Online car-hailing order risk control method and system based on acoustic event detection

By introducing acoustic event detection in the online ride-hailing industry, combining order data and itinerary recording, and using CRNN models to identify abnormal orders, the problem of high misjudgment rate in the existing technology is solved, and the accuracy and adaptability of abnormal order detection are improved. , reduces operating costs and improves service quality.

CN119047840APending Publication Date: 2024-11-29BEIJING BAILONG MAYUN TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411248467.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-06
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The existing technology has a high misjudgment rate in abnormal order detection in the online ride-hailing industry, which is difficult to adapt to complex and changeable scenarios, and traditional data characteristics are difficult to capture cheating, resulting in abnormal orders being misjudged as abnormal.

Method used

Using acoustic event detection methods, the itinerary recording of online ride-hailing orders is carried out in-depth analysis, combined with machine learning models, the audio feature matrix and order feature matrix are extracted, and the CRNN model is used to identify potential abnormal orders, which improves the recognition ability of the detection system. and accuracy.

Benefits of technology

It significantly reduces the misjudgment rate, improves the accuracy of abnormal order recognition, enhances the ability to adapt to complex scenarios, provides more reliable regulatory means, reduces labor costs, and improves user experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119047840A_ABST
    Figure CN119047840A_ABST
Patent Text Reader

Abstract

The invention discloses an online car-hailing order risk control method and system based on acoustic event detection. The invention relates to the field of data processing and intelligent transportation. Order data D, including order context information a, passenger information b, driver information c and journey recording d, are acquired in real time from a database of the online car-hailing platform; preprocessing the order data D, including data cleaning (removing invalid or wrong data); for an order i of which the probability alpha is greater than a preset threshold (such as 80%), inputting an audio feature matrix B of travel recording into the trained CRNN sound event detection model to generate a prediction result z; according to the method, the acoustic event features are introduced, and other context information of the order is combined, so that the order data can be analyzed more comprehensively, and the abnormal order can be identified more accurately. The online car-hailing platform is helped to find and handle potential fraud or illegal behaviors in time, and rights and interests of passengers and drivers are guaranteed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of data processing and intelligent transportation, and specifically to a calculation method for identifying abnormal online ride-hailing orders in an online ride-hailing aggregation platform based on online ride-hailing trip recordings and partial order information, and particularly to an online ride-hailing order risk control method and system based on acoustic event detection. Background Art

[0002] In the current online car-hailing industry, the core purpose of abnormal order detection technology is to effectively identify and handle orders that may violate service rules or involve fraud. Existing technologies mainly cover two categories of methods: rule-based abnormal order detection and machine learning-based abnormal order detection.

[0003] Rule-based abnormal order detection usually relies on a series of preset judgment criteria, which are formulated based on historical experience and industry knowledge to filter out orders that do not conform to normal patterns. However, the limitation of this method is that it has limited adaptability to complex and changing actual scenarios.

[0004] In contrast, the abnormal order detection method based on machine learning has shown greater flexibility and potential. By utilizing a large amount of historical order data, a model that can automatically identify abnormal patterns is trained. In practical applications, machine learning models can comprehensively consider data features in multiple dimensions such as order information, driver information, and passenger information to make more accurate abnormal judgments. However, in real scenarios, due to the complexity of passenger behavior, driver behavior, and the external environment, there are often some real orders that appear abnormal in data but are actually normal. Such orders are easily misjudged as abnormal in machine learning models that rely solely on statistical data features, which in turn leads to customer complaints from drivers.

[0005] The inventor believes that if the analysis results of the audio recordings of online ride-hailing order itineraries are integrated into the abnormal order detection system. Through in-depth analysis of the itinerary recordings, more cheating behaviors and abnormal patterns that are difficult to capture through traditional data features can be revealed. For example, specific acoustic events between passengers and drivers can be key clues to identifying abnormal orders. By combining the analysis results of itinerary recordings with machine learning models, the recognition ability of the abnormal order detection system can be significantly enhanced, and its adaptability to complex and changing scenarios can be improved. It can not only reduce the misjudgment rate, but also maximize the accuracy of abnormal order identification, thereby providing a more reliable and effective means of supervision for the online ride-hailing industry.

[0006] To this end, the present invention proposes a risk control method and system for online car-hailing orders based on acoustic event detection. Summary of the invention

[0007] In view of this, the present invention hopes to provide a risk control method and system for online car-hailing orders based on acoustic event detection to solve or alleviate the technical problems existing in the prior art, namely, how to apply acoustic event features to an abnormal order detection method based on machine learning, and at least provide a beneficial option for this; the technical solution of the present invention is implemented as follows:

[0008] First, the risk control method for online car-hailing orders based on acoustic event detection:

[0009] 1. Overview:

[0010] The present invention aims to comprehensively use big data analysis and advanced machine learning technology to achieve deep mining and intelligent processing of order data on online ride-hailing platforms. Key features are extracted through feature engineering to construct order feature matrices and audio feature matrices. The trained order data risk prediction model and CRNN sound event detection model are used to predict cheating risks and analyze sound events on orders, so as to accurately identify potential abnormal orders. Combined with the model prediction results and sound event analysis results, the orders are judged to be abnormal.

[0011] (II) Technical solution:

[0012] 2.1 Step S1, data collection and preprocessing:

[0013] Order data D is obtained in real time from the database of the online car-hailing platform, including order context information a, passenger information b, driver information c and trip recording d.

[0014] The order data D is preprocessed, including data cleaning (removing invalid or erroneous data), data conversion (converting the original data into a format suitable for analysis) and data standardization (such as normalization).

[0015] 2.2 Step S2, Feature Engineering:

[0016] Perform feature extraction on the preprocessed order data D and construct the order feature matrix A;

[0017] The acoustic features of the trip recording d are extracted and concatenated into an audio feature matrix B.

[0018] 2.2.1 Step S200, construct order feature matrix A:

[0019] Including order context features a', passenger statistics features b', driver statistics features c', supply and demand features e' at the time of the order, and marketing activities d' involved in the order:

[0020]

[0021] 2.2.2 Step S201, constructing the audio feature matrix B:

[0022] Including calculating the Mel cepstral coefficient d1, zero crossing rate d2, spectral centroid d3, chromaticity spectrum d4, spectral contrast d5 and spectral bandwidth speech d6 of the recording signal:

[0023]

[0024] 2.3 Step S3, model prediction:

[0025] The constructed order feature matrix A is input into the trained order data risk prediction model to predict the probability α that the order has a cheating risk; for order i whose probability α is greater than a preset threshold (such as 80%), the audio feature matrix B of its trip recording is input into the trained CRNN sound event detection model to generate a prediction result z.

[0026] 2.3.1 Step S300, perform probability prediction:

[0027] The order data risk prediction model is a deep belief network (DBN), which includes:

[0028] a. Input layer: input order feature matrix A;

[0029] b. Each hidden layer [RBM1, RBM1, ..., RBM n ]:

[0030]

[0031] c. Hidden layer: It is a logistic regression layer in the following form:

[0032] α=σ(W out h n +b out );

[0033] Among them, h1,h2,…,h n is the output of each hidden layer, each hidden layer is an RBM (Boltzmann machine). W1,W2,…,W n and b1,b2,…,b n are the weights and biases of each RBM layer. σ is the sigmoid function. W out and b out are the weights and biases of the output layer. α is the probability that the predicted order output by the model has a risk of cheating.

[0034] The training method of the deep belief network (DBN) is as follows: pre-train each layer of RBM of DBN separately to initialize the weights. Fine-tune using the back propagation algorithm, and the trained DBN model can output the probability α of predicting the risk of cheating in an order based on the input order feature matrix A.

[0035] 2.3.2 Step S301, determine the target sound event S:

[0036] S3010, input layer:

[0037] The order feature matrix A is converted into a log-Mel spectrogram X to capture the changes in frequency and time of the audio signal.

[0038] S3011, Convolutional Neural Network (CNN) layer:

[0039] It includes 6*6 convolution blocks, each of which contains a dilated convolution, a batch normalization layer, and a ReLU activation layer.

[0040] For input X, the output Yl of the l-th convolutional block is calculated as:

[0041] Yl = ReLU(BN(DC(X;θ l )));

[0042] Among them, θ l Represents the parameters of the lth convolutional block, DC is the dilated convolution operation, BN is the batch normalization operation, and ReLU is the activation function.

[0043] S3012, recurrent neural network (RNN) layer:

[0044] 3*3 bidirectional Bi-GRU layers are used to capture both forward and backward timing information. Each Bi-GRU layer is followed by a dropout layer to reduce overfitting. For the output Yl of the convolutional neural network layer, the output H of the mth Bi-GRU layer m The calculation is as follows:

[0045] H m =D(Bi-GRU(Yl;φ m ));

[0046] Among them, φ m Represents the parameters of the mth Bi-GRU layer, Bi-GRU represents the operation of the bidirectional gated recurrent unit, and D is the dropout operation used to reduce overfitting.

[0047] S3013, Execute Attention Mechanism:

[0048] Add an attention layer after the RNN layer to weight the output features of the RNN by calculating the weight matrix. The output Z of the attention layer is calculated as follows:

[0049]

[0050] Where H is the output of the RNN layer, T is the number of time frames, and α t is the attention weight of the tth frame, which is calculated by the following activation function:

[0051]

[0052] Among them, e t is the attention score of the tth frame, calculated through a fully connected layer.

[0053] S3014, output layer:

[0054] That is, a fully connected layer (FC) followed by a classifier layer with a softmax activation function is used to predict which sound event categories exist in the audio.

[0055] For the output Z of the attention layer, the prediction result z of the output layer is calculated as follows:

[0056] z=softmax(W·Z+b);

[0057] Among them, W and b are the weight and bias of the fully connected layer respectively, and softmax is the activation function used to convert the output into a probability distribution.

[0058] 2.4 Step S4, result analysis and decision making:

[0059] According to the prediction result z, determine whether the target sound event S exists in the recording event segment to which the audio feature matrix B belongs, analyze the sound event sequence in the order itinerary recording, and determine whether the order meets the preset normal order rules. Use the following conditional statement to make a judgment:

[0060]

[0061] The above AnalyzeAndDecide process accepts three parameters: prediction result z, audio feature matrix B, and normal order rule NO. The process first assumes that there is no target sound event and traverses each sound event in the normal order rule. If the prediction result z contains a sound event S, the corresponding flag E is set to true and the loop is exited. Finally, the order status is judged according to the value of flag E. If the event exists, the order status is "normal", otherwise it is "suspicious".

[0062] (III) Technical problems to be solved:

[0063] Voice data is introduced as an important feature to determine whether an order is abnormal. In the process of identifying abnormal orders, the voice information in the trip is fully utilized. By using the CRNN (convolutional recurrent neural network) network to recognize and detect voice, the actual abnormal behavior of the order can be captured more accurately. This not only improves the accuracy of abnormal identification, but also provides a more comprehensive and accurate basis for the detection of abnormal orders.

[0064] Second, the online car-hailing order risk control system based on acoustic event detection:

[0065] like Figure 6 As shown, the system includes a processor and a memory connected to the processor, wherein the memory stores program instructions, and when the program instructions are executed by the processor, the processor executes the risk control method for online car-hailing orders as described above. At the same time, the processor is connected to:

[0066] (1) A data collection and preprocessing module for obtaining order data in real time from the database of the online ride-hailing platform. Its tasks are:

[0067] (1.1) Data cleaning: remove invalid, erroneous or redundant data.

[0068] (1.2) Data denoising: De-noising the recorded data to improve the quality of the sound signal.

[0069] (1.3) Data formatting: Convert data into a format suitable for analysis, such as a unified file format, data encoding, etc.

[0070] (2) A feature extraction module that uses signal processing technology to extract key acoustic event features from the trip recording: including Mel cepstral coefficient d1, zero-crossing rate d2, spectral centroid d3, chromaticity spectrum d4, spectral contrast d5, and spectral bandwidth voice d6. At the same time, combined with other contextual information of the order, a multi-dimensional order feature vector is constructed. Its tasks are:

[0071] (2.1) Feature selection: Select appropriate features for extraction based on business requirements and data characteristics.

[0072] (2.2) Feature conversion: Convert the original features into a format more suitable for machine learning model processing.

[0073] (3) Model module for performing prediction tasks, whose tasks are:

[0074] (3.1) Data partitioning: Divide the dataset into training set, validation set and test set.

[0075] (3.2) Model evaluation: Use the validation set to evaluate model performance and adjust the model based on the evaluation results.

[0076] (3.3) Model iteration: Continuously optimize the model based on the performance and feedback of the test set.

[0077] (4) For new order data, the trained model is used to make real-time predictions to determine whether the order has abnormal behavior. The anomaly detection module: Combine the prediction results and order characteristics to conduct in-depth analysis of abnormal orders to determine the type and cause of the anomaly. Based on the analysis results, the platform is provided with decision support, including triggering alarms, suspending services, and / or further review. Its tasks are to:

[0078] (4.1) Real-time guarantee: Ensure that the model prediction speed meets the real-time requirements.

[0079] (4.2) Accuracy assessment: Perform accuracy assessment on the prediction results to ensure the reliability of the recognition results.

[0080] (4.3) Decision rule formulation: Formulate clear decision rules based on business needs.

[0081] Compared with the prior art, the present invention has the following beneficial effects:

[0082] 1. Improve the accuracy of abnormal order identification: By introducing acoustic event features and combining other contextual information of the order, the present invention can more comprehensively analyze order data and more accurately identify abnormal orders. This helps online ride-hailing platforms to promptly detect and handle potential fraud or violations, and protect the rights and interests of passengers and drivers.

[0083] 2. Reduce labor costs: Traditional abnormal order detection methods usually require a lot of manpower, including order labeling, data analysis, etc. The solution of the present invention significantly reduces the demand for human resources, improves detection efficiency, and reduces operating costs through automated data processing and machine learning model prediction.

[0084] 3. Improve user experience: The present invention helps to improve the service quality of online ride-hailing platforms by accurately identifying and processing abnormal orders, providing users with a safer and more reliable service experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0086] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0087] Figure 2 Schematic diagram of a sound event detection method in Embodiment 1 of the present invention;

[0088] Figure 3 A schematic diagram of the CRNN network of the present invention using the RNN structure to make decisions on sound event recognition frame by frame based on context information, and finally using a fully connected layer to perform multi-classification prediction;

[0089] Figure 4 This is a schematic diagram of obtaining an audio feature matrix by analyzing and transforming an audio signal, and aligning the labeled sound event data with the width of the audio feature matrix by data stretching in the first embodiment of the present invention;

[0090] Figure 5 This is a schematic diagram of the overall process of identifying abnormal orders in the first embodiment of the present invention;

[0091] Figure 6 It is a schematic diagram of the system composition of the present invention. DETAILED DESCRIPTION

[0092] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below;

[0093] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0094] Explanation of relevant terms:

[0095] (1) Tenant: the online ride-hailing company to which the driver belongs;

[0096] (2) Aggregation platform: an online ride-hailing platform composed of multiple tenants;

[0097] (3) Order characteristics: some characteristic data of the order during the order placement process, such as the pick-up / drop-off distance, pick-up / drop-off time, estimated price, the time the passenger waits for a response after placing the order, the time the driver waits for the passenger after arriving at the destination, etc.

[0098] (4) Ordering process: The ordering process of online ride-hailing orders is divided into the following steps: passenger places order → order response / driver accepts order → driver arrives at the order starting point → passenger gets on the car → starts the trip → arrives at the order destination → driver confirms the fare → passenger pays;

[0099] (5) Transaction system: The overall process of acquiring online ride-hailing orders, matching drivers and passengers, navigation, order pricing, post-order charging, and commission collection in online ride-hailing aggregation platforms;

[0100] (6) Risky orders on the transaction line: fake orders completed by drivers using technical or regulatory means in order to improve system scores, good reviews, or rewards, etc.

[0101] (7) Driver App: software developed by the platform and used by drivers to receive / grab online ride-hailing orders, record trip information, and calculate fares;

[0102] (8) Trip recording: For safety reasons, the driver app will start recording the trip in the background from the moment the driver receives the online ride-hailing order until the driver drops the passenger off at the destination.

[0103] (9) Sound event: an event that produces sound, such as people talking, rain, thunder, or doors opening and closing;

[0104] (10) Sound event detection algorithm: The purpose is to identify the type of sound event in an audio.

[0105] and the start and end points of the sound events in the recording;

[0106] (11) Mel Frequency Cepstral Coefficients (MFCC):

[0107] A feature extraction technique commonly used in speech signal processing;

[0108] (12) Zero Crossing Rate (ZCR): A basic characteristic in audio signal processing, which refers to the number of times a signal switches from a positive half cycle to a negative half cycle or from a negative half cycle to a positive half cycle within a period of time, that is, the number of times the signal crosses the zero axis;

[0109] (13) Spectral Centroid: A feature that describes the center position of the spectral distribution of an audio signal. It is obtained by calculating the weighted average of the spectral density of the signal, where the weight is a function of frequency.

[0110] (14) Chroma Spectral Feature: A method used to describe the spectral characteristics of an audio signal. Based on the concept of musical scale in music theory, the spectrum of an audio signal is divided into 12 chromas, each of which corresponds to a pitch.

[0111] (15) Spectral Contrast: A statistic used to describe the spectral characteristics of an audio signal, reflecting the contrast or difference between different frequency components in the signal spectrum.

[0112] (16) Spectral Bandwidth: A parameter that describes the spectral characteristics of an audio signal, indicating the breadth or range of energy distribution in the signal spectrum;

[0113] (17) Surcharges: parking fees before driving, highway tolls during driving, and other additional fees that passengers need to pay;

[0114] (18) Listening for orders: The online taxi driver is online and available to accept orders, idle, and waiting for the system to assign or play orders;

[0115] (19) Main order: On the aggregation platform, the online ride-hailing order placed by the passenger is the main order, which includes the passenger's starting point and destination;

[0116] (20) Sub-orders: In the aggregation platform, passengers will select multiple tenants when placing an order. After placing the order, the system will send an order to each tenant to match the drivers in the tenant. These orders are called sub-orders.

[0117] (21) Response: After a passenger places an order, the system matches the order with a driver and notifies the passenger and driver of the matching result;

[0118] (22) System dispatch: After the online ride-hailing system matches the online ride-hailing order with a driver, it dispatches the order to the driver and notifies the driver to pick up the passenger;

[0119] (23) Supply-demand ratio: the ratio of the number of online ride-hailing drivers available to take orders to the number of passengers currently trying to hail a ride.

[0120] Embodiment 1: Figures 1 to 5 As shown, this embodiment will provide a specific implementation method of the risk control method for online car-hailing orders based on acoustic event detection;

[0121] In this embodiment, regarding step S1, data collection and preprocessing:

[0122] Obtain order data D from the database of the online car-hailing platform in real time, including:

[0123] (1) Order context information a (such as order time, starting point, end point, etc.);

[0124] (2) Passenger information b (such as passenger ID, number of historical orders, etc.);

[0125] (3) Driver information c (such as driver ID, number of historical orders, rating, etc.);

[0126] (4) Itinerary recording (audio file);

[0127] Further, perform data preprocessing:

[0128] S100, data cleaning: removing invalid or erroneous order records.

[0129] S101, Data transformation: Convert datetimes to timestamps and encode categorical variables to numeric values.

[0130] S102, data standardization: normalize the numerical features.

[0131] In this embodiment, regarding step S2, feature engineering: feature extraction is performed on the preprocessed order data D to construct an order feature matrix A; acoustic feature extraction is performed on the trip recording d and they are spliced ​​into an audio feature matrix B.

[0132] Specifically, in step S200, an order feature matrix A is constructed:

[0133] S2000, order context feature a': including order timestamp and travel distance;

[0134] S2001, passenger statistical characteristics b': including the number of historical orders of passengers and the average order amount;

[0135] S2002, driver statistical characteristics c': including the driver's historical order count and average rating;

[0136] S2003, supply and demand characteristics e': including the supply and demand ratio when the order is issued;

[0137] S2004, marketing activity characteristics d': including whether the order participates in promotion activities;

[0138]

[0139] Specifically, step S201 constructs an audio feature matrix B: including calculating the Mel-frequency cepstral coefficient d1, zero-crossing rate d2, spectral centroid d3, chromaticity spectrum d4, spectral contrast d5 and spectral bandwidth speech d6 of the recording signal:

[0140]

[0141] In this embodiment, regarding step S3, model prediction: the constructed order feature matrix A is input into the trained order data risk prediction model to predict the probability α that the order has a cheating risk; for order i whose probability α is greater than a preset threshold (such as 80%), the audio feature matrix B of its itinerary recording is input into the trained CRNN sound event detection model to generate a prediction result z.

[0142] Specifically, step S300 performs probability prediction: the order data risk prediction model is a deep belief network (DBN), including:

[0143] a. Input layer: input order feature matrix A;

[0144] b. Each hidden layer [RBM1, RBM1, ..., RBM n ]:

[0145]

[0146] c. Hidden layer: It is a logistic regression layer in the following form:

[0147] α=σ(W out h n +b out );

[0148] Among them, h1,h2,…,h n is the output of each hidden layer, each hidden layer is an RBM (Boltzmann machine). W1,W2,…,W n and b1,b2,…,b n are the weights and biases of each RBM layer. σ is the sigmoid function. W out and b out are the weights and biases of the output layer. α is the probability that the predicted order output by the model has a risk of cheating.

[0149] It should be pointed out that the training method of the deep belief network (DBN) is to pre-train each layer of RBM of DBN separately to initialize the weights. The back propagation algorithm is used for fine-tuning. The trained DBN model can output the probability α of predicting the risk of cheating in an order based on the input order feature matrix A.

[0150] Specifically, step S301, determining the target sound event S, includes:

[0151] S3010, input layer: convert the order feature matrix A into a log-Mel spectrogram X to capture the changes in frequency and time of the audio signal.

[0152] S3011, Convolutional Neural Network (CNN) layer: includes 6*6 convolution blocks, each of which contains dilated convolution, batch normalization layer and ReLU activation layer. For input X, the output Yl of the lth convolution block is calculated as:

[0153] Yl = ReLU(BN(DC(X;θ l )));

[0154] Among them, θ l Represents the parameters of the lth convolutional block, DC is the dilated convolution operation, BN is the batch normalization operation, and ReLU is the activation function.

[0155] S3012, recurrent neural network (RNN) layer: 3*3 bidirectional Bi-GRU layer for capturing both forward and backward timing information. Each Bi-GRU layer is followed by a dropout layer to reduce overfitting. For the output Yl of the convolutional neural network layer, the output H of the mth Bi-GRU layer m The calculation is as follows:

[0156] H m =D(Bi-GRU(Yl;φ m ));

[0157] Among them, φ m Represents the parameters of the mth Bi-GRU layer, Bi-GRU represents the operation of the bidirectional gated recurrent unit, and D is the dropout operation used to reduce overfitting.

[0158] Specifically, S3013, execute the attention mechanism: add an attention layer after the RNN layer, and weight the output features of the RNN by calculating the weight matrix. The output Z of the attention layer is calculated as follows:

[0159]

[0160] Where H is the output of the RNN layer, T is the number of time frames, and α t is the attention weight of the tth frame, which is calculated by the following activation function:

[0161]

[0162] Among them, e t is the attention score of the tth frame, calculated through a fully connected layer.

[0163] Specifically, S3014, output layer:

[0164] That is, the fully connected layer (FC) is followed by a classifier layer with a softmax activation function, which is used to predict which sound event categories exist in the audio. For the output Z of the attention layer, the prediction result z of the output layer is calculated as follows: z = softmax(W·Z+b);

[0165] Among them, W and b are the weight and bias of the fully connected layer respectively, and softmax is the activation function used to convert the output into a probability distribution.

[0166] In this embodiment, regarding step S4, result analysis and decision making:

[0167] According to the prediction result z, whether the target sound event S exists in the recording event segment to which the audio feature matrix B belongs is obtained, and the sound event sequence in the order itinerary recording is analyzed to determine whether the order complies with the preset normal order rules. The specific rule needs to accept three parameters: the prediction result z, the audio feature matrix B, and the normal order rule NO. The process first assumes that there is no target sound event and traverses each sound event in the normal order rule. If a sound event S is included in the prediction result z, the corresponding flag E is set to true and the loop is exited. Finally, the order status is determined according to the value of the flag E. If the event exists, the order status is "normal", otherwise it is "suspicious".

[0168] For example, there is an order. After the DBN model predicts, the cheating risk probability α is 85%, which exceeds the threshold, so it enters the CRNN sound event detection model. The model prediction result z shows that the recording contains the sound event of "driver and passenger talking", which is an expected event in the normal order rules. Therefore, after the above process, the order is judged as "normal".

[0169] In this embodiment, the Python execution framework of the above solution is as follows:

[0170]

[0171]

[0172]

[0173]

[0174] In the above program: The DBN model is a deep neural network used to predict the probability of cheating risk in an order. It contains multiple hidden layers, and the last layer is a logistic regression layer that outputs the probability of cheating risk.

[0175] The CRNN model is a convolutional recurrent neural network for detecting sound events in audio. It contains convolutional layers to extract features, recurrent layers to capture timing information, and finally fully connected layers for classification.

[0176] Based on the prediction results of the DBN model, suspicious orders are identified. The CRNN model is used to detect sound events for suspicious orders. Further analysis or decision-making is carried out based on the detection results.

[0177] Embodiment 2: This embodiment will further provide a training method for the CRNN sound event detection model described in Embodiment 1:

[0178] The first step is data preparation and preprocessing:

[0179] (1) Data collection: Order recording data within 30 days must be obtained. This data covers sound records in various actual scenarios.

[0180] (2) Data labeling: 200 audio recordings were randomly selected from the collected audio recordings for manual labeling. The labeling process recorded in detail the specific locations of the five sound events, namely, [wind noise during driving], [navigation voice], [driver getting on and off the vehicle voice], [driver-passenger conversation], and [door opening and closing sound] in each recording.

[0181] (3) Feature Engineering: Feature extraction of recorded data, including calculation of Mel-frequency cepstral coefficients, zero-crossing rate, spectral centroid, chromaticity spectrum, spectral contrast, spectral bandwidth and other dimensions, to fully capture the characteristics of the sound signal.

[0182] The second step is to construct the feature matrix:

[0183] (1) Feature splicing: Align the extracted features in the time frame rate and splice them into an audio feature matrix with a dimension of (n*time frame rate).

[0184] (2) Annotation result mapping: The audio annotation result is stretched and transformed into a result matrix of (5*time frame rate) so as to correspond to the audio feature matrix.

[0185] (3) Matrix segmentation: Align the audio feature matrix and the result matrix with the time frame rate as the axis, and perform equidistant segmentation to obtain matrix data with dimensions of (n100) and (5100) as input for model training.

[0186] The third step is CRNN model construction and training:

[0187] (1) Network design: A CRNN network as described in Example 1 is constructed, which combines convolutional layers, recurrent layers, and fully connected layers to capture the spatiotemporal features in the sound signal.

[0188] (2) Model training: Use the segmented data to train a multi-classification model so that the model can identify whether the five target sound events exist in each slice data.

[0189] Step 4: Order data risk prediction model training:

[0190] (1) Data collection and feature engineering: Obtain order data within 180 days and perform feature engineering on the order data to construct multi-dimensional features including order context features, passenger statistics features, driver statistics features, supply and demand features, and marketing campaign features.

[0191] (2) Model training: Extract some order data for labeling, combine the labeling results with the features counted by feature engineering, and train the order data risk prediction model.

[0192] Step 5: Data prediction and abnormal order identification:

[0193] (1) Real-time prediction: Perform feature engineering on order data after the order is completed, and combine it with the order data risk prediction model to predict the probability of fraud risk in real time.

[0194] (2) Sound event detection: Feature engineering is performed on order trip recordings with a predicted fraud risk probability higher than 80%, and a sound event detection model is used to determine whether each target sound event exists in the recording event segment to which each feature matrix belongs.

[0195] All the above embodiments only express the implementation methods of the relevant practical applications of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

[0196] For those skilled in the art, it can be further appreciated that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0197] At the same time, those skilled in the art can understand that all or part of the processes in all the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

Claims

1. A risk control method for online car-hailing orders based on acoustic event detection, characterized in that: The implementation steps include: S1, obtain order data D, including order context information a, passenger information b, driver information c and trip recording d; S2, extract features from the preprocessed order data D and construct an order feature matrix A; Extract acoustic features from the trip recording d and concatenate them into an audio feature matrix B; S3, input the constructed order feature matrix A into the trained order data risk prediction model to predict the probability α that the order has a cheating risk; For order i with probability α greater than the preset threshold, the audio feature matrix B of its trip recording is input into the trained CRNN sound event detection model to generate the prediction result z; S4, based on the prediction result z, obtain whether there is a target sound event S in the recording event segment to which the audio feature matrix B belongs, analyze the sound event sequence in the order itinerary recording, and determine whether the order meets the preset normal order rules.

2. The risk control method for online car-hailing orders according to claim 1, characterized in that: In S2, the order feature matrix A includes order context features a', passenger statistical features b', driver statistical features c', supply and demand features e' at the time of the order, and marketing activities d' involved in the order: The audio feature matrix B includes the Mel cepstral coefficient d1, zero crossing rate d2, spectral centroid d3, chromaticity spectrum d4, spectral contrast d5 and spectral bandwidth voice d6 of the calculated recording signal:

3. The risk control method for online car-hailing orders according to claim 1, characterized in that: In S3, the order data risk prediction model is a deep belief network, including: a. Input layer: input order feature matrix A; b. Each hidden layer [RBM1, RBM1, ..., RBM n ]: c. Hidden layer: It is a logistic regression layer in the following form: α=σ(W out h n +b out ); Among them, h1,h2,…,h n is the output of each hidden layer, each hidden layer is an RBM; W1,W2,…,W n and b1,b2,…,b n are the weights and biases of each RBM layer; σ is the sigmoid function; W out and b out are the weights and biases of the output layer.

4. The online car-hailing order risk control method according to claim 1 or 3, characterized in that: In S3, the CRNN sound event detection model includes: Input layer: Convert the order feature matrix A into the logarithmic Mel spectrum graph X; Convolutional Neural Network Layer: For input X, the lth convolutional block calculates: Yl6ReLU(BN(DC(X)θ l ))) Among them, θ l Represents the parameters of the lth convolutional block, DC is the dilated convolution operation, BN is the batch normalization operation, and ReLU is the activation function; Recurrent Neural Network Layer: For the output Yl of the convolutional neural network layer, the output H of the mth Bi-GRU layer m The calculation is as follows: H m =D(Bi-GRU(Yl;φ m )); Among them, φ m represents the parameters of the mth Bi-GRU layer, Bi-GRU represents the operation of the bidirectional gated recurrent unit, and D is the dropout operation; Output layer: It is a classifier layer with a fully connected layer followed by a softmax activation function, which is used to predict which sound event categories exist in the audio. For the output Z of the attention layer, the prediction result z of the output layer is calculated as follows: z=softmax(W·Z+b); Among them, W and b are the weight and bias of the fully connected layer respectively, and softmax is the activation function used to convert the output into a probability distribution.

5. The risk control method for online car-hailing orders according to claim 4, characterized in that: In the convolutional neural network layer, a 6*6 convolution block is included, each block contains a hole convolution batch normalization layer and a ReLU activation layer.

6. The risk control method for online car-hailing orders according to claim 4, characterized in that: In the recurrent neural network layer, a 3*3 bidirectional Bi-GRU layer is included for simultaneously capturing forward and backward timing information; each Bi-GRU layer is followed by a dropout layer to reduce overfitting.

7. The risk control method for online car-hailing orders according to claim 4, characterized in that: After the recurrent neural network layer, an attention layer is also included to weight the output features of the RNN by calculating the weight matrix; the output Z of the attention layer is calculated as follows: Where H is the output of the RNN layer, T is the number of time frames, and α t is the attention weight of the tth frame, which is calculated by the following activation function: Among them, e t is the attention score of the tth frame, calculated through a fully connected layer.

8. The risk control method for online car-hailing orders according to claim 1, characterized in that: The determination method in S4 is as follows: First, assume that there is no target sound event, and traverse each sound event in the normal order rule; if the prediction result z contains a sound event S, set the corresponding flag E to true and jump out of the loop; finally, judge the order status according to the value of flag E. If the event exists, the order status is "normal", otherwise it is "suspicious".

9. The risk control system for online car-hailing orders based on acoustic event detection is characterized by: The system includes a processor and a memory connected to the processor, wherein program instructions are stored in the memory. When the program instructions are executed by the processor, the processor executes the online car-hailing order risk control method as described in any one of claims 1-8.

10. The online car-hailing order risk control system based on acoustic event detection according to claim 9, characterized in that: A data collection and preprocessing module for obtaining order data in real time from the database of the online car-hailing platform; A feature extraction module that uses signal processing techniques to extract key acoustic event features from trip recordings; Model modules for performing prediction tasks; An anomaly detection module that uses the trained model to make real-time predictions and determine whether there are abnormal behaviors in the order.