Task-driven FMEA risk assessment method and device for robot-assisted rehabilitation of children with cerebral palsy

Abstract

The invention relates to a task-driven FMEA risk assessment method and device for robot-assisted rehabilitation of children with cerebral palsy. The method comprises the following steps: acquiring a task sequence and a failure mode sequence according to a plurality of stages of a robot-assisted rehabilitation process of children with cerebral palsy, then respectively constructing a task relation network and a failure mode relation network, and generating a task node vector and a failure mode node vector by using a graph embedding algorithm; calculating the importance of the failure mode by using a TKL-FML double-layer network based on RWR in combination with a clustering algorithm and a DEMATEL method, determining an S score, and forming an original decision matrix together with an O score and a D score based on historical maintenance data; calculating a subjective weight by using a q-order orthogonal fuzzy set, calculating an objective weight by using a CRITIC method, and calculating an optimal combination weight by using a game theory; and sorting the failure modes by using a TODIM method, and determining the risk priority of each failure mode. Compared with the prior art, the method has the advantages of improving the accuracy and reliability of specific task risk assessment and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of robot-assisted rehabilitation, and in particular to a task-driven FMEA risk assessment method and device for robot-assisted rehabilitation of children with cerebral palsy. Background Art

[0002] Children with cerebral palsy often require specialized rehabilitation services that must fully consider their unique physical conditions and developmental needs. Robot-assisted rehabilitation provides a controlled, repeatable, and adjustable intervention method that can be fine-tuned to the specific requirements of the patient. The intricate interaction between the unique conditions of children with cerebral palsy and treatment procedures requires a thorough risk assessment to ensure safety and effectiveness. However, existing technologies focus more on the physical performance and basic operational risks of robots, while ignoring the impact of specific rehabilitation tasks on children's abilities and training processes, resulting in inaccurate risk assessments.

[0003] Failure mode and effects analysis (FMEA) is an effective risk assessment tool that was first used in the aerospace industry in the 1960s to improve safety during the design phase. The main goal of FMEA is to minimize the occurrence of failure modes (FMs) by evaluating, quantifying and ranking the risk level of potential failure modes. This process is usually completed by a team of experts who will list potential failure modes and evaluate them based on three risk factors (RFs): severity (S), occurrence (O) and detectability (D).

[0004] Nowadays, FMEA is widely used in many fields such as automobile manufacturing and medical equipment to evaluate and reduce potential risks, thereby improving the reliability of products and services. However, the application of FMEA for risk factor assessment in the field of robot-assisted rehabilitation still has the following problems:

[0005] (1) On the one hand, many methods rely too much on subjective judgment, resulting in analysis results being affected by the personal experience and opinions of experts, increasing uncertainty; on the other hand, driven analysis that relies solely on objective data cannot fully reflect the complexity and interrelationships between FMs, thus limiting the comprehensive understanding of potential risks. This contradiction between subjectivity and objectivity challenges the accuracy and effectiveness of risk assessment.

[0006] (2) Traditional FMEA methods usually treat each FM as an independent event, ignoring the interdependence and relationship between FMs. This approach may cause the occurrence of one FM to affect other FMs in a complex system, resulting in a chain reaction, and thus failing to accurately reflect the overall risk level of the system.

[0007] (3) Traditional FMEA weighting methods usually only consider a single perspective, either subjective or objective. Subjective weighting methods rely on the experience and judgment of experts. Although they can cope with complex situations, they are easily affected by bias or individual differences, resulting in inconsistent results. Objective weighting methods, on the other hand, rely entirely on data-driven methods, ignoring the uncertainty in actual operations and the field experience of experts. This single perspective often fails to fully and accurately reflect the true importance of each RFs, leading to deviations and misjudgments in risk assessment.

[0008] (4) Fuzzy sets transform the qualitative expression of the importance of RF by experts into quantitative expression, but they are easily affected by bias in the expert judgment process. This bias may lead to information distortion in the decision-making process, thereby affecting the accuracy of risk assessment. For example, intuitionistic fuzzy sets (IFS) and Pythagorean fuzzy sets (PFS) cannot accurately model risk information due to their inherent limitations in dealing with higher degrees of uncertainty and ambiguity.

[0009] (5) The traditional risk priority number (RPN) method cannot effectively distinguish the risk levels of different failure modes (FMs). Even if the RPN values ​​of different FMs are the same, they cannot accurately reflect their actual risk levels. In addition, the impact of experts' psychological fluctuations on priority results is often ignored, which may lead to biased and inaccurate decision-making. Summary of the invention

[0010] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a task-driven FMEA risk assessment method and device for robot-assisted rehabilitation of children with cerebral palsy, which can fully consider the impact of specific rehabilitation tasks on children's abilities and training process, and effectively improve the accuracy of risk assessment in the robot-assisted rehabilitation process of children with cerebral palsy.

[0011] The purpose of the present invention can be achieved by the following technical solutions:

[0012] According to a first aspect of the present invention, there is provided a task-driven FMEA risk assessment method for robot-assisted rehabilitation of children with cerebral palsy, comprising the following steps: a failure mode identification step, according to multiple stages of the robot-assisted rehabilitation process for children with cerebral palsy, after acquiring the task sequence and the failure mode sequence, respectively constructing a task relationship network and a failure mode relationship network, and then using a graph embedding algorithm to generate task node vectors and failure mode node vectors and perform normalization, wherein the multiple stages are pre-acquired, and the correlation between the task node vectors and the failure mode node vectors is determined by an expert to obtain an initial TKL-FML mapping matrix; a risk factor assessment step, based on the standardized task node vectors and failure mode Node vector, using the pre-built RWR-based TKL-FML two-layer network combined with clustering algorithm and DEMATEL method, calculate the importance of the failure mode, determine the S score, and together with the O score and D score based on historical maintenance data, form the original decision matrix of the failure mode, wherein the two-layer network is constructed based on the TKL-FML mapping matrix; risk factor weighting step, based on the original decision matrix, using q-order orthogonal fuzzy sets to calculate subjective weights, using the CRITIC method to calculate objective weights, and then using game theory to calculate the optimal combination weights; failure mode sorting step, based on the optimal combination weights, using the TODIM method to sort the failure modes and determine the risk priority of each failure mode.

[0013] As a preferred technical solution, in the failure mode identification step, the failure mode sequence is obtained using a pre-built SHELL model.

[0014] As a preferred technical solution, in the failure mode identification step, the task node vector and the failure mode node vector are obtained using the DeepWalk method, and the specific process includes: using a random walk mechanism to sample the nodes in the task relationship network and the failure mode relationship network respectively; using a pre-built skip-word model to process the sampled task nodes and failure mode nodes; and encapsulating the processed task nodes and failure mode nodes into corresponding low-dimensional embedding vectors respectively.

[0015] As a preferred technical solution, in the risk factor assessment step, a pre-constructed RWR-based TKL-FML two-layer network is used in combination with a clustering algorithm and a DEMATEL method to calculate the importance of the failure mode and determine the S score. The specific process includes: constructing a random walk matrix in the TKL-FML two-layer network; based on the random walk matrix, calculating the cosine similarity between each failure mode, and then obtaining the cosine similarity matrix of the failure mode; clustering the cosine similarity matrix using a K-means algorithm to obtain a direct relationship matrix of the failure mode; after standardizing the direct relationship matrix of the failure mode, a total relationship matrix is ​​generated, and the importance of the failure mode is calculated in combination with the pre-acquired influence degree data and the influence degree data, and the S score is determined.

[0016] As a preferred technical solution, in the risk factor evaluation step, the O score is determined based on the failure rate calculated based on the total number of failure modes each year, and the D score is determined based on the standard deviation of the failure interval time, and the total number of failure modes and the failure interval time are both derived from pre-acquired historical maintenance data.

[0017] As a preferred technical solution, in the risk factor weighting step, the subjective weight is calculated using a q-order orthogonal fuzzy set, and the specific process includes: defining a q-order orthogonal pair fuzzy weighted aggregation operator to obtain a linguistic variable table of a q-order orthogonal fuzzy set; multiple decision makers use the linguistic variable table to individually evaluate the weight of the risk factor, and different decision makers correspond to different weights; using the q-order orthogonal pair fuzzy weighted aggregation operator to integrate the individual evaluation results of each decision maker to obtain a comprehensive weight evaluation; after converting the comprehensive weight evaluation into an exact value, the standardized subjective weight of the risk factor is obtained.

[0018] As a preferred technical solution, in the risk factor weighting step, the CRITIC method is used to calculate the objective weight, and the specific process includes: standardizing the original decision matrix to obtain a standardized decision matrix; determining the reference risk factor and the target risk factor, and calculating the correlation coefficient between the two; based on the correlation coefficient, calculating the contrast and conflict score of the target risk factor; based on the contrast and the conflict score, calculating the objective weight of the target risk factor.

[0019] As a preferred technical solution, in the risk factor weighting step, game theory is used to calculate the optimal combination weight. The specific process includes: defining the combination weight, constructing an objective function, and the objective function is used to minimize the deviation between the combination weight and the subjective weight and the objective weight; based on the objective function, solving the pre-constructed optimal condition linear equation to obtain the combination coefficient, and then calculating the optimal combination weight.

[0020] As a preferred technical solution, the failure mode ranking step specifically includes: a. determining the target failure mode and the reference failure mode; b. determining the maximum combined weight among all risk factors based on the optimal combined weight; c. calculating the relative weight of the risk factor based on the optimal combined weight and the maximum combined weight; d. calculating the relative advantage between the target failure mode and the reference failure mode under the risk factor based on the relative weight, and then obtaining the comprehensive advantage between the two; e. calculating the global advantage of the target failure mode based on the comprehensive advantage; repeating a to e to obtain the global advantage of all failure modes, sort them, and determine the risk priority of each failure mode.

[0021] According to a second aspect of the present invention, there is provided a task-driven FMEA risk assessment device for robot-assisted rehabilitation of children with cerebral palsy, comprising a memory, a processor, and a program stored in the memory, wherein the method described is implemented when the processor executes the program.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. The method provided by the present invention is a task-driven FMEA method that combines expert subjective judgment with objective process description. The correlation between the task node vector and the failure mode node vector is determined by experts, which can reflect the characteristics of the expert-driven method and maintain the professionalism of the system. At the same time, the bidirectional mapping of the task node vector and the failure mode node vector is also oriented to objective tasks that can be clearly defined and quantified, which meets the main goal of system design - ensuring the successful execution of tasks. It can not only maintain the objectivity and systematicness of risk assessment, but also integrate the impact of task execution on failure mode assessment. The combination of expert subjective judgment and objective process description is simpler and more efficient than other analysis methods, and can be flexibly applied to the environment of standardized task procedures, thereby improving the overall risk management capabilities;

[0024] 2. The present invention uses the TKL-FML two-layer network analysis model based on RWR to construct the association network between failure modes and tasks, and combines the graph embedding algorithm DeepWalk with the random walk algorithm RWR to quantify the interaction between different failure modes. The model uses graph theory and data mining technology to analyze the impact path of the failure mode and identify potential chain reactions, ensuring that all relevant failure modes and their interactions are considered in the evaluation. It can more comprehensively understand and quantify the association between different failure modes and avoid the risk underestimation caused by the traditional FMEA ignoring the chain effect;

[0025] 3. The present invention adopts a fusion method combining subjective and objective weighting, combining the subjective judgment of experts with data-driven objective analysis such as CRITIC, so as to measure the relative importance of different RFs more comprehensively and accurately, improve the accuracy of risk assessment, and avoid the problem of insufficient or excessive weighting in traditional RPN;

[0026] 4. The present invention uses q-order orthogonal fuzzy sets with adjustable parameters, namely q-ROFS, to calculate the subjective weight of RF, which helps to capture and express fuzzy and uncertain risk information more robustly. It will also study the uncertainty of expert opinions by changing parameter values, thereby further improving the accuracy and reliability of risk assessment and reducing the information distortion problem caused by subjective bias in traditional methods;

[0027] 5. The present invention combines FMEA with the TODIM method to more comprehensively reflect the risk preferences of decision makers, and considers the impact of psychological factors on FM ranking results by adjusting parameter values, so that the risk levels of different FMs can be more reasonably distinguished, reducing the deviation caused by psychological factors and subjective judgment in traditional methods, and further effectively improving the accuracy and reliability of risk assessment;

[0028] 6. The present invention constructs a task-driven FMEA model by combining specific rehabilitation tasks. It can conduct risk assessment according to the characteristics of different types of tasks, refine the definition of rehabilitation tasks, and identify potential risks in each task based on the operating steps in the specific tasks. It can decompose each rehabilitation task into different subtasks and conduct independent risk analysis on each subtask, which significantly improves the accuracy of risk assessment, and can identify risks for individualized tasks to avoid excessive or insufficient risk prevention measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A schematic diagram of the framework flow of the method provided by the present invention;

[0030] Figure 2 A schematic diagram of the process of DeepWalk converting an undirected network graph and node information into a vector in an embodiment of the present invention;

[0031] Figure 3 is a sensitivity analysis result of the sorting results of different θ in the embodiment of the present invention;

[0032] Figure 4 is the sorting result of FM under different q in the embodiment of the present invention;

[0033] Figure 5 This is a ranking comparison based on four methods in the embodiments of the present invention. DETAILED DESCRIPTION

[0034] Aiming at the specific population of children with cerebral palsy, the present invention constructs a task-based FMEA model by combining specific rehabilitation tasks to perform targeted risk assessment according to the characteristics of different task types; at the same time, the method proposed in the present invention can effectively solve the problems existing in the application of the existing FMEA method in the field of robot-assisted rehabilitation for risk factor assessment as pointed out in the background technology. The specific principles are as follows:

[0035] (1) Traditionally, FMEA prioritizes failure modes by calculating the risk priority number (RPN), which is the product of the S, O, and D scores. However, this method is difficult to fully reflect the uncertainty inherent in the subjective judgment of experts in the actual evaluation process. To address this problem, researchers in the field of fuzzy theory have proposed a variety of semantic variables to represent the uncertainty in expert evaluation, such as interval numbers, belief structures, and fuzzy membership. Although these methods have improved the flexibility of evaluation to a certain extent, the impact of expert subjectivity has not been fully resolved because the scoring of RFs still mainly relies on the judgment of experts. To address this limitation, more and more researchers have begun to explore data-driven models and use RFs derived from objective data for analysis to improve the reliability of risk assessment. However, there are also problems with relying solely on data-driven methods. The data itself may be defective or biased and cannot fully capture all potential failure connections in a complex system. Therefore, relying solely on data-driven methods may miss key risks or lead to misleading evaluation results.

[0036] To overcome the above limitations, the present invention aims to combine expert-driven and data-driven methods, and propose a task-driven framework that combines expert subjective judgment with objective process description to improve the comprehensiveness and objectivity of risk assessment in FMEA, thereby improving the accuracy and effectiveness of risk assessment of robot-assisted rehabilitation tasks for children with cerebral palsy. Specifically, the task process description emphasizes the systematic division of operation procedures to ensure accurate identification of key links in robot-assisted rehabilitation, and the interaction between people, equipment and environment in the rehabilitation system can be analyzed through the SHELL model to help identify FMs and their structures, and the relationship between tasks and failures is defined by the expert-driven method in the rehabilitation field.

[0037] At the same time, the two-layer network is used to analyze the interconnections between tasks-tasks, FM-FMs, and tasks-FMs, so as to obtain the severity assessment of risk factors, that is, the S score, and provide a basis for the O and D scores through historical maintenance data. The S score is obtained by analyzing the network analysis of the complex relationship between tasks and FMs, which is a data mining technology. The technology consists of a task layer (TKL) and a failure mode layer (FML). In this framework, the relationship between the task nodes in the TKL and the failure mode nodes in the FML is defined by experts, which reflects the characteristics of the expert-driven method and maintains the professionalism of the system. In addition, the framework is oriented to objective tasks that can be clearly defined and quantified, which is in line with the main goal of system design - ensuring the successful execution of tasks. It not only maintains objectivity and systematicity, but also integrates the impact of task execution on FM evaluation. In addition, network analysis helps to consider the interactions between FMs, thereby improving the accuracy of severity assessment, especially when FMs are interrelated.

[0038] (2) Traditional FMEA often treats each FM as an independent event, ignoring the interdependence and correlation between different FMs. In a complex system, the occurrence of one FM may affect and trigger other FMs, resulting in a chain reaction. For example, the occurrence of a certain failure may trigger a series of reactions, leading to the occurrence of other FMs. If these interactions are not considered, risk assessment may not accurately reflect the overall risk level of the system.

[0039] In order to more comprehensively assess system risk, it is crucial to consider the interrelationships between FMs. It is necessary to introduce methods such as network analysis models to better understand and quantify the associations between FMs, so as to avoid underestimating the cumulative effect of risks. In order to further optimize the method, the present invention adopts link prediction techniques, such as the random walk with restart (RWR) algorithm, to evaluate the probability of node connections in dynamic network structures. Although RWR is widely used in fields such as sorting, recommendation systems, and community segmentation, it mainly focuses on the random walk of nodes and ignores the potential structural similarities between different nodes during the walk. In order to solve this problem, the present invention combines the graph embedding algorithm DeepWalk with RWR. DeepWalk can convert network data into low-dimensional vectors and maintain the similarity of network structures. Through this enhanced two-layer model, the similarity of FMs can be quantified based on the relationship between tasks, between FMs, and between tasks and FMs, so as to accurately calculate the S score.

[0040] (3) Traditional RPN cannot accurately reflect the differences in importance of different RFs, leading to miscalculations in risk prioritization. This limitation stems from the fact that RPN usually uses a simple multiplication formula and fails to consider the subtle differences between different RFs and their impact in specific situations. As a result, important risks may be underestimated, while less important risks may be overemphasized, distorting the decision-making process and affecting resource allocation.

[0041] In order to optimize the RF weights, it is crucial to integrate the subjective and objective weight fusion methods. By combining expert judgment and data-driven methods, the present invention can understand the importance of each RF in a more detailed way and improve the accuracy of risk assessment. Specifically, a fusion method combining subjective and objective weighting is adopted to combine the subjective judgment of experts with data-driven objective analysis such as CRITIC (an objective weighting method), so as to measure the relative importance of different RFs more comprehensively and accurately. Through expert judgment, the professional knowledge of situations and complex problems is considered to make up for the deficiencies in data analysis. At the same time, the CRITIC method is used to analyze the volatility and correlation between RFs to ensure that the weights of each factor are more objective and reasonable. In this way, the limitations brought about by relying solely on subjective or objective methods can be avoided, and an organic combination of the two can be achieved.

[0042] (4) Fuzzy sets convert the qualitative expression of the importance of RF by experts into quantitative expression, but are easily affected by bias in the expert judgment process. In this regard, the present invention uses q-ROFS with adjustable parameters to calculate the subjective weight of RF, which helps to capture and express fuzzy and uncertain risk information more robustly, and will also study the uncertainty of expert opinions by changing parameter values. The parameter q in q-ROFS is used to adjust the division and representation of fuzzy sets, affecting the expression and processing of information. By adjusting the parameter q, a more comprehensive information expression space is allowed, thereby improving the accuracy and reliability of risk assessment and reducing the information distortion problem caused by subjective bias in traditional methods.

[0043] (5) The traditional RPN method cannot distinguish risk levels, and even if their RPN values ​​are the same, they cannot reflect their different risk levels. To overcome these limitations, the multi-criteria decision making (MCDM) method was developed. In previous studies, various MCDM methods have been combined with FMEA to improve its performance, such as DEMATEL (decision experiment and evaluation test method), PROMETHEEII, MULTIMOORA, VIKOR, TODIM, etc. Among them, TODIM (Portuguese abbreviation for "interactive and multi-criteria decision making") uses prospect theory to reflect the psychological reactions of decision-makers (DMs) to gains and losses under uncertainty. This method has been regarded as a more accurate risk assessment in many fields. However, it still does not consider the impact of fluctuations in expert psychological factors on priority results.

[0044] Based on this, the present invention combines FMEA with TODIM to more comprehensively reflect the decision maker's preference for risk, and considers the impact of changes in expert psychological preferences on the FMs ranking results by adjusting parameter values. The TODIM method is based on prospect theory and can reflect the decision maker's psychological response to gains and losses under uncertainty. By integrating the idea of ​​multi-criteria decision-making, it provides a more flexible risk assessment framework. Combined with FMEA, different FMs can be ranked more accurately. At the same time, by adjusting the parameter values ​​in TODIM, the impact of changes in expert psychological preferences on priority results can be analyzed, thereby improving the scientific nature of decision-making.

[0045] In view of the above principle analysis, the present invention provides a task-driven FMEA risk assessment method for robot-assisted rehabilitation of children with cerebral palsy, including: a failure mode identification step, according to multiple stages of the robot-assisted rehabilitation process of children with cerebral palsy, after obtaining the task sequence and the failure mode sequence, respectively constructing a task relationship network and a failure mode relationship network, and then using a graph embedding algorithm to generate task node vectors and failure mode node vectors and perform standardized processing, wherein multiple stages are pre-acquired, and the correlation between the task node vectors and the failure mode node vectors is determined by experts to obtain an initial TKL-FML mapping matrix; a risk factor assessment step, based on the standardized task node vectors and failure mode node vectors The importance of failure modes is calculated by using the pre-built RWR-based TKL-FML two-layer network combined with clustering algorithm and DEMATEL method, and the S score is determined. The S score is then combined with the O score and D score based on historical maintenance data to form the original decision matrix of the failure mode, where the two-layer network is constructed based on the TKL-FML mapping matrix; the risk factor weighting step, based on the original decision matrix, uses the q-order orthogonal fuzzy set to calculate the subjective weight of the risk factor, uses the CRITIC method to calculate the objective weight, and then uses game theory to calculate the optimal combination weight; and the failure mode sorting step, based on the optimal combination weight, uses the TODIM method to sort the failure modes and determine the risk priority of each failure mode.

[0046] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0047] Example

[0048] like Figure 1 As shown, this embodiment proposes a task-driven FMEA risk assessment method for robot-assisted rehabilitation of children with cerebral palsy. Its risk assessment framework is based on the task-driven FMEA method and includes four components: identifying FM, evaluating RF, RF weighting and FM sorting. In the assessment system, the task {TK l |1≤l≤L} and FM{FM n |1≤n≤N}, and RF and DM are respectively represented as {RF m |1≤m≤M} and {DM k |1≤k≤K}.

[0049] 1. Task-driven failure mode identification (i.e. failure mode identification steps)

[0050] 1.1、Get tasks and FM

[0051] In the initial stage, the whole rehabilitation process is divided into four main stages, in which several tasks are further analyzed and the locations where failures may occur are determined. These tasks are consistent with the predefined generic task types (GTTs) cited in Table 1, reflecting their inherent problem structure and technical requirements. The task analysis and GTT selection are shown in Table 2. Through this classification, tasks are related to each other, task sequences (TKSs), TKS 1 ={TK 1 ,TK 5 ,TK 6 ,TK 8 These sequences help to build a network that encapsulates the relationships and dependencies between tasks, thus providing a systematic view of the operational framework. The FM in this rehabilitation work system. The model focuses on the interaction between the operation (L) and four elements: software (LS), hardware (LH), environment (LE) and other personnel (LL), as shown in Table 3. FM is classified according to these interactions, failure mode sequences (FMSs), such as FMS 1 ={FM 12 ,FM 13 ,FM 14 ,FM 15 ,FM 16 ,FM 17}, highlighting the interrelationships and potential impacts of FM within the system.

[0052] Table 1 General task types

[0053]

[0054]

[0055] Table 2 Task analysis and GTT selection

[0056]

[0057] Table 3 Failure modes of assisting the rehabilitation of children with cerebral palsy

[0058]

[0059] 1.2. Representing the node vector of tasks and FM through DeepWalk

[0060] Nodes in the same TKS or FMS may have similar underlying features and are connected by edges. Using this principle, we can use the functions in the NetworkX library in Python to draw undirected network graphs of tasks and FMs, and obtain task relationship networks and failure mode relationship networks.

[0061] DeepWalk is then used to generate node vector representations for the task and FM. The DeepWalk methodology consists of two main stages: first, nodes are sampled through a random walk mechanism; then, these nodes are processed using a skip-gram model to learn their vector representations, which are finally encapsulated into low-dimensional vectors. The functionality and application of DeepWalk are detailed in its Algorithm 1, as shown in Table 1-1. These sequences, TKSs and FMSs, are graphically represented in an undirected network graph (e.g. Figure 2 As shown in the figure, DeepWalk plays a key role in generating low-dimensional embedding vectors for these nodes. By comparing several sets of parameters, it is concluded that when the parameters are ω=4, d=8, γ=5, t=20, the calculation effect is better, the coverage of the network is improved, overfitting is avoided, the computational complexity is reduced, and meaningful node vector representation is generated. Using DeepWalk twice, the FM vector Φ is obtained. FM and the task node vector Φ TK , as shown in Tables 4 and 5. The experts determined the correlation between the task nodes and the FM nodes, and obtained the initial TKL-FML mapping matrix A as shown in Table 6. After normalizing these vectors, we get and It then serves as the basic input for the RWR-based two-layer network model.

[0062] Table 1-1 DeepWalk algorithm implementation

[0063]

[0064] Table 4 8-dimensional matrix Φ represented by FM node vector FM

[0065]

[0066]

[0067] Table 5 8-dimensional matrix Φ represented by task node vector TK

[0068]

[0069] Table 6 Initial TKL-FML mapping matrix

[0070]

[0071] 2. Risk factor assessment steps

[0072] 2.1. Severity (S score acquisition process)

[0073] (1) FMs similarity prediction using the RWR-based TKL-FML two-layer network

[0074] Let β be the probability that a particle jumps from TKL to FML or from FML to TKL, and construct the random walk matrix W of the wandering particles in the TKL-FML network:

[0075]

[0076] Taking FML as an example, construct the initial probability matrix

[0077]

[0078] The iterative process of the probability matrix of random walking particles in the “TKL-FML” network is:

[0079] P (t+1) =(1-π)W T P (t) +πP (0) (3)

[0080] Among them, π represents the probability of a free particle restarting random walk, and 1-π is the probability of continuing to walk in FML. When the probability change between iterations satisfies:

[0081] Δ=|P (t+1) -P (t) |≤ε (4)

[0082] , indicating that the calculation has reached a stable state and terminated. From the steady-state probability P (∞) Decomposition. The random walk model has been shown to be insensitive to its parameters. Therefore, this embodiment sets the default values ​​of the parameters to β = 0.5, π = 0.8, and ε = 10 -4 The probability of a link between two FMs can be represented by their cosine similarity, where a higher cosine similarity index indicates a greater likelihood of a link. The FM cosine similarity matrix generated by RWR is shown in Table 7. The cosine similarity matrix is ​​used to evaluate the interdependencies between complex factors, thereby being able to predict the potential impact of FMs on system performance.

[0083] Table 7 FMs cosine similarity matrix S

[0084]

[0085] (2) Calculation of FM importance

[0086] The K-means algorithm is used to cluster the cosine similarity values ​​with k=5, and the FMs direct relationship matrix Y=[y uv ] N×N , as shown in Table 8, where y uv Indicates FM u About FM v degree of impact.

[0087] Table 8 FMs direct relationship matrix Y

[0088]

[0089] DEMATEL uses graph theory and matrix tools to examine system components and establish cause-effect relationships between FMs. This analysis method uses a 5-level scale for measurement: 0-4 (0: no impact, 1: low impact, 2: medium impact, 3: high impact, 4: very high impact). The DEMATEL process unfolds as follows:

[0090] 1) Normalize the direct relationship matrix: sum each row and find the inverse of the maximum value: The standardized direct relationship matrix is: Y'=ρY.

[0091] 2) Generate the total relationship matrix T = Y'(Ι-Y') -1 : This matrix represents the combined effect of the influences between the various elements of the system, where I is the identity matrix.

[0092] 3) The degree of influence and the degree of being affected are determined by and express.

[0093] 4) Calculate FM u Importance

[0094]

[0095] Among them, FM u The S score is σ u .

[0096] 2.2. Incidence and detectability (O score and D score acquisition process)

[0097] Using historical maintenance data to identify RF can reduce subjective bias and enable DMs, decision makers, to more accurately assess potential FMs and implement preventive measures. Converting O to the total number of FMs per year provides a clear measure of the frequency of failures. Similarly, the standard deviation of the time between failures (SDTBF) can be used to effectively evaluate the D score in FMEA. SDTBF measures the variability of the time between failures, indicating the uncertainty in the fault detection process. A smaller SDTBF indicates a higher stability of fault detection, which enables timely detection and management of faults, thereby reducing risks. Conversely, a larger SDTBF indicates unstable fault detection, which may lead to undetected failures and increased risks. By accurately predicting the time of failure, preventive maintenance can be scheduled to address potential FMs before they occur, thereby minimizing their impact. The hospital's historical maintenance data provides the data source and, together with the above-mentioned S score, constitutes the original decision matrix for failure modes, as shown in Table 9.

[0098] Table 9 Original decision matrix D

[0099]

[0100] 3. Risk factor weighting steps

[0101] 3.1. Calculating subjective weights using q-order orthogonal fuzzy sets, i.e., q-ROFS

[0102] α k =(μ k ,ν k ), k = 1, 2, ..., K is a set of q-Rung orthopair fuzzy numbers (q-ROFNs), and ω = (ω 1 ,ω 2 ,...,ω K ) T is corresponding to α k The weight vector (j=1,2,...,K) satisfies ω k ∈[0,1] and The q-Rung orthopair fuzzyweighted aggregation (q-ROFWA) operator is defined as:

[0103]

[0104] In this embodiment, q is set to 3.

[0105] DM uses the q-ROFS language terminology (listed in Table 10) to define the RF m Weight The results of individual evaluations are shown in Table 11. Due to the different professional backgrounds of DMs, each DM is assigned a different weight: ω 1 =0.1,ω 2 =0.3,ω 3 =0.4,ω 4 =0.1,ω 5 =0.1. Represents the evaluation information of the kth DM on the mth RF. The q-ROFWA operator is used to integrate the weight evaluation of the DM's individual evaluation to obtain the comprehensive weight Evaluate. Use the formula The q-ROFS rating value Convert to exact value τ m The standardized subjective weight of RF is obtained as

[0106] Table 10q-Language variables in ROFN

[0107]

[0108] Table 11 Language expression of expert RF weight evaluation

[0109]

[0110] 3.2. Calculating objective weights using the CRITIC method

[0111] Step 1: Establish a decision matrix D = [d mn ] M×N , where: M is the number of RFs, N is the number of FMs, d mn is the score of the mth RF relative to the nth FM.

[0112] Step 2: Normalize RF using the following formula to establish a standardized decision matrix D' = [d' mn ] M×N :

[0113]

[0114] Among them, min{d mn} is the minimum value in the mth RF, max{d mn} is the maximum value in the mth RF.

[0115] Step 3: Calculate the correlation coefficient r between the i-th and m-th RF (i.e., the reference risk factor and the target risk factor) im :

[0116]

[0117] in and is the average of the normalized scores of the i-th and m-th RFs.

[0118] Step 4: Contrast σ of the mth RF m and the conflict score c m :

[0119] σ m = std(d' m1 ,d' m2 ,...,d' mN ) (9)

[0120]

[0121] Step 5: Calculate the objective weight of the mth RF:

[0122]

[0123] This formula combines the contrast and conflict scores of each RF to provide an objective weight.

[0124] 3.3. Calculate the optimal combination weights using game theory

[0125] In order to improve the rationality of the weight aggregation scheme, this embodiment calculates the combination coefficient of subjective and objective weights and the optimal combination weight based on the idea of ​​non-cooperative game theory. The principle is as follows:

[0126] Step 1: Construct the objective function.

[0127] Define the combination coefficient k 1 and k 2 , respectively represent the coefficients of subjective weight and objective weight. The combined weight is expressed as In order to minimize the deviation between the combined weight and the subjective weight and the objective weight, the following objective function is established:

[0128]

[0129] Step 2: Determine optimal conditions.

[0130] For a continuous function on a closed interval to obtain an extreme value, the first-order partial derivative in all directions must be zero. By deriving the objective function, the equation is established:

[0131]

[0132] Solving this set of linear equations, we obtain a set of combination coefficients under Nash bargaining.

[0133] Finally, the optimal combination weights can be calculated:

[0134]

[0135]

[0136] The final ensemble weights are summarized in Table 12.

[0137] Table 12 Weights of RFs

[0138]

[0139] 4. Failure mode ranking steps

[0140] This example uses the TODIM method to sort FMs. The loss attenuation coefficient θ is set to 200, which reflects the psychological preference of DM to deal with risks. The operation steps are as follows:

[0141] Step 1: From the standardized decision matrix D' = [d' mn ] M×N Start, where d' mn It is RF m Relative to FM n The standardized score is consistent with the standardized matrix in the above calculation of objective weights.

[0142] Step 2: Calculate the relative weight of RF.

[0143] Set RF m The optimal combination weight is ω m , the relative weights are calculated using the following equation:

[0144]

[0145] Among them, ω r is the maximum combined weight, which refers to the maximum value of the combined weights of the three risk factors RF, that is, ω r =max{ω m},m=1,2,3.

[0146] Step 3: Calculate the RF m Download FM n With FM v The relative advantage between the target failure mode and the reference failure mode.

[0147] Solution FM n Relative to Scheme FM v The advantages are:

[0148]

[0149] Wherein, θ represents the loss attenuation coefficient, θ>0. If 0<θ<1, the influence of loss increases, and if θ>1, on the contrary, when θ=1, the influence of loss and gain is ignored.

[0150] Step 4: Get FM n and FM v The comprehensive advantage between them can be calculated using the following equation:

[0151]

[0152] Step 5: Get the global dominance:

[0153]

[0154] According to the global dominance, the FMs were ranked and the risk priority of each FM was determined (Table 13).

[0155] Table 13 Risk priority ranking of FMEA

[0156]

[0157] Next, the effectiveness of the method provided in this embodiment is analyzed and verified.

[0158] 5. Sensitivity analysis

[0159] 5.1 Loss attenuation coefficient θ

[0160] The loss attenuation coefficient, denoted by θ in the TODIM methodology, is crucial in describing the psychological characteristics of participants in a decision-making environment. This coefficient determines the boundaries of risk acceptance, thereby shaping individual preferences and decisions. Sensitivity analysis of θ reveals insights into how participants perceive and manage risk. Figure 3 It is shown that the ranking of FMs varies with different values ​​of θ and these changes do not significantly affect the overall engineering properties, demonstrating the robustness of the approach providing the best results at θ = 200.

[0161] 5.2 Parameter q in q-ROFS

[0162] The proposed FMEA method is affected by the parameter q in its implementation. The use of parameter q provides a wider description space for expert evaluation information, which is not possible with IFS and PFS. At the same time, the q value also provides greater flexibility for DM decision-making. The impact of parameter q on risk assessment results and the sensitivity analysis of these results are analyzed here. In order to illustrate the impact of q value on FM ranking results, a Figure 4The bar graph shown in Figure 2 shows that the x-axis represents FM and the y-axis represents the ranking results. Sensitivity analysis shows that the proposed method can consistently identify FMs with high risk levels regardless of changes in the q-value, thus demonstrating the robustness of the method to parameter fluctuations.

[0163] 6. Comparative analysis

[0164] The effectiveness of the proposed method is further verified by comparing it with existing technologies, including the traditional RPN method, the fully data-driven FMEA method, and the TOPSIS-FMEA method. Figure 5 The resulting FM rankings using these methods are shown. Although the trends are generally similar for all methods, differences in rankings still exist.

[0165] A completely data-driven FMEA approach assesses the severity of FMs based on mean time to repair (MTTR) and relies extensively on historical data and statistical records. This approach may overlook potentially serious FMs that have not yet occurred. In contrast, the approach proposed in this embodiment combines task-driven and SHELL model strategies with expert insights and objective data, improving adaptability to rapid technological and market changes. For example, the significant ranking difference of FM8: unexpected activation of the emergency stop system may stem from its rarity or absence in historical data, so the data-driven approach may underestimate the severity, resulting in a lower risk priority.

[0166] TODIM and TOPSIS used in MCDM differ in their evaluation focus. TOPSIS evaluates alternatives relative to an ideal solution, while TODIM emphasizes the subjective preferences and psychological state of the DM. The analysis shows that both methods provide comparable rankings, rooted in practical applications. Therefore, specific scenarios should be considered when choosing between them to enhance decision-making in areas such as robot-assisted rehabilitation.

[0167] TODIM is preferable in cases where there is a large degree of patient variability or where there is an over-reliance on expert experience, as it allows for better identification of potential risks. For those who prioritize solution outcomes, TOPSIS may be more appropriate for outcomes that are closest to the ideal rehabilitation solution. In a comparative analysis with the traditional RPN approach, significant differences in FM ranking were observed. One of the main goals of this study was to reduce subjectivity in the FMEA process; therefore, direct RPN scoring was not used. Instead, RPN scores were calculated using a 1-10 rating defined by experts for each RF for comparison, as shown in Table 14. This expert-driven scoring approach requires tailored adjustments to suit different datasets or domains, as the scalability proposed in Table 14 may not be universally applicable. In addition, the traditional RPN approach ignores the weights of the RFs, resulting in the same importance for each factor. Therefore, this approach may not accurately reflect reality.

[0168] In summary, the comparative analysis verifies that the proposed approach is feasible, effective, and more in line with real-world decision-making scenarios.

[0169] Table 14 Expert-driven RPN

[0170]

[0171] Furthermore, this embodiment also provides a task-driven FMEA risk assessment device for robot-assisted rehabilitation of children with cerebral palsy, comprising a memory, a processor, and a program stored in the memory, wherein the processor implements one or more steps of the aforementioned method when executing the program, and the specific process will not be repeated here.

[0172] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A task-driven FMEA risk assessment method for robot-assisted rehabilitation of children with cerebral palsy, characterized in that: The following steps are involved: Failure mode identification step, according to multiple stages of the robot-assisted rehabilitation process for children with cerebral palsy, after obtaining the task sequence and the failure mode sequence, respectively construct a task relationship network and a failure mode relationship network, and then use a graph embedding algorithm to generate a task node vector and a failure mode node vector and perform a standardized process, wherein the multiple stages are pre-acquired, and the correlation between the task node vector and the failure mode node vector is determined by an expert to obtain an initial TKL-FML mapping matrix; The risk factor evaluation step is based on the standardized task node vector and failure mode node vector, and uses the pre-built RWR-based TKL-FML two-layer network combined with the clustering algorithm and the DEMATEL method to calculate the importance of the failure mode, determine the S score, and together with the O score and D score based on the historical maintenance data, form the original decision matrix of the failure mode, wherein the two-layer network is constructed based on the TKL-FML mapping matrix; The risk factor weighting step is to calculate the subjective weights based on the original decision matrix using q-order orthogonal fuzzy sets, calculate the objective weights using the CRITIC method, and then calculate the optimal combination weights using game theory; The failure mode ranking step is to sort the failure modes based on the optimal combination weights using the TODIM method to determine the risk priority of each failure mode.

2. The task-driven FMEA risk assessment method for robot-assisted rehabilitation of children with cerebral palsy according to claim 1 is characterized in that: In the failure mode identification step, the failure mode sequence is obtained using a pre-built SHELL model.

3. The task-driven FMEA risk assessment method for robot-assisted rehabilitation of children with cerebral palsy according to claim 2 is characterized in that: In the failure mode identification step, the task node vector and the failure mode node vector are obtained using the DeepWalk method, and the specific process includes: Using a random walk mechanism to sample nodes in the task relationship network and the failure mode relationship network respectively; Use the pre-built word skipping model to process the sampled task nodes and failure mode nodes; The processed task nodes and failure mode nodes are encapsulated into corresponding low-dimensional embedding vectors respectively.

4. The task-driven FMEA risk assessment method for robot-assisted rehabilitation of children with cerebral palsy according to claim 1, characterized in that: In the risk factor assessment step, the importance of the failure mode is calculated and the S score is determined by using the pre-built RWR-based TKL-FML two-layer network combined with the clustering algorithm and the DEMATEL method. The specific process includes: Constructing a random walk matrix in the TKL-FML two-layer network; Based on the random walk matrix, the cosine similarity between the failure modes is calculated to obtain the cosine similarity matrix of the failure modes; Clustering the cosine similarity matrix using a K-means algorithm to obtain a failure mode direct relationship matrix; After the direct relationship matrix of the failure mode is standardized, a total relationship matrix is ​​generated, and the importance of the failure mode is calculated by combining the impact degree data and the impact degree data obtained in advance to determine the S score.

5. The task-driven FMEA risk assessment method for robot-assisted rehabilitation of children with cerebral palsy according to claim 4, characterized in that: In the risk factor assessment step, the O score is determined based on the failure rate calculated based on the total number of failure modes per year, and the D score is determined based on the standard deviation of the failure interval time, and the total number of failure modes and the failure interval time are both derived from pre-acquired historical maintenance data.

6. The task-driven FMEA risk assessment method for robot-assisted rehabilitation of children with cerebral palsy according to claim 1, characterized in that: In the risk factor weighting step, the subjective weight is calculated using q-order orthogonal fuzzy sets, and the specific process includes: Define the q-order orthogonal pair fuzzy weighted aggregation operator and obtain the linguistic variable table of the q-order orthogonal fuzzy set; Multiple decision makers use the language variable table to individually evaluate the weights of risk factors, and different decision makers correspond to different weights; The q-order orthogonal pair fuzzy weighted aggregation operator is used to integrate the individual evaluation results of each decision maker to obtain a comprehensive weighted evaluation; After converting the comprehensive weight assessment into an exact value, the standardized subjective weight of the risk factor is obtained.

7. The task-driven FMEA risk assessment method for robot-assisted rehabilitation of children with cerebral palsy according to claim 6, characterized in that: In the risk factor weighting step, the objective weight is calculated using the CRITIC method, and the specific process includes: Standardizing the original decision matrix to obtain a standardized decision matrix; Determine the reference risk factor and the target risk factor, and calculate the correlation coefficient between the two; Based on the correlation coefficient, calculating contrast and conflict scores of the target risk factor; Based on the contrast and the conflict score, an objective weight of the target risk factor is calculated.

8. The task-driven FMEA risk assessment method for robot-assisted rehabilitation of children with cerebral palsy according to claim 7, characterized in that: In the risk factor weighting step, game theory is used to calculate the optimal combination weight, and the specific process includes: Defining a combined weight and constructing an objective function, wherein the objective function is used to minimize the deviation between the combined weight and the subjective weight and the objective weight; Based on the objective function, the pre-constructed optimal condition linear equation is solved to obtain the combination coefficient, and then the optimal combination weight is calculated.

9. The task-driven FMEA risk assessment method for robot-assisted rehabilitation of children with cerebral palsy according to claim 1, characterized in that: The failure mode ranking step specifically includes: a. Determine the target failure mode and reference failure mode; b. Based on the optimal combined weight, determine the maximum combined weight among all risk factors; c. calculating the relative weights of the risk factors based on the optimal combination weight and the maximum combination weight; d. Based on the relative weight, calculate the relative advantage between the target failure mode and the reference failure mode under the risk factor, and then obtain the comprehensive advantage between the two; e. Based on the comprehensive dominance, calculate the global dominance of the target failure mode; Repeat a~e to obtain the global dominance of all failure modes, sort them, and determine the risk priority of each failure mode.

10. A task-driven FMEA risk assessment device for robot-assisted rehabilitation of children with cerebral palsy, comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 9 is implemented.

Images