Electricity fee collection system and method based on voice interaction

Abstract

The invention discloses an electricity fee collection system and method based on voice interaction, which belongs to the technical field of electricity fee collection, and solves the problem that when the existing method divides users into different electricity fee levels, a single electricity fee consumption index is not considered to be unable to fully reflect the payment credit status of the user, resulting in poor adaptability and robustness of the existing electricity fee collection system. The method comprises obtaining user electricity consumption related information, evaluating the user credit level based on the TOPSIS algorithm improved by K-means clustering, identifying the risk of voice interaction information and user credit level by a recovery risk assessment model, and outputting an electricity fee recovery assessment result; in the invention, K-means clustering combined with the TOPSIS algorithm can comprehensively consider multiple feature indicators to evaluate the user credit level, and at the same time, voice interaction information and user credit level are considered in the electricity fee recovery assessment, thereby providing a more accurate electricity fee recovery assessment result for electricity fee collection.

Description

Technical Field

[0001] The present invention belongs to the technical field of electricity bill collection, and in particular relates to an electricity bill collection system and method based on voice interaction. Background Art

[0002] The collection of electricity bills is a key assessment indicator for the power marketing department. The electricity bill business includes five major links: meter reading, accounting, collection, collection and accounting. The collection management of electricity bills refers to the process by which the power company ensures that users pay their electricity bills on time through a series of measures and processes. It not only involves the timely recovery of financial income, but also relates to the stability of power supply and the rational allocation of social resources.

[0003] Chinese patent CN118335055A discloses an electricity bill collection system and method based on intelligent voice, including collecting user information and obtaining user electricity bill information; users input personal information and electricity bill account information through voice to establish a user profile model; users are divided into different electricity bill levels according to their electricity bill consumption levels, and the number of electricity bill collection reminders is set; the user's electricity bill collection cycle is analyzed in combination with the user's electricity consumption and historical payment records, and the correction amount of the number of electricity bill collection reminders is obtained, while the change of the correction amount of the number of electricity bill collection reminders is monitored, and electricity bill collection reminders are automatically sent; based on the obtained user payment information, users with abnormal status are monitored through deep learning, but when the existing method divides users into different electricity bill levels according to their electricity bill consumption levels, it does not take into account that a single electricity bill consumption indicator cannot fully reflect the user's payment credit status, which makes the existing electricity bill collection system poor in adaptability and robustness, and cannot provide a more accurate and comprehensive evaluation of user electricity bill collection. In response to the above problems, we propose an electricity bill collection system and method based on voice interaction. Summary of the invention

[0004] The purpose of the present invention is to address the shortcomings of the prior art and provide an electricity bill collection system and method based on voice interaction, which solves the problem that the existing method divides users into different electricity bill levels according to their electricity bill consumption levels, but does not take into account that a single electricity bill consumption indicator cannot fully reflect the user's payment credit status, resulting in poor adaptability and robustness of the existing electricity bill collection system, and is unable to provide a more accurate and comprehensive assessment of user electricity bill collection.

[0005] The present invention is implemented as follows: a method for collecting electricity bills based on voice interaction, the method for collecting electricity bills based on voice interaction comprising:

[0006] Acquire user electricity usage-related information, normalize the user electricity usage-related information, and obtain a normalized set, wherein the user electricity usage-related information includes user electricity usage fee information, payment behavior information, historical arrears information, electricity usage type information, and user profile information;

[0007] Load the normalized set, evaluate the user's credit rating based on the TOPSIS algorithm improved by K-means clustering, and upload the user's credit rating to the blockchain platform;

[0008] Load the user's credit rating. The blockchain platform generates a voice interaction strategy based on the improved Tacotron2 model and the user's credit rating.

[0009] In response to the voice interaction strategy, voice interaction information is obtained, and the voice interaction information and the user's credit level are used as inputs. The voice interaction information and the user's credit level risk are identified based on a pre-built recovery risk assessment model, and an electricity fee recovery assessment result is output.

[0010] Preferably, the method for normalizing the user's electricity consumption related information specifically includes:

[0011] Load user electricity usage related information and process missing values ​​and abnormal values ​​of user electricity usage related information;

[0012] Obtain the user electricity consumption related information after the missing values ​​and abnormal values ​​are processed, and determine the benchmark threshold and limit value of the characteristic indicators in the user electricity consumption related information based on the hierarchical analysis method;

[0013] The reference threshold and limit value of the characteristic index are calculated by the following formula:

[0014] (1)

[0015] (2)

[0016] (3)

[0017] in, Characteristic index The reference threshold and limit value, represents the input value of the characteristic index, represents the input mean of the feature index, are the maximum and minimum values ​​of the characteristic index, respectively. Indicates the threshold correction value, Represents the characteristic index determined based on the hierarchical analysis method The subjective weight of Respectively represent the number of feature indicators and the weight ranking of feature indicators, They are the maximum eigenvalue of characteristic indicators and the correlation coefficient of indicators based on the hierarchical analysis method;

[0018] Loading the benchmark threshold and limit value of the characteristic index, normalizing the user's electricity consumption related information based on the benchmark threshold and limit value of the characteristic index, and outputting a normalized set;

[0019] When normalizing the user's electricity consumption-related information based on the benchmark threshold and limit value of the characteristic index, the characteristic index output is kept between 0 and 1 through the normalization operation, and the normalization operation is performed through the following expression:

[0020] (4)

[0021] in, represents the normalized value of the feature index, Characteristic index The reference threshold and limit value.

[0022] Preferably, the method for evaluating user credit rating based on the TOPSIS algorithm improved by K-means clustering specifically includes:

[0023] Load the normalized set, calculate the characteristic indicator output correlation value based on the Gumbel-Copula function, and use the characteristic indicator output correlation value to describe the correlation between the characteristic indicator and the user's credit rating output;

[0024] Among them, the characteristic index output correlation value is calculated by the following expression:

[0025] (5)

[0026] In the formula, Indicates the output correlation value of characteristic index, is the estimated degrees of freedom parameter of the Gumbel-Copula function, Characteristic index and The marginal distribution function of

[0027] Obtain the output correlation value of the characteristic indicator, objectively weight the characteristic indicator based on the indicator correlation weight determination method, and obtain the objective weight of the characteristic indicator;

[0028] The objective weight of the characteristic indicator is calculated by the following formula:

[0029] (6)

[0030] In the formula, Indicates the output correlation value of characteristic index, represents the objective weight of the feature index, are the normalized value and standard deviation of the characteristic index respectively;

[0031] Load the objective weight of the characteristic index, and calculate the optimal weight of the index based on the coupling of the subjective weight and objective weight of the characteristic index by Moral game;

[0032] Among them, the optimal weight calculation method of the indicator is:

[0033] (7)

[0034] in, represents the optimal weight of the indicator, is the optimal value of the indicator based on Moral game, ;

[0035] Based on the K-means clustering algorithm, the user payment similarity matrix is ​​constructed and the indicator center clustering point is determined. The user payment similarity matrix is ​​multiplied by the optimal weight of the indicator to obtain a weighted standardized matrix.

[0036] A weighted normalized matrix is ​​obtained, and the TOPSIS algorithm determines positive and negative ideal solutions based on the weighted normalized matrix;

[0037] Load the positive and negative ideal solutions, calculate the Euclidean distance between the indicator center cluster point and the positive and negative ideal solutions, determine the progress of the posting based on the Euclidean distance between the indicator center cluster point and the positive and negative ideal solutions, and use the progress of the posting as the output of the user's credit level.

[0038] Preferably, the improved Tacotron2 model includes a word embedding module, a spectrum generation network, and a vocoder, wherein the word embedding module includes a hierarchical input layer, a text capture layer, and a text fusion layer, and asymmetric convolution is introduced in the text fusion layer. The asymmetric convolution includes three groups of convolution kernels, and the sizes of the convolution kernels are 1×3, 3×1, and 3×3, respectively. The spectrum generation network consists of a speech encoder and a speech decoder, wherein the speech encoder includes a bidirectional LSTM layer and a HSFs layer. The bidirectional LSTM layer is used to identify the text fusion result and generate the final hidden feature representation based on the text fusion result. The HSFs layer is used to express the overall speech characteristics of the text fusion result. The speech decoder consists of two groups of multi-layer attention perception mechanisms, two fully connected layers, a linear mapping layer, and a BP neural network fusion device. The BP neural network fusion device is used to output the Melton spectrum. The multi-layer attention perception mechanism and the fully connected layer correspond to the bidirectional LSTM layer and the HSFs layer, respectively. The Levenberg-Marquard algorithm is used to train the spectrum generation network during training of the improved Tacotron2 model.

[0039] Preferably, the blockchain platform generates a method for generating a voice interaction strategy based on an improved Tacotron2 model combined with a user's credit rating, specifically including:

[0040] Obtain the user's credit rating and improve the word embedding module in the Tacotron2 model to capture the corresponding collection text information, collection frequency, and collection priority in the blockchain platform based on the user's credit rating;

[0041] The text fusion layer fuses the features of collection text information, collection frequency, and collection priority based on asymmetric convolution, and outputs the text fusion result;

[0042] Obtain the text fusion result, the bidirectional LSTM layer recognizes the text fusion result, generates the final hidden feature representation based on the text fusion result, and the HSFs layer synchronously expresses the overall speech characteristics of the text fusion result;

[0043] Load the text fusion results to generate the final hidden layer feature representation and the overall speech characteristic expression of the text fusion results, align the text sequence and speech frame based on the multi-layer attention perception mechanism, and output the Melton spectrum;

[0044] The mel spectrum is converted into an audio waveform based on the vocoder to generate a voice interaction strategy.

[0045] Preferably, when the recycling risk assessment model is pre-constructed, a convolutional neural network is used as the initial model. The initial model consists of an input layer, a convolution module, a pooling layer and an output layer. The convolution module is improved by introducing a preprocessing layer Conv1 and a residual convolution network Conv2_x into the convolution module, freezing the pooling layer, and replacing the pooling layer with a feature fusion layer. The feature fusion layer consists of an SPPF module and a compression excitation module. The number of channels of the SPPF module is 256 and the parameter is 5. It is used to extract the convolution fusion result of the convolution fusion network Conv2_x input. The compression excitation module is used to infer the feature fusion result of the SPPF module input. The user's credit rating is used as prior information to determine the electricity fee recovery assessment value. The activation function of the convolution module is the Sigmoid activation function, and the activation function of the feature fusion layer is the SiLU activation function.

[0046] Preferably, the method for identifying voice interaction information and user credit rating risks based on a pre-built recovery risk assessment model specifically includes:

[0047] Load the voice interaction information and user credit rating. The input layer performs discrete Fourier transform on the voice interaction information based on framing and windowing to obtain the Fourier transform set corresponding to the voice interaction information.

[0048] Acquire a Fourier transform set, perform equivalent weighted filtering on the Fourier transform set based on an equivalent weighted filter, filter out interfering harmonics in the Fourier transform set, output a weighted filtering result, perform discrete cosine transform on the weighted filtering result, and obtain a static speech signal;

[0049] After the equivalent weighted filter performs equivalent weighted filtering on the Fourier transform set, the weighted filtering result is expressed as:

[0050] (8)

[0051] (9)

[0052] (10)

[0053] in, represents the weighted filtering result, , are the sampling angle differences of the equivalent weighted filter for the positive and negative frequencies of the harmonics, is the amplitude of the interfering harmonic, is the filtering times of equivalent weighted filter, represents the order of the equivalent weighted filter, is the input representation of the Fourier transform set, , are the sampling frequency and fundamental frequency of the Fourier transform set, respectively. is the phase of the interfering harmonic, represents the frequency-shifted signal in Fourier transform concentration;

[0054] Load the static speech signal, the preprocessing layer Conv1 converts the static speech signal into a static speech vector, the residual convolution network Conv2_x performs convolution fusion on the static speech vector, and outputs the convolution fusion result;

[0055] The convolution fusion result is obtained. The SPPF module extracts the convolution fusion result of the convolution fusion network Conv2_x input. The compression incentive module uses the user's credit level as prior information and determines the electricity fee recovery assessment value based on the convolution fusion result of the convolution fusion network Conv2_x input extracted by the SPPF module.

[0056] The electricity recovery assessment value is calculated using the following formula:

[0057] (11)

[0058] in, To recover the assessed value of electricity bills, is the user's credit rating, Extract the convolution fusion result of the convolution fusion network Conv2_x input for the SPPF module. is the bias vector of the SiLU activation function, Represents the SiLU activation function.

[0059] On the other hand, the present invention also provides an electricity fee collection system based on voice interaction, and the electricity fee collection system based on voice interaction specifically includes:

[0060] An information acquisition module is used to acquire user electricity usage related information, and normalize the user electricity usage related information to obtain a normalized set;

[0061] The rating evaluation module loads the normalized set, evaluates the user's credit rating based on the TOPSIS algorithm improved by K-means clustering, and uploads the user's credit rating to the blockchain platform;

[0062] The blockchain platform is used to load the user's credit rating. The blockchain platform generates a voice interaction strategy based on the improved Tacotron2 model and the user's credit rating;

[0063] The recycling assessment module responds to the voice interaction strategy, obtains voice interaction information, takes the voice interaction information and the user's credit level as input, identifies the voice interaction information and the user's credit level risk based on a pre-built recycling risk assessment model, and outputs the electricity fee recycling assessment result.

[0064] Preferably, the information acquisition module includes:

[0065] A preprocessing unit, which loads the user's electricity usage related information and processes the missing values ​​and abnormal values ​​of the user's electricity usage related information;

[0066] An index analysis unit, used to obtain the user electricity consumption related information after the missing values ​​and abnormal values ​​are processed, and determine the reference threshold and limit value of the characteristic index in the user electricity consumption related information based on the hierarchical analysis method;

[0067] The normalization unit loads the reference threshold and limit value of the characteristic index, performs normalization operation on the user's electricity consumption related information based on the reference threshold and limit value of the characteristic index, and outputs a normalized set.

[0068] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0069] In the embodiment of the present invention, the TOPSIS algorithm improved based on K-means clustering is used to evaluate the user credit level, so that the K-means clustering combined with the TOPSIS algorithm can comprehensively consider multiple feature indicators to evaluate the user credit level, and at the same time, the voice interaction information and the user credit level are considered in the electricity bill recovery evaluation, so as to provide a more accurate electricity bill recovery evaluation result for electricity bill collection, reduce the electricity bill collection cost and improve the intelligence level of electricity bill collection, and overcome the problem that when the existing method divides users into different electricity bill levels according to the user's electricity bill consumption level, a single electricity bill consumption indicator is not considered to be unable to fully reflect the user's payment credit status, so that the existing electricity bill collection system has poor adaptability and robustness, and cannot provide a more accurate and comprehensive evaluation of the user's electricity bill collection.

[0070] In an embodiment of the present invention, the user electricity consumption related information is normalized so as to convert the user electricity consumption related information from different sources and formats into a unified standard form. Based on the hierarchical analysis method, the relative importance of each characteristic indicator in the decision-making can be clarified, thereby providing support for determining the benchmark threshold and limit value to facilitate subsequent analysis and processing.

[0071] In the embodiment of the present invention, based on the coupling of subjective weights and objective weights of characteristic indicators by Moral game, by combining subjective weights and objective weights, the experience of decision makers and the analysis results of objective data can be more comprehensively considered, thereby improving the quality of decision-making. At the same time, the user payment similarity matrix is ​​multiplied by the optimal weight of the indicator to obtain a weighted standardized matrix, and the improvement of the TOPSIS algorithm is completed. Then, the positive and negative ideal solutions are determined, and the distance between the evaluation object and the positive and negative ideal solutions is calculated by Euclidean distance. The progress is used as the output of the user's credit level, so that the improved TOPSIS algorithm can comprehensively consider multiple indicators, so that the assessment of the user's telecommunications credit level is more comprehensive, and the deviation caused by human subjective judgment is reduced.

[0072] In an embodiment of the present invention, a voice interaction strategy is generated based on an improved Tacotron2 model combined with a user credit rating. The improved Tacotron2 model consists of a word embedding module, a spectrum generation network, and a vocoder, and the spectrum generation network consists of a speech encoder and a speech decoder. A bidirectional LSTM layer and an HSFs layer are introduced into the speech encoder, so that the overall speech characteristics of the text fusion result can be expressed, the audio synthesis quality and speed are guaranteed, and more personalized services can be provided at the same time, so that the collection voice is more in line with the actual situation and needs of the user, and the frequency and intensity of the collection are reduced, thereby reducing interference to the user and improving user satisfaction.

[0073] In an embodiment of the present invention, a recycling risk assessment model is provided. The recycling risk assessment model uses a convolutional neural network as an initial model, introduces a preprocessing layer Conv1 and a residual convolutional network Conv2_x, and also introduces a feature fusion layer composed of an SPPF module and a compression excitation module. The combination of the SPPF module and the compression excitation module increases the adaptability of the recycling risk assessment model to different types of features and improves the generalization ability of the recycling risk assessment model, so that the recycling risk assessment model can use the user's credit level as prior information and combine voice interaction information to ensure the accuracy of the electricity bill recovery assessment value. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a schematic diagram of the implementation flow of the method for collecting electricity bills based on voice interaction provided by the present invention.

[0075] Figure 2 The present invention shows a schematic diagram of the implementation process of the method for normalizing user electricity consumption related information.

[0076] Figure 3 The figure shows a schematic diagram of the implementation process of the method for evaluating user credit rating based on the TOPSIS algorithm improved by K-means clustering.

[0077] Figure 4 An architectural diagram of the improved Tacotron2 model is shown.

[0078] Figure 5 The figure shows a flowchart of the implementation of a method for generating a voice interaction strategy based on the blockchain platform based on the improved Tacotron2 model and the user's credit rating.

[0079] Figure 6 The present invention shows a schematic diagram of the implementation process of the method for identifying voice interaction information and user credit rating risks based on a pre-built recovery risk assessment model.

[0080] Figure 7 It is a structural schematic diagram of the electricity bill collection system based on voice interaction provided by the present invention. DETAILED DESCRIPTION

[0081] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of this application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0082] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0083] When existing methods divide users into different electricity fee levels according to their electricity consumption levels, they do not consider that a single electricity consumption indicator cannot fully reflect the user's payment credit status, which makes the existing electricity fee collection system have poor adaptability and robustness, and cannot provide a more accurate and comprehensive assessment of user electricity fee collection. To address the above problems, we propose an electricity fee collection system and method based on voice interaction. In short, when the method is implemented, the user's electricity consumption related information is first obtained, the user's electricity consumption related information is normalized to obtain a normalized set, and then the user's credit level is evaluated based on the TOPSIS algorithm improved by K-means clustering, and the user's credit level is uploaded to the blockchain platform 300. The blockchain platform 300 generates a voice interaction strategy based on the improved Tacotron2 model combined with the user's credit level. Finally, the voice interaction information and the user's credit level are used as input, and the voice interaction information and the user's credit level risk are identified based on the pre-built recovery risk assessment model. In the embodiment of the present invention, the TOPSIS algorithm improved based on K-means clustering is used to evaluate the user credit level, so that the K-means clustering combined with the TOPSIS algorithm can comprehensively consider multiple feature indicators to evaluate the user credit level, and at the same time, the voice interaction information and the user credit level are considered in the electricity bill recovery evaluation, so as to provide a more accurate electricity bill recovery evaluation result for electricity bill collection, reduce the electricity bill collection cost and improve the intelligence level of electricity bill collection, and overcome the problem that when the existing method divides users into different electricity bill levels according to the user's electricity bill consumption level, a single electricity bill consumption indicator is not considered to be unable to fully reflect the user's payment credit status, so that the existing electricity bill collection system has poor adaptability and robustness, and cannot provide a more accurate and comprehensive evaluation of the user's electricity bill collection.

[0084] The embodiment of the present invention provides a method for collecting electricity bills based on voice interaction. Figure 1 The present invention shows a schematic diagram of the implementation process of the method for collecting electricity bills based on voice interaction, wherein the method for collecting electricity bills based on voice interaction specifically includes:

[0085] S10, obtaining user electricity usage related information, normalizing the user electricity usage related information to obtain a normalized set;

[0086] It should be noted that the user's electricity usage-related information includes, but is not limited to, the user's electricity usage fee information, payment behavior information, historical arrears information, electricity usage type information, and user profile information;

[0087] The user's electricity cost information and historical arrears information may be information within the past three billing cycles, wherein the user's electricity cost information includes but is not limited to the average receivable electricity charges, the coefficient of variation of the average receivable electricity charges, the change in the receivable electricity charges, the month-on-month change in the latest billing cycle's receivable charges, and the month-on-month change in the previous billing cycle's receivable charges; the payment behavior information includes but is not limited to the average collection time, payment time, payment method, payment account, the coefficient of variation of the average collection time, the change in the collection time, and the change in the collection time in the latest billing cycle. Electricity usage type information includes industrial electricity, commercial electricity, residential electricity, and agricultural electricity. User profile information includes user name, change record, legal representative, contact number, address, bank account opening, electrical equipment model, power and other information. Historical arrears information includes but is not limited to arrears frequency, average total arrears days, sum of arrears days, average arrears amount, average penalty for breach of contract, whether there is any arrears behavior, change in the total number of arrears days in the latest billing cycle, change in the total number of arrears days in the previous billing cycle, and change in the total number of arrears days.

[0088] S20, loading the normalized set, evaluating the user credit rating based on the TOPSIS algorithm improved by K-means clustering, and uploading the user credit rating to the blockchain platform 300;

[0089] S30, loading the user credit rating, the blockchain platform 300 generates a voice interaction strategy based on the improved Tacotron2 model and the user credit rating;

[0090] S40, in response to the voice interaction strategy, voice interaction information is obtained, the voice interaction information and the user's credit rating are used as input, the voice interaction information and the user's credit rating risk are identified based on a pre-built recovery risk assessment model, and an electricity fee recovery assessment result is output.

[0091] It should be noted that, in response to the voice interaction strategy, when acquiring the voice interaction information, the follow-up acquisition of the voice interaction information can be achieved based on Automatic Speech Recognition and Audio to Audio.

[0092] In the embodiment of the present invention, the TOPSIS algorithm improved based on K-means clustering is used to evaluate the user credit level, so that the K-means clustering combined with the TOPSIS algorithm can comprehensively consider multiple feature indicators to evaluate the user credit level, and at the same time, the voice interaction information and the user credit level are considered in the electricity bill recovery evaluation, so as to provide a more accurate electricity bill recovery evaluation result for electricity bill collection, reduce the electricity bill collection cost and improve the intelligence level of electricity bill collection, and overcome the problem that when the existing method divides users into different electricity bill levels according to the user's electricity bill consumption level, a single electricity bill consumption indicator is not considered to be unable to fully reflect the user's payment credit status, so that the existing electricity bill collection system has poor adaptability and robustness, and cannot provide a more accurate and comprehensive evaluation of the user's electricity bill collection.

[0093] The embodiment of the present invention provides a method for normalizing user electricity consumption related information. Figure 2 The following is a schematic diagram of a method for normalizing user electricity usage-related information. The method for normalizing user electricity usage-related information specifically includes:

[0094] S101, loading user electricity usage related information, and processing missing values ​​and abnormal values ​​of the user electricity usage related information;

[0095] In this embodiment, when processing missing values ​​and abnormal values ​​of user electricity consumption related information, the missing value processing method can be deletion or interpolation, and the abnormal value processing method can be deletion, replacement, or isolation forest method.

[0096] S102, obtaining the user electricity consumption related information after the missing values ​​and abnormal values ​​are processed, and determining the reference threshold and limit value of the characteristic index in the user electricity consumption related information based on the hierarchical analysis method. The reference threshold and limit value of the characteristic index usually refer to the minimum value and maximum value used to map the data to a specific range (such as 0 to 1 or -1 to 1). By setting the reference threshold, all data can be converted to a unified scale for comparison and analysis;

[0097] In the embodiment of the present invention, the user electricity consumption related information of different sources and formats is converted into a unified standard form to facilitate subsequent analysis and processing, and the relative importance of each characteristic indicator in the decision-making can be clarified based on the hierarchical analysis method, thereby providing support for determining the benchmark threshold and limit value.

[0098] The reference threshold and limit value of the characteristic index are calculated by the following formula:

[0099] (1)

[0100] (2)

[0101] (3)

[0102] in, Characteristic index The reference threshold and limit value, represents the input value of the characteristic index, represents the input mean of the feature index, are the maximum and minimum values ​​of the characteristic index, respectively. Indicates the threshold correction value, Represents the characteristic index determined based on the hierarchical analysis method The subjective weight can be 0.1-0.9. Respectively represent the number of feature indicators and the weight ranking of feature indicators, They are the maximum eigenvalue of the characteristic index and the index correlation coefficient based on the hierarchical analysis method. In this embodiment, when calculating the maximum eigenvalue of the characteristic index, the characteristic matrix can be decomposed into the product of two matrices Q and R, and then the maximum eigenvalue is iteratively calculated using this decomposition, and the index correlation coefficient can be 0.5-0.9;

[0103] S103, loading the reference threshold and limit value of the characteristic index, normalizing the user's electricity consumption related information based on the reference threshold and limit value of the characteristic index, and outputting a normalized set;

[0104] When normalizing the user's electricity consumption-related information based on the benchmark threshold and limit value of the characteristic index, the characteristic index output is kept between 0 and 1 through the normalization operation, and the normalization operation is performed through the following expression:

[0105] (4)

[0106] in, represents the normalized value of the feature index, Characteristic index The reference threshold and limit value.

[0107] In an embodiment of the present invention, the user electricity consumption related information is normalized so as to convert the user electricity consumption related information from different sources and formats into a unified standard form. Based on the hierarchical analysis method, the relative importance of each characteristic indicator in the decision-making can be clarified, thereby providing support for determining the benchmark threshold and limit value to facilitate subsequent analysis and processing.

[0108] The embodiment of the present invention provides a method for evaluating user credit rating based on the TOPSIS algorithm improved by K-means clustering. Figure 3The schematic diagram of the implementation process of the method for evaluating the user credit rating based on the TOPSIS algorithm improved by K-means clustering is shown. The method for evaluating the user credit rating based on the TOPSIS algorithm improved by K-means clustering specifically includes:

[0109] S201, loading the normalized set, calculating the characteristic indicator output correlation value based on the Gumbel-Copula function, and using the characteristic indicator output correlation value to describe the correlation between the characteristic indicator and the user's credit rating output;

[0110] It should be noted that the characteristic indicators include but are not limited to the type of arrears, the number of arrears, the arrears mark, the voltage level, the equipment type, and the electricity usage type. The Gumbel-Copula function is a probability distribution model for describing the dependency relationship between random variables, and is used to capture the dependency relationship between multidimensional random variables, especially the tail correlation. In this embodiment, the characteristic indicator output correlation value is calculated based on the Gumbel-Copula function, so as to accurately describe the positive and negative correlation of multiple groups of characteristic indicators on the user's credit level, thereby ensuring the accuracy of the user's credit level assessment. The indicator output correlation value refers to the positive and negative correlation of multiple groups of characteristic indicators on the user's credit level.

[0111] Among them, the characteristic index output correlation value is calculated by the following expression:

[0112] (5)

[0113] In the formula, Indicates the output correlation value of characteristic index, is the estimated degrees of freedom parameter of the Gumbel-Copula function. In this embodiment, the estimated degrees of freedom parameter is 0.5-0.95. Characteristic index and The marginal distribution function of

[0114] S202, obtaining a characteristic indicator output correlation value, objectively weighting the characteristic indicator based on an indicator correlation weight determination method, and obtaining an objective weight of the characteristic indicator;

[0115] The objective weight of the characteristic indicator is calculated by the following formula:

[0116] (6)

[0117] In the formula, Indicates the output correlation value of characteristic index, represents the objective weight of the feature index, are the normalized value and standard deviation of the characteristic index respectively;

[0118] S203, loading the objective weight of the characteristic index, and calculating the optimal weight of the index based on the coupling of the subjective weight and the objective weight of the characteristic index by Moral game;

[0119] In the embodiment of the present invention, based on the coupling of subjective weights and objective weights of characteristic indicators by Moral game, by combining subjective weights and objective weights, the experience of decision makers and the analysis results of objective data can be more comprehensively considered, thereby improving the quality of decision-making. At the same time, the user payment similarity matrix is ​​multiplied by the optimal weight of the indicator to obtain a weighted standardized matrix, and the improvement of the TOPSIS algorithm is completed. Then, the positive and negative ideal solutions are determined, and the distance between the evaluation object and the positive and negative ideal solutions is calculated by Euclidean distance. The progress is used as the output of the user's credit level, so that the improved TOPSIS algorithm can comprehensively consider multiple indicators, so that the assessment of the user's telecommunications credit level is more comprehensive, and the deviation caused by human subjective judgment is reduced.

[0120] Among them, the optimal weight calculation method of the indicator is:

[0121] (7)

[0122] in, represents the optimal weight of the indicator, is the optimal value of the indicator based on Moral game, ;

[0123] S204, constructing a user payment similarity matrix based on a K-means clustering algorithm and determining the indicator center clustering point, multiplying the user payment similarity matrix by the indicator optimal weight to obtain a weighted standardized matrix;

[0124] S205, obtaining a weighted normalized matrix, and the TOPSIS algorithm determines positive and negative ideal solutions based on the weighted normalized matrix;

[0125] S206, loading positive and negative ideal solutions, calculating the Euclidean distances between the indicator center cluster points and the positive and negative ideal solutions, determining the progress of the posting based on the Euclidean distances between the indicator center cluster points and the positive and negative ideal solutions, and using the progress of the posting as the output of the user credit rating.

[0126] In the embodiment of the present invention, Figure 4 As shown, an architecture diagram of the improved Tacotron2 model is shown, and the improved Tacotron2 model includes a word embedding module, a spectrum generation network, and a vocoder.

[0127] The word embedding module is connected to the spectrum generation network, and the spectrum generation network is connected to the vocoder. The vocoder can be a Griffin-Lim vocoder. When improving the Tacotron2 model training, the Jmeter tool is used for testing, and the training rounds are 100-200 rounds.

[0128] Among them, the word embedding module includes a hierarchical input layer, a text capture layer, and a text fusion layer. Asymmetric convolution is introduced in the text fusion layer. The asymmetric convolution includes three groups of convolution kernels, and the convolution kernel sizes are 1×3, 3×1, and 3×3 respectively. The hierarchical input layer is connected to the text capture layer, and the text capture layer is connected to the text fusion layer. The text capture layer captures the corresponding collection text information, collection frequency, and collection priority in the blockchain platform 300 based on the user credit level. In this embodiment, different collection frequencies and collection priorities are set based on different user credit levels. The user credit level is 1-5, with 5 being the highest. The corresponding collection frequency can be once a week, the corresponding collection frequency for level 4 can be once every 3 days, and the corresponding collection frequency for level 3 can be once every 2 days; the corresponding collection frequency for level 2 can be once a day, and the corresponding collection frequency for level 1 can be once a day.

[0129] The spectrum generation network consists of a speech encoder and a speech decoder, wherein the speech encoder includes a bidirectional LSTM layer and a HSFs layer. The bidirectional LSTM layer is used to identify the text fusion results and generate the final hidden feature representation based on the text fusion results. The HSFs layer is used to express the overall speech characteristics of the text fusion results. The speech decoder consists of two groups of multi-layer attention perception mechanisms, two fully connected layers, a linear mapping layer, and a BP neural network fusion device. The BP neural network fusion device is used to output the Melton spectrum. The multi-layer attention perception mechanism and the fully connected layer correspond to the bidirectional LSTM layer and the HSFs layer respectively. The Levenberg-Marquard algorithm is used to train the spectrum generation network when training the improved Tacotron2 model.

[0130] In an embodiment of the present invention, a voice interaction strategy is generated based on an improved Tacotron2 model combined with a user credit rating. The improved Tacotron2 model consists of a word embedding module, a spectrum generation network, and a vocoder, and the spectrum generation network consists of a speech encoder and a speech decoder. A bidirectional LSTM layer and an HSFs layer are introduced into the speech encoder, so that the overall speech characteristics of the text fusion result can be expressed, the audio synthesis quality and speed are guaranteed, and more personalized services can be provided at the same time, so that the collection voice is more in line with the actual situation and needs of the user, and the frequency and intensity of the collection are reduced, thereby reducing interference to the user and improving user satisfaction.

[0131] The embodiment of the present invention provides a method for the blockchain platform 300 to generate a voice interaction strategy based on an improved Tacotron2 model combined with a user's credit rating. Figure 5The schematic diagram of the implementation process of the blockchain platform 300 generating a voice interaction strategy method based on the improved Tacotron2 model combined with the user credit rating is shown. The method of the blockchain platform 300 generating a voice interaction strategy based on the improved Tacotron2 model combined with the user credit rating specifically includes:

[0132] S301, obtaining the user's credit rating, improving the word embedding module in the Tacotron2 model to capture the corresponding collection text information, collection frequency, and collection priority in the blockchain platform 300 based on the user's credit rating;

[0133] S302, the text fusion layer fuses the features of the collection text information, collection frequency, and collection priority based on asymmetric convolution, and outputs the text fusion result;

[0134] S303, obtaining the text fusion result, the bidirectional LSTM layer recognizes the text fusion result, generates the final hidden layer feature representation based on the text fusion result, and the HSFs layer synchronously expresses the overall speech characteristics of the text fusion result;

[0135] S304, loading the text fusion result to generate the final hidden layer feature representation and the overall speech characteristic expression of the text fusion result, aligning the text sequence and the speech frame based on the multi-layer attention perception mechanism, and outputting the Melton spectrum;

[0136] S305, converting the Mel spectrum into an audio waveform based on a vocoder to generate a voice interaction strategy.

[0137] In this embodiment, when aligning text sequences and speech frames based on a multi-layer attention perception mechanism, the weight distribution in the attention mechanism is used to find the optimal alignment relationship between text units and speech frames through a dynamic programming algorithm. The aligned text and speech features are then weighted and fused to generate a unified representation.

[0138] The embodiment of the present invention provides a method for identifying voice interaction information and user credit rating risks based on a pre-built recovery risk assessment model. Figure 6 The schematic diagram of the implementation process of the method for identifying the risk of voice interaction information and user credit rating based on the pre-built recycling risk assessment model is shown. The method for identifying the risk of voice interaction information and user credit rating based on the pre-built recycling risk assessment model specifically includes:

[0139] S401, loading voice interaction information and user credit rating, the input layer performs discrete Fourier transform on the voice interaction information based on framing and windowing, and obtains a Fourier transform set corresponding to the voice interaction information;

[0140] In the embodiment of the present invention, frame windowing is an important step in speech signal processing. It reduces spectrum leakage and boundary effects and improves the accuracy of spectrum calculation by dividing the continuous speech signal into short-term stable signal blocks and introducing smooth transition at the frame edge. This process not only adapts to the time-varying characteristics of human speech signals, but also facilitates subsequent feature extraction and analysis.

[0141] S402, obtaining a Fourier transform set, performing equivalent weighted filtering on the Fourier transform set based on an equivalent weighted filter, filtering out interference harmonics in the Fourier transform set, outputting a weighted filtering result, and performing discrete cosine transform on the weighted filtering result to obtain a static speech signal;

[0142] After the equivalent weighted filter performs equivalent weighted filtering on the Fourier transform set, the weighted filtering result is expressed as:

[0143] (8)

[0144] (9)

[0145] (10)

[0146] in, represents the weighted filtering result, , are the sampling angle differences of the equivalent weighted filter for the positive and negative frequencies of the harmonics, is the amplitude of the interfering harmonic, is the filtering times of the equivalent weighted filter. In this embodiment, the filtering times can be 3-8 times. represents the order of the equivalent weighted filter, is the input representation of the Fourier transform set, , are the sampling frequency and fundamental frequency of the Fourier transform set, respectively. is the phase of the interfering harmonic, represents the frequency-shifted signal in Fourier transform concentration;

[0147] S403, loading a static speech signal, the preprocessing layer Conv1 converts the static speech signal into a static speech vector, the residual convolution network Conv2_x performs convolution fusion on the static speech vector, and outputs the convolution fusion result;

[0148] It should be noted that Conv2_x solves the gradient vanishing problem in deep networks by introducing skip connections, making the recycling risk assessment model deeper and thus extracting more complex features. Conv1, as a preprocessing layer, can effectively normalize the input data, reduce noise, and improve the generalization ability of the model.

[0149] S404, obtaining the convolution fusion result, the SPPF module extracts the convolution fusion result of the convolution fusion network Conv2_x input, and the compression incentive module uses the user's credit level as prior information and determines the electricity fee recovery assessment value based on the convolution fusion result of the convolution fusion network Conv2_x input extracted by the SPPF module;

[0150] The electricity recovery assessment value is calculated using the following formula:

[0151] (11)

[0152] in, To recover the assessed value of electricity bills, is the user's credit rating, Extract the convolution fusion result of the convolution fusion network Conv2_x input for the SPPF module. is the bias vector of the SiLU activation function, Represents the SiLU activation function.

[0153] In this embodiment, when the recycling risk assessment model is pre-constructed, a convolutional neural network is used as the initial model. The initial model consists of an input layer, a convolution module, a pooling layer and an output layer. The convolution module is improved, and a preprocessing layer Conv1 and a residual convolution network Conv2_x are introduced into the convolution module. The pooling layer is frozen, and the pooling layer is replaced by a feature fusion layer. The feature fusion layer consists of an SPPF module and a compression excitation module. The number of channels of the SPPF module is 256, and the parameter is 5. It is used to extract the convolution fusion result of the convolution fusion network Conv2_x input. The compression excitation module is used to infer the feature fusion result of the SPPF module input. The user credit level is used as prior information to determine the electricity fee recovery assessment value. The activation function of the convolution module is the Sigmoid activation function, and the activation function of the feature fusion layer is the SiLU activation function.

[0154] In an embodiment of the present invention, a recycling risk assessment model is provided. The recycling risk assessment model uses a convolutional neural network as an initial model, introduces a preprocessing layer Conv1 and a residual convolutional network Conv2_x, and also introduces a feature fusion layer composed of an SPPF module and a compression excitation module. The combination of the SPPF module and the compression excitation module increases the adaptability of the recycling risk assessment model to different types of features and improves the generalization ability of the recycling risk assessment model, so that the recycling risk assessment model can use the user's credit level as prior information and combine voice interaction information to ensure the accuracy of the electricity bill recovery assessment value.

[0155] The embodiment of the present invention provides an electricity bill collection system based on voice interaction. Figure 7The schematic diagram of the structure of the electricity fee collection system based on voice interaction is shown, and the electricity fee collection system based on voice interaction specifically includes:

[0156] The information acquisition module 100 is used to acquire the user's electricity usage related information, and normalize the user's electricity usage related information to obtain a normalized set;

[0157] The rating evaluation module 200 loads the normalized set, evaluates the user credit rating based on the TOPSIS algorithm improved by K-means clustering, and uploads the user credit rating to the blockchain platform 300;

[0158] The blockchain platform 300 is used to load the user credit rating. The blockchain platform 300 generates a voice interaction strategy based on the improved Tacotron2 model and the user credit rating;

[0159] The recycling assessment module 400 obtains voice interaction information in response to the voice interaction strategy, takes the voice interaction information and the user's credit rating as input, identifies the voice interaction information and the user's credit rating risk based on a pre-built recycling risk assessment model, and outputs an electricity fee recycling assessment result.

[0160] In this embodiment, the information acquisition module 100 includes:

[0161] The pre-processing unit 110 loads the user's electricity usage related information and processes the missing values ​​and abnormal values ​​of the user's electricity usage related information;

[0162] The index analysis unit 120 is used to obtain the user power consumption related information after the missing values ​​and abnormal values ​​are processed, and determine the reference threshold and limit value of the characteristic index in the user power consumption related information based on the hierarchical analysis method;

[0163] The normalization unit 130 loads the reference threshold and the limit value of the characteristic index, performs a normalization operation on the user's electricity consumption related information based on the reference threshold and the limit value of the characteristic index, and outputs a normalized set.

[0164] It should be noted that the electricity bill collection system based on voice interaction provided in the embodiment of the present invention corresponds to the above-mentioned electricity bill collection method based on voice interaction. The explanations, examples, beneficial effects and other parts of the relevant contents can refer to the corresponding contents in the electricity bill collection method based on voice interaction, and will not be repeated here.

[0165] In this embodiment, the blockchain platform 300 can be a database using distributed storage, without centralized nodes, which can effectively prevent information tampering, thereby ensuring information security. The information acquisition module 100 stores the processed data in the blockchain platform 300 through the intranet service.

[0166] In summary, the present invention provides an electricity bill collection system and method based on voice interaction. In an embodiment of the present invention, a TOPSIS algorithm improved based on K-means clustering is used to evaluate the user's credit level, so that K-means clustering combined with the TOPSIS algorithm can comprehensively consider multiple feature indicators to evaluate the user's credit level. At the same time, voice interaction information and user credit level are simultaneously considered in the electricity bill collection evaluation, thereby providing a more accurate electricity bill collection evaluation result for electricity bill collection, reducing the electricity bill collection cost while improving the intelligence level of electricity bill collection, and overcoming the problem that the existing method divides users into different electricity bill levels according to their electricity bill consumption level, but does not consider that a single electricity bill consumption indicator cannot fully reflect the user's payment credit status, resulting in poor adaptability and robustness of the existing electricity bill collection system, and cannot provide a more accurate and comprehensive evaluation of the user's electricity bill collection.

[0167] It should be noted that, for the above-mentioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described order of actions, because according to the present invention, some steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0168] In the several embodiments provided in the present application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are only schematic, such as the division of the above-mentioned units, which is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be in the form of telecommunication or other forms.

[0169] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on these 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. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A method for collecting electricity bills based on voice interaction, characterized in that: The electricity fee collection method based on voice interaction includes: Acquire user electricity usage-related information, normalize the user electricity usage-related information, and obtain a normalized set, wherein the user electricity usage-related information includes user electricity usage fee information, payment behavior information, historical arrears information, electricity usage type information, and user profile information; Load the normalized set, evaluate the user's credit rating based on the TOPSIS algorithm improved by K-means clustering, and upload the user's credit rating to the blockchain platform; Load the user's credit rating. The blockchain platform generates a voice interaction strategy based on the improved Tacotron2 model and the user's credit rating. In response to the voice interaction strategy, voice interaction information is obtained, and the voice interaction information and the user's credit rating are used as inputs, and the voice interaction information and the user's credit rating risk are identified based on a pre-built recovery risk assessment model, and an electricity fee recovery assessment result is output; The blockchain platform generates a voice interaction strategy based on the improved Tacotron2 model combined with the user's credit rating, which specifically includes: Obtain the user's credit rating and improve the word embedding module in the Tacotron2 model to capture the corresponding collection text information, collection frequency, and collection priority in the blockchain platform based on the user's credit rating; The text fusion layer fuses the features of collection text information, collection frequency, and collection priority based on asymmetric convolution, and outputs the text fusion result; Obtain the text fusion result, the bidirectional LSTM layer recognizes the text fusion result, generates the final hidden feature representation based on the text fusion result, and the HSFs layer synchronously expresses the overall speech characteristics of the text fusion result; Load the text fusion results to generate the final hidden layer feature representation and the overall speech characteristic expression of the text fusion results, align the text sequence and speech frame based on the multi-layer attention perception mechanism, and output the Melton spectrum; Convert the Mel spectrum into audio waveform based on the vocoder to generate a voice interaction strategy; The method for identifying voice interaction information and user credit rating risks based on the pre-built recovery risk assessment model specifically includes: Load the voice interaction information and user credit rating. The input layer performs discrete Fourier transform on the voice interaction information based on framing and windowing to obtain the Fourier transform set corresponding to the voice interaction information. Acquire a Fourier transform set, perform equivalent weighted filtering on the Fourier transform set based on an equivalent weighted filter, filter out interfering harmonics in the Fourier transform set, output a weighted filtering result, perform discrete cosine transform on the weighted filtering result, and obtain a static speech signal; After the equivalent weighted filter performs equivalent weighted filtering on the Fourier transform set, the weighted filtering result is expressed as: （8） （9） （10） in, represents the weighted filtering result, , are the sampling angle differences of the equivalent weighted filter for the positive and negative frequencies of the harmonics, is the amplitude of the interfering harmonic, is the filtering times of equivalent weighted filter, represents the order of the equivalent weighted filter, is the input representation of the Fourier transform set, , are the sampling frequency and fundamental frequency of the Fourier transform set, respectively. is the phase of the interfering harmonic, represents the frequency-shifted signal in Fourier transform concentration; Load the static speech signal, the preprocessing layer Conv1 converts the static speech signal into a static speech vector, the residual convolution network Conv2_x performs convolution fusion on the static speech vector, and outputs the convolution fusion result; The convolution fusion result is obtained. The SPPF module extracts the convolution fusion result of the convolution fusion network Conv2_x input. The compression incentive module uses the user's credit level as prior information and determines the electricity fee recovery assessment value based on the convolution fusion result of the convolution fusion network Conv2_x input extracted by the SPPF module. The electricity recovery assessment value is calculated using the following formula: （11） in, To recover the assessed value of electricity bills, is the user's credit rating, Extract the convolution fusion result of the convolution fusion network Conv2_x input for the SPPF module. is the bias vector of the SiLU activation function, Represents the SiLU activation function.

2. The method for collecting electricity bills based on voice interaction according to claim 1, characterized in that: The method for normalizing the user's electricity consumption related information specifically includes: Load user electricity usage related information and process missing values ​​and abnormal values ​​of user electricity usage related information; Obtain the user electricity consumption related information after the missing values ​​and abnormal values ​​are processed, and determine the benchmark threshold and limit value of the characteristic indicators in the user electricity consumption related information based on the hierarchical analysis method; The reference threshold and limit value of the characteristic index are calculated by the following formula: （1） （2） （3） in, Characteristic index The reference threshold and limit value, represents the input value of the characteristic index, represents the input mean of the feature index, are the maximum and minimum values ​​of the characteristic index, respectively. Indicates the threshold correction value, Represents the characteristic index determined based on the hierarchical analysis method The subjective weight of Respectively represent the number of feature indicators and the weight ranking of feature indicators, They are the maximum eigenvalue of characteristic indicators and the correlation coefficient of indicators based on the hierarchical analysis method; Loading the benchmark threshold and limit value of the characteristic index, normalizing the user's electricity consumption related information based on the benchmark threshold and limit value of the characteristic index, and outputting a normalized set; When normalizing the user's electricity consumption-related information based on the benchmark threshold and limit value of the characteristic index, the characteristic index output is kept between 0 and 1 through the normalization operation, and the normalization operation is performed through the following expression: （4） in, represents the normalized value of the feature index, Characteristic index The reference threshold and limit value.

3. The method for collecting electricity bills based on voice interaction according to claim 2, characterized in that: The method for evaluating user credit rating based on the TOPSIS algorithm improved by K-means clustering specifically includes: Load the normalized set, calculate the characteristic indicator output correlation value based on the Gumbel-Copula function, and use the characteristic indicator output correlation value to describe the correlation between the characteristic indicator and the user's credit rating output; Among them, the characteristic index output correlation value is calculated by the following expression: （5） In the formula, Indicates the output correlation value of characteristic index, is the estimated degrees of freedom parameter of the Gumbel-Copula function, Characteristic index and The marginal distribution function of Obtain the output correlation value of the characteristic indicator, objectively weight the characteristic indicator based on the indicator correlation weight determination method, and obtain the objective weight of the characteristic indicator; The objective weight of the characteristic indicator is calculated by the following formula: （6） In the formula, Indicates the output correlation value of characteristic index, represents the objective weight of the feature index, are the normalized value and standard deviation of the characteristic index respectively; Load the objective weight of the characteristic index, and calculate the optimal weight of the index based on the coupling of the subjective weight and objective weight of the characteristic index by Moral game; Among them, the optimal weight calculation method of the indicator is: （7） in, represents the optimal weight of the indicator, is the optimal value of the indicator based on Moral game, .

4. The method for collecting electricity bills based on voice interaction according to claim 3, characterized in that: The method for evaluating user credit rating based on the TOPSIS algorithm improved by K-means clustering specifically includes: Based on the K-means clustering algorithm, the user payment similarity matrix is ​​constructed and the indicator center clustering point is determined. The user payment similarity matrix is ​​multiplied by the optimal weight of the indicator to obtain a weighted standardized matrix. A weighted normalized matrix is ​​obtained, and the TOPSIS algorithm determines positive and negative ideal solutions based on the weighted normalized matrix; Load the positive and negative ideal solutions, calculate the Euclidean distance between the indicator center cluster point and the positive and negative ideal solutions, determine the progress of the posting based on the Euclidean distance between the indicator center cluster point and the positive and negative ideal solutions, and use the progress of the posting as the output of the user's credit level.

5. The method for collecting electricity bills based on voice interaction according to claim 1, characterized in that: The improved Tacotron2 model includes a word embedding module, a spectrum generation network, and a vocoder, wherein the word embedding module includes a hierarchical input layer, a text capture layer, and a text fusion layer. Asymmetric convolution is introduced in the text fusion layer. The asymmetric convolution includes three groups of convolution kernels, and the sizes of the convolution kernels are 1×3, 3×1, and 3×3, respectively. The spectrum generation network consists of a speech encoder and a speech decoder, wherein the speech encoder includes a bidirectional LSTM layer and a HSFs layer. The bidirectional LSTM layer is used to identify the text fusion result and generate the final hidden layer feature representation based on the text fusion result. The HSFs layer is used to express the overall speech characteristics of the text fusion result. The speech decoder consists of two groups of multi-layer attention perception mechanisms, two fully connected layers, a linear mapping layer, and a BP neural network fusion device. The BP neural network fusion device is used to output the Melton spectrum. The multi-layer attention perception mechanism and the fully connected layer correspond to the bidirectional LSTM layer and the HSFs layer, respectively. The Levenberg-Marquard algorithm is used to train the spectrum generation network during training of the improved Tacotron2 model.

6. The method for collecting electricity bills based on voice interaction according to claim 1, characterized in that: When the recycling risk assessment model is pre-constructed, a convolutional neural network is used as the initial model. The initial model consists of an input layer, a convolution module, a pooling layer and an output layer. The convolution module is improved, and a preprocessing layer Conv1 and a residual convolution network Conv2_x are introduced into the convolution module. The pooling layer is frozen, and the pooling layer is replaced by a feature fusion layer. The feature fusion layer consists of an SPPF module and a compression excitation module. The number of channels of the SPPF module is 256, and the parameter is 5. It is used to extract the convolution fusion result of the convolution fusion network Conv2_x input. The compression excitation module is used to infer the feature fusion result of the SPPF module input. The user credit level is used as prior information to determine the electricity fee recovery assessment value. The activation function of the convolution module is the Sigmoid activation function, and the activation function of the feature fusion layer is the SiLU activation function.

7. A system for collecting electricity bills based on voice interaction, used to implement the method for collecting electricity bills based on voice interaction as claimed in any one of claims 1 to 6, characterized in that: The electricity fee collection system based on voice interaction specifically includes: An information acquisition module is used to acquire user electricity usage related information, and normalize the user electricity usage related information to obtain a normalized set; The rating evaluation module loads the normalized set, evaluates the user's credit rating based on the TOPSIS algorithm improved by K-means clustering, and uploads the user's credit rating to the blockchain platform; The blockchain platform is used to load the user's credit rating. The blockchain platform generates a voice interaction strategy based on the improved Tacotron2 model and the user's credit rating; The recycling assessment module responds to the voice interaction strategy, obtains voice interaction information, takes the voice interaction information and the user's credit level as input, identifies the voice interaction information and the user's credit level risk based on a pre-built recycling risk assessment model, and outputs the electricity fee recycling assessment result.

8. The electricity bill collection system based on voice interaction as claimed in claim 7, characterized in that: The information acquisition module comprises: A preprocessing unit, which loads the user's electricity usage related information and processes the missing values ​​and abnormal values ​​of the user's electricity usage related information; An index analysis unit, used to obtain the user electricity consumption related information after the missing values ​​and abnormal values ​​are processed, and determine the reference threshold and limit value of the characteristic index in the user electricity consumption related information based on the hierarchical analysis method; The normalization unit loads the reference threshold and limit value of the characteristic index, performs normalization operation on the user's electricity consumption related information based on the reference threshold and limit value of the characteristic index, and outputs a normalized set.

Images