An abnormality detection and fault prediction method based on multi-source operation data credible fusion of fat-reducing equipment

Abstract

The application discloses a kind of based on the abnormality detection and fault prediction method of multi-source operation data credible fusion of fat-reducing equipment, comprising: obtaining the multi-source operation data of fat-reducing equipment and attaching operation phase mark and windowing processing;Establish the health baseline of different operation phases;From multi-source operation data, extract features, and calculate the credibility weight of each data source;Based on weight, the multi-source operation characteristics are handled by gate, then combined with operation phase to splice and obtain fusion features, input abnormality detection model, output overall abnormality score and the abnormal contribution of each data source;If overall abnormality score exceeds threshold or safety abnormal event is detected, introduce cross-module attribution model, output suspected fault module, fault type and attribution confidence;According to pre-set degradation evaluation period, the degradation state sequence of each module is constructed, and through staged monotonic degradation-discrete time survival joint model based on attribution constraint, predictive maintenance window is generated;Output hierarchical disposal strategy.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis and predictive maintenance technology for weight loss equipment. Specifically, it relates to an anomaly detection and fault prediction method based on the reliable fusion of multi-source operating data from weight loss equipment. Background Technology

[0002] Existing energy-based fat reduction devices typically include an energy output unit, a treatment execution unit, a cooling unit, a temperature monitoring unit, and a control unit. The energy output unit applies laser, radio frequency, ultrasound, or other forms of therapeutic energy to the target treatment area; the cooling unit controls the surface temperature of the treatment area; the temperature monitoring unit acquires temperature information of the treatment area; and the control unit controls the treatment energy and cooling parameters. Some automated fat reduction devices also include a motion execution mechanism to adjust the position, posture, and movement path of the treatment head relative to the target treatment area.

[0003] With the increasing automation and system integration of weight loss equipment, the equipment typically includes energy output links, cooling execution links, temperature monitoring links, motion control links, power supply links, and communication links. Existing equipment can generally perform self-tests and threshold alarms based on current, voltage, temperature, position status, or communication status, and perform protective operations such as reducing energy output, pausing treatment, or emergency shutdown when operating parameters exceed preset safety thresholds.

[0004] However, most existing equipment self-testing and alarm methods set fixed thresholds for single sensor data or single device modules, primarily used to identify relatively obvious anomalies such as overcurrent, overvoltage, overheating, and communication interruptions. For slow degradation issues such as performance degradation of energy output units, gradual blockage of cooling channels, decreased efficiency of cooling units, wear of motion actuators, sensor drift, and gradually increasing communication delays, the relevant operating parameters usually have not yet exceeded the fixed alarm thresholds in the early stages, making timely identification difficult.

[0005] Furthermore, fat reduction devices exhibit different normal operating characteristics at different stages of operation, including before, during, and after treatment. Specifically, device self-check, area acquisition, and treatment head positioning belong to the pre-treatment stage; energy output, path execution, synchronous cooling, and short-term pause belong to the during-treatment stage; and stopping energy output, continuous cooling, actuator return to position, and data saving belong to the post-treatment stage.

[0006] Furthermore, an abnormality in a single operating parameter does not necessarily indicate a equipment malfunction. For example, an abnormal temperature response in the treatment area may be caused by energy output attenuation, abnormal cooling intensity, changes in treatment head distance, movement of the treated object, or a malfunction in the temperature monitoring unit. Existing detection methods based on single-parameter thresholds struggle to differentiate between different causes of abnormalities by utilizing the correlation between energy output, cooling execution, motion control, and treatment temperature response.

[0007] Existing remote operation and maintenance systems typically upload equipment operating parameters, alarm logs, and maintenance records, allowing maintenance personnel to remotely troubleshoot after a fault occurs. However, these systems primarily focus on post-fault status queries and auxiliary diagnostics, lacking early abnormal pattern recognition, cross-module fault attribution, and remaining uptime prediction for weight loss equipment during the treatment phase. Furthermore, confirmation results after equipment repair, cleaning, calibration, and component replacement are usually only stored as maintenance records and are not fully utilized to correct the equipment's health baseline and fault prediction results at different operating stages.

[0008] Therefore, there is a need for an anomaly detection and fault prediction method suitable for fat reduction equipment. This method can determine the current operating stage based on the equipment control status, identify early anomalies by utilizing the correlation between control commands, actual equipment execution results, and thermal response of the treatment area, and output suspected fault modules, fault types, and predictive maintenance information based on data quality anomalies, patient disturbances, and equipment component failures. It can also achieve closed-loop optimization based on maintenance feedback. Summary of the Invention

[0009] To address the shortcomings of existing technologies, such as reliance on a single threshold for equipment self-inspection, difficulty in identifying early signs of failure, insufficient cross-module anomaly attribution capabilities, remote diagnosis biased towards post-incident investigation, and maintenance feedback unable to continuously optimize the detection model, this invention aims to overcome these deficiencies by proposing an anomaly detection and fault prediction method based on the reliable fusion of multi-source operational data from weight-loss equipment. This method upgrades equipment operation and maintenance from passive alarm to proactive prediction, from single-point detection to multi-source fusion, from manual investigation to intelligent attribution, and from post-incident maintenance to predictive maintenance.

[0010] In view of this, the present invention proposes an anomaly detection and fault prediction method based on the reliable fusion of multi-source operating data of a weight loss device, comprising: Step 1: Acquire multi-source operational data of the fat reduction device and attach operational stage markers, including pre-treatment stage, during-treatment stage, and post-treatment stage; Step 2: Perform time synchronization, data cleaning, and windowing processing on multi-source operational data; Step 3: Based on the operating stage, equipment configuration, treatment parameters, and historical normal operation data of the fat reduction equipment, establish health baselines for different operating stages; Step 4: Extract features from the multi-source running data after windowing processing in Step 2, and calculate the credibility weight of each data source; Step 5: Based on the credibility weights of each data source obtained in Step 4, the multi-source running features are gating processed, and then combined with the running stages to obtain fused features. The fused features are input into the pre-trained stage condition credibility gating dual-branch temporal anomaly detection model, and the overall anomaly score and anomaly contribution of each data source in the current time window are output. Step 6: If the overall anomaly score of the current time window exceeds the anomaly judgment threshold corresponding to the current running stage or a safety anomaly event is detected, the pre-trained cross-module attribution model is introduced to output the suspected fault module, fault type and attribution confidence. Step 7: According to the preset degradation assessment cycle, based on the output of Step 5 and / or the output of Step 6, construct the degradation state sequence of each module of the fat reduction equipment, and generate a predictive maintenance window through a staged monotonic degradation-discrete-time survival joint model based on attribution constraints. Step 8: Output the tiered handling strategy.

[0011] As an improvement to the above method, the multi-source operating data in step 1 includes at least two of the following: energy output link data, cooling link data, treatment head sensing link data, motion execution link data, power supply link data, communication link data, and system log data.

[0012] As an improvement to the above method, the health baseline in step 3 includes at least the normal range of features, the trend of feature changes, and the normal correlation between modules, wherein the normal correlation between modules is used to describe the normal correspondence between control commands, actual execution results of the device, and response of the treatment area; For the s-th operational phase and the p-th combination of treatment parameters, the equipment health baseline Represented as:

[0013] in, This represents the mean vector of each feature at this stage. This represents the covariance matrix or normal fluctuation range. Indicates the range of slope of characteristic change. This represents a cross-module association matrix.

[0014] As an improvement to the above method, step 4 includes: Extract time-domain features, frequency-domain features, trend features, image features, log features, and cross-module consistency features from the windowed multi-source runtime data; The credibility score of the x-th data source within the k-th window is calculated using the following formula. :

[0015] in, Indicates data integrity score, Indicates the signal stability score. This indicates the equipment's self-test score. Indicates the quality of time synchronization. Indicates the quality of image or sensor data. ~ This represents the corresponding weight coefficient, which is used to normalize the credibility scores of each data source to obtain the credibility weight. :

[0016] Where N represents the number of data sources, To prevent tiny constants with a denominator of zero.

[0017] As an improvement to the above method, the stage condition confidence-gated dual-branch temporal anomaly detection model in step 5 includes: a data source feature encoding module, a shared temporal encoding module, a reconstruction branch, a prediction branch, an association consistency calculation module, and an output module; wherein, the data source feature encoding module and the shared temporal encoding module are connected in series and then connected in parallel to the reconstruction branch, the prediction branch, and the association consistency calculation module, and then the output module is connected in series. The data source feature encoding module is used to encode the multi-source operating features respectively to obtain the local representation of each data source; The shared temporal coding module employs a temporal convolutional network with causal convolution and dilated convolution structures to encode the fused features of multiple consecutive time windows, extracting the short-term variation features and long-term variation trends of equipment operating parameters. The reconstruction branch is used to reconstruct the multi-source running features within the current time window based on the hidden features output by the shared temporal coding module, and to calculate the reconstruction error. The prediction branch is used to predict the multi-source operation characteristics of the next time window based on the time series characteristics of the current time window and those before it, and to calculate the prediction error after the data of the next time window arrives. The correlation consistency calculation module is used to calculate the cross-module correlation consistency error based on the standardized correlation residuals and weights of different modules of the weight loss equipment within the current time window. The output module is used to obtain the overall anomaly score of the current time window by weighted summation based on the reconstruction error, prediction error, and cross-module correlation consistency error; it is also used to calculate the anomaly contribution of each data source based on the proportion of each data source in the reconstruction error, prediction error, and cross-module correlation consistency error.

[0018] As an improvement to the above method, the cross-module attribution model in step 6 is a stage-conditional physical topology graph attention network; the processing includes: Construct a directed device association graph based on the control and information transmission relationships between modules. :

[0019] In the formula, Represents a set of module nodes. This represents the set of directed associated edges enabled in the current running phase s; For each directed association edge activated in the current stage, the attention coefficient between modules is calculated based on the attribution input features of adjacent module nodes, cross-module association residuals, and the current running stage, thereby obtaining the features of each module after cross-module information propagation. Based on the graph propagation characteristics of each module node, calculate the probability that the i-th module is a suspected faulty module within the k-th time window. for:

[0020]

[0021] In the formula, This represents the characteristics of the i-th module node after graph information propagation in the k-th time window. and These represent the trainable weight vector and bias corresponding to the module failure probability calculation, respectively. This represents the fault source score of the i-th module in the k-th time window. Indicates the first The fault source score of each module in the k-th time window. Indicates the number of device modules; For each module of the weight loss device, set a corresponding fault type classifier and output the probability of the q-th fault type of the i-th module. :

[0022] In the formula, and Let the weight matrix and bias vector of the fault probability type classifier for the i-th module be represented respectively. This represents the number of fault types corresponding to the i-th module; Softmax is the normalization exponential function. Let the matching degree between the i-th module's q-th type fault and the current anomaly mode be . Then the corrected attribution probability for:

[0023] In the formula, The fusion coefficient between the attention network output and the abnormal pattern matching result of the stage conditional physical topology graph is represented. The corrected attribution probabilities for all modules and fault types are normalized to obtain the final attribution probabilities. :

[0024] In the formula, Indicates the first The number of fault types corresponding to each module This indicates a small positive number that prevents the denominator from being zero. Indicates the summation index of the fault type; The failure attribution confidence level is obtained according to the following formula. : .

[0025] As an improvement to the above method, step 6 further includes: When the calculated fault attribution confidence is lower than the preset confidence threshold, or the difference between the two fault types with the highest attribution probabilities is less than the preset distinction threshold, multiple candidate fault types and probabilities are output, and a manual review prompt is generated.

[0026] As an improvement to the above method, step 7, which involves constructing the degradation state sequence of each module of the fat reduction device, specifically includes: For the i-th module of the weight loss device, during the n-th degradation evaluation period, construct the module degradation input features. :

[0027] In the formula, This indicates the cumulative load characteristics of the module. Indicates the abnormal trend characteristics of the module. Indicates fault attribution characteristics, Indicates historical preservation characteristics; The degradation features were encoded separately for the pre-treatment, treatment, and post-treatment stages, and then concatenated to obtain the joint degradation features. ; The joint degradation features from multiple consecutive degradation evaluation periods are input into a causal temporal convolutional encoder to obtain the temporal degradation features of the i-th module in the n-th evaluation period. ; Degraded input features and temporal degradation characteristics The degradation state characteristics corresponding to this assessment period are obtained by splicing them together. Multiple consecutive degradation state features are arranged in the order of the evaluation cycle to obtain a degradation state sequence.

[0028] As an improvement to the above method, the attribution constraints in step 7 specifically include: Use the final attribution probability obtained in step 6 Temporal degradation characteristics Constraints are applied to the attribution gating coefficient of the i-th module. Represented as:

[0029] Temporal degradation characteristics after attribution constraints Represented as:

[0030] In the formula, Indicates the basic retention factor. This represents the adjustment parameter by which the fault attribution results affect the degradation characteristics.

[0031] As an improvement to the above method, the processing procedure of the phased monotonically degenerate-discrete-time survival joint model in step 7 includes: Based on the temporal degradation characteristics after attribution constraints Calculate the degradation increment for the current degradation assessment period. for:

[0032] In the formula, , and These represent the trainable weights and biases corresponding to the degradation increment calculation of the i-th module, respectively. Represents a non-negative activation function; When maintenance is not performed:

[0033] After performing maintenance:

[0034] In the formula, This represents the degradation state of the i-th module at the end of the (n-1)-th degradation evaluation period. This represents the degradation state of the i-th module at the end of the n-th degradation evaluation period. Let represent the residual degradation coefficient of the i-th module after maintenance is performed in the n-th degradation assessment cycle, and satisfy . ; The conditional probability that the i-th module will experience a maintenance failure in the r-th prediction interval in the future. Represented as:

[0035] In the formula, This represents the conditional probability that the i-th module will experience a maintenance-required failure in the r-th prediction interval, provided that no maintenance-required failure has occurred previously. This indicates the temporal degradation characteristics after the attribution constraint ends; Indicates the current degradation state; This represents the position code of the r-th prediction interval in the future; Indicates the preset planned usage intensity; and These represent the trainable weights and biases of the survival prediction output layer, respectively. The probability that the i-th module will remain in a maintenance-free state for the first r prediction intervals. for:

[0036] Probability of cumulative failures :

[0037] The estimated remaining run time of the i-th module Represented as:

[0038] In the formula, v Indicates the index of the prediction interval for product calculation. Indicates the maximum number of prediction intervals. Indicates the runtime corresponding to a single prediction interval; Before generating the predictive maintenance window, a first maintenance risk threshold is preset. Second maintenance risk threshold And satisfy:

[0039] Determine the starting position of the maintenance window for the i-th module. for:

[0040] Determine the end position of the maintenance window for the i-th module. for:

[0041] Thus, the predictive maintenance window of the i-th module is obtained. for:

[0042] in, This represents the baseline time corresponding to the end of the nth degradation assessment period. Indicates the time length corresponding to a single prediction interval Compared with the prior art, the advantages of the present invention are: 1. This invention establishes corresponding health baselines for the pre-treatment, treatment, and post-treatment stages of the fat reduction device. Furthermore, it combines treatment energy output, cooling intensity, movement speed, treatment head distance, and treatment area area to call the corresponding stage-condition health baseline. This can avoid false alarms caused by differences in data distribution under different operating stages and treatment parameters, and improve the pertinence and accuracy of anomaly detection.

[0043] 2. This invention uses data integrity, time synchronization quality, signal stability, device self-test results, and image quality to calculate the reliability threshold value of each data source, and distinguishes between data quality anomalies and device status anomalies. This can reduce the interference of data loss, time asynchrony, noise increase, image blurring, and sensor acquisition anomalies on the detection results, and improve the stability and robustness of multi-source data fusion.

[0044] 3. This invention calculates the overall anomaly score by jointly considering reconstruction error, prediction error, and cross-module correlation consistency error. This not only identifies anomalies that have significantly deviated from the normal range, but also identifies early anomalies where a single parameter has not yet exceeded a fixed threshold, but the correlation between control commands, actual equipment execution results, and treatment area responses has changed. This improves the ability to identify slow degradation faults and potential fault precursors.

[0045] 4. This invention constructs a stage-condition physical topology diagram based on the module control relationship, energy transfer relationship, and information transfer relationship of the fat reduction device at different operating stages. It then utilizes a graph attention network to jointly analyze the energy output link, cooling link, treatment head sensing link, motion execution link, power supply link, and communication link. This avoids relying solely on a single abnormal parameter for fault diagnosis and improves the accuracy of attributing suspected faulty modules and fault types. Attached Figure Description

[0046] Figure 1 This is a flowchart of the overall process for abnormal detection and fault prediction of weight loss equipment; Figure 2 This is a flowchart of the phased baseline and anomaly detection process; Figure 3 It is a flowchart for generating cross-module fault attribution and predictive maintenance windows. Detailed Implementation

[0047] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0048] Example 1 This invention provides an anomaly detection and fault prediction method based on the reliable fusion of multi-source operational data from a weight loss device, such as... Figure 1 As shown, the method includes the following steps: S1: Multi-source operational data acquisition and operational phase marking Acquire multi-source operational data of the fat reduction device during one or more treatment sessions, and attach operational phase markers to the multi-source operational data. The operational phases include the pre-treatment phase, the treatment phase, and the post-treatment phase.

[0049] The fat reduction device includes one or more of the following: an energy output unit, a treatment execution unit, a treatment head sensing unit, a control unit, a cooling unit, a motion execution mechanism, a power management unit, and a communication unit. The energy output unit is used to apply laser, radio frequency, ultrasound, or other forms of therapeutic energy to the target treatment area.

[0050] The multi-source operational data includes two or more of the following: energy output link data, cooling link data, treatment head sensor link data, motion execution link data, power link data, communication link data, and system log data. The system selects the corresponding data source for collection based on the specific structure and functional configuration of the fat reduction device; it is not required that different types of fat reduction devices have the same functional modules.

[0051] The energy output link data includes one or more of the following: therapeutic energy output command, energy output unit drive parameters, actual therapeutic energy output value, output duration, output frequency, output duty cycle, output uniformity index, and output attenuation rate.

[0052] The cooling link data includes one or more of the following: cooling control commands, cooling level, cooling unit drive parameters, actual cooling output, cooling medium flow rate, cooling air velocity, cooling response delay, and cooling rate of the treatment area.

[0053] The treatment head sensor link data includes one or more of the following: treatment area surface temperature, temperature change rate, infrared temperature image, visible light image, image clarity, image brightness stability, treatment head distance, contact status, and sensor data missing rate.

[0054] The motion execution link data includes one or more of the following: target position, actual position, target attitude, actual attitude, position error, attitude error, motion speed, drive current, drive temperature, vibration amplitude, tracking lag, repetitive positioning deviation, and emergency stop trigger record.

[0055] The power link data includes one or more of the following: power supply voltage, power supply current, output power, power fluctuation, overvoltage record, and overcurrent record.

[0056] The communication link data includes one or more of the following: communication delay, data packet loss rate, data retransmission rate, data transmission time interval, control command confirmation time, and communication interruption record.

[0057] The system log data includes one or more of the following: device status switching events, operation events, alarm events, abnormal events, event occurrence time, event duration, and event level.

[0058] In one specific embodiment, the fat reduction device is a non-contact laser fat reduction robot. In this case, the energy output link data includes one or more of the following: treatment energy output command, energy output unit drive parameters, actual treatment energy output, spot uniformity index, and laser output power attenuation rate; the motion execution mechanism is a robotic arm, and the motion execution link data includes one or more of the following: robotic arm joint current, joint temperature, end-effector position error, end-effector posture error, tracking lag, vibration amplitude, and repetitive positioning deviation.

[0059] The pre-treatment stage refers to the operational phase from when the fat reduction device starts the treatment task to when the treatment officially begins, including one or more of the following operations: device self-check, treatment area information collection, treatment path generation, treatment parameter setting, and treatment head positioning.

[0060] The treatment phase refers to the operational phase from the formal start of the treatment task to its end or termination, including one or more operations such as treatment energy output, treatment process monitoring, treatment area cooling, movement of the treatment actuator, position adjustment between treatment areas, and short-term pauses during the treatment process.

[0061] The post-treatment phase refers to the operational phase from the end or termination of the current treatment task to the completion of the finishing operations of the fat reduction equipment after treatment. It includes one or more of the following operations: stopping the output of treatment energy, continuing to cool the treatment area, controlling the treatment head to exit the treatment area, controlling the motion actuator to return to its original position, saving operational data, and generating maintenance evaluation results.

[0062] The current operational stage is determined based on the treatment task status flags. When the treatment task has not yet officially started, the fat reduction equipment is designated as the pre-treatment stage; when the treatment task has officially started and no treatment completion or termination event has been received, the fat reduction equipment is designated as the in-treatment stage; when a treatment completion, manual termination, manual cancellation, or abnormal termination event is received, the fat reduction equipment is designated as the post-treatment stage.

[0063] The treatment task status flag is generated based on the device control status word, treatment start event, treatment end event, and treatment task execution status, and is verified by one or more of the following: energy output enable status, cooling unit status, treatment task execution status, and motion actuator status.

[0064] S2: Multi-source data time synchronization, cleaning, and windowing processing The multi-source operational data collected in step S1 undergoes time synchronization, missing value processing, acquisition pseudo-anomaly identification, data quality marking, sampling frequency unification, and windowing processing. Data exceeding the sensor's physical measurement range, communication padding values, duplicate data, timestamp errors, and obvious acquisition errors are marked or corrected, while the original data is retained. Deviation data that may be caused by equipment failure is not deleted but is judged by the subsequent stage's conditional confidence-gated dual-branch timing anomaly detection model. Since the sampling frequencies of the energy output, cooling section, image acquisition unit, infrared imaging unit, motion actuator controller, and power management module are different, the system can use unified timestamps, interpolation, resampling, event alignment, or sliding time window aggregation to map data from different sources into the same operational window.

[0065] For continuous time-series data, a sliding window of length T is used for segmentation; for image or infrared image data, image features within the corresponding time window are extracted; for system log data, log events are converted into structured features such as event codes, event frequencies, event durations, and event levels. Let the multi-source data within the k-th time window be represented as... ,in This represents energy output link data. This indicates cooling link data. This represents the sensor link data of the treatment head. This represents motion execution link data. Indicates power link data, This represents communication and log data.

[0066] S3: Establishing a phased equipment health baseline Based on the operating stage, equipment structure configuration, treatment parameters, and historical normal operation data of the fat reduction equipment, establish equipment health baselines corresponding to the pre-treatment, treatment, and post-treatment stages.

[0067] Data used to establish equipment health baselines comes from historical treatment tasks where equipment self-tests passed, no alarms or safety interlocks were triggered, no maintenance confirmation faults occurred, and the operating results were confirmed to be normal, as well as from equipment calibration tests or post-maintenance retests. Operating data with missing data, communication interruptions, or maintenance confirmation faults are not used to establish equipment health baselines.

[0068] Based on the specific structural configuration of the fat reduction device, determine the device links involved in establishing the health baseline. When the fat reduction device is equipped with an energy output unit, establish a normal correlation between energy output commands, drive parameters, and actual energy output; when the fat reduction device is equipped with a cooling unit, establish a normal correlation between cooling commands, actual cooling output, and the temperature response of the treatment area; when the fat reduction device is equipped with a motion actuator, establish a normal correlation between the target motion state and the actual position, posture, speed, and drive parameters; when the fat reduction device is equipped with a treatment head sensing unit, establish a normal range of variation for sensor data such as temperature, image, distance, or contact status.

[0069] The treatment parameters include one or more of the following: treatment energy output, cooling intensity, movement speed, treatment head distance, treatment area, and treatment duration. Historical normal operation data are divided into different categories according to the preset range of each treatment parameter, and equipment health baselines are established for different operating stages and categories.

[0070] For the s-th operating phase and the p-th type of treatment parameters, the device health baseline includes at least the normal range of features, the trend of feature changes, and the normal correlation across modules.

[0071] Among them, the normal range of features is used to describe the normal value range and fluctuation degree of each operating feature under the corresponding stage and parameter; the feature change trend is used to describe the normal slope, change rate or change direction of the operating feature over time; and the normal cross-module correlation is used to describe the normal correspondence between control commands, actual equipment execution results and treatment area response.

[0072] In the pre-treatment phase, the device health baseline includes at least the normal range of device self-test status, sensor data quality, communication status, treatment head position or distance, and motion actuator positioning status.

[0073] During the treatment phase, the device health baseline includes at least the normal correlation between the treatment energy output command and the actual treatment energy output, the normal correlation between the actual treatment energy output and the temperature response of the treatment area, and the normal correlation between the cooling output and the cooling response of the treatment area when a cooling unit is provided; when a motion actuator is provided, it also includes the normal correlation between the target motion state and the actual motion state.

[0074] In the post-treatment phase, the device health baseline includes at least the normal range of the treatment energy off state, the temperature recovery process of the treatment area, the motion actuator return state, the communication upload state, and the operation data storage state.

[0075] When the current treatment condition matches an existing condition category, the corresponding operating stage and the corresponding condition category of the equipment health baseline are invoked; when the current treatment condition is between two or more existing condition categories, the adjacent health baselines are interpolated or the health baseline with the closest parameter distance is selected based on the parameter distance between the current treatment condition and each existing condition category to obtain the equipment health baseline corresponding to the current operating window.

[0076] When the current treatment condition exceeds the applicable range of the existing health baseline, instead of directly using the distant health baseline for anomaly judgment, the current operating data is marked as an unknown condition, and one or more of the following operations are performed: restricted operation, manual confirmation, or supplementary calibration.

[0077] For the s-th operating phase and the p-th type of treatment parameter combination, the device health baseline can be expressed as:

[0078] in This represents the mean vector of each feature at this stage. This represents the covariance matrix or normal fluctuation range. Indicates the range of slope of characteristic change. This represents a cross-module association matrix.

[0079] S4: Multi-source feature extraction and credibility weight calculation Time-domain features, frequency-domain features, trend features, image features, log features, and cross-module consistency features are extracted from the windowed multi-source operational data. Data source reliability is calculated based on data integrity, time synchronization quality, signal stability, device self-test results, and image quality for each data source. The deviation of each data source from the stage health baseline and the degree of inconsistency in the correlation between different data sources are used as anomaly detection features input to step S5, and are not directly used to reduce data source reliability. The reliability score of the x-th data source within the k-th window is calculated. It can be represented as:

[0080] in, Indicates data integrity score, Indicates the signal stability score. This indicates the equipment's self-test score. Indicates the quality of time synchronization. Indicates the quality of image or sensor data. - These are expressed as weighting coefficients. The credibility scores for each data source are normalized to obtain the credibility weights:

[0081] Where N represents the number of data sources, To prevent tiny constants with a denominator of zero.

[0082] S5: Anomaly Detection in a Stage-Conditional Credibility-Gated Two-Branch Temporal Model Based on the multi-source operation characteristics, the credibility weights of each data source, and the current operation stage obtained in step S4, a stage-condition credibility-gated dual-branch time-series anomaly detection model is constructed, which outputs the overall anomaly score of the current time window and the anomaly contribution of each data source.

[0083] For the k-th time window, the feature of the i-th data source extracted in step S4 is represented as: The corresponding credibility weight is represented as By applying the aforementioned confidence weights to the features of each data source, the following results are obtained:

[0084] In the formula, This represents the features of the i-th data source after confidence gating. When a data source has missing data, asynchronous sampling, signal drift, increased noise, or degraded image quality, its corresponding confidence weight is reduced, thereby reducing the impact of the data source on the anomaly detection results.

[0085] Based on the current operational phase, a phase feature vector ms is selected, and the device operation fusion feature is constructed according to the following formula:

[0086] In the formula, Concat represents the feature concatenation operation, n represents the amount of data source, d(k) represents the deviation of each feature in the current window from the corresponding stage health baseline, and e(s) represents the stage code of the current running stage.

[0087] The phase feature selection vector is used to highlight key monitoring features at different operational phases. Specifically, the pre-treatment phase focuses on selecting features such as device self-test status, infrared imaging quality, treatment head distance, power status, and communication status; the intra-treatment phase focuses on selecting features such as treatment energy output commands, energy output unit drive parameters, actual treatment energy output, cooling commands, cooling execution feedback, motion actuator trajectory, treatment area surface temperature, temperature change rate, and communication latency; and the post-treatment phase focuses on selecting features such as treatment energy shutdown status, cooling recovery response, operational data storage status, and the trend of abnormal score changes during the current treatment.

[0088] The stage-conditional confidence-gated dual-branch temporal anomaly detection model includes a data source feature encoding module, a shared temporal encoding module, a reconstruction branch, and a prediction branch.

[0089] The data source feature encoding module encodes the gating features of the energy output link, cooling link, treatment head sensing link, motion execution link, power supply link, and communication and log links to obtain local representations of each data source. The shared temporal encoding module uses a temporal convolutional network with causal convolution and dilated convolution structures to encode the fused features of multiple consecutive time windows, extracting the short-term variation features and long-term trends of the device operating parameters.

[0090] The reconstruction branch reconstructs the multi-source running features within the current time window based on the hidden features output by the shared temporal coding module, and calculates the reconstruction error:

[0091] In the formula, This represents the reconstruction feature of the i-th data source output by the reconstruction branch. This represents the weighted reconstruction error for the current time window.

[0092] The prediction branch predicts the multi-source operating characteristics of the next time window based on the current time window and the time series characteristics before it, and calculates the prediction error after the data of the next time window arrives:

[0093] In the formula, This represents the predicted feature of the i-th data source in the next time window of the prediction branch output. This represents the weighted prediction error for the next time window.

[0094] To identify early anomalies where individual operating parameters have not yet exceeded thresholds but the correlation between different device links has already changed, cross-module correlation consistency error is further calculated based on the current operational phase. .

[0095]

[0096] In the formula, This represents the standardized correlation residuals of modules i and j within the k-th time window. This represents the weight of the associated residual.

[0097] In the pre-treatment phase, the cross-module correlation consistency error includes at least the deviation between the self-test command and the self-test feedback, and the deviation between image acquisition and image quality feedback. During the treatment phase, the cross-module correlation consistency error includes at least the deviation between the energy output power and the energy output unit drive parameters, the deviation between the actual treatment energy output and the temperature rise response of the treatment area, and the deviation between the cooling command and the cooling execution feedback. In the post-treatment phase, the cross-module correlation consistency error includes at least the deviation between the energy output shutdown command and the actual treatment energy output, the deviation between the cooling command and the temperature drop response of the treatment area, and the deviation between the motion actuator return command and the actual return position.

[0098] The overall anomaly score for the current k-th time window Calculate according to the following formula:

[0099] In the formula , and Let represent the coefficients corresponding to the reconstruction error, prediction error, and cross-module correlation consistency error, and satisfy . .

[0100] Abnormality thresholds were established for each of the pre-treatment, during-treatment, and post-treatment phases, respectively. , and Select the corresponding judgment threshold based on the current operating stage. ,when The k-th time window is identified as an abnormal window. The k-th time window is determined as a normal window, and different abnormal judgment thresholds are used for different running stages. A uniform threshold is not used to judge all running stages.

[0101] The stage-conditional confidence-gated two-branch temporal anomaly detection model is trained using confirmed normal historical operating data. During model training, by minimizing the reconstruction error, prediction error, and cross-module correlation consistency error of the normal operating data, the model learns the normal temporal variation patterns and cross-module correlations of the fat reduction equipment under different operating stages and treatment conditions.

[0102] To avoid false alarms caused by momentary fluctuations in a single time window, a general abnormal event is generated when the overall abnormal score continuously exceeds the corresponding stage's abnormal judgment threshold for a preset number of time windows; a safety abnormal event is directly generated when the operating parameters exceed the preset hardware safety threshold or trigger an emergency stop.

[0103] Based on the proportion of each data source in the reconstruction error, prediction error, and cross-module correlation consistency error, calculate the anomaly contribution of the i-th data source in the k-th time window:

[0104] In the formula, , and These represent the reconstruction error, prediction error, and association consistency error corresponding to the i-th data source, respectively. To prevent the denominator from being too small, the overall anomaly score, the anomaly contribution of each data source, the current operating stage, and the multi-source features within the anomaly window are output to step S6 for cross-module fault attribution.

[0105] S6: Cross-module fault attribution based on stage-conditional physical topology graph attention network When the overall anomaly score is output in step S5 Exceeding the anomaly detection threshold corresponding to the current running stage Alternatively, when a security anomaly is detected, the multi-source operational characteristics, confidence weights, various errors, anomaly contribution, and current operational stage within the anomaly time window are input into the cross-module attribution model. This model is a stage-conditional physical topology graph attention network, which outputs the suspected faulty module, fault type, and attribution confidence.

[0106] Based on the structure and function of the fat reduction equipment, the equipment is divided into energy output link nodes, cooling link nodes, treatment head sensing link nodes, motion execution link nodes, power supply link nodes, and communication and logging link nodes.

[0107] The attribution input features of the i-th module node within the k-th time window can be represented as:

[0108] In the formula, This represents the temporal features of the i-th module output by the data source feature encoding module in step S5. This represents the data credibility weight corresponding to the i-th module. This represents the abnormal contribution of the i-th module. , and These represent the reconstruction error, prediction error, and association consistency error corresponding to the i-th module, respectively. This indicates the deviation of the current operational characteristics from the corresponding health baseline. This indicates the current runtime stage code.

[0109] Based on the control and information transmission relationships between device modules, construct a directed device association graph:

[0110] In the formula, Represents a set of module nodes. This represents the set of directed edges enabled during the current operational phase s. In the pre-treatment phase, edges are primarily enabled between power supply, communication, treatment head sensing, and motion execution links to detect anomalies during device self-testing, image acquisition, and treatment head positioning. During the treatment phase, edges are enabled between energy output, cooling, treatment head sensing, motion execution, power supply, and communication links to assess consistency between treatment energy output, cooling execution, motion execution, and the temperature response of the treatment area. In the post-treatment phase, edges related to energy output shutdown, motion execution mechanism repositioning, and communication are primarily enabled to determine the device's finalization process.

[0111] For each directed association edge enabled in the current stage, the attention coefficient between modules is calculated based on the attribution input characteristics of adjacent module nodes, cross-module association residuals, and the current running stage.

[0112] The initial attention value between the i-th module node and the j-th module node is represented as:

[0113] In the formula, Represents the characteristic transformation matrix, This represents the trainable attention parameters. Indicates feature splicing, This represents the correlation residual between the i-th module and the j-th module.

[0114] Normalizing the initial attention value yields:

[0115] In the formula This represents the set of modules that have a valid connection with the i-th module in the current running phase.

[0116] The feature representation of the i-th module node after cross-module information propagation is:

[0117] In the formula and These represent the transformation matrices for the node's own characteristics and the characteristics of its neighboring nodes, respectively. This represents the activation function.

[0118] Through the above processing, fault attribution not only considers the degree of abnormality of a single module itself, but also considers whether other modules that have a control relationship with this module have corresponding abnormalities.

[0119] Based on the graph propagation characteristics of each module node, calculate the probability that each module is a fault source:

[0120] in:

[0121] In the formula, This represents the probability that the i-th module is a suspected faulty module within the k-th time window. This represents the characteristics of the i-th module node after graph information propagation in the k-th time window. and These represent the trainable weight vector and bias corresponding to the module failure probability calculation, respectively. This represents the fault source score of the i-th module in the k-th time window. Indicates the first The fault source score of each module in the k-th time window. Indicates the number of device modules; Furthermore, a corresponding fault type classifier is set for each module, outputting the probability of each fault type under that module:

[0122] In the formula, and Let the weight matrix and bias vector of the fault probability type classifier for the i-th module be represented respectively. This represents the number of fault types corresponding to the i-th module; Softmax is the normalization exponential function. To reduce erroneous attributions that rely solely on neural network outputs, the probability and type of faulty modules are corrected based on preset cross-module anomaly patterns.

[0123] Let the matching degree between the i-th module's q-th type fault and the current anomaly mode be . The corrected attribution probability is then expressed as:

[0124] In the formula, This represents the fusion coefficient between the output of the graph attention model and the result of the abnormal pattern matching.

[0125] The corrected attribution probabilities for all modules and fault types are normalized to obtain the final attribution probabilities:

[0126] In the formula, Indicates the first The number of fault types corresponding to each module It represents a tiny positive number that prevents the denominator from being zero.

[0127] Calculate the failure attribution confidence score based on the final attribution probability:

[0128] When the attribution confidence is lower than the preset confidence threshold, or the difference between the two types with the highest attribution probabilities is less than the preset distinction threshold, a unique fault conclusion will not be output directly. Instead, multiple candidate fault types and probabilities will be output, and a manual review prompt will be generated.

[0129] The cross-module attribution model is trained using historical fault samples confirmed through maintenance or fault reproduction experiments. Each fault sample includes at least the fault occurrence stage, multi-source operational data within the anomaly window, the actual fault module, and the actual fault type. During model training, the actual fault module and actual fault type are used as supervision labels, and the model parameters are optimized using fault module classification loss, fault type classification loss, and topology consistency constraint loss. For fault types where sufficient real fault samples are not yet available, training samples can be obtained through equipment calibration experiments, controlled fault injection experiments, or component accelerated degradation experiments. Figure 2 The diagram shown is a flowchart of the phased baseline and anomaly detection process.

[0130] S7: Predictive Maintenance Window Generation Based on the overall anomaly score, anomaly contribution of each module, reconstruction error, prediction error and cross-module correlation consistency error output in step S5, and the suspected fault module, fault type and final attribution probability output in step S6, a degradation state sequence of each equipment module is constructed. Then, through a staged monotonic degradation-discrete-time survival joint model based on attribution constraints, the probability of each equipment module experiencing a maintenance-required fault, the remaining runnable time and predictive maintenance window are predicted.

[0131] A complete treatment task, a preset runtime, or a preset number of time windows are used as a degradation assessment cycle. For the nth degradation assessment cycle, the abnormal detection results of the pre-treatment, treatment, and post-treatment stages are statistically analyzed to obtain the stage degradation characteristics corresponding to each module.

[0132] For the i-th device module in the n-th degradation evaluation period, construct the module degradation input features:

[0133] In the formula, This indicates the cumulative load characteristics of the module. Indicates the abnormal trend characteristics of the module. Indicates fault attribution characteristics, It indicates the characteristics of historical preservation.

[0134] The cumulative load characteristics of the module are obtained based on one or more of the following: the cumulative running time, cumulative output energy, and cumulative number of task executions of the i-th module in the n-th degradation assessment period; The abnormal trend characteristics of the module are obtained from one or more of the following: the overall abnormal score output in step S5, the abnormal contribution of the i-th module, the reconstruction error, and the cross-module correlation consistency error, the maximum value, the average value, the number of times the corresponding abnormal judgment threshold is exceeded, and the duration within the n-th degradation evaluation period. The fault attribution features are obtained based on one or more of the final attribution probability and attribution confidence corresponding to the i-th module output in step S6. The historical maintenance characteristics are obtained based on one or more of the following: maintenance type, component replacement status, calibration status, and post-maintenance retest results of the i-th module. The degradation characteristics in the pre-treatment, treatment, and post-treatment stages were encoded separately, resulting in:

[0135]

[0136]

[0137]

[0138] In the above formula, , , and This represents the feature encoding units for the pre-treatment, treatment, and post-treatment stages, as well as the combined degradation features. By encoding the degradation features of the three stages separately, it avoids treating pre-treatment self-check abnormalities, treatment execution abnormalities, and post-treatment repositioning or cooling recovery abnormalities as the same degradation pattern.

[0139] The joint degradation features from multiple consecutive degradation evaluation periods are input into a causal temporal convolutional encoder to obtain the historical degradation of the i-th module:

[0140] In the formula, This indicates the number of historical assessment periods used for degradation trend analysis. This represents the temporal degradation characteristics of the i-th module up to the n-th evaluation period.

[0141] The temporal degradation characteristics are constrained based on the fault attribution probability output in step S6. The attribution gating coefficient of the i-th module is expressed as:

[0142] In the formula, This represents the final attribution probability of the i-th module's q-th type of fault in the n-th degradation assessment period.

[0143] The temporal degradation characteristics after the attribution constraint is completed are represented as follows:

[0144] In the formula, Indicates the basic retention factor. This represents the adjustment parameter by which the fault attribution results affect the degradation characteristics.

[0145] Through the attribution constraint, modules that are continuously identified as suspected sources of failure in step S6 receive higher attention in predictive maintenance, while modules that show abnormal responses simply due to the propagation of failures in other modules are prevented from being directly identified as the main maintenance targets.

[0146] Calculate the degradation increment for the current degradation assessment period based on the temporal degradation characteristics after attribution constraints:

[0147] In the formula, This represents the amount of degradation added to the i-th module during the n-th degradation assessment period, and satisfies... .

[0148] When maintenance is not performed:

[0149] After performing maintenance:

[0150] in:

[0151] Input the module’s current degradation state, the temporal degradation characteristics after attribution constraints, the encoding of the future prediction time interval, and the planned usage intensity into the discrete-time survival prediction unit.

[0152] The conditional probability that the i-th module will require maintenance in the r-th prediction interval is expressed as:

[0153] In the formula, This represents the conditional probability that the i-th module will experience a maintenance-required failure in the r-th prediction interval, provided that no maintenance-required failure has occurred previously. This represents the position code of the r-th prediction interval in the future; This indicates the preset future usage intensity, including energy, treatment frequency, or duration of exercise by the motor actuator.

[0154] The probability representation of the i-th module remaining in an unmaintained state for the first r prediction intervals:

[0155] Probability of cumulative failures:

[0156] The estimated remaining runtime of the i-th module is expressed as:

[0157] In the formula, Indicates the maximum number of prediction intervals. This indicates the runtime corresponding to a single prediction interval.

[0158] Before generating the predictive maintenance window, a first maintenance risk threshold is preset. Second maintenance risk threshold And satisfy:

[0159] Determine the starting position of the maintenance window for the i-th module:

[0160] Determine the end position of the maintenance window for the i-th module:

[0161] Predictive maintenance window for the i-th module:

[0162] For samples that have already experienced maintenance failures, the actual evaluation period of the failure is used as the failure time label; for samples that have not yet experienced failures by the end of data collection, they are used as right-censored samples for model training.

[0163] For a sample that has failed, its discrete-time survival loss is expressed as:

[0164] In the formula, This represents the predicted interval of the actual maintenance-required fault that occurs in the i-th module.

[0165] For right-censored samples, the discrete-time survival loss is expressed as:

[0166] In the formula, This indicates the prediction interval for the last confirmed right-censored sample that does not require maintenance.

[0167] The total model loss includes at least the discrete-time survival loss and the degenerate monotonicity constraint loss:

[0168] In the formula, Represents discrete-time survival loss. This represents the monotonicity constraint loss in the degenerate state. This represents the loss due to the consistency constraint of fault attribution. and This represents the corresponding loss weight.

[0169] After maintenance personnel complete the inspection or repair, they obtain the actual faulty module, actual fault type, maintenance operation, component replacement information, pre- and post-repair test results, and post-repair retest results. They then use the maintenance feedback to update the residual degradation coefficient, discrete-time survival prediction model parameters, and maintenance risk thresholds for each module.

[0170] Step S7 outputs the cumulative probability of future failures, estimated remaining uptime, and predictive maintenance window for each device module. For example... Figure 3 The diagram shown is a flowchart for generating cross-module fault attribution and predictive maintenance windows.

[0171] S8: Tiered Handling Strategy and Maintenance Feedback Closed-Loop Optimization Based on the overall anomaly score and the anomaly contribution of each module output in step S5, the suspected faulty modules, fault types and attribution confidence output in step S6, and the future fault risk, remaining uptime and predictive maintenance window output in step S7, and combined with the current pre-treatment, treatment or post-treatment stage, the corresponding equipment handling strategy and maintenance suggestions are generated.

[0172] The treatment strategies include one or more of the following: observation and recording, prompting for review, re-self-testing, recalibration, limiting equipment operation, reducing treatment energy output, enhancing cooling, suspending treatment, shutting off treatment energy output, controlling the motion actuator to stop or retract, and triggering a safety shutdown.

[0173] Based on the severity of the anomaly and its impact on treatment safety, the response strategy is divided into observation level, early warning level, restricted operation level, and safety interlock level.

[0174] When the anomaly score only shows a short-term, slight increase and does not continuously exceed the anomaly judgment threshold for the corresponding stage, or when the confidence level of fault attribution is low, the anomaly is identified as an observation-level anomaly. The system records the corresponding anomaly time, anomaly characteristics, and related equipment modules, without temporarily changing the equipment's operating status, and continues detection in subsequent time windows.

[0175] When the anomaly score continuously exceeds the anomaly judgment threshold for the corresponding stage, but has not yet reached the immediate safety risk condition, the anomaly is identified as a warning-level anomaly. The system outputs anomaly prompts, suspected faulty modules, and suggested inspection items to the operator, and performs re-collection of data, re-self-test, recalibration, or manual review according to the anomaly type.

[0176] When the fault attribution results indicate significant degradation of the equipment module, or when step S7 predicts that the module is nearing its predictive maintenance window, the anomaly is classified as a restricted-operational-level anomaly. Based on the faulty module and fault type, the system generates recommendations such as reducing treatment energy output, shortening single-run time, limiting high-load operation, prohibiting the initiation of new treatment tasks, or scheduling maintenance after the current treatment ends.

[0177] When overvoltage, overcurrent, overtemperature, treatment head distance exceeding the safe range, motion actuator collision, critical communication interruption, emergency stop triggering, or other safety interlock events occur, the anomaly is identified as a safety interlock level anomaly. Safety interlock level anomalies do not rely on the model judgment results of steps S5 to S7; they directly shut down the treatment energy output, maintain or enhance cooling according to the anomaly type, control the motion actuator to stop movement or withdraw from the treatment area, and simultaneously prevent the equipment from automatically resuming operation, awaiting manual inspection and confirmation.

[0178] During the pre-treatment phase, if any abnormalities are detected in the device's self-test, sensors, treatment head distance, or critical communication, the device will be prohibited from entering the treatment phase. The operator will be prompted to re-test, reposition, recalibrate, or check the corresponding device modules. Treatment will only begin after the abnormalities are resolved and the device meets the treatment start-up conditions again.

[0179] During treatment, if any abnormalities are detected in energy output, cooling execution, motion of the motion actuator, or temperature monitoring, corresponding measures will be taken according to the severity of the abnormality. For minor abnormalities, the treatment energy output can be reduced, cooling can be enhanced, or the operator can be prompted to verify the abnormality. For abnormalities that may affect treatment safety but have not yet triggered the hardware safety interlock, the energy output will be suspended and cooling will be maintained, pending system retesting or manual confirmation. For abnormalities that meet the safety interlock conditions, a safety shutdown will be immediately executed.

[0180] When the confidence level of fault attribution is lower than the preset requirement, or when the attribution probabilities of multiple fault types are close, the system does not directly perform irreversible component replacement or repair operations. Instead, it outputs multiple candidate faults and their corresponding inspection sequences. When treatment safety is involved, a conservative approach is taken based on the candidate fault with the higher risk.

[0181] In the post-treatment phase, maintenance recommendations are generated based on the abnormal events recorded during the treatment, the fault attribution results, and the predictive maintenance window. These recommendations include continued observation, inspection of the cooling unit, inspection of the power supply and communication module, and replacement of one or more corresponding components.

[0182] When multiple modules exhibit anomalies simultaneously, prioritize processing the modules with higher impact on treatment safety, higher confidence in fault attribution, or those expected to enter the maintenance window earlier. The maintenance recommendation should include the suggested maintenance time, the suggested maintenance target, the primary basis for the anomaly, and the estimated remaining uptime.

[0183] After the equipment completes inspection, calibration, repair, or component replacement, maintenance feedback information is obtained. This feedback information includes the actual faulty module, the actual fault type, whether it was a false alarm, the maintenance operation performed, information on replaced components, pre- and post-maintenance test results, and post-repair retest results.

[0184] Based on the maintenance feedback, the equipment health baseline, data source reliability, anomaly detection model, cross-module attribution model, and phased monotonic degradation-discrete-time survival joint model are updated.

[0185] When maintenance results confirm that an anomaly is a real fault, the corresponding anomaly time window, fault module, and fault type are used as confirmed fault samples to improve the ability to identify and attribute the same anomaly pattern.

[0186] When maintenance results confirm that an anomaly is a false alarm, a patient being moved, an environmental change, or a data quality anomaly, the corresponding record will be used as a non-equipment fault sample to reduce the probability of false alarms in similar situations.

[0187] After the equipment has been calibrated, repaired, or had its components replaced, the health status of the corresponding module is re-established based on the retest data after the repair, and the degradation status and remaining runnable time prediction results in step S7 are corrected.

[0188] Equipment health baselines and model parameters should be updated only after maintenance results have been confirmed. The updated health baselines and model parameters should be put into operation only after offline verification or authorized confirmation, and the previous model version should be retained for rollback in case of performance anomalies after the update. Hardware safety thresholds such as overvoltage, overcurrent, overtemperature, mechanical impact, and emergency stop are not automatically adjusted based on model output; they can only be modified according to equipment rated parameters, safety specifications, or authorized manual settings.

[0189] Through the above-mentioned graded handling and maintenance feedback loop, the equipment can take different handling methods according to the stage of anomaly occurrence, the severity of the anomaly, the fault attribution results, and the future fault risk, and continuously correct the anomaly detection, fault attribution, and predictive maintenance results using real maintenance results.

[0190] Example 2 Embodiment 2 of the present invention provides an anomaly detection and fault prediction method based on the reliable fusion of multi-source operating data of a weight loss device, specifically including: Step 1: Data Acquisition of Weight Loss Equipment and Multi-Source Operation The fat reduction device is a non-contact laser fat reduction robot. The device includes a laser energy output unit, a cooling unit, a treatment head sensing unit, a robotic arm motion actuator, a power management unit, and a communication control unit.

[0191] The laser energy output unit is used to output therapeutic laser according to control commands and record laser power commands, laser drive current, actual laser output power, output duration and output power change rate.

[0192] The cold air cooling unit is used to cool the target treatment area and records the cooling level command, fan drive current, actual cold air velocity, cooling response delay, and cooling rate of the treatment area.

[0193] The treatment head sensing unit includes an infrared thermal imager, a visible light camera, and a distance sensor, used to acquire surface temperature of the treatment area, temperature change rate, infrared temperature image, image clarity, image brightness stability, and treatment head distance.

[0194] The robotic arm motion actuator is used to control the position, orientation, and motion path of the treatment head relative to the treatment area, and to record the target position, actual position, target orientation, actual orientation, end-effector position error, end-effector orientation error, motion speed, joint current, joint temperature, and tracking hysteresis.

[0195] The power management unit records supply voltage, supply current, output power, power fluctuations, overvoltage records, and overcurrent records. The communication control unit records communication delay, data packet loss rate, data retransmission rate, command confirmation time, and communication interruption records.

[0196] In this embodiment, laser control data, cooling control data, and power supply data are collected according to a preset cycle; infrared thermal images are acquired according to their image acquisition frequency; and robotic arm control data is acquired according to the robotic arm controller's original sampling frequency. Each data set is appended with a timestamp corresponding to a unified device clock.

[0197] Step 2: Identification during the runtime phase Establish a treatment task identifier for each treatment task and generate a treatment task status flag.

[0198] When the device has created the treatment task but has not yet received the formal instruction to start treatment, the current stage is marked as the pre-treatment stage. The pre-treatment stage includes device self-test, treatment area acquisition, path generation, treatment parameter setting, and treatment head positioning.

[0199] When a formal instruction to begin treatment is received, and no completion, cancellation, or abnormal termination event has been received for the current treatment task, the current stage is marked as the treatment in progress stage. During treatment, events such as switching between treatment loops, brief cessation of energy output, repositioning of the treatment head, or normal pauses do not change the treatment in progress stage marker.

[0200] When a treatment completion, manual termination, manual cancellation, or abnormal termination event is received, the current stage is marked as the post-treatment stage. The post-treatment stage includes stopping laser output, continuing cooling, the robotic arm exiting the treatment area and returning to its original position, saving operational data, and generating maintenance assessment results.

[0201] The stage markers are verified based on the laser enable status, cooling unit status, robotic arm task execution status, and safety interlock status. When the control status indicates that the treatment is in progress, but the actual execution feedback of the laser, cooling, or robotic arm is inconsistent with the control status and continues for more than a preset time, a stage execution inconsistency anomaly is generated.

[0202] Step 3: Data preprocessing and time window construction First, time synchronization is performed on all data sources. For continuous data with different sampling frequencies, resampling and interpolation are used to convert it to a unified time axis; for image data, image quality features and temperature field features corresponding to the unified time window are extracted; for log data, it is converted into event type, event quantity, event duration, and event level.

[0203] In one implementation, a time window with a length of 10 seconds and a sliding step size of 1 second is used. Within each time window, laser link characteristics, cooling link characteristics, treatment head sensing link characteristics, robotic arm motion link characteristics, power link characteristics, and communication and log link characteristics are formed respectively.

[0204] Data exceeding the sensor's physical measurement range, communication padding values, duplicate data, and timestamp errors are flagged as acquisition anomalies, and the original data is retained. Deviation data that may be caused by equipment degradation or component failure is not deleted during the preprocessing stage.

[0205] Step 4: Phase – Establishment of Operating Condition Health Baseline Historical treatment tasks that passed equipment self-tests, did not trigger alarms or safety interlocks, did not exhibit maintenance-confirmed faults, and whose operating results were confirmed to be normal were selected as the health baseline sample.

[0206] The healthy baseline samples were initially categorized into pre-treatment, treatment, and post-treatment stages. During the treatment stage, the treatment conditions were further classified according to preset parameter ranges, including laser energy output, cooling level, robotic arm movement speed, treatment head distance, and treatment area area.

[0207] For example, laser energy output can be divided into low, medium, and high ranges; cooling intensity can be divided into low, medium, and high ranges; and robotic arm movement speed can be divided into low, medium, and high ranges. Corresponding operating condition categories should be established based on the actual equipment's calibrated and used parameter combinations; it is not required to establish separate health baselines for all theoretical parameter combinations.

[0208] For each operating stage and condition category, the normal mean, normal fluctuation range, normal trend of each characteristic, as well as the normal correlation between control commands, equipment execution results, and treatment area response are statistically analyzed to form the corresponding stage-condition health baseline.

[0209] For the current k-th time window, the system first reads the running stage marker to obtain the current running stage; then it determines the current working condition category based on the current treatment parameters, and uses the running stage and working condition category as a joint index to call the corresponding health baseline.

[0210] When the current treatment parameter matches an existing operating condition category, the corresponding health baseline is directly invoked; when the current treatment parameter is between two existing operating condition categories, the nearest baseline is selected based on the parameter distance or the adjacent baseline is interpolated; when the current treatment parameter exceeds the range of existing baselines, it is marked as an unknown operating condition, and an abnormal conclusion is not drawn directly based on a health baseline that is far away.

[0211] Step 5: Data Reliability Calculation Calculate a confidence gating value between 0 and 1 independently for each data source.

[0212] Data integrity is determined based on the ratio between the actual amount of data received and the theoretical amount of data within the current time window. Time synchronization quality is determined based on data timestamp deviation and sampling interval stability. Signal stability is determined based on noise level, continuous repetition values, sensor saturation, and short-term jumps. Device self-test score is determined based on the self-test results of the corresponding sensor or device module. Image quality score is determined based on image sharpness, noise, frame loss rate, and the effective pixel ratio of the temperature field.

[0213] The degree of deviation of each data source from the healthy baseline, as well as the degree of inconsistency in the correlation between different device modules, are not used to reduce the credibility of the data source, but rather as anomaly detection features input into subsequent models.

[0214] Step 6: Stage Conditional Two-Branch Timing Anomaly Detection The operational characteristics of each data source are multiplied by the corresponding confidence gating value to obtain the gated data source characteristics. Then, the gating characteristics, the current stage encoding, and the deviation characteristics relative to the healthy baseline are concatenated to form the fused characteristics of the current time window.

[0215] The shared temporal coding module in this embodiment employs a temporal convolutional network. This temporal convolutional network includes multiple layers of causal convolutions and dilated convolutions, used to encode fused features from multiple consecutive time windows.

[0216] The shared encoding results are input into the reconstruction branch and the prediction branch, respectively. The reconstruction branch is used to reconstruct the multi-source running features within the current time window, and the prediction branch is used to predict the multi-source running features within the next time window.

[0217] For each cross-module relationship existing in the current stage, the expected response of the downstream module is predicted based on the control instructions, execution feedback and current treatment status of the upstream module, and the standardized correlation residual between the actual response and the expected response is calculated.

[0218] For example, during the treatment phase, the actual laser output power is predicted based on the laser power command and laser drive current; the temperature rise response of the treatment area is predicted based on the actual laser output power, robotic arm movement speed, treatment head distance, and cooling status; and the actual cold air output and cooling response are predicted based on the cooling command and fan drive current.

[0219] The overall anomaly score is obtained by weighting and combining the reconstruction error, prediction error, and cross-module correlation consistency error.

[0220] The stage-conditional confidence-gated two-branch temporal anomaly detection model is trained using confirmed normal historical operating data. After training, normal validation data from the pre-treatment, during-treatment, and post-treatment stages are input into the model, and the preset high quantile of the anomaly score of the normal validation data in each stage is used as the anomaly judgment threshold for the corresponding stage.

[0221] When the overall anomaly score of the current time window exceeds the anomaly determination threshold for the current stage, the time window is marked as an anomaly window. When the overall anomaly score exceeds the threshold for a preset number of consecutive time windows, a general anomaly event is generated.

[0222] Step 7: Cross-module fault attribution When a general abnormal event is detected, the laser link, cooling link, treatment head sensing link, robotic arm motion link, power link, and communication and log link are respectively treated as nodes in the device association diagram.

[0223] Directed association edges are set up based on the actual control relationship, energy transfer relationship, and information transfer relationship of the equipment. The power supply link is connected to the laser link, cooling link, and robotic arm motion link respectively; the communication link is connected to each execution link respectively; the laser link, cooling link, and robotic arm motion link are connected to the treatment head sensing link respectively.

[0224] By combining the temporal characteristics, confidence gating values, anomaly contribution, reconstruction error, prediction error, correlation consistency error, and stage-encoded input graph attention network of each module, the probability of each module acting as a fault source and the probability corresponding to each fault type are calculated.

[0225] This embodiment also sets a preset fault mode for correcting the model output.

[0226] When the laser drive current remains normal, but the actual laser output power and the temperature rise response of the treatment area decrease simultaneously, the attribution probability of laser output attenuation, optical path transmission loss, or optical window contamination is increased.

[0227] When the cooling command and fan drive current remain normal, but the actual cold air output and the cooling rate of the treatment area decrease simultaneously, the attribution probability of cooling channel blockage or fan efficiency reduction increases.

[0228] When multiple execution modules experience instruction response delays simultaneously, and no corresponding anomalies are reported in the local operation feedback of each module, the probability of attributing the communication link anomaly to the cause is increased.

[0229] When the infrared temperature image shows a positional shift, but the laser, cooling, and robotic arm are operating normally, the anomaly is preferentially marked as a movement of the treatment object, reducing the probability of attributing the fault to equipment component failure.

[0230] The fault attribution model outputs suspected fault modules, candidate fault types, and attribution confidence scores. When the attribution confidence score is lower than a preset requirement, it outputs multiple candidate fault types and suggested inspection order, but does not output a unique fault conclusion.

[0231] Step 8: Predictive Maintenance Window Generation A single complete treatment task is used as a degradation assessment cycle. For each device module, the overall anomaly score, module anomaly contribution, reconstruction error, prediction error, correlation consistency error, fault attribution probability, cumulative operating load, and historical maintenance records from multiple recent treatment tasks are statistically analyzed to form a module degradation characteristic sequence.

[0232] The cumulative load of the laser link includes the cumulative laser output time and cumulative output energy; the cumulative load of the cooling link includes the cumulative cooling operation time and the number of fan start-stop cycles; and the cumulative load of the robotic arm motion link includes the cumulative motion mileage and the cumulative number of path executions.

[0233] The degradation features of modules from multiple consecutive degradation assessment cycles are input into a causal temporal convolutional encoder to calculate the degradation increment of each module. When no maintenance is performed, the degradation increment is accumulated to the degradation state of the previous cycle; after cleaning, calibration, repair or component replacement is completed, the residual degradation state of the module is corrected according to the maintenance type and the results of the retest after repair.

[0234] Input the module’s current degradation state, temporal degradation characteristics and expected future usage intensity into the discrete-time survival prediction unit to calculate the conditional probability of maintenance-required failures occurring in each prediction interval, the cumulative probability of future failures and the expected remaining runnable time.

[0235] When the cumulative probability of future failures reaches the first maintenance risk threshold, the starting point of the predictive maintenance window is determined; when the cumulative probability of future failures reaches the second maintenance risk threshold, the ending point of the predictive maintenance window is determined.

[0236] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for anomaly detection and fault prediction based on reliable fusion of multi-source operational data from a weight-loss device, comprising: Step 1: Acquire multi-source operational data of the fat reduction device and attach operational stage markers, including pre-treatment stage, during-treatment stage, and post-treatment stage; Step 2: Perform time synchronization, data cleaning, and windowing processing on multi-source operational data; Step 3: Based on the operating stage, equipment configuration, treatment parameters, and historical normal operation data of the fat reduction equipment, establish health baselines for different operating stages; Step 4: Extract features from the multi-source running data after windowing processing in Step 2, and calculate the credibility weight of each data source; Step 5: Based on the credibility weights of each data source obtained in Step 4, the multi-source running features are gating processed, and then combined with the running stages to obtain fused features. The fused features are input into the pre-trained stage condition credibility gating dual-branch temporal anomaly detection model, and the overall anomaly score and anomaly contribution of each data source in the current time window are output. Step 6: If the overall anomaly score of the current time window exceeds the anomaly judgment threshold corresponding to the current running stage or a safety anomaly event is detected, the pre-trained cross-module attribution model is introduced to output the suspected fault module, fault type and attribution confidence. Step 7: According to the preset degradation assessment cycle, based on the output of Step 5 and / or the output of Step 6, construct the degradation state sequence of each module of the fat reduction equipment, and generate a predictive maintenance window through a staged monotonic degradation-discrete-time survival joint model based on attribution constraints. Step 8: Output the tiered handling strategy.

2. The method of claim 1, wherein the method is characterized by, The multi-source operational data in step 1 includes at least two of the following: energy output link data, cooling link data, treatment head sensing link data, motion execution link data, power supply link data, communication link data, and system log data.

3. The method of claim 1, wherein the method further comprises: The health baseline in step 3 includes at least the normal range of features, the trend of feature changes, and the normal correlation between modules. The normal correlation between modules is used to describe the normal correspondence between control commands, actual execution results of the device, and response of the treatment area. For the s-th operational phase and the p-th combination of treatment parameters, the equipment health baseline Represented as: in, This represents the mean vector of each feature at this stage. This represents the covariance matrix or normal fluctuation range. Indicates the range of slope of characteristic change. This represents a cross-module association matrix.

4. The anomaly detection and fault prediction method based on reliable fusion of multi-source operating data of weight loss equipment according to claim 1, characterized in that, Step 4 includes: Extract time-domain features, frequency-domain features, trend features, image features, log features, and cross-module consistency features from the windowed multi-source runtime data; The credibility score of the x-th data source within the k-th window is calculated using the following formula. : in, Indicates data integrity score, Indicates the signal stability score. This indicates the equipment's self-test score. Indicates the quality of time synchronization. Indicates the quality of image or sensor data. ~ This represents the corresponding weight coefficient, which is used to normalize the credibility scores of each data source to obtain the credibility weight. : Where N represents the number of data sources, To prevent tiny constants with a denominator of zero.

5. The anomaly detection and fault prediction method based on reliable fusion of multi-source operating data of weight loss equipment according to claim 1, characterized in that, The stage condition confidence-gated dual-branch temporal anomaly detection model in step 5 includes: a data source feature encoding module, a shared temporal encoding module, a reconstruction branch, a prediction branch, an association consistency calculation module, and an output module; wherein, the data source feature encoding module and the shared temporal encoding module are connected in series and then connected in parallel to the reconstruction branch, the prediction branch, and the association consistency calculation module, and finally the output module is connected in series. The data source feature encoding module is used to encode the multi-source operating features respectively to obtain the local representation of each data source; The shared temporal coding module employs a temporal convolutional network with causal convolution and dilated convolution structures to encode the fused features of multiple consecutive time windows, extracting the short-term variation features and long-term variation trends of equipment operating parameters. The reconstruction branch is used to reconstruct the multi-source running features within the current time window based on the hidden features output by the shared temporal coding module, and to calculate the reconstruction error. The prediction branch is used to predict the multi-source operation characteristics of the next time window based on the time series characteristics of the current time window and those before it, and to calculate the prediction error after the data of the next time window arrives. The correlation consistency calculation module is used to calculate the cross-module correlation consistency error based on the standardized correlation residuals and weights of different modules of the weight loss equipment within the current time window. The output module is used to obtain the overall anomaly score of the current time window by weighted summation based on the reconstruction error, prediction error, and cross-module correlation consistency error; it is also used to calculate the anomaly contribution of each data source based on the proportion of each data source in the reconstruction error, prediction error, and cross-module correlation consistency error.

6. The anomaly detection and fault prediction method based on reliable fusion of multi-source operating data of weight loss equipment according to claim 1, characterized in that, The cross-module attribution model in step 6 is a stage-conditional physical topology graph attention network; the processing includes: Construct a directed device association graph based on the control and information transmission relationships between modules. : In the formula, Represents a set of module nodes. This represents the set of directed associated edges enabled in the current running phase s; For each directed association edge activated in the current stage, the attention coefficient between modules is calculated based on the attribution input features of adjacent module nodes, cross-module association residuals, and the current running stage, thereby obtaining the features of each module after cross-module information propagation. Based on the graph propagation characteristics of each module node, calculate the probability that the i-th module is a suspected faulty module within the k-th time window. for: In the formula, This represents the characteristics of the i-th module after the graph information propagation in the k-th time window. and These represent the trainable weight vector and bias corresponding to the module failure probability calculation, respectively. This represents the fault source score of the i-th module in the k-th time window. Indicates the first The fault source score of each module in the k-th time window. Indicates the number of device modules; For each module of the weight loss device, set a corresponding fault type classifier and output the probability of the q-th fault type of the i-th module. : In the formula, and Let the weight matrix and bias vector of the fault probability type classifier for the i-th module be represented respectively. This represents the number of fault types corresponding to the i-th module; Softmax is the normalization exponential function. Let the matching degree between the i-th module's q-th type fault and the current anomaly mode be . Then the corrected attribution probability for: In the formula, The fusion coefficient between the attention network output and the abnormal pattern matching result of the stage conditional physical topology graph is represented. The corrected attribution probabilities for all modules and fault types are normalized to obtain the final attribution probabilities. : In the formula, Indicates the first The number of fault types corresponding to each module This indicates the prevention of small positive numbers with a denominator of zero. Indicates the summation index of the fault type; The failure attribution confidence level is obtained according to the following formula. : 。 7. The anomaly detection and fault prediction method based on reliable fusion of multi-source operating data of weight loss equipment according to claim 6, characterized in that, Step 6 also includes: When the calculated fault attribution confidence is lower than the preset confidence threshold, or the difference between the two fault types with the highest attribution probabilities is less than the preset distinction threshold, multiple candidate fault types and probabilities are output, and a manual review prompt is generated.

8. The anomaly detection and fault prediction method based on reliable fusion of multi-source operating data of weight loss equipment according to claim 1, characterized in that, The degradation state sequence of each module of the fat reduction device in step 7 specifically includes: For the i-th module of the weight loss device, during the n-th degradation evaluation period, construct the module degradation input features. : In the formula, This indicates the cumulative load characteristics of the module. Indicates the abnormal trend characteristics of the module. Indicates fault attribution characteristics, Indicates historical preservation characteristics; The degradation features were encoded separately for the pre-treatment, treatment, and post-treatment stages, and then concatenated to obtain the joint degradation features. ; The joint degradation features from multiple consecutive degradation evaluation periods are input into a causal temporal convolutional encoder to obtain the temporal degradation features of the i-th module in the n-th evaluation period. ; Degraded input features and temporal degradation characteristics The degradation state characteristics corresponding to this assessment period are obtained by splicing them together. Multiple consecutive degradation state features are arranged in the order of the evaluation cycle to obtain a degradation state sequence.

9. The anomaly detection and fault prediction method based on reliable fusion of multi-source operating data of fat reduction equipment according to claim 8, characterized in that, The attribution constraints in step 7 specifically include: Use the final attribution probability obtained in step 6 Temporal degradation characteristics Constraints are applied to the attribution gating coefficient of the i-th module. Represented as: Temporal degradation characteristics after attribution constraints Represented as: In the formula, Indicates the basic retention factor. This represents the adjustment parameter by which the fault attribution results affect the degradation characteristics.

10. The anomaly detection and fault prediction method based on reliable fusion of multi-source operating data of fat reduction equipment according to claim 9, characterized in that, The processing steps of the phased monotonically degenerate-discrete-time survival joint model in step 7 include: Based on the temporal degradation characteristics after attribution constraints Calculate the degradation increment for the current degradation assessment period. for: In the formula, , and These represent the trainable weights and biases corresponding to the degradation increment calculation of the i-th module, respectively. Represents a non-negative activation function; When maintenance is not performed: After performing maintenance: In the formula, This represents the degradation state of the i-th module at the end of the (n-1)-th degradation evaluation period. This represents the degradation state of the i-th module at the end of the n-th degradation evaluation period. Let represent the residual degradation coefficient of the i-th module after maintenance is performed in the n-th degradation assessment cycle, and satisfy . ; The conditional probability that the i-th module will experience a maintenance failure in the r-th prediction interval in the future. Represented as: In the formula, This represents the conditional probability that the i-th module will experience a maintenance-required failure in the r-th prediction interval, provided that no maintenance-required failure has occurred previously. This indicates the temporal degradation characteristics after the attribution constraint ends; Indicates the current degradation state; This represents the position code of the r-th prediction interval in the future; Indicates the preset planned usage intensity; and These represent the trainable weights and biases of the survival prediction output layer, respectively. The probability that the i-th module will remain in a maintenance-free state for the first r prediction intervals. for: Probability of cumulative failures : The estimated remaining run time of the i-th module Represented as: In the formula, v Indicates the index of the prediction interval for product calculation. Indicates the maximum number of prediction intervals. Indicates the runtime corresponding to a single prediction interval; Before generating the predictive maintenance window, a first maintenance risk threshold is preset. Second maintenance risk threshold And satisfy: Determine the starting position of the maintenance window for the i-th module. for: Determine the end position of the maintenance window for the i-th module. for: Thus, the predictive maintenance window of the i-th module is obtained. for: in, This represents the baseline time corresponding to the end of the nth degradation assessment period. This indicates the time length corresponding to a single prediction interval.

Images