LONG DURATION ALARM SEQUENCE PREDICTED WITH Bi-LSTM AND OPERATOR SEQUENCE RECOMMENDATIONS THEREOF

Abstract

To provide a method and system for long duration alarm sequence predictions from past alarm sequences using bidirectional long short-term memory (Bi-LSTM) and for operator sequence recommendations thereof.SOLUTION: The method comprises: generating an alarm dataset for a facility from historical alarm data; preprocessing the alarm dataset to generate a training dataset; training a bidirectional long short-term memory (Bi-LSTM) on a training dataset for predicting an alarm sequence for the facility; generating a mapping table using the preprocessed historical alarm data; predicting, by the trained Bi-LSTM, a future output alarm sequence for a next plurality of hours based on an immediate input alarm sequence; and recommending an operator action sequence among multiple operator action sequences for the predicted future output alarm sequence.SELECTED DRAWING: Figure 2B

Description

[Technical field]

[0001] (CROSS REFERENCE TO RELATED APPLICATIONS) This application claims priority to Indian Application No. 202321042213 filed on June 23, 2023.

[0002] The present embodiment generally relates to the field of system-based automatic alarms, and more specifically, to a method and system for predicting long-term alarm sequences from past alarm sequences using Bidirectional Long-Short-Term Memory (Bi-LSTM) and proposing operator sequences therefor. [Background technology]

[0003] Alarm management systems are an integral part of any manufacturing or processing facility. Equipment defects or inefficient production can be identified by alarms. These alarms can be triggered by deviations of process parameters from their operating or safety margins. Once an operator has taken the relevant action in response to the triggered alarm to mitigate the defect or to return the process to a desired state, the alarm can be put into a dormant state. If future alarm sequences are known in advance, then operators will be better prepared to take appropriate action. Furthermore, prediction of future alarm sequences helps operators understand the propagation of deviations in the equipment. Therefore, prediction of future alarm sequences can help operators to mitigate faults and manage the efficiency of equipment. Alarm prediction or technical problem prediction is limited to predicting the next alarm using past alarm sequences. The behavior is also directed towards predicting some frequent or infrequent alarms using past sequences, which is a multi-step classification problem. However, this approach does not address the prediction of the complete alarm sequence. In practice, some attempts to focus on predicting alarm sequences are unsupervised. Having knowledge of the next alarm or a few important alarms is not enough to ensure equipment within a facility, minimal breakdowns, or minimal downtime for maintenance. Knowing the entire alarm sequence can aid in proactive measures and dramatically reduce failures and maintenance downtime, effectively enabling the functional continuity of equipment.

[0004] The process of sequence-from-sequence prediction has been well explored in the literature over the past decade due to problems in specific domains such as language translation. Although long-short-term memory (LSTM) has been recently applied for next alarm prediction, LSTM-based approaches limit sequence prediction accuracy because the LSTM architecture only uses past information while training the model.

[0005] Furthermore, the delay or time window for alarm prediction provided by existing methods is very small, on the order of minutes. It can be well understood that smaller and larger prediction windows pose challenges to meeting the best preventative measures. This is due to the technical limitations of LSTMs to effectively learn dependencies along the input-output sequence in order to provide longer prediction horizons. Therefore, addressing the technical challenges that would allow longer prediction windows by operators is desirable for better action or response planning. Summary of the Invention [Means for solving the problem]

[0006] SUMMARY OF THE DISCLOSURE The embodiments of the present disclosure provide technical improvements as solutions to one or more of the above-mentioned technical problems identified by the inventors in conventional systems.

[0007] For example, according to one embodiment, a method is provided for prediction of long term alarm sequences from past alarm sequences. The method comprises generating an alarm dataset for a facility from historical alarm data including timestamps of triggered alarms, alarm IDs, alarm descriptions, and associated devices within the facility, the facility being monitored via multiple sensors to obtain the historical alarm data, and each data point in the alarm dataset having a past input alarm sequence and a future output alarm sequence each including single or multiple occurrences of a unique alarm identified by the alarm ID.

[0008] The method further comprises pre-computing the alarm dataset to generate a training dataset, the pre-computing comprising: removing chattering; removing non-critical alarms; resizing the past input alarm sequences to an optimal input sequence length; and resizing the future output alarm sequences to an optimal output sequence length.

[0009] The method further comprises training a bidirectional long short-term memory (Bi-LSTM) on a training dataset to predict alarm sequences for the facility for multiple time durations, wherein the sum of the unique alarms in the historical alarm data represents a plurality of features for the Bi-LSTM, and the optimal input sequence length is an input to the Bi-LSTM for generating the optimal output sequence length predicted after post-training by the Bi-LSTM.

[0010] Additionally, the method includes generating a mapping table using the pre-calculated historical alarm data of the equipment to map a plurality of operator action sequences to a plurality of associated characteristic alarms.

[0011] Further, the method includes predicting, for the equipment, a future output alarm sequence for a next plurality of times based on instantaneous alarm sequences stored in the recent past, by the trained Bi-LSTM executing on the one or more hardware processors.

[0012] The method further comprises suggesting an operator operation sequence among the plurality of operator operation sequences for a predicted future output alarm sequence, the suggested operator operation sequence comprising associating the predicted future output alarm sequence with a specific alarm having a highest matching sequence score (MSS) among the plurality of specific alarm sequences.

[0013] In another aspect, a system for prediction of long term alarm sequences from past alarm sequences is provided. The system includes a memory for storing instructions, one or more input / output (I / O) interfaces, and one or more hardware processors connected to the memory via the one or more input / output (I / O) interfaces, wherein the one or more hardware processors are configured, in accordance with the instructions, to generate an alarm data set for a facility from historical alarm data including timestamps of triggered alarms, alarm IDs, alarm descriptions, and associated devices within the facility, wherein the facility is monitored via multiple sensors to obtain the historical alarm data, and wherein each data point in the alarm data set has a past input alarm sequence and a future output alarm sequence, each including single or multiple occurrences of a unique alarm identified by the alarm ID.

[0014] The one or more hardware processors pre-compute the alarm data set to generate a training data set, the pre-computing including chatter removal, removal of non-critical alarms, resizing the past input alarm sequences to an optimal input sequence length, and resizing the future output alarm sequences to an optimal output sequence length.

[0015] Further, the one or more hardware processors are configured to train a bidirectional long-short-term memory (Bi-LSTM) on a training dataset to predict alarm sequences for the facility for multiple time durations, wherein the sum of the unique alarms in the historical alarm data represents a plurality of features for the Bi-LSTM, and the optimal input sequence length is an input to the Bi-LSTM for generating the optimal output sequence length predicted after post-training by the Bi-LSTM.

[0016] Further, the one or more hardware processors generate a mapping table using the pre-calculated historical alarm data of the equipment to map a plurality of operator action sequences to a plurality of associated specific alarms.

[0017] Further, the one or more hardware processors predict, for the equipment, future output alarm sequences for a next plurality of times based on instantaneous alarm sequences stored in the recent past, by the trained Bi-LSTM executing on the one or more hardware processors.

[0018] Further, the one or more hardware processors are configured to suggest an operator operation sequence among the plurality of operator operation sequences for a predicted future output alarm sequence, wherein the suggested operator operation sequence is associated with the predicted future output alarm sequence and a specific alarm having a highest matching sequence score (MSS) among the plurality of specific alarm sequences.

[0019] In another aspect, a non-transitory machine-readable information storage medium is provided that includes one or more instructions that, when executed by one or more hardware processors, cause a method of prediction of long term alarm sequences from past alarm sequences.

[0020] The method comprises generating an alarm dataset for a facility from historical alarm data including timestamps of triggered alarms, alarm IDs, alarm descriptions, and associated devices within the facility, the facility being monitored via multiple sensors to obtain the historical alarm data, and each data point in the alarm dataset having a past input alarm sequence and a future output alarm sequence each including single or multiple occurrences of a unique alarm identified by the alarm ID.

[0021] Further, the method comprises pre-computing the alarm dataset to generate a training dataset, the pre-computing comprising chatter removal, removal of non-critical alarms, resizing the past input alarm sequences to an optimal input sequence length, and resizing the future output alarm sequences to an optimal output sequence length.

[0022] The method further comprises training a bidirectional long short-term memory (Bi-LSTM) on a training dataset to predict alarm sequences for the equipment for multiple time durations, wherein the sum of the unique alarms in the historical alarm data represents a plurality of features for the Bi-LSTM, and the optimal input sequence length is an input to the Bi-LSTM for generating the optimal output sequence length of the future output alarm sequence predicted after post-training by the Bi-LSTM.

[0023] The method further comprises generating a mapping table using the pre-calculated historical alarm data of the equipment to map a plurality of operator action sequences to an associated plurality of unique alarms.

[0024] Further, the method comprises predicting, for the equipment, a future output alarm sequence for a next plurality of times based on instantaneous input alarm sequences stored in the recent past, by the trained Bi-LSTM executing on the one or more hardware processors.

[0025] The method further comprises suggesting an operator operation sequence among the plurality of operator operation sequences for a predicted future output alarm sequence, wherein the suggested operator operation sequence is associated with the predicted future output alarm sequence and a specific alarm having a highest matching sequence score (MSS) among the plurality of specific alarm sequences. It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed.

[0026] The accompanying drawings, which are incorporated in and constitute a part of this disclosure, illustrate example embodiments and, together with the detailed description, serve to explain the disclosed principles. [Brief description of the drawings]

[0027] [Figure 1A] FIG. 1A is a functional block diagram of a system for prediction of long-term alarm sequences from past alarm sequences using a Bidirectional Long Short-Term Memory (Bi-LSTM) Neural Network (NN) and its operator sequence suggestions, according to some embodiments of the present disclosure. [Figure 1B] FIG. 1B illustrates an architectural and process overview of the system of FIG. 1A, according to some embodiments of the present disclosure. [Figure 2A] FIG. 2A is a flow diagram illustrating a method for predicting long-term alarm sequences from past alarm sequences and proposing operator sequences thereof using the system depicted in FIG. 1A and FIG. 1B according to some embodiments of the present disclosure. [Figure 2B] FIG. 2B is a flow diagram illustrating a method for predicting long-term alarm sequences from past alarm sequences and proposing operator sequences thereof using the system depicted in FIG. 1A and FIG. 1B according to some embodiments of the present disclosure. [Figure 3A] FIG. 3A is an alarm prediction generated by the prior art. [Figure 3B] FIG. 3B is an alarm prediction generated by the prior art. [Figure 4] FIG. 4 depicts a long-term future output alarm sequence predicted by the system of FIGS. 1A and 1B based on received past input alarm sequences, according to some embodiments of the present disclosure. [Diagram 5] FIG. 5 depicts an architecture of a Bi-LSTM having an encoder and a decoder for predicting future output alarm sequences according to some embodiments of the present disclosure.

[0028] It should be appreciated by those skilled in the art that the several block diagrams presented herein represent conceptual views of illustrative systems and devices incorporating principles presented herein. Similarly, it will be understood that some flowcharts, flow diagrams, and the like may be substantially represented on a computer-readable medium and represent various processes that may be executed by a computer or processor, whether or not a computer or processor is explicitly depicted. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0029] Exemplary embodiments are described with reference to the accompanying drawings. In the drawings, the left-most digit(s) of a reference number identifies the first occurrence of the reference number. Wherever convenient, the same reference numbers will be used throughout the drawings to refer to the same or like parts. Although disclosed examples and features are described herein, modifications, adaptations, and other approaches can be made without departing from the scope of the disclosed embodiments.

[0030] Most of the techniques used by state-of-the-art alarm prediction systems rely on long-short-term memories (LSTMs). Conventional methods predict a single or a few critical alarms. One technical paper, “Learning operation strategies from alarm management systems by temporal pattern mining and deep learning” by Dorgo et al., attempts to predict alarm sequences, however, also relies on LSTMs. During model training, the LSTM architecture uses only past information, and therefore, the prediction accuracy is relatively low. Furthermore, unlike bidirectional LSTM, LSTM does not efficiently learn dependencies along the input and output sequences and therefore cannot predict alarm sequences for longer prediction horizons. The benchmark data sets used in the above efforts are simulated data sets, unlike real data sets, and are not complex. The number of defects introduced to generate alarm events is 11, and the example alarm sequences have a maximum sequence length of 5, which is very small.

[0031] This improves the learning of the algorithm and makes it more accurate in predicting alarm sequences. The benchmark datasets used in the above mentioned publications are simulated datasets and are not complex. The number of defects introduced to generate an alarm event is 11. The alarm sequence examples in the above literature have a maximum sequence length of 5, which is very small.

[0032] Embodiments of the present disclosure provide a method and system for prediction of long-term alarm sequences from past alarm sequences using a Bidirectional Long Short-Term Memory (Bi-LSTM) Neural Network (NN) and its operator sequence suggestions. The Bi-LSTM encoder-decoder technique uses the true output sequence as input to the decoder at every time step during training. This allows Bi-LSTM to efficiently learn dependencies along the input-output sequences and predict longer alarm sequences. The system disclosed herein is capable of predicting future output alarm sequences several hours into the future, for example 480 minutes.

[0033] Additionally, the disclosed system and method provides a mapping table with each unique alarm sequence mapped to an associated operator action sequence. The predicted future output alarm sequences are evaluated against each unique alarm sequence to identify the closest unique alarm sequence using the Matching Sequence Score (MSS) disclosed in the method. Known sequence evaluation metrics used in language translation models, such as bilingual evaluation (BLEU) scores and edit distance scores, cannot be applied to alarm sequence evaluation. BLEU scores and edit distance scores are used in the field of language translation, where words and phrases follow certain patterns. These scores are based on an N-gram module that is highly specialized for language translation tasks. Likewise, it does not aid in the evaluation in the area of ​​alarm sequence prediction. One of the prior art techniques suggests a metric for prediction of alarm sequences, however, it only considers the longest sequences. The longest sequence will not yield results if one false alarm is predicted between two long sequences.

[0034] The MSS disclosed in the method allows for a comparison between two alarm sequences and allows for mapping a predicted alarm sequence to the closest existing defined alarm sequence, well associated with the tested operator action defined in the mapping table. Thus, ensuring that the most appropriate action is taken for the alarm sequence predicted by the system 100.

[0035] The operator sequence in the mapping table that most closely corresponds to the best alarm sequence is suggested to the operator to ensure the best possible prevention, or Preventive measures are implemented to avoid malfunction or failure of one or more devices within the facility being monitored.

[0036] Referring now to the drawings, and more particularly to Figures 1A through 5, in which like reference numerals indicate corresponding features consistent throughout the drawings, preferred embodiments are shown and will be described with respect to the context of example systems and / or methods set forth below.

[0037] FIG. 1A is a functional block diagram of a system 100 for predicting long-term alarm sequences from past alarm sequences using a bidirectional long-short-term memory (Bi-LSTM) neural network (NN) and its operator sequence suggestions, according to some embodiments of the present disclosure.

[0038] In one embodiment, the system 100 includes a processor 104 operatively connected to a communications interface device, or input / output (I / O) interface 106 , and one or more data storage device memories 102 . System 100 , having one or more hardware processors, is configured to perform the functions of one or more of the functional blocks of system 100 .

[0039] With reference to the configuration of system 100 , in one embodiment, processor 104 may be one or more hardware processors 104 . In one embodiment, the one or more hardware processors 104 may be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuits, and / or devices that process signals based on operational instructions. Among other functionality, one or more hardware processors 104 are configured to retrieve and execute computer-readable instructions stored in memory 102 . In one embodiment, the system 100 may be implemented in a variety of computer systems, including laptop computers, notebook computers, portable devices such as mobile phones, workstations, mainframe computers, servers, and the like.

[0040] The I / O interface 106 can include a variety of software and hardware interfaces, such as, for example, a web interface, a graphical user interface such as the displays depicted in the system architecture and process overview of FIG. 1B. The I / O interface 106 can facilitate multiple communications within a wide variety of network N / W and protocol types, including wired networks, such as LAN, cable, and wireless networks, such as WLAN, mobile telephone, and the like. In one embodiment, I / O interface 106 may include one or more terminals for connecting to multiple output devices or to other servers or devices.

[0041] The memory 102 may include, for example, volatile memory, such as static random access memory (SRAM) and dynamic random access memory (DRAM), and / or known computer readable media, such as read only memory (ROM), erasable writeable ROM, flash memory, hard disks, optical disks, known computer readable media, and non-volatile memory, such as magnetic tape.

[0042] In one embodiment, the memory 102 includes a number of modules 110 . For example, a Bi-LSTM having an encoder and a decoder, an MSS calculation module for calculating an MSS for alarm sequence evaluation, a pre-calculation module, etc. (not shown) as depicted in FIG. 1B and described in FIG. 5. The multiple modules 110 contain programs or applications or coded instructions that complement the functions performed by the system 100 for performing the different steps involved in the process of alarm sequence prediction from the use of past alarm sequences and operator sequence suggestions performed by the system 100. The modules 110 can include, among other things, routines, programs, objects, components, and data structures that perform particular tasks or implement particular abstract data types. The modules 110 may also be used for signal processing, node machines, logic circuits, and / or other devices or components that manipulate signals based on operational instructions. Furthermore, the modules 110 may be implemented as hardware, as computer readable instructions executed by one or more hardware processors 104, or as a combination thereof. The plurality of modules 110 may include various sub-modules (not shown).

[0043] Additionally, the memory 102 may include information related to the input / output of each step executed by the processor 104 of the system 100 and the methods of the present disclosure.

[0044] In addition, the memory 102 includes a database 108 . The database can store historical alarm data including an update history of triggered alarms, alarm IDs, alarm descriptions, and associated equipment in the facility, pre-computed data generated from the historical alarm data by the process module, a training dataset, an inspection dataset, and a test dataset generated from the pre-computed for training, inspection, and inspection of the Bi-LSTM. The database (or repository) 108 may contain abstracted portions of refinement code and data processed, received, or generated as a result of execution of modules within the modules 110.

[0045] It should be noted that although database 108 is shown within system 100, in other embodiments database 108 may be implemented external to system 100 and communicatively connected thereto. The data contained within the external database can be updated periodically. For example, new data can be added to the database (not shown in FIG. 1A) and / or existing data can be modified and / or unnecessary data can be deleted from the database. As one example, the data may be stored in external systems such as a Lightweight Directory Access Protocol (LDAP) directory and a Relational Database Management System (RDBMS). The functionality of the system 100 configuration will now be described with reference to FIGS. 1B through 5.

[0046] 2A-2B are flow diagrams illustrating a method 200 of predicting long-term alarm sequences from past alarm sequences and proposing operator sequences thereof using the system shown in FIGS. 1A and 1B according to an embodiment of the present disclosure. In one embodiment, the system 100 includes one or more data storage devices or memories 102 operably coupled to the processor 104 and configured to store instructions for executing the steps of the method 200 by the processor or one or more hardware processors 104. The steps of the method 200 of the present disclosure will now be described with reference to the configuration or block diagram of the system 100 shown in FIGS. 1A and 1B, and the steps of the flow diagram shown in FIGS. 2A and 2B. Although process steps, method steps, or techniques may be described in a sequential order, these processes, methods, and techniques may be configured to function out of order. In other words, the sequence or order of steps that may be described does not necessarily indicate a requirement that the steps be performed in that order. The steps of processes described herein may be performed in any order practical. Moreover, some steps may be performed simultaneously.

[0047] Referring to the steps of method 200, in step 202 of method 200, the one or more hardware processors 104 are configured to generate an alarm data set for the equipment from the historical alarm data. The historical alarm data includes a chronological data set of triggered alarms, the corresponding timestamps, the alarm ID, a description of the alarm, and the associated device of the facility. The facility includes multiple devices and is monitored through multiple sensors to capture historical alarm data. A snippet of the historical alarm data is shown in Table 1 below.

[0048] [Table 1]

[0049] Each data point in the alarm data set has a past input alarm sequence and a future output alarm sequence that includes one or more occurrences of a particular alarm identified by an alarm ID. A snippet of the input and output alarm sequence data set is shown in Table 2 below.

[0050] [Table 2]

[0051] In step 204 of the method 200, the one or more hardware processors 104 are configured to pre-compute an alarm data set to generate a training data set. Similarly, evaluation and testing datasets for the Bi-LSTM are generated. Pre-computation includes chatter removal, removal of non-critical alarms, resizing past input alarm sequences to an optimal input sequence length, resizing future output alarm sequences to an optimal output sequence length, with padding and truncation where necessary. The proper lengths of the input and output sequences are specified. The fair sequence length is the sequence length for which only 5% of the data points have a sequence length greater than the fair sequence length. A fragment of the training data after precomputation is shown in Table 3 below. Processed assessment and laboratory data show similar trends.

[0052] [Table 3]

[0053] At step 206 of method 200, the one or more hardware processors 104 are configured to train a Bi-LSTM to predict alarm sequences for the equipment for multiple time durations using the training dataset. The total number of distinct alarms in the historical alarm data represents the multiple features for the Bi-LSTM, and the optimal input sequence length is the input of the Bi-LSTM to generate the optimal output sequence length of the predicted future output alarm sequence after post-training by the Bi-LSTM. For example, if the number of unique alarms in the dataset (n_features) is 155, The length of the past input alarm sequence (n_inputs) is 109. Assume that the future output alarm sequence length (n_outputs) is 20. As shown in Figure 5, a Bi-LSTM with an encoder-decoder architecture is used for sequence-to-sequence prediction. An example of sequence-to-sequence conversion is language conversion in the field of natural language processing. Past input alarm sequence [a 10 ,a7,a2,a5] is the future output alarm sequence [a9,a7,a2,a 10 ] is used to predict. Training and testing the encoder is similar. The decoder is trained using the true output as input to predict the next alarm. A test using the decoder takes all sequences as input to predict future sequence lengths.

[0054] The list of optimized HyperTune parameters is presented below and was experimentally determined to give the best performance of Bi-LSTM when applied to alarm sequences. i Number of LSTM units (n_uints) = 128 ii Number of epochs = 30 iii Batch size (batch_size) = 128 iv activation function(activation)=relu activation fun(x)=max(0, x) The activation and loss functions used here are well known in the ML domain and will not be described for brevity. The variables of the activation function are in the form of the input to the neuron, (w+b*x), b=bias, w=weight, and are similarly known in the art. v dropout = 0.3 vi Loss function = weighted focal loss(y pred ,α,γ)=-α(1-y pred ) γ log(y pred ) α is a weighting factor and gamma is a convergence factor. vii optimizer='rmsprop'

[0055] In step 208 of method 200, the one or more hardware processors 104 are configured to generate a mapping table using the pre-calculated historical alarm data of the equipment to map a plurality of unique alarm sequences to a plurality of associated operator action sequences. A database of alarm sequences and corresponding operator actions is created from past knowledge and stored in database 108 . A fragment of the database is shown in the mapping table of FIG. 1B. The predicted alarm sequence is compared with alarm sequences from a database using a matching sequence length algorithm. The algorithm finds the closest matching alarm sequence to predict the alarm sequence and suggests corresponding operator actions from the database.

[0056] In step 210 of method 200, the one or more hardware processors 104 are configured to predict, during the inferencing stage, future output alarm sequences for the equipment for the next multiple times based on instantaneous input alarm sequences stored in the recent past, using the trained Bi-LSTM. The predicted alarm sequence is shown in FIG. 4 and, for comparison, conventional alarm prediction, in which the next alarm or few important alarms are limited, is shown in FIGS. 3A and 3B, respectively.

[0057] Table 4 below provides the Bi-LST performance of system 100 for the training and testing datasets.

[0058] [Table 4]

[0059] In step 212 of the method 200, the one or more hardware processors 104 are configured to suggest an operator action sequence for the predicted future output alarm sequence. The proposed operator action sequence is associated with the specific alarm sequence having the highest matching sequence score (MSS) with the predicted future output alarm sequence among the multiple specific alarm sequences.

[0060] The MSS for each of the multiple alarm sequences is i) a sum of a plurality of alarm sequence lengths within a range of predicted future output alarm sequences that match at least a portion of an alarm sequence within a range of distinctive alarm sequences from the longest matching sequence to the smallest matching sequence having a sequence length equal to or greater than 2; ii) a characteristic alarm sequence length; and It is calculated as a percentage of

[0061] Additionally, the MSS is used as a reference value for evaluating actual alarm sequences (ground truth) against predicted future output alarm sequences. it is, i) a sum of a plurality of alarm sequence lengths within a range of predicted future output alarm sequences that match at least a portion of an alarm sequence within a range of an actual alarm sequence from the longest matching sequence to the smallest matching sequence having a sequence length equal to or greater than 2; ii) the actual alarm sequence length; and It is calculated as a percentage of For example, MSS (Match Sequence Score) can be calculated as shown in Equation 1.

[0062]

number

[0063] Actual Sequence = [26, 31, 26, 11, 98, 11, 98, 11, 98, 45, 11, 98] Predicted Sequence = [26, 31, 26, 11, 98, 45, 11] Longest matching sequence or Longest Match length (LML) = 26, 31, 26, 11, 98 2nd matching sequence length or 2nd LML 45, 11 [ =2] Actual sequence length = 12 Then, the Match Sequence Score is 0.58 according to equation (2). Here, 0 indicates a perfect mismatch and 1 indicates a perfect match.

[0064]

number

[0065] Known sequence evaluation metrics used in language translation models, such as bilingual evaluation (BLEU) scores and edit distance scores, cannot be applied to alarm sequence evaluation. BLEU scores and edit distance scores are used in the field of language translation, where words and phrases follow certain patterns. These scores are based on an N-gram module that is highly specialized for language translation tasks. Likewise, it does not aid in the evaluation in the area of ​​alarm sequence prediction. One of the prior art techniques suggests a metric for prediction of alarm sequences, however, it only considers the longest sequences. As mentioned before, the longest sequence will not yield results if one false alarm is predicted between two long sequences. For example, the actual sequence is [1,2,3,4,5,6,7,8,9,10] and the predicted sequence is [1,2,3,4,5,20,7,8,9,10]. In the above example, the longest matching sequence is 5 with a total length of 10. However, the two sequences above match except for one alarm. The metric disclosed in the method similarly gives more weight to the second longest sequence length, providing a better match.

[0066] The MSS disclosed in the method allows for a comparison between two alarm sequences and allows for mapping the predicted alarm sequence to the closest existing defined alarm sequence, which is well associated with the tested operator action defined in the mapping table. Thus, ensuring that the most appropriate action is taken for the alarm sequence predicted by the system 100.

[0067] Thus, the methods and systems disclosed herein provide a mechanism for prediction of alarm sequences having improved accuracy and longer term alarm event sequences enabling better monitoring and control of equipment functioning with adequate and appropriate operator action beyond the unwanted events associated with the predicted alarms that may occur.

[0068] The written specification describes the subject matter herein to enable any person skilled in the art to make and use the embodiments. The scope of the subject embodiments is defined by the claims, and may include other modifications that occur to those skilled in the art. If the claim has similar elements that are not literally different from the claims, or if the claim has equivalent elements that are unrealistically different from the claims, Such other modifications are intended to be within the scope of the following claims.

[0069] It will be understood that the scope of protection extends to computer readable means having such programs and messages in addition, such computer readable storage means including program code means for the embodying of one or more steps of the method when the program is executed on a server, a mobile device, or any suitable programmable device. The hardware device can be any type of device that can be programmed, such as, for example, any type of computer such as a server, a personal computer, etc., or a combination thereof. The device may also be implemented, for example, by hardware means, such as an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), or a combination of hardware and software means, such as an ASIC and an FPGA, or by at least one microprocessor and at least one memory having a software processing arrangement configured to process the data. Thus, the means may include both hardware and software means. The method embodiments disclosed herein can be implemented in hardware and software. The device may also include software means. Alternatively, embodiments may be implemented on different hardware devices, for example multiple CPUs.

[0070] The embodiments herein may include hardware and software elements. Embodiments implemented in software include, but are not limited to, firmware, resident software, microcode, etc. The functionality performed by the various configurations described herein may be implemented in other elements or combinations of elements. For purposes of this specification, a computer usable or computer readable medium can be made into an apparatus that can contain, store, communicate, propagate, or transfer a program for use with or in connection with an instruction execution system, apparatus, or device.

[0071] The depicted steps are presented for purposes of illustrating the exemplary embodiments shown, and it is to be anticipated that ongoing technological developments may result in changes to the methods, particularly the functions performed. These examples are presented here for purposes of illustration and not limitation. Moreover, the boundaries of functional building blocks are arbitrarily defined herein for convenience of explanation. Alternate boundaries can be defined so long as the specified functions and relationships thereof are appropriately performed. Alternatives (including equivalents, extensions, variations, and deviations of those described herein) will be apparent to one of ordinary skill in the relevant art(s) based on the teachings contained herein. Such alternatives fall within the scope of the disclosed embodiments. Additionally, the terms "comprising," "having," "containing," and "including" and other similar forms are intended to be equivalent in meaning and are intended to be open-ended and not to be an exhaustive list meaning that the item or items following them are limited to only such item or items or the listed item or items. It should also be noted that, as used herein and in the claims, the singular forms "a," "an," and "the" include plural references unless the context clearly dictates otherwise.

[0072] Furthermore, one or more computer readable mediums may be utilized to implement embodiments consistent with this specification. A computer-readable storage medium refers to any type of physical memory capable of being loaded into a processor and from which information or data can be read. Thus, a computer-readable storage medium may store instructions for execution by one or more processors, including instructions for causing the processors to perform steps or acts consistent with the embodiments described herein. It should be understood that the term "computer-readable medium" includes tangible items and excludes carrier waves and transient signals, ie, non-transient. For example, random access memory (RAM), read only memory (ROM), volatile memory, non-volatile memory, hard drives, CD-ROMs, DVDs, flash devices, disks, and other known physical storage media.

[0073] It is intended that the disclosure and examples be considered as exemplary only, with the true scope of the disclosed embodiments being indicated in the following claims.

Claims

1. A processor-implemented method (200) for prediction of an alarm sequence, the method comprising: generating (202) via one or more hardware processors an alarm data set for a facility from historical alarm data including timestamps of triggered alarms, alarm IDs, alarm descriptions, and associated devices within the facility, the facility being monitored via a number of sensors to obtain the historical alarm data, each data point in the alarm data set having a past input alarm sequence and a future output alarm sequence each including one or more occurrences of a specific alarm identified by the alarm ID; pre-computing (204) the alarm data set to generate a training data set via the one or more hardware processors, the pre-computing step including chatter removal, non-critical alarm removal, resizing the past input alarm sequences to an optimal input sequence length, and resizing the future output alarm sequences to an optimal output sequence length; training (206) via the one or more hardware processors a bidirectional long short-term memory (Bi-LSTM) on a training data set to predict alarm sequences for the facility for durations spanning multiple hours, wherein the sum of the distinct alarms in the historical alarm data represents a plurality of features for the Bi-LSTM, and the optimal input sequence length is an input to the Bi-LSTM for generating the optimal output sequence length of the future output alarm sequences predicted after post-training by the Bi-LSTM; generating (208) a mapping table using the pre-calculated historical alarm data of the equipment to map a plurality of unique alarm sequences to a plurality of associated operator action sequences via the one or more hardware processors; During the inferencing step, predicting (210) a future output alarm sequence for the next plurality of times for the equipment based on instantaneous input alarm sequences stored in the recent past, by the trained Bi-LSTM executed on the one or more hardware processors; proposing (212) an operator operation sequence among the plurality of operator operation sequences for a predicted future output alarm sequence via the one or more hardware processors, the proposed operator operation sequence being associated with a specific alarm sequence having a highest matching sequence score (MSS) among the predicted future output alarm sequence and the plurality of specific alarm sequences; Equipped with The way the processor is implemented.

2. The MSS for each of the plurality of alarm sequences comprises: i) a sum of a number of alarm sequence lengths within the range of the predicted future output alarm sequences that match at least a portion of an alarm sequence within a range of unique alarm sequences from the longest matching sequence to the smallest matching sequence having a sequence length of 2 or more; ii) the unique alarm sequence length; and It is calculated by evaluating the ratio of The processor-implemented method of claim 1 .

3. The MSS is used as a reference value for the predicted future output alarm sequence; i) a sum of a number of alarm sequence lengths within the range of the predicted future output alarm sequences that match at least a portion of an alarm sequence within a range of actual alarm sequences, the sum being from the longest matching sequence to the shortest matching sequence having a sequence length of 2 or more; ii) the actual alarm sequence length; and It is calculated by evaluating the ratio of The processor-implemented method of claim 2 .

4. The Bi-LSTM includes an encoder and a decoder with a LSTM equal to 128, a number of epochs equal to 30, a batch size equal to 128, and an experimentally obtained dropout layer equal to 0.3, and the Bi-LSTM includes a relu activation function, a weighted focal loss function, an rmsprop optimizer, and Used for The processor-implemented method of claim 1 .

5. A system (100) for prediction of an alarm sequence, said system (100) comprising: A memory (102) for storing instructions; one or more input / output (I / O) interfaces (106); one or more hardware processors (104) coupled to the memory (102) via the one or more input / output (I / O) interfaces (106); Equipped with The instructions cause one or more hardware processors (104) to: generating an alarm data set for a facility from historical alarm data including timestamp alarm IDs of triggered alarms, alarm descriptions, and associated devices within the facility, the facility being monitored via multiple sensors to obtain the historical alarm data, each data point in the alarm data set having a past input alarm sequence and a future output alarm sequence each including single or multiple occurrences of a unique alarm identified by the alarm ID; pre-computing the alarm data set to generate a training data set, the pre-computing comprising chatter removal, non-critical alarm removal, resizing the past input alarm sequences to an optimal input sequence length, and resizing the future output alarm sequences to an optimal output sequence length; training a bidirectional long short-term memory (Bi-LSTM) on a training data set to predict alarm sequences for the installation for multiple time durations, where the sum of the distinctive alarms in the historical alarm data represents a plurality of features for the Bi-LSTM, and the optimal input sequence length is an input of the Bi-LSTM to generate the optimal output sequence length predicted after post-training by the Bi-LSTM; generating a mapping table using the pre-calculated historical alarm data of the equipment to map a plurality of operator operation sequences to a plurality of associated characteristic alarms; During the inferencing step, predicting, for the equipment, a future output alarm sequence for the next plurality of times based on instantaneous alarm sequences stored in the recent past, by the trained Bi-LSTM executed on the one or more hardware processors; Suggesting an operator operation sequence among the plurality of operator operation sequences for a predicted future output alarm sequence, the suggested operator operation sequence associating the predicted future output alarm sequence with a specific alarm having a highest matching sequence score (MSS) among the plurality of specific alarm sequences; configured to: system.

6. The MSS for each of the plurality of alarm sequences comprises: i) a sum of a number of alarm sequence lengths within the range of the predicted future output alarm sequences that match at least a portion of an alarm sequence within a range of unique alarm sequences from the longest matching sequence to the smallest matching sequence having a sequence length of 2 or more; ii) the unique alarm sequence length; and It is calculated by evaluating the ratio of The system of claim 5.

7. The MSS is used as a reference value for the predicted future output alarm sequence; i) a sum of a number of alarm sequence lengths within the range of the predicted future output alarm sequences that match at least a portion of an alarm sequence within a range of actual alarm sequences, the sum being from the longest matching sequence to the shortest matching sequence having a sequence length of 2 or more; ii) the actual alarm sequence length; and It is calculated by evaluating the ratio of The system of claim 6.

8. The Bi-LSTM includes an encoder and a decoder with a LSTM equal to 128, a number of epochs equal to 30, a batch size equal to 128, and an experimentally obtained dropout layer equal to 0.3, and the Bi-LSTM includes a relu activation function, a weighted focal loss function, an rmsprop optimizer, and Used for The system of claim 5.

9. One or more non-transitory machine-readable information storage media containing one or more instructions that, when executed by one or more hardware processors, generating an alarm data set for a facility from historical alarm data including timestamps of triggered alarms, alarm IDs, alarm descriptions, and associated devices within the facility, the facility being monitored via a number of sensors to obtain the historical alarm data, each data point in the alarm data set having a past input alarm sequence and a future output alarm sequence each including one or more occurrences of a unique alarm identified by the alarm ID; pre-computing the alarm data set to generate a training data set, the pre-computing comprising chatter removal, removal of non-critical alarms, resizing the past input alarm sequences to an optimal input sequence length, and resizing the future output alarm sequences to an optimal output sequence length; training a bidirectional long short-term memory (Bi-LSTM) on a training data set to predict alarm sequences for the equipment for multiple time durations, wherein the sum of the distinctive alarms in the historical alarm data represents a plurality of features for the Bi-LSTM, and the optimal input sequence length is an input of the Bi-LSTM to generate the optimal output sequence length of the future output alarm sequence predicted after post-training by the Bi-LSTM; generating a mapping table using the pre-calculated historical alarm data of the equipment to map a plurality of operator action sequences to an associated plurality of specific alarms; During the inferencing step, predicting, for the facility, a future output alarm sequence for the next plurality of times based on an instantaneous input alarm sequence stored in the recent past, by the trained Bi-LSTM executed on the one or more hardware processors; Suggesting an operator operation sequence among the plurality of operator operation sequences for a predicted future output alarm sequence, the suggested operator operation sequence being associated with the predicted future output alarm sequence and a specific alarm having a highest matching sequence score (MSS) among the plurality of specific alarm sequences; Cause One or more non-transitory machine-readable information storage media.

10. The MSS for each of the plurality of alarm sequences comprises: i) a sum of a plurality of alarm sequence lengths within the range of the predicted future output alarm sequences that match at least a portion of an alarm sequence within a range of unique alarm sequences from the longest matching sequence to the smallest matching sequence having a sequence length equal to or greater than 2; ii) the unique alarm sequence length; and It is calculated by evaluating the ratio of 10. One or more non-transitory machine-readable information storage media according to claim 9.

11. The MSS is used as a reference value for the predicted future output alarm sequence; i) a sum of a plurality of alarm sequence lengths within the range of the predicted future output alarm sequences that match at least a portion of an alarm sequence within a range of actual alarm sequences, the sum being from the longest matching sequence to the shortest matching sequence having a sequence length equal to or greater than 2; ii) the actual alarm sequence length; and It is calculated by evaluating the ratio of 11. One or more non-transitory machine-readable information storage media according to claim 10.

12. The Bi-LSTM includes an encoder and a decoder with a LSTM equal to 128, a number of epochs equal to 30, a batch size equal to 128, and an experimentally obtained dropout layer equal to 0.3, and the Bi-LSTM includes a relu activation function, a weighted focal loss function, an rmsprop optimizer, and Used for 10. One or more non-transitory machine-readable information storage media according to claim 9.

Images