Coding and decoding architecture-based electricity handling service model construction method

Abstract

The invention discloses a power handling service model construction method based on a coding and decoding architecture, a power handling service model comprises an encoder and a decoder, the encoder extracts features from multi-modal corpus data, the encoder destroys the features and generates to-be-processed data, the decoder receives the to-be-processed data output by the encoder, and according to a preset rule, the to-be-processed data is stored in the decoder. The training process of the electricity handling service model comprises the steps that multi-modal corpus data is enhanced and clipped, the multi-modal corpus data is divided into phrase material data and long corpus data according to the clipping length, a low-order model of the electricity handling service model is constructed, and the low-order model of the electricity handling service model is constructed. Inputting the phrase material data into a low-order model of the electricity handling service model for training, inputting the long corpus data into the electricity handling service model for training, and obtaining an output difference value between the low-order model of the electricity handling service model and the electricity handling service model; and adjusting the power handling service model according to the difference between the output difference value and the difference value of the low-order model of the corresponding power handling service model.

Description

Technical Field

[0001] The present invention relates to the field of electricity service and machine learning technology, and more specifically, to a method for constructing an electricity service model based on a codec architecture. Background Art

[0002] Electricity services still rely on the traditional manual review process, which leads to low efficiency in electricity business processing. At present, when customers handle electricity services, they need to consult through the 95598 phone or the State Grid online customer service robot, but due to the cumbersomeness of single-mode operation and the lack of intelligent interaction capabilities, the customer experience is greatly limited. In addition, when customers handle business, they need to upload various documents, including text information, image files, etc. The review of this information is mainly completed manually, which is not only time-consuming, but also prone to human errors. The huge workload of data processing and review in each link of the electricity process leads to high work pressure, low work efficiency, extended customer waiting time, and poor user experience. The existing electricity service model fails to fully utilize the advantages of modern information technology, especially with the increasing popularity of mobile Internet today, users have higher expectations for the convenience and intelligence of services, and traditional service methods obviously cannot meet these needs. Summary of the invention

[0003] In order to solve the problem that electricity service relies on manual review, the present invention provides a method for constructing an electricity service model based on a codec architecture.

[0004] The technical solution of the present invention is as follows:

[0005] A method for constructing an electricity service model based on a codec architecture, wherein the electricity service model includes an encoder and a decoder.

[0006] The encoder extracts features from the multimodal corpus data, and the encoder destroys the features to generate data to be processed.

[0007] The decoder receives the data to be processed output by the encoder and generates corresponding output information according to preset rules.

[0008] The electricity service model obtains multimodal corpus data in the electricity service scenario and uses the multimodal corpus data as training data for training. The training process includes

[0009] The multimodal corpus data is enhanced and trimmed. According to the trimming length, the multimodal corpus data is divided into short corpus data and long corpus data.

[0010] Construct a low-level model of the electricity service model.

[0011] The short corpus data is used to train the low-level model of the power service model, and the long corpus data is used to train the power service model.

[0012] Obtain the output difference value of the low-order model of the electricity service model and the electricity service model,

[0013] The electricity service model is adjusted according to the difference between the output difference value and the difference value of the corresponding low-order model of the electricity service model.

[0014] The above-mentioned method for constructing an electricity service model based on a codec architecture,

[0015] The means of multimodal corpus data enhancement processing include

[0016] Background noise is added by randomly calling up background sound data from the existing background sound database and mixing it into the multimodal corpus data.

[0017] Speech rate adjustment: adjust the rate of multimodal corpus data according to a preset rate multiple or a preset rate.

[0018] Volume change: adjust the volume of multimodal corpus data according to a preset volume multiple or a preset volume value.

[0019] Reverberation processing: call out the reverberation data from the existing reverberation processing database and add the reverberation data to the multimodal corpus data.

[0020] Spectral masking, deleting or neutralizing the content of a random interval of multimodal corpus data,

[0021] Language habit simulation, identifying the content of multimodal corpus data, retrieving data with the same content from the existing voice habit database, and replacing the multimodal corpus data.

[0022] Emotional expression simulation adjusts the parameter attributes of multimodal corpus data according to the preset emotional expression parameter relationship.

[0023] In the above-mentioned method for constructing an electricity service model based on a codec architecture, the enhanced multimodal corpus data is cut according to a preset length to form short corpus data and long corpus data, and an association is established between the short corpus data, the long corpus data and the original multimodal corpus data.

[0024] In the above-mentioned method for constructing an electricity service model based on a codec architecture, before the short corpus data and the long corpus data are put into the model training, the short corpus data and the long corpus data are both subjected to spectrogram transformation processing, and the processing process includes

[0025] Construct a smoothing function so that the short corpus data and the long corpus data are multiplied by the smoothing function respectively.

[0026] Perform fast Fourier transform on the short and long corpus data after smoothing.

[0027] The energy spectra of the short corpus data and the long corpus data are calculated respectively, and the energy spectra are logarithmically calculated to obtain the spectrograms of the short corpus data and the long corpus data.

[0028] The above-mentioned method for constructing a power service model based on a codec architecture uses short text data for model training. The training process includes:

[0029] Obtain a true random number generator, generate multiple true random numbers, and use the true random numbers as a guide to select a region in the short material data to establish a reconstruction region.

[0030] Supplement the reconstructed area based on the content before and after the reconstructed area,

[0031] The reconstructed short text data is input into the low-order model of the electricity service model for calculation to generate the first prediction value.

[0032] Furthermore, the true random number generator generates a true random number for determining the size of the reconstruction area interval and a true random number for determining the starting position of the reconstruction area.

[0033] Furthermore, the process of supplementing the reconstruction region includes

[0034] The value of the previous frame of the reconstructed area is A, and the value of the next frame of the reconstructed area is B. The interpolation value for any time point t in the reconstructed area is

[0035]

[0036] Among them, V t is the insertion value at any time point t in the reconstruction area, t s is the end time point of the reconstruction area, t l is the starting time point of the reconstruction area.

[0037] Furthermore, the generation process of the low-level model of the electricity service model includes

[0038] Get the training data set, which includes input values, model prediction values, and actual output values.

[0039] Define the loss function,

[0040] Based on the electricity service model, the electricity service model is trained using the data of the training data set by descending the gradient of the loss function. When the gradient of the damage function drops to the set threshold, the low-order model of the electricity service model is obtained.

[0041] The above-mentioned method for constructing an electricity service model based on a codec architecture smoothes the output difference value, eliminates the fluctuation of the output difference value by using a smoothing function or a weighted average method, and obtains a stable difference value.

[0042] The electricity service model is adjusted according to the difference between the stable difference value and the difference value of the corresponding low-order model of the electricity service model.

[0043] The above-mentioned method for constructing an electricity service model based on a codec architecture, the multimodal corpus data includes voice data, text data whose content is converted into voice data through TTS technology, image data whose information is extracted through OCR technology and converted into voice data through TTS technology, and video data from which audio data can be extracted.

[0044] The present invention according to the above scheme has the beneficial effect that:

[0045] 1. The multimodal corpus used in the training of the electricity service model enhances the original data by adding background noise, adjusting speech speed, changing volume, reverberation processing, spectrum masking, language habit simulation and emotional expression simulation, making the data richer and enhancing the robustness of the electricity service model in the face of actual application scenarios.

[0046] 2. During the model training process, a true random number generator is used to create reconstruction areas and for the incorporation and superposition of existing data, which increases the randomness of the training data and helps prevent overfitting.

[0047] 3. The electricity service model adopts a smoothing processing mechanism for the output difference value. By using a smoothing function or a weighted averaging method, the fluctuation of the output difference value is eliminated to obtain a more stable difference value. The stable difference value is used to adjust the electricity service model to ensure that the difference value between the electricity service model and its low-order model tends to a stable state, making the electricity service model more stable during the training process and able to more smoothly tend to the ideal set state, thereby improving the reliability and robustness of the overall system.

[0048] 4. Short corpus data is used to train the low-level model of the electricity service model, while long corpus data is used to train the complete model. The training can anticipate scenarios of processing data of different lengths.

[0049] 5. The training of the electricity service model involves two-layer training, namely, the low-level model of the electricity service model and the electricity service model itself. Extending a comparison model from itself is conducive to the comparison and reference of output data, which helps the model converge faster and improves training efficiency and resource utilization efficiency. Layered training helps to improve the overall stability and generalization ability of the electricity service model. DETAILED DESCRIPTION

[0050] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0051] The electricity service model includes encoder and decoder.

[0052] The encoder extracts features from the multimodal corpus data, and the encoder destroys the features to generate data to be processed.

[0053] The decoder receives the data to be processed output by the encoder and generates corresponding output information according to preset rules.

[0054] The training process of the electricity service model is as follows.

[0055] Obtain multimodal corpus data in the electricity service scenario, and enhance and trim the multimodal corpus data.

[0056] The multimodal corpus data includes multiple types of data, which together provide various information about electricity services in order to better understand and process user requests and service needs. Specifically, the multimodal corpus data includes the following data types.

[0057] (1) Text data: such as electronic spreadsheets, application forms, contract documents and other written materials submitted by users; chat records and email exchanges in customer service; opinions and feedback about electricity services posted by users on social media platforms, etc.

[0058] (2) Voice data: recordings of calls made by users to customer service hotlines; user voice commands converted by automatic speech recognition (ASR); sound clips collected by smart speakers or smart home devices, etc.

[0059] (3) Image data: photos of identity documents and property certificates uploaded by users; pictures of fault sites taken by maintenance personnel; photos of promotional posters and bulletin boards in office halls, etc.

[0060] (4) Video data: educational videos and operating guides released by power companies; videos of fault phenomena recorded by users; user behavior videos captured by surveillance cameras, such as operation videos of self-service terminals, etc.

[0061] (5) Sensor data: electricity consumption data recorded by smart meters; household energy usage data collected by Internet of Things (IoT) devices; anomalies detected by sensors on public facilities, etc.

[0062] In the present invention, the data used in constructing the electricity service model is mainly voice data. In addition, text data can be converted into voice data through TTS technology (text to speech technology), image data can be extracted from the text content in the image through OCR technology (optical character recognition technology), and then converted into voice data through TTS technology, and audio data can also be extracted from video data.

[0063] The acquired voice data is enhanced to make it closer to the actual electricity service scenario, or to make the voice data cover more different groups of people, be more diversified, and expand the voice data database. The means of enhancement include background noise addition (adding background noise in different scenarios to the voice data, such as office environment noise, other people's conversations, traffic noise, etc., to simulate a real call environment), speech rate adjustment (adjusting the playback speed of the voice data to simulate the difference in speaking speed between different people, such as the speaking speed of men and women, and the speaking speed difference between different age groups), volume change (adjusting the volume of the voice data to better adapt to voices of different volume levels), reverberation processing (adding appropriate reverberation effects to simulate telephone lines or different room environments), spectrum masking (simulating voice loss, for example, after the voice volume is reduced to a certain level, masking all voice data with a volume below a certain threshold, or using true The random number generator generates any random number sequence, processes the specific rule signal according to the random number sequence, determines that a certain interval of the specific rule signal is the masking interval, and then identifies the data in the same interval of the voice data, neutralizes the data in the interval of the voice data, and masks the data in this interval), language habit simulation (simulates the use of different local accents, dialects and professional terms, calls relevant databases such as accents and dialects, and replaces the content in the original voice data), emotional expression simulation (through a large number of emotional voice databases, the frequency, volume and other voice attributes of each different emotional expression are obtained through content training, and the voice data is adjusted). The above-mentioned enhancement processing means can be arbitrarily selected from one or more combinations. The enhanced voice data greatly expands the original voice data volume. After the same voice data is subjected to different enhancement processing, more processed voice data is generated, thereby expanding the data volume and simulating more electricity service scenarios.

[0064] The enhanced multimodal corpus data is cut according to a preset length to form short corpus data and long corpus data. Short corpus data and long corpus data are two relative concepts. The length of short corpus data is shorter than that of long corpus data. The cut multimodal corpus data is associated with the original multimodal corpus data and can be traced through lists or attributes. That is, a short corpus data can correspond to a long corpus data, and then can correspond to the multimodal corpus data before cutting.

[0065] After clipping, the short corpus data and the long corpus data are both subjected to spectrogram transformation. Each corpus data is smoothed, and each corpus data is multiplied by the smoothing function to obtain the smoothed corpus data. Since the corpus data is the speech data formed after clipping, its beginning and end do not smoothly transition to the zero point, resulting in missing at the edge (starting point and ending point). The corpus data is multiplied by the smoothing function. Since the middle value of the smoothing function is large and the values ​​at both ends are close to zero, when the corpus data is multiplied by the smoothing function, the two ends of the original corpus data approach zero, the edge gradually decays, and the discontinuity at the beginning and end of the corpus data is greatly reduced, thereby reducing spectrum leakage. The smoothed corpus data is subjected to fast Fourier transform (FFT), the time domain signal is converted into a frequency domain signal, the complex spectrum is output, the amplitude spectrum or energy spectrum is calculated (amplitude spectrum, that is, the absolute value of FFT output, square), and the energy spectrum is taken logarithm to obtain the logarithmic energy spectrum. The logarithmic energy spectrum is plotted into an image, where the horizontal axis represents time, the vertical axis represents frequency, and the color or grayscale represents the energy intensity of different frequencies at different times, thus obtaining a spectrogram of the corpus data.

[0066] Both the short corpus data and the long corpus data are transformed into spectrograms and enter different training paths respectively.

[0067] The path to apply short text data for model training is as follows.

[0068] Get a true random number generator, and for a short sample data, generate a true random number for determining the size of the reconstructed area and a true random number for the starting coordinates of the reconstructed area, wherein the true random number for determining the size of the reconstructed area can be a certain ratio value, a certain width value, or a certain height value, and the true random number for the starting coordinates of the reconstructed area can be two true random numbers, representing the X and Y coordinates, respectively, or the nth value in sequence (similar to the parameter used to indicate the number on the left in the LEFT function of Excel) (n is a natural number). After all the true random numbers are determined, the position and size of the reconstructed area are both determined values, and the reconstructed area is compared with the short sample data, and the values ​​of the positions where the short sample data and the reconstructed area are compared are all modified to zero or null values. In this way, the processed short sample data is saved for the next step of processing.

[0069] The reconstructed area of ​​the short material data is trained and supplemented. The reconstructed area is filled with fixed, preset training supplement values ​​according to a certain rule, so that these obscured values ​​have meaning again. The values ​​filled into the reconstructed area are basically determined by the values ​​around the reconstructed area, and are filled into the reconstructed area by linear interpolation. Specifically, assuming that the value of the previous frame of the reconstructed area is A, and the value of the next frame of the reconstructed area is B, the interpolation value for any time point t in the reconstructed area is

[0070]

[0071] Among them, V t is the insertion value at any time point t in the reconstruction area, t s is the end time point of the reconstruction area, t l The training supplementary value determined by the linear interpolation method increases the correlation between the value of the reconstructed area and the value of the front and back ends, and makes the entire short material data more compact.

[0072] After the short material data supplemented by the training is saved, it is put into the low-order model of the electricity service model for training.

[0073] The low-order model of the electricity service model is a smaller model with higher computational efficiency formed after the electricity service model is compressed. The electricity service model obtains a training data set, and the content of this training data set includes input values, model prediction values ​​and actual output values, that is, a certain input content (including parameters, conditions, etc.) generates an actual output value under actual circumstances, and this input content is put into the electricity service model to generate a model prediction value. Based on the combination of these three values, there are multiple such combinations of values ​​to form a training data set. The electricity service model defines a loss function, which includes the actual loss difference and the model loss difference. The actual loss difference is expressed as the difference between the output value of the low-order model of the electricity service model and the actual output value for the same input value, and the model loss difference is the difference between the output value of the low-order model of the electricity service model and the model prediction value for the same input value. In actual operation, the actual loss difference and the model loss difference can be defined contents, and the loss function they constitute is a fixed function, or a function with a variable parameter of environmental change, or a function is used to associate the two values ​​of the actual loss difference and the model loss difference, so that the two values ​​are more balanced. After defining the loss function, based on the power service model, the power service model is trained using the data of the training data set by descending the gradient of the loss function. When the gradient of the damage function drops to the set threshold, the power service model obtained by immediate training is a low-order model.

[0074] On the other hand, long corpus data is directly put into the electricity service model for training.

[0075] The reconstructed short corpus data is input into the encoder of the low-order model of the electricity service model, and output by the low-order decoder of the low-order model of the electricity service model to obtain the first prediction value. The long corpus data is input into the encoder of the electricity service model, and output by the decoder of the electricity service model to obtain the second prediction value. The difference between the first prediction value and the second prediction value is calculated to obtain the contrast difference. The contrast difference is analyzed, and a data set or database is constructed with the contrast difference as the object. The contrast difference is smoothed by a smoothing function or a weighted average method to eliminate the fluctuation of the contrast difference data and obtain a stable difference. The electricity service model is adjusted in the reverse direction with the stable difference, so that the difference between the low-order model of the electricity service model and the electricity service model is a stable value.

[0076] Since the low-order model of the electricity service model is formed based on the electricity service model under the state of a set loss function, the stable difference between the low-order model of the electricity service model and the electricity service model is used to reversely adjust the electricity service model. In this way, the difference between the low-order model of the electricity service model and the electricity service model can be stabilized and tend toward the set loss function, or the ideal setting, thereby making the electricity service model more stable and tend toward the ideal value.

[0077] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing an electricity service model based on a codec architecture, characterized in that: The electricity service model includes encoder and decoder. The encoder extracts features from the multimodal corpus data, and the encoder destroys the features to generate data to be processed. The decoder receives the data to be processed output by the encoder and generates corresponding output information according to preset rules. The electricity service model obtains multimodal corpus data in the electricity service scenario and uses the multimodal corpus data as training data for training. The training process includes The multimodal corpus data is enhanced and trimmed. According to the trimming length, the multimodal corpus data is divided into short corpus data and long corpus data. Construct a low-level model of the electricity service model. The short corpus data is used to train the low-level model of the power service model, and the long corpus data is used to train the power service model. Obtain the output difference value of the low-order model of the electricity service model and the electricity service model, The electricity service model is adjusted according to the difference between the output difference value and the difference value of the corresponding low-order model of the electricity service model.

2. According to the method for constructing an electricity service model based on a codec architecture as described in claim 1, it is characterized in that: The means of multimodal corpus data enhancement processing include Background noise is added by randomly calling up background sound data from the existing background sound database and mixing it into the multimodal corpus data. Speech rate adjustment: adjust the rate of multimodal corpus data according to a preset rate multiple or a preset rate. Volume change: adjust the volume of multimodal corpus data according to a preset volume multiple or a preset volume value. Reverberation processing: call out the reverberation data from the existing reverberation processing database and add the reverberation data to the multimodal corpus data. Spectral masking, deleting or neutralizing the content of a random interval of multimodal corpus data, Language habit simulation, identifying the content of multimodal corpus data, retrieving data with the same content from the existing voice habit database, and replacing the multimodal corpus data. Emotional expression simulation adjusts the parameter attributes of multimodal corpus data according to the preset emotional expression parameter relationship.

3. According to the method for constructing an electricity service model based on a codec architecture as described in claim 1, it is characterized in that: The enhanced multimodal corpus data is cut into short corpus data and long corpus data according to a preset length, and an association is established between the short corpus data, the long corpus data and the original multimodal corpus data.

4. According to the method for constructing an electricity service model based on a codec architecture as described in claim 1, it is characterized in that: Before the short and long corpus data are put into the model training, the short and long corpus data are transformed into spectrograms. The processing process includes Construct a smoothing function so that the short corpus data and the long corpus data are multiplied by the smoothing function respectively. Perform fast Fourier transform on the short and long corpus data after smoothing. The energy spectra of the short corpus data and the long corpus data are calculated respectively, and the energy spectra are logarithmically calculated to obtain the spectrograms of the short corpus data and the long corpus data.

5. According to the method for constructing an electricity service model based on a codec architecture as described in claim 1, it is characterized in that: Apply short text data to train the model. The training process includes Obtain a true random number generator, generate multiple true random numbers, and use the true random numbers as a guide to select a region in the short material data to establish a reconstruction region. Supplement the reconstructed area based on the content before and after the reconstructed area, The reconstructed short text data is input into the low-order model of the electricity service model for calculation to generate the first prediction value.

6. According to the method for constructing an electricity service model based on a codec architecture as described in claim 5, it is characterized in that: The true random number generator generates a true random number for determining the size of the reconstruction area interval and a true random number for determining the starting position of the reconstruction area.

7. According to the method for constructing an electricity service model based on a codec architecture as described in claim 5, it is characterized in that: The process of reconstructing the region is supplemented by The value of the previous frame of the reconstructed area is A, and the value of the next frame of the reconstructed area is B. The interpolation value for any time point t in the reconstructed area is Among them, V t is the insertion value at any time point t in the reconstruction area, t s is the end time point of the reconstruction area, t l is the starting time point of the reconstruction area.

8. According to the method for constructing an electricity service model based on a codec architecture as described in claim 5, it is characterized in that: The generation process of the low-level model of the electricity service model includes Get the training data set, which includes input values, model prediction values, and actual output values. Define the loss function, Based on the electricity service model, the electricity service model is trained using the data of the training data set through the gradient descent of the loss function. When the gradient of the damage function drops to the set threshold, the low-order model of the electricity service model is obtained.

9. According to the method for constructing an electricity service model based on a codec architecture as described in claim 1, it is characterized in that: Smooth the output difference value, use a smoothing function or weighted average to eliminate the fluctuation of the output difference value, and obtain a stable difference value. The electricity service model is adjusted according to the difference between the stable difference value and the difference value of the corresponding low-order model of the electricity service model.

10. The method for constructing an electricity service model based on a codec architecture according to claim 1, characterized in that: The multimodal corpus data includes voice data, text data whose contents are converted into voice data by TTS technology, image data whose information is extracted by OCR technology and converted into voice data by TTS technology, and video data from which audio data can be extracted.

Images