Research on Data Cleaning Algorithm Based on Multi Type Construction Waste

Abstract

Owing to urbanization, the output of construction waste is increasing yearly. Garbage treatment plays a vital role in urban development and construction. The accuracy and integrity of data are important for the implementation of construction waste treatment. Abnormal detection and incomplete filling occur when traditional cleaning algorithms are used. To improve the cleaning of construction waste data, a data cleaning algorithm based on multi-type construction waste was presented in this study. First, a multi-algorithm constraint model was designed to achieve accurate matching between the cleaning content and cleaning model. Thereafter, a natural language data cleaning model was proposed, and the spatial location data were separated from the general data through the content separation mechanism to effectively frame the area to be cleaned. Finally, a time series data cleaning model was constructed. By integrating “check” and “fill”, large-span and large-capacity time series data cleaning was realized. This algorithm was applied to the data collected by the pilot cities, which had precision and recall rates of 93.87% and 97.90% respectively, compared with the traditional algorithm, ultimately exhibiting a certain progressiveness. The algorithm proposed herein can be applied to urban environmental governance. Furthermore, this algorithm can markedly improve the control ability and work efficiency of construction waste treatment, and reduce the restriction of construction waste on the sustainable development of urban environments.

1. Introduction

Owing to continued urban construction in recent years, the output of construction waste has accounted for more than 30% of the total urban waste. As a result, most cities have “garbage besieging the city”. Furthermore, the quality of data collected from construction waste is uneven [1], which markedly affects the development of construction waste treatment work. Therefore, determining how high-quality construction waste data can be obtained is the premise of construction waste treatment [2]. Accordingly, to improve the treatment capacity of construction waste and effectively solve the problem of garbage siege, data cleaning methods should be better studied for the problems existing in the data quality of construction waste.

At present, scholars worldwide have proposed a variety of data cleaning technologies to improve data quality. Schaub et al. used data mining and machine learning algorithms to fill in data and realize the automation of data cleaning, which markedly improved cleaning efficiency. Ma Fengshi et al. used the mathematical statistical method of mathematical extreme value, and judged abnormal data according to the characteristics of abnormal data appearing in the extreme value, which markedly improved the accuracy of outlier detection [3]. Yang Dongqing solved the problem in the process of data conversion, and developed a data cleaning tool [4]. Rubin sought to resolve the problem of missing data using the Bayesian logistic regression method for multiple interpolations, which makes up for the relative subjectivity of the single interpolation method and markedly increases the accuracy of data estimation [5]. Xu Sijia and others solved the problem of missing time series data by using a time series forecasting algorithm [6]. Using the traditional SNM algorithm, Hernandez proposed a multi-pass nearest neighbor (MPN) sorting algorithm, which further improved the efficiency of the weight judgment operation [7]. Huang Darong proposed a data cleaning model based on rough theoretical set, which addressed the complexity in the decision-making process, thereby markedly improving the cleaning efficiency [8]. Qin Hua, Su Yidan, Li Taoshen, et al. proposed a data cleaning method based on the genetic neural network to predict the value to be filled using nonlinear mapping of the neural network and global optimization of the genetic algorithm [9]. Song et al. improved the traditional Simhash algorithm by using the Delphi method and the TF-IDF algorithm, which markedly improved the detection accuracy of similar duplicate records in big data [10]. Zhang et al. constructed a wavelet neural network prediction model based on particle swarm optimization and achieved a significant improvement in the accuracy of data prediction [11].

The above cleaning techniques can only clean a single type of data; however, there are many types of construction waste data, severe false detection, and misfiling after cleaning. Accordingly, this study sought to propose a cleaning algorithm based on multi-type construction waste. First, a multi-algorithm constraint model was constructed to achieve the matching of cleaning content and cleaning algorithm. Thereafter, the natural language data were cleaned using the constructed natural language data cleaning model. Finally, a time series data cleaning model was constructed according to the characteristics of the construction waste time series data, and the detection and filling of time series data were realized. The experimental results show that the algorithm can effectively improve the integrity and accuracy of data cleaning while ensuring cleaning efficiency.

2. Materials and Methods

2.1. Algorithm Flow

The cleaning algorithm is mainly divided into three parts. The first part builds a multi-algorithm constraint model based on the TOPSIS algorithm in multiple attribute decision making (MADM), and takes the construction waste dataset as the input to obtain the cleaning content corresponding to the cleaning model for subsequent data cleaning. The second part of the algorithm is the natural language data cleaning model [12]. First, the content separation mechanism is used to construct the natural language data set to be cleaned. Thereafter, the N-gram algorithm is used to calculate the key value, and the Levenshtein distance algorithm is used to determine the similarity of the data to obtain “dirty data”. The third part of the algorithm is the time series data cleaning model. The model is composed of the Laida criterion and the improved long-short-term memory (LSTM) algorithm [13,14]. Using “check” and “fill”, the model can correct abnormal detection and missing filling of large-span and large-capacity time series data. Finally, clean construction waste data are obtained by sorting and summarizing the cleaning results. The specific process is shown in Figure 1.

2.2. Multi-Algorithm Constrained Model

Data cleaning can accurately find and correct identifiable errors in the data; however, too many data types will have a greater impact on the data cleaning effect. Therefore, solving the problem of data quality reduction caused by the diversity of data types is the main problem of data cleaning technology.

Herein, data cleaning of construction waste data was carried out, and large errors were found in the cleaning results. After analyzing the error of the construction waste data detection results, the difference in data attributes in the construction waste data of the same module (i.e., natural language data and time series data exist at the same time) leads to confusion in the matching of the cleaning model and data during cleaning. As a result, the cleaned data still possess “dirty data”. Therefore, the accurate matching of data cleaning content and data cleaning model is the premise for obtaining high-quality data.

In this study, the TOPSIS algorithm in MADM was employed to construct a multi-algorithm constraint model. Among them, the TOPSIS method, which is a multi-attribute decision-making method for sorting limited alternatives, was proposed by Hwang and Yoon. The principle of this method is to standardize the problem impact index, obtain the distance between the solution and the index’s positive and negative ideal solutions through the distance calculation method, and finally select the solution that is close to the positive ideal solution but far away from the negative ideal solution as the optimal solution. Accordingly, the final match between the data cleaning model and the data cleaning content is achieved by calculating the degree of closeness. The specific technical route is shown in Figure 2.

To obtain the attribute constraint matrix of the construction waste data, the initial construction waste data set was analyzed. By obtaining the column data type of the dataset, a multi-attribute constraint matrix X was constructed, as shown in Equation (1) [15].

$$\mathrm{X}=\left(\begin{array}{ccc}{\mathrm{x}}_{11}& \cdots & {\mathrm{x}}_{1\mathrm{p}}\\ ⋮& \ddots & ⋮\\ {\mathrm{x}}_{\mathrm{n}1}& \cdots & {\mathrm{x}}_{\mathrm{np}}\end{array}\right)$$

(1)

In the formula, each scheme in the scheme set of the multi-attribute decision scheme is ${\mathrm{X}}_{\mathrm{ip}}$.

To solve the problem of different trends of data indicators, the identification and normalization of indicator trends were carried out. There are four indicator trend types, as shown in

Table 1. Examples of the indicator trend type.

Indicator Trend Type	Indicator Features	Example
Very large indicator	The bigger the better	Grades, GDP, etc.
Very small indicator	The smaller the better	Expenses, defective rate, etc.
Intermediate indicator	The closer to a certain value the better	PH value for water quality assessment
Interval indicator	Within a certain range is better	Body temperature, etc.

[16].

After normalization, to realize the unification of the data format, the data was normalized, as shown in Equation (2) [17]. Finally, the normalized matrix Z was obtained, as shown in Equation (3).

$${\mathrm{Z}}_{\mathrm{ij}}=\frac{{\mathrm{x}}_{\mathrm{ij}}}{\sqrt{{\sum }_{\mathrm{i}}{\mathrm{x}}_{\mathrm{ij}}^2}}$$

(2)

In the equation: $\mathrm{i}=1, 2, 3\dots , \mathrm{n}$; j = $1, 2, 3\dots , \mathrm{p}$.

$$\mathrm{Z}=\left(\begin{array}{ccc}{\mathrm{z}}_{11}& \cdots & {\mathrm{z}}_{1\mathrm{p}}\\ ⋮& \ddots & ⋮\\ {\mathrm{z}}_{\mathrm{n}1}& \cdots & {\mathrm{z}}_{\mathrm{np}}\end{array}\right)$$

(3)

Based on the standardized matrix Z, the positive ideal solution ${\mathrm{Z}}^{+}$ and negative ideal solution ${\mathrm{Z}}^{-}$ of each index were obtained, as shown in Equations (4) and (5):

$${\mathrm{Z}}^{+}=\left\{\max ({\mathrm{z}}_{\mathrm{ij}}|\mathrm{i}=1,2,\dots \mathrm{n})\right\}$$

(4)

$${\mathrm{Z}}^{-}=\left\{\max \left({\mathrm{z}}_{\mathrm{ij}}\right)|\mathrm{i}=1,2,\dots \mathrm{n}\right\}$$

(5)

Finally, the Manhattan distance was used to calculate the distance between the object and the optimal value and the worst value [18].

According to the obtained distance, the corresponding closeness and index ranking were obtained, as shown in Equation (6).

$${\mathrm{C}}_{\mathrm{i}}=\frac{|{\mathrm{z}}_{\mathrm{ij}}-{\mathrm{z}}_{\mathrm{j}}^{-}|}{|{\mathrm{z}}_{\mathrm{ij}}-{\mathrm{z}}_{\mathrm{j}}^{+}|+|{\mathrm{z}}_{\mathrm{ij}}-{\mathrm{z}}_{\mathrm{j}}^{-}|}$$

(6)

In the equation, ${\mathrm{C}}_{\mathrm{i}}\in \left[0,1\right]$ is the matching degree between the cleaning model and the data; the smaller the value, the lower the match between the cleaning model and the data.

2.3. Natural Language Data Cleaning Model

At present, N-gram is the main algorithm for cleaning natural language data. The N-gram algorithm is a language model for the continuous recognition of a large vocabulary. The core idea of this algorithm is to assume that words in the content have a certain probability of occurrence. Thus, the preprocessed content is segmented, sliding cutting is performed according to a fixed size window, and a sequence of byte fragments of length N is fabricated. The frequency of occurrence of each byte fragment sequence is counted and filtered according to the preset threshold, and the probability of occurrence of the entire record is obtained by counting the probability of occurrence of byte fragments. However, as the natural language data of construction waste contain spatial location data and main cleaning data, the original repetition of spatial location data leads to the inaccurate similarity of the final data [19]. To solve this problem, a natural language data cleaning model based on the N-Gran algorithm was proposed in this study. Through the content separation mechanism designed in the model, the integrity and accuracy of the natural language cleaning of construction waste were improved. Figure 3 shows the overall technical flowchart of the natural language data cleaning model:

First, the construction waste data were analyzed, which revealed spatial location data in the data, and potential repeats of spatial location data in different data, resulting in incorrect similarity during cleaning [20]. Therefore, the content separation mechanism was adopted to obtain the main cleaning data set [21].

After obtaining the main cleaning data set, to ensure the validity of the data, the occurrence probability distribution between the words ${\mathsf{\omega }}_{\mathrm{i}}\left(\mathrm{i}\in \left(0,\mathrm{t}\right)\right)$ in the construction waste data record must be assumed to obey the n-1 order Markov model.

To markedly improve the data cleaning effect, the N value must be selected [22]. As the number of times of construction waste natural language data vocabulary composition is two2, the N value of two can maximize detection accuracy. The probability of occurrence of construction waste data records is set as $\mathrm{P}\left({\mathsf{\omega }}_{\mathrm{i}}\right)$ [23,24,25,26,27]. The equation is shown in (7):

$$\begin{gathered} \mathrm{p}\left(\mathrm{S}\right)=\mathrm{p}({\mathsf{\omega }}_1,{\mathsf{\omega }}_2,{\mathsf{\omega }}_3,\dots {\mathsf{\omega }}_{\mathrm{n}}) \\ =\mathrm{p}({\mathsf{\omega }}_1\left)\mathrm{P}\right({\mathsf{\omega }}_1|{\mathsf{\omega }}_2\left)\mathrm{p}\right({\mathsf{\omega }}_3|{\mathsf{\omega }}_1{\mathsf{\omega }}_2)\dots \mathrm{p}({\mathsf{\omega }}_{\mathrm{n}}|{\mathsf{\omega }}_1{\mathsf{\omega }}_2,\dots {\mathsf{\omega }}_{\mathrm{n}-1}) \\ \approx \mathrm{p}({\mathsf{\omega }}_1\left)\mathrm{p}\right({\mathsf{\omega }}_1|{\mathsf{\omega }}_2\left)\mathrm{p}\right({\mathsf{\omega }}_3|{\mathsf{\omega }}_2)\dots \mathrm{p}({\mathsf{\omega }}_{\mathrm{n}}\left|{\mathsf{\omega }}_{\mathrm{n}-1}\right) \end{gathered}$$

(7)

To calculate the repeatability of the data, similarity detection was performed on the data. The similarity detection algorithm adopts the Levenshtein Distance. The Levenshtein Distance was proposed by Levenshtein et al., and is a distance-based similar duplicate record matching algorithm. The core idea of this algorithm is to compare and analyze the two records, A and B, to find the difference between the two data characters. Assuming that the characters can be converted, the insertion, deletion, and replacement operations that must go from character A to character B must be recorded. The number of operations per operation is set to 1, and the number of operations required to go from A to B, which is called the edit distance, is counted. The fewer operations required, the higher the similarity. Accordingly, the similarity between the data was obtained by quantifying the similarity of the strings.

2.4. Time Series Data Cleaning Model

As the construction waste data are time series data, they are comparable to the same period data. Therefore, commonly used detection algorithms are used to construct the time series data for construction waste. For example, the detection algorithm based on the DBScan cluster is mainly based on the form of density detection, which cannot effectively utilize the time series attributes of the data. Furthermore, the construction waste data are large-span time series data [28]. Thus, the detection results of the data are poor. The statistical-based detection algorithm can have a better data detection effect through the comparative analysis of the time series [29,30]. Among them, the Pauta criterion is the most common and convenient method based on statistical detection algorithms. This method primarily seeks to determine whether all datasets differ from the dataset mean by more than three times the standard deviation. As this method is simple and convenient, and has the characteristics of better experimental results when the amount of data is large, the Pauta criterion method in statistics was adopted to detect and mark the time series data of the construction waste.

To improve the effectiveness of the filling results of the construction waste time series data, the data were divided according to the construction waste time series data attributes. The Laida criterion was used for the detection, and the sliding window was used for marking [31,32]. To solve the problem of long data time series of construction waste spatiotemporal data, the method of “detection first and then filling” was adopted for cleaning. As shown in Equation (8), the construction waste time series data set was pre-filled.

$${ \mathrm{F}}_{\mathrm{x}}=\frac{1}{3}{\sum }_{\mathrm{j}=\mathrm{i}}^{\mathrm{i}+3}{\mathrm{f}}_{\mathrm{j}}$$

(8)

To achieve accurate data filling, the LSTM algorithm was used for detection. LSTM is a recurrent neural network that was proposed by Hochreiter in 1997. This algorithm is a special network form of RNN that converts the tanh layer in the traditional recurrent neural network to include storage units and gates, modifying the structure of the memory cell, thereby overcoming the traditional RNN gradient diffusion and explosion problems. Its main process is described below.

First, the construction waste information that can be transmitted to the memory unit, which is controlled by the sigmoid function in the forget gate, is determined [33]. A value between 0 and 1 to ${\mathrm{f}}_{\mathrm{t}}$ is assigned based on the output ${\mathrm{h}}_{\mathrm{t}-1}$ at time t-1 and the input ${\mathrm{x}}_{\mathrm{t}}$ at time t. ${\mathrm{f}}_{\mathrm{t}}$ is mainly used to determine whether to fully or partially transmit the construction waste information learned at the previous instance [34]. The sigmoid function equation is shown in (9)

$$sigmoid\left(\mathrm{x}\right)=\frac{1}{1+{\mathrm{e}}^{-\mathrm{x}}}$$

(9)

The ${\mathrm{f}}_{\mathrm{t}}$ equation is shown in (10):

$${\mathrm{f}}_{\mathrm{t}} =\mathsf{\epsilon }({\mathrm{W}}_{\mathrm{f}}.{ \mathrm{x}}_{\mathrm{t}}+{\mathrm{W}}_{\mathrm{f}}.{ \mathrm{h}}_{\mathrm{t}-1}+{\mathrm{b}}_{\mathrm{f}})$$

(10)

where ${\mathrm{f}}_{\mathrm{t}}$ represents the forget gate, with a value range of [0, 1]; the symbol represents the sigmoid function; and ${\mathrm{f}}_{\mathrm{t}}$∈${\mathrm{R}}_{\mathrm{H}\ast \mathrm{d}}$, ${\mathrm{h}}_{\mathrm{t}}$∈${\mathrm{R}}_{\mathrm{H}}$, ${\mathrm{b}}_{\mathrm{f}}$∈${\mathrm{R}}_{\mathrm{H}}$; 1 indicates “completely reserved”, 0 indicates “totally abandoned”.

The next step involves the generation of the new construction waste data required for the update [35,36,37]. This step mainly depends on two aspects: the input gate uses the sigmoid function to decide which construction waste data must be updated; the new candidate value ${\mathrm{D}}_{\mathrm{t}}$ generated by the tanh layer is stored in the cell state [38], as shown in Equations (11)–(13).

$${\mathrm{i}}_{\mathrm{t}}=\mathsf{\epsilon }({\mathrm{W}}_{\mathrm{i}}.{\mathrm{x}}_{\mathrm{t}}+{\mathrm{W}}_{\mathrm{i}}.{ \mathrm{h}}_{\mathrm{t}-1}+{\mathrm{b}}_{\mathrm{i}})$$

(11)

$${\mathrm{D}}_{\mathrm{t}}=tanh({\mathrm{W}}_{\mathrm{c}}.{\mathrm{x}}_{\mathrm{t}}+{\mathrm{W}}_{\mathrm{c}}.{ \mathrm{h}}_{\mathrm{t}-1}+{\mathrm{b}}_{\mathrm{c}})$$

(12)

$$Tanh\left(\mathrm{x}\right)=\frac{{\mathrm{e}}^{\mathrm{x}}-{\mathrm{e}}^{-\mathrm{x}}}{{\mathrm{e}}^{\mathrm{x}}+{\mathrm{e}}^{-\mathrm{x}}}$$

(13)

where ${\mathrm{i}}_{\mathrm{t}}$ represents the input gate, and the value range is [0, 1].

An update of the old unit must be performed. First, the old cell state must be multiplied by ${\mathrm{f}}_{\mathrm{t}}$ to forget the unwanted construction waste data. Thereafter, ${\mathrm{i}}_{\mathrm{t}}$*${\mathrm{D}}_{\mathrm{t}}$ must be added to the result to obtain the candidate value ${\mathrm{C}}_{\mathrm{t}}$ [39]. The main Equation (14) is shown below:

$${\mathrm{C}}_{\mathrm{t}}={\mathrm{f}}_{\mathrm{t}}.{\mathrm{C}}_{\mathrm{t}-1}+{\mathrm{i}}_{\mathrm{t}}.{\mathrm{D}}_{\mathrm{t}}$$

(14)

Finally, the output of the model is determined, the initial output through a sigmoid layer is obtained, the tanh layer is used to perform the shrinking operation, the sigmoid tanh in pairs is multiplied, and a clean data set is obtained [40,41,42,43]. The ${\mathrm{o}}_{\mathrm{t}}$ and ${\mathrm{h}}_{\mathrm{t}}$ equations (i.e., (15) and (16)) are shown below:

$${\mathrm{o}}_{\mathrm{t}}=\mathsf{\epsilon }({\mathrm{W}}_0.{\mathrm{x}}_{\mathrm{t}}+{\mathrm{W}}_0.{\mathrm{h}}_{\mathrm{t}-1}+{\mathrm{b}}_0)$$

(15)

$${\mathrm{h}}_{\mathrm{t}}={\mathrm{o}}_{\mathrm{t}}.tanh\left({\mathrm{C}}_{\mathrm{t}}\right)$$

(16)

where ${\mathrm{o}}_{\mathrm{t}}$ is the output gate, and its value range is “(0, 1)”.

3. Results and Evaluation

3.1. Experimental Data

Construction waste is the collective term for muck, waste concrete, waste masonry, and other wastes generated during production activities related to demolition, construction, decoration, repair, and other construction activities. According to the source classification, construction waste can be divided into engineering muck, engineering mud, engineering waste, demolition waste and decoration waste, etc. Based on the composition, construction waste can be divided into muck, concrete block, crushed stone, brick and tile fragments, waste mortar, mud, asphalt blocks, waste plastics, waste metals, waste bamboo and wood, etc. The experimental data include all types of construction waste, and the data quality of the overall construction waste data is shown to be improved.

The main source of the data in this study is the construction waste management platform. Based on data collection from 35 pilot cities, a total of 5961 pieces of data were collected. According to the classification of the platform function modules, the data were divided into nine categories, as shown in Figure 4.

(1) Inventory inspection data: Statistics on the amount of data retained in construction waste. The data mainly include total garbage stock, engineering dregs, engineering mud, etc. (2) Meeting minutes data: relevant meeting minutes on construction waste management. The data mainly include the theme of the conference, the participants, and an overview of the content of the conference. (3) Leading group data: information on the leading groups of each pilot city. The data mainly include city, contact person, phone number, etc. (4) Implementation plan data: document records issued by construction waste management. The data mainly include city, time, program type, etc. (5) Landfill data: statistical data of landfill information. The data mainly include city, time, and project name. (6) Rules and regulations data: statistical data of the construction waste management work rules and regulations. The data mainly include city, time, number, etc. (7) Special planning data: statistical data of construction waste treatment planning documents. The main contents include city, time, planning type, etc. (8) Resource facility data: statistical data for information related to the resource facility. The data mainly include city, time, project name, etc. (9) Overall progress data: statistical data on the completion of construction waste treatment work. The data mainly include city, time, and work ratio.

3.2. Experimental Results

The algorithm proposed in this study detects the construction waste test data; the detection results are shown in

Table 2. The algorithm detection results.

	Inventory Check	Meeting Minutes	Leadership Team	Implementation	Landfill	Regulatory	Special Planning	Resource Facility	Overall Progress	Total
The amount of dirty data in the natural language data	85	70	15	112	213	102	255	159	51	1062
Number of errors in the natural language data	9	7	2	13	23	13	29	22	4	122
Correct number of natural language data	76	63	13	99	190	89	226	137	47	940
Missing natural language data	10	8	2	12	13	5	12	11	5	78
Natural language data accuracy	89.41%	90.00%	86.66%	88.39%	89.20%	87.25%	88.62%	86.16%	92.16%	88.65%
Natural language data recall	88.97%	95.97%	88.89%	89.68%	93.97%	94.88%	94.98%	92.86%	90.77%	92.33%
Amount of dirty data in the time series data model	82	86	32	187	278	115	411	96	48	1335
Number of error bars in the time series data model	8	8	2	23	33	10	52	11	4	151
The correct number of time series data models	74	78	30	164	245	105	359	85	44	1184
Time series data model missed check	6	6	3	12	19	8	27	6	4	91
Time series data model precision	89.11%	89.51%	89.96%	87.22%	87.63%	89.75%	85.22%	88.21%	91.51%	88.68%
Time series data model recall	92.96%	93.11%	92.92%	93.52%	90.54%	94.23%	92.48%	93.54%	92.46%	92.86%
The amount of dirty data in this algorithm	154	102	45	227	344	174	456	283	61	1846
The number of errors in the algorithm in this study	11	6	2	19	19	8	23	20	5	113
The correct number of algorithms in this study	143	96	43	208	325	166	433	263	56	1733
The algorithm in this study misses the check	2	3	1	3	8	6	8	4	2	37
The precision of the algorithm in this study	92.99%	94.52%	96.12%	91.48%	94.32%	95.23%	94.84%	93.25%	92.18%	93.88%
The recall rate of the algorithm in this study	98.81%	97.63%	98.71%	98.81%	97.18%	96.77%	97.96%	98.65%	96.93%	97.90%

.

Nine categories of data were selected as the detection data source to test the algorithm proposed in this study, the natural language data cleaning model, and the time series data cleaning model. Among them, the algorithm detected 1846 pieces of dirty data as a whole, including 1733 pieces of correct dirty data, 113 pieces of incorrect dirty data, and 37 pieces of missing data. Herein, the recall rate and precision rate were used to represent the accuracy of the algorithm. The precision rate was 93.88% and the recall rate was 97.90%.

The amount of dirty data in each module of the natural language data cleaning model was 1062, of which 940 were correct for the detection of dirty data, 122 were incorrect for the detection of dirty data, and 78 were missing overall. Herein, the recall rate and precision rate were used to represent the accuracy of the algorithm. The precision rate was 88.65% and the recall rate was 92.33%.

The number of dirty data detected by each module of the time series data cleaning model was 1335, of which 1184 were correct for the detection of dirty data, 151 were incorrect for the detection of dirty data, and 91 were missing overall. Herein, the recall rate and precision rate were used to represent the accuracy of the algorithm. The precision rate was 88.68% and the recall rate was 92.86%.

3.3. Result Analysis

To further reflect the advanced nature of the algorithm, the cleaning results of the algorithm were compared with those of a single cleaning model. The recall rate and precision rate were used to evaluate the cleaning accuracy of the construction waste natural language, and the average root mean square error (RMSE) and average MAPE indicators were used to evaluate the filling accuracy of the construction waste time series data of this algorithm.

3.3.1. Construction Waste Natural Language Data

The landfill data were employed as the experimental data, and 800 pieces of data were selected as the sample data. The similarity was calculated using the Smith–Waterman (S-W) algorithm and the Levenshtein distance algorithm, and the corresponding recall and precision were calculated. The comparison results are shown in

Table 3. Accuracy comparison results.

Cleaning Technology	100	200	300	400	500	600	700	800
Levenshtein Distance	0.667	0.692	0.716	0.73	0.747	0.754	0.768	0.78
S-W	0.7	0.739	0.757	0.775	0.784	0.788	0.796	0.807
This paper proposes a cleaning algorithm	0.875	0.889	0.894	0.919	0.929	0.945	0.957	0.891

.

In terms of precision, the cleaning accuracy increased from 0.875 to the highest of 0.961. The accuracy of the Levenshtein distance algorithm was as low as 0.667 and as high as 0.78. The precision of the Smith–Waterman algorithm ranged from 0.7 (the lowest) to 0.807 (the highest). The maximum difference between the accuracy of the cleaning algorithm and the Levenshtein distance algorithm was 0.208, and the minimum difference was 0.178. The maximum difference with the accuracy of the Smith–Waterman algorithm was 0.175, and the minimum difference was 0.137. The comparison results are shown in

Table 4. The full comparison results.

Cleaning Technology	100	200	300	400	500	600	700	800
Levenshtein Distance	0.687	0.711	0.715	0.727	0.735	0.742	0.759	0.764
S-W	0.727	0.731	0.738	0.745	0.757	0.762	0.775	0.782
This paper proposes a cleaning algorithm	0.842	0.857	0.864	0.873	0.883	0.895	0.911	0.911

.

In terms of recall, the cleaning accuracy increased from 0.8742 to the highest of 0.932. The accuracy of the Levenshtein distance algorithm ranged from 0.687 (the lowest) to 0.764 (the highest). The precision of the Smith–Waterman algorithm ranged from 0.727 (the lowest) to 0.782 (the highest).

The maximum difference between the accuracy of the cleaning algorithm proposed in this study and the Levenshtein distance was 0.168, the minimum difference was 0.146, the maximum difference between the precision of the Smith–Waterman algorithm was 0.15, and the minimum difference was 0.115.

Figure 5 and Figure 6 show the comparison of recall and precision changes. In this study, the method of sorting and then cleaning was adopted, and the accuracy of sorting was followed by further matching, which could markedly improve the accuracy of data cleaning. It not only fully meets the requirements of data cleaning technology, but also continues to increase with the continuous increase in data volume, and finally approaches a threshold.

3.3.2. Construction Waste Data Time Series Data

Using RMSE and MAPE as evaluation indicators, the data cleaning results of the time series data cleaning model proposed herein and the traditional LSTM data cleaning model were compared. The indicator value is inversely proportional to the precision of the data. As shown in

Table 5. Accuracy comparison data table.

Evaluation Parameters	Total Waste Stock		Engineering Muck		Engineering Mud		Engineering Waste		Demolition Garbage
	Algorithm of This Paper	LSTM	Algorithm of This Paper	LSTM	Algorithm of This Paper	LSTM	Algorithm of This Paper	LSTM	Algorithm of This Paper	LSTM
RMSE	11.59	20.85	13.24	25.85	12.59	26.84	11.51	20.86	9.597	18.85
MAP	8.84	16.24	9.84	18.24	10.86	19.74	8.94	16.24	6.84	14.24

, five types of data, including total waste, engineering muck, engineering mud, engineering waste, and demolition waste were tested.

Among the RMSE indicators, the engineering mud data had the highest improvement, with a decrease of 14.25, while the demolition garbage data had the lowest improvement, with a decrease of 9.253. Among the MAPE indicators, the engineering mud data had the highest improvement, with a decrease of 8.88, while the engineering waste data had the lowest improvement, with a decrease of 7.3. In each module, the time series data cleaning model proposed herein markedly improved the accuracy of RMSE and MAPE compared with that of the traditional LSTM data cleaning model.

The accuracy was compared with the average RMSE and MAPE. As shown in

Table 6. Comparison Table of Average Precision.

Evaluation Parameters	Fill Method
Average RMSE	Algorithm of this paper: 11.708	LSTM: 22.653
Average MAPE	Algorithm of this paper: 9.064%	LSTM: 16.942%

, the improved LSTM filling model had a higher accuracy.

As shown in Figure 7, the average accuracy evaluation of the model was compared. The improved LSTM filling model proposed herein is more conducive to the improvement of data quality than that of the traditional LSTM filling model.

3.3.3. Comparative Analysis of the Cleaning Effect

To reflect the overall improvement of the effect of the algorithm, the recall and precision rates of a single algorithm, a single cleaning model and the algorithm proposed herein are outlined, as shown in

Table 7. Average accuracy comparison data sheet.

N-gram Algorithm	Natural Language Data Cleaning Model	Multi-Algorithm Constraint Model + Time Series Data Cleaning Model	Natural Language Precision	Natural Language Recall	Time Series Data Precision	Time Series Data Recall
	×	×	81.73%	84.56%	N/A	N/A
	√	×	88.65%	92.33%	N/A	N/A
	√	√	94.87%	97.90%	97.13%	98.67%

.

Based on the N-gram algorithm, the natural language precision rate was 81.73%, and the natural language recall rate was 84.56%; the precision rate of the time series data was empty, and the recall rate of time series data was empty. After data cleaning based on the N-gram algorithm + natural language data cleaning model, the precision rate of natural language increased by 6.92% to 88.65%, and the recall rate of natural language increased by 7.77% to 92.33%. The precision of the time series data was empty, and the recall of time series data was empty. Based on the N-gram algorithm + natural language data cleaning model + multi-algorithm constraint model + time series data cleaning model, the natural language precision rate of the algorithm increased by 13.14% to 94.87%; the natural language recall rate increased by 13.34%, reaching 97.90%; the precision of the time series data was 97.13%, and the recall rate of the time series data was 98.67%.

4. Discussion

A multi-type construction waste data cleaning algorithm was proposed in this study. According to the diversity of the construction waste data, a natural language data cleaning model and a time series data cleaning model were designed to achieve accurate and complete cleaning of multi-type construction waste data. Comparative experiments with various algorithms revealed that our algorithm has certain advantages. Below, the overall structure and advantages of the algorithm, including the design and advantages of the cleaning model, the design and advantages of the multi-constraint combinatorial model, and the limitations of the algorithm, are discussed.

4.1. Cleaning Model Design and Advantages

The concept of data cleaning has a long history. For data cleaning research, scholars from various countries have proposed different cleaning algorithms. According to the cleaning content, these algorithms can be divided into time series data cleaning algorithms and natural language data cleaning algorithms. Among them, the algorithms for time series data cleaning mainly include the K-means clustering algorithm, Bayesian network, density-based local outlier factor detection algorithm, Detection algorithm based on data flow, etc. [44,45,46]. In contrast, the algorithms for natural language data cleaning mainly include the SNM algorithm, MPN algorithm, Distance based algorithm for similarity detection, etc. [47,48]. At present, most algorithms only perform data cleaning for a single data set; however, construction waste data are diverse, and the data include natural language data and numerical data, which leads to a poor cleaning effect of a single algorithm for construction waste data. According to the diversity of construction waste data, a natural language data cleaning model and a time series data cleaning model were proposed herein to ensure the accuracy and integrity of data cleaning.

4.2. Design and Advantages of the Multi-Algorithm Constrained Models

Construction waste data are diverse and mixed. Thus, in the same type of data, the data attributes of each piece of data include natural language data and time series data, and the proportion of natural language data and time series data in each piece of data is different. At present, there are many algorithms for data cleaning, such as the improved PNRS algorithm, Original operator cleaning algorithm based on similar connection, Cleaning algorithm based on support vector machine (SVM), etc. [49,50,51]. However, when data cleaning is performed, the cleaning model cannot be selected, which is a problem. Accordingly, this study sought to propose a multi-algorithm constraint model, which mainly uses the idea of MADM to achieve accurate matching between the data cleaning model and the data cleaning range, and finally realize the “one-to-one” cleaning mode, thereby markedly improving the construction waste data cleaning accuracy.

4.3. Limitations of the Algorithm

Based on experiments, the algorithm proposed herein is effective; however, some problems still exist. For example, although the algorithm cleans the natural language data and time series data when cleaning the construction waste data, the spatial location data existing in the natural language cannot be cleaned. Thus, further research on the spatial location data should be carried out in the future. Moreover, as the experimental data were collected at a certain period of time, there is still room for further expansion in the number of data sets. The effectiveness of the algorithm can be further verified by expanding the experimental data sets in the future.

5. Conclusions

Owing to the multi-type characteristics of construction waste data, this study sought to propose a data cleaning algorithm based on multi-type construction waste. The algorithm is mainly composed of three parts: the multi-algorithm constraint model, the natural language data cleaning model, and the time series data cleaning model. Based on traditional data cleaning, multiple cleaning models were combined, which effectively alleviated the poor cleaning effect caused by the many types of construction waste data. According to the test results of the experimental data set, the precision and recall of the algorithm reached 93.87% and 97.90%, respectively, indicating its ability to clean dirty data relatively completely and accurately. From the comparison experiment, the algorithm herein has remarkable improvement in precision and recall compared to that of a single model. Compared with the natural language data cleaning model, the precision and recall increased by 5.23% and 5.57%, respectively. Compared with the time series data cleaning model, the patent checking and recall rates increased by 5.2% and 5.04%, respectively.

In summary, the proposed algorithm has great improvement in cleaning integrity and accuracy, and its specific advantages are as follows.

The multi-algorithm constraint model realizes the matching between the cleaning algorithm and the cleaning content, thereby accurately matching the cleaning model and the cleaning content, and effectively improving the effect of data cleaning.

For traditional natural language data cleaning, combined with the characteristics of construction waste data, the spatial location data and general data are isolated, which remarkably improves the cleaning efficiency.

Based on the large vacancy of the time series data for construction waste data, the pre-fill method is used to preprocess the data set, which effectively improves the accuracy of data filling.

Although the experiments demonstrated that our algorithm has certain effectiveness, it still has some problems. For example, through further analysis of the construction waste data, some dirty data were found in the spatial location data. In the follow-up work, this problem will be evaluated according to the dirty data processing technology of spatial location information. Due to the increasing number of multi-type data, single data cannot meet the requirements of the data cleaning technology, and multi-type algorithms will be the main cleaning methods for future data cleaning algorithms.