A multi-dimensional information fusion evaluation method for intelligent product quality and safety risk

Abstract

The application discloses a kind of multi-dimensional information fusion evaluation methods of intelligent product quality safety risk, including computing raw material production data aggregation and deviation, carry out abnormal identification to obtain abnormal result, calculate potential hardware risk coefficient, calculate the first degree of influence and correct potential hardware risk coefficient to obtain hardware risk coefficient, carry out, according to user feedback matching abnormal result and carry out feature enhancement to obtain risk characteristics, carry out data quality evaluation and cross enhancement to obtain software integrated risk coefficient, construct risk correlation propagation network and intelligent product quality safety risk evaluation model, input the multi-dimensional information of the intelligent product to be evaluated into the intelligent product quality safety risk evaluation model to obtain intelligent product quality safety risk evaluation result.The method can not only improve the comprehensiveness, dynamics and accuracy of intelligent product risk evaluation, but also has important significance for improving the reliability of intelligent product and the level of industry risk control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of quality safety risk evaluation, and particularly relates to a multi-dimensional information fusion evaluation method for quality safety risk of intelligent products. BACKGROUND

[0002] With the rapid development of technologies such as the Internet of Things and artificial intelligence, intelligent products have been widely applied to key fields such as industrial control, smart home, medical health and transportation, and the quality safety thereof is directly related to user safety, data privacy and system reliability. Therefore, it is of great significance to build a quality safety risk evaluation method that can fuse multi-dimensional information and realize dynamic risk propagation analysis, so as to improve the reliability of intelligent products and the level of industry risk control.

[0003] At present, the quality risk evaluation of intelligent products mostly adopts methods based on statistical process control (SPC) or failure mode and effects analysis (FMEA). Although these methods are effective in specific scenarios, they still have obvious limitations. Firstly, most of the methods are only aimed at a single data source, and lack multi-dimensional collaborative analysis of raw material production, hardware, software, environment and user feedback. Secondly, traditional anomaly detection methods (such as threshold comparison and clustering analysis) cannot effectively identify the correlation and propagation effect between multi-source anomalies. Thirdly, the existing methods usually ignore the dynamic transmission mechanism of risks among system components, resulting in lagging and one-sided risk assessment results. By introducing information fusion technology and dynamic network model, the above defects can be overcome. Therefore, the present application provides a multi-dimensional information fusion evaluation method for quality safety risk of intelligent products, which fuses multi-dimensional information, constructs a multi-level risk feature extraction and anomaly recognition model, introduces a risk correlation propagation network to dynamically depict the transmission path of risks among multi-dimensional components, and finally realizes accurate prediction of risk level and probability through an attention mechanism and a bidirectional long short-term memory network, so as to significantly improve the comprehensiveness, dynamics and accuracy of risk evaluation and provide reliable technical support for quality management of intelligent products throughout their life cycle. SUMMARY

[0004] The present application aims to provide a multi-dimensional information fusion evaluation method for quality safety risk of intelligent products.

[0005] To achieve the above-mentioned purpose, the present application is implemented according to the following technical solutions:

[0006] The present application comprises the following steps:

[0007] Obtain raw material production data, hardware data, software integration data, environmental data and user feedback of historical intelligent products as multi-dimensional information of intelligent products and perform preprocessing;

[0008] Calculate the clustering degree of the raw material production data, obtain a first abnormal result by identifying anomalies based on the deviation degree of the raw material production data, and combine the clustering degree, the deviation degree, and the first abnormal result to form a potential hardware risk feature and calculate the potential hardware risk coefficient.

[0009] The hardware data is subjected to multi-sphere screening to obtain a second abnormal result. The influence of the first abnormal result on the second abnormal result is calculated to obtain a first influence degree. The potential hardware risk coefficient is corrected according to the first influence degree to obtain a hardware risk coefficient. The user feedback is matched with the second abnormal result to obtain hardware risk characteristics.

[0010] The software integration data is compared with the corresponding threshold to obtain the third anomaly result. The third anomaly result is then subjected to data quality evaluation and cross-enhancement to obtain the software integration risk coefficient and the third anomaly optimization result. The user feedback is then matched with the third anomaly optimization result to obtain the software integration risk characteristics.

[0011] The environmental data is compared with the environmental threshold to obtain a fourth abnormal result. The user feedback is matched with the fourth abnormal result to obtain environmental risk characteristics. The environmental risk coefficient is obtained based on the hardware risk characteristics and the software integration risk characteristics.

[0012] A risk association propagation network is constructed, and an intelligent product quality and safety risk assessment model is built based on the risk association propagation network. Multidimensional information of the intelligent product to be evaluated is input into the intelligent product quality and safety risk assessment model to obtain the intelligent product quality and safety risk assessment result.

[0013] Furthermore, the method for calculating the potential hardware risk coefficient includes:

[0014] The raw material production data includes raw material performance indicators, raw material prices, production machine parameters, and employee operating parameters;

[0015] Calculate the clustering degree of each raw material production data, and calculate the deviation of the raw material production data based on the historical average of raw material production data. Then, filter out abnormal raw material production data and their corresponding abnormal categories as the first abnormal results based on the deviation threshold of each raw material production data. The expression for the clustering degree of the raw material production data is:

[0016] ;

[0017] ;

[0018] in for Raw material production data The degree of aggregation, For statistical weighting, As confidence weights, cluster weight, as a distribution weight, , as the standard deviation and mean of the data set, , as the maximum and minimum values of the data set, as a smoothing coefficient, as the number of confidence levels, as the confidence boundary under the th confidence level, as the traditional clustering degree, as the number of subsets, as the KL divergence, as the distribution of the th subset, as the overall data distribution, as the Davies-Bouldin index, as the silhouette coefficient, as the Gini coefficient.

[0019] According to the standard value of each raw material production data, the deviation degree of the corresponding raw material production data is calculated, and the raw material production data whose deviation degree is greater than the corresponding deviation threshold is marked as an abnormal value, and the abnormal type is marked according to the raw material production data category, and the first abnormal result is determined according to the abnormal value and the abnormal type of the raw material production data; The first abnormal result includes abnormal category and corresponding abnormal data quantity, abnormal deviation cumulative value;

[0020] The aggregation degree, deviation degree and first abnormal result of the raw material production data are combined to form a potential hardware risk feature, and the raw material production data risk coefficient is calculated according to the hardware risk feature, and the mean value of the raw material production data risk coefficient is taken as the potential hardware risk coefficient, and the expression is:

[0021] ;

[0022] Wherein is the risk coefficient of the raw material production data , , is the basic risk weight, is the deviation degree of the data , is the abnormal type weight, is the sensitive index of abnormal value, is the number of abnormal data, is the th abnormal data, is the abnormal threshold.

[0023] Further, the method for obtaining the hardware risk coefficient and the hardware risk feature comprises:

[0024] The hardware data is subjected to multi-sphere screening to obtain abnormal data and corresponding abnormal categories, and a second abnormal result is determined according to the abnormal value and the abnormal type of the hardware data; the hardware data comprises a component failure rate and sensor accuracy; and the second abnormal result comprises an abnormal category and corresponding abnormal data quantity and abnormal deviation cumulative value;

[0025] A first influence degree is obtained by calculating the influence degree of the first abnormal result on the second abnormal result, and a hardware risk coefficient is obtained by correcting a potential hardware risk coefficient according to the first influence degree, and the expression is as follows:

[0026] ;

[0027] ;

[0028] ;

[0029] is the first influence degree, indicating the influence degree of the first abnormal result on the second abnormal result, is the first abnormal result category index, is the second abnormal result category index, is a category abnormal result occurs with a category abnormal result , is the average abnormal deviation when occurs together, is a time sensitivity coefficient, is a time decay factor, indicating the time decay characteristics associated with , is the mutual information associated with , is the entropy of a category abnormal result, is the entropy of a category abnormal result, is the time interval of the th common occurrence from the current time, is the abnormal deviation size at the th common occurrence, is the average abnormal deviation, is The class correction raw material production data risk coefficient is taken as the hardware risk coefficient, and the average of the correction raw material production data risk coefficients is taken as the hardware risk coefficient, The influence degree adjustment coefficient is obtained.

[0030] The user feedback use abnormality is matched with the second abnormality result. If there is a matching relationship, the corresponding abnormal data quantity and abnormal deviation cumulative value are enhanced using a matching weight, and the enhanced second abnormality result is taken as the hardware risk feature. If there is no matching relationship, the original second abnormality result is taken as the hardware risk feature. The matching weight is determined by domain knowledge.

[0031] Further, the method for obtaining the software integration risk coefficient and the software integration risk feature comprises:

[0032] The software integration data comprises one-class data and two-class data. The one-class data comprises module coupling degree, interface complexity, and test coverage. The two-class data comprises code vulnerability, code repair, and response delay.

[0033] The software integration data is compared with a corresponding threshold to obtain a third abnormality result. The third abnormality result comprises an abnormality category and corresponding abnormal data quantity and abnormal deviation cumulative value.

[0034] The abnormality result of the one-class data is taken as one-class third abnormality result, and the abnormality result of the two-class data is taken as two-class third abnormality result. The software integration data is evaluated according to the one-class third abnormality result and the two-class third abnormality result to obtain a first quality coefficient and a second quality coefficient. The first quality coefficient and the second quality coefficient are weighted to obtain the software integration risk coefficient.

[0035] The one-class third abnormality result is cross-enhanced according to the second quality coefficient to obtain one-class third abnormality optimization result. The two-class third abnormality result is cross-enhanced according to the first quality coefficient to obtain two-class third abnormality optimization result. The one-class third abnormality optimization result and the two-class third abnormality optimization result are adjusted and spliced to obtain third abnormality optimization result.

[0036] The user feedback use abnormality is matched with the third abnormality optimization result. If there is a matching relationship, the corresponding abnormal data quantity and abnormal deviation cumulative value are enhanced using a matching weight, and the enhanced third abnormality optimization result is taken as the software integration risk feature. If there is no matching relationship, the original third abnormality optimization result is taken as the software integration risk feature.

[0037] Further, the method for obtaining the environment risk coefficient and the environment risk feature comprises:

[0038] The environmental data is compared with a corresponding threshold to obtain a fourth abnormal result; the environmental data includes an operating environment and an Internet of Things environment; and the fourth abnormal result includes an abnormal category and corresponding abnormal data quantity and abnormal deviation cumulative value;

[0039] The user feedback is matched with the abnormal category of the fourth abnormal result, and if there is a matching relationship, the corresponding abnormal data quantity and abnormal deviation cumulative value are enhanced using a matching weight, and the enhanced fourth abnormal result is taken as an environmental risk feature; if there is no matching relationship, the original fourth abnormal result is taken as an environmental risk feature;

[0040] The influence degree of the environmental risk feature on the hardware risk feature is calculated to obtain a first environmental correlation degree, the influence degree of the environmental risk feature on the software integration risk feature is calculated to obtain a second environmental correlation degree, and a weighted value of the first environmental correlation degree and the second environmental correlation degree is taken as an environmental risk coefficient.

[0041] Further, the method for constructing the risk correlation propagation network comprises:

[0042] A raw material production risk layer, a hardware risk layer, a software integration risk layer and an environmental risk layer are set, wherein the raw material production risk layer only serves as a precursor layer of the hardware risk layer, data indexes of each layer are taken as network nodes, edge weights between nodes in the same layer are determined according to abnormal result entropy values of the nodes, and edge weights between nodes in different layers are determined according to node abnormal result mutual information, risk values of each layer and influence degrees of each layer, and the expression is:

[0043] ;

[0044] ;

[0045] ;

[0046] ;

[0047] wherein is an edge weight of a raw material production risk layer node and a hardware risk layer node , is an edge weight of a hardware risk layer node and a software integration risk layer node , is an edge weight of a hardware risk layer node and an environmental risk layer node , is an edge weight of a software integration risk layer node and an environmental risk layer node , is an edge weight of a raw material production risk layer node The risk factor, This is a set of nodes in the raw material production risk layer. For hardware risk factor, To account for the risk factor of software integration, For environmental risk coefficient, This refers to the mutual information between the two nodes corresponding to abnormal results. As a correlation moderating factor, As the first level of influence, The first environmental correlation, This represents the second environmental correlation degree;

[0048] Set the iterative update conditions for the risk propagation network, and update the risk coefficient of each layer according to the edge weights. The expression is:

[0049] ;

[0050] in for layer The iterative risk coefficient update includes a hardware risk layer, a software integration risk layer, and an environmental risk layer. This is the risk attenuation coefficient. For risk transmission coefficient, for The set of nodes in a layer For nodes The set of predecessor nodes, Predecessor node The risk coefficient of the layer in question For nodes and nodes edge weights, for Iterating predecessor node Corresponding abnormal data Risk information entropy, This corresponds to the number of abnormal data.

[0051] Furthermore, the method for obtaining the quality and safety risk assessment results of smart products includes:

[0052] The historical raw material production risk coefficient, hardware risk coefficient, software integration risk coefficient, environmental risk coefficient, potential hardware risk characteristics, hardware risk characteristics, software integration risk characteristics, environmental risk characteristics, and corresponding safety risks are used as the quality and safety risk set. The random forest algorithm is used to divide the quality and safety risk set into a training set and a test set in a 6:4 ratio. The training set is used to train the intelligent product quality and safety risk assessment model, and the test set is used to evaluate the performance of the intelligent product quality and safety risk assessment model.

[0053] The intelligent product quality safety risk evaluation model comprises a feature processing layer, a risk propagation layer, a prediction layer and an output layer; the feature processing layer respectively performs one-hot encoding on category type features, performs periodic encoding on time series features, splices and outputs the processed features to the risk propagation layer and the prediction layer; the risk propagation layer is embedded with a risk correlation propagation network, simulates the propagation path of risks among components of the system, dynamically updates the risk coefficients of each layer and outputs to the prediction layer; the prediction layer adopts a bidirectional long short-term memory network based on an attention mechanism, learns the complex nonlinear relationship between risk features and risk coefficients, and performs risk level classification and probability prediction; the output layer extracts risk propagation factors from the risk propagation layer according to the prediction result, and outputs the intelligent product quality safety risk evaluation result in combination with the risk level classification result and the probability prediction result;

[0054] The intelligent product quality safety risk evaluation model adopts a hybrid loss function to evaluate the difference between the safety risk prediction result and the real safety risk; the hybrid loss function comprises a classification loss, a regression loss and a risk consistency loss;

[0055] The multi-dimensional information of the intelligent product to be evaluated is input into the intelligent product quality safety risk evaluation model to obtain the intelligent product quality safety risk evaluation result.

[0056] The beneficial effects of the present application are:

[0057] The present application is a multi-dimensional information fusion evaluation method for intelligent product quality safety risks, which has the following technical effects compared with the prior art:

[0058] The present application can improve the data preprocessing capability and enhance the model adaptability in intelligent product quality safety risk evaluation through the steps of anomaly identification, risk matching, calculation of data aggregation degree and correlation degree, data quality evaluation and cross enhancement, and model construction, thereby improving the efficiency and precision of intelligent product quality safety risk, optimizing the intelligent product quality safety risk evaluation technology, processing multi-source heterogeneous data, capturing the correlation between different risk factors, realizing real-time dynamic risk assessment, and significantly improving the comprehensiveness, dynamics and accuracy of risk evaluation, providing reliable technical support for intelligent product whole life cycle quality management. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The present application is a multi-dimensional information fusion evaluation method for intelligent product quality safety risks, which has the following technical effects compared with the prior art: DETAILED DESCRIPTION

[0060] The present application is a multi-dimensional information fusion evaluation method for intelligent product quality safety risks, which has the following technical effects compared with the prior art:

[0061] The application discloses a multi-dimensional information fusion evaluation method for intelligent product quality safety risks.

[0062] As shown in the figure, in the embodiment, the following steps are included: Figure 1

[0063] The raw material production data, hardware data, software integration data, environment data and user feedback of the historical intelligent product are acquired as the multi-dimensional information of the intelligent product and are preprocessed;

[0064] The raw material production data aggregation degree is calculated, the first abnormal result is obtained according to the raw material production data deviation degree and abnormality recognition, the aggregation degree, the deviation degree and the first abnormal result are combined to form a potential hardware risk feature, and a potential hardware risk coefficient is calculated;

[0065] The hardware data is subjected to multi-sphere screening to obtain a second abnormal result, the first abnormal result is subjected to influence degree calculation on the second abnormal result to obtain a first influence degree, the potential hardware risk coefficient is corrected according to the first influence degree to obtain a hardware risk coefficient, and the hardware risk feature is obtained by matching the user feedback with the second abnormal result;

[0066] The software integration data is compared with a corresponding threshold value to obtain a third abnormal result, the third abnormal result is subjected to data quality evaluation and cross enhancement to obtain a software integration risk coefficient and a third abnormal optimization result, and the software integration risk feature is obtained by matching the user feedback with the third abnormal optimization result;

[0067] The environment data is compared with an environment threshold value to obtain a fourth abnormal result, the environment risk feature is obtained by matching the user feedback with the fourth abnormal result, and the environment risk coefficient is obtained according to the hardware risk feature and the software integration risk feature;

[0068] A risk correlation propagation network is constructed, an intelligent product quality safety risk evaluation model is constructed according to the risk correlation propagation network, the multi-dimensional information of the intelligent product to be evaluated is input into the intelligent product quality safety risk evaluation model, and an intelligent product quality safety risk evaluation result is obtained.

[0069] In the embodiment, the method for calculating the potential hardware risk coefficient comprises:

[0070] The raw material production data comprises raw material performance indexes, raw material prices, production machine parameters and employee operation parameters;

[0071] The aggregation degrees of the raw material production data are calculated, the deviation degrees of the raw material production data are calculated according to the mean values of the historical raw material production data, the raw material production abnormal data and corresponding abnormal categories are screened as the first abnormal result according to the deviation degree threshold values of the raw material production data, and the aggregation degree expression of the raw material production data is:​

[0072] ;

[0073] ;

[0074] wherein is raw material production data aggregation, is a statistical weight, is a confidence weight, a clustering weight, is a distribution weight, , is standard deviation and mean of the data set, , is maximum and minimum of the data set, is a smoothing coefficient, is a number of confidence levels, is a confidence boundary at the th confidence level, is a traditional clustering degree, is a number of subsets, is a KL divergence, is a distribution of the th subset, is an overall data distribution, is a Davies-Bouldin index, is a silhouette coefficient, is a Gini coefficient;

[0075] According to the standard value of each raw material production data, the deviation degree of the corresponding raw material production data is calculated, the raw material production data whose deviation degree is greater than the corresponding deviation threshold is marked as an abnormal value, and the abnormal type is marked according to the raw material production data category, and the first abnormal result is determined according to the abnormal value and the abnormal type of the raw material production data; The first abnormal result includes abnormal category and corresponding abnormal data quantity, abnormal deviation cumulative value;

[0076] The aggregation, deviation and first abnormal result of the raw material production data are combined to form a potential hardware risk feature, and the raw material production data risk coefficient is calculated according to the hardware risk feature, and the mean value of the raw material production data risk coefficient is taken as the potential hardware risk coefficient, and the expression is:

[0077] ;

[0078] wherein is raw material production data risk coefficient, , is a basic risk weight, is a deviation degree of data , is an abnormal type weight, is a sensitive index of abnormal value, is an abnormal data quantity, is the first abnormal data, is an abnormal threshold value;

[0079] In the actual evaluation, taking the quality and safety risk evaluation of the intelligent hotel delivery robot of brand A as an example, a group of historical intelligent product multi-dimensional information is processed, the deviation degree of each raw material production data is calculated according to “deviation degree=(raw material production data-raw material production data standard value) / raw material production data standard value”, the raw material production data greater than the deviation degree threshold value is marked as an abnormal value, and the abnormal type is marked according to the data category (only including raw material performance index, raw material price, production machine parameter, and employee operation parameter four categories). The number of abnormal data of the same category and the abnormal deviation cumulative value constitute an abnormal result of a category. Four kinds of abnormal results constitute a first abnormal result, wherein the abnormal deviation cumulative value takes the cumulative value of all “(abnormal data-abnormal threshold value) / abnormal threshold value”;

[0080] Taking the statistical weight / confidence weight / clustering weight / distribution weight as 0.3 / 0.3 / 0.3 / 0.2, the basic risk weight as 0.6 / 0.4, the smoothing coefficient as 10 -6 , the confidence level number as 3 (90% / 95% / 99%), the subset number as 5, the abnormal type weight as 0.8 / 0.9 / 0.7 / 1, the sensitive index as 1.2 / 1.5 / 0.8 / 2, the four categories of raw material production data aggregation degree (0.833 / 0.717 / 0.859 / 0.683), deviation degree (0.15 / 0.25 / 0.08 / 0.35), first abnormal result (abnormal data quantity, abnormal deviation cumulative value: 3 / 2 / 1 / 4, 0.45 / 0.6 / 0.2 / 1.2), the risk coefficient of four categories of raw material production data is calculated as 0.241 / 0.442 / 0.161 / 0.633, and the mean value is obtained. The potential hardware risk coefficient is 0.369.

[0081] In this embodiment, the method for obtaining the hardware risk coefficient and the hardware risk feature comprises:

[0082] The hardware data is subjected to multi-sphere screening to obtain abnormal data and corresponding abnormal categories, and a second abnormal result is determined according to the abnormal value and the abnormal type of the hardware data; the hardware data includes component failure rate and sensor accuracy; and the second abnormal result includes abnormal category and corresponding abnormal data quantity and abnormal deviation cumulative value. ​

[0083] The first influence degree of the first abnormal result on the second abnormal result is calculated to obtain a first influence degree, and the potential hardware risk coefficient is corrected according to the first influence degree to obtain a hardware risk coefficient, and the expression is:

[0084] ;

[0085] ;

[0086] ;

[0087] is the first influence degree, indicating the influence degree of the first abnormal result on the second abnormal result, is the first abnormal result category index, is the second abnormal result category index, is the abnormal result of the first category occurs together with the abnormal result of the second category , is the average abnormal deviation when occurs together, is the time sensitivity coefficient, is the time decay factor, indicating the time decay characteristics associated with , is the mutual information associated with , is the entropy of the abnormal result of the first category, is the entropy of the abnormal result of the second category, is the time interval between the th common occurrence and the current time, is the size of the abnormal deviation when the th common occurrence occurs, is the average abnormal deviation, is the risk coefficient of the modified raw material production data of the first category, and the average of the risk coefficient of the modified raw material production data is taken as the hardware risk coefficient, is the influence degree adjustment coefficient;

[0088] The user feedback usage abnormality is matched with the second abnormal result, if there is a matching relationship, the corresponding abnormal data quantity and abnormal deviation cumulative value are enhanced by using a matching weight, and the enhanced second abnormal result is taken as the hardware risk feature, if there is no matching relationship, the original second abnormal result is taken as the hardware risk feature; the matching weight is determined by domain knowledge;

[0089] In the actual evaluation, the hardware data is subjected to multi-sphere screening to obtain a second abnormal result (abnormal data quantity, abnormal deviation cumulative value: 5 / 3, 1.8 / 0.9), the time sensitivity coefficient is taken as 0.8, the influence degree adjustment coefficient is taken as 0.15, the entropy of the first abnormal result 0.85 / 0.78 / 0.92 / 0.75 and the entropy of the second abnormal result 0.88 / 0.82 are substituted into the formula to calculate the first influence degree 0.1841;

[0090] The first influence degree is used to correct the raw material production data risk coefficient 0.241 / 0.442 / 0.161 / 0.633 to obtain 2.77 / 0.508 / 0.185 / 0.728, and the average value is taken to obtain the hardware risk coefficient 0.424;

[0091] The user feedback abnormality is matched with the abnormal category of the second abnormal result, the user feedback "sensor precision abnormality" is matched with the second abnormal result category 2 "sensor precision abnormality", the matching weight is taken as 1.2, and the enhanced second abnormal result (abnormal data quantity, abnormal deviation cumulative value: 5 / 3.6, 1.8 / 1.08) is taken as the hardware risk feature.

[0092] In the embodiment, the method for obtaining the software integration risk coefficient and the software integration risk feature comprises:

[0093] The software integration data comprises one type of data and two types of data; the one type of data comprises module coupling degree, interface complexity and test coverage; the two types of data comprise code vulnerability, code repair and response delay;

[0094] The software integration data is compared with the corresponding threshold to obtain a third abnormal result; the third abnormal result comprises an abnormal category and corresponding abnormal data quantity and abnormal deviation cumulative value;

[0095] The abnormal result of the one type of data is taken as a first type of third abnormal result, the abnormal result of the two types of data is taken as a second type of third abnormal result, the software integration data is evaluated according to the first type of third abnormal result and the second type of third abnormal result, the first quality coefficient and the second quality coefficient are obtained, and the software integration risk coefficient is obtained by weighting the first quality coefficient and the second quality coefficient;

[0096] The first type of third abnormal result is cross-enhanced according to the second quality coefficient to obtain a first type of third abnormal optimization result, the second type of third abnormal result is cross-enhanced according to the first quality coefficient to obtain a second type of third abnormal optimization result, and the third abnormal optimization result is obtained by adjusting and splicing the first type of third abnormal optimization result and the second type of third abnormal optimization result;

[0097] The user feedback is matched with the third abnormality optimization result, if there is a matching relationship, the corresponding abnormal data quantity and abnormal deviation cumulative value are enhanced by using the matching weight, and the enhanced third abnormality optimization result is taken as the software integration risk feature, if there is no matching relationship, the original third abnormality optimization result is taken as the software integration risk feature;

[0098] In actual evaluation, the software integration data (50 groups) is compared with the corresponding threshold to obtain third abnormality results (abnormal data quantity, abnormal deviation cumulative value): the first type of third abnormality result is 8 / 12 / 5, 3.2 / 5.8 / 1.5, and the second type of third abnormality result is 15 / 6 / 10, 8.2 / 2.1 / 4.5;

[0099] According to the following formula, the data evaluation is performed on each software integration data respectively:

[0100] ;

[0101] Wherein is the quality coefficient of the software integration data index, is the evaluation weight, , is the abnormal data quantity and abnormal deviation mean of the software integration data index, , is the data total and abnormal deviation allowed value of the software integration data index;

[0102] The quality coefficients of the first type of data and the second type of data are averaged to obtain the first quality coefficient 0.821 and the second quality coefficient 0.786, and the software integration risk coefficient is calculated as =1-0.5(0.821+0.786)=0.1965;

[0103] The second quality coefficient 0.786 is used to cross-enhance the first type of third abnormality result, and the cross-enhancement step is [1+(1-quality coefficient)]*abnormality result, and the first type of third abnormality optimization result (abnormal data quantity, abnormal deviation cumulative value) is obtained: 9.71 / 14.57 / 6.07, 3.88 / 7.04 / 1.82; the first quality coefficient 0.786 is used to cross-enhance the second type of third abnormality result, and the second type of third abnormality optimization result (abnormal data quantity, abnormal deviation cumulative value) is obtained: 17.69 / 7.07 / 11.79, 9.68 / 2.48 / 5.31;

[0104] The third abnormal optimization result of a class is spliced with the third abnormal optimization result of a second class to obtain a third abnormal optimization result 9.71 / 14.57 / 6.07 / 17.69 / 7.07 / 11.79, 3.88 / 7.04 / 1.82 / 9.68 / 2.48 / 5.31. The user feedback of the response delay anomaly is matched with the response delay anomaly in the third abnormal optimization result successfully, and the matching weight is 1.2. The enhanced third abnormal optimization result 9.71 / 14.57 / 6.07 / 17.69 / 7.07 / 14.15, 3.88 / 7.04 / 1.82 / 9.68 / 2.48 / 6.37 is taken as the software integration risk feature.

[0105] In the embodiment, the method for obtaining the environment risk coefficient and the environment risk feature comprises:

[0106] The environment data is compared with the corresponding threshold to obtain a fourth abnormal result. The environment data comprises an operation environment and an Internet of Things environment. The fourth abnormal result comprises an abnormal category and corresponding abnormal data quantity and abnormal deviation cumulative value.

[0107] The user feedback use abnormality is matched with the abnormal category of the fourth abnormal result. If there is a matching relationship, the corresponding abnormal data quantity and abnormal deviation cumulative value are enhanced by using a matching weight. The enhanced fourth abnormal result is taken as the environment risk feature. If there is no matching relationship, the original fourth abnormal result is taken as the environment risk feature.

[0108] The influence degree of the environment risk feature on the hardware risk feature is calculated to obtain a first environment correlation degree. The influence degree of the environment risk feature on the software integration risk feature is calculated to obtain a second environment correlation degree. The weighted value of the first environment correlation degree and the second environment correlation degree is taken to obtain the environment risk coefficient.

[0109] In actual evaluation, the environment data comprises an operation environment and an Internet of Things environment. The operation environment comprises temperature, humidity, electromagnetic, power and physical vibration. The Internet of Things environment comprises signal strength, data packet loss, network delay, connection interruption and protocol compatibility.

[0110] The environment data is compared with the corresponding threshold to obtain a fourth abnormal result (abnormal data quantity, abnormal deviation cumulative value): 6 / 4 / 3 / 8 / 5 / 12 / 9 / 7 / 15 / 5, 4.8 / 2.1 / 1.8 / 6.4 / 3.5 / 9.6 / 5.4 / 4.2 / 12 / 2.5. The user feedback abnormal data is matched with the network delay anomaly and the power stability anomaly, and the matching weight is 1.2. The enhanced fourth abnormal optimization result 6 / 4 / 3 / 9.6 / 5 / 12 / 9 / 9.6 / 15 / 5, 4.8 / 2.1 / 1.8 / 7.68 / 3.5 / 9.6 / 5.4 / 5.04 / 12 / 2.5 is taken as the environment risk feature.

[0111] The influence degree of the environmental risk feature on the hardware risk feature and the software integration risk feature is calculated respectively to obtain a first environmental correlation degree 0.78 and a second environmental correlation degree 0.82, and a weighted calculation (the weight is 0.1) obtains an environmental risk coefficient 0.16.

[0112] In the embodiment, the method for constructing the risk correlation propagation network comprises:

[0113] The raw material production risk layer, the hardware risk layer, the software integration risk layer and the environmental risk layer are set, wherein the raw material production risk layer only serves as a precursor layer of the hardware risk layer, the data indexes of each layer are taken as network nodes, the edge weight between nodes in the same layer is determined according to the abnormal result entropy value of each node, and the edge weight between nodes in different layers is determined according to the mutual information of the abnormal result of each node, the risk value of each layer and the influence degree of each layer, and the expression is:

[0114] ;

[0115] ;

[0116] ;

[0117] ;

[0118] wherein is the edge weight of the raw material production risk layer node and the hardware risk layer node , is the edge weight of the hardware risk layer node and the software integration risk layer node , is the edge weight of the hardware risk layer node and the environmental risk layer node , is the edge weight of the software integration risk layer node and the environmental risk layer node , is the risk coefficient of the raw material production risk layer node , is the raw material production risk layer node set, is the hardware risk coefficient, is the software integration risk coefficient, is the environmental risk coefficient, is the mutual information of the corresponding abnormal results of two nodes, is the correlation adjustment factor, is the first influence degree, is the first environmental correlation degree, is the second environmental correlation degree.

[0119] The risk correlation propagation network iteration update condition is set, and the risk coefficients of each layer are updated according to the edge weight, and the expression is:

[0120]

[0121] The iteration risk coefficient update includes the hardware risk layer, the software integration risk layer and the environment risk layer, is a risk attenuation coefficient, is a risk propagation coefficient, is a node set of the layer, is a predecessor node set of the node is the risk coefficient of the layer where the predecessor node is the edge weight of the node and the node is the risk information entropy of the abnormal data corresponding to the iteration predecessor node is the number of abnormal data;

[0122] In actual evaluation, the network node numbers of the raw material production risk layer, the hardware risk layer, the software integration risk layer and the environment risk layer are 4 / 2 / 6 / 10 respectively, and the edge weights of each network node are calculated according to the correlation adjustment factor 0.15, the first influence degree 0.1841, the first environment correlation degree 0.78 and the second environment correlation degree 0.82;

[0123] Taking the edge weight between nodes in the same layer in the hardware risk layer as an example, the entropy value of the abnormal result corresponding to the component failure rate and the sensor accuracy is 0.78, and the edge weight of the component failure rate node and the sensor accuracy node = entropy value * hardware risk coefficient = 0.424 * 0.78 = 0.331;

[0124] Taking the edge weight calculation between the raw material production risk layer node and the hardware risk layer node as an example, the mutual information between the node and the node is 0.6, and is calculated;

[0125] Taking the edge weight calculation between the hardware risk layer node and the software integration risk layer node as an example, the mutual information between the node and the node​​​​​​​​​​​ Mutual information is 0.7, and calculation ;

[0126] Take the edge weight calculation of the hardware risk layer node and the environment risk layer node as an example, the node and the node Mutual information is 0.6, and calculation ;

[0127] Take the edge weight calculation of the software integration risk layer node and the environment risk layer node as an example, the node and the node Mutual information is 0.5, and calculation ;

[0128] Take the hardware risk layer risk coefficient update as an example, the risk attenuation coefficient is 0.7, and the risk propagation coefficient is 0.3. All predecessor nodes of the hardware risk layer come from the raw material production layer (corresponding to the potential hardware risk coefficient is 0.369, the risk information entropy of each node is 0.5 / 0.6 / 0.4 / 0.7, the number of abnormal data is 3 / 2 / 1 / 4), the current hardware risk coefficient is 0.424, and the updated hardware risk coefficient is 0.461.

[0129] In the embodiment, the method for obtaining the intelligent product quality safety risk evaluation result comprises:

[0130] The historical raw material production risk coefficient, the hardware risk coefficient, the software integration risk coefficient, the environment risk coefficient, the potential hardware risk feature, the hardware risk feature, the software integration risk feature, the environment risk feature and the corresponding safety risk are taken as a quality safety risk set, the quality safety risk set is divided into a training set and a test set according to 6:4 by using a random forest algorithm, an intelligent product quality safety risk evaluation model is trained by using the training set, and the performance of the intelligent product quality safety risk evaluation model is evaluated by using the test set.

[0131] The intelligent product quality and safety risk evaluation model comprises a feature processing layer, a risk propagation layer, a prediction layer and an output layer; the feature processing layer respectively performs one-hot encoding on category type features, periodic encoding on time series features, splices and outputs the processed features to the risk propagation layer and the prediction layer; the risk propagation layer embeds a risk correlation propagation network, simulates the propagation path of risks among system components, dynamically updates the risk coefficients of each layer and outputs to the prediction layer; the prediction layer uses a bidirectional long short-term memory network based on an attention mechanism to learn the complex nonlinear relationship between risk features and risk coefficients, and performs risk level classification and probability prediction; the output layer extracts risk propagation factors from the risk propagation layer according to the prediction results, and outputs the intelligent product quality and safety risk evaluation results in combination with the risk level classification results and the probability prediction results;

[0132] The intelligent product quality and safety risk evaluation model uses a hybrid loss function to evaluate the difference between the safety risk prediction results and the actual safety risks; the hybrid loss function comprises a classification loss, a regression loss and a risk consistency loss;

[0133] The multi-dimensional information of the intelligent product to be evaluated is input into the intelligent product quality and safety risk evaluation model to obtain the intelligent product quality and safety risk evaluation results;

[0134] In actual evaluation, the multi-dimensional information of the intelligent product to be evaluated is processed in advance to obtain the raw material production risk coefficient, the hardware risk coefficient, the software integration risk coefficient, the environmental risk coefficient, the potential hardware risk feature, the hardware risk feature, the software integration risk feature and the environmental risk feature, which are input into the intelligent product quality and safety risk evaluation model;

[0135] In the feature processing layer, the numerical feature processing includes abnormal data quantity normalization and abnormal deviation value logarithmic transformation, and the category type feature processing includes one-hot encoding of abnormal types and periodic encoding of time features;

[0136] The risk propagation layer updates the raw material production risk coefficient, the hardware risk coefficient, the software integration risk coefficient and the environmental risk coefficient to obtain 0.255 / 0.467 / 0.152 / 0.658, 0.4756, 0.1863 and 0.1638;

[0137] The prediction layer performs risk prediction according to the updated risk coefficients and the risk features processed by the feature processing layer: hardware and environmental collaborative risk / high risk / 0.87, raw material supply chain risk / medium risk / 0.68;

[0138] The output layer locates the corresponding nodes of the risk correlation propagation network according to the prediction classification results, extracts the edge weights of each node and the predecessor node to obtain the risk propagation factors / edge weights: employee operation parameter risk / 0.32, environmental electromagnetic interference anomaly / 0.25 and hardware sensor precision risk / 0.18.

[0139] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1.A method for multi-dimensional information fusion evaluation of intelligent product quality safety risk, characterized in that, The method comprises the following steps: S1, obtaining raw material production data, hardware data, software integration data, environment data and user feedback of historical intelligent products as intelligent product multi-dimensional information and preprocessing; S2, calculating the aggregation degree of the raw material production data, obtaining a first abnormal result according to the raw material production data deviation degree and abnormal recognition, and grouping the aggregation degree, the deviation degree and the first abnormal result into potential hardware risk characteristics and calculating a potential hardware risk coefficient; S3, obtaining a second abnormal result by multi-sphere screening of the hardware data, calculating the influence degree of the first abnormal result on the second abnormal result to obtain a first influence degree, correcting the potential hardware risk coefficient according to the first influence degree to obtain a hardware risk coefficient, and matching the user feedback with the second abnormal result to obtain hardware risk characteristics; S4, comparing the software integration data with the corresponding threshold to obtain a third abnormal result, performing data quality evaluation and cross enhancement on the third abnormal result to obtain a software integration risk coefficient and a third abnormal optimization result, and matching the user feedback with the third abnormal optimization result to obtain software integration risk characteristics; S5, comparing the environment data with the environment threshold to obtain a fourth abnormal result, matching the user feedback with the fourth abnormal result to obtain environment risk characteristics, and obtaining an environment risk coefficient according to the hardware risk characteristics and the software integration risk characteristics; S6, constructing a risk correlation propagation network, constructing an intelligent product quality and safety risk evaluation model according to the risk correlation propagation network, inputting the multi-dimensional information of the intelligent product to be evaluated into the intelligent product quality and safety risk evaluation model to obtain an intelligent product quality and safety risk evaluation result. 2.The multi-dimensional information fusion evaluation method for intelligent product quality safety risk according to claim 1, characterized in that, The method for calculating the potential hardware risk coefficient comprises: The raw material production data comprises raw material performance indicators, raw material prices, production machine parameters and employee operation parameters; The aggregation degree of each raw material production data is calculated, the deviation degree of the raw material production data is calculated according to the mean value of the historical raw material production data, and the raw material production abnormal data and the corresponding abnormal category are screened according to the deviation degree threshold value of each raw material production data as the first abnormal result; the aggregation degree of the raw material production data is expressed as: ； ； wherein is raw material production data of the degree of aggregation, is a statistical weight, is a confidence weight, is a clustering weight, is a distribution weight, , is standard deviation and mean of the dataset, , is maximum and minimum of the dataset, is a smoothing coefficient, is a number of confidence levels, is a confidence boundary at the confidence level, is a traditional clustering degree, is a number of subsets, is a KL divergence, is a distribution of the subset, is an overall data distribution, is a Davies-Bouldin index, is a silhouette coefficient, is a Gini coefficient; The deviation degree of the corresponding raw material production data is calculated according to the standard value of each raw material production data, the raw material production data whose deviation degree is greater than the corresponding deviation degree threshold value is marked as an abnormal value, and the abnormal type is labeled according to the raw material production data category, the abnormal value and the abnormal type of the raw material production data are determined to obtain the first abnormal result; the first abnormal result comprises an abnormal category and the corresponding abnormal data quantity and abnormal deviation cumulative value; The aggregation degree, the deviation degree and the first abnormal result of the raw material production data are grouped into potential hardware risk characteristics, and the raw material production data risk coefficient is calculated according to the hardware risk characteristics; the mean value of the raw material production data risk coefficient is taken as the potential hardware risk coefficient, and the expression is as follows: ； in for Raw material production data The risk factor, , Basic risk weights, For data The degree of deviation Weights for abnormal types, A sensitivity index for outlier values. The number of abnormal data. For the first One abnormal data point. This is the abnormal threshold. 3.The multi-dimensional information fusion evaluation method for intelligent product quality safety risk according to claim 1, characterized in that, The method for obtaining the hardware risk coefficient and the hardware risk characteristics comprises: The hardware data is subjected to multi-sphere screening to obtain abnormal data and corresponding abnormal categories, and a second abnormal result is determined according to the abnormal values and types of the hardware data; the hardware data includes component failure rate and sensor accuracy; and the second abnormal result includes abnormal categories and corresponding abnormal data quantity and abnormal deviation cumulative value; A first influence degree is calculated according to the influence of the first abnormal result on the second abnormal result, and a hardware risk coefficient is obtained by correcting the potential hardware risk coefficient according to the first influence degree, and the expression is: ； ； ； wherein is a first influence degree, representing a degree of influence of the first abnormal result on the second abnormal result, is a first abnormal result category index, is a second abnormal result category index, is a first abnormal result category index, is a first abnormal result category index, is a first abnormal result category index, is a first abnormal result category index, is a first abnormal result category index, is a first abnormal result category index, is a first abnormal result category index, is a first abnormal result category index, is a time sensitivity coefficient, is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing is a time decay factor, representing The user feedback use abnormality is matched with the second abnormal result, if there is a matching relationship, the corresponding abnormal data quantity and abnormal deviation cumulative value are enhanced by using a matching weight, the enhanced second abnormal result is taken as a hardware risk feature, and if there is no matching relationship, the original second abnormal result is taken as a hardware risk feature; the matching weight is determined by domain knowledge. 4.The multi-dimensional information fusion evaluation method for intelligent product quality safety risk according to claim 1, characterized in that, The method for obtaining the software integration risk coefficient and the software integration risk feature comprises: The software integration data includes one type of data and two types of data; the one type of data includes module coupling degree, interface complexity and test coverage; and the two types of data include code vulnerability and code repair and response delay; A third abnormal result is obtained by comparing the software integration data with corresponding threshold values; the third abnormal result includes abnormal categories and corresponding abnormal data quantity and abnormal deviation cumulative value; The abnormal result of the one type of data is taken as one type of third abnormal result, the abnormal result of the two types of data is taken as two types of third abnormal result, data evaluation is performed on the software integration data according to the one type of third abnormal result and the two types of third abnormal result, a first quality coefficient and a second quality coefficient are obtained, and the software integration risk coefficient is obtained by weighting the first quality coefficient and the second quality coefficient; The one type of third abnormal result is cross-enhanced according to the second quality coefficient to obtain one type of third abnormal optimization result, the two types of third abnormal result are cross-enhanced according to the first quality coefficient to obtain two types of third abnormal optimization result, and the third abnormal optimization result is obtained by adjusting and splicing the one type of third abnormal optimization result and the two types of third abnormal optimization result; The user feedback use abnormality is matched with the third abnormal optimization result, if there is a matching relationship, the corresponding abnormal data quantity and abnormal deviation cumulative value are enhanced by using a matching weight, the enhanced third abnormal optimization result is taken as a software integration risk feature, and if there is no matching relationship, the original third abnormal optimization result is taken as a software integration risk feature. 5.The multi-dimensional information fusion evaluation method for intelligent product quality safety risk according to claim 1, characterized in that, The method for obtaining the environment risk coefficient and the environment risk feature comprises: Fourth abnormal results are obtained by comparing environment data with corresponding threshold values; the environment data includes operation environment and Internet of Things environment; and the fourth abnormal results include abnormal categories and corresponding abnormal data quantity and abnormal deviation cumulative value; The user feedback use abnormality is matched with the abnormal categories of the fourth abnormal result, if there is a matching relationship, the corresponding abnormal data quantity and abnormal deviation cumulative value are enhanced by using a matching weight, the enhanced fourth abnormal result is taken as an environment risk feature, and if there is no matching relationship, the original fourth abnormal result is taken as an environment risk feature. The influence degree of the environmental risk feature on the hardware risk feature obtains a first environmental correlation degree, the influence degree of the environmental risk feature on the software integration risk feature obtains a second environmental correlation degree, and a weighted value of the first environmental correlation degree and the second environmental correlation degree obtains an environmental risk coefficient. 6.The multi-dimensional information fusion evaluation method for intelligent product quality safety risk according to claim 1, characterized in that, The method for constructing the risk correlation propagation network comprises: The raw material production risk layer, the hardware risk layer, the software integration risk layer and the environmental risk layer are set, wherein the raw material production risk layer is only a precursor layer of the hardware risk layer, data indexes of each layer are taken as network nodes, edge weights between nodes in the same layer are determined according to abnormal result entropy values of the nodes, and edge weights between nodes in different layers are determined according to mutual information of abnormal results of the nodes, risk values of each layer and influence degrees of each layer, and an expression is as follows: ； ； ； ； in Risk layer node for raw material production Hardware risk layer nodes edge weights, Hardware risk layer node Software integration risk layer node edge weights, Hardware risk layer node With environmental risk layer nodes edge weights, For software integration risk layer nodes With environmental risk layer nodes edge weights, Risk layer node for raw material production The risk factor, This is a set of nodes in the raw material production risk layer. For hardware risk factor, To account for the risk factor of software integration, For environmental risk coefficient, This refers to the mutual information between the two nodes corresponding to abnormal results. As a correlation moderating factor, As the first level of influence, The first environmental correlation, This represents the second environmental correlation degree; An iteration update condition of the risk correlation propagation network is set, and risk coefficients of each layer are updated according to the edge weights, and an expression is as follows: ； in for layer The iterative risk coefficient update includes a hardware risk layer, a software integration risk layer, and an environmental risk layer. This is the risk attenuation coefficient. For risk propagation coefficient, for The set of nodes in a layer For nodes The set of predecessor nodes, Predecessor node The risk coefficient of the layer in question For nodes and nodes edge weights, for Iterating predecessor node Corresponding abnormal data Risk information entropy, This corresponds to the number of abnormal data. 7.The multi-dimensional information fusion evaluation method of intelligent product quality safety risk according to claim 1, characterized in that, The method for obtaining the intelligent product quality safety risk evaluation result comprises: The historical raw material production risk coefficient, the hardware risk coefficient, the software integration risk coefficient, the environmental risk coefficient, the potential hardware risk feature, the hardware risk feature, the software integration risk feature, the environmental risk feature and corresponding safety risks are taken as a quality safety risk set, the quality safety risk set is divided into a training set and a test set according to a ratio of 6:4 by using a random forest algorithm, an intelligent product quality safety risk evaluation model is trained by using the training set, and performance of the intelligent product quality safety risk evaluation model is evaluated by using the test set; The intelligent product quality safety risk evaluation model comprises a feature processing layer, a risk propagation layer, a prediction layer and an output layer; the feature processing layer respectively performs one-hot encoding on category type features, performs periodic encoding on time series features, splices and outputs processed features to the risk propagation layer and the prediction layer; the risk propagation layer is embedded with the risk correlation propagation network, simulates a propagation path of risks between components in a system, dynamically updates risk coefficients of each layer and outputs to the prediction layer; the prediction layer adopts a bidirectional long short-term memory network based on an attention mechanism, learns a complex nonlinear relationship between risk features and risk coefficients, performs risk grade classification and probability prediction; and the output layer extracts risk propagation factors from the risk propagation layer according to a prediction result, and outputs an intelligent product quality safety risk evaluation result in combination with a risk grade classification result and a probability prediction result. The intelligent product quality safety risk evaluation model evaluates differences between safety risk prediction results and real safety risks by using a hybrid loss function; and the hybrid loss function comprises a classification loss, a regression loss and a risk consistency loss. Multi-dimensional information of an intelligent product to be evaluated is input into the intelligent product quality safety risk evaluation model to obtain an intelligent product quality safety risk evaluation result.

Images