A multi-parameter fusion-based cobalt salt packaging machine equipment predictive maintenance system

Abstract

The application discloses a kind of cobalt salt packing machine equipment predictive maintenance systems based on multi-parameter fusion, it is related to industrial equipment predictive maintenance technical field, including signal acquisition unit, for collecting first sensing signal and temperature sensing signal;Processing module, for: when first sensing signal exceeds threshold value, according to current first sensing signal generates prediction temperature sequence;Locate temperature change rate extreme point in prediction temperature sequence, and extract verification interval based on extreme point;Matching is carried out between the prediction temperature sequence in verification interval and actual acquisition temperature sequence;If matching degree exceeds pre-setting signal threshold value, then determine real degradation, and determine first remaining life based on the degradation characteristics of real degradation stage.The application realizes reliably to the remaining life of packing machine equipment and is predicted in time.

Description

Technical Field

[0001] This invention relates to the field of predictive maintenance technology for industrial equipment, and in particular to a predictive maintenance system for packaging machine equipment based on multi-parameter fusion. Background Technology

[0002] Cobalt borylate is a substance composed of cobalt ions and acid radicals, and may also contain a certain proportion of other ions. Cobalt borylate is used to enhance the adhesion between rubber and metal, and has been widely recognized and applied in products such as radial tires and steel wire hoses. It is a rubber adhesive. Cobalt borylate packaging equipment is used to fill and package packaging barrels. As the main drive servo motor of the cobalt borylate packaging equipment, due to its requirements for high positioning accuracy and fast response speed, problems such as jamming and abnormal noise may occur. Therefore, how to improve the prediction of its remaining life is a major technical challenge.

[0003] Existing technologies typically predict the lifespan of servo motors based on data collected by sensors. For example, patent publication number "CN113505535B" discloses a method for predicting the lifespan of a motor based on a gated adaptive hierarchical attention unit network. This method collects vibration signals throughout the entire lifespan of a servo motor and predicts the remaining lifespan of the servo motor by processing the vibration signals.

[0004] However, signals collected by sensors, such as vibration or current signals, are prone to false faults over a period of time due to external environmental influences. For example, during high-speed operation of a packaging machine, external mechanical vibrations are transmitted to the motor housing, causing the vibration sensor to mistakenly collect abnormal signals not originating from the motor itself; or sudden temperature changes in the production workshop can cause variations in the motor winding resistance, leading to current signal drift. If false faults are not effectively identified, the predictive model may prematurely trigger maintenance commands, increasing downtime costs, or it may ignore genuine degradation signals, causing sudden failures.

[0005] Therefore, how to reliably predict the remaining lifespan of packaging machines in a timely manner has become an urgent technical problem to be solved. Summary of the Invention

[0006] The technical problem solved by this invention is that it is difficult to reliably predict the remaining lifespan of packaging equipment in a timely manner.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a predictive maintenance system for cobalt salt packaging machines based on multi-parameter fusion, comprising: a signal acquisition unit for acquiring a first sensing signal and a temperature sensing signal; and a processing module for: generating a predicted temperature sequence based on the current first sensing signal when the first sensing signal exceeds a threshold; locating extreme points of temperature change rate in the predicted temperature sequence and extracting a verification interval based on the extreme points; matching the predicted temperature sequence within the verification interval with the actual acquired temperature sequence; and determining a first remaining lifetime based on the degradation characteristics of the actual degradation stage if the matching degree exceeds a preset threshold.

[0008] Preferably, the signal acquisition unit is used to simultaneously acquire at least two heterogeneous sensing signals, the heterogeneous sensing signals including a first sensing signal and a temperature sensing signal, wherein the first sensing signal is a vibration signal reflecting the mechanical state of the motor or a current signal reflecting the electrical state.

[0009] Preferably, the processing module includes: a data preprocessing unit, configured to use the temperature sensing signal as a time reference, align the sampling timestamps of the first sensing signal and the temperature sensing signal using interpolation, and perform noise filtering on the first sensing signal and the temperature sensing signal to generate standardized time-series data; an anomaly triggering unit, configured to detect whether the first sensing signal experiences abnormal fluctuations through a dynamic threshold, and generate a lifetime prediction start command when an anomaly is triggered; and a cross-modal prediction unit, which, in response to the lifetime prediction start command, performs the following operations: based on a pre-trained time-series prediction model, according to the current... The first sensing signal generates a predicted temperature sequence for the next N sampling periods. The time-series prediction model is trained using the coupling relationship between the first sensing signal and the temperature signal in historical data. Extreme points of temperature change rate are located in the predicted temperature sequence, and a verification interval is extracted based on the critical time window of the extreme points. The predicted temperature sequence within the verification interval is matched with the actual collected temperature sequence. If the matching degree exceeds a preset threshold, it is determined to be a real degradation event. A lifetime decision unit is used to determine the first remaining lifetime by fusing the multidimensional degradation features of the first sensing signal and the temperature signal after determining it to be a real degradation event.

[0010] Preferably, the time-series prediction model in the cross-modal prediction unit is a bidirectional long short-term memory network based on an attention mechanism. The input layer of the time-series prediction model receives the first signal sliding window data of the previous T periods at the current time, and the output layer generates the predicted temperature sequence for the next N periods, where T>N. The network weights are trained by the dynamic coupling relationship between the first signal and the temperature signal in historical data.

[0011] Preferably, locating the extreme point of the rate of temperature change in the predicted temperature sequence and extracting the verification interval based on the critical time window of the extreme point includes: determining the first-order difference sequence of the predicted temperature sequence, locating the time point with the largest difference value as the starting point of the first verification window; using the starting point of the first verification window as a reference, extending forward by t1 periods and backward by t2 periods to generate the verification interval, wherein t1 < t2, t1 = τ / 3, t2 = 2τ, and τ is the thermal inertia time constant of the packaging machine equipment; or, locating the extreme point of the rate of temperature change in the predicted temperature sequence and extracting the verification interval based on the critical time window of the extreme point includes: determining the second-order difference sequence of the predicted temperature sequence, locating the curvature abrupt change point as the starting point of the second verification window; using the starting point of the second verification window as a reference, extending forward by t3 periods and backward by t3 periods to generate the verification interval, wherein t3 is the average value of t1 and t2.

[0012] Preferably, the step of matching the predicted temperature sequence and the actual collected temperature sequence within the verification interval, and determining it as a real degradation event if the matching degree exceeds a preset confidence threshold, includes: standardizing the predicted temperature sequence and the actual temperature sequence within the verification interval; constructing a cumulative distance matrix based on the standardized predicted temperature sequence and the actual temperature sequence; determining the minimum cumulative distance between the predicted temperature sequence and the actual temperature sequence based on the cumulative distance matrix; and determining a successful match when the minimum cumulative distance is less than the benchmark distance of similar faults in a preset degradation mode library.

[0013] Preferably, the signal acquisition unit is further configured to synchronously acquire the second sensing signal; the cross-modal prediction unit is further configured to verify, after determining that a real degradation event has occurred, whether the second sensing signal exhibits synchronous abnormal fluctuations within the critical time window; the lifetime decision unit is further configured to: after verifying that the second sensing signal exhibits synchronous abnormal fluctuations, count the number of synchronous abnormalities and the total number of sampling points of the second sensing signal within the critical time window; determine the remaining lifetime correction value based on the number of synchronous abnormalities, the total number of sampling points, and a preset lifetime decay coefficient; determine the second remaining lifetime based on the first remaining lifetime and the remaining lifetime correction value; and determine the second remaining lifetime as the remaining lifetime of the packaging machine equipment.

[0014] Preferably, the processing module further includes: a fault mode learning unit, which stores the phase difference features of vibration-current-temperature signals under preset typical faults, used to extract the phase difference features corresponding to the current vibration-current-temperature signal in the actual degradation stage, determine the closest fault mode through cosine similarity matching, and load the corresponding lifetime decay coefficient according to the matched fault mode.

[0015] The beneficial effects of this invention are as follows: When the first signal is abnormal, the future temperature sequence is predicted based on the first signal, and a verification point is selected from the time corresponding to the maximum rate of change in the future temperature sequence from the current time. The real-time temperature signal corresponding to the verification point is verified based on the verification point to eliminate false faults, thereby improving the reliability of the remaining lifespan of the packaging machine. The verification point is selected from the time corresponding to the maximum rate of change in the future temperature sequence from the current time, which can promptly distinguish whether the abnormal first signal indicates a real fault or a false fault in the packaging machine. If the verification is successful, the remaining lifespan of the packaging machine is predicted based on the real-time first signal. This allows for timely and reliable prediction of the remaining lifespan of the packaging machine when a real fault occurs. Attached Figure Description

[0016] Figure 1 A schematic diagram of the structure of a predictive maintenance system for packaging machine equipment based on multi-parameter fusion provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the basic process of a cross-modal prediction method provided in one embodiment of the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Some data in the embodiments of this application have been anonymized.

[0018] Example 1, referring to Figure 1 As an embodiment of the present invention, a predictive maintenance system for cobalt salt ( ) packaging machine equipment based on multi-parameter fusion is provided. The system 1 includes: The signal acquisition unit 10 is used to simultaneously acquire at least two heterogeneous sensing signals, including a first sensing signal and a temperature sensing signal. The first sensing signal is either a vibration signal reflecting the mechanical state of the motor or a current signal reflecting the electrical state.

[0019] Vibration sensors acquire vibration signals using a piezoelectric accelerometer (model PCB 352C33, range ±50g, frequency response 0.5-10kHz), mounted vertically on the non-drive end bearing housing of the motor to avoid signal attenuation caused by axial mounting. A magnetic base ensures tight contact between the sensor and the housing. The resonant frequency is >15kHz, and sensitivity is calibrated using a standard vibration table (frequency 10Hz-5kHz, acceleration 1g), with an error <±3%. Current sensors acquire current signals using a Hall effect closed-loop sensor (LEM LAH). The current sensor (100-P, accuracy ±0.5%) is wired by connecting three phase currents to three independent Hall sensors, with the neutral line grounded. The sensor output signal is transmitted through shielded twisted-pair cable, and the shielding layer is grounded at one end. The temperature sensor collects temperature signals (temperature status signals), using a PT100 platinum resistance thermometer (four-wire connection, accuracy ±0.1℃), embedded at the end of the motor stator winding, 10cm away from the coil connector to avoid direct contact with the copper conductor. The sensor surface is coated with thermally conductive silicone grease and wrapped with a ceramic fiber insulation layer to reduce the influence of ambient temperature.

[0020] The synchronous acquisition mechanism of the signal acquisition unit 10 can be achieved through sampling frequency, time scale alignment, etc. For example, sampling frequencies can be set to 10kHz for vibration signals, 2kHz for current signals, and 1Hz for temperature signals; microsecond-level time synchronization can be achieved using PTP (Precision Clock Protocol); the temperature signal is compensated for by a first-order hysteresis model (τ=3 seconds) due to response delay. The first-order hysteresis model is as follows: , where the time constant =3 seconds, calibrated through a step response experiment, so that the temperature signal response delay is reduced from 5 seconds to within 0.5 seconds.

[0021] When processing the acquired vibration signal, the RMS value of the rolling window (window length 200ms, step size 50ms) can be calculated to reduce the data dimensionality and obtain the vibration signal transmitted to the processing module 20. When processing the current signal, the effective value per cycle (50Hz) can be calculated, and the harmonic distortion rate (THD) and the amplitude of the 5th and 7th harmonics can be extracted simultaneously to obtain the current signal transmitted to the processing module 20. When processing the temperature signal, the influence of ambient temperature (reference temperature = 25℃) can be removed, the temperature rise ΔT can be calculated, and the temperature signal transmitted to the processing module 20 can be obtained.

[0022] The processing module 20 includes a data preprocessing unit 21 to a lifetime decision unit 24, and may further include a fault mode learning unit 25.

[0023] The data preprocessing unit 21 is used to align the sampling timestamps of the first sensing signal and the temperature sensing signal with the temperature sensing signal as the time reference by interpolation, and to filter the first sensing signal and the temperature sensing signal and noise to generate standardized time series data.

[0024] Using the temperature signal as a reference, the time axes of vibration and current signals are aligned by cubic spline interpolation. Wavelet thresholding is applied to the vibration signal (Symlet wavelet basis, decomposition level 5), and Kalman filtering is applied to the current signal to eliminate random fluctuations.

[0025] The abnormal triggering unit 22 is used to detect whether the first sensing signal has abnormal fluctuations through a dynamic threshold, and to generate a lifetime prediction start command when an abnormality is triggered.

[0026] For example, if the RMS value exceeds twice the baseline for 5 consecutive cycles (baseline = historical average of normal data), vibration abnormality can be determined. The RMS mean and standard deviation are calculated hourly for a rolling window (window length 1 hour), and the baseline value = mean + 3σ. If the harmonic distortion rate (THD) exceeds 5% for 3 consecutive cycles, current abnormality can be determined. The THD threshold is dynamically adjusted according to the load rate (when the load rate is >80%, the THD threshold is relaxed to 7%). Any abnormal signal generates a lifetime prediction command, and the abnormal signal is identified as the first sensing signal, and the other signal is identified as the second sensing signal. If both are abnormal, the vibration signal can be identified as the first sensing signal first, and S110~S130 can be executed. Then, the current signal can be identified as the first sensing signal, and S110~S130 can be executed. Alternatively, the one with more significant abnormal fluctuation can be identified as the first sensing signal.

[0027] The cross-modal prediction unit 23, in response to the lifetime prediction initiation command, executes the cross-modal prediction method. For details, please refer to [link to relevant documentation]. Figure 2 The cross-module prediction method includes S110~S130.

[0028] S110, based on a pre-trained time-series prediction model, generates a predicted temperature sequence for the next N sampling periods based on the current first sensing signal. The time-series prediction model is trained by the coupling relationship between the first sensing signal and the temperature signal in historical data.

[0029] Specifically, the time-series prediction model in the cross-modal prediction unit 23 is a bidirectional long short-term memory network based on the attention mechanism. The input layer of the time-series prediction model receives the first signal sliding window data of the previous T periods at the current time, and the output layer generates the predicted temperature sequence for the next N periods, where T>N. The network weights are trained by the dynamic coupling relationship between the first signal and the temperature signal in the historical data.

[0030] The model framework of the time series prediction model is as follows: Input layer, which receives the sliding window data of the first sensing signal (vibration / current); Bi-LSTM layer, which is a bidirectional long short-term memory network with 64 hidden layers and a dropout rate of 0.2; Attention layer: calculates the weights of the time steps; Output layer, which generates the predicted temperature sequence for the future.

[0031] The following is an example of a time series prediction model structure.

[0032] The input layer consists of vibration signals that have been denoised and RMS-calculated as standardized time-series data, with a dimension of 50×1 (50-second sliding window). The input current signals are the fundamental RMS value and harmonic characteristics (5th and 7th harmonic amplitudes), with a dimension of 50×3. The input data is Z-score standardized, and the sliding window is configured with a window length of T=50 seconds, a sliding step size of 1 second, and an overlap rate of 98% to ensure temporal continuity.

[0033] Bi-LSTM (Bidirectional Long Short-Term Memory) layer: The network structure includes hidden layers, gating mechanisms, and Dropout mechanisms. The hidden layer is configured as a single-layer bidirectional LSTM, with 64 dimensions of hidden units in each direction, resulting in a total output dimension of 128. The gating mechanism uses Sigmoid as the activation function for the input, forget, and output gates, and the candidate memory unit activation function can be set to Tanh. The Dropout mechanism adds a Dropout layer after the LSTM layer output, with a dropout rate of 0.2%.

[0034] The temporal information of the Bi-LSTM layer is transmitted by processing the temporal data sequentially from t=0 to t=49 through a forward LSTM; then processing the temporal data in reverse order from t=49 to t=0 through a backward LSTM; and finally concatenating the forward and backward hidden states at each time step to form a 128-dimensional feature vector.

[0035] The attention layer trains the weight matrix and bias based on the 128-dimensional feature vector sequence (50 time steps × 128 dimensions) output by Bi-LSTM.

[0036] The output layer consists of a fully connected layer and an activation function. The fully connected layer takes a context vector c (128 dimensions) as input and outputs a dimension of 30 (corresponding to the prediction for the next 30 seconds). It uses linear activation (no activation function) and directly outputs the predicted temperature value. The output layer's output dimension is 30×1, representing the predicted temperature value per second for the next 30 seconds. The output layer also denormalizes the predicted values ​​to restore the actual temperature dimensions.

[0037] The model was trained using data from 100 identical packaging machines, running for 3 months to 1 year, covering the following operating conditions: normal operation (60% of data), bearing wear (20% of data, wear severity categorized as mild, moderate, and severe), winding short circuit (15% of data, short-circuit resistance 1Ω~5Ω), and heat dissipation failure (5% of data, fan failure or duct blockage). The collected data was divided into three sets: 70 sets of packaging machine data (70%) for training, 15 sets of packaging machine data (15%) for validation, and 15 sets of motor data (15%) for testing. The actual temperature sequence (1Hz sampling) for the next 30 seconds was used as the label data, and vibration / current signals were labeled for each 50-second window. Furthermore, data augmentation can be performed, for example, by adding Gaussian noise with μ=0 and σ=0.1 to the input signal, injecting pulse noise with an amplitude of ±0.5 at random time points, simulating sensor transient faults, randomly stretching or compressing the timing length (±10%), enhancing the model's adaptability to variable speed conditions, randomly sampling a 45~55 second sub-window from the original 50 second window and filling it to 50 seconds (zero padding at the edges), etc.

[0038] By capturing forward and backward dependencies through Bi-LSTM, strengthening key features through attention mechanism, and improving generalization through Dropout, the test set error is reduced by 40%, and the model is lightweight (less than 100,000 parameters), making it suitable for real-time inference on embedded devices.

[0039] S120 locates the extreme points of temperature change rate in the predicted temperature sequence and extracts the verification interval based on the critical time window of the extreme points.

[0040] In one alternative approach, S120 includes sub-steps S121-S122: S121, determine the first-order difference sequence of the predicted temperature sequence, and locate the time point with the largest difference value as the starting point of the first verification window.

[0041] S122, taking the starting point of the first verification window as the reference, extend forward by t1 period and backward by t2 period to generate the verification interval, where t1 < t2, t1 = τ / 3, t2 = 2τ, and τ is the thermal inertia time constant of the packaging machine equipment.

[0042] This method uses the first derivative method to locate the point of maximum temperature rise rate, which is to locate the maximum value of the first derivative. For example, the extended window is [τ / 3, 2τ], where τ = 15 seconds (thermal inertia constant).

[0043] The thermal inertia constant can be determined by applying a step load (50%→100%) to the motor, recording the temperature rise curve, and fitting the temperature rise curve to determine τ=15 seconds (R²=0.98).

[0044] In another alternative approach, S120 includes sub-steps S123-S124: S123, determine the second-order difference sequence of the predicted temperature sequence, and locate the curvature abrupt change point as the starting point of the second verification window.

[0045] S124, taking the starting point of the second verification window as the reference, extends forward by t3 periods and backward by t3 periods to generate the verification interval, where t3 is the average value of t1 and t2.

[0046] This method uses the second derivative method to locate the point of curvature change, which is to locate the maximum value of the second derivative, for example, the symmetric extended window [τ-5, τ+5] seconds.

[0047] The selection of the first and second verification windows can be based on the fault mode. Specifically, in S110, the servo motor fault diagnosis method disclosed in patent CN118501692B is used simultaneously to determine the fault mode, which will not be elaborated here. If the temperature changes rapidly in the fault mode, such as a winding short circuit or severe bearing wear, the predicted temperature sequence will show obvious changes in the early stage. In this case, the first verification window can be selected to accurately and quickly capture the changes in the predicted temperature sequence through the first derivative. If the temperature changes slowly, the predicted temperature sequence will show obvious changes in the middle and later stages, such as slight bearing wear, and the changes in the predicted temperature sequence will be gradual. In this case, the second verification window can be selected to accurately and timely capture the changes in the predicted temperature sequence through the second derivative.

[0048] Vibration signals can capture physical anomalies such as wear and impact in mechanical parts (e.g., bearing spalling, gear tooth breakage), and early faults can be detected through time-frequency analysis (STFT / wavelet transform). Current signals can reflect electrical conditions (e.g., winding short circuits, harmonic distortion), and power supply interference or load imbalance can be detected through THD analysis. Temperature signals can directly characterize thermal conditions, but they exhibit hysteresis (e.g., frictional losses require heat accumulation to trigger a temperature rise). Therefore, if prediction is made directly based on temperature signals, it is easy to wait until the late stage of the fault before making a relatively accurate prediction. If prediction is made directly based on the first sensor signal (vibration signal or current signal), false alarms are likely to occur frequently. Therefore, the sensitivity of the first sensor signal to fault occurrence and the accuracy of the temperature signal to fault response can be utilized. In response to abnormal fluctuations in the first sensor signal, the temperature sequence can be predicted to obtain a predicted temperature sequence. The accuracy of the feedback from the first sensor signal can be verified based on the predicted temperature sequence, thereby achieving accurate and timely identification of faults in packaging machine equipment.

[0049] By using the time point with the largest localization difference as the starting point of the first verification window or the localization curvature abrupt change point as the starting point of the second verification window, the accuracy of the first sensing signal feedback can be verified based on the predicted temperature sequence this morning, while ensuring verification accuracy.

[0050] S130, the predicted temperature sequence within the verification interval is matched with the actual collected temperature sequence. If the matching degree exceeds the preset confidence threshold, it is determined to be a real degradation event.

[0051] In one implementation, the degree of matching between the predicted temperature sequence and the actual acquired temperature sequence can be determined by DTW. Specifically, S130 includes sub-steps S131~S134: S131, standardize the predicted temperature sequence S1 and the actual temperature sequence SS2 within the verification interval.

[0052] Z-score normalization is performed on the predicted temperature sequence and the actual temperature sequence within the validation interval.

[0053] S132, construct the cumulative distance matrix based on the normalized predicted temperature sequence and the actual temperature sequence.

[0054] When calculating the cumulative distance matrix, path constraints are applied, which means that the path is only allowed to move in the upper right direction to ensure that the temperature sequence is monotonically increasing.

[0055] Construct the cumulative distance matrix D(i,j)=|S1(i)-S2(j)|+min(D(i-1,j),D(i,j-1),D(i-1,j-1)).

[0056] S133, determine the minimum cumulative distance between the predicted temperature sequence and the actual temperature sequence based on the cumulative distance matrix.

[0057] The minimum cumulative distance is the element D(m,n) at the bottom right of the cumulative distance matrix D(i,j).

[0058] S134, when the minimum cumulative distance is less than the baseline distance of the same type of fault in the preset degradation mode library, the matching is determined to be successful.

[0059] For example, if the normalized distance D1=D(n,m) / max(n,m)≤0.7, then the match is considered successful, and the fault reflected by the fluctuation of the first sensor signal is a real fault, that is, the packaging machine equipment has experienced a real degradation event.

[0060] The lifetime decision unit 24 is used to determine the first remaining lifetime by fusing the multidimensional degradation features of the first sensing signal and the temperature signal after determining that a real degradation event has occurred.

[0061] Vibration and temperature signals can be used to calculate the influence coefficient of temperature rise on insulation life based on the Arrhenius equation; current and temperature signals can be used to correlate current anomalies with temperature rise through the Joule heating model.

[0062] The lifetime calculation is based on vibration and temperature signals using the Arrhenius model. The vibration signal reflects the degree of mechanical wear, while the temperature signal reflects the temperature rise caused by friction. The influence of temperature on the lifetime of insulating materials can be quantified using the Arrhenius equation. The remaining lifetime RUL1 = L0 × exp(-Ea / (k×Tbaseline²) × ΔTcorrection), where L0 is the initial lifetime (provided by the manufacturer, e.g., 1000 hours), Ea is the material activation energy (e.g., 0.8 eV, calibrated experimentally), k is the Boltzmann constant (8.617e-5 eV / K), Tbaseline is the reference temperature (e.g., 298K, i.e., 25℃), and ΔTcorrection is the combined correction value of the actual temperature rise and vibration energy (ΔTcorrection = measured temperature rise + α × vibration energy integral, α is the experimental calibration coefficient). For example, if L0 = 1000 hours, the measured temperature rise ΔT = 10℃, the vibration energy integral is calculated to be 50 (m / s²·s), and α = 0.1, then the ΔT correction = 10 + 0.1 × 50 = 15℃. Substituting into the formula, we get RUL1 ≈ 327 hours.

[0063] Based on the Joule heating model and the lifetime calculation using current and temperature signals, current anomalies (such as harmonic distortion) lead to additional Joule heating, accelerating material aging. The lifetime formula is: Remaining lifetime RUL2 = L0 / (1 + β × Ieffective² × ΔTtotal), where Ieffective is the effective current value (unit: ampere-ampere A), and β is the aging factor (e.g., 0.002 A). - 2℃ - ¹(calibrated experimentally), ΔTtotal is the total temperature rise (measured temperature rise + additional temperature rise), and the additional temperature rise is ΔTadditional = γ × THD × Ieffective² × t (γ is the thermal resistance coefficient, THD is the harmonic distortion rate, and t is the duration of the anomaly). For example, if Ieffective = 5A, THD = 10%, t = 10 hours, and γ = 0.05, then ΔTadditional = 0.05 × 10 × 5² × 10 = 125℃·h. Assuming the measured temperature rise is 8℃, the total temperature rise ΔTtotal = 8 + 125 = 133℃, substituting into the formula, we get RUL2 ≈ 278 hours.

[0064] The α, β, and γ coefficients mentioned above can all be determined through accelerated aging experiments and regression analysis. For example, different current and vibration conditions can be simulated in a constant temperature chamber, the material failure time can be recorded, and the coefficient values ​​can be fitted.

[0065] In another implementation, the signal acquisition unit 10 is also used to synchronously acquire the second sensor signal. The cross-modal prediction unit 23 is also used to verify whether the second sensing signal exhibits synchronous abnormal fluctuations within the critical time window after it is determined to be a real degradation event.

[0066] The method for determining the second sensing signal and the method for judging the synchronous abnormal fluctuation of the second sensing signal can be referred to the description in the abnormal triggering unit 22 above, and will not be repeated here.

[0067] Lifetime decision unit 24 is also used to execute S241~S244: S241, after verifying that the second sensor signal has abnormal synchronization fluctuations, count the number of synchronization anomalies and the total number of sampling points of the second sensor signal within the critical time window.

[0068] S242 determines the remaining lifetime correction value based on the number of synchronization anomalies, the total number of sampling points, and the preset lifetime decay coefficient.

[0069] S243, determine the second remaining lifetime based on the first remaining lifetime and the remaining lifetime correction value.

[0070] S244, the second remaining life is determined as the remaining life of the packaging machine equipment.

[0071] ;in, For the second remaining lifetime, For the first remaining lifetime, To preset the lifespan degradation coefficient, To synchronize the number of exceptions, This represents the total number of sampling points.

[0072] For determining the preset lifetime attenuation coefficient, the processing module 10 also includes: a fault mode learning unit 25, which stores the phase difference features of vibration-current-temperature signals under preset typical faults. It is used to extract the phase difference features corresponding to the current vibration-current-temperature signal in the actual degradation stage, determine the closest fault mode through cosine similarity matching, and load the corresponding lifetime attenuation coefficient according to the matched fault mode.

[0073] The phase difference Δφ between the current vibration signal or current signal and the temperature signal can be calculated using the cross-correlation function. For example, in some typical fault modes, the bearing wear phase difference characteristic is Δφ=30°±5°, and the winding short circuit phase difference characteristic is Δφ=120°±10°. By extracting the current signal Δφ in real time, the closest fault mode is matched using cosine similarity, and the corresponding attenuation coefficient (such as bearing wear coefficient=1.2) is applied.

[0074] This application embodiment, when the first signal is abnormal, predicts the future temperature sequence based on the first signal, and selects a verification point from the time corresponding to the maximum rate of change in the future temperature sequence from the current time. The real-time temperature signal corresponding to the verification point is then verified to eliminate false faults, improving the reliability of the remaining lifespan of the packaging machine. The selection of the verification point from the time corresponding to the maximum rate of change in the future temperature sequence allows for timely differentiation between a real and false fault in the packaging machine caused by the abnormal first signal. If the verification passes, the remaining lifespan of the packaging machine is predicted based on the real-time first signal. This allows for timely and reliable prediction of the remaining lifespan of the packaging machine when a real fault occurs.

[0075] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A predictive maintenance system for cobalt salt packaging machines based on multi-parameter fusion, characterized in that, include: The signal acquisition unit is used to acquire the first sensor signal and the temperature sensor signal; Processing module, used for: When the first sensing signal exceeds the threshold, a predicted temperature sequence is generated based on the current first sensing signal; Locate the extreme points of temperature change rate in the predicted temperature sequence, and extract the validation interval based on the extreme points; The predicted temperature sequence within the verification interval is matched with the actual collected temperature sequence; If the matching degree exceeds the preset confidence threshold, it is determined to be a true degradation, and the first remaining lifetime is determined based on the degradation characteristics of the true degradation stage.

2. The system according to claim 1, characterized in that, The signal acquisition unit is used to simultaneously acquire at least two heterogeneous sensing signals, including a first sensing signal and a temperature sensing signal, wherein the first sensing signal is a vibration signal reflecting the mechanical state of the motor or a current signal reflecting the electrical state.

3. The system according to claim 2, characterized in that, The processing module includes: The data preprocessing unit is used to align the sampling timestamps of the first sensing signal and the temperature sensing signal with the temperature sensing signal as the time reference by interpolation, and to perform noise filtering on the first sensing signal and the temperature sensing signal to generate standardized time-series data. An abnormal triggering unit is used to detect whether the first sensing signal has abnormal fluctuations through a dynamic threshold, and to generate a lifetime prediction start command when an abnormality is triggered. The cross-modal prediction unit, in response to the lifetime prediction initiation command, performs the following operations: Based on a pre-trained time-series prediction model, a predicted temperature sequence for the next N sampling periods is generated according to the current first sensing signal. The time-series prediction model is trained by the coupling relationship between the first sensing signal and the temperature signal in historical data. The extreme points of temperature change rate are located in the predicted temperature sequence, and the verification interval is extracted based on the critical time window of the extreme points; The predicted temperature sequence within the verification interval is matched with the actual collected temperature sequence. If the matching degree exceeds the preset confidence threshold, it is determined to be a real degradation event. The lifetime decision unit is used to determine the first remaining lifetime by fusing the multidimensional degradation features of the first sensing signal and the temperature signal after determining that a real degradation event has occurred.

4. The system according to claim 3, characterized in that, The time-series prediction model in the cross-modal prediction unit is a bidirectional long short-term memory network based on an attention mechanism. The input layer of the time-series prediction model receives the first signal sliding window data of the previous T periods at the current time, and the output layer generates the predicted temperature sequence for the next N periods, where T>N. The network weights are trained by the dynamic coupling relationship between the first signal and the temperature signal in historical data.

5. The system according to claim 4, characterized in that, The step of locating the extreme points of temperature change rate in the predicted temperature sequence and extracting the verification interval based on the critical time window of the extreme points includes: Determine the first-order difference sequence of the predicted temperature sequence, and locate the time point with the largest difference value as the starting point of the first verification window; Based on the starting point of the first verification window, a verification interval is generated by extending forward by period t1 and backward by period t2, where t1 < t2, t1 = τ / 3, t2 = 2τ, and τ is the thermal inertia time constant of the packaging machine. or, The step of locating the extreme points of temperature change rate in the predicted temperature sequence and extracting the verification interval based on the critical time window of the extreme points includes: The second-order difference sequence of the predicted temperature sequence is determined, and the curvature abrupt change point is located as the starting point of the second verification window. Based on the starting point of the second verification window, a verification interval is generated by extending forward by t3 periods and backward by t3 periods, where t3 is the average value of t1 and t2.

6. The system according to claim 5, characterized in that, The step of matching the predicted temperature sequence within the verification interval with the actual collected temperature sequence, and determining it as a real degradation event if the matching degree exceeds a preset confidence threshold, includes: The predicted and actual temperature sequences within the validation interval are standardized. Construct a cumulative distance matrix based on the standardized predicted temperature sequence and the actual temperature sequence; The minimum cumulative distance between the predicted temperature sequence and the actual temperature sequence is determined based on the cumulative distance matrix. A match is considered successful when the minimum cumulative distance is less than the baseline distance for the same type of fault in the preset degradation mode library.

7. The system according to claim 6, characterized in that, The signal acquisition unit is also used to synchronously acquire the second sensor signal; The cross-modal prediction unit is also used to verify whether the second sensing signal exhibits synchronous abnormal fluctuations within the critical time window after it is determined to be a real degradation event. The lifespan decision unit is also used for: After verifying that the second sensing signal exhibits abnormal synchronization fluctuations, the number of synchronization anomalies and the total number of sampling points of the second sensing signal within the critical time window are counted. The remaining lifetime correction value is determined based on the number of synchronization anomalies, the total number of sampling points, and the preset lifetime decay coefficient. The second remaining lifetime is determined based on the first remaining lifetime and the remaining lifetime correction value; The second remaining lifespan is determined as the remaining lifespan of the packaging machine equipment.

8. The system according to claim 7, characterized in that, The processing module further includes: The fault mode learning unit stores the phase difference features of vibration-current-temperature signals under preset typical faults. It is used to extract the phase difference features corresponding to the current vibration-current-temperature signal in the actual degradation stage, determine the closest fault mode through cosine similarity matching, and load the corresponding lifetime decay coefficient according to the matched fault mode.

Images