Predictive maintenance for large medical imaging systems

Abstract

A predictive maintenance alert device (40) includes a server computer (42) operably connected with a computer network (46) to receive timestamped machine log data (30) and timestamped service log data (32) from a medical imaging device (10) and to send a maintenance alert (44) to a service center (12). The predictive maintenance alert method includes: deriving a feature from received log data; and applying a set of models (64) of the group of components to the derived features to generate a maintenance alert. Each model may include a heterogeneous model that includes a machine learning analysis model (70) representing the set of components, the machine learning analysis model having a statistical remaining useful life model (72) of components embedded in the set of components. Each component may belong to a single group of components. Some derived features may be constructed fault mode features or constructed fault solution features.

Description

[0001] This application is a divisional application of patent application 201880045935.8, whose application date is July 5, 2018 and whose invention name is “Predictive maintenance for large medical imaging systems”. Technical Field

[0002] The following generally relates to the field of medical imaging system (including image-guided therapy, iGT) maintenance, the field of medical imaging system fault prediction, and related fields. Background Art

[0003] Medical imaging devices include very complex systems such as magnetic resonance imaging (MRI) devices, ultrasound imaging devices, digital radiography (DR) devices, transmission computed tomography (CT) imaging devices, emission imaging systems such as positron emission tomography (PET) imaging devices, and gamma cameras for single photon emission computed tomography (SPECT) imaging, hybrid systems that provide multiple modalities in a single device such as PET / CT or SPECT / CT imaging devices, and imaging devices designed to guide biopsies or other interventional medical flow procedures, commonly referred to as image-guided therapy (iGT) devices. These are exclusively illustrative examples only.

[0004] Modern medical imaging equipment presents unusual challenges from a maintenance perspective. These imaging devices can be very complex, for example, in some medical imaging systems there may be approximately 19,000 service parts that need to be maintained. These devices are also used for an increasing variety of medical imaging types and procedures: for example, while iGT machines have traditionally been used for a limited number of procedures (such as diagnostic procedures in preparation for open heart surgery), iGT machines are now used for a wider range of minimally invasive procedures (such as heart valve replacement). The complexity of the machines can also be understood in terms of data generation: for example, the medical imaging equipment installed base may generate 10TB / year of log data, which is 2-4GB per day (mostly in the form of compressed text files).

[0005] Another difficulty in diagnosing problems with medical imaging equipment is that some data sources may have data quality issues. For example, uncertainty or quality in the content of service call data may be associated with human determinants (in this case, decisions are made in the field based on factors that cannot always be quantified (e.g., customer relationships)). In addition, when using past service history to predict future problem resolution, some machine problems may be resolved in multiple ways (calibration vs. part replacement), resulting in ambiguity.

[0006] New and improved systems and methods are disclosed below. Summary of the invention

[0007] In one disclosed aspect, a non-transitory storage medium stores instructions that can be read and executed by an electronic processor to perform a predictive maintenance alert method, the method comprising: receiving time-stamped machine log data and time-stamped service log data of a medical imaging device via an electronic network; deriving features based on the received time-stamped machine log data and time-stamped service log data; applying a set of models of component groups to the derived features to generate a maintenance alert, wherein each component of the medical imaging device is exclusively a member of a single component group; and sending the generated maintenance alert to a service center via the electronic network.

[0008] In another disclosed aspect, a predictive maintenance alert device includes a server computer operatively connected to a computer network to receive time-stamped machine log data and time-stamped service log data from a medical imaging device via the electronic network, and transmit a maintenance alert to a service center via the electronic network. A non-transitory storage medium stores instructions that can be read and executed by the server computer to perform a predictive maintenance alert method, the method including deriving features based on the received time-stamped machine log data and time-stamped service log data, and applying a set of models of component assemblies to the derived features to generate a maintenance alert. Each model in the set of models of component assemblies includes a heterogeneous model, the heterogeneous model including a machine learning analysis model representing the component assembly, the machine learning analysis model having a statistical remaining useful life model of components in the component assembly embedded therein.

[0009] In another disclosed aspect, a predictive maintenance alert method includes: receiving time-stamped machine log data and time-stamped service log data of a medical imaging device on a server computer via an electronic network; deriving features based on the received time-stamped machine log data and time-stamped service log data, including deriving at least one of: (i) constructed failure mode features, each feature including a combination of two or more features that together represent a failure mode of a component, and (ii) constructed failure solution features, each feature including a combination of two or more features that together represent a failure of the component and a solution to the failure of the component; applying a set of models of a component group to the derived features to generate a maintenance alert; and transmitting the generated maintenance alert from the server computer to a service center via the electronic network. The deriving and applying are suitably performed by the electronic server.

[0010] One advantage resides in providing maintenance alerts for complex medical imaging systems that often include ten thousand or more parts.

[0011] Another advantage resides in providing such maintenance alerts while reducing component level and component group level ambiguity.

[0012] Another advantage resides in providing such maintenance alerts while reducing component-level and component-group-level dependencies.

[0013] Another advantage resides in providing maintenance alerts that integrate machine learning analytical models at a component group level with statistical remaining useful life models of components.

[0014] Another advantage resides in providing such maintenance alerts with features built in to effectively capture and address failure modes.

[0015] Another advantage resides in providing such maintenance alerts using features built in to efficiently capture and handle different fault resolutions.

[0016] Another advantage resides in providing maintenance alerts that maximize correct maintenance alerts while limiting false maintenance alerts, thereby providing effective alerts without increasing the burden on service personnel.

[0017] A given embodiment may provide none of the aforementioned advantages, provide one, two, more, or all of the aforementioned advantages, and / or may provide other advantages, as will become apparent to one of ordinary skill in the art upon reading and understanding this disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention may take the form of various components and arrangements of components and various steps and arrangements of steps. The accompanying drawings are for the purpose of illustrating preferred embodiments only and should not be construed as limiting the present invention. In the accompanying drawings presenting log or service call data, specific identification information has been edited by using an overlaid edit box.

[0019] Figure 1 A predictive maintenance alert device and its surroundings are schematically illustrated, including a medical imaging device for which a maintenance alert is generated and a service center to which the maintenance alert is communicated.

[0020] Figure 2 An illustrative snippet of machine log data is shown.

[0021] Figure 3 Illustrative service log data for a service call is shown.

[0022] Figure 4 Illustrative service log data for another service call is shown.

[0023] Figure 5 A process for training a component assembly model of a predictive maintenance alarm device according to claim 1 is schematically shown.

[0024] Figure 6 Shown by Figure 1 An illustrative display of a maintenance alert generated by the predictive maintenance alert device and appropriately displayed on a display at the service center. DETAILED DESCRIPTION

[0025] In maintenance services, three typical approaches are: reactive maintenance, proactive maintenance, and predictive maintenance. The differences between these three approaches have to do with who initiates the call. In reactive maintenance, the customer initiates the call; whereas in proactive or predictive maintenance, the service provider will initiate the call. Predictive and proactive maintenance rely on statistical, machine learning, data mining, and / or optimization models that are developed using historical data related to the equipment to be maintained. For medical imaging devices, this data typically falls into the following categories: service call logs (containing information about when the customer called to report a maintenance-related issue and what actions were taken to resolve the issue); and machine log data (here broadly understood to cover usage data, i.e., information about how often each part of the system is used, and sensor data, i.e., information extracted from sensors, such as temperature, etc.).

[0026] Predictive maintenance of medical imaging equipment is particularly challenging due to the complexity of the equipment and the large amount of log data generated. To illustrate, some image-guided therapy (iGT) systems manufactured by Royal Philips AG include approximately 19,000 different parts to maintain. In terms of data logging, the installed base of medical imaging equipment may generate 10TB of data per year, or 2-4GB per day. These data volumes are for compressed data, mostly in the form of compressed text files.

[0027] In the predictive maintenance device and method embodiments disclosed herein, these difficulties are effectively addressed. The complexity of the hardware is addressed by identifying related groups of components that can be optimally modeled together (often at least 10,000 components that make up a medical imaging device). Since calls with common causes can be identified, the ambiguity of service calls can be reduced, which can enhance the robustness of predicted maintenance alerts (for example, achieving a "right first time" response to customer calls). The processing of large and heterogeneous features can be addressed by identifying relevant features and data resources related to the current component group. The "context" processing around machine log messages is partially addressed by using constructed features that combine features (such as log messages) that together represent failure modes or failure solutions. This helps to eliminate the ambiguity of specific errors in machine logs, which may be interpreted differently depending on the errors or messages around them. In some embodiments, a heterogeneous predictive maintenance model is used, and the predictive alerts can be made clearer through statistical models.

[0028] To address the challenges associated with providing useful maintenance alerts for complex medical imaging equipment, the disclosed predictive maintenance alert apparatus and method receives inputs including service call logs and machine log data (including usage data and sensor data). Groups of components (e.g., the ~19,000 components included in the iGT system) that are best grouped together in a predictive model are determined (i.e., they are associated with common root causes). In a training phase, relevant service calls for the group of components are found, and appropriate features are selected and / or established (e.g., built features that capture both failures and their solutions, and specific failure modes of components). The group of components model is then optimized.

[0029] To provide an illustrative example, consider the component group of the foot switch of an iGT device. The foot switch is a component that controls fluorescent inspection and exposure when using the iGT system. This group of components may include a cable to which the foot switch is connected to the foot switch, a component that enables X-rays in response to the operation of the foot pedal, and possibly other related components. To build a model for this "foot switch" component group, first select the components in the group. This includes hardware (e.g., selected component ids covering 15 years of hardware development) and software (e.g., selected units covering 15 years of software development). Next, identify calls related to the foot switch component group in the service call data (again, these calls may contain specific components attached to the foot switch, such as an attached cable, which, when disconnected, can generate similar error messages in the machine log file). The constructed features can identify, for example, situations where the foot switch is used but the X-ray is not enabled (i.e., the foot switch component is associated with a specific failure mode). The frequency and repetition of this constructed feature are optimized for model performance. The model is then optimized for performance based on business needs and domain expertise.

[0030] refer to Figure 1, schematically illustrates an exemplary predictive maintenance alert device and its associated surrounding environment, which includes a medical imaging device 10 (for which a maintenance alert is generated) and a maintenance center 12 in communication therewith. By way of non-limiting example, the illustrative medical imaging device 10 is a hybrid imaging system including a first imaging modality 14 and a second imaging modality 16 served by a common patient couch or transport device 18. For example, the first modality 14 may be transmission computed tomography (CT) imaging, and the second modality 16 may be positron emission tomography (PET) imaging (thus, the medical imaging device 10 would be a PET / CT imaging device). These are exclusively examples only, and more generally, the medical imaging device that is monitored to issue predictive maintenance alerts may be (by way of further non-limiting illustration) a magnetic resonance imaging (MRI) device, a CT imaging device, an emission imaging device such as a PET imaging device, or a gamma camera for single photon emission computed tomography (SPECT) imaging, a hybrid imaging device such as PET / CT or SPECT / CT, or an imaging device designed for guiding a biopsy or other interventional medical procedures, commonly referred to as an image-guided therapy (iGT) device.

[0031] Service Center 12 Figure 1 1 is represented by a service center workstation or service center computer 20, which includes a display 22 and one or more user input devices (e.g., an illustrative keyboard 24 and a mouse 26). The service center workstation or computer 20 can perform various service support functions, such as mapping the locations of medical institutions and various medical imaging devices that are serviced by the service center 12, coordinating, tracking and recording service calls performed by human service agents, queues or docking service requests received from customers, etc. In connection with the predictive maintenance alert device disclosed herein, the service center workstation or computer 20 is programmed to receive and display maintenance alerts generated by the predictive maintenance alert device. The service center personnel can then decide whether to follow up on any specific maintenance alert received from the predictive maintenance alert device.

[0032] Continue to refer Figure 1 The medical imaging device 10 or a system or device associated with the medical imaging device 10 generates a machine log 30 (again, as used herein, the machine log 30 may include medical imaging device usage log data and / or sensor data collected by sensors of the medical imaging device). Similarly, a maintenance service log 32 is maintained, which records information about service calls, i.e., maintenance performed on the medical imaging device 10.

[0033] Brief reference Figure 2, shows an illustrative snippet 30f of a possible form of a machine log 30. In this illustrative example, each event (or message) occupies one row of the machine log data. The leftmost column, labeled "EventTimeStamp," stores the timestamp of the event or message that was logged. The "eventID" column stores a unique identifier for each machine log entry. The "Description" column stores a semi-structured text description of the event or message. The "AdditionalInfo" column stores further information, again in a semi-structured text format. The "EventCategory" column stores the classification of each event or message in the machine log. Figure 2 The example of is only one non-limiting illustrative example, and the machine log data may take various other forms, such as including additional, fewer, and / or different columns, different syntax, etc. Typically, each entry of the machine log 30 is time-stamped and includes a text-based description of the event or message, which may include information such as sensor data, detected component failures, entries recording the start of an imaging process, and subsequent entries recording operations performed during the imaging process, etc.

[0034] refer to Figure 3 , shows an illustrative entry 32f1 of a possible format for a service log 32 for a particular illustrative service call. As can be seen, each service call is again timestamped (the term "CallOpenDate" is used here; "CallCloseDate" is also contemplated), and a semi-structured text description of the service call is stored, including the information collected and the remedial actions taken to resolve the issue.

[0035] refer to Figure 4 , shows another illustrative entry 32f2 of a possible format for a service log 32 for another illustrative service call. Again, the service call entry is timestamped by "CallOpenDate" and stores a semi-structured text description of the service call. Figure 3 and Figure 4The examples are exclusively non-limiting illustrative examples, and the service log data may take a variety of other forms, such as including additional, less and / or different information in different formats and / or using different syntax. Typically, each entry in the service log 32 is timestamped and includes a text-based description of the service call, including any information collected and the remedial actions taken to resolve the problem. In some cases, if the service call does not actually resolve the problem, the service call entry may not include a solution to the problem. In addition, the service call entry may contain irrelevant information - for example, the service personnel may perform routine maintenance operations that are unrelated to the component failure that originated the service call. There may also be "noise" in the sense that in some cases the remedial actions taken may be "wrong", that is, they do not actually resolve the root cause of the component failure that was the basis for the service call.

[0036] Return to reference Figure 1 , the predictive maintenance alert device 40 includes a server computer 42 that executes instructions stored in a non-transitory storage medium (not shown), which can be read and run by the server computer 42 to perform the predictive maintenance alert method disclosed herein. The server computer 42 can be implemented in different ways, such as being implemented as a single server computer, or being implemented as two or more server computers that are operatively connected to perform concurrent processing to implement the predictive maintenance alert method, or being implemented as a combination of two or more computers that are operatively connected in an ad hoc manner to perform the predictive maintenance alert method (e.g., cloud computing resources). Although server computers (or groups or ad hoc combinations of server computers) are generally preferred because they can provide a large amount of computing power, in other embodiments, it is contemplated that the disclosed predictive maintenance alert method can be implemented on a desktop computer or other electronic processor if it has sufficient computing power. The non-transitory storage medium can be, for example, a hard disk drive, RAID or other magnetic storage medium; flash memory, solid state drive (SSD) or other electronic storage medium; optical disk or other optical storage medium; various combinations thereof; and the like.

[0037] The predictive maintenance alert device 40 processes the machine log 30 and the service log 32 of the medical imaging device 10 to generate a maintenance alert 44, which is sent to the service center 12 (e.g., more specifically to the service center workstation or computer 20 in the illustrative embodiment). To this end, an electronic processor (an exemplary server computer) 42 is operatively connected to an electronic network 46 to receive the time-stamped machine log data 30 and the time-stamped service log data 32 via the electronic network 46 and send the maintenance alert 44 to the service center 12 via the electronic network 46. The electronic network 46 is connected to the service center 12 via the electronic network 46. Figure 146 is schematically indicated by data flow box arrows, and in physical implementation may be implemented in various ways as a hospital data network, the Internet, a data networking infrastructure provided by a commercial Internet Service Provider (ISP), and / or other. Different data networking hardware may be employed for data flowing between the medical imaging device 10 and the predictive maintenance alert device 40, on the one hand, and between the predictive maintenance alert device 40 and the service center 12, on the other hand. For example, the former may include a hospital data network connected to the Internet, which is connected to a local area network of which the server computer 42 is a part, while the latter may (again, by way of non-limiting illustrative example) include only the local area network hosting the server computer 42 when the server computer is located at the service center 12, or may also include a commercial ISP if located elsewhere. These are just illustrative examples of some possible implementations of the non-limiting electronic network 46.

[0038] like Figure 1 As further schematically shown in FIG. 4 , the service center 12 or server computer or other computing resource may optionally perform processing of the maintenance alert 44. For example, an electronic service schedule 48 may be implemented on a server or other computer to automatically or semi-automatically schedule a service call based on the maintenance alert 44. For example, a proposed service call may be added to the service schedule 48, which must be accepted by the customer via a web-based customer interface to the service schedule 48 or the like. The customer may choose to reschedule or cancel the proposed service call, depending on the availability of the service personnel. Additionally or alternatively, an electronic inventory 49 may be implemented on such a server or other computer to automatically or semi-automatically order parts that are needed (or may be needed) to resolve the maintenance alert 44, and / or to restock these parts in anticipation of a maintenance call to resolve the maintenance alert. The electronic inventory 49 may also ship parts from physical inventory to a customer site in preparation for an anticipated service call. A maintenance personnel may be required to approve the parts order or shipment, and the customer is preferably not billed for parts unless and until the parts are installed in the customer's equipment.

[0039] In a typical embodiment, the logs 30, 32 are uploaded to the predictive maintenance alert device 40 on a daily basis, for example, each morning when the machine 10 is brought online to provide imaging services. Figure 110; however, the logs may be stored in any network-accessible server, RAID array, cloud storage cabinet, etc. In some embodiments, the logs may be maintained on a server or other network-based resource maintained by the manufacturer or service provider of the medical imaging device. To save bandwidth and speed up data transmission, a data compressor 50 is executed on a computer (not shown, disposed on or operably connected to the source storing the logs) to compress the data logs 30, 32 using any suitable compression algorithm (e.g., zip compression or a variation thereof, tar compression or a variation thereof, etc.). Even with such compression, the compressed data records 30, 32 may constitute a large amount of data for a large medical imaging device. In the predictive maintenance alert device 40, the server computer 42 executes a data decompressor 44 to decompress the compressed log data received at the predictive maintenance alert device 40 to reproduce (a copy of) the time-stamped machine log data 30 and the time-stamped service log data 32 at the predictive maintenance alert device 40. Although data compression is preferred to reduce bandwidth, increase data transfer speeds, and reduce load on the electronic network 46, it is contemplated to omit the data compression components 50, 52 and instead transmit the log data 30, 32 in an uncompressed form.

[0040] The predictive maintenance alert device 40 generates a maintenance alert 44 for the medical imaging device 10 by performing a time-stamped feature extraction 60 (optionally including derived constructed features 62, 63 as disclosed herein) based on the received time-stamped machine log data 30 and the time-stamped service log data 32, and applying a set 64 of models of component groups to derive features to generate the maintenance alert 44. By way of non-limiting illustration, in Figure 1 , the set of models 64 of component assemblies is depicted as containing N models for N component assemblies.

[0041] Feature extraction 60 is suitably performed by including parsing machine logs 30 (see, e.g., Figure 2 ) and service log 32 (see Figure 3 and Figure 4 ) (and optionally additional information). Typically, the choice of function depends on the terminology and syntax used by the machine and service logging software. Some features may be numeric; for example, a machine log entry reporting a temperature sensor reading may be expressed as a (sensor identifier, <temperature value>) pair, where <temperature value> indicates the temperature read by the sensor. Some features may be binary, or may take a value selected from a discrete set of possible values: for example, in Figure 4In the service log fragment 32f2 of FIG. 32f3 , a possible feature might be (table transverse brake component status = inoperative), where, for this type of feature, the two possible values ​​are typically "operative" or "inoperative". Parsing of the semi-structured text includes identifying component names or identifiers, and associating log entries containing those names or identifiers with the named or identified components. The extracted features may also be categorized by event category or other information. The timestamp of each entry converted to a feature is preferably retained as the timestamp of the feature.

[0042] The constructed features 62, 63 are constructed as two or more constituent features, where each constituent feature is a combination extracted from a log entry. These combinations are selected to be more informative than individual features extracted from a single log entry. From another perspective, the constructed features provide additional scenarios that can provide information for predicting incipient maintenance issues.

[0043] By way of illustration, one type of constructed feature is to establish a failure mode feature 62 that suitably includes a combination of two or more features that together represent a failure mode of the component. For example, considering the illustrative foot switch component mentioned above, the constructed failure mode feature 62 can be constructed as a combination of a log entry indicating activation of the foot switch and a subsequent entry indicating that the X-ray is not enabled. This constructed feature captures the failure mode of the foot switch, where activating the foot switch fails to perform the intended function of enabling the X-ray. In constructing this constructed feature, timestamps are preferably taken into account, that is, the constructed feature preferably requires that the timestamp of the log entry indicating the X-ray is after the log entry indicating that the foot switch has been activated, and in addition, it may also be required that there are no timestamped entries in between indicating operations that will disable the X-ray.

[0044] As another illustrative example, another type of constructed feature is a constructed fault solution feature 63, which suitably includes a combination of two or more features that together represent a fault of a component and a solution to the fault of the component. Figure 3 The service log fragment 32f1 can establish a constructed fault solution feature 63, which includes a combination of a feature indicating a fault "tabletop movement difficulty" (or in other embodiments, a more general term such as "tabletop movement problem") and a solution indicating "clean rails". Figure 4 The maintenance log fragment 32f2 can establish a constructed fault solution feature 63, which includes a combination of the feature (workbench lateral brake assembly status = not working) and the feature "lateral brake assembly has been replaced".

[0045] The model set 64 of component groups has specific features for maintenance prediction of large medical imaging equipment. The assignment of components to component groups is preferably performed to group together components related to common root causes. In some embodiments, each component of the medical imaging device is exclusively a member of a single component group and is therefore modeled by a single model in the model set 64. For example, if two different models generate a maintenance alert for the same component, this approach can reduce ambiguity that may result.

[0046] In other contemplated embodiments, some components may be assigned to two or more different groups and thus may be modeled by different models. For example, considering the example of a foot pedal operable to enable X-ray, a component that receives input from the foot switch and generates a signal to enable X-ray may effectively be included in both the foot switch component group and the X-ray component group. In this case, the membership of a component in multiple groups is preferably kept low, e.g., preferably no more than two or three groups.

[0047] like Figure 1 As schematically shown in , in some embodiments, at least one model (and preferably all models) in a set of models 64 of a component assembly includes a heterogeneous model comprising a machine-learned analytical model 70 representing the component assembly, with embedded statistical remaining useful life models 72 for the components of the component assembly. This type of heterogeneous model advantageously utilizes machine learning to provide the analytical model 70 to empirically capture the complex interrelationships between the components of the component assembly, while also using the remaining useful life model 72 to capture the statistical failure probability and assign a time frame when the problematic component is likely to fail.

[0048] refer to Figure 5 , describes a suitable method for building a collection 64 of models. The illustrative method is implemented by a server computer 42 executing instructions read from a non-transitory storage medium, but may alternatively be implemented on other computers. Inputs include an imaging system specification 80 (e.g., identification of components of a component group whose models are being trained along with other significant parameters of the medical imaging device), training log data 82, and component life data 84. The training log data 82 includes time-stamped machine log data and time-stamped service log data, and may be obtained from the medical imaging device 10 for which the collection of models is intended, and / or may be obtained from one or more other medical imaging devices having the same type of features.

[0049] In operation 86, component groups are defined. This is typically a manual operation performed by system engineers who have expertise as to which components function together as sub-components as sub-systems of a medical imaging device. Operation 88 parses the training log data 82. The parsing may be similar to the time-stamped feature extraction 60 already described, but optionally more features may be extracted than the features of extraction 60 to perform statistically based feature selection, for example, based on the discriminability of the features. Operation 88 also extracts positive training data and negative training data from the parsed training log data. Positive training data includes training data with a timestamp before a component failure of the component group, while negative training data includes training data with a timestamp after a component failure of the component group, or training data from a medical imaging device where a component failure of the component group has never occurred.

[0050] In operation 90, machine learning is applied to train an analytical model using the positive training data and the negative training data. Operation 88 optionally includes a feature selection stage 92 that selects a subset of the features extracted in operation 88 based on relevance or discriminability for a given model. For example, relevance may be determined based on whether a feature is associated with any component in a component group, while discriminability may be determined based on whether the feature occurs more frequently in the positive training data relative to the negative training data, or vice versa. (Conversely, if a feature occurs with similar frequency in both the positive and negative training data, its discriminability may be low and may be discarded during feature selection).

[0051] The component life data 84 is used to construct statistical remaining useful life models 72 for the components in the component group. These component-level statistical remaining useful life models 72 are embedded in the analytical model 70 for the component group, which is then trained through machine learning to optimize the model relative to the target 94. Various target formats can be used; however, in a preferred embodiment, the target includes maximizing true positive maintenance alerts for the component group (where "true positive" means that the model correctly predicts maintenance operations that actually occurred in the positive training data) subject to an upper limit on false positive maintenance warnings for the component group (where "false positive" means that the model incorrectly predicts maintenance alerts that did not appear in the negative training data). This choice of target 94 advantageously maximizes the rate of "correct" maintenance alerts (measured in true positive rate) that are useful in driving maintenance activities while avoiding issuing more than a threshold number of false positive maintenance alerts that could otherwise overwhelm the service center's resources 12. The trained model then forms a previously referenced Figure 1 The described predictive maintenance alert device 40 employs a collection of models 64 of component assemblies.

[0052] Return to reference Figure 1 And further reference Figure 6 ,exist Figure 6 It is shown in Figure 1 An illustrative display (referred to as a "monitoring panel" 96) of possible maintenance alerts generated by the predictive maintenance alert device 40 and appropriately displayed on the display 22 at the service center 12. Figure 6 As shown, the two columns on the left identify the "Country" and "Site name" of the medical imaging device to which the maintenance alert belongs. (It should be understood that Figure 1 The predictive maintenance alert device 40 can be used not only for Figure 1 The monitoring and issuing of maintenance alerts for a single illustrative medical imaging device 10 is shown, but can be used for an entire fleet of medical imaging devices that can be deployed in different hospitals around the world. Generally speaking, the medical imaging devices may be of different types, such as CT, MRI, iGT, etc.). Figure 1 , the contents of the "Site name" column are deleted. As indicated by the "Category," "Device type," and "System #" columns, various other information about each alarm may be provided. For example, the "Device type" column identifies the brand and model of the medical imaging device (the illustrative "Allura XPer" refers to a Philips Allura Xper FD20 / 10 dual-wing hybrid cardiovascular X-ray system), and the "System #" is the serial number or other unique identifier of the particular instance or installation. The "Aggregate Title" identifies the component group to which the maintenance alarm belongs (e.g., the Geometry Syncnet component group in the illustrative example). The column titled "Priority" serves two purposes: the label on each icon identifies the priority, and the icon itself can be selected by clicking the mouse 26 (or by touching the touch screen, or by some other user input device) to bring up a pop-up window providing information about the details of the maintenance alarm (e.g., the expected failure mode and the expected remedial actions required, and possibly other relevant information, such as the expected time range for failure obtained from the statistical component life model 72). Finally, the last column titled “Last Alert Date” indicates the date the maintenance alert was issued.

[0053] The present invention has been described with reference to the preferred embodiments. By reading and understanding the aforementioned detailed description, those skilled in the art may conceive of various modifications and variations. The purpose is that the present invention is understood to include all such modifications and variations as long as they fall within the scope of the claims or their equivalents.

Claims

1. A non-transitory storage medium storing instructions capable of being read and executed by an electronic processor (42) to perform a predictive maintenance alert method, the method comprising: receiving time-stamped machine log data (30) and time-stamped service log data (32) for a medical imaging device (10) via an electronic network (46); deriving features based on the received time-stamped machine log data and the time-stamped service log data; applying a set of models (64) of component groups to the derived features to generate maintenance alerts (44), each model in the set of models corresponding to a component group, wherein each component of the medical imaging device is exclusively a member of a single component group, training the models in the set of models using machine learning to optimize an objective (94) comprising maximizing true positive maintenance alerts for the component group subject to an upper bound on false positive maintenance alerts for the component group, the models in the set of models being trained using positive training data and negative training data; and The generated maintenance alert is transmitted to a service center (12) via the electronic network.

2. The non-transitory storage medium according to claim 1, wherein: At least one model in the set (64) of models of component assemblies comprises a heterogeneous model, the heterogeneous model comprising a machine learning analytical model (70) representing the component assembly, the machine learning analytical model having an embedded statistical remaining useful life model (72) of the components in the component assembly.

3. The non-transitory storage medium according to any one of claims 1 to 2, wherein: The derivation of features includes deriving constructed failure mode features (62), each constructed failure mode feature including a combination of two or more features that together represent a failure mode of the component.

4. The non-transitory storage medium according to any one of claims 1 to 2, wherein: The derivation of features includes deriving constructed fault solution features (63), each constructed fault solution feature including a combination of two or more features, the combination of two or more features together representing a fault of a component and a solution to the fault of the component.

5. The non-transitory storage medium according to any one of claims 1 to 2, wherein: The stored instructions are readable and executable by the electronic processor (42) to further perform a machine learning method operating on training data (82), the training data comprising time-stamped machine log data and time-stamped service log data for at least one of the medical imaging device and one or more other medical imaging devices of the same type, the machine learning method comprising: extracting positive training data from the training data to train a model in the set of models of a component group, wherein the positive training data includes training data having a timestamp prior to a failure of a component in the component group; extracting negative training data from the training data to train a model in the set of models of the component set, wherein the negative training data includes training data having a time stamp after a failure of a component in the component set or training data from a medical imaging device in which a failure of the component in the component set has never occurred; and Machine learning is applied to train the model using the positive training data and the negative training data.

6. The non-transitory storage medium according to any one of claims 1 to 2, wherein: The received time-stamped machine log data (30) includes medical imaging device usage log data and sensor data collected by sensors of the medical imaging device.

7. A predictive maintenance alarm device comprising: a server computer (42) operatively connected to an electronic network (46) to receive time-stamped machine log data (30) and time-stamped service log data (32) from a medical imaging device (10) via the electronic network and to transmit maintenance alerts (44) to a service center (12) via the electronic network; as well as A non-transitory storage medium storing instructions that can be read and executed by the server computer to perform a predictive maintenance alert method, the predictive maintenance alert method comprising: deriving features based on the received time-stamped machine log data and the time-stamped service log data; and Applying a set of component assembly models (64) to the derived features to generate the maintenance alert, each model in the set of models corresponds to a component group, wherein each component of the medical imaging device (10) is exclusively a single component a member of a group, and wherein each model in the set of models of the component group Both include heterogeneous models, the heterogeneous models including machine learning representations of the component groups The machine learning analysis model (70) has the A statistical remaining useful life model (72) of the components in the component group, using a machine learning to train said model in said set of models to optimize an objective (94), The target includes a limit on false positive maintenance alerts for the component group. The model is based on the principle of maximizing true positive maintenance alerts for the component group. The model in the combination is trained using positive training data and negative training data.

8. The predictive maintenance alert device of claim 7, wherein: The derivation of features includes deriving constructed failure mode features (62), each constructed failure mode feature including a combination of two or more features that together represent a failure mode of the component.

9. The predictive maintenance alert device according to any one of claims 7-8, wherein: The derivation of features includes deriving constructed fault solution features (63), each constructed fault solution feature including a combination of two or more features, the combination of two or more features together representing a fault of a component and a solution to the fault of the component.

10. The predictive maintenance alert device according to any one of claims 7-8, wherein: The stored instructions are readable and executable by the server computer (42) to further perform a machine learning method operating on training data (82), the training data comprising time-stamped machine log data and time-stamped service log data for at least one of the medical imaging device and one or more other medical imaging devices of the same type, the machine learning method comprising: extracting positive training data from the training data to train a model in the set of models of a component group, wherein the positive training data includes training data having a timestamp prior to a failure of a component in the component group; extracting negative training data from the training data to train a model in the set of models of a component group, wherein the negative training data includes training data having a time stamp after a failure of a component in the component group or training data from a medical imaging device in which a failure of the component in the component group has never occurred; and Machine learning is applied to train the model using the positive training data and the negative training data.

11. The predictive maintenance alert device according to any one of claims 7-8, wherein: The collection (64) of component group models models at least 10,000 components of the medical imaging device.

12. The predictive maintenance alert device according to any one of claims 7-8, wherein: The predictive maintenance alert method further includes at least one of the following: Maintaining an electronic service schedule (48), including scheduling service calls to remedy generated maintenance alerts; as well as A maintenance electronic list (49) includes ordering parts for the medical imaging device (10) required in anticipation of remedying the generated maintenance alert.

Images