Effective extraction and fusion method of multi-source data at the traffic edge in multi-source sensing environment

Abstract

The invention relates to the technical field of data fusion, and discloses an effective extraction and fusion method for multi-source data of traffic edge in a multi-source sensing environment, comprising the following steps: step S01: inputting target data and performing preliminary processing; step S02: performing data feature extraction and classification identification on the target data; step S03: obtaining continuous fusion data features; step S04: obtaining discrete fusion data features; step S05: obtaining final fusion data; by providing step S03, step S04 and step S05, traffic data is divided into continuous data and discrete data, classified and processed, and then fused; when dealing with traffic data in a multi-source sensing environment, different fusion methods are adopted for different types of data for corresponding preliminary fusion, and then dimension reduction is performed to fuse the continuous data with the discrete data, thereby improving the accuracy and adaptability of data fusion, providing convenience for subsequent use of data, and laying a foundation.

Description

Technical Field

[0001] The present invention relates to the field of data fusion technology, and more specifically to a method for effectively extracting and fusing multi-source data at the traffic edge in a multi-source sensing environment. Background Art

[0002] With the rapid development of intelligent transportation systems, multi-source sensors are widely used in traffic monitoring, vehicle tracking, traffic flow analysis and other fields. These sensors can collect a variety of environmental parameters including sound, image, speed, direction, etc., providing rich data support for traffic management. Therefore, extracting effective information from massive multi-source data and realizing data fusion are crucial to the field of intelligent transportation.

[0003] Publication number CN103838772B discloses a multi-source traffic data fusion method, including: acquiring traffic information data from three data sources: mobile phones, floating vehicles, and traffic video images; performing pixel-level fusion on the traffic information data to eliminate unqualified information from the traffic information data; performing feature-level fusion on the traffic information data from each of the three data sources to generate traffic status information for a road section from the three different data sources; performing decision-level fusion on the traffic status information for the road section from the three different data sources to generate consistent traffic status description information for the road section; and outputting the traffic status description information for the road section. The present invention acquires traffic information data from multiple data sources, performs three-level fusion, and generates the final traffic status of the road section, enabling more accurate determination of road traffic conditions.

[0004] In the existing technology, data fusion is mainly divided into data-level fusion, feature-level fusion and decision-level fusion, etc. However, traffic data is complex, diverse and of different types. When dealing with traffic data in a multi-source sensing environment, there are still problems such as low fusion accuracy and poor real-time performance. Secondly, for different traffic problems, different traffic data need to be extracted, fused and analyzed accordingly. For example, when analyzing traffic congestion, it is necessary to extract and fuse traffic flow data and traffic signal data accordingly; when analyzing traffic management conditions, it is necessary to extract, fuse and analyze road condition data, meteorological information data and air quality data accordingly. If all traffic data are fused, the amount of calculation will be too large due to too much data, and there will be a lot of irrelevant data, resulting in inaccurate analysis results. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides an effective extraction and fusion method for multi-source data of traffic edge in a multi-source sensing environment to solve the problems existing in the above-mentioned background technology.

[0006] The present invention provides the following technical solution: a method for effectively extracting and fusing multi-source data from traffic edges in a multi-source sensing environment, comprising the following steps:

[0007] Step S01: Input target data and perform preliminary processing: the target data is the traffic data that needs to be extracted and integrated;

[0008] Step S02: Input the target data into the constructed data recognition model to extract data features and classify the data into two categories: continuous data and discrete data;

[0009] Step S03: Perform continuous class fusion processing on the data features of the continuous data, and build a continuous fusion model to obtain continuous fusion data features;

[0010] Step S04: performing discrete class fusion processing on the data features of discrete data to obtain discrete fusion data features;

[0011] Step S05: performing dimensionality reduction processing on the continuous fusion data features and the discrete fusion data features, and then performing final fusion to obtain final fusion data.

[0012] Preferably, the preliminary processing includes cleaning the target data, removing outliers and missing values, and performing data alignment operations on the target data, including time alignment and space alignment, to unify the target data into the same time dimension and space dimension;

[0013] The target data includes video surveillance data, vehicle information data, traffic flow data, vehicle driving status data, environmental monitoring data, traffic accident data, traffic violation data and road facility status data; the video surveillance data includes video surveillance content, the vehicle information data includes vehicle number, vehicle model, vehicle color and vehicle speed, the traffic flow data includes the number of vehicles passing through a certain time period, the vehicle driving status data includes vehicle acceleration and steering angle, the environmental monitoring data includes road conditions and air quality, the traffic accident data includes the time, location and number of traffic accidents, the traffic violation data includes the number of illegal parking and vehicle speeding information, and the road facility status data includes road lighting status, traffic signal light status and signboard status.

[0014] Preferably, the constructed data recognition model includes an input gate, a forget gate and an output gate;

[0015] Denote the current input as x t , that is, the input at time t; the hidden layer vector at time t is represented as h t , the cell state at time t is expressed as C t , and the final output is expressed as y t;but: ;

[0016] ;

[0017] ;

[0018] Among them, I t is the input gate, f t For the forget gate, O t is the output gate; W I is the weight matrix of the input gate, W f is the weight matrix of the forget gate, W O is the weight matrix of the output gate; b I is the bias of the input gate, b f is the bias of the forget gate, b O is the bias of the output gate; σ is the Sigmoid function;

[0019] At time t, the update formula of the cell state is: ,in, represents the Hadamard product, is the updated cell state, and the formula is expressed as: , where tanh is the hyperbolic tangent activation function, W C is the weight matrix of the cell state, b C is the bias of the cell state;

[0020] The formula for the hidden layer vector is: ;

[0021] The final output formula is expressed as: , where W y is the weight matrix of the output layer, b y is the bias of the output layer.

[0022] Preferably, the specific method of obtaining the continuous fusion data features is:

[0023] Divide the continuous data into several data sets, and then divide the data sets into training part and output part. For each data set in the training part, use regression analysis method to build a single model to describe the pattern of the data in the data set; for each single model, calculate the model weight;

[0024] Based on the calculated model weights, the individual models are fused to obtain a continuous fusion model. The output part is input into the continuous fusion model, and the output obtained is the continuous fusion data feature LR.

[0025] Preferably, the calculation formula of the model weight is: ; Among them, ωm is the weight of the single model, k is the adjustment parameter of the model weight, and its value range is [0, 1], n m is the number of continuous data in the dataset corresponding to the single model m; N is the number of data in the target data; d m is the error evaluation index of the single model m, c and a are positive real number parameters;

[0026] SC m It is the output value of the single model for the input sample, and the input sample is the data of the output part.

[0027] Preferably, the specific method of obtaining the discrete fusion data features is:

[0028] The discrete data is constructed into a decision table, which consists of a condition attribute set and a decision attribute set. The condition attribute set includes several condition attribute value vectors, and the decision attribute value set includes several decision attribute values. Each target object in the decision table corresponds to a condition attribute value vector and a decision attribute value. All target objects constitute a target object set U. The condition attribute set is the various discrete data, and the decision attribute set is the traffic problem. If the target data collection area is divided into Q, there are Q target objects.

[0029] For each attribute subset in the conditional attribute set B, the target objects in the target object set U are grouped according to the value of the attribute subset, and each group forms an equivalence class; the attribute subset is composed of several conditional attribute value vectors;

[0030] For any attribute subset X, calculate the upper and lower deterministic sets of the attribute subset under the conditional attribute value vector in the conditional attribute set; the upper deterministic set includes all target objects that completely belong to the attribute subset, and the lower deterministic set includes the equivalence class in which some target objects belong to the attribute subset;

[0031] Through the cardinality of the upper determination set and the lower determination set, that is, the number of elements, the roughness of the attribute subset to the decision attribute value g under the conditional attribute value vector of the conditional attribute set is calculated, that is, the degree to which it cannot be completely determined by the conditional attribute set.

[0032] Preferably, the calculation formula of the roughness is: ;

[0033] Among them, β B(X) represents the roughness of X under the conditional attribute value vector in B, B X - is the lower deterministic set of X under the conditional attribute value vector in B, B X +is the upper deterministic set of X under the conditional attribute value vector in B; λ is the adjustment parameter, and its value range is [0, 1]; S(xi∈X) is the characteristic function, which takes the value 1 when xi belongs to X, otherwise it takes the value 0; w i is the weight of the i-th target object xi; γ is the time adjustment parameter, T X is the average timestamp of the data in the attribute subset X, T is the current time; θ is a decay coefficient used to control the rate of time decay;

[0034] Based on the roughness, calculate the dependency of the decision attribute value g on the conditional attribute value vector in B; Among them, D B(g) is the dependency of the decision attribute value g on the conditional attribute value vector in B;

[0035] Define the importance of a single conditional attribute value vector q to the decision attribute value g as the dependency when q is included, minus the dependency when q is not included; sort all conditional attribute value vectors q from large to small according to their importance values ​​to obtain the sorted attributes; set an importance threshold YU, traverse the sorted attributes, if the corresponding importance is greater than YU, then retain q in B; if the importance is less than or equal to YU, then remove q from B and regard it as a redundant attribute. The remaining attributes in B are the core attribute set, and the other removed attributes are regarded as redundant; use the core attribute set to construct a new decision table; record it as a simplified decision table.

[0036] Preferably, the decision rule is extracted from the simplified decision table, the antecedent of the decision rule is a combination of condition attribute value vectors, and the consequent is the corresponding decision attribute value;

[0037] Calculating the confidence of a decision rule ; Among them, r1 represents the set of target objects that fully support the decision rule r, r2 represents the set of target objects that partially support the decision rule; μ1 and μ2 are confidence adjustment parameters, and Gr is the number of conditional attribute values ​​of the decision rule;

[0038] Compute coverage of decision rules: Among them, FH r To cover the function, ; Among them, η is the coverage adjustment parameter, the value range is [0, 1]; φ is the exponential decay rate parameter;

[0039] The comprehensive weight ZQ(r) of the decision rule is calculated based on the coverage and confidence of the decision rule. ; Among them, ε is the comprehensive weight parameter;

[0040] Get fused data features ; Where JC(r) is the vector of decision rules, which is a 0-1 vector with a dimension equal to the number of all different conditional attribute values;

[0041] The fused data features are mapped into numerical feature vectors, and the comprehensive weights of the decision rules are used as the component values ​​of the feature vectors.

[0042] Preferably, the specific method of obtaining the final fusion data is:

[0043] The fused continuous data and discrete data are regarded as a data cluster respectively. For each data cluster, the corresponding continuous fused data features and discrete fused data features are mapped to a low-dimensional space, that is, dimensionality reduction is performed;

[0044] For every two data points in the same data cluster, the eigenvectors v and z are calculated, and the comprehensive similarity K(vz) between v and z is calculated. The eigenvector similarity matrix K is constructed by taking the comprehensive similarity K(vz) as an element. ; Where N(v) is the k-nearest neighbor set of the feature vector v; N(z) is the k-nearest neighbor set of the feature vector z; α is an exponential parameter that can adjust the degree of influence of the number of shared nearest neighbors on similarity; τ is the intra-cluster distance adjustment parameter that flexibly controls the degree of influence of the position within the cluster on similarity; Mv is the distance from the feature vector v to the cluster center of the data cluster to which it belongs; Mz is the distance from the feature vector z to the cluster center of the data cluster to which it belongs;

[0045] Randomly initialize H low-dimensional data points in a preset low-dimensional space. For every two low-dimensional data points in the low-dimensional space, calculate the comprehensive similarity between them, and use the comprehensive similarity as an element to construct a low-dimensional space similarity matrix Z. Calculate the KL divergence between the low-dimensional space similarity matrix Z and the eigenvector similarity matrix K. Use the gradient descent method to iteratively optimize the position of the low-dimensional data points to minimize the KL divergence, so that the similarity matrix Z in the low-dimensional space is as close as possible to the similarity matrix K in the high-dimensional space. After the iterative optimization converges, the final low-dimensional data point is obtained, which corresponds to the data in the data cluster, and the dimensionality reduction is completed.

[0046] The continuous data features after dimensionality reduction are concatenated and fused with the discrete data features to obtain the final fused data.

[0047] Technical effects and advantages of the present invention:

[0048] The present invention is provided with step S03, step S04 and step S05, which is conducive to performing continuous-type fusion processing on the data features of continuous data, constructing a continuous fusion model to obtain continuous fusion data features, performing discrete-type fusion processing on the data features of discrete data, obtaining discrete fusion data features, and then performing dimensionality reduction processing on the continuous fusion data features and the discrete fusion data features, and then performing final fusion to obtain final fusion data; dividing traffic data into continuous data and discrete data, performing classification processing and then fusing them, when dealing with traffic data in a multi-source sensing environment, adopting different fusion methods for different types of data to perform corresponding preliminary fusion, and then performing dimensionality reduction to fuse the continuous data with the discrete data, thereby improving the accuracy and adaptability of data fusion, providing convenience for subsequent use of data, and laying a foundation. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of the method for effectively extracting and fusing multi-source data from traffic edges in a multi-source sensing environment of the present invention. DETAILED DESCRIPTION

[0050] The technical solutions of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the present invention. In addition, the forms of the various structures described in the following embodiments are merely examples. The method for effectively extracting and fusing multi-source data at the edge of traffic in a multi-source sensing environment involved in the present invention is not limited to the various structures described in the following embodiments. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0051] like Figure 1 As shown, the present invention provides an effective extraction and fusion method for multi-source data of traffic edge in a multi-source sensing environment, comprising the following steps:

[0052] Step S01: Input target data and perform preliminary processing: the target data is traffic data that needs to be extracted and integrated; the preliminary processing includes but is not limited to cleaning the target data, eliminating outliers and missing values, and performing data alignment operations on the target data, including time alignment and space alignment, unifying the target data to the same time dimension and space dimension. Since the target data contains data collected at different time points and by different edge devices, it is necessary to unify the time and space dimensions of the target data; the target data includes but is not limited to video surveillance data, vehicle information data, traffic flow data, vehicle driving status data, environmental monitoring data, traffic accident data, traffic violation data, and road facility status data; the video surveillance data includes but is not limited to video surveillance content, and the vehicle information data includes but is not limited to The traffic flow data includes but is not limited to the number of vehicles passing through in a certain time period, the vehicle driving status data includes but is not limited to vehicle acceleration, steering angle, etc., the environmental monitoring data includes but is not limited to road conditions, air quality, etc., the traffic accident data includes but is not limited to the time, location, and number of traffic accidents, the traffic violation data includes but is not limited to the number of illegal parking and vehicle speeding information, and the road facility status data includes but is not limited to road lighting status, traffic signal light status, signboard status, etc.; the target data can be unified in time dimension through interpolation method, and the target data can be mapped to the same spatial coordinate system to reflect the spatial relationship and position relationship between the target data, thereby unifying the spatial dimension;

[0053] Step S02: Input the target data into the constructed data recognition model for data feature extraction and classification recognition, and divide the data into two categories: continuous data and discrete data; the data recognition model is specifically a recurrent neural network model; its purpose is that the weight parameters in the recurrent neural network model are shared, which helps to reduce the number of parameters of the model and improve computational efficiency. Weight sharing enables the recurrent neural network model to use the same weight parameters for calculation when processing data at different time steps, thereby reducing the complexity and computational cost of the model. Through its loop structure, it can also capture the time series information and context information in the data, so it can remember past information and use this information for current decision-making, so that it can better understand the context information when processing data, and can also be flexibly adjusted according to different input and output structures to adapt to different application scenarios; therefore, it is more adaptable when facing a variety of complex and different traffic data. At the same time, for traffic data, past traffic data can be applied to current decisions, thereby improving the accuracy of classification of complex and diverse traffic data;

[0054] Step S03: Perform continuous class fusion processing on the data features of the continuous data, and build a continuous fusion model to obtain continuous fusion data features;

[0055] Step S04: performing discrete class fusion processing on the data features of discrete data to obtain discrete fusion data features;

[0056] Step S05: performing dimensionality reduction processing on the continuous fusion data features and the discrete fusion data features, and then performing final fusion to obtain final fusion data.

[0057] In this embodiment, it should be specifically explained that the constructed data recognition model includes an input gate, a forget gate, and an output gate. Using the LSTM model as the data recognition model can control the information of the cell state through the LSTM design gate structure. This structure can effectively filter out the characteristics of long-time step data, thereby being able to solve the problem of vanishing gradients in the original RNN. These three gates have their own functions. The input gate determines the retention of current information, the forget gate determines the retention of previous information to discard useless state information, and the output gate determines the retention of updated state information.

[0058] Denote the current input as x t , that is, the input at time t; the hidden layer vector at time t is represented as h t , the cell state at time t is expressed as C t , and the final output is expressed as y t ;but: ;

[0059] ;

[0060] ;

[0061] Among them, I t is the input gate, f t For the forget gate, O t is the output gate; W I is the weight matrix of the input gate, W f is the weight matrix of the forget gate, W O is the weight matrix of the output gate; b I is the bias of the input gate, b f is the bias of the forget gate, b O is the bias of the output gate; σ is the activation function, which is the Sigmoid function;

[0062] At time t, the update formula of the cell state is: ,in, represents the Hadamard product, is the updated cell state, and the formula is expressed as: , where tanh is the hyperbolic tangent activation function, W C is the weight matrix of the cell state, b C is the bias of the cell state;

[0063] The formula for the hidden layer vector is: ;

[0064] The final output formula is expressed as: , where W y is the weight matrix of the output layer, b y is the bias of the output layer;

[0065] Collect historical data to train the data recognition model. 75% of the data can be used as the training set and 25% as the validation set to train and optimize the data recognition model.

[0066] In this embodiment, it should be specifically explained that the specific method of obtaining the continuous fusion data features is:

[0067] Divide the continuous data into several data sets, and divide the data sets into training part and output part. For each data set in the training part, use the regression analysis method to build a single model to describe the pattern of the data in the data set; for each single model, calculate the model weight ; Among them, ω m is the weight of the single model, k is the adjustment parameter of the model weight, the value range is [0, 1], which controls the relative importance of the two weight items; n m is the number of continuous data in the dataset corresponding to the single model m; N is the number of data in the target data; d m is the error evaluation index of the single model m, such as root mean square error, R square, etc.; c and a are positive real number parameters used to control the shape of the function;

[0068] Based on the calculated model weights, the individual models are fused to obtain a continuous fusion model, and the output part is input into the continuous fusion model. The output obtained is the continuous fusion data feature LR;

[0069] SC m It is the output value of the single model for the input sample, and the input sample is the data of the output part.

[0070] In this embodiment, it should be specifically explained that the specific method of obtaining the discrete fusion data features is:

[0071] The discrete data is constructed into a decision table, which consists of a condition attribute set and a decision attribute set. The condition attribute set includes several condition attribute value vectors, and the decision attribute value set includes several decision attribute values. Each target object in the decision table corresponds to a condition attribute value vector and a decision attribute value. All target objects constitute a target object set U. The condition attribute set is the various discrete data, and the decision attribute set is the traffic problem. If the target data collection area is divided into Q, there are Q target objects.

[0072] For each attribute subset in the conditional attribute set B, the target objects in the target object set U are grouped according to the value of the attribute subset, and each group forms an equivalence class; the attribute subset is composed of several conditional attribute value vectors;

[0073] For any attribute subset X, calculate the upper and lower deterministic sets of the attribute subset under the conditional attribute value vector in the conditional attribute set; the upper deterministic set includes all target objects that completely belong to the attribute subset, and the lower deterministic set includes the equivalence class in which some target objects belong to the attribute subset;

[0074] Through the cardinality of the upper and lower determination sets, i.e. the number of elements, the roughness of the attribute subset to the decision attribute value g under the conditional attribute value vector of the conditional attribute set is calculated, i.e. the degree to which it cannot be completely determined by the conditional attribute set;

[0075] The calculation formula of the roughness is: ;

[0076] Among them, β B(X) represents the roughness of X under the conditional attribute value vector in B, B X - is the lower deterministic set of X under the conditional attribute value vector in B, B X + is the upper deterministic set of X under the conditional attribute value vector in B; λ is a tuning parameter used to control the influence of the weighted term, usually in the range of [0, 1]; S(xi∈X) is an indicator function, which takes the value 1 when xi belongs to X and 0 otherwise; w i is the weight of the i-th target object xi; γ is the time adjustment parameter, which is used to control the influence of the time factor, T X is the average timestamp of the data in the attribute subset X, T is the current time; θ is a decay coefficient used to control the rate of time decay, usually a positive value;

[0077] Based on the roughness, calculate the dependency of the decision attribute value g on the conditional attribute value vector in B;

[0078] Among them, D B(g)is the dependency of the decision attribute value g on the conditional attribute value vector in B;

[0079] The importance of a single conditional attribute value vector q to the decision attribute value g is defined as the dependence when q is included minus the dependence when q is not included. The larger the importance value, the greater the influence of attribute a on the determination of g.

[0080] Sort all conditional attribute value vectors q by their importance values ​​from large to small to obtain the sorted attributes; set an importance threshold YU and traverse the sorted attributes. If the corresponding importance is greater than YU, q is retained in B; if the importance is less than or equal to YU, q is removed from B and regarded as a redundant attribute. The remaining attributes in B are the core attribute set, and the other removed attributes are regarded as redundant; use the core attribute set to construct a new decision table, which is recorded as the reduced decision table;

[0081] The decision rules are extracted from the simplified decision table using the MODLEM rule extraction algorithm. The antecedent of the decision rule is the combination of the condition attribute value vectors, and the consequent is the corresponding decision attribute value.

[0082] Calculating the confidence of a decision rule ; Among them, r1 represents the set of target objects that fully support the decision rule r, r2 represents the set of target objects that partially support the decision rule; μ1 and μ2 are confidence adjustment parameters used to control the impact of partially supported target objects and rule-covered target objects on confidence; Gr is the number of conditional attribute values ​​of the decision rule;

[0083] Compute coverage of decision rules: Among them, FH r To cover the function, ; Among them, η is the coverage adjustment parameter, the value range is [0, 1]; φ is the exponential decay rate parameter, which is a positive real number;

[0084] The comprehensive weight ZQ(r) of the decision rule is calculated based on the coverage and confidence of the decision rule. ; Among them, ε is the comprehensive weight parameter, which is used to control the relative importance of confidence and coverage in the comprehensive weight;

[0085] Get fused data features ; Where JC(r) is the vector of decision rules, which is a 0-1 vector with a dimension equal to the number of all different conditional attribute values;

[0086] The fused data features are mapped into numerical feature vectors, and the comprehensive weights of the decision rules are used as the component values ​​of the feature vectors.

[0087] In this embodiment, it should be specifically explained that the specific method of obtaining the final fusion data is:

[0088] The fused continuous data and discrete data are regarded as a data cluster respectively. For each data cluster, the corresponding continuous fused data features and discrete fused data features are mapped to a low-dimensional space, that is, dimensionality reduction is performed to reduce the dimension of the data.

[0089] For every two data points in the same data cluster, the eigenvectors v and z are calculated, and the comprehensive similarity K(vz) between v and z is calculated. The eigenvector similarity matrix K is constructed by taking the comprehensive similarity K(vz) as an element.

[0090] ; Where N(v) is the k-nearest neighbor set of the feature vector v; N(z) is the k-nearest neighbor set of the feature vector z; α is an exponential parameter that can adjust the degree of influence of the number of shared nearest neighbors on similarity; τ is the intra-cluster distance adjustment parameter that flexibly controls the degree of influence of the position within the cluster on similarity; Mv is the distance from the feature vector v to the cluster center of the data cluster to which it belongs; Mz is the distance from the feature vector z to the cluster center of the data cluster to which it belongs;

[0091] Randomly initialize H low-dimensional data points in a preset low-dimensional space. For every two low-dimensional data points in the low-dimensional space, calculate the comprehensive similarity between them, and use the comprehensive similarity as an element to construct a low-dimensional space similarity matrix Z. Calculate the KL divergence between the low-dimensional space similarity matrix Z and the eigenvector similarity matrix K. Use the gradient descent method to iteratively optimize the position of the low-dimensional data points to minimize the KL divergence, so that the similarity matrix Z in the low-dimensional space is as close as possible to the similarity matrix K in the high-dimensional space. After the iterative optimization converges, the final low-dimensional data point is obtained, which corresponds to the data in the data cluster, and the dimensionality reduction is completed.

[0092] By optimizing the position of data in low-dimensional space, the similarity in low-dimensional space can preserve the local and global structure in high-dimensional space as much as possible;

[0093] The continuous data features after dimensionality reduction are concatenated and fused with the discrete data features to obtain the final fused data.

[0094] In this embodiment, it should be specifically explained that the difference between this embodiment and the prior art mainly lies in that this embodiment has step S03, step S04 and step S05, which is conducive to performing continuous fusion processing on the data features of continuous data, building a continuous fusion model to obtain continuous fusion data features, performing discrete fusion processing on the data features of discrete data, obtaining discrete fusion data features, and then performing dimensionality reduction processing on the continuous fusion data features and the discrete fusion data features, and then performing final fusion to obtain final fusion data; dividing traffic data into continuous data and discrete data, performing classification processing and then performing fusion, when dealing with traffic data in a multi-source sensing environment, adopting different fusion methods for different types of data to perform corresponding preliminary fusion, and then performing dimensionality reduction to fuse the continuous data with the discrete data, thereby improving the accuracy and adaptability of data fusion, providing convenience for subsequent use of data, and laying a foundation.

[0095] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0096] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. An effective method for extracting and fusing multi-source data at the edge of traffic in a multi-source sensing environment, characterized by: The following steps are involved: Step S01: input target data and perform preliminary processing: the target data is the traffic data that needs to be extracted and integrated; Step S02: Input the target data into the constructed data recognition model to extract data features and classify the data into two categories: continuous data and discrete data; Step S03: Perform continuous class fusion processing on the data features of the continuous data, and construct a continuous fusion model to obtain continuous fusion data features; Step S04: performing discrete class fusion processing on the data features of discrete data to obtain discrete fusion data features; Step S05: performing dimensionality reduction processing on the continuous fusion data features and the discrete fusion data features, and then performing final fusion to obtain final fusion data; The method of obtaining discrete fusion data features is: Construct a decision table, which consists of a condition attribute set and a decision attribute set. For each attribute subset in the condition attribute set B, it consists of several condition attribute value vectors. For any attribute subset X, calculate the upper determination set and lower determination set of the attribute subset under the condition attribute value vector in the condition attribute set. Through the cardinality of the upper determination set and the lower determination set, that is, the number of elements, calculate the roughness of the attribute subset under the condition attribute value vector of the condition attribute set to the decision attribute value g, that is, the degree to which it cannot be completely determined by the condition attribute set. The calculation formula of the roughness is: ； Among them, β B(X) represents the roughness of X under the conditional attribute value vector in B, B X - is the lower deterministic set of X under the conditional attribute value vector in B, B X + is the upper deterministic set of X under the conditional attribute value vector in B; λ is the adjustment parameter, and its value range is [0, 1]; S(xi∈X) is the characteristic function, which takes the value 1 when xi belongs to X, otherwise it takes the value 0; w i is the weight of the i-th target object xi; γ is the time adjustment parameter, T X is the average timestamp of the data in the attribute subset X, T is the current time; θ is a decay coefficient; Based on the roughness, calculate the dependency of the decision attribute value g on the conditional attribute value vector in B; ; Among them, D B(g) is the dependency of the decision attribute value g on the conditional attribute value vector in B; Define the importance of a single conditional attribute value vector q to the decision attribute value g as the dependency when q is included, minus the dependency when q is not included; sort all conditional attribute value vectors q from large to small according to their importance values ​​to obtain the sorted attributes; set an importance threshold YU, traverse the sorted attributes, if the corresponding importance is greater than YU, then keep q in B; if the importance is less than or equal to YU, then remove q from B and regard it as a redundant attribute, and the remaining attributes in B are the core attribute set, and the other removed attributes are regarded as redundant; use the core attribute set to build a new decision table; recorded as a simplified decision table.

2. The method for effectively extracting and fusing multi-source data at the edge of traffic in a multi-source sensing environment according to claim 1 is characterized by: The preliminary processing includes cleaning the target data, removing outliers and missing values, and performing data alignment operations on the target data, including time alignment and space alignment, to unify the target data into the same time dimension and space dimension; The target data includes video surveillance data, vehicle information data, traffic flow data, vehicle driving status data, environmental monitoring data, traffic accident data, traffic violation data and road facility status data; the video surveillance data includes video surveillance content, the vehicle information data includes vehicle number, vehicle model, vehicle color and vehicle speed, the traffic flow data includes the number of vehicles passing through a certain time period, the vehicle driving status data includes vehicle acceleration and steering angle, the environmental monitoring data includes road conditions and air quality, the traffic accident data includes the time, place and number of traffic accidents, the traffic violation data includes the number of illegal parking and vehicle speeding information, and the road facility status data includes road lighting status, traffic signal light status and signboard status.

3. The method for effectively extracting and fusing multi-source data at the edge of traffic in a multi-source sensing environment according to claim 1 is characterized by: The constructed data recognition model includes an input gate, a forget gate and an output gate; Represent the current input as x t , which is the input at time t; the hidden layer vector at time t is represented as h t , the cell state at time t is represented as C t , and the final output is represented as y t ;but: ； ； ； Among them, I t is the input gate, f t For the forget gate, O t is the output gate; W I is the weight matrix of the input gate, W f is the weight matrix of the forget gate, W O is the weight matrix of the output gate; b I is the bias of the input gate, b f is the bias of the forget gate, b O is the bias of the output gate; σ is the Sigmoid function; At time t, the update formula of the cell state is: ,in, represents the Hadamard product, is the updated cell state, and the formula is expressed as: , where tanh is the hyperbolic tangent activation function, W C is the weight matrix of the cell state, b C is the bias of the cell state; The formula for the hidden layer vector is: ; The final output formula is expressed as: , where W y is the weight matrix of the output layer, b y is the bias of the output layer.

4. The method for effectively extracting and fusing multi-source data at the edge of traffic in a multi-source sensing environment according to claim 1 is characterized by: The specific method of obtaining the continuous fusion data features is: Divide the continuous data into several data sets, and divide the data sets into training part and output part. For each data set in the training part, use regression analysis method to build a single model to describe the pattern of data in the data set. For each monomer model, calculate the model weight; Based on the calculated model weights, the monomer models are fused to obtain a continuous fusion model, and the output part is input into the continuous fusion model, and the output obtained is the continuous fusion data feature LR.

5. The method for effectively extracting and fusing multi-source data at the edge of traffic in a multi-source sensing environment according to claim 4 is characterized by: The calculation formula of the model weight is: ; Among them, ω m is the weight of the single model, k is the adjustment parameter of the model weight, and its value range is [0, 1], n m is the number of continuous data in the data set corresponding to the single model m; N is the number of data in the target data; d m is the error evaluation index of the single model m, c and a are positive real number parameters; SC m It is the output value of the single model for the input sample, and the input sample is the data of the output part.

6. The method for effectively extracting and fusing multi-source data at the edge of traffic in a multi-source sensing environment according to claim 1 is characterized by: The specific method of obtaining the discrete fusion data features is: The discrete data is constructed into a decision table, which consists of a condition attribute set and a decision attribute set. The condition attribute set includes a number of condition attribute value vectors, and the decision attribute value set includes a number of decision attribute values. Each target object in the decision table corresponds to a condition attribute value vector and a decision attribute value. All target objects constitute a target object set U. The condition attribute set is various types of discrete data, and the decision attribute set is a traffic problem. If the target data collection area is divided into Q, there are Q target objects. For each attribute subset in the conditional attribute set B, the target objects in the target object set U are grouped according to the value of the attribute subset, and each group forms an equivalence class; the attribute subset is composed of a number of conditional attribute value vectors; For any attribute subset X, calculate the upper and lower determination sets of the attribute subset under the conditional attribute value vector in the conditional attribute set; the upper determination set contains all target objects that completely belong to the attribute subset, and the lower determination set contains the equivalence class in which some target objects belong to the attribute subset; Through the cardinality of the upper and lower determination sets, that is, the number of elements, the roughness of the attribute subset to the decision attribute value g under the conditional attribute value vector of the conditional attribute set is calculated, that is, the degree to which it cannot be completely determined by the conditional attribute set.

7. The method for effectively extracting and fusing multi-source data at the edge of traffic in a multi-source sensing environment according to claim 1 is characterized by: The simplified decision table extracts a decision rule, wherein the antecedent of the decision rule is a combination of conditional attribute value vectors, and the consequent is a corresponding decision attribute value; Calculating the confidence of a decision rule ; Among them, r1 represents the set of target objects that fully support the decision rule r, r2 represents the set of target objects that partially support the decision rule; μ1 and μ2 are confidence adjustment parameters, and Gr is the number of conditional attribute values ​​of the decision rule; Compute coverage of decision rules: Among them, FH r To cover the function, ; Wherein, η is the coverage adjustment parameter, and its value range is [0, 1]; is the exponential decay rate parameter; The comprehensive weight ZQ(r) of the decision rule is calculated based on the coverage and confidence of the decision rule. ; Among them, ε is the comprehensive weight parameter; Get fusion data features ; Where JC(r) is the vector of decision rules, which is a 0-1 vector with a dimension equal to the number of all different conditional attribute values; The fused data features are mapped into numerical feature vectors, and the comprehensive weights of the decision rules are used as the component values ​​of the feature vectors.

8. The method for effectively extracting and fusing multi-source data at the edge of traffic in a multi-source sensing environment according to claim 1 is characterized by: The specific method of obtaining the final fusion data is: The fused continuous data and discrete data are regarded as a data cluster respectively. For each data cluster, the corresponding continuous fused data features and discrete fused data features are mapped to a low-dimensional space, that is, dimensionality reduction is performed; For every two feature vectors v and z corresponding to the same data cluster, calculate the comprehensive similarity K(vz) between v and z; use the comprehensive similarity K(vz) as an element to construct the feature vector similarity matrix K; ; Where N(v) is the k-nearest neighbor set of the feature vector v; N(z) is the k-nearest neighbor set of the feature vector z; α is an exponential parameter that can adjust the degree of influence of the number of shared nearest neighbors on similarity; τ is the intra-cluster distance adjustment parameter that flexibly controls the degree of influence of the intra-cluster position on similarity; Mv is the distance from the feature vector v to the cluster center of the data cluster to which it belongs; Mz is the distance from the feature vector z to the cluster center of the data cluster to which it belongs; Randomly initialize H low-dimensional data points in a preset low-dimensional space. For every two low-dimensional data points in the low-dimensional space, calculate the comprehensive similarity between them, and use the comprehensive similarity as an element to construct a low-dimensional space similarity matrix Z; calculate the KL divergence between the low-dimensional space similarity matrix Z and the eigenvector similarity matrix K, and use the gradient descent method to iteratively optimize the position of the low-dimensional data points to minimize the KL divergence, so that the similarity matrix Z in the low-dimensional space is as close as possible to the similarity matrix K in the high-dimensional space; after the iterative optimization converges, the final low-dimensional data point is obtained, which corresponds to the data in the data cluster, that is, the dimensionality reduction is completed; The continuous data features after dimensionality reduction are concatenated and fused with the discrete data features to obtain the final fused data.

Images