Predict the priority of end-users’ maintenance requests and the required technical staff through LSTM and Bi-LSTM recurrent neural networks Open Access

Information technology (IT) systems are assuming a key role in facility management (FM), as they can increase the efficiency of FM processes by acquiring, organizing and exploiting accurate and reliable information during the whole building life cycle (Ahmed et al., 2017; Ebbesen, 2016; Gunay et al., 2019a; Maslesa and Jensen, 2019). The introduction of IT systems into FM started more than two decades ago (Maslesa and Jensen, 2019), focusing on operational and tactical levels to solve specific FM tasks, but IT became essential also to support strategic FM decisions. IT systems and technologies in FM can be conceptually organized into seven categories (Ebbesen, 2016; Maslesa and Jensen, 2019): data repositories/containers/viewers, interoperability, workflow systems, facilities intelligence systems, sensor, mobile and real-time location systems, field data capture systems and communication systems. In this overall scheme, workflow systems are essential as they allow digitizing workflows related to FM by using different tools: computerized maintenance management system (CMMS), computer-aided facilities management (CAFM), integrated workplace management system (IWMS) and enterprise resource planning (ERP).

Among the workflow systems, CMMS is mainly devoted to the management of information on maintenance issues (Maslesa and Jensen, 2019), as it contains the history of the maintenance activities, the systems requiring corrective actions, the priority assigned in the past to specific issues and the organization involved in the maintenance activity (Bortolini and Forcada, 2020; Marocco and Garofolo, 2021; Pishdad-Bozorgi et al., 2018). Then, a CMMS is a rich source of information, including work orders (WOs), asset histories, parts inventories, maintenance personnel management and the calculation of maintenance metrics. It can be actively used to both manage the daily FM activity and support strategic decisions (Bortolini and Forcada, 2018; Gunay et al., 2019a; Maslesa and Jensen, 2019; Pärn et al., 2017).

However, a CMMS is often configured as a large, unstructured and amorphous data set (Gunay et al., 2019a). It mainly contains textual descriptions, then text mining algorithms can be applied to extract useful information (Bortolini and Forcada, 2020). As a consequence, part of the FM research focused on the development of methods to extract textual information from CMMS (Baek et al., 2021) and to develop predictive methods based on machine learning (ML) (D’Orazio et al., 2022; Marocco and Garofolo, 2021; Mo et al., 2020), to improve the maintenance process. Text-mining methods have been recently proposed to analyze CMMS databases and extract information in relation to different maintenance tasks, e.g. for the analysis of fault frequency in building components and systems (Gunay et al., 2019b), or to evaluate where the fault happened and to identify the most problematic building elements and systems (Marocco and Garofolo, 2021). Furthermore, the research investigated the automatic detection of anomalies using neural-based ML methods (Du et al., 2017) and of WOs priority (Bortolini and Forcada, 2020). The automation of the attribution process through ML, i.e. scheduling and staff assignment, has been also investigated (El-Dash, 2007; Gutjahr and Reiter, 2010; Wu and Sun, 2006) and a set of classifier models evaluated to predict subcategories (McArthur et al., 2018). In this sense, recent research provided a prediction model that automatically assigns staff in response to unstructured textual requests by applying three different ML methods (Mo et al., 2020): logistic regression, naïve Bayes and support vector machine (Baek et al., 2021; Çınar et al., 2020; McArthur et al., 2018; Mo et al., 2020; Žižka et al., 2019). Sentiment analysis methods were also used to develop predictive methods for the maintenance priority assignment (D’Orazio et al., 2022) and to improve the management of contemporary on-demand corrective actions (Chanter and Swallow, 2007) in large and complex organizations.

Several recent works in the field address relevant contexts such as health-care facilities and hospitals, in view of the critical implications of faults in their daily operation, also based on ML methods for priority assessment (Alanazi, 2022). In particular, priority assessment for the maintenance of medical equipment has been analyzed by different works (Hutagalung and Hasibuan, 2019; Shamayleh et al., 2023; Zamzam et al., 2021), while more limited attention is paid to the other work categories connected to general building operation.

Recently, a particular type of recurrent neural network (RNN), named long short-term memory (LSTM RNN), received particular attention due to its ability to acquire information from sequences of data (Siami-Namini et al., 2019), instead of single values, also in different FM fields (Jung et al., 2021; Luo and Oyedele, 2021). (Unidirectional) LSTM and bidirectional LSTM (Bi-LSTM) RNN have been recently applied to analyze the feasibility of anomaly analysis using deep learning, based on data collected by IoT devices on the conditions of indoor spaces (Jung et al., 2021) and to improve predictive maintenance approaches, by predicting failures events (Tessoni and Amoretti, 2022). These RNNs mimic the human brain; hence, they have been applied also to natural language. Indeed, RNN architecture “reflects a characteristic of natural language that a sentence consists of successive words” (Baek et al., 2021). In fact, the meaning of a sentence is also related to the phrase construction and the position of the words in the sentence.

However, despite the rapid increase of the works in this field and the promising results obtained through ML methods, LSTM and Bi-LSTM RNN in particular, further research is still needed. The application of ML methods to the natural language is a complex task (D’Orazio et al., 2022; Gunay et al., 2019a; Kim et al., 2022; McArthur et al., 2018), requiring alone or combined pre-processing actions (i.e. stopword remotion, symbol and punctuation filtering, stemming, lemmatization, etc.) that can have a strong impact on the accuracy of the results (Mo et al., 2020).

This paper provides a further contribution in this field, applying LSTM and Bi-LSTM RNN to predict the priority of corrective maintenance tasks and the related technical staff to assign, directly from the written natural language in the end-users’ requests (i.e. the text of e-mails). The paper then aims to develop a predictive approach based on RNN to be applied to end-users’ maintenance requests continuously collected by facility managers in CMMSs. It additionally investigates and then compares the RNN reliability on a selected case study (a hospital, considered a particularly sensitive context).

According to a data-driven approach to FM (Gunay et al., 2019a; Ma et al., 2020), the work also aims at using the maintenance requests data set to have a clearer overview of corrective maintenance needs and related performances in terms of work category, assigned priority, target time and effective intervention time, to verify the most relevant elements in the building and how the current FM relates to the requests.

The work entailed three main steps addressing the main research goals highlighted before, as summarized in the graphical framework in Figure 1.

The first phase involved statistical analysis of the original data set on maintenance requests, according to a data-driven approach to FM (Section 4.2). This step allowed detecting the statistical distributions of the maintenance requests by work category, assigned priority (in terms of intervention target time) and effective time to reply and complete the related maintenance interventions. Correlation analyses were also performed to check the correspondence between assigned priority and effective intervention time.

The second phase concerned text pre-processing of the end-users’ maintenance requests, to arrange textual sequences into the data set for LSTM and Bi-LSTM RNN training (Section 4.3). In particular, pre-filtering actions were performed to prepare the texts for the successive training, testing and validation actions on both LSTM and Bi-LSTM RNN. All these actions were performed through MATLAB 2021b (https://it.mathworks.com/support/requirements/previous-releases.html; last access: 24/06/2022).

Finally, the third step entailed training, testing and validation of the Bi-LSTM and LSTM RNN, to predict the priority and the required staff for each end-user’s maintenance request (Section 4.4).

The data set is composed of end-users’ maintenance requests collected for 14 months (from January 2016 to February 2017) through a CMMS in a hospital (of about 300 beds and more than 2,000 staff people). The FM is provided by an external contractor. CMMS data mainly referred to the main hospital site comprising 14 different buildings/areas. Data not referring to the main site were discarded, obtaining 10253 unique records in the CMMS data set.

For each end-user’s request, the data set comprises a brief textual description of the maintenance issue collected via e-mail, data concerning its location (building, floor, room), the priority assigned by the technical FM staff (that is expressed in terms of target intervention time), the work category, the time to reply and the time to close the intervention (described in terms of opening date, response date and closing date) in respect to the target time. End-users are represented by the hospital staff.

The data set considers:

• six work categories (corresponding to different contracts), i.e. electrical, fabrics, management, mechanical, minor new works, waste disposal; and

• five main classes concerning the target intervention time, which expresses the maximum allowable time to complete the intervention when a request has been submitted, thus denoting the request priority. These classes are 30 min, 60 min, 12 h (720 min), 24 h (1,440 min) and “no time” (programmable interventions).

In particular, concerning the target intervention time, the label “no time” has been transformed into the numerical value corresponding to 1 week (10,080 min) to grant homogeneity and numerical consistency with the other classes and perform successive analyses. Timestamps of the time necessary to reply to each end-user’s request after the necessary checks and to perform the intervention have been also recorded in the original data set. The differences between timestamps were used to calculate the time (minutes) necessary to reply to each end-user’s request with respect to the target time, to solve the issue and to close the ticket.

The original data set was characterized by detecting statistical distributions of work categories and target intervention time (denoting assigned priority). Preliminarily, a Shapiro–Wilk test was performed to check the normality of the sample (Shapiro and Wilk, 1965). Then correlation analyses were performed, mainly focusing on “times to reply” versus “time to close the intervention,” through R statistical rel.4.3.

The second phase concerned the pre-processing actions necessary to arrange the data set to be trained with LSTM and Bi-LSTM RNN. In the original data set (see Section 4.2), each brief textual description of the maintenance request usually comprised meaningful words (necessary to describe what and where happened), strengthening words (introduced by the end-users to add more or less strength to the request), complementary words (i.e. “the, in, into, please,” etc.), numbers, symbols and punctuations. Each word could be present in different forms, depending on the structure of the sentence (e.g. “urgent, urgently, urgency”). Moreover, all these elements can be arranged in different ways, thus providing different possible meanings to sentences, even if composed of the same words.

Then, the recognition process can be very complex and influenced by the data set itself. The original data set has been hence pre-processed through the functions of the MATLAB Text analytics toolbox, also according to previous works (Banchs, 2021), to solve this issue and ensure a reliable LSTM and Bi-LSTM RNN application. Textual data were filtered performing a remotion of: very common words (stop-words); the most frequent out-of-vocabulary words (i.e. words and internal codes used to identify the specific areas or rooms of the building); symbols and punctuation; low length words (< two letters). Finally, a lemmatization process has been performed to group together the inflected forms of a word (dimensionality reduction), obtaining the data set to be tokenized and trained. Finally, according to Section 4.2, the outcoming data set for Bi-LSTM and LSTM RNN training comprises a list of end-user’s maintenance requests, associated with the related priority label, expressed according to target time (30, 60, 720, 1,040,10,080 min) and to the required staff label, expressed according to the work categories (electrical, fabrics, management, mechanical, minor new works, waste disposal).

The last phase involved the training and testing of two Bi-LSTM and two LSTM RNN, and their accuracy evaluation in terms of automatic recognition of end-users’ maintenance requests priority and technical staff to assign.

A neural network (NN) is a classifier that generally minimizes the mean square error between the predicted output and the expected value (Singh et al., 2014). An NN includes several layers, and each layer is composed of one or more nodes. Each node receives input and sends an output based on the weight assigned to each node. During the training process, the weight of each node is changed to minimize the error in the prediction. Then, the NN is directly trained from the data, based on the prediction errors.

To perform the training and test process, the data set has been divided into three parts. The two main parts are “training” (70% of the sentences) and “testing” (20% of the sentences). The third part of the data set (10% of sentences) has been used to separately check the ability of the Bi-LSTM and LSTM RNN to predict priority and technical staff assignment, according to criteria defined in Section 4.3.

Sentences have been tokenized (each sentence was divided into word units) by MATLAB tokenizer (which uses rules based on Unicode standard Annex #29 and ICU tokenizer), and converted into sequences of numeric indices. Sentences have been truncated to have word vectors of the same length. Different sequence lengths were considered (15, 30, 60, 120) to take into account the effect of alternative truncation choices. A length of 30 words was finally assumed for the analysis and discussion of results, depending on the overall statistics of the word sample for each end-user’s maintenance request in the data set.

The word embedding layer was put to the dimension of 50, and the number of hidden units was set to 80. The number of iterations to train the data set has been put to 31,200, corresponding to 50 epochs. Each “epoch” measures the number of times that the learning algorithm works through the entire training data set.

As said above, 10% of the data set has been used to separately evaluate how the LSTM and the Bi-LSTM RNN can automatically recognize the priority and technical staff to assign. Results were analyzed and compared through the following indicators (Gonçalves et al., 2013; Ribeiro et al., 2016):

• accuracy, i.e. the number of elements correctly classified with respect to the total number of elements;

• recall, i.e. the ratio of the number of elements correctly classified to the number of known elements in each class;

• precision, i.e. the ratio of the number of elements correctly classified to the total predicted in each class; and

• f1-score, i.e. the harmonic mean between both precision and recall.

The Bi-LSTM and LSTM RNN training and testing process, and the calculation of accuracy, precision, recall and f1-score have been performed through scripts in MATLAB 2021b code.

In the considered hospital case study, in 14 months, 14,577 correction actions were requested by end-users. The maximum number of maintenance requests has been recorded in January 2017, with a mean of 43 requests/day.

Figure 2 shows the distribution of the interventions by work categories (leading to specific staff required) and priority (expressed in terms of target priority time) assigned by the FM technical staff. In Figure 2 (left), the “fabrics” category includes the interventions required on the building components such as doors, windows, etc., which are usually the most frequent. Second, “electrical” and “mechanical” categories refer to the interventions on the equipment (electrical, HVAC, etc.) and represent about half of the whole maintenance requests number, also because they are more affected by the occupants’ interactions during the building operation. The “waste disposal” category comprises the interventions required for the waste disposal apparatus. The “new minor works” category includes the requests for modification to the current building systems status; hence, it cannot be directly considered as a corrective maintenance category. The categories “electrical,” “mechanical” and “fabrics,” when summed, entail more than 95% of the intervention requests, thus suggesting that most of the maintenance efforts should be carried out considering technical staff operating in these fields.

In Figure 2 (right), the number of interventions classified with a target intervention time equal to 720 min (12 h) is prevalent, followed by the interventions with a priority time equal to 60 min (1 h) and 1,440 min (24 h). The number of interventions with a response time < 30 min is very limited. There is also a significant number of delayable interventions that can be part of programmed maintenance tasks, which have been herein included in the class marked by “10080 minutes” (ideally equal to one week). Results from target intervention time analysis imply that the staff assignment should be essentially completed within one working day, to better face the priority of end-users’ requests according to the current FM contracts.

Figure 3(a) and 3(b) show the density function of the time necessary, respectively, to reply to each request and to close the intervention ticket (at the end of the work). As expected, considering the contractual constraints, the time necessary to reply and intervene after the ticket emission is very low (median = 16 min). The time necessary to close the ticket (at the end of the work) is higher. The median value is 4,006 min (2.78 days).

Figure 3(c) shows the distribution of the time necessary to close the ticket (end of the work) for the four main categories of corrective maintenance works previously identified, which are “electrical,” “fabrics,” “mechanical” and “waste disposal.” Distributions are similar, but the median values are lower for the electrical and mechanical categories, than for the fabrics category. This is presumably due to possible interruption of critical services in case of issues on electrical or mechanical plants in a hospital, as well as on the additional efforts due to fabrics-related intervention, that mainly involve building components rather than systems.

Thanks to the normality of the sample, the correlation analysis between the time to reply and to close the intervention ticket has been performed. Correlation is very low (Spearman’s rho = –0.18), showing that the time necessary to close the intervention depends not only on a quick reply.

The distribution of the length (number of words) of the sentences in the data set is characterized by a mean value equal to nine words and a median value equal to seven words. Long documents (e.g. more than 100 words) relating to the same end-users’ requests limitedly occur, and about 98% of the sample is composed of sentences with less than 30 words. Moreover, the number of words to be considered also reflects the necessity to include all the useful data according to the general sentence structure. In fact, sentences are composed at least of two or three paragraphs, including the description of the role of the writer (generally, within the first paragraph, which can also contain some initial greetings) and then the information on the request that can be considered essential for predictions. If present, the last paragraph generally contains the final greetings.

Figure 4 shows the four confusion matrices, representing the number of sentences for each “predicted” class (x-axis) against the “real” class assigned by the FM technical staff in the original dataset (y-axis), considering a sequence length equal to 30 (which is consistent in respect to Section 5.1 results on document length).

Classes are related to priority (Figure 4, left) and staff assignment with respect to the specific work categories (Figure 4, right). Results related to Bi-LSTM are in the upper part of the figure, while those related to LSTM are in its lower part. The model accuracy is obtained by summing the correct predictions (values on the diagonal of the matrix in Figure 4) and dividing them by the total sum of the sentences. Extended results for both work category and priority prediction are reported in the supplementary material, including the influence of the sequence length on the whole accuracy, and the related confusion matrices and precision, recall and F1-score data.

Concerning priority assignment, the whole accuracy obtained is 0.9364 (Bi-LSTM) and 0.9388 (LSTM) considering a sequence length equal to 30. As documented in the supplementary material, considering different sequence lengths, the use of Bi-LSTM slightly improves the accuracy (+3.54% with respect to the worst prediction case for LSTM) but only when sequence lengths > 60 are considered. In the other cases, Bi-LSTM and LSTM performances are essentially the same. The confusion matrices show that the trained Bi-LSTM and LSTM RNN better predict sentences characterized by mean (Class 3) and low (Class 4) priority. Similar results are shown by the indicators in Table 1 for the sequence length equal to 30. Sentences characterized by the highest priority (Class 1– 30 min) show a lower F1-score (0.82), both for Bi-LSTM and LSTM RNN. For this class, either the models seem to fail in priority prediction although results could be influenced by the specific sub-sample size of the true classes. In fact, Class 1 is characterized by the lowest dimension in terms of the number of sentences (85) with respect to the whole sample. On the contrary, F1-score related to Class 5 is quite low in respect of the other classes with a higher number of sentences. In this case, the results are probably affected by the fact that this class includes delayable interventions.

Results on technical staff assignment (Figure 4, right) confirm the general trend also observed for priority assignment. In particular, the final whole accuracy obtained is equal to 0.9707 (Bi-LSTM) and 0.9681 (LSTM), considering a sequence length equal to 30, thus providing similar results for the two RNNs. Anyway, the use of Bi-LSTM generally slightly improves the accuracy in respect to LSTM application (that is for sequence lengths = 60).

Table 2 shows the precision, accuracy and F1-score values for both Bi-LSTM and LSTM RNN for technical staff assignment, considering the related work categories. Slightly better results are obtained with Bi-LSTM compared to LSTM RNN, even if all the values are very high. Clear evidence of the influence of the sample dimension arises, thus confirming what is noticed for priority assignment. As shown by Figure 4 and Table 2, Classes 5 and 6 (respectively, “minor new works” and “waste disposal”) are characterized by a limited number of cases with respect to the other three main classes (“electrical,” “fabric,” “mechanical”). As for priority predictions, they are thus characterized by relatively lowest values in both RNN cases. Anyway, indicators confirm that both LSTM and Bi-LSTM can be reliably used for work categories assignment support.

The outcomes of this work underline the general reliability of LSTM and Bi-LSTM RNN in priority and staff assignment tasks, as demonstrated by the obtained indicators, and mainly by accuracy and F1-scores. These RNNs can support, at least, the technical staff members in:

• understanding the less urgent interventions, thus allowing them to focus on checking the other required tasks; and

• automatically predict the work categories in a quite homogeneous way, thus providing a good preliminary check for staff assignment tasks.

Staff assignment seems to be generally more reliable than priority assignment, while Bi-LSTM seems to be slightly more reliable than LSTM considering the whole accuracy.

However, the authors are aware that some limitations could have influenced these work outcomes.

In this sense, the work investigates a single case study, dealing with hospital and health-care facilities. Although the sample dimension is relevant and comparable to other previous works (D’Orazio et al., 2022; Gunay et al., 2019a; Kim et al., 2022; McArthur et al., 2018), the classes distribution for priority and work categories in the input data set are quite heterogeneous, as shown by Figure 2. Such a condition could affect the prediction indicators for both LSTM and Bi-LSTM, demonstrating their generally higher ability in recognizing proper classes for the most populated samples. On the contrary, as expected, the models seem to fail in the case of classes characterized by the lowest number of sentences. Anyway, the generally high reliability of the tested RNN is still demonstrated. Such results and these insights encourage future efforts linked to the extension of the tests to other wider data sets, using the same structure of this work, and also involving different case studies to compare differences between them.

Additional aspects can be connected to the specific methods adopted in this work, and in particular to the selected consolidated pre-processing techniques. To this end, future research could evaluate the impact of different pre-processing methods (e.g. stemming versus lemmatization) on the Bi-LSTM and LSTM application, given the encouraging outcomes of this research. Additional ML methods (D’Orazio et al., 2022; Gunay et al., 2019a; Kim et al., 2022; McArthur et al., 2018) could be then tested according to the proposed approach, to evaluate their reliability by involving larger data sets and different application contexts.

This paper demonstrates that Bi-LSTM and LSTM RNN can be effectively implemented into CMMS by collecting end-users’ maintenance requests, to support FM in detecting current performances and trends on maintenance needs and in the automatic assessment of maintenance requests by untrained end-users, thus improving the management process of building stocks.

Considering the tested data set, the related pre-processing actions and a sequence length relevant with respect to the communication structure, results obtained in this study show the ability of Bi-LSTM and LSTM RNN to properly and automatically recognize the priority of the corrective maintenance interventions and the technical staff to assign, with an accuracy of respectively 93.64% (Bi-LSTM) and 93.88% (LSTM) for the priority assignment and 97.07% (Bi-LSTM) and 96.81% (LSTM) for the work categories connected to the staff assignment. Slightly better results are obtained by training Bi-LSTM RNN with respect to the technical staff assignment prediction, while differences in priority assignment performances between Bi-LSTM and LSTM RNN are reduced.

The accuracy of priority assignment seems to be most affected by the original data set used in this work, which is characterized by the predominance of a single priority class.

The authors are aware that further efforts are needed to extend the validation of the proposed data-driven and text-mining approach to FM, but the application to a relevant context such as a hospital already demonstrates the capability of the methodology from a technical perspective. In this sense, the research outcomes encourage the use of IT systems for FM tasks, by also exploiting powerful data sources such as written texts and e-mail communications even by end-users who are not trained to properly describe corrective maintenance actions. Such information, collected during the operative phase of a building and stored in CMMSs, can be used to develop models able to predict maintenance actions to perform, thanks to ML methods and especially LSTM and Bi-LSTM RNN architectures. These techniques can be used to screen collected data for what concern both what happened and which member of the maintenance team should be involved in the corrective action (i.e. work category assignment) and how to better plan the intervention depending on its urgency (i.e. priority assignment). This could effectively reduce the involvement of technicians in repetitive and time-consuming actions, while they can be focused on the effective organization of the intervention, thus improving efforts on the most urgent tasks.

In fact, the priority of end-users’ corrective maintenance requests and the category of the technical staff to assign are used by facility managers to define the order to follow in the maintenance activities and the related physical, technical and management sources to use, to ensure business continuity as well as possible. In this sense, approaches implementing LSTM and Bi-LSTM RNN could be very important for large organizations where critical services are present (i.e. hospitals, public administrations, universities, etc.) and where the number of contemporary end-users’ maintenance requests can be very high, due to the dimension and complexity of the building stock.

The supplementary material for this article can be found online.