Massage instrument operation state monitoring method and system based on deep learning
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGZHOU QILING INFORMATION TECH CO LTD
- Filing Date
- 2025-10-09
- Publication Date
- 2026-06-26
AI Technical Summary
Traditional massager status monitoring technology suffers from problems such as limited coverage of single-parameter monitoring, weak anti-interference ability, simple data preprocessing, low model training efficiency and easy overfitting, and insufficient anomaly localization accuracy, resulting in missed detection and high maintenance costs in complex fault scenarios.
Data is collected from multiple sources of sensors. Through sliding window denoising and multi-source parameter collaborative standardization, combined with timestamp time-series alignment technology, a time-series attention fusion model based on Transformer architecture is constructed. The model is trained using a coupling loss function and Adam optimizer, and an early stopping mechanism is introduced to output abnormal states and locate specific faults.
It improves the accuracy and reliability of massager operation status assessment, especially in the case of combined fault scenarios of sudden temperature rise and abnormal vibration, it can accurately identify fault types and generate structured reports, thereby reducing maintenance costs.
Smart Images
Figure CN121256569B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent monitoring technology, specifically to a method and system for monitoring the operating status of a massager based on deep learning. Background Technology
[0002] With the popularization of health consumption concepts, the demand for massage devices as home health care equipment continues to grow. Their core function is to simulate the effect of manual massage to relieve muscle fatigue through mechanical vibration, pressure application, and temperature control. However, massage devices are prone to malfunctions during long-term operation due to problems such as motor overload, component wear, or poor heat dissipation. These problems can range from affecting the user experience to causing equipment damage or even safety hazards. Therefore, real-time monitoring of the operating status of massage devices and early identification of potential faults are crucial to ensuring the reliability of the equipment.
[0003] However, traditional massager status monitoring technology mainly relies on single-parameter threshold alarms, such as judging whether the device is abnormal by independent data from pressure or temperature sensors. This method has significant limitations: First, single-parameter monitoring cannot cover complex fault scenarios. For example, motor overload may simultaneously cause temperature rise and abnormal vibration, but traditional methods can only detect a single parameter exceeding the standard, easily missing complex faults. Second, data preprocessing is simple and has weak anti-interference ability. Raw sensor data often contains noise, missing values, or time misalignment. Traditional moving average or filtering methods are difficult to adapt to the differences in the scale of multi-source data, leading to distorted analysis results. Third, the constructed model has low training efficiency and is prone to overfitting. Traditional LSTM or CNN models require a lot of computing resources and have a slow convergence speed when processing long time series data. The lack of an early stopping mechanism in the training process easily leads to poor model performance on the test set. Fourth, the anomaly localization accuracy is insufficient. Traditional methods can only output "whether it is abnormal" and cannot locate specific fault parameters (such as excessive pressure or abnormal temperature), increasing maintenance costs.
[0004] Therefore, a method and system for monitoring the operating status of massage devices based on deep learning were developed. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for monitoring the operating status of massagers based on deep learning. This invention constructs a multi-parameter real-time monitoring system for the operating status of massagers by deploying multi-source sensors, breaking through the limitations of traditional single-parameter monitoring, which has limited coverage and weak anti-interference ability. In the data preprocessing stage, sliding window denoising and multi-source parameter collaborative standardization formulas are used, combined with timestamp time sequence alignment technology, to effectively solve the analysis bias caused by differences in dimensions, noise interference, and time misalignment of multi-source data. The contribution of each parameter to the device status is dynamically quantified, avoiding evaluation distortion caused by differences in dimensions between parameters. Furthermore, the data analysis capability is improved by constructing a time sequence attention fusion model, especially in the case of combined fault scenarios of sudden temperature rise and abnormal vibration, which improves the accuracy and reliability of device operating status assessment.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On the one hand, a deep learning-based method for monitoring the operating status of a massage device, the specific steps of which are as follows:
[0007] Data acquisition steps: Multi-source sensors are deployed on the massager to collect raw data including pressure data, motor current data, vibration frequency data and temperature data in real time during the operation of the massager, and transmit them to the data preprocessing step;
[0008] Data preprocessing steps: Receive raw data, clean the raw data by denoising with a sliding window, filling in missing values and removing outliers, then standardize the cleaned data using a multi-source parameter collaborative standardization formula, and align the data according to timestamps to form standardized time-series data;
[0009] Model building and training steps: Build a temporal attention fusion model based on the Transformer architecture, divide the standardized time series dataset into training and test sets according to the proportion, train by coupling the loss function formula and the optimizer, and introduce an early stopping mechanism to optimize and store the parameters;
[0010] Real-time monitoring and anomaly identification steps: Receive standardized time-series data and input it into the time-series attention fusion model, output the running status evaluation result, compare the evaluation result with the preset threshold, trigger an alarm when an abnormal state occurs, locate abnormal parameters by combining attention weights and send abnormal information.
[0011] Furthermore, in the data acquisition step, the multi-source sensors include a piezoresistive pressure sensor, a Hall current sensor, a piezoelectric vibration sensor, and a thermocouple temperature sensor. Specifically, the piezoresistive pressure sensor is arranged on the surface of the massage head of the massager to collect pressure data and corresponding timestamps during massager operation; the Hall current sensor is arranged in the power supply circuit of the massager motor to collect motor current data and corresponding timestamps during massager operation; the piezoelectric vibration sensor is arranged on the vibration motor housing of the massager to collect vibration frequency data and corresponding timestamps during massager operation; and the thermocouple temperature sensors are respectively arranged at the motor windings and housing of the massager to collect temperature data and corresponding timestamps during massager operation.
[0012] Furthermore, in the data preprocessing step, the cleaned data is standardized using a multi-source parameter collaborative standardization formula, which is as follows: ,in, For the first The standardized results of each parameter For the first The original values of the parameters, The first The mean and standard deviation of the original values of each parameter were obtained from historically collected data. For the first The self-weighting coefficients of each parameter are fixed values preset based on the degree of influence of the parameter on the equipment status. This represents the total number of parameters, with a value of 4, corresponding to pressure, current, vibration frequency, and temperature. For the first The original values of the parameters, and , , The first The mean and standard deviation of the original values of each parameter. For the first With the The correlation coefficient of each parameter.
[0013] Furthermore, in the model construction and training steps, a Transformer architecture temporal attention fusion model is constructed, comprising an input embedding layer, a multi-head self-attention layer, an encoder layer, a feature fusion layer, and an output layer. The input embedding layer converts standardized temporal data into 256-1024 dimensional feature vectors through linear transformation, while adding positional encoding information to preserve temporal features. The multi-head self-attention layer sets 4-16 parallel attention heads, each of which captures the correlation weights of different parameters in the temporal dimension through a multi-source temporal attention weight formula, and then fuses multi-view features through concatenation and linear transformation. The encoder layer contains 4-8 stacked encoder units to perform deep extraction of correlated features. The feature fusion layer dynamically assigns weights to each parameter feature based on attention weights. The output layer consists of 1-3 fully connected layers and a softmax activation function, outputting three types of operational state evaluation results: normal, mildly abnormal, or severely abnormal.
[0014] Furthermore, in the model construction and training steps, the standardized time-series dataset is divided into training and test sets in a ratio of 6:4 to 8:2. During the training phase, the loss is calculated by coupling the loss function formula, and the model parameters are iteratively optimized by the Adam optimizer. At the same time, an early stopping mechanism is introduced, which automatically stops training when the test set loss does not decrease for 3-8 consecutive rounds. After training is completed, the performance of the time-series attention fusion model is verified by the test set, and the optimized time-series attention fusion model parameters are stored.
[0015] Furthermore, in the model construction and training steps, each attention head captures the correlation weights of different parameters in the temporal dimension through a multi-source temporal attention weight formula, which is as follows: ,in, For the first Time of the first Attention weights for each parameter, For the first Dimensional weights of each parameter, For the first The temporal position weight of each moment. These are the weight coefficients for the parameter dimensions. For the total number of parameters, This is the total timing length. , These are the parameter index and the time series index, respectively. For the first The time decay factor at time t is calculated using the formula: , The time series indicates the specific moment when the attention weights are being calculated. For the first The time decay factor at time.
[0016] Furthermore, in the model construction and training steps, the loss is calculated during the training phase using a coupling loss function formula, which is: ,in, This represents the total loss value. For traditional cross-entropy loss, The coupling error coefficient is... The correlation coefficients for the predicted i-th and j-th parameters are... The correlation coefficients of the i-th and j-th parameters in the actual data. This represents the total number of parameters.
[0017] Furthermore, in the real-time monitoring and anomaly identification step, standardized time-series data is received, a time-series attention fusion model is invoked for inference calculation, and three types of operating status evaluation results (normal, slightly abnormal, or severely abnormal) and attention weight values of each parameter are output. When a slightly abnormal or severely abnormal evaluation result is output, the actual values of the single parameter (pressure, motor current, vibration frequency, and temperature) at the corresponding moment in the evaluation result are compared with the threshold values of each parameter in the preset threshold library to determine whether a single parameter exceeds the threshold. The threshold values of each parameter in the preset threshold library are set by the massager at the factory. When the evaluation result is consistent with the result of exceeding the threshold of a single parameter, an LED alarm light is triggered to flash at a frequency of 0.5-2Hz and a buzzer is triggered to sound an intermittent alarm. The abnormal parameter is located by combining the attention weight values of each parameter, and an anomaly information report containing the anomaly type, abnormal parameter, occurrence time, and current parameter value is generated and sent to the user terminal.
[0018] On the other hand, a deep learning-based massage device operation status monitoring system includes:
[0019] Data acquisition module: Multi-source sensors are installed on the massager to collect raw data, including pressure data, motor current data, vibration frequency data and temperature data, in real time during the operation of the massager, and then transmit them.
[0020] Data preprocessing module: Receives raw data, cleans the raw data by denoising, filling in missing values and removing outliers using a sliding window, then standardizes the cleaned data using a multi-source parameter collaborative standardization formula, and aligns the data according to timestamps to form standardized time-series data;
[0021] Model building and training module: Construct a temporal attention fusion model based on the Transformer architecture, divide the standardized time series dataset into training and test sets according to the proportion, train it by coupling the loss function formula and the optimizer, and introduce an early stopping mechanism to optimize and store the parameters.
[0022] Real-time monitoring and anomaly identification module: Receives standardized time-series data and inputs it into the time-series attention fusion model, outputs the operation status evaluation result, compares the evaluation result with the preset threshold, triggers an alarm when an abnormal state occurs, and locates abnormal parameters by combining attention weights and sends abnormal information.
[0023] Compared with existing technologies, this deep learning-based method and system for monitoring the operating status of massage devices has the following advantages:
[0024] I. This invention constructs a multi-parameter real-time monitoring system for the operating status of a massager by deploying multiple source sensors. This system overcomes the limitations of traditional single-parameter monitoring, which has limited coverage and weak anti-interference capabilities. In the data preprocessing stage, a sliding window denoising and multi-source parameter collaborative standardization formula are used, combined with timestamp time sequence alignment technology. This effectively solves the analysis bias caused by differences in dimensions, noise interference, and time misalignment of multi-source data. The contribution of each parameter to the device status is dynamically quantified, avoiding evaluation distortion caused by differences in dimensions between parameters. Furthermore, the data analysis capability is improved by constructing a time sequence attention fusion model. Especially in the case of combined fault scenarios of sudden temperature rise and abnormal vibration, the accuracy and reliability of device operating status assessment are improved.
[0025] Second, this invention optimizes the training process of the temporal attention fusion model by coupling the loss function formula with the Adam optimizer. Combined with an early stopping mechanism, it avoids overfitting and shortens the training time. In the anomaly identification stage, the temporal attention fusion model outputs the attention weight values of each parameter and adopts a "double-layer verification" mechanism: first, it narrows down the fault range through state evaluation results, and then compares the single parameter threshold to effectively confirm the specific fault type. In addition, the anomaly information report generation mechanism integrates the anomaly type, occurrence time, parameter values, and attention weights to generate a structured report and push it to the user terminal. At the same time, it triggers intermittent alarms from LED lights and buzzers, which facilitates quick focusing on the fault source based on the anomaly information report, providing an efficient solution for the intelligent maintenance of massagers.
[0026] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description
[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0028] Figure 1 This is a flowchart of a deep learning-based method for monitoring the operating status of a massage device.
[0029] Figure 2 This is a framework diagram of a deep learning-based massage device operation status monitoring system.
[0030] Figure 3 This is a flowchart of the real-time monitoring and anomaly identification process in a deep learning-based massager operation status monitoring method. Detailed Implementation
[0031] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below. Example 1:
[0032] Data acquisition steps: In the scenario of monitoring the operating status of a home-use cervical spine massager, a piezoresistive pressure sensor is placed on the surface of the massage head of the massager to collect real-time pressure data and corresponding timestamps when the massage head presses on the cervical spine; a Hall current sensor is connected in series in the power supply circuit of the built-in vibration motor of the massager to collect current data and corresponding timestamps when the motor is working; a piezoelectric vibration sensor is attached to each vibration motor housing to collect vibration frequency data and corresponding timestamps; a thermocouple temperature sensor is installed at each motor winding and massager main unit housing to collect temperature data (winding temperature, housing temperature) and corresponding timestamps, and the collected raw data is transmitted to the data preprocessing step, such as... Figure 1 As shown.
[0033] Data preprocessing steps: After receiving the raw data, a sliding window with a window size of 5 seconds is used to remove environmental interference noise; missing values caused by brief sensor offline periods are filled in; and outliers in data such as pressure and current (e.g., pressure suddenly exceeding 80N, current abruptly dropping to 0) are removed; subsequently, standardization is performed using a multi-source parameter collaborative standardization formula, which is as follows: ,in, For the first The standardized results of each parameter For the first The original values of the parameters, The first The mean and standard deviation of the original values of each parameter were obtained from historically collected data. For the first The self-weighting coefficients of each parameter are fixed values preset based on the degree of influence of the parameter on the equipment status. This represents the total number of parameters, with a value of 4, corresponding to pressure, current, vibration frequency, and temperature. For the first The original values of the parameters, and , , The first The mean and standard deviation of the original values of each parameter. For the first With the The correlation coefficients of each parameter are determined; and based on timestamps, the four types of parameter data (pressure data, motor current data, vibration frequency data, and temperature data) are time-series aligned to form standardized time-series data, such as... Figure 3 As shown.
[0034] Model Construction and Training Steps: A Transformer architecture temporal attention fusion model is constructed, comprising an input embedding layer, a multi-head self-attention layer, an encoder layer, a feature fusion layer, and an output layer. The input embedding layer transforms standardized temporal data into a 512-dimensional high-dimensional feature vector using linear transformation and adds sinusoidal positional encoding to preserve temporal features. The multi-head self-attention layer uses eight parallel attention heads. Each attention head captures associated weights using a multi-source temporal attention weight formula, and then fuses multi-view features through concatenation and linear transformation. The multi-source temporal attention weight formula is as follows: ,in, For the first Time of the first Attention weights for each parameter, For the first Dimensional weights of each parameter, For the first The temporal position weight of each moment. These are the weight coefficients for the parameter dimensions. For the total number of parameters, This is the total timing length. , These are the parameter index and the time series index, respectively. For the first The time decay factor at time t is calculated using the formula: , The time series indicates the specific moment when the attention weights are being calculated. For the first The temporal decay factor at each time step; the encoder layer contains 6 stacked encoder units to deeply extract associated features; the feature fusion layer dynamically allocates feature weights for each parameter based on attention weights; the output layer consists of 2 fully connected layers and a softmax activation function, outputting three types of operational state evaluation results: normal, mildly abnormal, or severely abnormal.
[0035] The standardized time-series dataset is divided into training and testing sets in a 7:3 ratio. During the training phase, the loss is calculated using the coupling loss function formula, which is as follows: ,in, This represents the total loss value. For traditional cross-entropy loss, The coupling error coefficient is... The correlation coefficients for the predicted i-th and j-th parameters are... The correlation coefficients of the i-th and j-th parameters in the actual data. The total number of parameters is given; the Adam optimizer is used to iteratively optimize the model parameters; and an early stopping mechanism is introduced, which automatically stops training when the test set loss does not decrease for 5 consecutive rounds. After verifying the model performance, the optimized parameters are stored.
[0036] Real-time monitoring and anomaly identification steps: Standardized time-series data is received and input into a trained time-series attention fusion model. The model outputs evaluation results of "normal," "mild anomaly," and "severe anomaly," along with the attention weights of each parameter. The actual values of individual parameters such as pressure, motor current, vibration frequency, and temperature at the corresponding time in the evaluation results are compared with the threshold values of each parameter in a preset threshold library to determine if any single parameter exceeds the threshold. The threshold values of each parameter in the preset threshold library are set at the factory of the massager. When the time-series attention fusion model outputs "severe anomaly," and the winding temperature exceeds the threshold, with high current and temperature weights, an LED light flashes at 2Hz, a buzzer sounds an intermittent alarm, and an anomaly information report containing "severe anomaly - motor overheating," "abnormal parameters: winding temperature exceeds threshold, motor current exceeds threshold," and "occurrence time" is generated and sent to the user terminal.
[0037] In summary, for home-use cervical massagers, sensors are placed in locations such as the massage heads and motor power supply circuits to collect and transmit data such as pressure and current in real time. After preprocessing such as sliding window denoising and interpolation completion, standard data is formed using a multi-source parameter collaborative standardization formula and time-series alignment. A temporal attention fusion model is then constructed, with the dataset divided in a 7:3 ratio. The model is trained using a coupled loss function formula and the Adam optimizer, and an early stopping mechanism is introduced. During real-time monitoring, the temporal attention fusion model outputs three types of state results. In case of anomalies, alarms are triggered via LEDs and buzzers. By combining attention weights to locate the problem and generate a report, the model is sent to the user terminal. This allows for accurate and timely monitoring of the massager's status, providing a safe operating guarantee for home users. Example 2:
[0038] Data acquisition steps: In the scenario of monitoring the operating status of a commercial foot massager, a piezoresistive pressure sensor is placed on the surface of the foot massage rollers of the commercial foot massager to collect pressure data and corresponding timestamps during massage; a Hall current sensor is installed in the power supply circuit of the drive motor of the massager to collect motor current data and corresponding timestamps; a piezoelectric vibration sensor is fixed on each drive motor housing to collect vibration frequency data and corresponding timestamps; thermocouple temperature sensors are installed on each motor winding and the outer shell of the massager to collect winding temperature and shell temperature and corresponding timestamps, and the collected raw data is transmitted to the data preprocessing step.
[0039] Data preprocessing steps: After receiving the raw data, a sliding window with a window size of 10s is used to remove environmental interference noise; missing values caused by sensor malfunctions are filled in; and outliers such as pressure (e.g., sudden change to 150N) and vibration frequency (e.g., sudden change to 10Hz) are removed. Then, the data is processed using a multi-source parameter co-normalization formula, which is as follows: Based on timestamps, the four types of data are time-series aligned to form standardized time-series data.
[0040] Model Construction and Training Steps: A Transformer architecture temporal attention fusion model is constructed, comprising an input embedding layer, a multi-head self-attention layer, an encoder layer, a feature fusion layer, and an output layer. The input embedding layer converts standardized data into 1024-dimensional feature vectors and adds positional encoding to preserve temporal information. The multi-head self-attention layer uses 16 parallel attention heads. Each attention head captures associated weights using a multi-source temporal attention weight formula, and then fuses multi-view features through concatenation and linear transformation. The multi-source temporal attention weight formula is as follows: The encoder layer contains eight stacked encoder units that perform deep extraction of associated features; the feature fusion layer assigns parameter weights based on attention weights; the output layer consists of three fully connected layers and a softmax activation function, outputting three types of operational state evaluation results: normal, mildly abnormal, or severely abnormal.
[0041] The standardized time-series dataset is divided into training and testing sets in an 8:2 ratio. During the training phase, the loss is calculated using a coupling loss function formula, which is as follows: The model parameters are iteratively optimized using the Adam optimizer. An early stopping mechanism is also introduced, which stops training when the test set loss does not decrease for 8 consecutive rounds. After verifying the model performance, the optimized parameters are stored.
[0042] Real-time monitoring and anomaly identification steps: Standardized time-series data is received and input into a trained time-series attention fusion model. The model outputs three types of operational status evaluation results and parameter attention weights. The actual values of individual parameters (pressure, motor current, vibration frequency, and temperature) at the corresponding time in the evaluation results are compared with the threshold values of each parameter in a preset threshold library to determine if any single parameter exceeds the threshold. The threshold values of each parameter in the preset threshold library are set at the factory of the massager. When the model outputs "mild anomaly," and the comparison reveals that the roller pressure exceeds the threshold, and the attention weight shows a higher weight for the pressure parameter, an LED light flashes at 0.5Hz, a buzzer sounds once every 3 seconds, and an anomaly information report containing "mild anomaly - high roller pressure," "abnormal parameter: roller pressure exceeds threshold," and "occurrence time" is generated and sent to the user terminal. Figure 3 As shown.
[0043] In summary, for commercial foot massagers, sensors are deployed in components such as the massage rollers and drive motors to collect and transmit data on pressure, vibration frequency, and other parameters. After sliding window denoising and outlier removal, standard data is generated using a multi-source parameter collaborative standardization formula and time-series alignment. A temporal attention fusion model is then constructed, with the dataset divided in an 8:2 ratio. This model is trained using a coupled loss function formula and the Adam optimizer, and an early stop mechanism is implemented. During real-time monitoring, abnormal conditions trigger audible and visual alarms. Weighted problem localization is used to generate reports that are sent to the user's terminal. This approach improves the operational efficiency of commercial massagers and reduces operational losses caused by downtime due to malfunctions.
[0044] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A method for monitoring the operating status of a massager based on deep learning, characterized in that, The specific steps of this method are as follows: Data acquisition steps: Multi-source sensors are deployed on the massager to collect raw data including pressure data, motor current data, vibration frequency data and temperature data in real time during the operation of the massager, and transmit them to the data preprocessing step; Data preprocessing steps: The raw data is received, and cleaned by using a sliding window to denoise, fill in missing values, and remove outliers. Then, the cleaned data is standardized using a multi-source parameter co-standardization formula, which is: ,in, For the first The standardized results of each parameter For the first The original values of the parameters, The first The mean and standard deviation of the original values of each parameter were obtained from historically collected data. For the first The self-weighting coefficients of each parameter are fixed values preset based on the degree of influence of the parameter on the equipment status. For the total number of parameters, For the first The original values of the parameters, and , , The first The mean and standard deviation of the original values of each parameter. For the first With the The correlation coefficients of each parameter are determined; and the data are time-series aligned based on timestamps to form standardized time-series data. Model Construction and Training Steps: A temporal attention fusion model based on the Transformer architecture is constructed. The standardized temporal dataset is proportionally divided into training and test sets. Training is performed by coupling the loss function formula and the optimizer, and an early stopping mechanism is introduced to optimize and store the parameters. Each attention head captures the correlation weights of different parameters in the temporal dimension through a multi-source temporal attention weight formula. The multi-source temporal attention weight formula is as follows: ,in, For the first Time of the first Attention weights for each parameter, For the first Dimensional weights of each parameter, For the first The temporal position weight of each moment. These are the weight coefficients for the parameter dimensions. For the total number of parameters, This is the total timing length. , These are the parameter index and the time series index, respectively. For the first The time decay factor at time t is calculated using the formula: , For time sequence moments, For the first The time decay factor at any given moment; Real-time monitoring and anomaly identification steps: Receive standardized time-series data and input it into the time-series attention fusion model, output the running status evaluation result, compare the evaluation result with the preset threshold, trigger an alarm when an abnormal state occurs, locate abnormal parameters by combining attention weights and send abnormal information.
2. The massager operation status monitoring method based on deep learning according to claim 1, characterized in that, In the data acquisition step, the multi-source sensors include a piezoresistive pressure sensor, a Hall current sensor, a piezoelectric vibration sensor, and a thermocouple temperature sensor. Specifically, the piezoresistive pressure sensor is positioned on the surface of the massage head of the massager to collect pressure data and corresponding timestamps during operation; the Hall current sensor is positioned in the power supply circuit of the massager motor to collect motor current data and corresponding timestamps during operation; the piezoelectric vibration sensor is positioned on the vibration motor housing of the massager to collect vibration frequency data and corresponding timestamps during operation; and the thermocouple temperature sensors are positioned on the motor windings and housing of the massager to collect temperature data and corresponding timestamps during operation.
3. The massage device operation status monitoring method based on deep learning according to claim 1, characterized in that, In the model construction and training steps, a Transformer architecture temporal attention fusion model is constructed, comprising an input embedding layer, a multi-head self-attention layer, an encoder layer, a feature fusion layer, and an output layer. The input embedding layer converts standardized temporal data into 256-1024 dimensional feature vectors through linear transformation, while adding positional encoding information to preserve temporal features. The multi-head self-attention layer sets 4-16 parallel attention heads, and each attention head captures the correlation weights of different parameters in the temporal dimension through a multi-source temporal attention weight formula, and then fuses multi-view features through concatenation and linear transformation. The encoder layer contains 4-8 stacked encoder units, which perform deep extraction of associated features; The feature fusion layer dynamically assigns weights to each parameter feature based on attention weights; the output layer consists of 1-3 fully connected layers and a softmax activation function, outputting three types of operational state evaluation results: normal, mildly abnormal, or severely abnormal.
4. The massage device operation status monitoring method based on deep learning according to claim 3, characterized in that, In the model construction and training steps, the standardized time series dataset is divided into training and test sets in a ratio of 6:4 to 8:
2. During the training phase, the loss is calculated by coupling the loss function formula, and the model parameters are iteratively optimized by the Adam optimizer. At the same time, an early stopping mechanism is introduced, which automatically stops training when the test set loss does not decrease for 3-8 consecutive rounds. After training is completed, the performance of the time series attention fusion model is verified by the test set, and the optimized time series attention fusion model parameters are stored.
5. The massage device operation status monitoring method based on deep learning according to claim 4, characterized in that, In the model construction and training steps, the loss is calculated during the training phase using a coupling loss function formula, which is: ,in, This represents the total loss value. For traditional cross-entropy loss, The coupling error coefficient is... The correlation coefficients for the predicted i-th and j-th parameters are... The correlation coefficients of the i-th and j-th parameters in the actual data. This represents the total number of parameters.
6. The method for monitoring the operating status of a massager based on deep learning according to claim 1, characterized in that, In the real-time monitoring and anomaly identification step, standardized time-series data is received, a time-series attention fusion model is called to perform inference calculations, and three types of operating status evaluation results (normal, mildly abnormal, or severely abnormal) and attention weight values of each parameter are output. When outputting mild or severe anomaly assessment results, the actual values of individual parameters such as pressure, motor current, vibration frequency, and temperature at the corresponding time in the assessment results are compared with the threshold values of each parameter in the preset threshold library to determine whether any single parameter exceeds the threshold. The threshold values of each parameter in the preset threshold library are set by the massager at the factory. When the assessment result is consistent with the result of exceeding the threshold of a single parameter, the LED alarm light is triggered to flash at a frequency of 0.5-2Hz and the buzzer alarm is triggered intermittently. The abnormal parameter is located by combining the attention weight value of each parameter, and an anomaly information report containing the anomaly type, abnormal parameter, occurrence time, and current parameter value is generated and sent to the user terminal.
7. A deep learning-based massager operation status monitoring system, the system being applicable to the deep learning-based massager operation status monitoring method according to any one of claims 1-6, characterized in that, The system includes: Data acquisition module: Multi-source sensors are installed on the massager to collect raw data, including pressure data, motor current data, vibration frequency data and temperature data, in real time during the operation of the massager, and then transmit them. Data preprocessing module: Receives raw data, cleans the raw data by denoising, filling in missing values and removing outliers using a sliding window, then standardizes the cleaned data using a multi-source parameter collaborative standardization formula, and aligns the data according to timestamps to form standardized time-series data; Model building and training module: Construct a temporal attention fusion model based on the Transformer architecture, divide the standardized time series dataset into training and test sets according to the proportion, train it by coupling the loss function formula and the optimizer, and introduce an early stopping mechanism to optimize and store the parameters. Real-time monitoring and anomaly identification module: Receives standardized time-series data and inputs it into the time-series attention fusion model, outputs the operation status evaluation result, compares the evaluation result with the preset threshold, triggers an alarm when an abnormal state occurs, and locates abnormal parameters by combining attention weights and sends abnormal information.