1. Introduction
In recent years, with the rapid development of information technology such as "Internet
+", cloud computing, AI technology, big data, and other information technology, enterprises
are facing increasing pressure to "digital" transformation [1]. Financial management is an important part of enterprise management, its management
level directly affects the various aspects of enterprise production and operation
and further affects the economic benefits of enterprises. However, the current financial
management of most enterprises has low accounting efficiency, visualization is not
high, the interaction between financial information and business information is not
strong, and other issues, are far from meeting the needs of enterprise development
of financial information. Therefore, to promote the enterprise to "digital" transformation,
there is an urgent need to promote the enterprise financial management of "digital",
industry financial integration is the general trend of transformation and development
[2].
In the era of digitization, the use of emerging technology to explore the value behind
financial data has attracted extensive attention from all walks of life. At the theoretical
level, Zhang et al. analyzed that the financial analysis field is currently faced
with the opportunity of mixing and matching various types of data, such as the coexistence
of financial data and non-financial data, structured data, and unstructured data [3]. Data is complex and diverse, to revealing the connection behind this massive data,
to providing advice for the development of the enterprise's financial strategy will
be the future development trend in the field of financial analysis. Business data
analysis is based on business intelligence (BI) tools, after data exploration, data
aggregation, data pre-processing, and then data analysis and mining, and finally through
the visualization design to show the results of the analysis. Du et al. based on the
analysis of the RPA application basis and Gome application scenarios, constructed
applications including accounting, financial analysis, internal control, and intelligent
upgrading of the financial sharing center, and elaborated that the application of
financial robots can optimize the company's financial analysis system, achieve multi-level
business analysis, and improve the efficiency of financial analysis [4,5]. Cheng based on robotic process automation (RPA), from a variety of aspects to collect
and analyze the enterprise's financial data, according to the user's requirements
to provide multi-dimensional financial analysis and financial report generation process
optimization and improvement, pointed out that the application of RPA technology for
the analysis and prediction of the enterprise's financial condition of the business
is more comprehensive and scientific [6]. RPA financial data analysis robot relies on RPA technology, combined with the automation
of the financial data analysis process and the formation of a kind of intelligent
analysis software. It follows the established rules and procedures, using robotic
process automation technology, through simulation, enhancement, and expansion of the
interaction process between analysts and computer systems, automated collection, cleaning,
analysis, and visualization of financial data, and automatically generates financial
data analysis reports, assisting financial analysts to efficiently complete the analysis
of the task in the analysis of the higher degree of standardization, repetitiveness,
and workload of the larger content [7].
In the study of neural network models for intelligent recognition involved in RPA,
in 2008, Hiroshi Ueda investigated a parallel algorithm of infinite density matrix
rearranging group (iDMRG) applied to one-dimensional quantum systems [8]. It combines a variant of Hida's iDMRG applied to stochastic one-dimensional spin
systems with McCulloch's wavefunction prediction, which can efficiently parallelize
shared and distributed storage systems and provide a large number of physical quantities,
such as the total energy, the nearest-neighbor spin bond strengths, the spin-spin
correlation function and its correlation lengths, and there is no finite-size effect,
and the wavefunction-predicted variant will parallelize the Lanczos method in iDMRG
by a factor of about 3. Li et al. proposed a novel and robust binary coding strategy
that does not require the user to have domain knowledge of CNNs [9]. Based on this, a new quantum behavioral evolution strategy is proposed to ensure
the validity of the evolved CNN structure, which provides better performance and immunity
to perturbation than the traditional method. Iris et al. analyzed a new quantum machine
learning model based on the convolutional neural network [10], which is a quantum convolutional neural network (QCNN) that uses only O (log(N))
variational parameters, allowing it to be efficiently trained and implemented on realistic,
near-term quantum devices. This QCNN system combines multiscale entanglement, renormalization,
and quantum error correction. Leonard et al. made polynomial improvements to known
algorithms for quantum linear systems and proposed the solution of dense linear systems
on a quantum computer [11]. Clader et al. developed a state preparation routine that initializes a general state,
integrating a quantum-compatible preconditioner that greatly extends the number of
problems capable of exponential speedups beyond classical linear system solvers [12,13]. Li et al. creatively accomplished the combination of parametric quantum circuits
with convolutional neural networks. The recognition model constructed by this scheme
achieved considerable experimental results [14,15]. Similar to the traditional DCNN, the model has the structure of a quantum convolutional
layer and a quantum classification layer. Inspired by variational quantum algorithms,
a quantum-classical hybrid training scheme is given for parameter updating of QDCNN
[16-18]. Maxwell et al. verified the feasibility of the method by using random quantum lines
to locally transform the input information [19,20]. Nathan et al. verified the feasibility of the method by building an efficient system
of linear equations solving algorithms [21-23] that efficiently determine the quality of the least-squares fit on exponentially
large data sets, and also efficiently find a succinct function that approximates the
fitted data and limits the approximation error [24-28], which can be used to provide efficient parameter estimation of the quantum state
when the input data is a pure quantum state [29,30]. Therefore, in this paper, a novel optimized quantum neural network model is developed
based on RPA robot visualization technology and applied to the study of financial
data analysis platform construction.
2. RPA Financial Data Analysis Robot Modeling
The model construction of the financial data analysis robot is a guideline for intelligent
financial analysis, which specifies the construction goal of the robot and its specific
application, and at the same time explores the data processing mechanism and visualization
generation mechanism of the RPA robot technology to guarantee the feasibility, rationality
and compliance of the system implementation. Financial data analysis is essentially
a process of data collection, data screening, data calculation, and data analysis.
Based on the features and advantages of RPA+AI technology processing, fast deployment,
and multi-end compatibility, the framework model of the financial data analysis robot
is constructed. The framework model includes an infrastructure layer, data layer,
service layer, platform layer, and application layer, as shown in Fig. 1. The infrastructure layer is the foundation of the data layer, service layer, platform
layer, and application layer, providing basic services for the other layers, including
servers, network, storage, data control, information security, etc., which guarantees
the security of the operating environment of the financial data analysis robot.
Fig. 1. RPA Robotic Modeling Framework for Financial Data Analysis.
The data layer is the data foundation of the financial data analysis robot, including
the acquisition of raw financial data, the mechanism of data collection and processing,
and the formation of financial data storage, data dictionaries data tables, and other
documents. The data types of raw financial data are structured data, semi-structured
data, and unstructured data, and the data sources are varied, including data information
from various information systems of the enterprise, as well as relevant paper documents
and electronic files. Based on the non-invasive and easy-to-deploy features of RPA,
it can integrate the data and information between different systems, open the gap
between different systems, realize the information interaction between systems, then
extract the required financial data and information and store them in the financial
database or form data dictionaries, data tables and other documents through the data
collection and processing mechanism. These data constitute the underlying data of
the financial data analysis robot, providing data protection for the work of the financial
robot.
The service layer provides RPA functional components, Al technology, visualization
services, and data analysis report services for financial data analysis robots. This
layer realizes the demanded applications through automation functional components
such as interface operation automation, Excel automation, Word automation, email automation,
browser automation, data processing automation, application program interactions,
etc., and integrates artificial intelligence technology in the development process
of financial data analysis robots. This layer realizes the required applications through
automation function components such as interface operation automation, Excel automation,
Word automation, browser automation, data processing automation, application interaction,
etc. At the same time, it integrates artificial intelligence technology in the development
of financial data analysis robots. For example, Optical Character Recognition (OCR)
technology, which analyzes and processes data in images to obtain textual information
in the data; Natural Language Processing (NLP) technology, which realizes the computer's
accurate understanding of human language and enables it to communicate and interact
with people in a natural way; Speech Recognition (ASR), which converts voice signals
into corresponding text; and TTS technology, i.e., Text-to-Speech (TTS) technology,
which converts textual information into natural, fluent, and more natural textual
information. converted into natural, fluent, and more human-like speech. This layer
also provides visualization services and data analysis report services, pre-setting
Excel analysis template files containing visualization graphics and Word analysis
report template files containing title page, table of contents, summary, and body,
to realize the specific functions of the financial data analysis robot, such as data
entry, processing, calculations, visualization and automatic generation of analysis
reports. The platform layer includes three parts: process design platform, robot program,
and management control platform. The process design platform includes business process
optimization, script development, test runs, debugging, and error correction, and
is mainly responsible for designing the financial data analysis robot. The robot program
is deployed after the process design, and through the execution of console commands,
it realizes the functions of data extraction and processing, data analysis, and generation
of analysis reports. The management and control platform are capable of real-time
monitoring of robot operation, process triggering, human-machine interaction, and
presentation and log analysis of the operation results of the financial data analysis
robot, providing a strong guarantee for the development and application of the financial
data analysis robot back. The application layer is the specific application of RPA
and AI technology in the field of financial data analysis, is the robot realizes the
automation of specific application scenarios, which consists of a cluster of analysis
robots organically combined. In terms of financial data analysis work, it can realize
automated analysis of office expense data, automated analysis of travel expense data,
automated analysis of bank deposit data, and so on.
3. RPA Construction Under Hybrid Quantum Convolutional Neural Network Modeling
Quantum convolutional neural networks are characterized by exponentially scaled memory
capacity, fast processing datasets, good stability of training models, and high learning
efficiency. It has a unique advantage in processing large-scale image data. It lays
a solid technical foundation for the further construction of ultra-high-speed information
processing systems. The quantum convolutional neural network for image data can efficiently
extract image feature values and significantly reduce the difficulty of the prediction
model.
The model of the quantum neural network includes five main aspects: input data, quantum
state preparation, design mode of the quantum convolutional neural network, network
optimization, and final output data. The financial system structure of the quantum
convolutional neural network is shown in Fig. 2.
Fig. 2. Diagram of the financial system architecture of the quantum convolutional neural network.
Fig. 3. Circuit structure of RPA quantum convolutional neural network.
Given the quantum convolutional neural network model hierarchy division, the quantum
convolutional neural network can be divided into a quantum input layer, a quantum
convolutional layer, a pooling layer, a fully connected layer, and an output layer,
as shown in Fig. 3, which displays the circuit diagram structure of a specific quantum convolutional
network. It can be seen that the input data of the quantum convolutional neural network
is based on quantum states. The convolutional layer of this division model consists
of several double quantum gates U, Considering the probabilistic nature of quantum
mechanics, a single quantum bit measurement is generally taken to calculate the probability
of the ground state to calculate the cross-entropy loss.
The Hybrid Quantum Convolutional Neural Network model (HQCNN) designed in this paper
is a new model constructed by combining the quantum convolutional layer with the classical
pooling layer and the classical fully connected layer. The input of the model needs
to be converted from classical data to quantum state and the quantum convolutional
layer is done by the quantum gate operation. The convolution is done by using the
step size of 1 in the quantum convolutional operation in HQCNN. A classical image
is converted into the quantum state for input and training. The quantum convolution
layer is followed by a classical pooling layer connected to a classical fully connected
layer and the output of the pooling layer is represented as:
The output of the fully-connected-layer is represented as follows:
where Z denotes the output of the jth node of the k+1st layer, w denotes the weight
of the ith node of the kth layer connecting the jth node of the k+1st layer, and b
denotes the bias value. For the inverse adjustment of the quantum convolutional layer,
the parameter to be adjusted is the angular value of the single quantum bit gate R
that constitutes the quantum convolutional layer with the following equation:
Further update the formula to:
Where t denotes the number of iterations and n represents the learning rate.
After that, the established model is needed and validated. Firstly, the validation
is carried out by comparing CNN with HQCNN, and the parameter design of the two models
is shown in Table 1.
Table 1. Parameter Settings for the Two Models.
|
RPA Network Architecture
|
Quantum Layer Number (Physics)
|
Core Size
|
Pooling Method
|
Global Connection Layer Book
|
|
CNN
|
3
|
5*5,3*3
|
Maximum pooling
|
2
|
|
HQCNN
|
1
|
2*2
|
Maximum pooling
|
2
|
After repeated experiments, the comparisons of loss values, training set accuracy,
and test set accuracy for 100 iterations of the two models are shown in Table 2. As shown, it can be seen that the loss value of HQCNN is lower than that of the
traditional model by 0.0056%, while the training and test accuracies have been improved
by 0.93% and 0.36%, respectively.
Table 2. Comparison of Experimental Results under the RPA Network.
|
RPA Network Architecture
|
Loss Value In %
|
Training Accuracy %
|
Test Accuracy %
|
|
CNN
|
0.0162
|
97.35
|
96.45
|
|
HQCNN
|
0.0106
|
98.28
|
96.81
|
|
difference (the result of subtraction)
|
-0.0056
|
+0.93
|
+0.36
|
In this case, the loss values for 100 iterations of both models and the accuracy model
tested are shown in Fig. 4:
It can be seen that the hybrid quantum convolutional neural network model improves
in both training set accuracy and test set accuracy. The loss value of the hybrid
quantum convolutional neural network model fluctuates less after 60 iterations and
remains stable after 90 iterations. As can be seen from Fig. 4, the accuracy rate of the hybrid quantum convolutional neural network model is higher
than that of the classical convolutional neural network model after 40 iterations,
and it converges more quickly than the classical convolutional neural network model.
In addition, considering the advantage of the particle swarm algorithm in neural network
systems, we combine the particle swarm algorithm into HQCNN with superior performance
and carry out the application of financial system recognition based on RPA for the
robot visualization platform.
Fig. 4. Loss values after 100 iterations and accuracy curves for the test.
4. Models and Comparisons
According to the research on the hybrid quantum convolutional neural network model,
the input classical data size is preferably in the form of n$^{2}$, but in this paper,
when the dataset space does not match, the dataset has to be further segmented. The
dataset is a CAPTCHA image containing 4 characters, therefore, the image of the financial
system is to be divided equally into different parts and trained separately. Because
in the dataset, there is a great deal of stickiness between the characters of the
CAPTCHA, the characters of the segmented image are not all complete, and here it is
necessary to manually eliminate the images with extremely incomplete segmentation
results, remove these images that may have a greater impact on the training results,
and use the remaining data as a dataset for model training. Due to the serious stickiness
between characters in the financial system dataset, it is not suitable to use the
connected domain method for image segmentation. As shown in Fig. 5, for the characters with serious adhesion, the pixel projection method is used to
segment the characters can get better results. The core idea of the pixel projection
method is to calculate the number of points in each column of the two-dimensional
matrix of the image whose value is not 0, which is denoted as m. The value of m between
non-adhesive characters is 0, and the value of m in the character region will be very
large, and the value of m between characters that are slightly adhered to will be
smaller than the value of m in the character region, so a better segmentation can
be obtained by selecting the appropriate threshold value. The characters with extremely
serious adhesion will be segmented into one character, at this time, the width of
the segmented text image is larger than the width of the image that is segmented into
one character, so the image with a width larger than a certain threshold value will
be cut isometrically, that is, the CAPTCHA image segmentation can be completed. The
threshold value of the pixel projection method is set to 13, and isometric cutting
is performed when the image width is greater than 30.
Fig. 5. Schematic representation of some characters before and after image segmentation.
Fig. 6. Loss curves for three RPA network models for recognition validation.
Fig. 7. Accuracy curves and comparison of three RPA network models for recognition verification.
To convert the classical data into quantum state data, the image is first compressed
to the size of 8$^{2}$ in quantum bit coding mode. The classical representation of
the image is a two-dimensional array, and the value x in the array, which takes the
value range between [0,1], is subjected to quantum bit coding, which is firstly mapped to between [0, ${\pi}$/2],
and then refined by its conversion to quantum state. Finally, the optimized and improved
PSO-HQCNN is compared with CNN and HQCNN = network model for identification and verification,
and the corresponding calculation results. As can be seen from Fig. 6, PSO-HQCNN is slightly smaller than the loss values of CNN and HQCNN. the loss value
curve of the CNN network model decreases more slowly compared to that of the HQCNN
and PSO-HQCNN network models, and the loss curve fluctuates more. In terms of loss
decline speed and the degree of fluctuation, there is no big difference between HQCNN
and PSO-HQCNN in terms of the loss curve.
As can be seen from the three model recognition verification accuracy curves in Fig. 7, PSO-HQCNN has the best recognition effect, with a recognition accuracy rate of up
to 96.85%, followed by HQCNN, which has a recognition accuracy rate of 96.20%, and
CNN, which has the lowest recognition accuracy rate of 95.65%. In terms of convergence
speed, PSO-HQCNN can accomplish faster convergence in achieving a shorter number of
iterations, HQCNN convergence speed is in the middle, and the CNN recognition model
has the slowest convergence speed. The comparison of the training time of the three
network models of CNN, HQCNN, and PSO-HQCNN is shown in Fig. 7. The training time comparison graph of the three models shows that PSO-HQCNN has
the longest training time, which takes 23.6 hours, HQCNN has the shortest training
time, which takes only 1 hour, and CNN has a training time of 1.2 hours. the PSO-HQCNN
model gets an improvement in terms of accuracy at the cost of time. In addition, the
financial system executes the automatic identification system process 200 times and
can successfully identify the financial system 182 times with an average success rate
of 91%, which can meet the requirements of the automatic identification system.
In terms of analyzing complementary aspects of practical applications. When performing
financial data analysis of RPA robot visualization for HQCNN, it is usually necessary
to write a financial data analysis report to present the results of the analysis and
facilitate the reader's use. The financial data analysis report will present the current
situation, problems, causes, and analysis conclusions of the enterprise's finances,
which is one of the main tools to help financial management and decision-makers recognize
the current situation of the enterprise's finances, understand the problems, master
the information, and use the information to make decisions. The data analysis report
automation stage is an essential stage of the financial data analysis robot, in the
specific application before the need to develop an analysis report template file,
management costs financial data analysis report structure is divided into the title
page, the table of contents page, summary, and the body. The title page title should
be concise, and straightforward to mark the content of the analysis report; The directory
is a tool to reveal the analysis report, to help us understand the main content of
the financial data analysis report; a summary is an excerpt of the important points
of the analysis, and its basic elements include the background, purpose, ideas and
conclusions, specifically, the main object and scope of the work of the analysis of
the financial data, the results of the analysis and the important conclusions reached;
the main text is the core part of the financial data analysis report. The financial
data analysis report is the core part of the report, it will be a systematic and comprehensive
expression of the process and results of data analysis, the report body is mainly
divided into the analysis of the background and purpose of the report, the analysis
of ideas, analysis of the content of the report, conclusions and recommendations of
the four parts.
The process of data analysis report automation is based on data analysis, data visualization
mechanism, predefined generating report templates and rules for graphical presentation,
and text output of overhead data analysis. A large number of repetitive and mechanical
work in the data analysis report can be completed by RPA, such as automatically filling
in the date of the report issued, issued by the person, setting the standard analysis
report mode, and text requirements. The route to achieve analysis report automation
by the management cost data analysis robot is mainly accomplished through interface
operation automation, Excel automation, Word automation application program interaction,
and other functional components. Through macro commands to play back the operation
steps, reduce the time spent on complex and repetitive steps, move the visualization
charts in the Excel analysis file to the Word data analysis report, and output the
text information to the corresponding position in the data analysis report template
file through the pre-set analysis rules. At the same time, the overhead data analysis
robot can set different analysis rules according to different analysis contents. For
example, trend analysis can set the range of reasonable differences to analyze the
fluctuations of management expenses, and which periods are not within the range of
reasonable fluctuations require special attention and targeted analysis of their specific
operations to find out the reasons for the differences. In summary, as shown above,
exploring the application of RPA robot visualization to achieve the process can provide
useful exploration and feasible paths for the automated process processing of enterprise
financial data analysis. In this case, the key code for graphical visualization is
shown below
(1) Prepare a Dom with size (width and height) for ECharts
<div id="main" style=" width: 600px; height: 400px;"></div>
(2) Initialization
var myChart = echarts. init (document. getElementByld ('main'));
(3) Configure chart properties and data using setoption.
myChart. setOption ({title title configuration: data1 data matching: xAxis, yAxis
horizontal and vertical coordinate settings;
(4) Ajax asynchronous get data, configure the data to match the chart data
S.ajax ({type: "post", async: true, URL: "", data: {}, dataType: "JSON", success:
function(result){if (result) {for (var i=0; i<result.length; i++){names. push (result[i].name);
// take the attribute categories one by one and fill them into the corresponding data
array}}
5. Conclusions
This study develops novel RPA techniques in financial data analysis based on quantum
neural network modeling. The results of the study show that:
The loss value of HQCNN is 0.0056% lower than that of the conventional model, while
the training and testing accuracy is improved by 0.93% and 0.36%, respectively. In
addition, the hybrid quantum convolutional neural network model of this RPA has less
fluctuation in the loss value after 60 iterations and remains stable after 90 iterations.
Particle swarm optimization algorithm is combined into HQCNN to convert classical
data into quantum state data for joint optimization. The recognition verification
accuracy graph shows that SO-HQCNN has the best recognition effect with a recognition
accuracy rate of 96.85%, HQCNN is second with its recognition accuracy of 96.20%,
and CNN has the lowest recognition accuracy of 95.65%.
PSO-HQCNN has the longest training time of 23.6 hours, HQCNN has the shortest training
time of 1 hour, and CNN has a training time of 1.2 hours. The PSO-HQCNN model gets
an improvement in accuracy at the cost of time.
The financial system executes the automatic identification system process 200 times,
and the number of times it can successfully identify the financial system is 182 times,
and the average success rate is 91%, which can meet the requirements of the automatic
identification system\textemdash{}in summary, as shown in the above, through the quantum
neural network model established by the RPA robot visualization of the application
of the implementation of the process, contributing to the efficiency of the automated
process processing of enterprise financial data analysis.