In this paper, manufacturing cloud services are divided into hardware and software cloud services. Hardware cloud services produce and manufacture the equipment, whereas software cloud services deliver the software resources. To develop a QoS-aware model for manufacturing cloud services, we first define the information of the manufacturing cloud service.

Definition 1. The QoS information of a manufacturing cloud service S can be modeled as SQoS={pprice, ptime, prel, pava, phon} , where pprice is the price of the cloud service. If S is a hardware cloud service, the price includes the outsourcing and transportation costs. Denoting the current average outsourcing and transporting costs of a single product by Cr  and Ct respectively, we have pprice=Cr+Ct ; if S is a software cloud service, then pprice is the fee of using service S.

ptime represents the time interval between the calling of service S and the receiving of a response. This value indicates the responsiveness of the cloud service to a user’s request. If S is a hardware cloud service, the processing and transportation times of the task product are Tproc and Ttrans respectively, and Ptime=Tcom+Tdelay ; if S is a software cloud service requiring computational time Tcom , the time delay between issuing the call command and the start of the S execution is Tdelay , so we have ptime=Tcom+Tdelay . The above time and price samplings and measurement methods obviously differ from those of traditional services; in particular, they fully consider the business background of the cloud service as a product manufacturing service.

The reliability of the service ( prel ) defines the ability of the cloud service to ∑service(pava) is the probability of normal service operation, and the honesty of the service ( phon ) defines the extent to which the user complies with the agreement after the service is fulfilled. The latter is determined by average evaluation. Suppose that service S is called K times during time interval ( t1 , t2 ). Let the number of normal responses be Kn , the time of no failures be tn , and the users’ evaluations at each time be Ei . The above three indices are then defined as

(1) prel=Kn/K

(2) pava=tn/(t2−t1); 

(3) phon=1Kn∑i=1KnEi.

In the proposed method, dishonest evaluations are detected by a monitor placed in the cloud manufacturing service platform, which collects, checks, and verifies the feedback data of users. Data which are too large or too small, which likely represent the malicious attack data, are filtered out. Whether data represent a malicious attack is judged by user collaborative filtering. Evaluations given by the same users using the same services are expected to vary only slightly.

Suppose that user X and user Y  access the same service  S={s1, s2, …, sk} . The honest evaluations of the k services given by users X and Y  can be expressed as vector spaces Ux:{(s1, x1),(s2, x2),(s3, x3), …, (sk, xk)} and Uy:{(s1, y1), (s2, y2), (s3, y3), …, (sk, yk)} , respectively. Usually, the degree of similarity between two vectors is given by the Euclidean Distance between the vectors. Defining the Euclidean Distance between the vector spaces Ux and Uy as d (Ux, Uy) , the distance between users X and Y  is given by:(4) d(Ux,Uy)=∑i=1k(xi−yi)2

The above method gives the set of users U=1m∑i=1mhi{u1, u2, …um} similar to user X. Suppose that m′′X−type′′ users access a service. If user A evaluates the honesty of a new service s as hx , and the evaluation given by the i−th user is hi , then the average evaluation is ha=1m∑i=1mhi and the honesty of this user’s evaluation is computed as:(5) Hon={0,               |hx−ha|/ha>11−|hx−ha|ha, |hx−ha|/ha≤1

The final evaluation of the users, given by hx × Hon , is then written into the QoS database by the monitor. Therefore, the setting of the user evaluation honesty can adjust or weaken the subjective factors of the users, and filter the outliers that might signify malicious attack data.

The comprehensive QoS indicators of manufacturing cloud services are derived from the historical records of each evaluation indicator. The collection sources can be divided into two categories: QoS indicators of the network performance of the software cloud services (such as service computing time and network transmission delay), and QoS indicators of the manufacturing resources of hardware cloud services. The former can be extracted from the QoS database of the corresponding cloud manufacturing service platform, and the latter are generally provided by users (service providers, service users and platform operators) of the cloud manufacturing platform. The price of outsourcing, transportation, and other services can be directly obtained from the user input values on the manufacturing service platform.

Let S={s1, s2, …, sm} be a set of candidate services providing similar services in the manufacturing cloud pool. The QoS attribute value of each candidate service is inserted into a vector pi={pi1, pi2, …, pin} . The QoS information of the relevant candidate services when selecting cloud manufacturing services is then expressed as:(6) P=[p1p2…pm]=[p11, p12, …, p1np21, p22, …, p2n…, …, …, …,pm1, pm2, …, pmn].

To comprehensively evaluate a cloud service, we must quantify the individual QoS indicators. As different QoS attributes have very different values and measurement units, they cannot be directly calculated. Instead, their values must be normalized to facilitate the multi-objective decision-making. This paper uses the following unified quantitative utility function [26]:

(7) Pij={pjmax−pijpjmax−pjmin,    pjmax−pjmin≠01,                  pjmax−pjmin=0, 

(8) Pij′={pij−pjminpjmax−pjmin,    pjmax−pjmin≠01,                  pjmax−pjmin=0, 

where pjmax and pjmin represent the maximum and minimum values of the j−th attribute among all candidate services in service class S, respectively. The QoS attributes of each candidate service can be divided into forward and reverse attributes. Forward attribute implies that a larger value corresponds to better service quality, such as reliability, availability, and service honesty. This can be calculated using Eq. (8). Reverse attribute implies that a smaller value corresponds to better service quality, such as price and service time. This can be calculated using Eq. (7). After applying the unified quantization to Eq. (6), the following matrix is obtained:(9) P′=[P1P2…Pm]=[P11, P12, …, P1nP21, P22, …, P2n…, …, …, …,Pm1, Pm2, …, Pmn]

Before applying RS theory, we must cluster the QoS attribute values of each manufacturing cloud service. This paper adopts the k−center clustering method [29]. To reduce the interferences from the inevitable noises in actual manufacturing businesses, the RSs were processed by Ziarko’s rough-set model with variable precision [30]. This method relaxes the approximate boundaries of standard RSs, extending their upper and lower values to a defined precision level β∈[0, 0.5) . When β=0 , the rough-set model with variable precision reduces to the classical rough-set model.

Let the tetrad I=(U, a=C∪D, V, F) be the QoS information decision-making system, and U be the non-empty subset of the finite domain of the instance object, where U={x1, x2, .., xm} represents the m history records. Also let A be a non-empty finite set of QoS attributes, C={a1, a2, .., an−1} be the condition attribute set of QoS, D={an} be the decision attribute set, V be the value domain of the attribute set, and F be the information function of each attribute mapping to the value domain.

Definition 2. Let X and Y be non-empty subsets in a finite domain. If X⊆Y where ⊆ is a partially ordered relationship, then(10) c(X, Y)={1−X∩Y|X|, X≠ϕ0, X=ϕ where |X| is the cardinality of the set, defining the number of objects in the equivalent class. c(X, Y) represents the relative error classification rate of X with respect to Y. If 0≤β<0.5 , then there exists a majority inclusion relationship Y⊇β⁡X⇔c(X, Y)≤β . When c(X, Y)=0 , there exists a standard inclusion relationship from Y to X.

In the information decision-making system of QoS, this paper sets X=U/C={X1, X2, …, X|U/C|} , defining the equivalent class divided by the domain U according to the conditional attribute set C. Meanwhile, Y=U/D={Y1, Y2, …, Y|U/D|} is the equivalent class divided by the domain U according to the decision attribute set D.

Definition 3. Given a β∈[0, 0.5) , the lower approximation β of Yj  with respect to the conditional attribute set C is given by:(11) PosXkβ=∪{Y∈ U/D | c(Xi, Xj)≤β}

The positive region PosXkβ(Yj) of β , also denoted as Cβ(Yj) , is the set of lower approximations {Cβ(Y1), Cβ(Y2), …, Cβ(Y|U/D|)} of all decision classes Yj , representing the probability distribution of U/C in each decision class.

Definition 4. The information amount of a conditional attribute [30] reflects the ability of the attribute to classify data objects. The greater the amount of information, the stronger is the ability to classify objects. The information amount of conditional attribute αk , is calculated as:(12) γ(αk)=1|U|2∑i=1|U/αk||Xi|2, αk∈C where Xi is an equivalence class of the conditional attribute αk in U, and |Xi| is the cardinality of that equivalence class.

Definition 5. The dependency degree [30] of a conditional attribute reflects the dependency degree of the decision-attribute classification of that attribute. The greater the dependency degree, the more critical is the attribute. In this paper, the dependency degree of conditional attribute αk is expressed as:(13) λ(αk)=∑j=1|U/D||Posαkβ(Yj)||U|

The information amount and dependency of conditional attributes represent different aspects of the attribute importance, and must be comprehensively evaluated. In this paper, both aspects are considered equally important. The dependence and information amount of each attribute, calculated by Eqs. (12) and (13) respectively, provide scattered information. This information must then be normalized to give the importance (weight) of the conditional attribute. This article proposes the following normalization method:(14) wk=(γ(ak)+λ(ak))/2∑k=1n−1(γ(ak)+λ(ak))/2, with, ∑k=1n−1wk=1

Let Z={z1, z2, …, zn} be the user-assigned weight of each QoS attribute, and W={w1, w2, …, wn} be the weight calculated by the RS. S={s1, s2, …, sm} is known as the candidate set of entities providing similar services in the manufacturing cloud pool. The QoS attribute value of each candidate service is compiled into the vector Ps={p1, p2, …, pn} . The main steps of the proposed optimization algorithm for manufacturing cloud services are listed below.

Assisted by the telecom cloud computing platform, a prototype of the cloud manufacturing service test platform was designed using jdkl5+MyEclipse2019.4.0 as the Integrated Drive Electronics development environment, Tomcat 8.5 as the server, and MySQL 8.0 and Sybase16.0 as the database and its design tools, respectively. The call of computer nodes to cloud manufacturing service platform services was simulated in Mpiblast, a common distributed application software. Because the description and definition of manufacturing cloud services are highly autonomous and diverse, no common service benchmark database has been recognized by the vast majority of scholars. Therefore, a standard test set is currently lacking. Most service-selection tests of manufacturing cloud services instead use randomly generated test data. The cloud manufacturing service test platform established in this paper includes 1200 manufacturing cloud services for testing.

As an example, we consider production by a door handle manufacturer. The final production process is electroplating the door handle To meet the production requirements, 2,0000 door handles must be electroplated within 20 days. The manufacturer publishes these requirements to the cloud manufacturing service platform. The QoS attribute value of each service is extracted from the service platform QoS database. After receiving a service request, the platform-related module queries the hardware cloud service resource library in the manufacturing cloud pool. Suppose that five similar hardware cloud services in the pool can meet the functional requirements of the manufacturer. The manufacturer bases its decision on the cost, time, reliability, availability, and honesty indicators of the five services. The services and their QoS indicator values are listed in Table 1.

The cost in Table 1 is obtained by summing the processing costs and per-kilometer transportation costs of a single product. The time is the number of days required for processing and transporting the product. Service 2 has a reliability value of 40/42, indicating that this service was called 42 times from the QoS database, and provided 40 normal responses. In a practical scenario, the historical record of Service S2 revealed 42 bids for its service, and 40 completed contracts among those bids. The availability describes the probability of a service’s failure-free time during the effective period of the database record. Honesty (range [0, 5]) is the average evaluation obtained after completing all bid-winning projects. The honesty data are corrected by the QoS monitor in the cloud manufacturing service platform (see Eq. (5)). The cost and time indicators are normalized by Eq. (2), and the other indicators are processed by Eq. (3). The normalized data are listed in Table 2.

Based on the attribute weight calculation of the RS, the weight ratio of each attribute was obtained by analyzing the historical records. The presented experiment accessed the historical records of 5116 manufacturing resources in the manufacturing cloud pool. To demonstrate the reasoning process, we randomly selected eight records as the reasoning sample. The cost and time factors likely differ among the processed products, and the data are not directly comparable. The measurement parameter was the ratio of the bid quote in the record to the bid price of the current business, and the time factor was analyzed in terms of the advance time ratio. For example, service S1 in Table 1 participated in the bidding three days in advance of its manufacturing time (10 d), so its time factor was 0.3. The bidding history of the eight randomly selected manufacturing-resource records are listed in Table 3.

For processing by the RS (which can process only clustered data), the data in the historical database were clustered by the k−center clustering method described in [29]. The number of clusters was set to 5, and the clustered data are shown in Table 4.

According to the data in Table 4, the bid-winning decision-making class was Y1=U/d1={x1, x2, x4, x8} , and the bid-losing decision-making class was Y2=U/d2={x3, x5, x6, x7} . For conditional attribute α5 , the records were divided into equivalent classes U/α5={{x1, x8}{x2, x4, x6}, {x3, x5, x7}} , and the information amount of this conditional attribute was calculated by Eq. (12): γ(α5)=44/128 . The information amounts of the other attributes were similarly obtained as γ(α1)=64/128 , γ(α2)=36/128 , γ(α3)=28/128 , and γ(α4)=68/128 .

Setting β=0.4 in Eq. (11), the β -lower approximate distributions of Y1 and Y2 with respect to conditional attribute α5 were calculated as PosX5β(Y1)={x1, x8} and PosX5β(Y2)={x3, x5, x7} , respectively. Eq. (13) gives the dependency degree of conditional attribute α5 as λ(α5)=10/16 . Similarly, we obtained λ(α1)=16/16 , λ(α2)=6/16 , λ(α3)=4/16 , λ(α4)=14/16 . By Eq. (11), we then obtained w1=0.31 , w2=0.15 , w3=0.10 , w4=0.29 , and w5=0.15 . After analyzing a large number of collected experimental data and considering the suggestions of manufacturing experts, the pre-entered attribute preference weights were set to Z={0.4, 0.1, 0.1, 0.2, 0.2} . Using the data information in Table 2, the similarities of the services were calculated by Algorithm 1, thus obtaining sim(S1)=0.78 , sim(S2)=0.82 , sim(S3)=0.59 , sim(S4)=0.60 , and sim(S5)=0.73 . After ranking the performances of the services as S2>S1>S5>S4>S3 , service s2 was selected as the optimal choice.

If the weights of the QoS indicators are ignored or set to the same value, the data of Table 1 yield sim(S1)=0.73 , sim(S2)=0.63 , sim(S3)=0.45 , sim(S4)=0.55 , sim(S5)=0.76 , and the services are ranked as S5>S1>S2>S4>S3 . Service S5 receives the best performance score because it requires a shorter time and has a high contract completion rate, but its cost is high and its honesty is only average. Overall, S5 gives a balanced performance. When the weights are the weights Z preset by the user, the data of Table 1 give sim(S1)=0.76 , sim(S2)=0.74 , sim(S3)=0.45 , sim(S4)=0.55 , and sim(S5)=0.68 . In this case, the services are ranked S1>S2>S5>S4>S3 . Service S1 outperforms S2 and S5 because the cost factor carries a high weight. The weight of the cost factor was obviously larger in Z than in W (the weight vector mined by the RS), but the weights of availability and honesty were not distinguished in Z and W. When considering the cost factor, the availability weight was slightly higher than that of honesty, consistent with real production, and confirming the higher objectivity of mining the weights from historical data over relying on user evaluations. In the experiment, Service S2 outperformed Service S1 because the price, availability, and honesty weights were all adjusted to make it more appealing.

In traditional information retrieval, the performance of retrieval algorithms is commonly estimated by the precision. In the present experiment, the performance of the selection algorithm was analyzed by a variant of precision called the success rate. With the aim of recommending optimal services to users, the selection success rate is defined as follows:

(15) success rate=number of accepted recommendationstotal number of recommendations×100%

This experiment measured the success rates of three selection methods: the equal-weight selection method (in which the conditional attribute weights are evenly distributed and the QoS value is neither updated nor corrected), the user fixed-weight selection method (in which the weights are preset by the user, and the QoS value is neither updated nor corrected) and the hybrid fixed-weight selection method (the proposed method, in which the user-set weights are combined with the rough-set weights). The QoS value in the hybrid method is adjusted by honesty detection and an information correction mechanism. Simulations were executed on 20 manufacturing operations, each considered by 100 users for service selection. The success rates obtained in the comparison experiment are plotted in Fig. 1.

As shown in Fig. 1, the user fixed-weight selection method was more successful than the equal-weight service-selection method in most cases, indicating the importance of considering the users’ expectation and their preferred evaluation factors. As the equal-weight selection method does not consider the weight differences between the QoS performance indicators, its success rate was relatively low and unstable. Meanwhile, the user fixed-weight selection method is blinded to some extent by the preferences of the weight setter, so its success rate fluctuated greatly. Initially, the hybrid fixed-weight selection method did not outperform the user fixed-weight selection method, because it lacked a historical record in the earliest stages. As the historical database extended, the success rate of the hybrid method gradually increased, indicating that the weights became increasingly more objective. The stability of the success rate also improved over time. The proposed method clearly outperformed the other two methods in terms of selection success rate.

For the dishonesty evaluation, we simulated 100 randomly selected services evaluated by up to 200 dishonest users. The selection success rates of the three selection methods are compared in Fig. 2.

As shown in Fig. 2, increasing the number of dishonest users reduced the success rates of the equal-weight and user fixed-weight selection methods, because neither method adopts an honesty information detection and correction mechanism. The equal-weight selection method was especially sensitive to dishonest evaluations, because its honesty weight was relatively large. The hybrid fixed-weight selection method introduces a QoS information correction mechanism, and collects the user evaluation data by a monitor placed on the cloud manufacturing service platform. The low quality of their evaluation data reduces the credibility of dishonest users; accordingly, these users score very lowly in the QoS honesty calculation and their evaluation information will be filtered out. Therefore, increasing the number of dishonest users barely affected the success rate of the hybrid fixed-weight selection method.