Abstract
Comprehensive performance evaluation is an important basis for improving the training effect of enterprise employees and the effective allocation of enterprise resources. Based on AHP and BP neural network theory, this paper constructs a comprehensive performance evaluation method for enterprises, AHP is used to calculate the weight of the index, and then the importance index is screened. The model proposes a conceptual model of comprehensive performance of manufacturing enterprises from the support layer, core layer, and promotion layer and constructs a manufacturing system from horizontal and vertical. The influencing factors of comprehensive performance solve the quantification problem of enterprise comprehensive performance evaluation and have obvious guiding value for the research on the integration mode and path of industrialization and industrialization of regional manufacturing enterprises. In the simulation process, the weight of each index in the evaluation system is first determined by the analytic hierarchy process; then the evaluation index membership score table is established, and fuzzy mathematics is used to calculate the expert’s score, so as to solve the problem caused by the intermediate value. The uncertainty caused by the jump is finally established by the analytic hierarchy process, and the neural network is used to simulate the sample. The experimental results show that by using AHP to collect training samples for neural network evaluation, the comprehensive performance evaluation system has good fitness and achieves the best comprehensive consideration of accuracy and training time when there are 17 hidden layer neurons. The maximum relative error is 1.64%, which is much lower than the general accuracy requirement of 5%, which effectively improves the performance and calculation accuracy of the network.
1. Introduction
The comprehensive performance evaluation service has developed rapidly at home and abroad and has received more and more attention [1]. The essence of comprehensive performance evaluation is an innovative product produced by the combination of services, which is a multisubject and multiwin service model [2]. The huge effect of this realistic multiwin model on the participants has made many companies flock to carry out this business. However, when third-party logistics companies carry out comprehensive performance evaluation business, the weak concept of risk management and the low level of risk evaluation also restrict comprehensive performance evaluation [3–5]. R&D’s role in enterprise activities has become increasingly obvious. At the same time, R&D activities are also an important part of scientific research activities and are the foundation of innovation [6, 7].
The comprehensive scientific and technological performance is a special and important resource. Scientific evaluation of them is the premise of recruitment and selection [8], the basis for improving the training effect and the effective allocation of comprehensive performance, and an important factor in formulating employee career planning is given. From the existing research results, it can be seen that most high-tech enterprises do not pay enough attention to the comprehensive performance of science and technology [9], which leads to unreasonable evaluation indicators, single evaluation standards, and unscientific evaluation methods, which brings great influence to the development and management of comprehensive performance. There are big problems [10], such as reduced overall performance satisfaction, or even resignation. In the long run, the management of the enterprise will fall into chaos, and the goals of the enterprise will be difficult to achieve. Therefore, both internationally and domestically, the quantity and structure of R&D resources are regarded as the core indicators for measuring the comprehensive national strength [11]. R&D has the characteristics of exploratory, creative, uncertain, and risky and is crucial to the productivity that transforms scientific and technological achievements into reality [12–14].
This paper mainly studies the subjective and objective AHP-BP model in the field of comprehensive performance evaluation risk assessment. This paper obtains relevant data through the investigation and analysis of the comprehensive performance of science and technology of a high-tech enterprise and uses the established model to conduct simulation research based on MATLAB software to verify the scientificity of the method. In terms of the risk assessment index system, this paper tries to establish a more comprehensive and practical index system: the first part includes the first three chapters, mainly on the research status, significance, index system construction, and the relevant theoretical basis of model construction. The practicability and accuracy of the paper in the third part is mainly about the shortcomings of this paper and the prospect of future research. Although 26 indicators are selected in this paper, they are all traditional financial indicators. The extent to which visualization should be optimal in both horizontal and vertical directions is not reflected in the text. Too high a degree of visualization will lead to a higher level of complexity in the organization, and too low a degree of visualization may affect the organization’s ability to sense. Due to their own limitations, traditional financial indicators pay more attention to the historical information of enterprises. Therefore, it is difficult to fully reveal the potential development capabilities of enterprises.
2. Related Work
On the basis of understanding the current research status at home and abroad, combined with the characteristics of high-tech enterprises’ comprehensive performance of science and technology, so as to provide scientific and technological comprehensive performance evaluation tools for enterprise managers. To form a more objective and scientific understanding of the current state of enterprise comprehensive performance of science and technology [15], to provide a basis for the development planning and strategy formulation of enterprise comprehensive performance.
Human capital property rights are the rights of its owners to dispose of or utilize their own human capital in order to obtain benefits, including the right to possess, use, dispose of, and benefit from human capital. Some western economists put forward the human capital property right incentive theory from the perspective of human capital participating in income distribution. Profit sharing is a typical contractualization of human capital property rights. Liang and Li [16] believed that compared with the traditional performance evaluation system, 360-degree evaluation has its advantages and is more comprehensive and accurate. Huang et al. [17] research believes that when the main purpose of the 360-degree evaluation is to serve the development of employees and provide help for employees’ careers, the evaluation of evaluators will be more objective and fair. Zhang et al. [18] believe that 360-degree evaluation uses multiple perspectives to evaluate employee performance, especially in providing feedback and guidance, allocating bonuses and opportunities, and avoiding evaluation errors. Cai et al. [19] believe that, according to the characteristics of R&D employees, a four-layer KPI evaluation index system for R&D employees is designed, which combines performance result orientation and workload orientation for evaluation, while realizing the evaluation of the R&D personnel of high-tech enterprises at a single time point, through the multistage information aggregation method of dual incentive control lines, the performance of the R&D personnel of high-tech enterprises can be evaluated in a complete R&D project cycle.
The research group led by Qian [20] is mainly based on the job analysis method of competency characteristics and uses empirical evaluation to construct the structure of the competency characteristics of regional enterprise senior managers, etc. and proposes to first determine a set of general basic competencies, and then these competencies are tailored to specific roles, thereby defining performance levels for each competency. Scholars applied the competency model to the administrative management professional personnel training program and put forward several suggestions for formulating the professional personnel training program. This paper believes that competency is closely related to the job performance of enterprise employees [21]. The use of the competency model can predict the future work performance of enterprise employees and can distinguish the outstanding performers and the average ones in the enterprise. It is multilevel, multidimensional, cross-organizational, linked to task scenarios, and dynamic [22]. Through the performance-oriented comprehensive performance management system, the R&D personnel are implemented performance management, and targeted and personalized incentive measures are taken [23–25].
3. Construction of an Enterprise Comprehensive Performance Evaluation Method Based on AHP and BP Neural Network
3.1. Analytic Hierarchy Architecture
The factor analysis hierarchy process can be divided into R-type factor analysis and Q-type factor analysis according to the different research objects to find several common factors that control all variables through the study of the correlation matrix or the internal dependence of the covariance matrix of the variables through the study of the internal structure of the similar matrix of the samples, to find out the main analytic hierarchy process factors that control all the samples.
There are many calculation methods for calculating the single-level sorting of each layer to the previous layer, such as the sum method, root method, characteristic root method, and least square method. The existing research and application show that the root method is accurate and the evaluation effect is better. Therefore, this paper intends to use the root method to calculate the relative weight. It can be seen that the coefficient of variation of the inventory comprehensive performance evaluation rate is 1.74, and the variation coefficient of the business cycle is 1.37, indicating that the inventory comprehensive performance evaluation rate has a greater degree of variation and reflects a relatively large amount of information. Therefore, this paper chooses the comprehensive performance evaluation rate of inventory as one of the representative indicators.
Here, K represents knowledge of comprehensive performance; S represents skills of comprehensive performance; A represents the ability of comprehensive performance; I represents the mediating variable, including motivation and attitude; B represents the behavior of comprehensive performance. The theory holds that behavioral change is the result of a series of jobs and is the primary concern of comprehensive performance development like changes in knowledge, skills, and abilities enable changes in behavior, and the translation of this possibility into reality requires the role of mediating variables such as motivation and attitude. DEA deals with the same type of DMU and indicators with the same input and output, so the DEA method can perform a time series analysis; that is, each time point is a decision-making unit; it can also perform a cross-sectional analysis; that is, each individual are decision-making units.
Figure 1 further judges the influence of indicators by using multiple indicators, if the unit of measurement is the same as the mean, the standard deviation can be directly used for comparison. Among the other five indicators of their own kind, the multiple correlation coefficients of nonperforming assets ratio, long-term asset suitability ratio, accounts payable turnover ratio, comprehensive performance evaluation rate of fixed assets, and comprehensive performance evaluation rate of accounts receivable are 0.266, respectively. Sometimes a single neural network works better than a modular neural network for some data. The reasons are analyzed as follows: the insufficient number of samples after the modules are divided makes the neural network inferior to a single neural network, so such a problem can only be solved by ensuring a number of standard samples. Therefore, this paper selects the comprehensive performance evaluation rate of fixed assets and the comprehensive performance evaluation rate of accounts receivable as reflecting the enterprise for the representative indicator of comprehensive performance evaluation capabilities.
3.2. BP Neural Network Topology
Before using large size kernels, GoogLeNet scales the computation by adding a bottleneck layer with 11 convolution kernels. It uses sparse connections to overcome the problems of information redundancy and high computational cost by omitting irrelevant feature maps. In addition, GoogLeNet innovatively uses global average pooling in the last layer to reduce the density of connections. These parameters are adjusted which resulted in a significant reduction in the number of parameters from 40 million to 5 million. If it is the Inception module of GoogLeNet without dimension reduction, it is the Inception module after adding dimension reduction in Figure 2, subcriteria layers should be further decomposed.
The judgment matrix and scale meaning are as shown, where the middle value of the two values indicates the importance between the two. The neural network will inevitably lose some image feature information due to the deepening of the number of network layers. At the same time, the low-level feature information cannot be directly transmitted to the high-level, and the network cannot learn more robust features. The higher the number of network layers, the better the fitting ability will be, but it may cause the gradient to disappear or the gradient to explode, and the network cannot continue to learn effective features and the network training stops.
The six regions are convolved to extract feature information and then cascaded to extract higher-level feature information so that each pixel can jointly represent and exchange information to ensure richer and more robust feature information. The model in this chapter not only has fewer network layers and fewer parameters but also has a higher accuracy of image classification.
In the task of neural network processing image classification, the conventional operation is to preprocess the image first, then input the whole image to the convolutional neural network model and then output the classification. However, since the category targets of different images may have regional differences, different categories may have highly similar feature information, so the accuracy of image classification only by global features is not high.
The indicators of each enterprise in the neural network sample use standardized data, and the target data use the data evaluated by AHP. The specific evaluation steps are as follows: determine the evaluation index set and output set and determine the number of nodes in the input layer, output layer, and hidden layer. Initialize the weights and thresholds of the neural network nodes. The commonly used method is to assign a random number between 0 and 1. Input neural network learning step size, error target, momentum coefficient, maximum number of iterations, etc. Input sample data, the indicators of the data are standardized data, and the target output is the AHP evaluation value. The neural network propagates forward and calculates the node output values of the input layer, output layer, and hidden layer, respectively.
The network often uses sigmoid log or tangent activation functions and linear functions. When it is desired to limit the output of the network, such as limiting the output between 0 and 1, the logarithmic s-type activation function should be used in the output layer. In general, the s-type activation function is used in the hidden layer, while the output layer uses a linear activation function.
When analyzing the dendrogram in the cluster analysis results obtained by using the systematic clustering method, a threshold value needs to be set manually because the category number system will not be given. Therefore, the threshold value can be determined according to the needs of the research setting, which in turn determines the number of subcategories. According to the cluster analysis results in the above figure, the threshold value selected in this paper is about 22. This paper divides the above 10 indicators into 6 subcategories. Period, current assets, comprehensive performance evaluation rate of total assets, and cash comprehensive performance evaluation period are classified into one category, and the other five indicators are each classified into one category.
3.3. Analysis of Performance Evaluation Indicators
For performance evaluation indicators, quantitative indicator evaluation does not require the participation of multiple evaluators but is obtained by a special responsible person such as an HR assessment specialist based on the real performance of the R&D personnel and comparing it with the performance plan. The performance indicators in this paper are quantitative indicators of R&D personnel, and their scores are related to the corporate objectives of the indicators. The expected objectives of R&D personnel include the target threshold value and the target challenge value. If the target threshold value cannot be reached, the assessment is 0. The basic requirements are for employees, and the target challenge value is the cut-off point of the performance full score of 1, which requires employees to make more efforts to achieve. The samples of 8 belong to ordinary modules and are divided into STU2, and these data are put into Matlab and then simulated, and STU1, STU2, and a single neural network are compared, respectively. In most scenarios, the larger the coefficient, the faster the collection, the smaller the risk of bad debts, and the stronger the ability to repay short-term debts. SSE represents the sum of squares of errors, SSW refers to the sum of squares of network weights, ENP represents effective weights and thresholds, and TPR refers to the number of training steps.
The application value of research results can be divided into parts in Figure 3 for statistics: one is the patent score, that is, regardless of the field and type of the patent, each patent is awarded 5 points; the second is the achievement score; that is, the patent application and the economic value generated in real life is recorded as 1 point for every 20,000 yuan of benefits. The calculation method of multi-person cooperation refers to the calculation of the paper or monograph. The evaluation of qualitative indicators is generally completed by an expert group composed of multiple evaluators. The members of the expert group get the corresponding weights according to their roles with the R&D personnel and calculate the fuzzy membership degree according to the different evaluation levels. Firstly, the actual situation of model initialization is analyzed, then the assessment span is determined, and the weights of assessment personnel are allocated. The members of the assessment team divide the level of the k-th R&D personnel. For example, the indicators can be divided into A, B, C, D, and E grades.
It can be seen that the multiple correlation coefficients of the comprehensive performance evaluation rate of inventory and the operating cycle in the indicators in Figure 4 are 0.963 and 0.966, respectively. After calculating with AHP, the adaptive force evaluation of each manufacturing enterprise can be obtained, which can be expressed as it, the neural network provides training samples. After data processing and normalization or fuzzification, the data is input, and simulation experiments are carried out. After calculation, 8 groups were selected as training data. As long as the index information of the R&D personnel to be evaluated is input into the system, accurate evaluation results can be obtained. We can also multiply the result by 10 or 100 according to the individual needs of the enterprise and restore it to the competency evaluation score suitable for the enterprise. And, with the increase of learning samples, the accuracy of evaluation can be further improved, so it has wide applicability. The results of the multi-index comprehensive evaluation realized by the BP neural network are convincing. It overcomes the influence of human factors, ambiguity, and randomness on evaluation and is an intelligent comprehensive evaluation method.
3.4. Comparison of Network Data Fitness
After investigation, we obtained the values of network data fitness among 12 evaluation indicators. Among them, the economic benefit index generated by patent and achievement application has not obtained comprehensive information due to the difficulty of investigation, so this paper will not calculate it for the time being. See below for the survey data. Since the index c itself is a value between 0 and 1, no normalization is required, and the other index values are between 0 and 100, it is divided by 100 to obtain the normalized data. In this paper, the first 12 samples of the samples are selected as training samples to test the BP neural network. After comparative analysis, it is determined that the number of nodes in the hidden layer is 8, the convergence speed is fast, and the error can meet the requirements. The output value of the network after training is: 0.9701, 0.7801, 0.6294, 078012, 0.6084, 0.6620, 0.58971, 0.8380, 0.6193, 0.6281, 0.7185, and 0.58900.
Figure 5 selects 91 indicators that can basically cover all the information of listed companies. According to the principle of maximum membership, we can see from the judgment matrix month that the evaluation of enterprise B’s manufacturing system self-adaptation is medium. Among its first-level indicators, flexibility is good, visualization is medium, and self-knowledge is medium. Among them, in the evaluation of flexibility, it can be seen from the data that the evaluation is not consistent in this indicator (between good and medium, slightly biased and good). Scoring rules: for quantitative competency indicators, the comprehensive performance management personnel will compare the employee’s performance data with the performance target schedule, analyze, and finally obtain the precise score. Qualitative competency indicators are divided into 5 grades: A, B, C, D, and E. The corresponding grades are evaluated by a 360-degree assessment of the competency of R&D personnel, including experts from the company’s R&D personnel competency evaluation team. For the index rating, the corresponding scores are 10, 8, 6, 4, and 2, respectively. After the rating, fuzzy membership processing is performed, and the scores of the quantitative and qualitative indicators in Table 1 are standardized to [0, 1].
This paper analyzes the application data of R&D personnel in a high-tech enterprise. Quantitative indicators are mainly derived from statistics on the R&D statistical records of the high-tech enterprise, and qualitative indicators are mainly derived from the evaluation records and interviews of employees of the enterprise. The original data of the enterprise are shown in Figure 6. Before training the network, it is necessary to initialize the thresholds and weights of the network. The command newff to establish the network will directly initialize the thresholds and weights of the network when the network is established. Since the input data and output data are both between 0 and 1, the Logsig transformation function is applied to both the hidden layer.
Divide the value data of the 13 groups of R&D personnel competency indicators in Part 01 into the parts in Figure 6, select the first 7 groups of data in Part 01 and the 13 groups of data in Part 02 as learning samples, train neurons to connect weights and thresholds, select after the 01 part, 6 groups of data are checked and tested. When the number of training reaches 1 to 150, the training meets the required accuracy. It can be seen that the maximum relative error between the expected output and the training result is 0.15%, which meets the required accuracy. It can clearly reflect the importance of various indicators, including the application of information technology, the degree of perfection of information communication mechanism, the degree of information sharing, the real-time monitoring of WIP, RM, and FM, and the weights of the utilization rate of production forms and electronic case indicating that the lower-level indicators in the visualization have an impact on the adaptive force of the manufacturing system. This method is very convenient to apply, and the evaluation result can be obtained by inputting the information of the object to be evaluated.
4. Application and Analysis of an Enterprise Comprehensive Performance Evaluation Method Based on AHP and BP Neural Network
4.1. AHP and BP Neural Network Data Pooling
DDU (Decomposition Decision Unit) is a decision subunit that decomposes AHP and BP neural network data. AHP is mainly used to normalize data, make evaluations, generate training samples and weights of various indicators, and input them into different data processing in STU. Among them, STU is a subtask processing unit (Subtask Processing Unit), and the number of STUs is determined by DDU. The selected neural network implements the function of STU, and the neural network is trained according to the data sent by DDU. The data sent in different ranges are trained at the same time using different neural networks. In this paper, MATLAB software is used to program, the financial data of 20 three-level indicators of the previous 15 listed companies are used as the input of the network, the comprehensive score value is used as the output, and the dimensionless data is used to calculate the model. The comprehensive evaluation results and network fitting errors are shown in Figure 7.
The judgment matrix is obtained by sending out questionnaires to experts to collect the data. In this paper, the calculation process and numerical model of AHP are introduced. Combined with the obtained judgment matrix, the relative weights and total weights of each layer of indicators are calculated, and the RI and CI of the weights are tested to judge their validity. By distributing the judgment matrix questionnaire to 8 experts, the judgment matrix is constructed, the weight value obtained by each expert’s judgment matrix is calculated, and then the average value of all weights is calculated by the average method to determine the indicators at all levels. In production flexibility, the application of information technology, the perfection of information communication mechanism, and judgment matrix for the degree of information sharing. From the table, it can be seen that the application of information technology has the largest weight, followed by the perfection of the information communication mechanism and the degree of information sharing. Among them, in the evaluation of comprehensive performance, it can be seen from the data that the evaluation is not consistent in this indicator (between good and medium, slightly biased and good).
By analogy, the weights and consistency tests of the judgment matrices of the remaining experts are calculated. Figure 8 uses the average algorithm to take into account the opinions of each expert, weakening the bias caused by personal subjectivity, and finally obtains the ranking of the weights of each single layer. Assuming that external factors (such as social environment, economic environment, and market environment) remain unchanged, according to the principle of index system construction, five categories of second-level indicators and corresponding 20 third-level indicators are selected, and AHP is used to construct the level of financing ability evaluation. structural model. The weight of each indicator is calculated based on the expert scores. The weights of the usage rate account for 0.1291, 0.0645, 0.0645, 0.0644, and 0.0645, respectively, and these weights are the largest values in the total ranking of the hierarchy, indicating that the lower-level indicators in the visualization have an impact on manufacturing. The overall performance of the system has a greater impact.
4.2. Simulation Realization of Enterprise Comprehensive Performance Evaluation
Based on 20 business sample data that have occurred in company A as the original data, this paper evaluates the accuracy and usability of the model through the evaluation of the AHP-BP model. According to the risk management rules of company A, the risk level of loan companies is generally divided into five grades: excellent, good, medium, poor, and extremely poor. The data involved in this paper are both quantitative and qualitative. Quantitative data can be obtained directly, while qualitative data are scored by experts to quantify qualitative data. Experts’ scoring of quantitative data is based on the actual performance of related companies and the risk classification standards of company A. The expert scores are [0–10], “extremely poor” is [0–2), “poor” is [2–4), “moderate” is [4–6), and “good” is [6–8), “excellent” is [8–10]. According to the original data of company A and expert scores, the risk assessment samples are shown in Table 2.
In the Matlab programming of this article, the error precision is set to 0.15, the maximum number of iterations is 5000 times, the step size is 0.05, the momentum coefficient is 0.25, and the initial weight and threshold are all assigned gradients with random numbers between [0, 1], and validation check are all system defaults. Too few neurons in the hidden layer will make the trained neural network not “robust” enough, and at the same time, the fault tolerance is poor, and there are too many neurons. It will make the network training time too long, and the error may not be the best. The most commonly used membership functions in current research are trapezoidal distribution, triangular distribution, rectangular distribution, and normal distribution. In the specific application process, the appropriate function can be selected as the membership function according to the actual situation of the research object, and the undetermined parameters in the membership function can be determined. After calculation, 8 groups are selected from Figure 9 as training data.
The model in this paper has tested the network training situation from 14 to 20 hidden layer neurons, and the accuracy is all up to the requirements, but although increasing the number of neurons will improve the accuracy, it will also change the training time and error. The best combination of accuracy and training time is achieved when there are layers of neurons. The maximum relative error of the neural network simulation is 1.64%, which is far below the general accuracy requirement of 5%, and the neural network evaluation performs well. According to the simulation diagram, the risk level of each assessed enterprise can be clearly displayed: No. 17 enterprise is “good”, No. 18 and 19 enterprises are “medium”, and No. 20 enterprise is “excellent”. In DDU, the agility value is 0. The membership value of each index is calculated by the method of a trigonometric linear function, and finally, the evaluation result is obtained. Then, the obtained data is used as a sample for neural network training to be simulated and tested. From the analyzed data, the effect of the modular neural network is better.
According to the steps of the analytic hierarchy process, first determine the index weight and obtain the index weight from the formula: W = [0.0292, 0.0118, 0.0719, 0.0252, 0.0444]. Experts scored each sample in Figure 10 according to the original data and enterprise risk assessment standards and obtained the AHP evaluation score of each sample according to the above weights. The scores in the original table will no longer be included in the evaluation, and only some raw data will be scored. From the image before improvement, we can see that the fitness rises relatively slowly, and even the maximum fitness is still declining at the beginning, and it falls into a local optimal solution later. This coefficient is the actual embodiment of the main content of operating profit and is very important to the current assets of the sample. If it is measured by time, it means the time from the establishment of the account receivable to the account. It can be seen from the table that the improved BP algorithm can obtain a high-quality solution in a short time, greatly improve the convergence speed of the search, and significantly improve the global optimization performance.
5. Conclusion
Based on the research on modern comprehensive performance evaluation methods, this paper determines the evaluation method combining AHP and BP neural network. Firstly, the AHP and BP neural networks are summarized, respectively, and their basic principles and application processes are introduced; secondly, the process of calculating index weights and screening important indicators is proposed; finally, based on the screening results, BP neural network is used to carry out scientific case. In the comprehensive performance evaluation, the BP neural network is trained first with the learning samples, and then the performance of the network is verified with the test samples; the simulation research proves that the method is effective. In the comprehensive performance indicators, the relevant research on agility indicators is referenced, which shows that comprehensive performance and agility have something in common, and the research on comprehensive performance is the inheritance and development of agility. The advantage of the neural network is that according to the operation performance of the manufacturing system in the actual process, through self-learning and self-correction, the evaluation results are more in line with the actual situation. In future research, attention should be paid to finding more reasonable and feasible methods to process the data more reasonably. The research method in this paper can effectively absorb the advantages of the two methods, simplify the difficulty of system decision-making, and improve the efficiency and accuracy of evaluation.
Data Availability
The data used to support the findings of this study are available from the author upon request.
Conflicts of Interest
The author declares that there are no conflicts of interest.
Acknowledgments
This work was supported by the 2019 School Level Teaching and Research Project: Audit virtual simulation experiment teaching center (Project no: ch19xnfz02).