EP4681040A1 - Fault detection modeling - Google Patents

Abstract

An example method of predicting the state of an instrumented mechanical component is described herein. The example method includes receiving mechanical data, where the mechanical data includes measurements of a least one mechanical parameter associated with an instrumented mechanical component; inputting the mechanical data into a trained machine learning model; and predicting, using the trained machine learning model, a state of the instrumented mechanical component.

Description

FAULT DETECTION MODELING

RELATED APPLICATION

[0001] This application claims priority to, and the benefit of, U.S. Provisional Patent Application No. 63/452,789, filed March 17, 2023, entitled “FAULT DETECTION MODELING,” which is incorporated by reference herein in its entirety.

BACKGROUND

[0002] Different types of forces on a system or device can have different effects and cause different types of stresses and damage to the mechanical components of that system. For example, sudden impacts can tend to cause certain types of mechanical failures, that might affect certain components. Conversely, prolonged stresses, wear, and/or strains can cause different types of failures that affect different components.

[0003] Detection of wear or damage is useful for maintaining or designing systems and devices, including mechanical systems and devices. For example, in a healthcare setting, medical devices can require consistent maintenance and pose significant safety risks if they fail unexpectedly.

[0004] Therefore, improved systems, methods, and devices for maintaining, designing, and modeling devices and systems can benefit the maintenance and reliability of those systems, including the mechanical durability of those systems.

SUMMARY

[0005] Methods and systems for predicting faults are disclosed herein.

[0006] An example method is described herein. The example method includes: receiving mechanical data, where the mechanical data includes a least one mechanical parameter associated with an instrumented mechanical component; inputting the mechanical data into a trained machine learning model; and predicting, using the trained machine learning model, a state of the instrumented mechanical component.

[0007] In some implementations, the state of the instrumented mechanical component includes at least one of: healthy, bent, cracked, secondary failure, or multiple faults. [0008] In some implementations, the at least one mechanical parameter includes stress data and vibration data.

[0009] In some implementations, the machine learning model includes fuzzy logic.

[0010] In some implementations, the machine learning model includes a supervised machine learning model.

[0011] In some implementations, the supervised machine learning model is a decision tree or a support vector machine classifier.

[0012] In some implementations, the supervised machine learning model is a multilayer perceptron.

[0013] In some implementations, the instrumented mechanical component includes a caster.

[0014] In some implementations, the instrumented mechanical component includes a wheel.

[0015] An example method for training a machine learning model, the method including: performing fatigue testing on an instrumented mechanical component; measuring mechanical data of the instrumented mechanical component, where the mechanical data includes a least one mechanical parameter associated with an instrumented mechanical component; assessing a state of the instrumented mechanical component; creating a dataset including the mechanical data and the state of the instrumented mechanical component; and training the machine learning model to predict the state of the instrumented mechanical component using the dataset.

[0016] In some implementations, the state of the instrumented mechanical component includes at least one of: healthy, bent, cracked, secondary failure, or multiple faults.

[0017] In some implementations, the at least one mechanical parameter includes stress data and vibration data.

[0018] In some implementations, the instrumented mechanical component includes a caster.

[0019] In some implementations, the instrumented mechanical component includes a wheel.

[0020] In some implementations, the machine learning model includes fuzzy logic. [0021] In some implementations, the machine learning model includes a supervised machine learning model.

[0022] In some implementations, the supervised machine learning model is a decision tree or a support vector machine classifier.

[0023] In some implementations, the supervised machine learning model is a multilayer perceptron.

[0024] An example system is described herein. The example system includes an assistive device, including a mechanical component; a mechanical sensor configured to measure at least one mechanical parameter associated with the mechanical component; and a processor and a memory, the memory operably coupled to the processor, the memory having computerexecutable instructions stored thereon that, when executed by the processor, cause the processor to: receive data from the mechanical sensor; input the data into a trained machine learning model; and predict, using the trained machine learning model, a state of the mechanical component using a trained machine learning model.

[0025] In some implementations, the processor and the mechanical sensor are operably coupled by a wireless communication link.

[0026] In some implementations, the mechanical component is a caster or a wheel.

[0027] In some implementations, the sensor is a strain sensor or a vibration sensor.

BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The components in the drawings are not necessarily to scale relative to each other. Like reference numerals designate corresponding parts throughout the several views.

[0029] FIG. 1 illustrates an example method of predicting a state of a mechanical component.

[0030] FIG. 2A illustrates an example method of training a machine learning model to predict the state of a mechanical component.

[0031] FIG. 2B illustrates an example method of training a machine learning model to predict the state of a mechanical component including preprocessing and model validation.

[0032] FIG. 3 illustrates a system block diagram of an example implementation.

[0033] FIG. 4 is an example computing device. [0034] FTG. 5 illustrates an example wheelchair and caster, according to implementations of the present disclosure.

[0035] FIG. 6 illustrates an example stem fracture and bent stem of a caster.

[0036] FIG. 7 illustrates an example installation of a strain gauge and accelerometer, according to implementations of the present disclosure.

[0037] FIG. 8 illustrates an example experimental results including failure modes and fatigue damage, according to implementations of the present disclosure.

[0038] FIG. 9 illustrates an example test fixture for recording experimental data from casters, according to implementations of the present disclosure.

[0039] FIG. 10 illustrates an example method for training a machine learning model using data collected from healthy stems and faulty stems, according to implementations of the present disclosure.

[0040] FIG. 11 illustrates an example stress number (SN) curve for bending stress and number of stress cycles, according to implementations of the present disclosure.

[0041] FIG. 12 illustrates a table of example results including SumSq, DF, MeanSq, F and pValue, according to implementations of the present disclosure.

[0042] FIG. 13 A illustrates an example analysis of patient age and drive-per-month, according to an example implementation of the present disclosure.

[0043] FIG. 13B illustrates a regression fit for adjusted data, including a 95% confidence bounds, according to an example implementation of the present disclosure.

[0044] FIG. 14 illustrates example effects of predictor variables on time to failure for example implementations of the present disclosure.

[0045] FIG. 15 illustrates an example analysis of chair age and drive-per month, according to an example implementation of the present disclosure.

[0046] FIG. 16 illustrates an example analysis of user sex and drive-per month, according to an example implementation of the present disclosure.

[0047] FIG. 17 illustrates an example neural network that can be used to predict failure of a system, for example a wheelchair, according to implementations of the present disclosure.

[0048] FIG. 18 illustrates example training, validation, and test learning curves for the neural network illustrated in FIG. 17. [0049] FTG. 19 illustrates example machine learning training over 1 1 epochs, according to implementations of the present disclosure.

[0050] FIG. 20 illustrates an error histogram of neural network training, according to implementations of the present disclosure.

[0051] FIG. 21A illustrates an example analysis of adjusted time to failure for a young patient group, before outlier removal, according to implementations of the present disclosure.

[0052] FIG. 2 IB illustrates an example analysis of adjusted time to failure for a young patient group, after outlier removal, according to implementations of the present disclosure.

[0053] FIG. 22 illustrates an example parallel plot analysis of drive per month vs. risk group where Early repair = 1 represents high risk, and Early Repair = 0.

[0054] FIG. 23 illustrates study parameters for a telematics study, according to implementations of the present disclosure.

[0055] FIG. 24 illustrates an example analysis of telematics data for a low-risk group and a high-risk group, according to implementations of the present disclosure.

[0056] FIG. 25 illustrates an example fractured wheelchair caster stem bolt, according to implementations of the present disclosure.

[0057] FIG. 26 illustrates elastic Mean Strain Correlation to Plastic Deformation in an example stem bolt fatigue test. R²=0.99, according to implementations of the present disclosure.

DETAILED DESCRIPTION

[0058] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. Methods and materials similar or equivalent to those described herein can be used in the practice or testing of the present disclosure. As used in the specification, and in the appended claims, the singular forms “a,” “an,” “the” include plural referents unless the context clearly dictates otherwise. The term “comprising” and variations thereof as used herein is used synonymously with the term “including” and variations thereof and are open, non-limiting terms. The terms “optional” or “optionally” used herein mean that the subsequently described feature, event or circumstance may or may not occur, and that the description includes instances where said feature, event or circumstance occurs and instances where it does not. Ranges may be expressed herein as from "about" one particular value, and/or to "about" another particular value. When such a range is expressed, an aspect includes from the one particular value and/or to the other particular value. Similarly, when values are expressed as approximations, by use of the antecedent "about," it will be understood that the particular value forms another aspect. It will be further understood that the endpoints of each of the ranges are significant both in relation to the other endpoint, and independently of the other endpoint. While implementations will be described for measuring/detecting the state of casters used for an assistive device such as a wheelchair, it will become evident to those skilled in the art that the implementations are not limited thereto, but are applicable for any other type mechanical fault detection/prediction. This may include, but is not limited to, predicting the state of other mechanical components, non-limiting examples of which include any kind of wheel, strut, and/or other support.

[0059] The term “artificial intelligence” is defined herein to include any technique that enables one or more computing devices or comping systems (i.e., a machine) to mimic human intelligence. Artificial intelligence (Al) includes, but is not limited to, knowledge bases, machine learning, representation learning, and deep learning. The term “machine learning” is defined herein to be a subset of Al that enables a machine to acquire knowledge by extracting patterns from raw data. Machine learning techniques include, but are not limited to, logistic regression, support vector machines (SVMs), decision trees, Naive Bayes classifiers, and artificial neural networks. The term “representation learning” is defined herein to be a subset of machine learning that enables a machine to automatically discover representations needed for feature detection, prediction, or classification from raw data. Representation learning techniques include, but are not limited to, autoencoders. The term “deep learning” is defined herein to be a subset of machine learning that enables a machine to automatically discover representations needed for feature detection, prediction, classification, etc. using layers of processing. Deep learning techniques include, but are not limited to, artificial neural network or multilayer perceptron (MLP).

[0060] Machine learning models include supervised, semi-supervised, and unsupervised learning models. In a supervised learning model, the model learns a function that maps an input (also known as feature or features) to an output (also known as target or targets) during training with a labeled data set (or dataset). In an unsupervised learning model, the model learns patterns (e.g., structure, distribution, etc.) within an unlabeled data set. In a semisupervised model, the model learns a function that maps an input (also known as feature or features) to an output (also known as target or targets) during training with both labeled and unlabeled data.

[0061] Described herein are systems and methods for predicting the state of mechanical components. As used herein, the state of a mechanical component can refer to any description of the condition of the mechanical component, e.g., the presence or absence of a defect, and/or a degree of wear, effectiveness, etc. Non-limiting examples of states of a mechanical component can include the presence, absence, or degree of any or all of: healthy, bent, cracked, secondary failure, or multiple faults.

[0062] An example method 100 for predicting the state of a mechanical component is shown in FIG. 1. The example method 100 can include receiving mechanical data at step 102. The mechanical data can include measurements of one or more mechanical parameter(s) associated with an instrumented mechanical component. The present disclosure contemplates that different mechanical data can be used to predict different types of failures of different types of mechanical components. As a non-limiting example, mechanical data including the distance traveled (e.g., milage) can be used to predict the state of a powered vehicle (e.g., a powered wheelchair). As another non-limiting example, mechanical data including the

[0063] The mechanical component can include any type of mechanical component. Non-limiting examples of mechanical components include casters, or wheels. FIG. 5 illustrates an example wheelchair 500 including a caster 510. The caster 510 includes a upper stem bearing 512, a stem housing 514, and a stem 516. The stem 516 rotates on the bearing 512 to allow the caster 510 to spin in the direction that the wheelchair is traveling. A fork 518 couples the stem 516 to an axel 520 that the wheel 522 can spin around. Because the stem 516 couples the caster 510 to the wheelchair 500, the stem 516 can receive stresses from the caster 510 and therefore the stem 516 is subject to fatigue both over time and in response to shocks (e.g., impacts caused by an uneven surface or collisions).

[0064] It should be understood that “mechanical component” as used herein, can refer to a single mechanical component or assemblies of separate mechanical components. For example, in some implementations of the present disclosure, the mechanical component can include a system of mechanical components (e.g., a powered or unpowered wheelchair). In implementations of the present disclosure where the mechanical component includes a system of mechanical components, the systems and methods described herein can be configured to predict or estimate the state of the system, one or more of the mechanical components that make up the system, and/or both the system and the components of the system. Alternatively, in some implementations of the present disclosure, the mechanical component can be a single mechanical component.

[0065] It should also be understood that the mechanical component can include electrical and/or chemical elements. For example, in some implementations the mechanical component can be a battery.

[0066] It should be understood that an instrumented mechanical component can include the mechanical component and one or more instruments for measuring mechanical parameters (e.g., strain or vibration). This disclosure contemplates that the instruments can be located both outside the mechanical component, as well as instruments that are integrated into the mechanical component itself. In other words, the instruments can be any instrument configured to measure a mechanical parameter of the mechanical component. This can include instruments that are coupled to the mechanical components, both indirectly and directly. The instruments described herein can be sensors including, but not limited to, temperature sensors, strain sensors, displacement sensors, velocity sensors, and accelerometers. Non-limiting examples that can be used in implementations of the present disclosure include triaxial accelerometers, torsion strain gauges and axial bending strain gauges.

[0067] An example instrumented mechanical component is a stem housing 514 including a strain gauge 702 as shown in FIG. 7. The strain gauge can detect or measure strain in the stem housing 514, and the detected or measured strain can be used as mechanical parameters. Optionally, an accelerometer can be coupled to a wheelchair 500 adjacent to the stem housing 514. The accelerometer can measure or detect acceleration of the wheelchair 500. The acceleration measured by the accelerometer can be a mechanical parameter used by implementations of the present disclosure to predict a state of mechanical components of the wheelchair 500. For example, the acceleration of the wheelchair 500 can be used by implementations of the present disclosure to predict a failure of the stem housing 514, stem bearing 512 or stem 516 caused by accelerations over time.

[0068] The instruments in the instrumented mechanical component can be configured to measure any mechanical parameter that facilitates assessment of the mechanical component. Non-limiting examples of mechanical parameters include stress data and vibration data. In some implementations, the mechanical parameters can alternatively or additionally include measures of impact/shock, temperature changes, and any other mechanical property.

[0069] At step 104, the method can include inputting the mechanical data into a trained machine learning model. The trained machine learning model is operating in inference mode. In other words, the trained machine learning model has already been trained with a dataset. For example, the trained machine learning model can be a machine learning model trained according to the method described with reference to FIG. 2A or FIG. 2B. Optionally, in some implementations, the trained machine learning model is a supervised learning model such as a decision tree. A decision tree (DT) model is a supervised learning model that uses a hierarchal tree structure including a root node, branches, internal nodes, and leaf nodes to predict a target. This disclosure contemplates that the DT model can be implemented using a computing device (e.g., a processing unit and memory as described herein). DT models can be used for classification and regression tasks. DT models are trained with a data set (also referred to herein as a “dataset”) to maximize or minimize an objective function, for example a measure of the DT model’s performance, during training. DT models are known in the art and are therefore not described in further detail herein.

[0070] Optionally, in some implementations, the trained machine learning model is a supervised learning model such as or support vector machine (SVM). A support vector machine (SVM) is a supervised learning model that uses statistical learning frameworks to predict the probability of a target. This disclosure contemplates that the SVMs can be implemented using a computing device (e.g., a processing unit and memory as described herein). SVMs can be used for classification and regression tasks. SVMs are trained with a data set (also referred to herein as a “dataset”) to maximize or minimize an objective function, for example a measure of the SVM’s performance, during training. SVMs are known in the art and are therefore not described in further detail herein.

[0071] Optionally, in some implementations, the trained machine learning model is a supervised learning model such as an artificial neural network (ANN), and optionally a deep neural network. An artificial neural network (ANN) is a computing system including a plurality of interconnected neurons (e.g., also referred to as “nodes”). This disclosure contemplates that the nodes can be implemented using a computing device (e.g., a processing unit and memory as described herein). The nodes can be arranged in a plurality of layers such as input layer, output layer, and optionally one or more hidden layers. An ANN having hidden layers can be referred to as deep neural network or multilayer perceptron (MLP). Each node is connected to one or more other nodes in the ANN. For example, each layer is made of a plurality of nodes, where each node is connected to all nodes in the previous layer. The nodes in a given layer are not interconnected with one another, i.e., the nodes in a given layer function independently of one another. As used herein, nodes in the input layer receive data from outside of the ANN, nodes in the hidden layer(s) modify the data between the input and output layers, and nodes in the output layer provide the results. Each node is configured to receive an input, implement an activation function (e.g., binary step, linear, sigmoid, tanH, or rectified linear unit (ReLU) function), and provide an output in accordance with the activation function. Additionally, each node is associated with a respective weight. ANNs are trained with a dataset to maximize or minimize an objective function. In some implementations, the objective function is a cost function, which is a measure of the ANN’s performance (e.g., error such as LI or L2 loss) during training, and the training algorithm tunes the node weights and/or bias to minimize the cost function. This disclosure contemplates that any algorithm that finds the maximum or minimum of the objective function can be used for training the ANN. Training algorithms for ANNs include, but are not limited to, backpropagation.

[0072] It should be understood that decision trees, SVMs and neural networks are provided only as examples. According to supervised learning, the machine learning model “learns” a function that maps an input (sometimes referred to herein as the “features”) to an output (sometimes referred to herein as the “target”) based on a data set, which includes a plurality of samples (e.g., the mechanical data described herein) tagged with one or more labels (e.g., the state of the mechanical component described herein), during model training mode.

[0073] At step 106, the method can include predicting, using the trained machine learning model, a state of the instrumented mechanical component. In the implementations described herein, the prediction (i.e., output of the trained machine learning model) is a classification into one of a plurality of states of the mechanical component. As described above, non-limiting examples of states of the mechanical component can include healthy, bent, cracked, secondary failure, or multiple faults. Although the trained machine learning models described herein are classifiers, the present disclosure contemplates that the trained machine learning models can output a risk of mechanical fault (e.g., a regression). [0074] The machine learning model trained by the method 100 can be trained to predict faults in complex systems like mechanical assemblies where conventional fault prediction algorithms or heuristics can fail. Additionally, implementations of the trained machine learning models described herein can predict multiple states of the mechanical component in addition to simple failure and wear, and therefore can be used to determine when maintenance should be performed on a mechanical component. An example neural network 1700 for predicting faults in complex systems (e.g. wheelchair failure) is shown in FIG. 17. The neural network 1700 includes an input layer 1702, two hidden layers 1710 and 1712 that apply the ReLU activation function, and an output layer 1714. Input layer 1702 includes 4 nodes, hidden layer 1710 includes 10 nodes, hidden layer 1712 includes 3 nodes, and output layer 1714 includes 1 node. The nodes in each of these layers may include weights (W) and/or biases (b). As shown in FIG. 17, features (e g. four features) are input into the neural network 1700 at the input layer 1702, and the output (i.e. prediction) from the output layer 1714 is received at block 1704. It should be understood that the neural network architecture, including the number of layers and nodes, shown in FIG. 17 is provided only as an example. This disclosure contemplates using neural networks with different architecture to predict faults. Training the neural network 1700 and operating the neural network 1700 in inference mode is described herein with reference to example 3, for example. It should be understood that embodiments of the present disclosure include neural networks with more than three layers and fewer than three layers. It should also be understood that neural networks according to embodiments of the present disclosure can have any number of weights and/or biases for each layer of the neural network.

[0075] With reference to FIG. 2A, a method 200 for training a machine learning model, is shown. The method 200 can include performing fatigue testing on an instrumented mechanical component at step 202. During fatigue testing, the sensors are used to collect mechanical data as described below.

[0076] In some implementations, the machine learning model can include fuzzy logic. Alternatively or additionally, the machine learning model can be a supervised learning model. Non-limiting examples of supervised learning models include decision trees, support vector machine classifiers, and artificial neural networks.

[0077] The method 200 can also include measuring mechanical data of the instrumented mechanical component at step 204. The mechanical data can include a least one mechanical parameter associated with an instrumented mechanical component. The mechanical data, mechanical component, and mechanical parameters can include any of the mechanical data, mechanical components, or mechanical parameters described with reference to FIG. 1.

[0078] At step 206 the method can include assessing a state of the instrumented mechanical component. As described above, non-limiting examples of “states” can include the presence, absence, or degree of any or all of: healthy, bent, cracked, secondary failure, or multiple faults.

[0079] At step 208, the method can further include creating a dataset including the mechanical data and the state of the instrumented mechanical component. For each sample in the dataset, the mechanical data are the features and the state of the instrumented mechanical component are the labels.

[0080] At step 210, the method can include training the machine learning model to predict the state of the instrumented mechanical component using the dataset. This disclosure contemplates training the machine learning model using techniques known in the art. For example, the step of training the machine learning model can include minimizing or maximizing an objective function. Optionally, the objective function is a cost function such as an error between the machine learning model's prediction and ground truth (e.g. the labels). Parameters associated with the machine learning model are adjusted during training responsive to the objective function.

[0081] With reference to FIG. 2B, an example method 250 for training a machine learning model is shown, according to an example implementation of the present disclosure. The method 250 can include preprocessing at step 252, training at step 254, and validation at step 256.

[0082] At step 252, preprocessing can include data segmentation, data labeling, and feature extraction.

[0083] At step 254, training can include training a fault detection model to classify the failure (or absence of failure) or any other state of a mechanical component.

[0084] At step 256, validation can include using a data split and/or confusion matrix to evaluate the performance of the model and detect misclassifications.

[0085] FIG. 3 illustrates an example of a system 300 that the method 100 shown in FIG. 1 can be used in. The system 300 can include an assistive device 302 that can include a mechanical component 304. In some implementations, the assistive device 302 is a wheelchair. The assistive device 302 can further include a mechanical sensor 306, which can be arranged in/on/adjacent to the mechanical component 304, so that the mechanical component 304 can be considered an instrumented mechanical component, as described herein with reference to FIGS. 1A-2B. A processor 308 can be operably coupled to the mechanical sensor 306, so that the processor 308 can receive data from the mechanical sensor 306. In some implementations, the processor 308 can be physically connected to the mechanical sensor 306 (e.g., by wire(s), or by being part of the sensor module). Alternatively or additionally, the processor 308 can be in operably connected to the sensor via a communication link. This disclosure contemplates the communication link is any suitable communication link. For example, a communication link may be implemented by any medium that facilitates data exchange including, but not limited to, wired, wireless and optical links. Optionally, the processor 308 can be in operably connected to the sensor using a wireless communication link. It should be understood that the system 300 can include any number of processors 308, mechanical sensors 306, assistive devices 302, and mechanical components 304. It should also be understood that the assistive device 302 described in FIG. 3 could be any other mechanical system or device, and that the wheelchair is only a nonlimiting example.

[0086] In some implementations of the present disclosure, the system can be modeled using finite element models of one or more parts of the system (e.g., any mechanical component 304, including a caster stem bolt used to connect a caster to a wheelchair). The present disclosure can include algorithms that can predict failure of the system 300 based on the finite element models. Optionally, the mechanical sensor can be configured to measure strain on a caster stem bolt. Alternatively or additionally, the mechanical sensor can include more than one strain sensor configured to measure local mechanical strains at different locations on a mechanical component 304 (e.g., along the length and/or width of a caster and/or caster stem bolt).

[0087] The processor 308 can include any/all of the components of the computing device 400 illustrated in FIG. 4 and described herein. The processor 308 can have a memory with computer readable instructions that cause the processor to receive data from the mechanical sensor 306. The mechanical sensor 306 can be configured to use a trained machine learning model (e.g., a machine learning model trained according to the methods 200, 250 illustrated in FIGS. 2A-2B) to predict a state of the mechanical component 304. [0088] In the non-limiting example implementation shown in FIG. 3, the mechanical component 304 can be a caster that is part of the wheelchair, and the sensor can be a vibration sensor that detects vibration on the caster. The processor 308 can predict, using the trained machine learning model and vibration data received from the vibration sensor, the state of the caster. The state of the caster can then be used to determine when to replace the caster, when to perform maintenance, when the caster presents a danger to the user, or to make any other decision.

[0089] Example 1

[0090] An example implementation of the present disclosure includes systems and methods to detect wheelchair failures. Wheelchair parts break frequently, causing significant safety risks and preventing people with mobility disabilities from performing activities of daily living. [ 1 ]-[ 10]. Specialty repair of wheelchair parts is expensive and time consuming. Therefore predicting future failures can be used to improve preventative maintenance and wheelchair design.

[0091] Among wheelchair parts, the smaller front wheels, called casters take the brunt of road bumps and potholes and account for the highest i.e., one-third of all part failures[l 1], [12], An example wheelchair and caster are shown in FIG. 5. Caster stem fractures, for example as illustrated in FIG. 6, can result in a wheelchair tipping and the user falling out of the chair. A quarter of tips result in injuries or bruises [11], leading to a downward spiral of economic and health outcomes. Users miss workdays while out-of-pocket repair bills range from $6 to $4000 [5], Also, public health expenditures are doubling up [13], Wheelchair failure constrains users to their beds^{6 10} or an inadequate loaner chair [14], [15], No wonder wheelchair failures are associated with pressure injuries (bed sores), depression and rehospitalization [14], Evidencebased fault detection modeling can prevent and address these and problems with existing systems and methods. Condition monitoring technologies, including fault-detection modeling, adapted to wheelchairs and other assistive technologies can benefit users of assistive technologies.

[0092] A study described herein includes an example implementation configured to detect upcoming failures and/or provide users alerts of maintenance events.

[0093] The study can characterize the phenomenon of failure. Daily use of a wheelchair contributing to fatigue, faulty behavior and ultimately, failure of parts, needs to be measured. The study can model usage and fatigue parameters to detect the onset of a part’s faulty behavior. Once fault is detected, a user can resolve the fault by conducting maintenance.

[0094] A logger can measure road shocks to casters, for example using a strain gauge and/or accelerometer as illustrated in FIG. 7. Example data is illustrated in FIG. 8. Road shock data was transformed into a kinematic model that depicted caster movement on a wheelchair in motion. A matching lab-based model was built by subjecting casters to shocks from obstacles of varying heights and shapes on a new laboratory test bed^33,33.

[0095] Like shocks, surface abrasion and humidity are two other factors that affect fatigue. Their physical effects on casters were studied from more than n=100 failed samples (FIG. 6). Accordingly, the laboratory exposures of the two factors were tuned and integrated into the kinematic shock model for testing protocol development³⁶. The example protocol described herein for caster testing demonstrated high reliability (FIG. 8), efficacy (screen low-quality casters), and effectiveness (reproducibility of high-risk community failures) [33], [37],

[0096] Development of a caster fault detection model. A caster breaks in many ways. Stem fracture is the riskiest caster failure, according to experts and failure incident reports from wheelchair repair centers [36], It causes the caster to snap, injure a wheelchair userl 1, and damage other wheelchair parts. Secondary caster failures of parts like caster wheel or fork, supported by the stem, increase bending stress and cause rapid fatigue, leading to stem fracture. The models described herein can prioritize the detection of the stem fracture from a host of secondary caster failures.

[0097] The study herein leverages the maintenance research advances in automotive and aviation areas that use machine learning methods for fault detection. The study can include caster fatigue testing in the laboratory to generate shock data to develop the model. The study can further include deploying instrumented casters to wheelchair users and validating the model with high-risk users.

[0098] Protocol. The study can include build models for a lightweight manual wheelchair design whose casters are prone to stem fractures according to the repair registry [18] and testing studies [33], [37], Instrumented casters can undergo fatigue testing as shown in FIG. 9. Test load will be the maximum weight specified by the manufacturer. At least five samples of each caster design will witness two shock exposures - a 0.5-inch bump (standard shock) and a 0.75-inch bump (high shock) at a standard speed of Im/s. Stem fracture is the endpoint of the test. The study can document caster component conditions periodically and replace components if their function seizes.

[0099] Instrumentation. Four equidistant axial bending strain gauges and a torsion strain gauge will be installed on the stem, not on the stem housing as shown in Fig 3. They will measure stress and torsion. Universal testing machines and hydraulic actuators will perform static and dynamic calibrations. Weight on the caster will be measured and correlated with stress when the caster is at rest. A triaxial accelerometer will be installed above the gauges to monitor vibrations. The sensors will be connected to slip rings. Our custom-built data logger will amplify and record data from the sensors for offline data analysis.

[00100] Model development approach. A fault detection model built on machine learning techniques is described herein and configured to detect the state of the stem and decide whether the stem is healthy or faulty (bent or cracked). FIG. 10 illustrates an example process of developing the machine learning process using stem shock data. The instrumentation can, for example, capture stress and vibrations at 1000Hz. Bandpass filters of suitable order and cutoff will condition the signal. Standard algorithms can extract time and frequency domain features from the signals in the preprocessing stage. Time-domain features include means, standard deviation, range, and skewness. Frequency domain features include poles, zeroes, power spectral density, signal-to-noise ratio, and peaks. Selective features, their combination or residuals, will be used as condition monitors based on how their behavior changes as the stem deforms.

[00101] The study can identify fault values of the condition monitors unique to different stem states (healthy, bent, cracked). These values can depend on the material elasticity, stem dimensions and assembly. The example implementation can include feature matching between accelerometer vibration signatures and strain gauge stress events. It will assist with deploying the instrumentation in stage 2-3 field trials and for future implementation. Optionally testing can be performed over substantial time intervals to collect the data samples between the three stem states and intervene for maintenance. Crack propagation to fracture may not be instantaneous.

[00102] In some implementations of the present disclosure, machine learning techniques such as fuzzy logic, decision tree, or support vector machine classifiers can be used to train the decision model and classify stem states. Optionally, the training can include generating at least 50-100 samples per stem state. The study can partition them randomly into 80% training and 20% testing samples and conduct cross-validation of the model results. No stem will be tested against itself during validation. The present disclosure contemplates that a combination of machine learning techniques can be used in some implementations of the present disclosure.

[00103] The study can include a confusion matrix. A confusion matrix can include a multiclass contingency table comparing predicted and actual fault states will indicate misclassified states. The matrix can be inspected; if fault values of condition monitors overlap for misclassified states, the monitors can be analyzed and revised to improve accuracy. Also, misclassifications per failure evidence from studies can be punished. False-negative cases can be analyzed and minimized as much as possible. The model can be trained iteratively to reduce misclassification costs and improve performance.

[00104] Sample Size Estimation. With a sample size of 5 casters in each shock exposure group, the study included have an 81% power to detect a difference (Pi-Po) of 0.55 using an alpha level of 0.05. The study assumed that the population proportion under the null hypothesis (Po) is 0.1. The effect size of 0.55 is based on previous wheelchair testing results.

[00105] Heavy impacts versus shocks. Incident reports show casters break from heavy impacts during extreme off-road use, road accident, mishandling, or misuse by a novice user. Such heavy impact events may confound a fault detection model. The study can generate a stress-number of stress cycles (SN) curve (e.g., the graph shown in FIG. 11) using a classical approach [47], Fracture limits for each obstacle exposure should mimic the reliability of damage values shown in Fig 4. A Brinell Hardness test can output ultimate tensile strength (UTS) [48], Each stem material (n=5 samples) can have a unique SN curve. Beyond the curve, stresses in the Unsafe Region can be the heavy impacts, where the stem typically breaks or deforms depending on material elasticity. The study can identify such heavy impact events.

[00106] The study (e.g., the training of a machine learning model) can include measuring the weight on casters and/or wheelbase dimensions will be measured to estimate differences in the design of a wheelchair. Data can be collected by data loggers and/or sensors on casters. The present disclosure contemplates that the data can be collected over Wi-Fi or by any other network connection, or by using any type of display.

[00107] Implementations of the present disclosure can be configured to evaluate data when users are completing a course. Wheelchair users can complete a simulated road course which can be set up at any location. They can maneuver through road loads, including a series of filleted square wave obstacles and curb descent differing in heights. The study can rate user skills and verify the appropriateness of instrumentation installation and reliability of data collection. Optionally, the logger collects data only when the wheelchair is rolling and/or when vibrations match the O.5-5OHz seen on vehicles. The data logger will also record the time of wheelchair use, travel distance and speed.

[00108] During monthly follow-up visits, the study can document the physical symptoms of the stem conditions and other components.

[00109] Adaptation of fault detection model. The study can detect coexisting caster failures in this stage because of myriad factors in the community and shock monitoring on the wheelchair system. Optionally, the study can include the stage 1 model for additional multi class failure detection. Optionally, the study can classify five fault states - 1) healthy stem (no fault), 2) bent stem, 3) cracked stem, 4) secondary caster failures and 5) multiple faults. The multiple faults category can be a single fault flag to designate a defective caster or wheelchair.

[00110] Fault values of condition monitors curated from time and frequency domain features could shift with new community data. Therefore, the study can label such parameters for different failure states of the caster components. FIG. 10 illustrates a model training and validation process that can be used to perform iterative modeling to ensure the accuracy for predicting stem failure states is high (e.g., >80%). The study can examine and minimize misclassifications and false negatives. For root causes of failure other than fatigue, like heavy impacts, the study can stratify the model depending on how root causes and other user characteristics contribute to fatigue.

[00111] Additional Analysis. Health, social and financial costs of failures can be computed. Regression analysis will evaluate the relationship between adverse consequences as dependent variables and user demographics and usage factors as predictors. These predictors can be part of a proportional hazards model to correlate with survival time for casters. Further, the study can compare and contrast outcomes and reliability indicators (availability, survival analysis) across caster designs and user diagnoses. Caster faults correlate with battery faults[49], and so, the study can perform such correlations with 2/3rds, i.e., the rest of the wheelchair part failures.

[00112] Implementations of the present disclosure can include methods of risk determination. Risk determination can be used to prioritize maintenance for high-risk cases. Failure mode and effect analysis (FMEA) [51 ] is a systematic technique to identify risk of faults. The FMEA generates a criticality number - a product of severity and occurrence of faults. The severity of faults is the cost of consequences and will be ranked by the subjects. Occurrence is the frequency of each of the five faults.

[00113] Model efficacy testing. The study of the example implementation expects a 70% reduction in caster failures.

[00114] Implementations of the present disclosure can include using the the monitoring and failure prediction as a tool for the maintenance and prevention of wheelchair failures and adverse consequences to the user. Optionally, this tool can be integrated into digital technologies, serving as an artificial intelligence solution. Using digital tools for prevention of failures can reduce or eliminate the need for maintenance visits. Additionally, automatically detecting maintenance events can be used to verify claims of damage and preemptively schedule repairs.

[00115] Example 2

[00116] A driving cause of wheelchair caster stem bolt failures is shock loading from the caster hitting irregularities in the traveled surface. These impacts produce high peak stresses in the bolt that are below the ultimate tensile strength of the bolt. Repeat impacts on the order of hundreds or thousands, eventually lead to a mechanical failure mode called low cycle fatigue.

[00117] The life cycle of a caster stem bolt that results in fracture can be described in three core stages, (i) Full load carrying capacity, no structural defects, (ii) Full load carrying capacity, microscopic crack formation (ii) Reduced load carrying capacity, large crack structures, significant plastic deformations. Early detection and characterization of this damage would allow for preemptive replacement or initiation of the processes needed to replace the damage part either eliminating or at least reducing downtime associated with failures.

[00118] Wheelchair stem bolt failures can be the result of repeated high force impacts originating from the caster wheel hitting irregularities on the traveled surface. The failures are bolt fractures concentrated around the stem bolt and lower ball bearing contact interface and result in a permanently deformed bolt structure even before total fracture occurs. This deformation can be quantified via strain measurement which in turn can be leveraged to indicate when the bolt has permanently deformed and is at the onset of losing all load carrying capability. A measured strain response in lab can simulate wheelchair caster stem bolt failures and quantify specific signal signatures and thresholds that indicate the onset of complete fracture.

[00119] Measuring strain response in a lab can include developing a FEM (Finite Element Model) of the wheelchair stem bolt to determine critical locations for failure, stresses, evaluate bolt modification impacts on stress. The study can further include designing and fabricating a test fixture to hold a caster stem bolt in position, so that repeatable data can be collected (e.g., for training machine learning models described herein). FEM can be used to determine measured strains under a variety of gauge lengths and areas. FEM can, for example, find minimum gauge length and area that produces expected strains below the fatigue strengths of the strain gauges. This can ensure the strain gauge survives the cyclic loading until the point of bolt fracture even under high applied local strains that cause fracture of the stem bolt. Alternatively or additionally, a modified caster stem bolt can be used to allow for in-situ strain gauging monitoring the bending strains while minimizing the maximum bending stress increase to due modifications

[00120] As a non-limiting example, the study can include low cycle SN testing on 20 baseline caster stem bolts, and 20 modified caster stem bolts, allowing the effect of modifications to be measured. As another non-limiting example, the study can include performing low cycle fatigue testing on the instrumented stem bolts to monitor bending strain, bending displacement, and applied load at sufficiently high sample rates until the bolt fails. As yet another example, the study can include evaluating measured strain, displacement and applied load, for signatures indicating the onset of stem bolt failure and developing an algorithm and thresholds that provide quantitative feedback on the rate and risk level of stem bolt failure.

[00121] Wheelchair caster shimmy occurrence frequency and oscillation energy is known to be increased when caster stem bolt assemblies are damaged. Theoretical sources of the increased shimmy behavior are increased clearances in bearing supports reducing overall stiffness of the assembly and reduced effective trail length of caster due to deformation in the stem bolt assembly. However, there is a need for data and/or a model that specifically looks at the influence of damaged stem and caster assemblies on the caster shimmy behavior. The study therefore contemplates experimentally investigating the effect of damaged wheelchair stem bolts on wheelchair caster shimmy behavior including changes in the critical caster load, speed and disturbance magnitude which trigger the self-excited vibratory behavior in the castor that is known as wheelchair shimmy. This can allow correlation of the caster shimmy phenomena to stem bolt failure and provide further measurable metrics that indicate the onset of stem bolt failure (e.g., using the machine learning techniques described in the present disclosure).

[00122] The study can include creating a set of damaged caster stem bolts.

[00123] The algorithms described herein can be used for precise deformation and damage metrics for each individual test that removes some effects of statistical variation in stem bolt low cycle fatigue lifetimes. This can be done on modified and/or instrumented stem bolts. Stem bolt damage can be applied as a function of prediction intervals based on statistical regressions of the stem bolt SN lifetimes. This can be performed using statistically significant quantities of unmodified stem bolts.

[00124] The study can further including designing and fabricating a caster walker capable of creating caster shimmy with introductions of surface anomalies creating impact and swivel in the caster. Alternatively or additionally, the study can include instrumenting the caster with triaxial acceleration measurements on the caster fork. Instrument the caster walker with high-speed video. Alternatively or additionally, the study can include installing undamaged and damaged stem bolts into the caster walker and/or measuring critical shimmy speed as a function of caster load for damaged and undamaged stem bolts.

[00125] Strain measurement and plastic deformation are direct indicators of mechanical damage, while wheelchair shimmy metrics are indirect secondary effects of mechanical damage. However, strain measurement can require a modification to the standard stem bolt part to be implemented. Acceleration is a parameter that can be measured in-situ without significant modification of current stem bolt assembly designs. This parameter is an indicator of shock or impact to the wheelchair. From acceleration transducers can be cheap and easy to implement. Acceleration levels measured in an impact can be calibrated to the stresses induced on the caster stem bolt. With that calibration, standardized bolt fatigue data combined with cumulative damage metrics can be utilized to gain an estimate of cyclic load history on the stem bolt using only acceleration metrics. Knowledge of the cyclic load history can be a strong indicator of bolt structural integrity and remaining life.

[00126] In an example implementation, the system can include installing a modified stem bolt and installing small gauge length strain gauges designed to measured very local strains to get maximum bending strain with a high degree of accuracy. [00127] Alternatively or additionally, an example implementation can include a caster walker, where strain gauged stem bolts and accelerometers are installed into the caster walker. The accelerometers can optionally be deployed on the caster housing.

[00128] The modified stem bolt and/or caster walker can be impacted to measure bending strains in the stem bolt and/or acceleration of the caster housing. The bending strains and/or acceleration can be used to model the relationship between bending strains and/or acceleration, or between bending strains and/or acceleration and failures. Optionally the bending strains and/or acceleration data include peak bending strain and/or peak acceleration data.

[00129] In some implementations of the present disclosure, a regression can be performed of the peak strain to peak acceleration levels to create a proof-of-concept calibration between acceleration and strain levels.

[00130] Accelerometers can be mounted to casters as direct measurement of caster shimmy dynamics. Implementations of the present disclosure can perform a cross correlation of acceleration and caster shimmy, which can in turn increase the robustness of measured shimmy metrics to indicate the onset of stem bolt failure. An analytical dynamic model for wheelchair caster shimmy which includes the effects of deformations and clearances due to mechanical damage in a wheelchair stem bolt assembly can be used. The measured caster shimmy data incorporating damaged stem bolts will be used to validate the model. In turn the model will allow for prediction on the effects of caster design changes on critical speeds of wheelchair shimmy and enable the use of caster shimmy as a diagnostic tool on these new designs with less overall cost and time.

[00131] The present disclosure further contemplates using a lumped parameter analytical description of the mass, stiffness, and damping interactions within a wheelchair caster. Terms can be derived and included into the model to account for modifications to geometry from bolt deformations, and stiffness from clearances.

[00132] The present disclosure further contemplates using measurements of the caster stem bolt deformations and bearing clearances from damaged assemblies to quantity damage inputs into the model. Optionally, the model can simulate wheelchair caster shimmy critical speed under various levels of damage.

[00133] Existing literature (Mhatre et al., 2020; Mhatre et al., 2017) and fractography work (FIG. 25) on stem bolt failures taken from wheelchairs in the community supports the theory that the fracture mechanism is low cycle bending fatigue. On the fracture surface, a fatigue crack can be seen originating from the tensile side of the stem bolt, and the profde view shows the significant plastic deformation of the bolt, resulting in a curved profile in the failed specimen.

[00134] A significant hurdle in strain-based field monitoring of low cycle fatigue is that the strain gage sensor itself will not survive the fatigue process and produce unreliable data as damage progresses (Dowling, 1977). Further research shown in FIG. 26, shows that strain gages mounted elastically on a specially modified caster stem bolt can be used to monitor bending strain in low-cycle fatigue structures and will show a mean offset when significant damage has occurred. This finding is supported by work done on structural beams (Heise Jr, 1965). Laboratory fatigue tests show that the permanent deformation of the stem bolt (damage) correlates almost perfectly with the specially mounted strain gauge.

[00135] Example 3

[00136] A study was performed using a dataset containing 119 patients' wheelchair usage information. The dataset included 5 features: age of the patient, sex of the patient, age of the wheelchair at the beginning of the study, drives per month, and time to failure.

[00137] The study performed a regression analysis on our dataset to determine which features have a significant impact on predicting wheelchair failure. The results indicate that patient sex does not have a significant effect on failure prediction as shown in FIG. 16. The biggest predictor is drive-per-month, followed by the patient's age and, finally, the age of the wheelchair. RMSE: 4.0415 [Months], FIG. 13A illustrates example regression analysis results. FIG. 13B illustrates an example regression fit with adjusted data and 95% confidence bounds, plotting adjusted time to failure vs. an example adjusted model. A table of example statistical results including SumSq, DF, MeanSq, F and pValue are illustrated in FIG. 12.

[00138] FIGS. 14 and 15 illustrate example experimental results from the study. FIG. 14 illustrates example effects of predictor variables on time to failure for example implementations of the present disclosure. FIG. 15 illustrates an example analysis of chair age and drive-per month, according to an example implementation of the present disclosure.

[00139] The study included neural network training. The study trained an example neural network shown in FIG. 17 on the example dataset to predict wheelchair failure. The example neural network included a 2-layered MLP with ReLU activation function and Adam optimizer. The results show that an RMSE error of 3.9780 can be achieved with the example model. FIG. 18 illustrates example training, validation, and test learning curves for the neural network illustrated in FIG. 17. FIG. 19 illustrates example machine learning training over 11 epochs and FIG. 20 illustrates an error histogram of neural network training.

[00140] FIGS. 21A-22 illustrate additional experimental results from the study. FIG. 21 A illustrates an example analysis of adjusted time to failure for a young patient group, before outlier removal, according to implementations of the present disclosure. FIG. 21B illustrates an example analysis of adjusted time to failure for a young patient group, after outlier removal, according to implementations of the present disclosure. FIG. 22 illustrates an example parallel plot analysis of drive per month vs. risk group where Early repair = 1 represents high risk, and Early Repair = 0.

[00141] The study further included an analysis of younger patient group fit (Age <40). Another set of models were fitted to the younger patient group. The example described herein included two approaches: (1) Linear regression fit without outlier removal; (2) Linear regression fit with outlier removal.

[00142] The study used the scaled Median Absolute Deviations (MAD) measure to detect outliers. In the example, a value is considered an outlier if it exceeds the median by more than three times the scaled Median Absolute Deviations (MAD). The Median Absolute Deviation (MAD) is a measure that is calculated for a vector A, which has a finite length and consists of N individual scalar observations. MAD is given as:

[00143] MAD = median (I A_t — median (.4) I), for i = 1,2, ... , N.

[00144] The outlier removal improved the regression fit significantly. The RMSE (in months) dropped from 2.45 months to 1.28 months. The table illustrates details of the fit for before and after outlier removal respectively:

Estimated Coefficients: SE tStat pValue

(Intercept) 2.8985

Age 0.11615 0.096073 1.209 0.25447

Age_of_chair 0.0063911 0.052509 0.12172 0.90554

Drive_per_month -0.093337 0.033924 -2.7514 0.020431 Sex m 1.0147 1.3353 0.75996 0.46482

[00145] The regression fit to the young patient group after outlier removal gives the lowest RMSE error of 1.28 months. The other parameters of the study were: Number of observations: 15, Error degrees of freedom: 10, Root Mean Squared Error: 2.45, R-squared: 0.506, Adjusted R-Squared: 0.309, F-statistic vs. constant model: 2.56, and p-value =0.104.

[00146] The study further included an investigation of the mean and standard deviation of high-risk and low-risk groups, where the high-risk mean was: 36.49, the high-risk standard deviation was 21.23, the low-risk mean was 22.26, and the low-risk standard deviation was 17.83.

[00147] The study shows that it is possible to predict wheelchair failure of the whole group with an RMSE error of —4 months using regression analysis and an RMSE error of —3.9 using neural network training.

[00148] In the example data set the best fit was achieved in the younger patient group after outlier removal. It was possible to predict wheelchair failure in the younger patient group with an RMSE error of 1.28 months.

[00149] The study included case studies focused on the application of predictive modeling for wheelchair failure. The goal was to predict the time to the first wheelchair failure, which can be used to schedule preventative maintenance and avoid unexpected breakdowns.

[00150] The model was trained on a dataset of 12 patients aged below 40 and tested on 2 patients. The features included in the model were the age of the patient, sex of the patient, age of the wheelchair, and average drive per month.

[00151] Test Patient 1 : Age: 35; Sex: Female; Age of chair: 7 years; Drive per month: 34; Time to first wheelchair failure: 2 months; Predicted first failure lower bound: 2.8 months

[00152] The model described herein predicted the first failure with an error of less than a month. If the proposed model had been utilized, a preventative maintenance visit could have prevented the failure this patient faced.

[00153] Test Patient 2; Age: 36; Sex: Female; Age of chair: 23 years; Drive per month: 23; Time to first wheelchair failure: 4 months; Predicted first failure lower bound: 3.34 months. [00154] Similar to the first patient, the model described herein predicted the first failure with an error of less than a month. If the proposed model had been utilized, a preventative maintenance visit could have prevented the failure this patient faced.

[00155] In conclusion, predicting wheelchair failure is an important task that can help prevent accidents and ensure patient safety. The study shows that it is possible to predict wheelchair failure with regression analysis and training and deploying neural networks and/or other machine learning models.

[00156] Example 4

[00157] A wheelchair is a mobility device designed to assist people who have difficulty walking or moving around due to various reasons such as disability, injury, or illness. Wheelchairs come in various types and designs, including manual wheelchairs that require physical effort to move, and electric or power wheelchairs that are motorized and operated with a joystick or other control mechanism. Mobility is essential for human health and well-being. For people who need a wheelchair to engage in everyday activities, an appropriately designed and well-maintained wheelchair is essential to ensure physical and mental health and development, function, participation and inclusion. WHO has estimated 80 million people (or 1% of the world’s population) are likely to require a wheelchair to assist their mobility, with this number set to grow as the global population ages and other trends such as an increase in chronic health conditions continue. Wheelchairs play a crucial role in enhancing the mobility and independence of individuals with mobility challenges, allowing them to participate more fully in daily activities and navigate their environment. While wheelchairs are generally designed with safety in mind, accidents and injuries can still occur. Wheelchair parts break frequently, resulting in falling and injuries for many wheelchair users around the world.

[00158] The breaking and malfunctioning of wheelchair parts can be concerning for several reasons, and while they might not always lead to severe health consequences, they can pose risks and contribute to potential health issues. Failure of critical wheelchair parts such as back wheels, front casters and brakes may result in a tip-over, which might induce injuries such as bruises, sprains, fractures, or more serious injuries. Experiencing injury resulting from a broken wheelchair can be traumatic, leading to fear and anxiety about using the wheelchair or navigating certain environments. This emotional impact can affect a person's mental well-being and confidence. Injuries, even minor ones, can potentially lead to secondary complications. For example, if a person injures their arm during a fall, it may affect their ability to self-propel the wheelchair, leading to changes in mobility patterns. A broken wheelchair can disrupt a person's daily activities and independence, making it more challenging to navigate their environment and engage in routine tasks.

[00159] To minimize the risk of injuries and financial loss, it is crucial for wheelchair users to receive proper training on wheelchair use and maintenance. As regular check-ups with healthcare professionals can help address any emerging health issues, regular maintenance checks on the wheelchair to ensure its stability and functionality are equally important for the wellbeing of a wheelchair user. Wheelchair maintenance involves regular checks and care to ensure that the wheelchair remains in good working condition. Proper maintenance helps prevent breakdowns, ensures user safety, and extends the overall lifespan of the wheelchair. Some essential aspects of wheelchair maintenance include inspecting frame and components, checking for loose fasteners, inspecting the wheels for wear and tear, examining the spokes and rims, checking seating and positioning, examining wheelchair brakes to ensure they engage and release properly, maintaining wheelchair bearings, clean and inspect the caster wheels for debris or damage, inspecting footrests and armrests and cleaning the wheelchair regularly.

[00160] Scheduling professional inspections with a qualified wheelchair technician or healthcare professional to identify and address any issues is critically important to maintain the functionality of wheelchairs and promote wellbeing for wheelchair users. The question comes to when should such maintenance and servicing be scheduled? Most wheelchair users wait passively for wheelchair parts to break before they contact maintenance service. In this case damage and injuries might have already happened. This study’s objective is to test the feasibility of the usage data-driven wheelchair servicing concept and evaluate the relationship between wheelchair usage measures like travel distance, speed and shocks, and failure measures like failure frequency. The study’s central hypothesis is that wheelchair usage measures are correlated with failure measures. A secondary objective of the study was to assess participants’ preferences for new technology for wheelchair servicing.

[00161] Methods

[00162] Study Design and Ethics [00163] A descriptive, cross-sectional study design was implemented for 1 year to address the study’s objective. Wheelchair usage was monitored for a month using a sensor mounted on the rear axel of the wheelchair (“WheelTrak”). Data collection on user demographic characteristics, wheelchair design characteristics, and participant preferences on servicing technology was conducted using surveys. Study procedures and data analysis were conducted at The Ohio State University. Informed consent was obtained from all participants.

[00164] Study Participation

[00165] Participants who met the inclusion criteria: 18 years of age and older and used a manual wheelchair as a primary means of mobility, were screened and recruited. Participants with cognitive impairments that limit full study participation were excluded from the study. Recruitments were pursued through posting study flyers at affiliated clinical facilities, and word of mouth for users participating in ongoing studies at the institute. Wheelchair users who wish to participate in the study were contacted via phone and email to be screened against the inclusion and exclusion criteria.

[00166] Study Procedures

[00167] Study participants went through two study visits for data collection. Visits were scheduled in a private clinical room at the Ohio State Martha Morehouse Outpatient Care facility. During the first visit, weight measurements and parts inspection of the participant’s wheelchair were conducted. With the participant sitting in their wheelchair, weight measurements of the whole wheelchair and only the front casters of the chair were taken on an electrical weight scale. After recording the above measurements, the participant was asked to transfer out of the chair in a designated seating area, and the weight of the wheelchair without the participant was taken and recorded. Wheelchair inspection of frames, casters, brakes, drive wheels, and axles was conducted using instructions noted in the validated Wheelchair Maintenance Assessment Tool (W-MAT). The dimensions of front casters and drive wheels were recorded. The sensor was mounted on the wheelchair's rear axle with zip-ties. Surveys included demographic information about the participant, wheelchair manufacturer/model, part failures, maintenance, and repairs observed over the past year was emailed to the participant. The sensor would collect wheelchair usage parameters for a month. A second visit for sensor retrieval was scheduled at the end of the first visit.

[00168] Surveys [00169] Demographic variables included age, sex, racial identity, education, presence of familial or caregiver support, difficulty walking 100m, number of years spent in the current wheelchair, training status on skills and maintenance, typical travel distance/day, and transportation use. Wheelchair manufacturer, model, maintenance activity routine, and failures in the past year were recorded. When failures were reported, participants were asked whether the following adverse consequence(s) occurred: no consequences occurred, stranded, injured, missing work/school, or missing medical appointment (M. Toro et al., 2016). In the second survey, at t=8 months, maintenance preferences using technology, in addition to failure and consequences since the first visit were recorded. Participants’ willingness to buy and use the servicing technology, and other payer suggestions were noted. Both study surveys took 15-20 minutes on average to complete.

[00170] Sensor Instrumentation and Data Collection

[00171] Previous research on wheelchair caster quality testing standards demonstrated that outdoor shocks can be recorded and replicated onto laboratory-based testing equipment to reproduce caster failures seen in the community (Mhatre et al., 2020). This study extends this validated approach of data collection and employs refined instrumentation. The Arduino Nano BLE Sense replaced the off-the-shelf sensor from the previous study. The D-size alkaline batteries were replaced with a 3.7V, 6000mAh Lithium-polycarbonate battery. Wheelchair usage data from the sensor was collected on the SD card. The Arduino incorporates a 9-degree-of-freedom inertial measurement unit. Instantaneous accelerations that occur when the wheelchair wheel strikes an obstacle were recorded by the accelerometer at 100Hz. The magnetometer monitors a magnet on the rear wheel and computes travel distance and speed. The sensor’s battery capability allows seven days of data collection. The participants are advised to charge the sensor overnight using a magnetic charger that’s connected to the sensor to enable the sensor to continue recording data in a month.

[00172] Data Analysis

[00173] Descriptive statistics (mean ± standard deviation) were performed on user characteristics, wheelchair-related variables, failures and consequences. Part failures were classified by risk. High-risk failures included those that could cause the user to get injured or other wheelchair parts to be damaged. Others were grouped as low-risk. [00174] The sensor data was retrieved from the SD card and loaded into data analysis software. Data retrieved from the sensor includes milliseconds since the device started, travel distance, speed, acceleration and gyroscope data in x, y and z directions. Referring to methods from wheelchair standards development (Mhatre et al., 2020), vertical acceleration values greater than five times the RMS were computed and designated as fatigue-inducing shocks. Correlations between the continuous usage and failure data were assessed using Pearson’s correlation coefficient.

[00175] Maintenance patterns were computed. The frequency of maintenance activity reported by participants was ordered with no/not possible = 0, daily = 5, weekly = 4, monthly = 3, quarterly = 2, and yearly = 1, and a cumulative total maintenance score for each participant was computed. Correlations between the ordinal maintenance score and training status were tested using point-biserial correlation, and maintenance score and failure measures were tested using Spearman’s rho.

[00176] Descriptive, shock and correlation analyses were performed in Microsoft Excel and MATLAB R2023a and SPSS 28.0 respectively (IBM Corp., 2021; Microsoft Corporation, 2023; The MathWorks Inc., 2022). Since this was a feasibility testing study with a smaller sample size, a less conservative a-level of 0.1 was selected.

[00177] Results

[00178] Recruitment Results

[00179] Thirty-three participants were screened for the study and thirty were eligible for the study. Twenty participants completed the study with sensor intact. Two reported sensors were broken and requested repair, and three participants conducted two months of data collection. Five participants returned sensors damaged, hanging or out of battery. Three participants never returned the sensor.

[00180] Part Failure Outcomes

[00181] At least 215 wheelchair part failures were reported over the past 12 months. Participants experienced 13.4±14.8 self-reported part failures and 4.7±4.8 high-risk failures. High-risk failures observed were broken axle lock, caster wheel, caster fork, axle bearings, brakes, and push rims, and loose brakes. Low-risk failures observed were worn-out tires, seat, cushions, foot supports, leg supports, lateral supports, trunk supports, and sagging upholstery. Tires and brakes, axile bearings, seat and upholstery, foot/leg/arm/back support and fasteners saw more than 2 failures for 4 participants. Adverse consequences were experienced from high-risk failures. Six users experienced adverse consequences in the past 12 months. Although no participant was injured, 3 adverse events led to participants being stranded and 22 adverse events led to missed work, school or medical appointments.

[00182] Usage-failure Relationship

[00183] Caster fractures and other wheelchair failures are highly correlated with shocks and travel distance (R² > 0.75). Wheelchairs with less than 7 days of data and data from broken sensors were omitted. Analysis with 23 users showed that travel distance as a usage factor correlates with total part failures, r(22) = .532, p< 01 and part fractures, r(22) = .601, p< 01. Separate one-way ANOVA testing showed that user’s age F(2,22)=3.68, p< 05 and tire damage status F(2,22)=9.63, p< 05 had a significant effect on caster’s damage status. At a relaxed p< 15, wheelchair type and weight on casters significantly affected caster damage. Travel distance, use of van or truck, and maintenance cumulative scores significantly affected wheel damage status. p< 05. Shock RMS significantly affected axle damaged status F(2,22)=5.33, p<.05. Also, brake damage status varied significantly with use of van or truck, functional mobility and falls frequency, p< 05. Part fractures correlated with adverse consequences, r(22) = .577, p<.01 and total weight, r(22)=.47, p< 05. Hierarchical linear regression showed that 75% of the variation in part fracture frequency is attributed to travel distance, wheelchair+user weight, and frame and wheel damage condition measured by W-MAT, F(2,14)=8.236, p< 001, R² =.746. The consequence of missing work, school, or medical appointments was associated with part fractures, maintenance responsibility and frame damage condition, F(3,19)=9.806, p< 001, R² =.608. Being stranded was associated with the damage condition of frames, wheels and tires, F(3,19)=4.568, p<.05, R² = 419. User’s age, wheelchair age and amount of maintenance explained 99.7% of the variance in functional mobility, F(3,26)=2948, p< 001, R² =.997. Falls out of the chair negatively correlated with the load on and diameter of casters, p<.05.

[00184] Participant Feedback on Wheelchair Servicing Technology

[00185] Twelve out of 25 participants would like to receive wheelchair service reminders. Nineteen out of 25 would like to use and 17 would like to buy WheelTrak. Ten participants noted an average $105 one-time purchase, while 5 preferred a monthly subscription of $10. [00186] Example 5; Telematics study

[00187] An example study analyzed telematics data, collecting telematics data from a set of wheelchair users to evaluate relationships between telematics data and predicted failures. The telematics data was collected by a telematics device. As used herein, telematics refers to a devices that capture on-board diagnostics for a vehicle of any kind (e.g., a wheelchair).

[00188] The study analyzed de-identified data that included postcode, wheelchair issued date, duration of power wheelchair use in years, disability type (congenital/acquired/progressive), date of birth or age, gender, and wheelchair repair history. The telematics device provided wheelchair drive time (hours), activity bouts and duration, power-up time (hours), fault code, and battery indicators of charging time, charging level (%), and voltage dip duration (hours) and level (volts). Repairs in both groups over the course of the evaluation, including information on dates of repairs, repair and replacement parts used, and costs of parts, labor, technician travel, and administration, were tracked in the study. FIG. 23 illustrates example attributes of the study groups, separated into a “telematics cohort” and a “standard care cohort” (i.e., without telematics).

[00189] Data Analysis. Prior to analysis, data cleaning was performed. Descriptive statistics were computed for demographic and wheelchair-related information and telematics parameters. The effectiveness of battery servicing was assessed by the percentage difference in repair frequency across batteries and other parts and repair costs. The percentage of users whose behavior changed after the training was computed. To evaluate associations between wheelchair usage measures, participant demographic characteristics, and repair measures, the Pearson correlation was used. The chi-square test evaluated the associations between disability type and other variables. Continuous variables were compared using paired t-tests, and categorical ones using the Chisquare tests. Linear regression was performed to test the prediction of repair measures based on correlated variables.

[00190] Findings from the linear regression were extended to test whether correlated independent variables can diagnose failure risk. Research evidence on power wheelchair repairs shows a high part repair frequency within 6 months of use. Participants who received the telematics device were grouped into two categories: one with multiple repairs within 6 months (high-risk) and the other with one or no repairs (low-risk) during the study. A receiver- operating-characteristic (ROC) curve was generated to assess the accuracy of independent variables in diagnosing highrisk participants. Those who participated for less than 6 months were omitted from this analysis.

[00191] Postcode was used to generate the urban and rural classification using the 2016, 6-fold Urban Rural Classification [16], In this classification, classes 1-2 are urban and 3-6 are rural areas. ANOVA tested differences in usage and repair measures between urban and rural areas. The level of significance was set a priori at a — .05. All statistical analyses were performed using SPSS 28.0.

[00192] Characteristics of participants and repairs are listed in FIG. 23. Electronic repairs included repairs witnessed with cables, power modules, batteries and battery chargers, remote controllers, and 217 actuators.

[00193] In the telematics cohort, the participation in the evaluation averaged 9.42 ± 4.03 months, with 31 measures. Participant's age and drive time per month were associated with total repairs ( (2,118) = 26.12, p < .01, R² = .31), electronic repairs (F(2,118) = 8.35, p < .01, R² = .12), battery repairs (F(2,118) = 5.59, p < .01, R² = .08), and caster repairs (F(2,118) = 13.6, p < .01, R² = .19) and total repair costs (F(2,118) = 26.97, p < .01, R² = .32). Regression modeling showed that drive time per month predicted time-to-failure for wheelchair parts (F(l,119) = 17.1, p < .01, R² = .13). In the standard care cohort, age was weakly associated with total repair costs (F(2,117) = 5.99, p < .05, R² = .05).

[00194] Diagnosing Failure Risk. Fifty-one participants experienced between 2 to 13 repairs within 6 months and were grouped into a high-risk group. The low-risk group included 46 participants who experienced one or no repair during the evaluation. FIG. 24 shows the distributions of drive time per month of both groups. The high-risk group drive time (36.42 ±21.29 h) was significantly higher than the low-risk group drive time (22.18 ± 17.82 h). When drive time per month was used to classify participants into two groups, the area under the ROC curve was 0.71 (95% CI 0.61 to 0.82; p < 0.001) (FIG. 24). At the cutoff of 15.42 h, drive time per month predicted repairs with 92% sensitivity and 50% specificity. The study shows that telematics can be used to successfully capture data that can predict failures of wheelchairs.

[00195] It should be appreciated that the logical operations described herein with respect to the various figures may be implemented (1) as a sequence of computer implemented acts or program modules (i.e., software) running on a computing device (e.g., the computing device described in FIG. 4), (2) as interconnected machine logic circuits or circuit modules (i.e., hardware) within the computing device and/or (3) a combination of software and hardware of the computing device. Thus, the logical operations discussed herein are not limited to any specific combination of hardware and software. The implementation is a matter of choice dependent on the performance and other requirements of the computing device. Accordingly, the logical operations described herein are referred to variously as operations, structural devices, acts, or modules. These operations, structural devices, acts and modules may be implemented in software, in firmware, in special purpose digital logic, and any combination thereof. It should also be appreciated that more or fewer operations may be performed than shown in the figures and described herein. These operations may also be performed in a different order than those described herein.

[00196] Referring to FIG. 4, an example computing device 400 upon which the methods described herein may be implemented is illustrated. It should be understood that the example computing device 400 is only one example of a suitable computing environment upon which the methods described herein may be implemented. Optionally, the computing device 400 can be a well-known computing system including, but not limited to, personal computers, servers, handheld or laptop devices, multiprocessor systems, microprocessor-based systems, network personal computers (PCs), minicomputers, mainframe computers, embedded systems, and/or distributed computing environments including a plurality of any of the above systems or devices. Distributed computing environments enable remote computing devices, which are connected to a communication network or other data transmission medium, to perform various tasks. In the distributed computing environment, the program modules, applications, and other data may be stored on local and/or remote computer storage media.

[00197] In its most basic configuration, computing device 400 typically includes at least one processing unit 406 and system memory 404. Depending on the exact configuration and type of computing device, system memory 404 may be volatile (such as random access memory (RAM)), non-volatile (such as read-only memory (ROM), flash memory, etc ), or some combination of the two. This most basic configuration is illustrated in FIG. 4 by dashed line 402. The processing unit 406 may be a standard programmable processor that performs arithmetic and logic operations necessary for operation of the computing device 400. The computing device 400 may also include a bus or other communication mechanism for communicating information among various components of the computing device 400.

[00198] Computing device 400 may have additional features/functionality. For example, computing device 400 may include additional storage such as removable storage 408 and non-removable storage 410 including, but not limited to, magnetic or optical disks or tapes. Computing device 400 may also contain network connect! on(s) 416 that allow the device to communicate with other devices. Computing device 400 may also have input device(s) 414 such as a keyboard, mouse, touch screen, etc. Output device(s) 412 such as a display, speakers, printer, etc. may also be included. The additional devices may be connected to the bus in order to facilitate communication of data among the components of the computing device 400. All these devices are well known in the art and need not be discussed at length here.

[00199] The processing unit 406 may be configured to execute program code encoded in tangible, computer-readable media. Tangible, computer-readable media refers to any media that is capable of providing data that causes the computing device 400 (i.e., a machine) to operate in a particular fashion. Various computer-readable media may be utilized to provide instructions to the processing unit 406 for execution. Example tangible, computer-readable media may include, but is not limited to, volatile media, non-volatile media, removable media and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. System memory 404, removable storage 408, and non-removable storage 410 are all examples of tangible, computer storage media. Example tangible, computer-readable recording media include, but are not limited to, an integrated circuit (e.g., field-programmable gate array or application-specific IC), a hard disk, an optical disk, a magneto-optical disk, a floppy disk, a magnetic tape, a holographic storage medium, a solid-state device, RAM, ROM, electrically erasable program read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices.

[00200] In an example implementation, the processing unit 406 may execute program code stored in the system memory 404. For example, the bus may carry data to the system memory 404, from which the processing unit 406 receives and executes instructions. The data received by the system memory 404 may optionally be stored on the removable storage 408 or the non-removable storage 410 before or after execution by the processing unit 406.

[00201] It should be understood that the various techniques described herein may be implemented in connection with hardware or software or, where appropriate, with a combination thereof. Thus, the methods and apparatuses of the presently disclosed subject matter, or certain aspects or portions thereof, may take the form of program code (i.e., instructions) embodied in tangible media, such as floppy diskettes, CD-ROMs, hard drives, or any other machine-readable storage medium where, when the program code is loaded into and executed by a machine, such as a computing device, the machine becomes an apparatus for practicing the presently disclosed subject matter. In the case of program code execution on programmable computers, the computing device generally includes a processor, a storage medium readable by the processor (including volatile and non-volatile memory and/or storage elements), at least one input device, and at least one output device. One or more programs may implement or utilize the processes described in connection with the presently disclosed subject matter, e.g., through the use of an application programming interface (API), reusable controls, or the like. Such programs may be implemented in a high level procedural or object-oriented programming language to communicate with a computer system. However, the program(s) can be implemented in assembly or machine language, if desired. In any case, the language may be a compiled or interpreted language and it may be combined with hardware implementations.

[00202] Although the subject matter has been described in language specific to structural features and/or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims.

[00203] References

[00204] 1. World Health Organization. Global Report on Assistive

Technology (2022).

[00205] 2. Borg, J., Khasnabis, C. Guidelines on the provision of manual wheelchairs in less-resourced settings (2008).

[00206] 3. McClure, L. A. et al. Arch. Phys. Med. Rehabil. 90, 2034-2038

(2009). [00207] 4. Worobey, L. et al. Am. J. Phys. Med. Rehabil. 91, 463-9 (2012).

[00208] 5. Worobey, L. et al. Arch. Phys. Med. Rehabil. 103, 779-789

(2022).

[00209] 6. Toro, M. et al. Arch. Phys. Med. Rehabil. 97, 1753-1760 (2016).

[00210] 7. Henderson, G. V et al. Disabil. Rehabil. Assist. Technol. 1-7

(2020).

[00211] 8. lezzoni, L. Millions rely on wheelchairs for mobility, but repair delays are hurting users (2022).

[00212] 9. World Health Organization. Factsheet on Assistive Technology

(2018).

[00213] 10. Veteran Affairs. VHA Audit of the Timeliness of VISN 7 Power

Wheelchair and Scooter Repairs (2018).

[00214] 11. Gaal, R. P et al. J. Rehabil. Res. Dev. 34, 58-71 (1997).

[00215] 12. James, A. M. et al. Assist. Technol. 1-9 (2022).

[00216] 13. Xiang, H. et al. Inj. Prev. 12, 8-11 (2006).

[00217] 14. Hogaboom, N. S. et al. Arch. Phys. Med. Rehabil. (2018).

[00218] 15. Boccardi, A., Mhatre, A. JMIR Aging (2022) [Preprint], JMIR

Preprints. 05/05/2022:39301

[00219] 16. Hansen, R. et al. Clin. Rehabil. 18, 631-639 (2004).

[00220] 17. Mhatre, A. et al. Spinal Cord (2021).

[00221] 18. James, A. et al. Arch. Phys. Med. Rehabil. 102, 1416-1419 (2021).

[00222] 19. Prajapati, A. et al. J. Qual. Maint. Eng. 18, 384-400 (2012).

[00223] 20. Center for Medicare and Medicaid Services. Medicare Coverage of

Durable Medical Equipment & Other Devices (2020).

[00224] 21. Ruffing, J. et al. Disabil. Rehabil. Assist. Technol. (2022) [In

Review],

[00225] 22. Sheldon, S., Jacobs, N. A. Report of a consensus conference on wheelchairs for developing countries. (2006).

[00226] 23. Model Systems Knowledge Translation Center. Maintenance

Guide for Users of Manual and Power Wheelchairs. (2018). [00227] 24. Toro, M. L et al. Disabil. Rehabil. Assist. Technol. 12, 843-851

(2017).

[00228] 25. Worobey, L. et al. Arch. Phys. Med. Rehabil. 97, e55 (2016). doi:

10.1016/j.apmr.2016.08.16.

[00229] 26. Worobey, L. A. et al. Arch. Phys. Med. Rehabil. 103, 790-797

(2022).

[00230] 27. D’Innocenzo, M. E. et al. PLoS One 16, e0228428 (2021).

[00231] 28. Toro Hernandez, M. L. Development, implementation, and dissemination of a wheelchair maintenance training program. (2015).

[00232] 29. Garcia, Y. et al. Glob. Perspect. Assist. Technol. 231 (2019).

[00233] 30. Toro, M. L. et al. BMC Health Serv. Res. 16, 26 (2016).

[00234] 31. Durable Medical Equipment Repair and Training Group (2022).

[00235] 32. Moubray, J. Reliability-centered maintenance (2001).

[00236] 33. Mhatre, A. et al. PLoS One 15, e0226621 (2020).

[00237] 34. Gonzalez, B., Schoendorfer D., & Mhatre A. in International

Seating Symposium (2019).

[00238] 35. Mhatre, A. et al. African J. Disabil. 6, 358 (2017).

[00239] 36. Mhatre, A. et al. Disabil. Rehabil. Assist. Technol. 1-10 (2019). doi: 10.1080/17483107.2018.1554714.

[00240] 37. Fried, J., Pearlman, J., Schmeler M., & Mhatre A. in RESNA 2022

Annual Conference (2022).

[00241] 38. Mhatre, A. et al. Spinal Cord (2021).

[00242] 39. ISO. ISO 7176-32: Test method for wheelchair castor assembly durability (2022).

[00243] 40. Giorgio Rizzoni. IEEE (2022). ORCID: 0000-0002-8397-7241.

[00244] 41. Laukonen, E. G. et al. IEEE Trans. Control Syst. Technol. 3, 347-

355 (1995).

[00245] 42. Yang, R. et al. J. Clean. Prod. 187, 950-959 (2018).

[00246] 43. Moreira, R. de P., Nascimento, C. L. in 2012 IEEE Aerospace

Conference 1-8 (2012).

[00247] 44. Koellner, A. et al. Procedia Eng. 72, 350-355 (2014). [00248] 45. Pearlman, J. L., Cooper, R. A., Duvall, J. A., & Mhatre, A. A.

(2018). U.S. Patent Application No. 15/544,109.

[00249] 46. Cipollini, F. et al. Ocean Eng. 149, 268-278 (2018).

[00250] 47. Zuo, F. J. et al. International Journal of Damage Mechanics 24,

767-784 (2014).

[00251] 48. Marcus, M. et al. in RESNA 2019 Annual Conference (2019).

[00252] 49. Mair, C. Applied internet-of things technology in the management of wheelchair maintenance. European Seating Symposium (2018).

[00253] 50. Pazzaglia, M. et al. PLoS One 8, e58312 (2013).

[00254] 51. Stamatis, D. H. Failure mode and effect analysis: FMEA from theory to execution (2003).

[00255] 52. Schnack, H. G., Kahn, R. S. Front. Psychiatry 7, 50 (2016).

[00256] 53. Hennink, M., Kaiser, B. N. Soc. Sci. Med. 292, 114523 (2022).

[00257] 54. Mhatre, A. et al. Assist. Technol. 28, 193-201 (2016).

[00258] 55. Mhatre, A. et al. African J. Disabil. 6, 15 pages (2017).

[00259] 56. VanSickle, D. P. et al. IEEE Trans. Rehabil. Eng. 8, 371-384

(2000).

[00260] 57. Mhatre, A. Development and validation of a wheelchair caster testing protocol (2018)

[00261] Boccardi, A., Wu, F., Pearlman, J., & Mhatre, A. (2022). Older

Wheelchair Users Recommend Age-Friendly Design Improvements to a Wheelchair Maintenance App: Mixed Methods Development Study. JMIR Aging, 5(4), e39301. https://doi.org/10.2196/39301

[00262] Boccardi Alyssa, Wu Fangzheng, Pearlman Jon, & Mhatre Anand. (2022).

Elderly wheelchair users recommend age-friendly design improvements to a mHealth wheelchair maintenance application. JMIR Aging, https://preprints.jmir.org/preprint/39301

[00263] Borg, J., & Khasnabis, C. (2008). Guidelines on the provision of manual wheelchairs in less-resourced settings. In World Health Organization. World Health Organization.

[00264] Consumer Right To Repair Powered Wheelchairs, (2022). https://leg.colorado.gov/bills/hb22-1031 [00265] D’Innocenzo, M. E., Pearlman, J. L., Garcia-Mendez, Y., Vasquez- Gabela, S., Zigler, C., Rosen, P., Dewi, E. EL, Praptoraharjo, L, & Mhatre, A. (2021). Exploratory investigation of the outcomes of wheelchair provision through two service models in Indonesia. PLOS ONE, 16(6), e0228428. https://doi.org/10.1371/joumal.pone.0228428

[00266] Fitzgerald, S. G., Collins, D. M., Cooper, R. A., Tolerico, M., Kelleher, A., Hunt, P., Martin, S., Impink, B., & Cooper, R. (2005). Issues in maintenance and repairs of wheelchairs: A pilot study. Journal of Rehabilitation Research and Development, 42(6), 853- 862.

[00267] Fried, J. J. (2022). Utilizing Community Data and Laboratory Testing to Raise the Quality of Wheelchair Casters.

[00268] Gaal, R. P., Rebholtz, N„ Hotchkiss, R. D., & Pfaelzer, P. F. (1997). Wheelchair rider injuries: causes and consequences for wheelchair design and selection. Journal of Rehabilitation Research and Development, 34(1), 58-71.

[00269] Hansen, R., Tresse, S., & Gunnarsson, R. K. (2004a). Fewer accidents and better maintenance with active wheelchair check-ups: a randomized controlled clinical trial. Clinical Rehabilitation, 18(6), 631-639. https://doi.org/10.1191/0269215504cr777oa

[00270] Hansen, R., Tresse, S., & Gunnarsson, R. K. (2004b). Fewer accidents and better maintenance with active wheelchair check-ups: a randomized controlled clinical trial. Clinical Rehabilitation, 18(6), 631-639. https://doi.org/10.1191/0269215504cr777oa

[00271] Henderson, G. V, Boninger, M. L., Dicianno, B. E., & Worobey, L. A. (2020). Type and frequency of wheelchair repairs and resulting adverse consequences among veteran wheelchair users. Disability and Rehabilitation: Assistive Technology, 1-7.

[00272] Hogaboom, N. S., Worobey, L. A., Houlihan, B. V, Heinemann, A. W., & Boninger, M. L. (2018). Wheelchair breakdowns are associated with pain, pressure injuries, rehospitalization, and self-perceived health in full-time wheelchair users with spinal cord injury. Archives of Physical Medicine and Rehabilitation. https://doi.Org/10.1016/j.apmr.2018.04.002

[00273] IBM Corp. (2021). IBM SPSS Statistics for Windows.

[00274] International Organization for Standardization. (2014). ISO - ISO

Standards - ISO/TC 173/SC 1 - Wheelchairs. [00275] James, A., Igede, P., Mhatre, A., Brienza, D , Pearlman, J., Karg, P., & Schmeler, M. (2021). Development of a Wheelchair Repair Registry. Archives of Physical Medicine and Rehabilitation, 102(7), 1416-1419. https://doi.Org/10.1016/j.apmr.2021.02.023

[00276] James, A. M., Pramana, G., Schein, R. M., Mhatre, A., Pearlman, J., Macpherson, M., & Schmeler, M. R. (2022). A Descriptive Analysis of Wheelchair Repair Registry Data. Assistive Technology, 1-9. https://doi.org/10.1080/10400435.2022.2044407

[00277] McClure, L. A., Boninger, M. L., Oyster, M. L., Williams, S., Houlihan, B., Lieberman, J. A., & Cooper, R. A. (2009). Wheelchair Repairs, Breakdown, and Adverse Consequences for People With Traumatic Spinal Cord Injury. Archives of Physical Medicine and Rehabilitation, 90(12), 2034-2038. https://doi.Org/10.1016/j.apmr.2009.07.020

[00278] Mhatre, A., Pearlman, J., Schmeler, M., Krider, B., & Fried, J. (2021a). Community -based wheelchair caster failures call for improvements in quality and increased frequency of preventative maintenance. Spinal Cord, https://doi.org/10.1038/s41393-021-00689- 3

[00279] Mhatre, A., Pearlman, J., Schmeler, M., Krider, B., & Fried, J. (2021b). Community-based wheelchair caster failures call for improvements in quality and increased frequency of preventative maintenance. Spinal Cord, https://doi.org/10.1038/s41393-021-00689- 3

[00280] Mhatre, A., Reese, N., & Pearlman, J. (2020). Design and evaluation of a laboratory-based wheelchair castor testing protocol using community data. PLOS ONE, 15(1), e0226621. https://doi.org/10.1371/journal.pone.0226621

[00281] Microsoft Corporation. (2023). Microsoft Excel.

[00282] Reese, N., & Rispin, K. (2015). Assessing Wheelchair Breakdowns In Kenya To Inform Wheelchair Test Standards For Low-Resource Settings. RESNA Annual Conference - 2015.

[00283] Rispin, K., Riseling, K., & Wee, J. (2018). A longitudinal study assessing the maintenance condition of cadres of four types of wheelchairs provided in low-resource areas. Disability and Rehabilitation. Assistive Technology, 13(2), 146-156. https://doi.org/10.1080/17483107.2017.1299805

[00284] Ruffing, J., Schein, R. M., Schmeler, M. R., & Mhatre, A. A. (2022). A Cross Sectional Descriptive Analysis of Complex Rehabilitation Technology (CRT) Provider Opinions on the Current State of Wheelchair Repair Services. Disability and Rehabilitation: Assistive Technology.

[00285] The MathWorks Inc. (2022). MATLAB (R2022b).

[00286] Toro, M. L., Bird, E., Oyster, M., Worobey, L., Lain, M., Bucior, S., Cooper, R. A., & Pearlman, J. (2017). Development of a wheelchair maintenance training programme and questionnaire for clinicians and wheelchair users. Disability and Rehabilitation. Assistive Technology, 12(8), 843-851. https://doi.org/10.1080/17483107.2016.1277792

[00287] Toro, M., Worobey, L., Boninger, M. L., Cooper, R. A., & Pearlman, J. (2016). Type and Frequency of Reported Wheelchair Repairs and Related Adverse Consequences Among People With Spinal Cord Injury. Archives of Physical Medicine and Rehabilitation, 97(10), 1753-1760. https://doi.org/10.1016/J APMR.2016.03.032

[00288] U.S. Centers for Medicare & Medicaid Services. (2022). Standard Documentation Requirements for All Claims Submitted to DME MACs (A55426). https://www.cms. gov/medicare-coverage-database/view/article.aspx?articleid=55426

[00289] Veterans Health Administration. (2018). Veterans Health Administration Audit of the Timeliness of VISN 7 Power Wheelchair and Scooter Repairs. https://www.va.gov/oig/pubs/VAOIG-16-04655-70.pdf

[00290] World Health Organization. (2022). Global Report on Assistive Technology, https://www.who. int/publications/i/item/9789240049451

[00291] World Health Organization. (2023). Wheelchair provision guidelines.

World Health Organization. https://www.who.int/publications/i/item/9789240074521

[00292] Worobey, L. A., Heinemann, A. W., Anderson, K. D., Fyffe, D., Dyson- Hudson, T. A., Berner, T., & Boninger, M. L. (2021). Factors Influencing Incidence of Wheelchair Repairs and Consequences Among Individuals with Spinal Cord Injury. Archives of Physical Medicine and Rehabilitation. https://doi.Org/https://doi.org/10.1016/j.apmr.2021.01.094

[00293] Worobey, L. A., Heinemann, A. W., Anderson, K. D., Fyffe, D., Dyson- Hudson, T. A., Berner, T., & Boninger, M. L. (2022). Factors Influencing Incidence of Wheelchair Repairs and Consequences Among Individuals with Spinal Cord Injury. Archives of Physical Medicine and Rehabilitation, 103(4), 779-789. https://doi.Org/https://doi.org/10.1016/j.apmr.2021.01.094 [00294] Worobey, L., Oyster, M., Nemunaitis, G , Cooper, R., & Boninger, M. L. (2012). Increases in wheelchair breakdowns, repairs, and adverse consequences for people with traumatic spinal cord injury. American Journal of Physical Medicine & Rehabilitation / Association of Academic Physiatrists, 91(6), 463-469. https://doi.org/10.1097/PHM.0b013e31825ab5ec

[00295] Worobey, L., Oyster, M., Pearlman, J., Gebrosky, B., & Boninger, M. L. (2014). Differences between manufacturers in reported power wheelchair repairs and adverse consequences among people with spinal cord injury. Archives of Physical Medicine and Rehabilitation, 95(4), 597-603. https://doi.Org/10.1016/j.apmr.2013. l l.022

[00296] Xiang, H., Chany, A.-M., & Smith, G. A. (2006a). Wheelchair related injuries treated in US emergency departments. Injury Prevention : Journal of the International Society for Child and Adolescent Injury Prevention, 12(1), 8-11. https://doi.org/10.! 136/ip.2005.010033

[00297] Xiang, H., Chany, A.-M., & Smith, G. A. (2006b). Wheelchair related injuries treated in US emergency departments. Injury Prevention : Journal of the International Society for Child and Adolescent Injury Prevention, 12(1), 8-11. https://doi.org/10.1136/ip.2005.010033

[00298] A. Mhatre, J. Ott, and J. Pearlman, “Development of wheelchair caster testing equipment and preliminary testing of caster models,” Afr J Disabil, vol. 6, p. 358, Sep. 2017, doi: 10.4102/ajod.v6i0.358

[00299] A. Mhatre, N. Reese, and J. Pearlman, “Design and evaluation of a laboratory -based wheelchair castor testing protocol using community data,” PLoS One, vol. 15, no. 1, p. e0226621, Jan. 2020, doi: 10.1371/joumal. pone.0226621

[00300] N. E. Dowling, “Performance of metal-foil strain gages during large cyclic strains: Accuracy and life are monitored for gages attached to axial specimens tested under large cyclic strains as in low-cycle fatigue,” Exp Meeh, vol. 17, pp. 193-197, 1977

[00301] R. E. Heise Jr, “Low-cycle fatigue-crack indications by strain gages operating in elastic strain fields: Response of resistance-type strain gages installed on the exterior surfaces of boxlike structures undergoing low-cycle fatigue failure is discussed in this paper,” Exp Meeh, vol. 5, no. 4, pp. 19A-24A, 1965

Claims

WHAT IS CLAIMED:

1. A method comprising: receiving mechanical data, wherein the mechanical data comprises at least one mechanical parameter associated with an instrumented mechanical component; inputting the mechanical data into a trained machine learning model; and predicting, using the trained machine learning model, a state of the instrumented mechanical component.

2. The method of claim 1, wherein the state of the instrumented mechanical component comprises at least one of healthy, bent, cracked, secondary failure, or multiple faults.

3. The method of claim 1 or claim 2, wherein the at least one mechanical parameter comprises stress data and vibration data.

4. The method of any one of claims 1-3, wherein the trained machine learning model comprises fuzzy logic.

5. The method of any one of claims 1-4, wherein the trained machine learning model comprises a supervised machine learning model.

6. The method of claim 5, wherein the supervised machine learning model is a decision tree or a support vector machine classifier.

7. The method of claim 5, wherein the supervised machine learning model is a multilayer perceptron.

8. The method of any one of claims 1-7, wherein the instrumented mechanical component comprises a caster.

9. The method of any one of claims 1-8, wherein the instrumented mechanical component comprises a wheel.

10. A method for training a machine learning model, the method comprising: performing fatigue testing on an instrumented mechanical component; measuring mechanical data of the instrumented mechanical component, wherein the mechanical data comprises at least one mechanical parameter associated with the instrumented mechanical component; assessing a state of the instrumented mechanical component; creating a dataset comprising the mechanical data and the state of the instrumented mechanical component; and training the machine learning model to predict the state of the instrumented mechanical component using the dataset.

11. The method of claim 10, wherein the state of the instrumented mechanical component comprises at least one of: healthy, bent, cracked, secondary failure, or multiple faults.

12. The method of claim 9 or claim 10, wherein the at least one mechanical parameter comprises stress data and vibration data.

13. The method of any one of claims 9-11, wherein the instrumented mechanical component comprises a caster.

14. The method of any one of claims 9-11, wherein the instrumented mechanical component comprises a wheel.

15. The method of any one of claims 9-13, wherein the machine learning model comprises fuzzy logic.

16. The method of any one of claims 9-14, wherein the machine learning model comprises a supervised machine learning model.

17. The method of claim 16, wherein the supervised machine learning model is a decision tree or a support vector machine classifier.

18. The method of claim 16, wherein the supervised machine learning model is a multilayer perceptron.

19. A system comprising: an assistive device comprising a mechanical component; a sensor configured to measure at least one mechanical parameter associated with the mechanical component; a processor and a memory operably coupled to the processor, the memory having computer-executable instructions stored thereon that, when executed by the processor, cause the processor to: receive data from the sensor; input the data into a trained machine learning model; and predict, using the trained machine learning model, a state of the mechanical component.

20. The system of claim 19, wherein the processor and the sensor are operably coupled by a wireless communication link.

21. The system of claim 19 or claim 20, wherein the mechanical component is a caster or a wheel.

22. The system of any one of claims 19-21, wherein the sensor is a strain sensor or a vibration sensor.