Fault tracing method for fruit and vegetable juice production line equipment

Abstract

The invention discloses a fruit and vegetable juice production line equipment fault tracing method, which comprises the following steps of: acquiring parameters such as temperature, pressure, vibration, rotating speed and motor current in real time through a multi-channel sensor, and establishing a working condition characteristic database by combining filtering, normalization, statistics and frequency domain characteristic extraction; based on a support vector regression algorithm, a nonlinear mapping model of working condition features and anomaly detection thresholds is constructed, and dynamic threshold adaptive output and real-time anomaly judgment for different working conditions are realized; according to a detection result, a fault signal is automatically triggered, a model is continuously incremented and trained, the adaptability to new working conditions is improved, the accuracy, intelligence and stability of equipment anomaly detection are effectively improved, misinformation and missing information can be reduced, and the automatic operation and maintenance level of a production line is enhanced.

Description

Technical Field

[0001] This invention relates to the field of monitoring and anomaly detection technology for fruit and vegetable juice production lines, and in particular to a method for tracing the source of equipment failures in fruit and vegetable juice production lines. Background Technology

[0002] With the increasing automation level and growing demand for intelligent manufacturing in the fruit and vegetable juice industry, equipment fault monitoring and anomaly detection technologies for fruit and vegetable juice production lines are receiving increasing attention from the industry. Existing equipment monitoring systems for fruit and vegetable juice production lines typically rely on multi-sensor signal acquisition, anomaly discrimination methods based on statistical analysis or traditional rule settings, and a small number of anomaly recognition models based on machine learning, in order to achieve real-time perception of equipment operating status and preliminary fault warnings. In current mainstream technical solutions, dynamic thresholding is widely used to improve the sensitivity and adaptability of anomaly detection. Its basic principle is to utilize the statistical distribution characteristics of historical equipment operating parameters, and automatically adjust the anomaly judgment threshold through methods such as sliding windows, clustering, or time-series segmentation, combined with empirical rules or traditional statistical indicators, to adapt to changes in equipment status at different stages, under different loads, and in different environments on the production line. This type of method can, to a certain extent, reduce the problems of missed detections and false alarms caused by fixed threshold settings, and is an important development direction in the field of intelligent monitoring of fruit and vegetable juice production lines. On the other hand, with the deepening application of big data and artificial intelligence technologies in the industrial field, some high-end production lines have begun to introduce machine learning models trained on historical large-sample data, such as support vector machines (SVM), decision trees, and clustering algorithms, to perform high-dimensional mapping of multi-source signal features and comprehensive judgment of multivariate anomalies, thereby achieving more intelligent equipment health management. These methods can automatically summarize the feature distribution patterns under typical operating conditions, providing data-driven power for the adaptive adjustment of anomaly detection thresholds, and are representative of modern smart factories. However, existing technologies still have significant shortcomings in terms of adaptability, real-time performance, and accuracy of anomaly detection under the complex and variable operating conditions of fruit and vegetable juice production lines: (1) Static or dynamic threshold methods that rely solely on historical data are difficult to respond to rapid changes in equipment operating conditions. When there are differences in raw material batches, adjustments to process parameters, or sudden changes in environmental temperature and humidity, the existing threshold adjustment mechanism lags significantly, leading to an increase in false alarms or serious missed alarms under new operating conditions, which affects the accurate location and rapid response of equipment faults. (2) For scenarios where the operating status of equipment fluctuates significantly or there are extreme working conditions, traditional dynamic threshold methods cannot make full use of real-time working condition characteristics and lack effective modeling of the mapping relationship between threshold and real working conditions under multi-dimensional data, resulting in coarse threshold adjustment granularity and limited prediction ability. (3) At present, most anomaly detection based on machine learning focuses on static model training and lacks a data closed-loop optimization mechanism that combines real-time working conditions and dynamic interaction. It is difficult to achieve rapid model adaptation to new working conditions and sudden anomalies, and the model generalization ability and practical application value are restricted. (4) Existing technologies often focus only on the monitoring and discrimination of a single physical quantity, failing to fully explore the complex coupling relationship between multiple sources of equipment signals and the fusion information of timing statistics and frequency domain characteristics, making it difficult to meet the fault tracing requirements of high reliability and low latency for complex production lines. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for tracing the source of equipment failure in a fruit and vegetable juice production line.

[0004] The technical solution of this invention is implemented as follows: A method for tracing the source of equipment failure in a fruit and vegetable juice production line, comprising: S1: Collect real-time operating parameters of multiple devices in the fruit and vegetable juice production line under different operating conditions, and record the corresponding operating condition label information. The operating parameters include temperature, pressure, vibration, speed and motor current signal. S2: Perform filtering, noise reduction, and normalization preprocessing on the collected operating parameters to eliminate sensor drift and environmental interference, and generate a standardized equipment operation dataset; S3: Based on the preprocessed equipment operation dataset, extract multi-dimensional operating condition feature vectors, which include the mean, variance, kurtosis, skewness, and frequency domain energy distribution index within the sliding window; S4: Based on historical abnormal samples and corresponding working condition feature vectors, construct a nonlinear mapping model between working condition features and abnormal detection thresholds. Train the model using support vector regression algorithm and output the historical optimal dynamic threshold. S5: Input the currently collected working condition feature vector into the nonlinear mapping model, and output an adaptive anomaly detection threshold that adapts to the current operating state in real time; S6: Based on the degree of deviation between the adaptive anomaly detection threshold and the current device operating parameters, determine whether the device is in an abnormal state. If the deviation exceeds the threshold, generate an anomaly marker signal. S7: If the abnormal state is determined, the equipment status abnormal signal is triggered and output to the fault tracing module, and the current working condition characteristics and detection threshold are recorded to the model update database. S8: Periodically perform incremental training on the nonlinear mapping model based on the newly added data in the model update database to improve the model's adaptability to new working conditions, while limiting the model update frequency to avoid resource overload.

[0005] The method for tracing equipment failures in a fruit and vegetable juice production line provided in this application has the following beneficial effects: (1) This invention collects multi-source operating parameters (including temperature, pressure, vibration, speed, motor current, etc.) and accurately labels the operating conditions, extracts multi-dimensional operating condition features reflecting the current operating status of the equipment, and constructs a nonlinear mapping model between the operating condition feature vector and the anomaly detection threshold. By utilizing the support vector regression algorithm and the dynamic input of the operating condition features, the real-time adaptive adjustment of the anomaly detection threshold is realized, effectively solving the problems of threshold lag and poor adaptability of static setting; (2) This invention introduces a nonlinear model, which can dynamically output the optimal anomaly detection threshold within milliseconds based on the actual working conditions at the production site. Compared with the previous method of setting thresholds using a single historical window, it supports precise parameter mapping for different equipment, different working conditions, and different operating segments, significantly improving the accuracy and sensitivity of anomaly judgment; (3) This invention designs an automatic reflow mechanism for abnormal working condition samples and an incremental model update mechanism, which can periodically and automatically use the latest working condition and detection threshold data to perform targeted incremental training on the model. By introducing sample quality scores and distribution difference measures, high-value training samples are screened, effectively avoiding model aging and continuously improving the generalization ability to new working conditions, extreme environments and abnormal patterns. After the model is updated, the threshold prediction error is significantly reduced, effectively adapting to the constantly changing production line operating environment; (4) The entire data flow of this invention adopts multi-channel synchronous acquisition, timestamp alignment, wavelet denoising, PCA dimensionality reduction, sliding window statistics and other multi-level preprocessing, which greatly ensures the quality of available data and computational efficiency. The model can run stably on embedded or field industrial servers through incremental batch training and resource consumption control. The update cycle and computing resources can be controlled as needed, avoiding excessive system load and greatly improving the engineering application feasibility of the solution. Attached Figure Description

[0006] Figure 1 This is a flowchart of a method for tracing the source of equipment failure in a fruit and vegetable juice production line according to the present invention; Figure 2 This is a sub-flowchart of a method for tracing equipment failures in a fruit and vegetable juice production line according to the present invention. Figure 3 This is another sub-flowchart of the method for tracing the source of equipment failure in a fruit and vegetable juice production line according to the present invention. Detailed Implementation

[0007] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0008] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0009] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having,” etc., specify the presence of the stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0010] Please see Figures 1-3 As shown, a method for tracing equipment failures in a fruit and vegetable juice production line includes: S1: Collect real-time operating parameters of multiple devices in the fruit and vegetable juice production line under different operating conditions, and record the corresponding operating condition label information. The operating parameters include temperature, pressure, vibration, speed and motor current signal. S2: Perform filtering, noise reduction, and normalization preprocessing on the collected operating parameters to eliminate sensor drift and environmental interference, and generate a standardized equipment operation dataset; S3: Based on the preprocessed equipment operation dataset, extract multi-dimensional operating condition feature vectors, which include the mean, variance, kurtosis, skewness, and frequency domain energy distribution index within the sliding window; S4: Based on historical abnormal samples and corresponding working condition feature vectors, construct a nonlinear mapping model between working condition features and abnormal detection thresholds. Train the model using support vector regression algorithm and output the historical optimal dynamic threshold. S5: Input the currently collected working condition feature vector into the nonlinear mapping model, and output an adaptive anomaly detection threshold that adapts to the current operating state in real time; S6: Based on the degree of deviation between the adaptive anomaly detection threshold and the current device operating parameters, determine whether the device is in an abnormal state. If the deviation exceeds the threshold, generate an anomaly marker signal. S7: If the abnormal state is determined, the equipment status abnormal signal is triggered and output to the fault tracing module, and the current working condition characteristics and detection threshold are recorded to the model update database. S8: Periodically perform incremental training on the nonlinear mapping model based on the newly added data in the model update database to improve the model's adaptability to new working conditions, while limiting the model update frequency to avoid resource overload. Step S1: Collect real-time operating parameters of multiple devices in the fruit and vegetable juice production line under different operating conditions, and record the corresponding operating condition label information. These operating parameters include temperature, pressure, vibration, speed, and motor current signals. Specifically, this includes: S1.1: Based on the key equipment configuration in the fruit and vegetable juice production line, deploy a multi-channel sensor network to synchronously collect temperature, pressure, vibration, speed and motor current signals to obtain raw operating data of the equipment under different operating conditions.

[0011] Based on the key equipment layout and monitoring requirements in the fruit and vegetable juice production line, a multi-channel synchronous acquisition strategy is adopted to obtain multi-dimensional operating signals of temperature, pressure, vibration, speed and motor current. A distributed sensor network deployment method is adopted (parameters: temperature sensor accuracy ±0.1℃, pressure sensor range 0-1MPa, vibration sensor frequency response range 0-10kHz, speed sensor resolution ≤1rpm, current sensor sampling frequency ≥10kHz) to realize the data acquisition function of channel access and parallel sampling of corresponding signals; Furthermore, by using a high-stability synchronous clock allocation technique (parameter: clock synchronization error ≤ 1μs), the sampling phase of the multi-channel sampling units is unified, and the original multi-source operating data with a strict time alignment reference is obtained. Furthermore, based on the electromagnetic interference shielded wiring method (parameter: double-layer shielding coefficient ≥60dB) and differential signal transmission protocol, complete data transmission from long-distance sensors to acquisition nodes is achieved, and a low-interference raw signal stream is generated; Furthermore, through the analog-to-digital conversion module (parameters: 16-bit sampling accuracy, sampling rate meeting Nyquist frequency requirements), various analog sensor signals are converted into digital sampling sequences, and raw running data packets in the digital domain are generated; By using multi-channel buffer storage and unified data indexing, the aforementioned digitization results are transformed into raw sampling datasets that can be used for subsequent timestamp alignment and operating condition labeling, thereby achieving the technical effects of synchronous acquisition of multi-source signals and high-precision time consistency. For example, on a production line with a daily output of 2000 liters of fruit and vegetable juice, a PT100 platinum resistance temperature sensor (accuracy ±0.05℃) is deployed at the inlet of the juicing unit; a pressure transmitter (range 0-0.6MPa, accuracy ±0.25%FS) is installed in the pressing chamber; vibration signals (frequency response 5Hz-10kHz) are collected at both ends of the main pressing shaft using piezoelectric accelerometers; a Hall effect speed sensor (resolution 0.5rpm) is installed at the end of the shaft; and a Hall current sensor (sampling frequency 20kHz) is connected in series in the main drive motor circuit. All sensors achieve a sampling time error of less than [value missing] through a synchronous clock module based on the IEEE 1588 Precision Time Protocol (PTP). Under three typical operating conditions—full load, half load, and no load—the system continuously collects data from each channel. The 1-second signal is converted into a 16-bit quantized digital signal sequence through analog-to-digital conversion. The original data packet specifies the sampling time, channel number, and corresponding physical quantities. Actual measurement data under this configuration shows that the synchronization deviation of the multi-source signals remains within a certain range under full production load conditions. Within a certain range, and without significant interference peak drift, it provides high-quality data input for subsequent timestamp alignment and working condition feature extraction; S1.2: Perform timestamp alignment processing on the collected raw running data to ensure the consistency of multi-source signals in the time dimension, thereby generating a structured raw dataset with time-series correlation characteristics; Based on the digital raw data packets obtained through multi-channel synchronous acquisition, a high-precision timestamp interpolation method is used (parameter: interpolation resolution is...). This enables precise compensation and alignment of time stamps for sampling points in each channel; Furthermore, a common clock signal reference comparison method is used (parameter: clock source stability is better than...). ), calculate the phase offset between the timestamp of each channel and the reference clock. And generate a time offset correction vector; Furthermore, a linear time offset compensation algorithm is adopted, based on the formula: in, The corrected sampling time. The original sampling time label, To update the time stamp for the channel's offset relative to the master clock; Furthermore, based on the millisecond-level synchronization window segmentation mechanism, the corrected sampling data of each channel is divided into equal-length data frames according to a unified time period, ensuring that the sampling values ​​of different physical quantities at corresponding times can form strictly one-to-one time-series samples. Furthermore, the time key index mapping table generation module assigns a frame number, start timestamp, and sampling interval metadata to each frame of data, thereby realizing the time dimension association between cross-channel and multi-source signals; Through the above timestamp alignment and framing processing, multi-source heterogeneous sampling data is transformed into a time-series dataset with strict time synchronization and structured index, enabling high-precision data docking for subsequent working condition annotation and feature extraction. For example, in a fruit and vegetable juice production line with a daily output of 2000 liters, the sampling frequency of the juicer temperature signal is: The pressure signal sampling frequency is The vibration signal sampling frequency is The sampling frequency of the rotation speed signal is The sampling frequency of the motor current signal is The sampling timestamps for each channel are marked by the master clock module based on the IEEE 1588 PTP protocol, and the temperature channel is obtained after phase offset detection. Pressure Channel Vibration channel Speed ​​channel Current channel After correcting the time stamps of all sampling points based on the compensation formula, a uniform frame length of 1 millisecond is used for data frame segmentation. The number of sampling points for each channel within each frame is automatically matched according to its sampling frequency, resulting in a synchronization deviation of less than [value missing] under full production load conditions. The structured multi-source time-series dataset provides stable and accurate input data for the labeling of the working conditions of S1.3; S1.3: Based on the mapping relationship between equipment operating status and process flow, define the operating condition labeling system, and label the operating condition status of the collected operating data segments to generate an operating data sample set with operating condition labels; S1.4: Perform data integrity verification on the operational data sample set with operating condition labels, identify and mark missing values, outliers and signal drift points to ensure the data quality of subsequent feature extraction; S1.5: Store the verified running data sample set and its corresponding operating condition label information into the distributed data acquisition database as the input data source for subsequent feature extraction and model training.

[0012] Step S2: The collected operating parameters are preprocessed by filtering, noise reduction, and normalization to eliminate sensor drift and environmental interference, generating a standardized equipment operation dataset. Specifically, this includes: S2.1: Perform wavelet threshold denoising on the collected temperature, pressure, vibration, speed and motor current signals to suppress high-frequency noise interference in the sensor signals and obtain the original equipment operation signal after denoising; For multi-source operating signals such as temperature, pressure, vibration, speed and motor current that have been timestamped and verified for data integrity, a wavelet threshold denoising method is used (parameters: wavelet basis selection is Daubechies 8th order, decomposition level is 5 levels, threshold strategy is Stein's Unbiased Risk Estimate - SURE soft threshold) to achieve non-stationary suppression of high-frequency noise components. Furthermore, through multi-scale wavelet decomposition, the input signal is... Decomposed into low-frequency approximation coefficients at each decomposition level. With multiple sets of high-frequency detail coefficients The decomposition process satisfies: To achieve detailed separation of signals at different scales in order to selectively reduce noise components; Furthermore, based on each set of detail coefficients Calculate the standard deviation of noise The robust estimation formula using MAD (Median Absolute Deviation) is adopted: Obtain the noise level estimation result of the current decomposition layer; Furthermore, based on the noise standard deviation With the sample length of this layer Adaptive thresholding in computational layer The optimal threshold parameter is obtained; Furthermore, the SURE soft-threshold compression function is used to compress the detail coefficients. Denoising is performed to achieve flexible suppression of high-frequency noise while maintaining the main signal components without distortion; By using the wavelet reconstruction algorithm, the approximation coefficients of each decomposition layer and the thresholded detail coefficients are inversely transformed to output the denoised original device operation signal data sequence, achieving the technical effect of wideband noise suppression and signal detail preservation. For example, in an anomaly detection optimization scenario for a 2000-liter-per-day fruit and vegetable juice production line, the vibration signal sampling frequency is: Hz, the sampling frequency of the motor current signal is Hz, respectively, selected the Daubechies 8th order wavelet basis for 5-level decomposition, and the noise standard deviation of the third level of the vibration signal. The estimated value is Sample length for Then the threshold The calculation is as follows: After applying soft threshold compression based on this threshold and removing high-frequency spike noise, the noise power decreased by approximately [amount missing] in the spectral domain. dB, effectively improving the signal-to-noise ratio and providing high-quality signal input for S2.2 smoothing filtering and subsequent feature extraction; S2.2: Based on the original equipment operation signal after noise reduction, the sliding window mean filtering algorithm is used to smooth the signal in order to eliminate transient interference caused by equipment start-up or load changes, and obtain a smoothed equipment operation signal sequence. S2.3: Perform standardization processing based on zero mean and standard deviation normalization on the smoothed equipment operation signal sequence to eliminate the differences in the dimensions of different sensors and generate normalized equipment operation signal data with uniform dimensions; S2.4: Based on normalized equipment operation signal data, a sliding window mechanism is used to extract the signal statistical features within the window, including the window mean, variance, maximum and minimum values, to generate a preliminary structured feature vector; S2.5: Perform principal component analysis (PCA) dimensionality reduction on the initial structured feature vectors to remove redundant features and retain the main variation patterns, and output the dimensionality-reduced standardized equipment operation feature dataset for use by the subsequent operating condition feature extraction module.

[0013] Step S3: Based on the preprocessed equipment operation dataset, extract multi-dimensional operating condition feature vectors. These feature vectors include the mean, variance, kurtosis, skewness, and frequency domain energy distribution indices within a sliding window. For example... Figure 2 As shown, it specifically includes: S3.1: Divide the preprocessed standardized equipment operation dataset into sliding windows. Each sliding window contains a continuous sequence of equipment operation parameters for a set time length in order to capture the dynamic changes of the equipment under different operating conditions. S3.2: Calculate the mean of the equipment operating parameter sequence within each sliding window to characterize the central trend of the equipment operating status within that time period and obtain the sliding window mean characteristic parameter; For the standardized equipment operating parameter sequence after S3.1 sliding window division, the arithmetic mean calculation method is used (parameters: window length L is 256-2048 sampling points, sliding step size Δ is 1-128 sampling points) to realize the quantitative characterization of the signal center trend within each sliding window; Furthermore, the window mean characteristic parameter is calculated using the arithmetic mean formula, which is: in, This represents the mean of the current sliding window. The value of the device operating parameter for the i-th sampling point within the window. This represents the total number of sampling points within the window length. Furthermore, based on the numerical stability optimization strategy, Kahan summation compensation operation is performed on the data within the window to reduce floating-point accumulation error and improve the numerical accuracy of mean calculation under long window conditions. Furthermore, by performing parallel mean calculations on single-channel and multi-channel signals, the mean features of different sensor types are vectorized and combined according to timestamps to generate multi-dimensional mean feature vectors, ensuring that the central trends of each operating parameter are comparable and usable on a unified time scale. Furthermore, to address potential window boundary effects, a mirror filling strategy is employed to supplement incomplete window data, ensuring that the mean calculation for each sliding window is based on sufficient sample capacity and avoiding feature shifts caused by incomplete data. The above algorithm is used to transform the original windowed sequence output by S3.1 into sliding window mean feature parameters that quantitatively describe the central trend of the equipment's operating status, thereby effectively capturing the stability level of operating parameters under different operating conditions. For example, in the vibration signal channel of a 2000-liter-per-day fruit and vegetable juice production line, the sliding window length L is set to... Number of sampling points (corresponding time length is) ms), sliding step size Δ is Each sampling point. A sequence of vibration acceleration samples within a given window. Let the total number of sampling points be... = According to the mean formula, the actual mean value of this window is calculated to be... The error after Kahan compensation is less than This indicates that the numerical stability meets the requirements. Under a certain abnormal operating condition, the average vibration value increased by approximately [value missing] due to bearing failure. The combined application of this mean feature with subsequent variance and skewness features can significantly improve the sensitivity and accuracy of anomaly detection. The multi-channel mean vector, after being fused with other statistical features, serves as the input to the standardized multi-dimensional operating condition feature vector of S3.7, providing a stable and effective description of the central trend for the nonlinear mapping model. S3.3: Based on the sliding window mean characteristic parameter, further calculate the variance of the equipment operating parameter sequence to reflect the degree of fluctuation of the equipment status within this time period and obtain the sliding window variance characteristic parameter; Based on the sliding window mean characteristic parameters obtained in S3.2 and the standardized equipment operating parameter sequence within the same window, variance calculation is performed on the sequence to quantify the degree to which the signal deviates from the central trend within that time period; The unbiased sample variance formula is used (parameter: window length). The mean is The variance characteristic parameters are calculated using the following mathematical relationship: in, For the first in the window Each sample value The total number of samples, The variance feature of the current sliding window; Furthermore, to avoid precision loss in long-window operations, a high-precision variance estimation is achieved through the Welford online algorithm (parameters: iteratively updating the mean and sum of squared deviations point by point), ensuring numerical stability during the calculation process. Furthermore, the variance of the multi-channel sensor signals is calculated separately, and feature vectorization fusion is achieved based on timestamps to generate multi-dimensional variance feature vectors, ensuring that the fluctuation characteristics of different physical quantity channels can be compared and analyzed on a unified time scale. Furthermore, for variance values ​​that may be affected by window boundary truncation, transient spikes are filtered out using an exponentially weighted moving average (parameter: decay factor α), thereby improving the smoothness and representativeness of the feature parameters. Through the above chain processing, the mean feature output by S3.2 is extended into a sliding window variance feature parameter that can characterize the stability of the operating state and sensitively reflect the degree of vibration or current fluctuation, providing an effective fluctuation description for the robustness of subsequent anomaly detection. For example, in the vibration acceleration signal analysis of a production line with a daily output of 2000 liters of fruit and vegetable juice, the sliding window length... = sampling points, sliding step size is Each sampling point corresponds to the mean in this window. = Calculate the sum of squared differences between each sampling point and the mean. Substituting into the unbiased sample variance formula: The variance was approximately Under bearing failure conditions, the variance value increases to [value missing] under the same window length. Compared to normal operating conditions, it improves by approximately %; apply a decay factor to the sequence After applying an exponentially weighted moving average, variance mutations caused by single shocks can be effectively suppressed, improving the accuracy of fault trend identification. When this variance feature is used in conjunction with features such as mean, kurtosis, and skewness, the F1 score for anomaly detection in multi-condition mixed environments is significantly improved. percentage points; S3.4: Perform kurtosis calculation on the sequence of equipment operating parameters within the sliding window to assess the steepness of its distribution pattern, thereby identifying whether there are abnormal fluctuations or extreme states and obtaining the kurtosis characteristic parameters of the sliding window. Based on the sliding window variance characteristic parameters generated by S3.3 and the standardized equipment operating parameter sequence within the same window, kurtosis calculation is performed on the sequence to evaluate the tail thickness and peak degree of the signal probability distribution; The unbiased kurtosis calculation formula is used (parameter: window length). The mean is The standard deviation is The calculation formula is: This function calculates the ratio of the fourth central moment to the fourth power of the standard deviation of the sample values ​​within a window. Furthermore, the original kurtosis value is corrected for deviation using the Fisher correction formula, which is: To achieve excess kurtosis quantization under a standard normal distribution benchmark, so as to facilitate horizontal comparison between different signal channels and operating conditions; Furthermore, an online fourth-order cumulant update algorithm is adopted (parameter: sliding window step size). Forgetting factor This enables real-time updates of kurtosis characteristics, avoiding repetitive full-window scan calculations in large-scale data stream processing, and improving computational efficiency and real-time performance. Furthermore, the kurtosis of the multi-channel sensor signals is calculated separately, and the results are vectorized and combined based on the same timestamp to generate a multi-dimensional kurtosis feature vector, thereby preserving the peak distribution characteristics of different physical quantities within the same operating window. Furthermore, to suppress kurtosis jumps caused by single extreme outliers, the kurtosis feature sequence is smoothed using median filtering (parameter: filter window length). This improves the stability and noise resistance of this feature during anomaly detection. Through the above processing method, the volatility description in the previous step is extended into a sliding window kurtosis feature parameter that can sensitively reflect the peak degree and tail weight of the distribution, thereby achieving accurate characterization and identification of non-Gaussian anomalous distributions. For example, in the analysis of motor current signals in a production line with a daily output of 2000 liters of fruit and vegetable juice, let the sliding window length be... = sampling points, step size = The average value within this window is calculated from the given sampling points. = Standard deviation = The summation result after calculating the fourth-order central moments is: , substitute into the kurtosis formula: Calculated ≈ After Fisher correction ≈ This indicates that the current signal under this operating condition exhibits a distinct peak distribution characteristic. Under conditions of minor bearing damage, the kurtosis value rises to [value missing]. This is 1.6 times that under normal operating conditions. After median filtering, the standard deviation of the kurtosis curve decreased by approximately [missing information]. This effectively improves the robustness and stability of subsequent anomaly detection based on multi-feature fusion; S3.5: Calculate the skewness based on the sequence of device operating parameters within the sliding window to measure the symmetry of its distribution pattern, identify whether there is abnormal behavior biased in a certain direction, and obtain the skewness characteristic parameters of the sliding window. S3.6: Perform Fast Fourier Transform (FFT) processing on the sequence of equipment operating parameters within the sliding window to extract frequency domain energy distribution features, identify the vibration or current response modes of the equipment at different frequencies, and obtain frequency domain energy distribution feature vectors; S3.7: Normalize the sliding window mean, variance, kurtosis, skewness, and frequency domain energy distribution characteristic parameters to eliminate the influence of different dimensions on feature fusion and generate a standardized multidimensional working condition feature vector. S3.8: Associate the standardized multidimensional working condition feature vectors with the corresponding working condition label information to construct a structured working condition feature database, providing input samples for subsequent nonlinear mapping model training.

[0014] Step S4: Based on historical abnormal samples and corresponding operating condition feature vectors, a nonlinear mapping model between operating condition features and abnormal detection thresholds is constructed. This model is trained using a support vector regression algorithm to output the historically optimal dynamic threshold. Figure 3 As shown, it specifically includes: S4.1: Perform feature matching processing on historical abnormal samples and corresponding working condition feature vectors to construct a supervised learning sample set for training nonlinear mapping models; For the multidimensional feature vectors and historical abnormal sample information in the structured working condition feature database output from step S3.8, a dual-index feature association method based on timestamps and working condition labels is adopted (parameter: time synchronization tolerance ε). Label consistency accuracy ≥ This enables accurate matching of historical abnormal samples with corresponding operating condition feature vectors; Furthermore, the feature consistency test algorithm (parameter: Euclidean distance threshold) is used. Matching ratio threshold ), calculate the distance metric between the feature vector of the abnormal sample and the feature vector of the same time window, ensure that the matching results meet the requirements of similarity and temporal consistency, and obtain a preliminary matching index table; Furthermore, an anomaly label cross-checking method is adopted (parameter: multi-source label conflict tolerance ≤ This allows for consistency comparison between manually annotated results and system-judged labels in historical anomaly event records. Inconsistent records are marked as requiring manual review, and data entries that do not meet the training sample quality standards are removed. Furthermore, through a sample quality scoring mechanism (parameter: scoring weight coefficient) (corresponding to data integrity, temporal consistency, and feature matching degree), calculate the comprehensive quality score for each matching record. and with Use a threshold to filter high-quality matching samples; Furthermore, dynamic normalization and feature alignment are performed on the selected matching samples (parameter: mean). With variance (Calculate by grouping according to working condition type) to ensure that all training features are input into the subsequent support vector regression model training process under the same dimensions and distribution conditions; Through the above chain processing, the key information in the structured chemical condition feature database and the historical abnormal event records are effectively integrated, and finally a high-quality, structured and consistent supervised learning training sample set is output, which achieves the technical effect of laying the data foundation for the construction of nonlinear mapping models. For example, in a bearing failure case of a 2000-liter-per-day fruit and vegetable juice production line, the historical anomaly sample was recorded at 10:23:45 on March 15, 2024, with the corresponding fault label being "main drive bearing inner ring wear," the operating condition label being "full load—high temperature—long-term operation," and the sampling clock frequency being [missing information]. Synchronization via timestamps, at ε= Matching multi-dimensional working condition feature vectors within the tolerance range The Euclidean distance between the event and the anomalous event characteristics is calculated as follows: in, Let i be the i-th feature component of the abnormal event sample. To match the corresponding feature components of the candidate samples. Calculated... If the value is less than δ, the match is confirmed as valid. Quality score result. If the value exceeds the screening threshold of 0.85, it is included in the training sample set. This feature vector is then Z-score standardized and used as input along with the corresponding historical best dynamic threshold. The pairing successfully constitutes an effective supervised learning sample for training the support vector regression model, enabling continuous optimization of the model's threshold prediction capability under this type of working condition. S4.2: Based on the supervised learning sample set, a nonlinear mapping relationship model between the working condition feature vector and the historical best dynamic threshold is constructed using the support vector regression algorithm, so as to realize the prediction output of the anomaly detection threshold under different working conditions; Based on the high-quality supervised learning sample set obtained through step S4.1, the Support Vector Regression (SVR) method is adopted (parameters: kernel function type = radial basis function kernel RBF, penalty factor). Kernel scale parameters (Determined by historical sample data distribution estimation), realizing the construction of a nonlinear mapping relationship from the working condition feature vector to the optimal dynamic threshold; Furthermore, by defining the regression function form of SVR: in, and For the Lagrange multipliers of the SVR model, For kernel function, For bias terms, Given any input working condition feature vector, These are support vector samples; Furthermore, the kernel function is defined as a radial basis kernel function: in, The scale of the feature space mapping is controlled, and the optimal initial value is obtained by back-calculating the feature mean square error of the sample set. Furthermore, an ε-insensitive loss function is used for model objective optimization; Furthermore, the Sequence Minimum Optimization (SMO) algorithm is used to decompose and solve the above Lagrange dual problem, and the results are iteratively updated. and The optimal set of support vectors and regression model parameters are obtained by satisfying the KKT conditions. Through the above SVR nonlinear modeling process, high-dimensional multi-source working condition features are mapped to the scalarized historical best anomaly detection threshold output, thereby achieving high-precision fitting of threshold prediction under complex working conditions. For example, in the historical dataset of anomalies in the main drive system of a 2000-liter-per-day fruit and vegetable juice production line, 200 high-quality matching samples are selected as the training set. The dimension of the operating condition feature vector is 5, and the mean feature range is... Variance feature range Kurtosis range .set up = , = , = The SMO solution converged on the 120th iteration, with support vectors accounting for 30% of the total samples. When the test set was input into the trained model, the root mean square error between the predicted threshold and the true historical best dynamic threshold was... The matching rate between the judgment results and historical anomaly detection records reached 97.2%, which is significantly better than the 89.5% achieved by using the fixed threshold method, thus improving the accuracy and generalization of threshold prediction under complex working conditions. S4.3: Perform hyperparameter optimization on the training process of the nonlinear mapping model to improve the model's generalization ability and prediction stability under different working conditions; Based on the support vector regression nonlinear mapping model and its initial parameter set constructed in step S4.2, a hyperparameter optimization method is adopted (parameter: penalty factor). Kernel scale parameters Insensitive interval This allows for the optimization and adjustment of the model's regression accuracy and prediction stability under different operating conditions. Based on grid search method (parameter search range: ∈ , ∈ , ∈ (Step size is logarithmic), traverse candidate hyperparameter combinations, train the model, and calculate the regression error index on the validation set. and The error matrix result is obtained. Using k-fold cross-validation (parameter: k= The folding method is hierarchical sampling to ensure that samples under different working conditions are evenly distributed in each fold, so as to evaluate the generalization performance of each hyperparameter combination and calculate its mean error and standard deviation to quantify the stability of model prediction. A Bayesian optimization algorithm (parameters: prior model is a Gaussian process, acquisition function is the desired improvement EI) is used to iteratively update the hyperparameter combination based on the initial grid search results, in order to maximize the generalization performance objective function. = ,in For stability weights, The standard deviation of the cross-validation error; Furthermore, by using the learning curve analysis method, the trends of training error and validation error under different hyperparameter configurations are fitted to determine whether underfitting or overfitting occurs, and hyperparameter convergence is determined in conjunction with the generalization performance index in order to select the globally optimal hyperparameter combination. Through the above hyperparameter optimization process, the parameter configuration of the initial nonlinear mapping model is transformed into an SVR model instance that achieves the optimal balance between generalization ability and prediction stability, thereby realizing a comprehensive improvement in the prediction performance of anomaly detection threshold under complex working conditions. For example, under the 5-dimensional operating condition input conditions of a production line with a daily output of 2000 liters of fruit and vegetable juice, the initial parameters are taken as follows: = , = , = As a baseline configuration, a grid search is employed. ∈ , ∈ , ∈ Eighty sets of parameters were sampled within the range, and 5-fold cross-validation was performed to calculate the RMSE and MAE for each set of parameters. After 15 rounds of Bayesian optimization iterations, the optimal parameters converged to... = , = , = The corresponding RMSE = MAE= Cross-validation standard deviation = The optimized model improved threshold prediction accuracy on the test set compared to the initial configuration. The percentage point improvement in prediction stability under extreme conditions of high temperature and high humidity and low temperature and low humidity significantly enhances the practicality and robustness of the dynamic threshold generation module. S4.4: The performance of the trained nonlinear mapping model is evaluated based on the cross-validation method to verify its fitting accuracy and anomaly threshold prediction ability on historical data. S4.5: Deploy the validated nonlinear mapping model into the real-time anomaly detection system as the core algorithm model of the dynamic threshold generation module to support subsequent adaptive threshold output based on the characteristics of the current operating conditions.

[0015] Step S5: Input the currently collected operating condition feature vector into the nonlinear mapping model, and output an adaptive anomaly detection threshold that adapts to the current operating state in real time. Specifically, this includes: S5.1: Obtain the extracted multi-dimensional operating condition feature vector under the current equipment operating status. The multi-dimensional operating condition feature vector includes the mean, variance, kurtosis, skewness and frequency domain energy distribution index within the sliding window, which are used as input parameters of the nonlinear mapping model to ensure that the model can perform dynamic inference based on real-time operating conditions. S5.2: A nonlinear mapping model based on the support vector regression algorithm is used to perform weighted feature fusion calculation on the input multidimensional working condition feature vector to obtain the output value of the anomaly detection threshold mapping function under the current working condition, so as to realize the real-time prediction of dynamic threshold. After obtaining the multidimensional feature vectors of the current working condition, a nonlinear mapping model based on the Support Vector Regression (SVR) algorithm is used (parameters: kernel function type = radial basis function kernel RBF, penalty factor). Kernel scale parameters Bias terms This enables weighted feature fusion calculation of input features, which is then mapped to the corresponding anomaly detection threshold prediction space. Furthermore, the input multidimensional working condition feature vector is mapped using a kernel function. The original feature space is mapped to a high-dimensional feature space, and the similarity between each input feature and the support vector sample is calculated using the radial basis kernel function. Furthermore, based on the support vector regression decision function and the weighted summation of kernel function outputs, the initial predicted value of the anomaly detection threshold under the current operating conditions is calculated. ; Furthermore, the kernel function similarity value is multiplied element-wise with the weight parameters of the corresponding support vector, and then accumulated item by item to obtain the comprehensive weighted feature fusion result, thereby obtaining a single scalar dynamic threshold prediction output. Furthermore, the calculated prediction threshold is subject to a numerical range constraint (parameter: lower limit). The upper limit is the maximum historical feasible threshold value), ensuring that the predicted value is within the range allowed by physical and technological conditions, and avoiding numerical anomalies caused by external interference; Through the above weighted feature fusion and nonlinear mapping operation, the high-dimensional multi-source working condition feature vector is transformed into the initial dynamic threshold prediction result corresponding to the current working condition, so as to realize the real-time threshold generation preparation for working condition adaptation. For example, in the juicer equipment of a production line with a daily output of 2000 liters of fruit and vegetable juice, the obtained 5-dimensional operating condition feature vector is as follows: Input to the parameter-optimized SVR model (assuming...) = , = , = After kernel function similarity calculation and weighted accumulation, the predicted output is obtained. The range of kernel function values ​​in the formula is within... Between these steps, the sums are multiplied by the weights and then a bias is added. = The final predicted value is obtained. This value is maintained after applying upper and lower bound constraints. This provides a stable input for the smooth processing of the next sub-step, effectively avoiding the problem of increased false alarm rate in the fixed threshold method under high temperature and high humidity conditions. S5.3: Perform sliding weighted average filtering on the initial dynamic threshold output by the nonlinear mapping model to eliminate threshold oscillations caused by instantaneous operating condition fluctuations, obtain a smoothed adaptive anomaly detection threshold, and improve the stability of anomaly judgment. The initial dynamic anomaly detection threshold under the current operating conditions output by the support vector regression model is applied using a moving weighted average filtering method (parameter: window length). For 3-20, weighted coefficient sequence ,satisfy Threshold sequence smoothing is implemented to reduce the impact of short-term fluctuations; Furthermore, the exponentially weighted moving average (EWMA) method (parameter: smoothing factor) is used. (0.1-0.9) to enhance the response to the latest data. The larger the value, the faster the filter responds to the latest threshold change; Furthermore, a variance control strategy within a sliding window (parameter: variance threshold) is employed. During the filtering process, the stability of the threshold is detected. When the variance within the window exceeds the set threshold, the weight distribution is dynamically adjusted to converge toward the mean in order to improve stability. Furthermore, the historical threshold sequence generated during the smoothing process is associated with and stored with the current operating condition label, so that the subsequent anomaly detection module can call the smoothing trend parameter under similar operating conditions, thereby improving consistency across time. By combining sliding weighted average filtering with exponential weighted recursion, the high-frequency oscillation component of instantaneous operating condition disturbances on the model output threshold is suppressed within a controllable range, thereby stabilizing the threshold signal at the input end of the anomaly judgment module and improving its anti-interference performance. For example, in the scenario of monitoring juicers on a production line with a daily output of 2000 liters of fruit and vegetable juice, the initial dynamic threshold sequence output by the support vector regression model is: Let the sliding weighted average window length be... = Weighting coefficient Calculate the smoothed value at the third time step using the formula: Furthermore, applying the EWMA method, the smoothing factor... = initial value = ,but = After repeated calculations over multiple time periods, the standard deviation of the output threshold sequence decreased from 0.0048 to 0.0023, successfully suppressing the threshold peak-valley shift caused by transient fluctuations in high temperature and humidity, ensuring that the subsequent anomaly detection module obtains a stable input signal; S5.4: Calculate the deviation between the smoothed adaptive anomaly detection threshold and the real-time sampled value of the current equipment operating parameters to obtain an index of the degree of abnormal deviation of the equipment operating status, so as to quantify the difference between the current operating status and the normal operating condition. S5.5: Based on the abnormal deviation index of the equipment's operating status and the preset judgment rules, execute the abnormal judgment logic, output the judgment result of whether the equipment is in an abnormal state, and transmit the adaptive abnormal detection threshold and the judgment result to the abnormal marking module in a synchronized manner to support the subsequent abnormal signal generation and fault tracing process.

[0016] Step S6: Based on the deviation between the adaptive anomaly detection threshold and the current device operating parameters, determine whether the device is in an abnormal state. If the deviation exceeds the threshold, generate an anomaly marker signal. Specifically, this includes: S6.1: Calculate the difference between the adaptive anomaly detection threshold output by the nonlinear mapping model and the current equipment operating parameters to obtain the real-time deviation value of the equipment operating status; S6.2: Perform time series statistical analysis on the real-time deviation value based on the sliding window mechanism, and calculate its mean and standard deviation within a set time window to assess the persistence and significance of the deviation; For the real-time deviation value sequence of equipment operating status, a sliding window mechanism is adopted (parameter: window length). Step length Time-series truncation processing is performed to achieve local time-domain aggregation of deviation values; Based on each extracted time window, the arithmetic mean method is used to obtain the central trend index of the deviation values ​​within the window. The calculation formula is as follows: in, The mean of the window. For the first One deviation value, The length of the window; Furthermore, the fluctuation intensity of the deviation value within the window is measured using the unbiased sample standard deviation calculation method. The calculation formula is as follows: in, The standard deviation of the window is used to characterize the dispersion of the deviation values; Furthermore, a persistence and significance evaluation vector is constructed by combining the window mean and standard deviation indicators. The formula is as follows: By comparison Deviation threshold from preset mean The relative relationship with the preset stability threshold enables the quantitative determination of the persistence of deviation and the significance of fluctuations in the input data preparation. By using the above-mentioned sliding window truncation, mean calculation, standard deviation calculation and feature vector construction, the real-time deviation value sequence is transformed into a statistical description that can reflect the abnormal stability characteristics of the current operating state, ensuring that a reliable quantitative basis is provided for determining the range of abnormal fluctuations; For example, in the scenario of monitoring the juicer in a production line with a daily output of 2000 liters of fruit and vegetable juice, the real-time sampling frequency of the deviation value is set to 1 Hz, and the sliding window length is... = Step length = The collected deviation sequence for a certain time period is Based on the mean calculation formula, we obtain... = Based on the standard deviation formula, we obtain = When the preset mean deviation threshold is And the stability threshold is When the mean of the window is higher than the mean deviation threshold and the standard deviation is lower than the stability threshold, it indicates that the deviation is a persistent slight deviation with low oscillation, which can provide input conditions with a low probability of false triggering for subsequent anomaly judgment steps. S6.3: Input the mean and standard deviation of the deviation within the time window into the multi-threshold judgment logic module, and determine whether the operating status of the equipment exceeds the normal fluctuation range based on the dynamic confidence interval; S6.4: When the equipment operating status is determined to be outside the normal fluctuation range, an abnormal equipment status marker signal is generated and the current operating condition feature label is attached for subsequent fault tracing and model updates. S6.5: Perform jitter removal and logical verification on the abnormal marker signal, and perform consistency verification based on the judgment results of multiple consecutive time windows to reduce the probability of false triggering and improve the reliability of abnormal judgment.

[0017] Step S7: If an abnormal state is determined, an abnormal equipment state signal is triggered and output to the fault tracing module, while simultaneously recording the current operating condition characteristics and detection threshold to the model update database. Specifically, this includes: S7.1: Based on the comparison result between the deviation value output by the anomaly detection module and the adaptive anomaly detection threshold, determine whether the current operating state of the device meets the anomaly triggering conditions. If it does, generate a device state anomaly event flag as the input basis for subsequent signal triggering and data recording. S7.2: Perform signal encapsulation processing on the generated equipment status abnormal event markers, and use industrial communication protocols to format and encapsulate the abnormal signals to generate abnormal status trigger signals that conform to the interface specifications of the fault tracing module; S7.3: The encapsulated abnormal state trigger signal is transmitted to the fault tracing module via industrial Ethernet to realize real-time notification of abnormal equipment events and automatic start of subsequent analysis processes; The abnormal state trigger signal encapsulated by the industrial communication protocol is used as the input data packet. The industrial Ethernet data link layer frame structure mapping method (parameters: MAC address mapping table, frame type field identifier) ​​is used to realize the physical link encapsulation and addressing mapping of the abnormal signal. Furthermore, the TCP / IP protocol stack transmission scheduling method (parameters: TCP port number, IP address, retransmission threshold) is used. This enables reliable connection establishment and segmented data splicing for abnormal signals at the transport layer, and obtains session data packets with transmission sequence numbers and acknowledgment mechanisms. Furthermore, a data integrity verification and CRC cyclic redundancy code generation method (parameter: CRC32 polynomial generator) is adopted to realize the bit error detection of abnormal signal data packets and generate a check code field to be appended to the end of the data frame to ensure that the data is not tampered with or lost during network transmission. Furthermore, through a real-time priority queue scheduling algorithm (parameter: priority level) to This enables abnormal state trigger signals to be forwarded with high priority over ordinary monitoring data, thereby reducing transmission latency and maintaining millisecond-level response performance. The fault tracing module interface uses a handshake confirmation mechanism (parameter: handshake timeout). ms, maximum number of retries The abnormal state trigger signal transmitted in the previous step is converted into an input event that the tracing module can parse, so as to realize the immediate notification of abnormal events and the automatic start of subsequent analysis programs. For example, in Nissan In a scenario for detecting malfunctions in the juicer of a fruit and vegetable juice production line, the industrial Ethernet transmission bandwidth is assumed to be... Mbps, the default address of the MAC address mapping table for tracing the source module is . - - - - - The TCP port number is set to The retransmission threshold is set to Next, the priority level is set to To ensure real-time response, the CRC32 generator formula is: When vibration deviation of the juicer spindle is detected... When the adaptive anomaly detection threshold is exceeded, an anomaly event flag is generated, and after the aforementioned link mapping, handshake, and priority scheduling, it is then... Within milliseconds, the data was successfully transmitted to the fault tracing module. The CRC verification result showed no errors, triggering the tracing module's real-time analysis process of the bearing lubrication status, ensuring the real-time performance and data integrity of the abnormal response process. S7.4: Pack the working condition feature vector corresponding to the current abnormal event and the adaptive abnormality detection threshold used to generate structured model update samples, which are used as new training data for incremental training of nonlinear mapping models. S7.5: Write the structured model update samples into the specified data table of the model update database, and add timestamp, condition label and anomaly type identifier fields to support data version control and sample validity management during subsequent incremental training.

[0018] Step S8: Periodically perform incremental training on the nonlinear mapping model based on newly added data in the model update database to improve the model's adaptability to new working conditions, while limiting the model update frequency to avoid resource overload. Specifically, this includes: S8.1: Perform data cleaning on the newly added working condition feature vectors and corresponding anomaly detection threshold samples in the model update database to remove duplicate samples, missing values ​​and outliers, and generate a high-quality incremental training sample set that can be used for incremental training. S8.2: Based on the cleaned incremental training sample set, calculate the distribution difference between it and the current nonlinear mapping model training dataset, and use the KL divergence analysis method to evaluate the potential contribution of the new samples to the model's generalization ability, so as to decide whether to perform model update; S8.3: When the distribution difference exceeds the preset threshold, an expanded training set is constructed by merging the incremental training sample set with the original training dataset, and the nonlinear mapping model is fine-tuned by the support vector regression incremental learning algorithm to adapt to the latest working condition change trend. Based on the condition that the completed distribution difference evaluation results exceed the preset threshold, the cleaned incremental training sample set and the original training dataset of the current nonlinear mapping model are merged and expanded (parameters: data merging strategy is timestamp ascending order, working condition label priority alignment rule) to achieve unified formatting and label alignment of different time series data; Furthermore, a sample weight allocation method (parameter: the weight coefficient calculation formula is the weighted product of the new sample proportion coefficient and the feature distribution sparseness coefficient) is used to calculate the training weight of each sample in the expanded training set, so as to give priority to the latest working condition features in the incremental learning algorithm; Furthermore, the Support Vector Regression incremental learning algorithm is invoked (parameters: the kernel function type is Radial Basis Function (RBF), and the penalty factor is...). Initialize to nuclear width Initialize to To avoid the risk of memory overflow caused by training the entire training set at once, the extended training set is loaded and trained in batches. Furthermore, a gradient update method based on sequence blocks is used (parameter: batch training sample size is set to...). The maximum number of iterations is set to The changes in the support vector set are calculated batch by batch, and a support vector truncation strategy is executed after each batch of training to control the model size and update efficiency. Furthermore, a parameter fine-tuning strategy is used (parameter: learning rate decay coefficient). Minimum learning rate threshold Key parameters during incremental training include and Adaptive adjustments are made to ensure that the model can balance its ability to fit historical and new operating conditions. Through the above incremental training and parameter fine-tuning, the extended training results are transformed into a set of nonlinear mapping model weights and support vectors that are highly matched with the latest working condition feature distribution, so as to achieve rapid adaptation of model structure and prediction performance to changes in working conditions. For example, in daily processing In the optimization scenario of malfunction detection of the juicing pump in a fruit and vegetable juice production line, the number of newly added samples in the model update database accounts for a significant portion of the original training dataset. The distribution dissimilarity was calculated using KL divergence. Exceeding the preset threshold Trigger the update process. Expand the total number of samples in the training set to... The RBF kernel function SVR incremental learning algorithm is adopted, with initial... = , = Batch size iteration The training round is completed, and the number of support vectors in the model is reduced from [previous round] using a support vector truncation strategy. Compress to The trained model's mean squared error on the latest validation set has been reduced to [percentage missing]. Compared to before the update, it has decreased. This effectively improves the prediction accuracy and robustness of anomaly detection thresholds under dynamic working conditions; S8.4: Validate the performance of the incrementally trained nonlinear mapping model. Calculate the model prediction error and stability index based on the reserved validation set. If the validation results meet the preset accuracy and robustness requirements, mark the model as the latest available model. S8.5: Based on the current CPU load rate and memory usage rate fed back by the system resource monitoring module, determine whether the resource constraints for model updates are met. If they are met, deploy the latest training model to the runtime environment and update the model version information and update timestamp to control the model update frequency and prevent system resource overload.

[0019] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0020] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for tracing the source of equipment failure in a fruit and vegetable juice production line, characterized in that, Includes the following steps: S1: Collect real-time operating parameters of multiple devices in the fruit and vegetable juice production line under different operating conditions, and record the corresponding operating condition label information; S2: Perform filtering, noise reduction, and normalization preprocessing on the collected operating parameters to generate a standardized equipment operation dataset; S3: Extract multi-dimensional operating condition feature vectors based on the standardized equipment operation dataset; S4: Based on historical abnormal samples and corresponding working condition feature vectors, construct a nonlinear mapping model between working condition features and abnormal detection thresholds, train the nonlinear mapping model using support vector regression algorithm, and output the historical optimal dynamic threshold. S5: Input the currently collected working condition feature vector into the nonlinear mapping model, and output an adaptive anomaly detection threshold that adapts to the current operating state in real time; S6: Based on the degree of deviation between the adaptive anomaly detection threshold and the current device operating parameters, determine whether the device is in an abnormal state. If the deviation exceeds the threshold, generate an anomaly marker signal. S7: If the abnormal state is determined, the equipment status abnormal signal is triggered and output to the fault tracing module, and the current working condition characteristics and detection threshold are recorded to the model update database. S8: Periodically perform incremental training on the nonlinear mapping model based on newly added data in the model update database.

2. The method for tracing equipment failures in a fruit and vegetable juice production line according to claim 1, characterized in that, Step S1 specifically includes: Based on the key equipment configuration in the fruit and vegetable juice production line, a multi-channel sensor network is deployed to synchronously collect temperature, pressure, vibration, speed and motor current signals to obtain raw operating data of the equipment under different operating conditions. The collected raw runtime data is timestamped to generate a structured raw dataset; Based on the mapping relationship between equipment operating status and process flow, a working condition labeling system is defined, and the working condition status is labeled on the collected operating data segments to generate a set of operating data samples with working condition labels. Perform data integrity verification on the operational data sample set with operating condition labels, and identify and mark missing values, outliers, and signal drift points; The verified operational data sample set and its corresponding operating condition label information are stored in the distributed data acquisition database.

3. The method for tracing equipment failures in a fruit and vegetable juice production line according to claim 2, characterized in that, In step S1, the multi-channel synchronous acquisition step uses the IEEE 1588 protocol to distribute the master clock signal, so that the sampling synchronization error of each channel is ≤1μs. The data is buffered by channel and archived with a unified index to obtain high-precision time consistency.

4. The method for tracing equipment failures in a fruit and vegetable juice production line according to claim 1, characterized in that, Step S2 specifically includes: The collected temperature, pressure, vibration, speed and motor current signals are subjected to wavelet threshold denoising to obtain the original equipment operation signal after denoising; Based on the denoised original equipment operation signal, a sliding window mean filtering algorithm is used to smooth the signal to obtain a smoothed equipment operation signal sequence. The smoothed equipment operation signal sequence is subjected to standardization processing based on zero mean and standard deviation normalization to generate normalized equipment operation signal data; Based on the normalized equipment operation signal data, a sliding window mechanism is used to extract the signal statistical features within the window and generate a preliminary structured feature vector. Principal component analysis is performed on the preliminary structured feature vectors to reduce dimensionality, and the dimensionality-reduced standardized equipment operation feature dataset is output.

5. The method for tracing equipment failures in a fruit and vegetable juice production line according to claim 4, characterized in that, The denoising process uses the Daubechies 8th order wavelet basis, with 5 decomposition layers, and the SURE soft thresholding strategy is used for threshold selection.

6. The method for tracing equipment failures in a fruit and vegetable juice production line according to claim 1, characterized in that, Step S3 specifically includes: The preprocessed standardized equipment operation dataset is divided into sliding windows, with each sliding window containing a continuous sequence of equipment operation parameters for a set time length; Calculate the mean value of the equipment operating parameter sequence within each sliding window to obtain the sliding window mean characteristic parameter; Based on the sliding window mean feature parameter, the variance of the equipment operating parameter sequence is calculated to obtain the sliding window variance feature parameter; Perform kurtosis calculation on the sequence of equipment operating parameters within the sliding window to identify whether there are abnormal fluctuations or extreme states, and obtain the kurtosis characteristic parameters of the sliding window. The skewness is calculated based on the sequence of device operating parameters within the sliding window to identify whether there is abnormal behavior biased in a certain direction, and to obtain the skewness feature parameters of the sliding window. Perform Fast Fourier Transform on the sequence of equipment operating parameters within the sliding window to extract frequency domain energy distribution features, identify the vibration or current response modes of the equipment at different frequencies, and obtain frequency domain energy distribution feature vectors. The sliding window mean, variance, kurtosis, skewness, and frequency domain energy distribution characteristic parameters are normalized to generate a standardized multidimensional working condition feature vector. The standardized multidimensional working condition feature vectors are associated with the corresponding working condition label information to construct a structured working condition feature database.

7. The method for tracing equipment failures in a fruit and vegetable juice production line according to claim 6, characterized in that, In step S3, the sliding window length is 256-2048 sampling points, the sliding step size is 1-128 sampling points, and the statistical features include the mean, variance, kurtosis, skewness, and frequency domain energy distribution value obtained by FFT for each window. The feature vector is normalized and PCA dimensionality reduction and then used as the input to the subsequent model.

8. The method for tracing equipment failures in a fruit and vegetable juice production line according to claim 1, characterized in that, Step S4 specifically includes: Feature matching is performed on historical abnormal samples and corresponding working condition feature vectors to construct a supervised learning sample set for training the nonlinear mapping model; Based on the supervised learning sample set, a nonlinear mapping relationship model between the working condition feature vector and the historical best dynamic threshold is constructed using the support vector regression algorithm. The training process of the nonlinear mapping model is subjected to hyperparameter optimization. The performance of the trained nonlinear mapping model is evaluated based on the cross-validation method to verify its fitting accuracy and anomaly threshold prediction ability on historical data. The validated nonlinear mapping model is then deployed to the real-time anomaly detection system.

9. The method for tracing equipment failures in a fruit and vegetable juice production line according to claim 1, characterized in that, Step S5 specifically includes: Obtain the extracted multidimensional operating condition feature vector under the current equipment operating status. The multidimensional operating condition feature vector includes the mean, variance, kurtosis, skewness, and frequency domain energy distribution index within the sliding window. Based on the constructed nonlinear mapping model, a weighted feature fusion calculation is performed on the input multidimensional working condition feature vector to obtain the output value of the anomaly detection threshold mapping function under the current working condition. The initial dynamic threshold output by the nonlinear mapping model is subjected to a sliding weighted average filter to obtain a smoothed adaptive anomaly detection threshold. The deviation between the smoothed adaptive anomaly detection threshold and the real-time sampled value of the current device operating parameters is calculated to obtain an index of the degree of abnormal deviation of the device operating status. Based on the abnormal deviation degree index and the preset judgment rules, the abnormal judgment logic is executed, the judgment result of whether the device is in an abnormal state is output, and the adaptive abnormal detection threshold and the judgment result are synchronously transmitted to the abnormal marking module.

10. A method for tracing equipment failures in a fruit and vegetable juice production line according to claim 1, characterized in that, In step S7, the abnormal signal is encapsulated using an industrial communication protocol, and data integrity is ensured through CRC32 cyclic redundancy check during transmission.

Images