Machine learning method and system

Abstract

The invention provides a machine learning method and system, and relates to the technical field of data processing, and the method comprises the steps: 1, collecting a real-time sensing data stream of industrial equipment at an edge calculation node, the data stream comprising a welding robot end temperature value, a driving motor current value and a mechanical arm vibration frequency value; and step 2, based on the real-time sensing data stream, extracting characteristic values of a starting positioning point, a path midpoint and an ending positioning point in the welding track through an edge intelligent processing unit, and dynamically generating an anomaly judgment threshold according to the characteristic values. The method can quickly respond to instantaneous abnormity, improves the progressive fault recognition and prediction capability, reduces the equipment fault risk and maintenance cost, and guarantees the production safety and quality.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a machine learning method and system. Background Art

[0002] In edge computing monitoring of industrial equipment (such as welding robots), it's common to use real-time sensor data (such as temperature, current, and vibration) for anomaly detection and immediate control (power outages, speed regulation, etc.). However, existing edge-side solutions have certain shortcomings when faced with progressive failures that indicate slow deterioration of equipment health (such as gradual changes in vibration characteristics caused by wear on a robot arm's bearings): Although most systems can record sensor data and control instructions, there is room for improvement in integrating and analyzing the "long-term relationship between real-time control operations (such as speed regulation) and equipment status (such as vibration changes)." In some cases, the historical data generated by control instructions to respond to instantaneous anomalies and their impact on equipment status (especially vibration patterns) are not stored in the system.

[0003] This may make it difficult for the diagnostic system to fully utilize the state evolution information of the equipment during long-term operation (especially after control intervention), and the early identification accuracy and prediction capabilities of progressive problems are sometimes limited, affecting the effectiveness of predictive maintenance. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a machine learning method and system to reduce the risk of equipment failure and ensure industrial production safety.

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows: In a first aspect, a machine learning method is provided, the method comprising: Step 1: Collect real-time sensor data streams of industrial equipment at the edge computing node, including the temperature value of the welding robot end, the current value of the drive motor, and the vibration frequency value of the robotic arm; Step 2: Based on the real-time sensor data stream, the edge intelligent processing unit extracts the characteristic values ​​of the starting point, path midpoint, and end point in the welding trajectory, and dynamically generates an abnormality judgment threshold based on the characteristic values; Step 3: Based on the abnormality determination threshold, a power cut-off instruction is triggered when the driving motor current value exceeds the threshold, or a welding speed adjustment instruction is generated when the vibration frequency value of the robot arm deviates from the normal range; Step 4: Aggregate the sensor data stream, abnormality judgment threshold, and welding speed adjustment instruction into time-series quasi-physical units according to the equipment entity and store them; Step 5: Based on the historical vibration frequency value sequence and the corresponding control instruction record in the pseudo-physical unit, a progressive fault diagnosis model is constructed and the decision parameters of the progressive fault diagnosis model are updated to obtain an updated progressive fault diagnosis model; Step 6: Based on the updated progressive fault diagnosis model, output a report on the evolution sequence of the vibration characteristics of the robotic arm.

[0006] In a second aspect, a machine learning system includes: An acquisition module is used to collect real-time sensor data streams from industrial equipment at edge computing nodes. The data streams include the temperature value of the welding robot end, the current value of the drive motor, and the vibration frequency value of the robotic arm. A threshold generation module is used to extract the characteristic values ​​of the starting point, path midpoint, and end point in the welding trajectory through the edge intelligent processing unit based on the real-time sensor data stream, and dynamically generate anomaly judgment thresholds based on the characteristic values; A control module is used to trigger a power cut-off instruction when the driving motor current value exceeds the threshold value based on an abnormality determination threshold, or to generate a welding speed adjustment instruction when the vibration frequency value of the robot arm deviates from the normal range; The storage module is used to aggregate the sensor data stream, abnormality judgment threshold and welding speed adjustment instructions into time-series quasi-physical units for storage according to the equipment entity; An updating module is used to construct a progressive fault diagnosis model and update the decision parameters of the progressive fault diagnosis model based on the historical vibration frequency value sequence and the corresponding control instruction record in the pseudo-physical unit to obtain an updated progressive fault diagnosis model; The report output module is used to output a report on the evolution sequence of vibration characteristics of the robotic arm based on the updated progressive fault diagnosis model.

[0007] According to a third aspect, a computing device includes: one or more processors; The storage device is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0008] In a fourth aspect, a computer-readable storage medium stores a program, which implements the method when executed by a processor.

[0009] The above solution of the present invention includes at least the following beneficial effects: By collecting real-time sensor data through edge computing nodes and dynamically generating abnormality judgment thresholds, it can quickly respond to transient abnormalities (such as overcurrent in the drive motor), promptly trigger safety instructions such as power cut-off, reduce the risk of equipment failure, and ensure industrial production safety; in response to the deviation of the vibration frequency of the robot arm, it generates welding speed adjustment instructions that match the degree of abnormality, realizes fine-grained adjustment of the equipment's operating status, and reduces product quality problems caused by abnormal vibration.

[0010] By aggregating sensor data, thresholds, and control instructions into time-series pseudo-physical units for storage according to device entities, structured data management is achieved, making it easier to trace the historical status of the device. A progressive fault diagnosis model is constructed and dynamically updated based on the historical data in the pseudo-physical units, fully integrating the correlation information between control intervention and state evolution during long-term operation of the equipment. This improves the early recognition accuracy and prediction capability of slow-degradation faults such as robot arm bearing wear, providing strong support for predictive maintenance. The output vibration feature evolution sequence report contains information such as feature changes, intervention correlation, and wear prediction, reducing equipment downtime and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a flowchart of a machine learning method provided by an embodiment of the present invention.

[0012] Figure 2 Schematic diagram of a machine learning system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0013] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0014] like Figure 1 As shown, an embodiment of the present invention provides a machine learning method, which includes the following steps: Step 1: Collect real-time sensor data streams of industrial equipment at the edge computing node, including the temperature value of the welding robot end, the current value of the drive motor, and the vibration frequency value of the robotic arm; Step 2: Based on the real-time sensor data stream, the edge intelligent processing unit extracts the characteristic values ​​of the starting point, path midpoint, and end point in the welding trajectory, and dynamically generates an abnormality judgment threshold based on the characteristic values; Step 3: Based on the abnormality determination threshold, a power cut-off instruction is triggered when the driving motor current value exceeds the threshold, or a welding speed adjustment instruction is generated when the vibration frequency value of the robot arm deviates from the normal range; Step 4: Aggregate the sensor data stream, abnormality judgment threshold, and welding speed adjustment instruction into time-series quasi-physical units according to the equipment entity and store them; Step 5: Based on the historical vibration frequency value sequence and the corresponding control instruction record in the pseudo-physical unit, a progressive fault diagnosis model is constructed and the decision parameters of the progressive fault diagnosis model are updated to obtain an updated progressive fault diagnosis model; Step 6: Based on the updated progressive fault diagnosis model, output a report on the evolution sequence of the vibration characteristics of the robotic arm.

[0015] In an embodiment of the present invention, by collecting real-time sensor data through edge computing nodes and dynamically generating abnormality judgment thresholds, it can quickly respond to transient abnormalities (such as overcurrent in the drive motor), promptly trigger safety instructions such as power cut-off, reduce the risk of equipment failure, and ensure industrial production safety; in response to the deviation of the vibration frequency of the robot arm, a welding speed adjustment instruction matching the degree of abnormality is generated, thereby achieving fine-grained adjustment of the equipment operating status and reducing product quality problems caused by abnormal vibration.

[0016] Aggregating sensor data, thresholds, and control instructions into time-series virtual units for storage by device entity enables structured data management and facilitates tracing of device historical status. A progressive fault diagnosis model is constructed and dynamically updated based on historical data in the virtual units. This fully integrates the correlation between control interventions and state evolution during long-term device operation, improving the early identification accuracy and prediction capabilities of slow-degradation faults such as robot arm bearing wear, providing strong support for predictive maintenance. The output vibration feature evolution sequence report includes information such as feature changes, intervention correlations, and wear predictions, reducing equipment downtime and maintenance costs.

[0017] In a preferred embodiment of the present invention, step 1 collects real-time sensor data streams of industrial equipment at the edge computing node, wherein the data streams include the temperature value of the welding robot end, the current value of the driving motor, and the vibration frequency value of the robotic arm; step 2 includes: Step 200, extracting the temperature change rate characteristics of the welding trajectory starting point, the current distribution statistical characteristics of the welding path midpoint, and the vibration frequency domain envelope characteristics of the welding end point based on the real-time sensor data stream; Step 201: Inversely weight the temperature change rate feature in combination with the real-time welding speed parameter to generate a temperature change weighted feature, forwardly weight the current distribution statistical feature in combination with the real-time welding material thickness parameter to generate a current distribution weighted feature, and forwardly weight the vibration frequency domain envelope feature in combination with the equipment cumulative working time parameter to generate a vibration frequency domain weighted feature. Step 202 : Dynamically generate an abnormality determination threshold by fusing the weighted features of temperature change, current distribution, and vibration frequency domain.

[0018] In this embodiment of the present invention, an edge computing node (such as an industrial-grade edge gateway or embedded controller) establishes a real-time communication link with the welding robot's sensor system and collects three types of core data according to preset rules. The specific process is as follows: Temperature value collection at the end of the welding robot: A high-precision thermocouple sensor (measuring range: -50°C to 1000°C) is installed at the end of the robot's welding gun. The sensor is connected to the edge node through an industrial bus (such as Modbus RTU). The edge node triggers collection at a sampling period of 50 milliseconds. During each collection, the original temperature signal is first filtered with a sliding window (the window size is 3 sampling points, and the average value is taken) to eliminate instantaneous jump values ​​caused by arc interference (if a single jump exceeds 50°C, it is judged as interference and replaced with the effective value of the previous moment). After collection, each temperature value is marked with a precise timestamp (accurate to milliseconds) and temporarily stored in the ring buffer of the edge node in the format of "acquisition time-temperature value".

[0019] Drive motor current collection: A Hall effect current sensor (measurement range 0-50A) is connected in series in the robot's drive motor power supply circuit. The analog signal output by the sensor is converted to a digital signal by a 16-bit analog-to-digital converter and then transmitted to the edge node. The sampling cycle is synchronized with temperature collection (50 milliseconds / time). During collection, the motor operation phase ("arc on," "welding," and "arc off" status signals output by the robot controller) is recorded simultaneously. For example, the current value is only included in the valid data when the status signal is "welding." The edge node clips outliers from the collected current values ​​(retaining data within the 0-40A range and recording any excess as 40A to avoid extreme values ​​caused by sensor failure) and stores them with an associated timestamp.

[0020] Collection of robotic arm vibration frequency values: A three-axis accelerometer (measurement range ±16g, sampling rate 1000Hz) is attached to the joints of the upper and lower arms of the robotic arm. The sensor sends vibration time domain data (1000 acceleration values ​​per second) to the edge node via a wireless transmission module (such as Wi-Fi6). After receiving the data, the edge node first preprocesses it: the gravitational acceleration component is removed (through high-pass filtering with a cutoff frequency of 0.5Hz). The preprocessed time domain signal is then framed (each frame contains 1024 sampling points with a frame overlap rate of 50%). A fast Fourier transform (FFT) is performed on each frame of data to convert it to the frequency domain to obtain a frequency-amplitude distribution in the range of 0-500Hz. Finally, the top five frequency values ​​with the highest amplitude (such as 20Hz, 50Hz, and 100Hz) are extracted from the frequency domain results as the robotic arm vibration frequency value for that sampling period and stored with an associated timestamp.

[0021] After the three types of data are collected, the edge node integrates them into a real-time sensor data stream according to the structure of "device ID-timestamp-temperature-current-vibration frequency" and continuously pushes them to the input buffer of the edge intelligent processing unit.

[0022] In step 200, when the robot coordinates reach the starting point of the welding trajectory (preset coordinates X0, Y0, Z0), the starting point marking is triggered. Temperature data from the six sampling periods (300 milliseconds) before and the six sampling periods (300 milliseconds) after the marking moment (a total of 13 temperature values) are selected and arranged in chronological order as T0, T1, ..., T12. The temperature differences between two adjacent moments are calculated: ΔT1 = T1 - T0, ΔT2 = T2 - T1, ..., ΔT12 = T12 - T11. Each difference corresponds to a 50 millisecond (0.05 second) interval. Therefore, the instantaneous temperature change rate is v1 = ΔT1 / 0.05, v2 = ΔT2 / 0.05, ..., v12 = ΔT12 / 0.05. The two instantaneous temperature change rates with the largest absolute values ​​are removed (to eliminate extreme interference). The arithmetic mean of the remaining 10 instantaneous temperature change rates is taken as the temperature change rate characteristic of the starting point, reflecting the temperature rise trend during the initial welding phase. When the robot coordinates reach 50% of the total trajectory length (preset coordinates Xm, Ym, Zm), it is marked as the midpoint of the path. The current data of the 12 sampling periods (600 milliseconds) before and the 12 sampling periods (600 milliseconds) after the marking moment (a total of 25 current values) are selected and recorded as I0, I1, ..., I24. These current values ​​are statistically calculated: average value = (I0 + I1 + ... + I24) / 25, reflecting the average motor load at the midpoint stage; variance = [(I0 - average value)² + (I1 - average value)² + ... + (I24 - average value)²] / 25, reflecting the degree of current fluctuation; range = max(I0, ..., I24) - min(I0, ..., I24), reflecting the range of current variation. The current values ​​are combined in the order of "average value - variance - range" as the statistical characteristics of the current distribution at the midpoint of the path.

[0023] Extraction of vibration frequency domain envelope features of the terminal positioning point: When the robot coordinates reach the end point of the welding trajectory (preset coordinate X e , Y e , Z e) is marked as the end positioning point; vibration frequency domain data from the 8 sampling periods (400 milliseconds) before and 8 sampling periods (400 milliseconds) after the marking moment are selected (a total of 17 sets of frequency-amplitude distributions), each set of data containing amplitude information from 0 to 500 Hz. For each frequency point (e.g., 10 Hz, 20 Hz, ..., 500 Hz), the maximum amplitude value from the 17 sets of data is extracted to form a "frequency-maximum amplitude" envelope. A baseline amplitude value is set (1.2 times the average amplitude of the robot's end point during normal welding over the past 30 days). All frequency points with amplitudes exceeding this baseline are filtered from the envelope and listed in a "frequency value-amplitude value" format as the vibration frequency domain envelope characteristics of the end positioning point, reflecting the abnormal vibration distribution at the end of welding.

[0024] Step 201: Based on the basic features extracted in step 200, combined with real-time process parameters and equipment status parameters, weighted features are generated through weight adjustment: Temperature variation weighted feature (reverse weight adjustment): Based on the real-time welding speed parameter (provided by the robot control system, in mm / s, ranging from 1 to 10 mm / s), the adjustment adopts the reverse logic of "the higher the speed, the lower the weight" (during high-speed welding, temperature changes rapidly, and the proportion of non-fault fluctuations is high). The welding speed baseline is set to 5 mm / s, and the weight coefficient is calculated: if the real-time speed V ≤ 5 mm / s, the weight coefficient Kt = 1; if the real-time speed V > 5 mm / s, the weight coefficient Kt = 5 / V (for example, when V = 8 mm / s, Kt = 5 / 8). The temperature change rate feature extracted in step 200 is multiplied by Kt to obtain the temperature variation weighted feature, which weakens the impact of non-fault temperature fluctuations during high-speed welding.

[0025] Current distribution weighted feature (forward weight adjustment): Adjustments are based on the real-time welding material thickness parameter (input from the production work order system, in mm, ranging from 0.5 to 10 mm), using a forward logic of "the thicker the material, the higher the weight" (thicker materials require higher currents, and current stability has a more significant impact on quality). A base material thickness value of 3 mm is set, and the weight coefficient is calculated: if the real-time thickness D ≤ 3 mm, the weight coefficient Kj = 1; if the real-time thickness D > 3 mm, the weight coefficient Kj = D / 3 (for example, if D = 6 mm, Kj = 6 / 3 = 2). The current distribution statistical features extracted in step 200 (mean, variance, and range multiplied by Kj, respectively) are combined to create a weighted current distribution feature, enhancing attention to current fluctuations when welding thick materials.

[0026] Vibration frequency domain weighted features (forward weight adjustment): Adjustments are based on the cumulative operating time parameter (the total robot operating time recorded by the edge node, measured in hours, ranging from 0 to 10,000 hours). A positive logic of "longer operating time, higher weight" is used (increasing cumulative operating time increases the risk of robot arm wear, and vibration characteristics require more attention). The cumulative operating time baseline is set at 1,000 hours, and the weight coefficient is calculated: if the cumulative operating time H ≤ 1,000 hours, the weight coefficient Kv = 1; if the cumulative operating time H > 1,000 hours, the weight coefficient Kv = H / 1,000 (for example, if H = 1,500 hours, Kv = 1,500 / 1,000). The vibration frequency domain envelope features extracted in step 200 (the amplitude value at each frequency point multiplied by Kv) are tabulated to generate vibration frequency domain weighted features, thereby improving the sensitivity of older equipment to vibration anomalies.

[0027] Step 202: The three weighted features obtained in step 201 are integrated to dynamically generate an abnormality determination threshold adapted to the real-time scenario. The process is as follows: Retrieve the three types of weighted feature data from the simulated memory area during the robot's normal welding over the past 20 days, and calculate the historical maximum and minimum values ​​of each feature (e.g., the historical range of the weighted feature of temperature change is 0.2-3.5°C / s, and the historical range of the average value of the weighted feature of current distribution is 5-25A, etc.). The current temperature change weighted feature, current distribution weighted feature, and vibration frequency domain weighted feature are mapped to the range of 0 to 1 according to the formula "(current value - historical minimum value) / (historical maximum value - historical minimum value)" to eliminate dimensional differences. Based on the normalized historical normal data, the normal fluctuation range of each feature is calculated: the mean μ and standard deviation σ of the historical data are taken to obtain the normal range of [μ-2σ, μ+2σ] (covering 95% of normal data). If a feature is continuously in the range of [μ+1.5σ, μ+2σ] (close to the historical upper limit) in the last 8 sampling periods, the upper limit of the feature is temporarily relaxed to μ+2.5σ (to avoid misjudgment due to short-term fluctuations). If a feature is in the range of [μ-σ, μ+σ] (fluctuation is stable) in the last 15 sampling periods, the upper and lower limits of the feature are tightened to [μ-1.5σ, μ+1.5σ] (to improve detection sensitivity).

[0028] Fusion weights are assigned to the three features (based on historical fault data verification, the weighted temperature change feature accounts for 30%, the weighted current distribution feature accounts for 40%, and the weighted vibration frequency domain feature accounts for 30%). The adjusted upper limits of the normal range of each feature are weighted and summed to obtain the final abnormality judgment threshold. For example, if the upper limit of the temperature feature is 0.8, the upper limit of the current feature is 0.7, and the upper limit of the vibration feature is 0.9, then the threshold = 0.8 × 30% + 0.7 × 40% + 0.9 × 30% = 0.79.

[0029] Through targeted sensor selection, sampling cycle design, and preprocessing (filtering and outlier handling), we ensure that temperature, current, and vibration data truly reflect the equipment status. Feature extraction focuses on the start, midpoint, and end points of the welding trajectory, avoiding data redundancy throughout the entire process. This ensures that the features are more closely aligned with the equipment's operating characteristics at each stage, enhancing their relevance to fault conditions. Reverse / forward weighting is performed based on real-time parameters such as welding speed, material thickness, and accumulated equipment time, enabling features to dynamically adapt to varying process conditions and equipment aging, reducing interference from non-fault-related factors. Dynamic adjustment of the threshold range based on historical data and real-time fluctuation trends preserves the baseline for normal conditions while accommodating short-term fluctuations or stable operation. This significantly reduces the risk of false positives (relaxed when approaching the upper limit) or false negatives (tightened when stable) caused by fixed thresholds, providing a precise basis for abnormal response. Accurate features, adaptive weighting, and dynamic thresholds make the data stored in the quasi-physical unit more valuable for analysis, providing high-quality training data for the progressive fault diagnosis model and improving the model's accuracy in identifying slowly degrading equipment conditions.

[0030] In a preferred embodiment of the present invention, step 3 includes: The driving motor current value is compared with the abnormality judgment threshold in real time. When the current value continues to exceed the abnormality judgment threshold within continuous sampling cycles, the power cut-off command is triggered; the envelope offset area of ​​the vibration spectrum and the abnormality judgment threshold is calculated based on the vibration frequency value of the robot arm; when the envelope offset area exceeds the preset tolerance threshold, a welding speed adjustment command is generated, and the adjustment amplitude of the welding speed adjustment command increases with the increase of the envelope offset area.

[0031] In the embodiment of the present invention, based on the real-time collected driving motor current value and the abnormality determination threshold generated in step 202, a safety instruction is triggered through continuous monitoring and logical judgment. The specific process is as follows: The edge computing node obtains the drive motor current value at a sampling period of 50 milliseconds (consistent with the current acquisition period). After each acquisition, the current current value is immediately compared with the abnormal judgment threshold. If the current value is ≤ the abnormal judgment threshold, it is marked as "normal state" and the continuous over-limit counter is reset (the counter value is reset to zero); if the current value is greater than the abnormal judgment threshold, it is marked as "over-limit state" and the continuous over-limit counter value is increased by 1 (for example, the first over-limit counter = 1, the second continuous over-limit counter = 2, and so on).

[0032] When the continuous over-limit counter reaches three times (i.e., three consecutive sampling cycles, totaling 150 milliseconds), it is determined to be a "persistent current anomaly." At this point, the control logic module of the edge computing node immediately generates a power-off command. The command format includes the device ID (e.g., "Welding Robot #001"), the command type ("Emergency Power Off"), the trigger reason ("Drive motor current has exceeded the limit three times in a row, current value XXA, threshold XXA"), and a timestamp. The generated power-off command is converted by the communication protocol conversion module into a protocol recognizable by the robot control system (e.g., Profinet) and transmitted to the robot's power control unit via Industrial Ethernet. Upon receiving the command, the power control unit immediately disconnects the main power supply circuit of the drive motor and returns a "command execution successful" feedback signal to the edge node. The edge node records the command trigger time and execution results, and synchronously stores them in the virtual memory area.

[0033] Extract the frequency domain data of the robot arm vibration from the real-time sensor data stream (from the 0-500 Hz spectrum converted by FFT in step 1, including the amplitude value of each frequency point); at the same time, convert the abnormality judgment threshold generated in step 202 into a "vibration frequency domain threshold envelope" - that is, according to the threshold value, generate a corresponding amplitude upper limit line in the 0-500 Hz frequency range (for example, the threshold is mapped to "the maximum allowable amplitude A at frequency f"). th (f)”). Calculate the “offset area” between the vibration spectrum and the threshold envelope (only the part of the spectrum exceeding the threshold is counted), specifically including: The 0-500Hz frequency range is divided into 50 intervals of 10Hz (i.e., 0-10Hz, 10-20Hz, ..., 490-500Hz). The intermediate frequency point of each interval (e.g., 5Hz, 15Hz...495Hz) is used as the calculation basis. For each intermediate frequency point, the actual amplitude A of the vibration spectrum is calculated. a (f) Amplitude A of the threshold envelope th (f) difference, if A a (f)≤A th (f), the difference is 0; if A a (f)>A th (f), the difference is A a (f)-A th (f) The width of each frequency interval is 10 Hz. Multiply the amplitude difference of the interval by 10 Hz to obtain the offset area of ​​a single interval. The offset areas of 50 intervals are accumulated, and the total is the "envelope offset area between the vibration spectrum and the abnormality judgment threshold" (unit: Hz•mm / s², the amplitude unit is mm / s²).

[0034] A tolerance threshold for the envelope offset area is preset (set based on historical normal welding data, such as 500Hz•mm / s²). If the calculated envelope offset area is ≤ the tolerance threshold, it is judged as a "minor vibration anomaly" and no adjustment command is generated. If the area is greater than the tolerance threshold, it is judged as a "significant vibration anomaly" and triggers the generation of a welding speed adjustment command. The adjustment amplitude is directly proportional to the envelope offset area. The specific rule is: for every 100Hz•mm / s² exceeding the tolerance threshold, the welding speed is reduced by 2% (the reference speed is the current welding speed V0). For example, if the envelope offset area = 700Hz•mm / s² (exceeding the tolerance threshold by 200Hz•mm / s²), the adjustment amplitude = 200 / 100×2% = 4%, that is, the new welding speed = V0×(1-4%). ​​To avoid the impact of too low speed on welding quality, the adjustment amplitude is set to an upper limit of 20% (i.e., a maximum reduction of 20%). Even if the offset area far exceeds the threshold, the speed will not be lower than 80% of the original speed. The control logic unit generates a welding speed adjustment command, which includes the equipment ID, target speed value, adjustment amplitude, trigger reason ("vibration envelope offset area XX, exceeds tolerance threshold XX") and timestamp. The command is converted into a format recognizable by the robot (such as ModbusTCP) by the communication protocol conversion unit and sent to the robot's motion control unit. The command information is also recorded in the virtual memory area.

[0035] Through the judgment logic of "continuous sampling cycle exceeding the limit" (rather than a single exceeding the limit), false triggers caused by instantaneous current fluctuations are effectively filtered out, ensuring that the power cut command is triggered only in the event of a persistent anomaly, which not only ensures equipment safety but also reduces unnecessary downtime losses. Based on the "envelope offset area" to quantify the degree of vibration anomaly, the welding speed adjustment range is proportional to the severity of the anomaly (the larger the area, the larger the adjustment range), avoiding "one-size-fits-all" adjustment, which can both suppress excessive vibration and minimize the impact on welding efficiency and quality. Power cutoff targets fatal current anomalies (rapid stop loss), while speed adjustment targets non-fatal vibration anomalies (dynamic optimization). The combination of these two strategies achieves a balance between "safety in emergencies and production in minor anomalies," improving the stability of industrial production.

[0036] In a preferred embodiment of the present invention, step 4 aggregates the sensor data stream, abnormality determination threshold, and welding speed adjustment instruction into time-series virtual units for storage according to the device entity; step 5 includes: Step 500 , constructing an initial progressive fault diagnosis model using the historical vibration frequency value sequence and the corresponding welding speed adjustment instruction record in the time-series virtual unit; Step 501, obtaining the vibration frequency value sequence and welding speed adjustment instruction record of the latest acquisition cycle, inputting the latest data into the initial progressive fault diagnosis model, updating the vibration characteristic decision parameters, and generating an updated progressive fault diagnosis model, specifically including: Step 5010, obtaining the spectrum envelope characteristic value of the latest vibration frequency value sequence; Step 5011, comparing the spectrum envelope characteristic value with the historical spectrum benchmark in the initial progressive fault diagnosis model to generate a characteristic offset data sequence; Step 5012: Process the characteristic offset data sequence, identify sampling periods that continuously exceed the offset threshold, and generate an excess period mark sequence; Step 5013: Adjust the vibration feature weight coefficient based on the duration of the overrun period mark sequence to generate a preliminary revised weight coefficient; Step 5014: extract the number of interventions in the welding speed adjustment instruction record, perform attenuation processing on the preliminary correction weight coefficient based on the intervention frequency, and generate an updated vibration characteristic decision parameter; Step 5015 , writing the updated vibration characteristic decision parameters into the initial progressive fault diagnosis model to generate an updated progressive fault diagnosis model.

[0037] In an embodiment of the present invention, multiple types of data related to device operation are aggregated by device entity and stored in a time series format as quasi-physical units. The specific process is as follows: The edge computing node first distinguishes different industrial devices by their unique device identifiers (such as the welding robot's serial number "WR-2023-005"), creates a separate storage directory for each device, and then categorizes the data to be stored: Sensor data stream: includes the welding robot end temperature value (every 50 milliseconds, with a timestamp), the drive motor current value (same frequency, with an operation stage mark), and the robot arm vibration frequency value (same frequency, with frequency domain characteristics); Abnormal determination threshold: the dynamic threshold generated in step 202 (updated every 1 second, with a generation timestamp and corresponding feature fusion logic description); Welding speed adjustment command: The command generated in step 3 (including command ID, trigger time, target speed, adjustment range, trigger reason, etc.).

[0038] Aggregate data using "time windows": Set each minute as a basic time window (this can be adjusted based on the equipment's operating rhythm, such as 5 minutes for longer welding cycles). Arrange the data within each window in ascending order by timestamp to form "simulated unit fragments." Each fragment contains: all sensor data within that window (structured in "timestamp-temperature-current-vibration frequency" format); the abnormality determination threshold generated within that window (if updated once per second, each 1-minute window contains 60 threshold records). The welding speed adjustment instructions triggered within this window (if any, sorted by trigger time, if not, marked as "no instruction"). The virtual unit of each device is stored in the hierarchical structure of "device ID-date-time window" (such as "WR-2023-005 / 20240901 / 08:00-08:01") and a dual index is established: Time index: records the start / end timestamps of each simulacrum unit segment, supporting queries by time period; Data type indexing: Distinguishes sensor data, thresholds, and instructions within each segment, enabling fast extraction by data type (e.g., retrieving adjustment instructions for a specific time period).

[0039] The storage medium uses the local solid-state drive (SSD) of the edge node, and is backed up daily to the cloud-based virtual memory area to ensure data security and long-term traceability.

[0040] Step 500, the process of building the initial progressive fault diagnosis model: retrieve the target device's time-series data from the past six months from the simulacrum unit, requiring it to cover the complete "no fault - slight wear - moderate wear" lifecycle (this must be confirmed in conjunction with the device maintenance log, such as including a single minor bearing wear repair record); extract the frequency domain data of the robot arm's vibration within each time window, focusing on the 50-200 Hz frequency band (based on research on the vibration characteristics of bearing wear, this frequency band is a wear-sensitive area), and splice it into a continuous sequence in chronological order (one data point per minute, including the average amplitude and main frequency values ​​within the 50-200 Hz window); extract instructions time-aligned with the vibration sequence, including "trigger timestamp, adjustment amplitude, and vibration offset area of ​​the corresponding window", and group them by "vibration window timestamp". Count the number of instructions and the average adjustment amplitude within each window (for example, "two instructions triggered within the window 08:00-08:01, average adjustment amplitude -5%"). Remove data from equipment downtime, calibration, and maintenance periods (identified through the "equipment status label" in the virtual unit, such as directly eliminating window data in the "downtime" state); if the amplitude of a window changes by more than 30% compared to the previous window (determined to be sensor interference), replace it with the average of the three previous and subsequent windows; establish a "vibration-instruction" time association table: that is, use the vibration window timestamp as the key to associate the vibration characteristics in the window with the adjustment instructions, ensuring that each vibration data can correspond to the intervention behavior of the same period (if there is no instruction, mark it as "no intervention").

[0041] Model input features and fault label definitions: Input features (all normalized to values ​​between 0 and 1) include core vibration features: average amplitude in the 50-200 Hz frequency range (F1), dominant frequency value (F2, such as 100 Hz), and daily amplitude growth rate (F3, i.e., average amplitude of the current day / average amplitude of the previous day - 1). Intervention-related features include the total number of welding speed adjustment instructions in the past three days (F4), average absolute value of the adjustment amplitude (F5), and average interval between two instructions (F6). Based on equipment maintenance logs and vibration characteristics, historical data is annotated into three states: State 1 (no wear): 50-200 Hz average amplitude < 120% of the historical fault-free period average, and no adjustment instructions; State 2 (mild wear): The average amplitude is 120%-150% of the historical mean, and adjustment instructions appear 1-2 times a week; State 3 (moderate wear): The average amplitude is 150%-200% of the historical mean, and adjustment instructions appear 3-5 times a week.

[0042] The system activates the "feature-state" co-occurrence analysis algorithm and automatically calculates the distribution range of each feature under different states: for 1000 state one data points, the system automatically calculates the proportion of samples with F1 < 0.3, F3 < 0.02, and F4 = 0 (92%), forming a feature association pattern for state one; for 800 state two data points, the system automatically calculates the proportion of samples with F1 between 0.3-0.5, F3 between 0.02-0.05, and F4 = 1-2 (88%), forming a feature association pattern for state two; the system automatically extracts the vibration data for state one for the previous three months from the preprocessed data, calculates the mean of the average amplitude of 50-200Hz (e.g., 0.04mm / s²), and stores it as the baseline value (B). By calculating the coefficient of variation of each feature in different states (i.e. first calculate the standard deviation and mean of all sample values ​​of each feature in state 1, state 2, and state 3 respectively, then calculate the coefficient of variation of the feature in each state according to "coefficient of variation = standard deviation / mean", and finally calculate the overall coefficient of variation of the feature in all states, which reflects the feature's ability to distinguish different states), the weight is automatically assigned: F1 (average amplitude): the highest coefficient of variation (0.42), automatically assigned weight 0.35; F3 (daily amplitude growth rate): the second highest coefficient of variation (0.31), automatically assigned weight 0.25; F4 (number of instructions): the coefficient of variation 0.28, automatically assigned weight 0.2; F2 (dominant frequency value): the coefficient of variation 0.15, automatically assigned weight 0.1; F5, F6: the coefficient of variation <0.1, automatically assigned weight 0.05 each; the total weight is automatically verified by the system to be 1, ensuring the self-consistency of the diagnostic logic.

[0043] Step 5010: For the vibration data of the "latest collection period" (preset to the last 30 minutes), the following process is automatically executed: 30 minutes of vibration frequency domain data are automatically retrieved from the real-time cache of the edge computing node and automatically segmented into 1 minute / frame (synchronized with the virtual unit time window). For each frame of data, the system automatically counts the amplitude values ​​of all frequency points in the 50-200Hz frequency band, extracting the "maximum amplitude" (envelope vertex) and "average amplitude" (overall energy) of this frequency band. The characteristic values ​​of each window are automatically generated in frequency order: "50-200Hz average amplitude (E1), maximum amplitude corresponding frequency (E2)" (for example, "window 09:00-09:01: E1=0.055mm / s², E2=110Hz"). The system automatically splices the features of 30 windows in ascending timestamp order to form the latest spectrum envelope feature sequence (with timestamps).

[0044] Step 5011: Call the historical spectrum benchmark stored in the initial model and automatically complete the following operations: Loading benchmark parameters: B1 = 0.04 mm / s² (50-200 Hz average amplitude benchmark), B2 = 100 Hz (maximum amplitude corresponding to frequency benchmark); for each window in the latest sequence, the offset is automatically calculated: average amplitude offset (D1) = (current E1 - B1) / B1 × 100% (positive values ​​indicate an increase above the benchmark); main frequency offset (D2) = current E2 - B2 (positive values ​​indicate a shift toward higher frequencies); according to the window timestamp sequence, the system automatically combines D1 and D2 to generate a "characteristic offset data sequence" (for example, "09:00-09:01: D1 = 37.5%, D2 = 10 Hz; 09:01-09:02: D1 = 40%, D2 = 12 Hz").

[0045] Step 5012: Automatically perform anomaly identification based on the correlation between feature offset and fault status: Load offset judgment threshold: D1 = 20% (the lowest offset in state two, i.e. 120% of B1), D2 = 5 Hz (a critical value based on wear characteristic statistics); the system automatically compares D1 and D2 in each window. When D1 > 20% and D2 > 5 Hz, it is automatically marked as an "abnormal window." The system automatically traverses the offset sequence. When three or more consecutive windows are "abnormal windows" and the average D1 value of these three windows is > 25%, it is automatically determined to be a "continuous abnormal period." The start / end window timestamp, duration, and average D1 / D2 value of each period are recorded (for example, "start 09:00-09:01, end 09:02-09:03, last 3 minutes, average D1 39%, average D2 11 Hz").

[0046] Step 5013: Dynamically adjust the weight of the vibration feature based on the duration of the overrun period to strengthen the influence of the abnormal feature. The process is as follows: Load the initial weights, including F1=0.35, F2=0.1 (inheriting the automatic allocation result of step 500), Automatically adjust according to preset logic: that is, for a duration of 1 minute (1 window), F1 weight = 0.35 + 0.35 × 10% = 0.385, F2 weight = 0.1 + 0.1 × 10% = 0.11; for a duration of 2 minutes (2 windows), F1 = 0.35 + 0.35 × 10% × 2 = 0.42, F2 = 0.1 + 0.1 × 10% × 2 = 0.12; for a duration of ≥ 5 minutes, the F1 upper limit is automatically locked at 0.5, and the F2 upper limit is locked at 0 .2 (to avoid excessive influence of features); in case of abnormal interruption, the weight automatically falls back to the initial value × 1.1 (for example, F1 = 0.35 × 1.1 = 0.385); the system traverses the latest 30 windows, collects the adjusted F1 and F2 weights corresponding to each window, and then calculates the arithmetic mean of the 30 F1 weight values ​​and the arithmetic mean of the 30 F2 weight values ​​respectively. The results are the "preliminary optimized weights" (for example, F1 = 0.42, F2 = 0.12).

[0047] In step 5014, the weight calibration is automatically completed in combination with the intervention frequency of the welding speed adjustment instruction: the system automatically extracts the adjustment instructions of the latest 30 minutes from the virtual unit, counts the total number of times (e.g., 6 times), and automatically calculates the intervention frequency = 6 times / 30 minutes = 0.2 times / minute; loads the baseline intervention frequency = 0.1 times / minute (the system automatically calculates the average based on historical normal production data), and automatically calculates it according to the formula "attenuation coefficient = 1 / (1+actual frequency / baseline frequency)" (when it is 0.2 times / minute, the attenuation coefficient = 1 / (1+0.2 / 0.1) = 1 / 3); automatically multiplies the preliminary optimization weight by the attenuation coefficient (e.g., F1 = 0.42×1 / 3 = 0.14, F2 = 0.12×1 / 3 = 0.04), and normalizes all feature weights (ensuring that the total is 1).

[0048] Step 5015, automatically perform parameter replacement and verification to complete the model update: automatically replace the corresponding parameters of the initial model with the updated weights of step 5014 (such as F1=0.14, F2=0.04), and retain other parameters (such as F3=0.25); the system automatically extracts the vibration characteristics of the "non-abnormal window" within the latest 30 minutes, and automatically calculates according to the formula "new benchmark = (old benchmark × 9 + latest non-abnormal feature mean) / 10" (such as old B1=0.04mm / s², latest non-abnormal E1 mean=0.042mm / s², new B1=0.0402mm / s²); automatically retrieve historical data from the past hour (including samples with known fault status, the status is determined by the setting The data is automatically annotated in the preparation log), and the data is input into the updated model to obtain the diagnosis results; the proportion of samples that are consistent with the diagnosis results of the statistical model and the actual status is calculated (for example, 55 out of 60 windows in 1 hour are consistent, and the accuracy rate is 55 / 60); if the accuracy rate is improved by ≥5% compared with the initial model (for example, from 85% to 92%), the update is confirmed to be effective; if the accuracy rate decreases (for example, from 85% to 82%), the baseline intervention frequency is automatically lowered (for example, from 0.1 times / minute to 0.08 times / minute), and steps 5014-5015 are re-executed; an "updated progressive fault diagnosis model" is generated, and the model parameters, update timestamp, and verification results are synchronously stored in the "model log" folder of the virtual unit.

[0049] Aggregating equipment entities and storing them in a time-series format solves the problem of fragmented and weakly correlated data, forming a complete timeline from sensor data to thresholds to commands. This provides high-quality, highly correlated historical data for model construction, while also improving data query and traceability efficiency. The model is constructed based on six months of historical equipment data, incorporating sensitive frequency bands and command intervention characteristics, enabling the initial model to accurately capture the characteristics of progressive failures in the equipment (e.g., wear vibration patterns of specific bearings). By comparing the latest vibration characteristics with historical benchmarks, weighting the duration of overshoots, and attenuating the intervention frequency, the model's decision parameters adapt in real time to changes in equipment status (e.g., increased wear, frequent interventions), maintaining a high recognition rate for progressive failures over the long term. Incorporating intervention information from welding speed control commands during model updates allows diagnostic results to reflect the impact of interventions on equipment status (e.g., whether vibration is alleviated after adjustment). This improves diagnostic practicality and provides a more practical basis for predictive maintenance.

[0050] In a preferred embodiment of the present invention, step 6 includes: Step 600 , using the updated progressive fault diagnosis model to process the historical vibration frequency value sequence, extracting the vibration spectrum envelope characteristic values ​​of continuous time stamps, and generating a characteristic evolution sequence sorted by acquisition time; Step 601: Locate the triggering time node of the welding speed adjustment instruction in the feature evolution sequence and generate a feature sequence with a time mark; Step 602, analyzing the temporal and spatial correlation between the degree of feature value offset and the adjustment instruction in the time-stamped feature sequence, and generating a predictive diagnosis result of the wear state of the robot arm bearing, specifically includes: Step 6020: Calculate the characteristic value change gradient of adjacent time nodes based on the time-stamped characteristic sequence to generate a characteristic gradient sequence; Step 6021, performing time window integration processing on the characteristic gradient sequence to generate a cumulative offset; Step 6022: locate the corresponding segment in the accumulated offset according to the triggering time node of the welding speed adjustment instruction, and extract the offset mean of the corresponding segment; Step 6023, calculating the bearing wear compensation coefficient by combining the offset mean value and the adjustment range of the welding speed adjustment command; Step 6024: Generate a predictive diagnosis result of the wear status of the robot arm bearing based on the accumulated offset and the bearing wear compensation coefficient; Step 603: Outputting a robot arm vibration diagnosis report including a time-stamped feature sequence, correlation analysis results, and wear prediction results.

[0051] In an embodiment of the present invention, historical data is processed by an updated progressive fault diagnosis model to extract the time-varying patterns of vibration characteristics. The specific process is as follows: The target device's time series data from the past seven days (covering a sufficient time span to reflect progressive changes) was retrieved from the simulacrum unit. The focus was on extracting the robotic arm's vibration frequency value sequence (one time window per minute, for a total of 7 × 24 × 60 = 10,080 windows). Based on the optimized sensitive frequency band (50-200 Hz) in the updated progressive fault diagnosis model, two types of spectral envelope feature values ​​were extracted from the vibration frequency domain data in each time window: the average amplitude of the 50-200 Hz frequency band (denoted as Favg, reflecting overall vibration energy); and the frequency value with the maximum amplitude within this band (denoted as Fpeak, reflecting the location of the dominant vibration frequency). The extracted Favg and Fpeak values ​​were sorted in ascending timestamp order (from the first window seven days ago to the most recent window), forming a structured sequence of "timestamp-Favg-Fpeak," also known as the "feature evolution sequence."

[0052] Step 601: Associate the welding speed adjustment instruction with the characteristic evolution sequence to identify the nodes where the intervention behavior affects the vibration characteristics. The specific process is as follows: From the "welding speed adjustment instruction" records of the virtual unit, extract the trigger timestamps (accurate to minutes) and adjustment ranges (such as -5%, -8%) of all instructions in the past 7 days to form a "instruction timestamp-adjustment range" list (for example: "2024-09-01 08:30--5%; 2024-09-01 14:15--8%").

[0053] Traverse the feature evolution sequence and find a time window that exactly matches the command timestamp (or an error of ≤1 minute). Add a command tag after the feature value in that window. For example, if the command corresponding to 2024-09-01 08:30 in the feature sequence is "08:30-0.05mm / s²-105Hz," it will be marked as "08:30-0.05mm / s²-105Hz-Command (-5%)." Unmarked window features are retained, ultimately forming a "time-stamped feature sequence" that not only captures the continuous vibration feature changes but also clearly identifies the trigger location and intensity of the intervention command.

[0054] Step 6020 reflects the degradation rate of the vibration characteristics by calculating the characteristic change rate of adjacent time nodes. The specific process is as follows: For time-stamped feature sequences, select two consecutive time windows (with a 1-minute interval between adjacent windows), for example, window t1 (08:30) and window t2 (08:31). For the average amplitude (Favg) of the 50-200 Hz frequency band: gradient Gavg = (Favg of window t2 - Favg of window t1) / time interval (1 minute). A positive value indicates an increase (deterioration) in amplitude, while a negative value indicates a decrease (improvement). For the frequency value with the largest amplitude in the 50-200 Hz frequency band (Fpeak): the gradient Gpeak = (Fpeak of the t2 window - Fpeak of the t1 window) / 1 minute. A positive value indicates that the main frequency is shifting toward high frequency (a typical feature of increased wear). The Gavg and Gpeak of all adjacent windows are concatenated in chronological order to form a "characteristic gradient sequence", for example: "08:30-08:31: Gavg = 0.002 mm / s² / minute, Gpeak = 1 Hz / minute; 08:31-08:32: Gavg = 0.001 mm / s² / minute, Gpeak = 0.5 Hz / minute".

[0055] Step 6021 quantifies the total feature offset within a period of time by integrating the time window of the gradient sequence. The specific process is as follows: Using a sliding time window (1-hour window size, i.e., 60 adjacent windows) with a sliding step of 10 minutes (ensuring coverage of all time periods and overlap to avoid missing information), the characteristic gradient sequence within each time window is accumulated: the cumulative Favg offset Savg = the sum of all Gavg values ​​within the window (reflecting the total increase in average amplitude over the 1-hour period); the cumulative Fpeak offset Speak = the sum of all Gpeak values ​​within the window (reflecting the total offset of the dominant frequency over the 1-hour period). The windows are sorted by their start time to generate a cumulative offset sequence of "time window - Savg - Speak", for example: "08:00-09:00: Savg = 0.1 mm / s², Speak = 50 Hz; 08:10-09:10: Savg = 0.11 mm / s², Speak = 52 Hz".

[0056] Step 6022: Calculate the cumulative offset of the positioning adjustment instruction trigger period and analyze the vibration changes before and after the intervention. The specific process is as follows: For each welding speed adjustment instruction (e.g., triggered at 08:30), its corresponding segment is defined as "30 minutes before the instruction to 30 minutes after the instruction" (a total of one hour, covering the entire impact period before and after the intervention), namely 08:00-09:00. The window corresponding to this time period (08:00-09:00) is found in the cumulative offset sequence, and its Savg and Speak values ​​are extracted. If the time period spans multiple windows (e.g., if the instruction is triggered at 08:55, the segment is 08:25-09:25, covering the 08:20-09:20 and 08:30-09:30 windows), the average of the Savg and Speak values ​​across these windows is taken as the segment offset. The arithmetic mean of the extracted segment offsets (e.g., Savg = 0.1mm / s² and 0.12mm / s²) is calculated as the "offset mean" for the segment corresponding to the instruction.

[0057] Step 6023, calculation of bearing wear compensation coefficient: This coefficient is used to quantify the effect of the welding speed adjustment instruction on the alleviation of bearing wear. The specific process is as follows: The larger the adjustment amplitude (e.g., the greater the speed reduction), the stronger the vibration suppression effect and the larger the compensation coefficient (indicating the percentage of wear reduction due to intervention). Based on the commanded adjustment amplitude (taken as its absolute value, e.g., -5% represents a 5% adjustment amplitude), a positive correlation is established between the amplitude and the compensation coefficient: compensation coefficient K = adjustment amplitude × 0.02 (based on the historical data trend that "every 1% increase in adjustment amplitude reduces the wear rate by approximately 2%)." K is then modified based on the mean offset value extracted in step 6022 (e.g., Savg mean = 0.11 mm / s²). If the mean offset value exceeds the historically normal offset value for the same period (e.g., 0.08 mm / s²), the intervention effect is limited, and K is multiplied by 0.8. If it is ≤ 0.08 mm / s², the intervention is effective, and K remains unchanged. For example, if the adjustment amplitude is 5% and the mean offset is 0.11 mm / s² (>0.08), then K = 5% × 0.02 × 0.8 = 0.0008 (i.e., 0.08%).

[0058] Step 6024: The accumulated offset and compensation coefficient are combined to determine the current wear state and predict future trends. The specific process is as follows: Take the total accumulated offset value (Savg) for the last 24 hours (e.g., 2.5 mm / s²) and subtract the total compensation coefficient for all adjustment commands during the same period (e.g., 0.3) to obtain the actual wear correlation value (2.2 mm / s²). Based on historical data from the updated progressive fault diagnosis model, a corresponding relationship between the actual wear correlation value and the wear level is established: 0-1.5 mm / s² represents State 1 (no significant wear); 1.5-3 mm / s² represents State 2 (slight wear); and 3 mm / s² represents State 3 (moderate wear). For example, an actual value of 2.2 mm / s² represents State 2. Calculate the actual wear-related value growth rate over the past seven days (e.g., an increase of 0.3 mm / s² per day). Combined with the current value (2.2), predict the time it will take to reach level 2 (3 mm / s²): (3 - 2.2) / 0.3 ≈ 2.7 days. This means that moderate wear is likely in about three days. This generates a predictive diagnostic result that includes "Current Wear Level (State 2), Forecast Level for the Next Three Days (State 3), and Recommended Maintenance Time (Within Three Days)."

[0059] Step 603: Integrate the above analysis results into an intuitive report. The specific content and format are as follows: The Favg and Fpeak values ​​at key time points (e.g., 0:00 a.m. each day, when the instruction was triggered) for the past seven days are displayed in a table, with the instruction tag annotated (e.g., "08:30-instruction-5%"). The "Favg change within one hour after the instruction was triggered" is displayed graphically (e.g., Gavg drops from 0.002 to 0.001 after the instruction, indicating that the intervention was effective). The mean compensation coefficients under different adjustment ranges are calculated (e.g., K averages 0.0007 at a 5% range and 0.0012 at a 8% range) to illustrate the relationship between intervention intensity and effect. The current level, future trend curve (marking the time point when level 2 is reached), and recommended maintenance measures (e.g., "Stop the machine and replace the joint bearing within three days") are presented in the form of text and charts. The report is generated in PDF format, named according to "equipment identification-date" (e.g., "welding robot-A3 workshop-012-20240908.pdf"), stored synchronously in the "diagnostic report" folder of the virtual unit, and pushed to the equipment management platform.

[0060] The feature evolution sequence intuitively displays the continuous change of vibration characteristics (amplitude, main frequency) over time, making equipment degradation trends (such as a slow increase in amplitude and a shift of main frequency toward high frequency) clear at a glance, facilitating the rapid identification of early signs of progressive failures. The time-stamped feature sequence clearly marks the triggering node of the adjustment instruction, allowing direct observation of the instruction's impact on the vibration characteristics (such as a slowdown in amplitude growth after the instruction), providing a basis for evaluating the effectiveness of interventions. Through gradient calculation, cumulative offset analysis, and compensation coefficient correction, the vibration characteristics are quantitatively associated with the wear state, which not only accurately determines the current wear level but also predicts future degradation time, providing a precise time window for predictive maintenance and reducing unplanned downtime. The report integrates the feature sequence, correlation analysis, and prediction results, with comprehensive and intuitive content, improving equipment management efficiency.

[0061] like Figure 2 As shown, an embodiment of the present invention further provides a machine learning system, comprising: An acquisition module is used to collect real-time sensor data streams from industrial equipment at edge computing nodes. The data streams include the temperature value of the welding robot end, the current value of the drive motor, and the vibration frequency value of the robotic arm. A threshold generation module is used to extract the characteristic values ​​of the starting point, path midpoint, and end point in the welding trajectory through the edge intelligent processing unit based on the real-time sensor data stream, and dynamically generate anomaly judgment thresholds based on the characteristic values; A control module is used to trigger a power cut-off instruction when the driving motor current value exceeds the threshold value based on an abnormality determination threshold, or to generate a welding speed adjustment instruction when the vibration frequency value of the robot arm deviates from the normal range; The storage module is used to aggregate the sensor data stream, abnormality judgment threshold and welding speed adjustment instructions into time-series quasi-physical units for storage according to the equipment entity; An updating module is used to construct a progressive fault diagnosis model and update the decision parameters of the progressive fault diagnosis model based on the historical vibration frequency value sequence and the corresponding control instruction record in the pseudo-physical unit to obtain an updated progressive fault diagnosis model; The report output module is used to output a report on the evolution sequence of vibration characteristics of the robotic arm based on the updated progressive fault diagnosis model.

[0062] It should be noted that this system is a system corresponding to the above method, and all implementation methods in the above method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0063] An embodiment of the present invention further provides a computing device comprising: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the above-described method. All implementations in the above-described method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0064] The embodiment of the present invention further provides a computer-readable storage medium storing instructions, which, when executed on a computer, causes the computer to execute the above-described method. All implementations in the above-described method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0065] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A machine learning method, characterized in that The method comprises: Step 1: Collect real-time sensor data streams of industrial equipment at the edge computing node, including the temperature value of the welding robot end, the current value of the drive motor, and the vibration frequency value of the robotic arm; Step 2: Based on the real-time sensor data stream, the edge intelligent processing unit extracts the characteristic values ​​of the starting point, path midpoint, and end point in the welding trajectory, and dynamically generates an abnormality judgment threshold based on the characteristic values; Step 3: Based on the abnormality determination threshold, a power cut-off instruction is triggered when the driving motor current value exceeds the threshold, or a welding speed adjustment instruction is generated when the vibration frequency value of the robot arm deviates from the normal range; Step 4: Aggregate the sensor data stream, abnormality judgment threshold, and welding speed adjustment instruction into time-series quasi-physical units according to the equipment entity and store them; Step 5: Based on the historical vibration frequency value sequence and the corresponding control instruction record in the pseudo-physical unit, a progressive fault diagnosis model is constructed and the decision parameters of the progressive fault diagnosis model are updated to obtain an updated progressive fault diagnosis model; Step 6: Based on the updated progressive fault diagnosis model, output a report on the evolution sequence of the vibration characteristics of the robotic arm.

2. The machine learning method according to claim 1, wherein: The step 2 includes: Based on the real-time sensor data stream, the temperature change rate characteristics of the starting point of the welding trajectory, the current distribution statistical characteristics of the midpoint of the welding path, and the vibration frequency domain envelope characteristics of the welding end point are extracted; The temperature change rate feature is combined with the real-time welding speed parameter to perform reverse weight adjustment to generate the temperature change weighted feature; the current distribution statistical feature is combined with the real-time welding material thickness parameter to perform forward weight adjustment to generate the current distribution weighted feature; the vibration frequency domain envelope feature is combined with the equipment cumulative working time parameter to perform forward weight adjustment to generate the vibration frequency domain weighted feature; The abnormality judgment threshold is dynamically generated by fusing the weighted features of temperature change, current distribution and vibration frequency domain.

3. The machine learning method according to claim 2, wherein: The step 3 comprises: The driving motor current value is compared with the abnormality judgment threshold in real time. When the current value continues to exceed the abnormality judgment threshold within continuous sampling cycles, the power cut-off command is triggered; the envelope offset area of ​​the vibration spectrum and the abnormality judgment threshold is calculated based on the vibration frequency value of the robot arm; when the envelope offset area exceeds the preset tolerance threshold, a welding speed adjustment command is generated, and the adjustment amplitude of the welding speed adjustment command increases with the increase of the envelope offset area.

4. The machine learning method according to claim 3, wherein: The step 5 comprises: An initial progressive fault diagnosis model is constructed using the historical vibration frequency value sequence and the corresponding welding speed adjustment instruction record in the time-series simulacrum unit. The vibration frequency value sequence and welding speed adjustment instruction record of the latest acquisition cycle are obtained, the latest data are input into the initial progressive fault diagnosis model, the vibration characteristic decision parameters are updated, and an updated progressive fault diagnosis model is generated.

5. The machine learning method according to claim 4, wherein: Obtain the vibration frequency value sequence and welding speed adjustment instruction record of the latest acquisition cycle, input the latest data into the initial progressive fault diagnosis model, update the vibration characteristic decision parameters, and generate an updated progressive fault diagnosis model, including: Obtain the spectrum envelope characteristic value of the latest vibration frequency value sequence; Comparing the spectrum envelope characteristic value with the historical spectrum benchmark in the initial progressive fault diagnosis model to generate a characteristic offset data sequence; Processing the characteristic offset data sequence, identifying sampling periods that continuously exceed the offset threshold, and generating an excess period marker sequence; Based on the duration of the excessive time period mark sequence, the vibration characteristic weight coefficient is adjusted to generate a preliminary correction weight coefficient; The number of interventions in the welding speed adjustment instruction record is extracted, and the preliminary correction weight coefficient is attenuated based on the intervention frequency to generate updated vibration characteristic decision parameters; The updated vibration characteristic decision parameters are written into the initial progressive fault diagnosis model to generate an updated progressive fault diagnosis model.

6. The machine learning method according to claim 5, wherein: The step 6 comprises: An updated progressive fault diagnosis model is used to process the historical vibration frequency value sequence, extract the vibration spectrum envelope feature values ​​of continuous time stamps, and generate a feature evolution sequence sorted by acquisition time; Locate the triggering time node of the welding speed adjustment instruction in the feature evolution sequence and generate a feature sequence with time mark; Analyze the spatiotemporal correlation between the degree of characteristic value offset and the adjustment instructions in the time-stamped characteristic sequence to generate predictive diagnostic results for the wear status of the robot arm bearing; Outputs a robot arm vibration diagnosis report containing time-stamped feature sequences, correlation analysis results, and wear prediction results.

7. The machine learning method according to claim 6, wherein: Analyze the temporal and spatial correlation between the degree of characteristic value offset and the adjustment instructions in the time-stamped characteristic sequence to generate predictive diagnostic results for the wear status of the robot arm bearing, including: Based on the time-stamped feature sequence, the feature value change gradient of adjacent time nodes is calculated to generate a feature gradient sequence; Perform time window integration processing on the characteristic gradient sequence to generate cumulative offset; Locate the corresponding section in the accumulated offset according to the triggering time node of the welding speed adjustment instruction, and extract the offset mean value of the corresponding section; The bearing wear compensation coefficient is calculated by combining the mean value of the offset and the adjustment range of the welding speed adjustment instruction; Based on the accumulated offset and the bearing wear compensation coefficient, a predictive diagnosis result of the wear status of the robot arm bearing is generated.

8. A machine learning system, which implements the method according to any one of claims 1 to 7, characterized in that: include: An acquisition module is used to collect real-time sensor data streams from industrial equipment at edge computing nodes. The data streams include the temperature value of the welding robot end, the current value of the drive motor, and the vibration frequency value of the robotic arm. A threshold generation module is used to extract the characteristic values ​​of the starting point, path midpoint, and end point in the welding trajectory through the edge intelligent processing unit based on the real-time sensor data stream, and dynamically generate anomaly judgment thresholds based on the characteristic values; A control module is used to trigger a power cut-off instruction when the driving motor current value exceeds the threshold value based on an abnormality determination threshold, or to generate a welding speed adjustment instruction when the vibration frequency value of the robot arm deviates from the normal range; The storage module is used to aggregate the sensor data stream, abnormality judgment threshold and welding speed adjustment instructions into time-series quasi-physical units for storage according to the equipment entity; An updating module is used to construct a progressive fault diagnosis model and update the decision parameters of the progressive fault diagnosis model based on the historical vibration frequency value sequence and the corresponding control instruction record in the pseudo-physical unit to obtain an updated progressive fault diagnosis model; The report output module is used to output a report on the evolution sequence of vibration characteristics of the robotic arm based on the updated progressive fault diagnosis model.

9. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images