay T a and record them at a weekly granularity.

[0071] Quality inspection report: Record the quality score Q a and quantify it from 0 to 100 points.

[0072] Resource usage conditions: Statistic the resource consumption rate C a and cost deviation B a etc. Among them, C a is the ratio of actual resource consumption to the plan, and B a is the percentage deviation of the actual cost from the budget.

[0073] Other data: Personnel change frequency P a and contract change times M a and risk event occurrence rate E a, Customer satisfaction S a , Number of legal disputes L a .

[0074] The multi-source heterogeneous data includes the project progress ratio R a , Quality score Q a , Resource consumption rate C a , Personnel change frequency P a , Cost deviation B a , Time delay T a , Number of contract changes M a , Risk event occurrence rate E a , Customer satisfaction S a and the number of legal disputes L a .

[0075] S2, Project process data augmentation

[0076] Generate samples based on the SMOTE algorithm with adaptive synthetic sampling to achieve project process data augmentation.

[0077] The process of generating samples based on the SMOTE algorithm with adaptive synthetic sampling includes the following steps:

[0078] S201, Initialize the dynamic parameter gating mechanism: Based on the dimension D of data project management and the preset number of features F fs , Randomly generate the initial weight matrix and the initial bias vector and use the feature variability threshold θ c as the constraint condition for subsequent sample generation.

[0079] The dimension D is determined by the number of original data features. When the number of features F fs = 10, set D to 0.5 times the number of features.

[0080] The initial weight matrix is initialized using the Xavier normal distribution, and its standard deviation is

[0081] The initial weight matrix has a dimension of D × F fs , where F fs is the preset 10 features.

[0082] The feature variability threshold θ c is determined by calculating the variance of the original dataset x c , where x c is a matrix containing N c samples.

[0083] S202. High-dynamic feature anomaly detection: For the resource consumption rate C a and the cost deviation B a , identify the critical points of abnormal resource allocation in the data item management process through the coefficient of variation of features CV c , where the coefficient of variation of features CV c is calculated based on the standard deviation σ c of the feature values and the ratio to the mean μ c .

[0084] By calculating the coefficient of variation of features, the key points of possible resource waste or cost out-of-control in the project process can be identified. When the coefficient of variation of these features exceeds the threshold θ c , mark these features as high-risk areas, so that when generating synthetic samples during data augmentation, synthetic samples related to these high-risk features are preferentially generated. For example, when the resource consumption rate fluctuates abnormally, more synthetic data related to resource management can be generated to enhance the sensitivity to such problems when training the model. The calculation formula of the coefficient of variation of features CV c is:

[0085]

[0086] where the standard deviation σ c and the mean μ c are calculated through the following formulas respectively:

[0087]

[0088] In the formula, N c is the total number of samples, x c,i is the i-th sample. When CV c exceeds θ c , mark the corresponding feature as a high-risk area.

[0089] S203. Dynamic evaluation of feature importance: Use the information gain ratio G c (f) to quantify the correlation between the frequency of personnel changes P a and the incidence rate of risk events E a , and dynamically adjust the interpolation weight λ c according to the information gain ratio G c to preferentially generate synthetic samples of high-risk features.

[0090] The adjustment formula of the interpolation weight λ c is:

[0091]

[0092] In the formula, ∑ f∈F G c(f) is the sum of the information gain ratios of all features, F is the set of all features,

[0093] G c (f) is the information gain ratio of feature f.

[0094] G c (f) is the information gain ratio of feature f. Using the information gain ratio can help dynamically adjust the interpolation weights, preferentially generate samples related to high-risk features, and by adjusting the sample generation strategy, enable the model to pay more attention to these key features, thereby improving the prediction accuracy. For example, when the incidence rate of risk events is high, automatically adjust the weights for generating synthetic samples to strengthen the model's prediction ability for potential risks, G c (f) is defined as:

[0095]

[0096] In the formula, H(Y) is the entropy of the project progress data, H(Y|f) is the conditional entropy given feature f, and H(f) is the entropy of feature f.

[0097] The information gain ratio G c When calculating (f), equal-width binning discretization is used for continuous features, and the number of bins k = 10 is set. For example, for the personnel change frequency P a and the incidence rate of risk events E a , when μ c <5, Poisson distribution fitting discretization is used.

[0098] S204. Synthetic sample generation and adjustment: According to the real-time feature distribution and the interpolation weight λ c , combined with the dynamic adjustment coefficient γ c , generate new samples through the interpolation formula where γ c is automatically adjusted according to the chaos degree of the project progress data to control the generation intensity of high-risk data.

[0099] The chaos degree of the project progress data (i.e., the fluctuations in project delays, quality scores, etc.) will affect the attenuation of the adjustment coefficient. When there are large fluctuations in the project progress, the adjustment coefficient will increase appropriately to generate more data samples that conform to high-risk features. Through synthetic sample generation and adjustment, the prediction ability for problems such as schedule delays and resource consumption can be effectively improved. For example, when the number of contract changes increases or the time delay is more serious, the system will expand the coverage of relevant data through the sample generation mechanism so that the model can better handle similar situations in the future.

[0100] The dynamic adjustment coefficient γ c is adjusted exponentially according to the entropy of the project progress data, and the calculation formula is:

[0101] γ c = γ 0 · e -βH(Y)

[0102] In the formula, γ 0 is the initial adjustment coefficient, γ 0 is set to 2, e is the natural constant, β is the attenuation factor, and the value range of the attenuation factor β is 0.1 ≤ β ≤ 0.5, and β is set to 0.2.

[0103] The formula for interpolating to generate a new sample is:

[0104] x c new = x c + γ c · λ c · (x' c - x c )

[0105] In the formula, x' c and x c are two randomly selected minority class samples, and x c new is the synthesized new sample.

[0106] The minority class sample selection mechanism is: among the k nearest neighbors (k = 5) of the same class, select adjacent samples according to the probability , where is the Euclidean distance between x' c and x c .

[0107] S205. Repeat the above steps until the preset number of synthesized samples n c is reached, which is used to train a risk prediction model to optimize resource allocation and schedule delay prediction.

[0108] In this embodiment, the stability of the adaptive SMOTE algorithm in the risk prediction task is verified under different data imbalance ratios. By comparing two key indicators, the F1 score and ROC-AUC, the processing ability of the algorithm for the class imbalance problem is evaluated. When the minority class ratio is less than 30%, the performance of all algorithms shows a significant decline, but the decline of the adaptive SMOTE is the smallest (F1 drops by about 15% and AUC drops by about 10%). In the high imbalance scenario (ratio < 20%), the F1 score of the adaptive SMOTE is 8 - 12 percentage points higher than that of the sub-optimal algorithm. The dynamic feature adjustment mechanism effectively maintains the correlation between features, making the AUC index always higher than other methods. The experimental results show that the adaptive SMOTE significantly improves the sensitivity of the model to high-risk features through the dynamic evaluation of feature importance and the adjustment coefficient mechanism, and can still maintain an AUC value of more than 85% in the extremely imbalanced scenario, verifying the effectiveness of the algorithm design, as Figure 1 shown.

[0109] S3. Risk prediction of the whole-process consulting project

[0110] Use a deep neural network classifier for risk prediction of the whole-process consulting project. The deep neural network classifier includes a temporal feature separation layer, which assigns independent neuron channels to the project progress ratio R a and the time delay T a and optimizes the weight update process through enhanced global search.

[0111] If the progress ratio of the project deviates from the expectation, it may lead to time delay. Therefore, by independently processing these characteristics, potential progress delay risks can be identified more accurately.

[0112] The temporal feature separation layer consists of two independent long short-term memory network modules, which respectively process the temporal features of the project progress ratio R a and the time delay T a . Each long short-term memory network module contains 32 hidden units, and the output is mapped to an independent neuron channel through a fully connected layer. Other temporal features (such as the personnel change frequency P a , the number of contract changes M a , etc.) are processed using a shared long short-term memory network module, and the number of its hidden units is 16.

[0113] The training process of the deep neural network classifier includes the following steps:

[0114] S301. Initialize the parameters of the deep neural network classifier: Construct a 6-layer fully connected deep neural network classifier. According to the preset number of features and the characteristics of engineering data, for the project progress ratio R a with temporal correlation and the time delay T aAllocate independent hidden layer neurons and randomly generate the initial weight matrix and the initial bias vector to match the feature dimension.

[0115] The 6-layer fully connected deep neural network is: a parallel cascade layer of the time series feature separation layer and the input layer (8 nodes), hidden layer 1 (64 nodes, ReLU), hidden layer 2 (32 nodes, ReLU), hidden layer 3 (16 nodes, tanh), hidden layer 4 (8 nodes), and output layer (3 nodes, Softmax). The dropout rate is set to 0.2.

[0116] The initial weight matrix has a dimension of F fs ×P fs where F fs = 10 is the preset number of features and P fs is the number of hidden layer neurons.

[0117] The initial bias vector has the same dimension as P fs which is consistent.

[0118] S302, Enhanced global search optimization: By calculating the Euclidean distance between the weight matrix and the best weight dynamically adjust the reflection intensity γ p and combine the hyperbolic tangent activation function to constrain the magnitude of the weight update ΔW p By dynamically adjusting the weight update, optimize the model's ability to capture the inflection point of schedule delay.

[0119] The dynamic adjustment of the reflection intensity can refine the magnitude of the weight update, making the model more sensitive when facing key features such as schedule delay. For example, when the project faces a large schedule deviation, the model can automatically enhance its attention to this feature, thereby improving the accuracy of risk prediction. The calculation formula for the reflection intensity γ p is:

[0120]

[0121] In the formula, κ p is the reflection intensity sensitivity parameter, exp is the natural exponential function, is the Euclidean distance between the weight at the t-th iteration and the best weight, is the weight at the current t-th iteration, is the best weight.

[0122] The reflection intensity sensitivity parameter κ p has a value range of 0.1 ≤ κ p≤0.5, when the training accuracy fluctuates by more than 5%, it is automatically adjusted, and the adjustment method is to shrink by a factor of 0.9.

[0123] Weight update amount ΔW p The calculation formula is:

[0124]

[0125] Among them, λ p is the learning rate of the deep neural network classifier, γ p is the reflection intensity, and tanh is the hyperbolic tangent activation function.

[0126] S303, Feature Sensitivity-Driven Weight Adjustment: Based on feature sensitivity and local curvature Optimize the weight update direction through the composite loss function L p and preferentially correct the neuron connection weights corresponding to highly sensitive features.

[0127] The composite loss function combines cross-entropy and root mean square loss to ensure that the model can focus on the characteristics that have a greater impact on the prediction results during training. For example, cost deviation (such as overspending) and customer satisfaction (such as negative feedback) help to address the characteristics that have a greater impact on the project progress or results, enabling the model to classify project risks more accurately. The composite loss function L p is defined as a weighted combination of cross-entropy loss and root mean square loss, and the calculation formula is:

[0128]

[0129] In the formula, is the true label of the i p th sample, is the predicted probability of the deep neural network classifier for the i p th sample, and N p is the number of batch samples input to the deep neural network classifier.

[0130] By calculating the sensitivity of each feature to the prediction result, it is possible to determine which features are most critical for project risk prediction. For example, if the sensitivity of a certain feature (such as project progress ratio or cost deviation) is particularly high, it indicates that this feature has a greater impact on the model result. At this time, the model can give priority to these features and optimize them more during training to improve the prediction accuracy. Feature sensitivity The calculation formula is:

[0131]

[0132] In the formula, is the ic The input value of a feature, and ya is the predicted probability of the sample by the deep neural network classifier.

[0133] S304, iteratively train until the stop condition is met: repeat the above steps to converge the model to the optimal parameters for risk prediction, and output the early warning of project cost overrun and the probability of schedule delay.

[0134] The inference process of the deep neural network classifier is the process of risk prediction for the whole-process consulting project. The trained deep neural network model will calculate and output a risk classification result. For example, for each input project data, the model will output 3 categories, including low risk, medium risk, and high risk.

[0135] As Figure 2 shown, the performance differences between this technology and the conventional long short-term memory network, random forest, and support vector machine in project risk prediction are compared to verify the effectiveness of the structural design for independently processing time series features and the dynamic weight optimization mechanism. The experimental results show that this technology is significantly superior to the comparison methods in terms of prediction accuracy and result stability. The network structure that separates and processes key time series features can capture the risk evolution law more accurately. However, the traditional models lack the ability to distinguish the importance of features and a dynamic optimization mechanism, resulting in insufficient recognition ability for complex risk patterns, verifying the key role of independent neuron channels and enhanced global search optimization in risk prediction effect.

[0136] As Figure 3 shown, by changing the scale of training data, the performance change trends of different models in the schedule delay prediction task are compared. The focus of the experiment is to evaluate the generalization ability of the model in the scenario of limited data, especially to verify the improvement effect of the enhanced global search optimization on feature sensitivity. The experimental results show that this technology can still maintain high prediction performance in the case of small samples, and shows a faster learning convergence speed as the data volume increases, indicating the enhanced learning ability of the dynamic reflection intensity mechanism for key features. In contrast, the traditional long short-term memory network has a single feature processing method and is difficult to effectively extract the risk association patterns in time series features when the data is insufficient. The machine learning models based on static rules have a significant lag in performance improvement due to the lack of a feature adaptation mechanism, reflecting the advantage of this technology in optimizing the risk inflection point capture ability through dynamic weight adjustment.

[0137] The above is only a preferred embodiment of the present invention and is not used to limit the protection scope of the present invention.

Claims

1. The whole-process consulting project risk prediction method based on artificial intelligence is characterized by: The following steps are involved: S1, project process data collection and annotation In the process of project management, we collect multi-source heterogeneous data such as progress reports, quality inspection reports, and resource usage, and manually annotate the collected project process data. The annotated categories include low risk, medium risk, and high risk. S2, project process data expansion The SMOTE algorithm based on adaptive synthetic sampling is used to generate samples and expand project process data; S3. Risk prediction of the whole process consulting project The deep neural network classifier is used to predict the risk of the whole consulting project. The deep neural network classifier contains a time series feature separation layer, which is the project progress ratio R a With time delay T a Assign independent neuron channels and optimize the weight update process through enhanced global search.

2. The method for predicting risk of a full-process consulting project based on artificial intelligence according to claim 1 is characterized in that: The multi-source heterogeneous data in S1 include the project progress ratio R a , quality score Q a , Resource consumption rate C a , Personnel turnover frequency P a 、Cost Deviation B a , time delay T a 、Number of contract changes M a , Risk event rate E a , Customer Satisfaction a and the number of legal disputes L a .

3. The method for predicting risk of a full-process consulting project based on artificial intelligence according to claim 1 is characterized in that: The process of generating samples by the SMOTE algorithm based on adaptive synthetic sampling in S2 includes the following steps: S201. Initialize dynamic parameter gating mechanism: dimension D and preset feature number F based on data project management fs , randomly generate the initial weight matrix and the initial bias vector And the feature variability threshold θ c As a constraint for subsequent sample generation; The dimension D is determined by the number of features of the original data. fs =10, set D to 0.5 times the number of features; Initial weight matrix Initialized using Xavier normal distribution, its standard deviation is S202, High Dynamic Feature Anomaly Detection: Targeting Resource Consumption Rate C a and cost deviation B a , through the characteristic coefficient of variation CV c Identify the critical points of abnormal resource allocation in the data project management process, where the characteristic coefficient of variation CV c The calculation is based on the standard deviation σ of the eigenvalues c With mean μ c The ratio of S203, dynamic evaluation of feature importance: using information gain ratio G c (f) Quantify the frequency of personnel changes P a The risk event rate E a The correlation, and according to the information gain ratio G c (f) Dynamically adjust the interpolation weight λ c , prioritize the generation of synthetic samples with high-risk features; S204, synthetic sample generation and adjustment: according to real-time feature distribution and interpolation weight λ c , combined with the dynamic adjustment coefficient γ c , generate new samples through interpolation formula where γ c Automatically adjust according to the chaos of project progress data to control the generation intensity of high-risk data; S205, repeat the above steps until the preset number of synthetic samples n is reached c , used to train risk prediction models to optimize resource allocation and schedule delay prediction.

4. The method for predicting risk of a full-process consulting project based on artificial intelligence according to claim 3 is characterized in that: The initial weight matrix in S201 The dimension is D×F fs , where F fs There are 10 preset features; feature Variability threshold θ c By calculating the original data set x c The variance of c To include N c The matrix of samples.

5. The method for predicting risk of a full-process consulting project based on artificial intelligence according to claim 3 is characterized in that: The characteristic coefficient of variation CV in S202 c The calculation formula is: The standard deviation σ c and mean μ c Calculated by the following formulas: Where N c is the total number of samples, x c,i is the i-th sample, when CV c Exceeding θ c , the corresponding feature is marked as a high-risk area.

6. The method for predicting risk of a full-process consulting project based on artificial intelligence according to claim 3 is characterized in that: The interpolation weight λ in S203 c The adjustment formula is: In the formula, ∑ f∈F G c (f) is the sum of the information gain ratios of all features, F is the set of all features, G c (f) is the information gain ratio of feature f; G c (f) is defined as: Where H(Y) is the entropy of the project schedule data, H(Y|f) is the conditional entropy given feature f, and H(f) is the entropy of feature f.

7. The method for predicting risk of a full-process consulting project based on artificial intelligence according to claim 3 is characterized in that: The dynamic adjustment coefficient γ in S204 c According to the entropy of the project progress data, exponential decay adjustment is performed, and the calculation formula is: c c =γ0·e -βH(Y) In the formula, γ0 is the initial adjustment coefficient, β is the attenuation factor, e is the natural constant, and the value range of the attenuation factor β is 0.1≤β≤0.5; The formula for interpolating to generate new samples is: x c new =x c +g c ·l c ·(x' c -x c ) In the formula, x' c and x c are two randomly selected minority class samples, x c new is a new synthetic sample.

8. The method for predicting risk of a full-process consulting project based on artificial intelligence according to claim 1 is characterized in that: The training process of the deep neural network classifier in S3 includes the following steps: S301, initialize the parameters of the deep neural network classifier: construct a 6-layer fully connected deep neural network classifier, and calculate the project progress ratio R with time series association according to the preset feature number and engineering data characteristics. a With time delay T a Assign independent hidden layer neurons and randomly generate initial weight matrices and the initial bias vector To match the feature dimensions; The 6-layer fully connected deep neural network is: parallel cascade layer of temporal feature separation layer and input layer, hidden layer 1, hidden layer 2, hidden layer 3, hidden layer 4, output layer, and the drop rate is set to 0.2; S302, Enhanced global search optimization: by calculating the weight matrix With optimal weight Euclidean distance, dynamically adjust the reflection intensity γ p , and combined with the hyperbolic tangent activation function to constrain the weight update amount ΔW p By dynamically adjusting the weight update amount, the model's ability to capture the inflection point of schedule delay is optimized; S303, feature sensitivity driven weight adjustment: based on feature sensitivity and local curvature Through the composite loss function L p Optimize the weight update direction and give priority to correcting the neuron connection weights corresponding to highly sensitive features; S304, iterative training until the stopping condition is met: repeat the above steps to make the model converge to the optimal parameters for risk prediction, and output the project cost overrun warning and schedule delay probability.

9. The method for predicting risk of a full-process consulting project based on artificial intelligence according to claim 8 is characterized in that: The initial weight matrix in S301 The dimension is F fs ×P fs , where F fs =10 is the preset feature number, P fs is the number of neurons in the hidden layer; Initial bias vector The dimension and P fs Consistency; The reflection intensity γ in S302 p The calculation formula is: In the formula, κ p is the reflection intensity sensitivity parameter, exp is the natural exponential function, is the Euclidean distance between the weight of the tth iteration and the optimal weight, is the weight of the current t-th iteration, is the optimal weight; Weight update amount ΔW p The calculation formula is: Among them, λ p is the learning rate of the deep neural network classifier, γ p is the reflection intensity, and tanh is the hyperbolic tangent activation function.

10. The method for predicting risk of a full-process consulting project based on artificial intelligence according to claim 8 is characterized in that: The composite loss function L in S303 p It is defined as a weighted combination of cross entropy loss and root mean square loss, and is calculated as: In the formula, For the i p The true labels of samples, is the deep neural network classifier for the i-th p The predicted probability of samples, N p is the number of batch samples input to the deep neural network classifier; Feature Sensitivity The calculation formula is: In the formula, For the i c The input value of the feature, ya is the predicted probability of the sample by the deep neural network classifier.

Images