Cable production equipment fault diagnosis method based on machine learning

Abstract

The invention discloses a cable production equipment fault diagnosis method based on machine learning. The method comprises the steps that multi-source time sequence data are collected and preprocessed to generate a standardized data set; executing fractional calculus operation to obtain a fractional response sequence and dynamic memory weight distribution; constructing long-term memory features and forming an exogenous input vector sequence; performing joint training on exogenous input and real observation to obtain a convergent improved NARX neural network model; inputting the exogenous input and historical output lagging sequence into the model, outputting a state prediction value and mapping the state prediction value into a fault risk index; generating a fractional order residual energy index and an abnormal score sequence; and judging and outputting the fault state and the degradation trend grade of the cable production equipment. According to the invention, by introducing the fractional calculus algorithm and improving the NARX neural network model, high-precision fault identification and degradation trend intelligent prediction of cable production equipment under complex working conditions are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and equipment fault diagnosis technology, and in particular to a fault diagnosis method for cable production equipment based on machine learning. Background Technology

[0002] With the development of intelligent manufacturing and industrial automation, the structural complexity and automation level of cable production equipment are constantly increasing. Because cable production is a continuous manufacturing process, cable production equipment is prone to problems such as aging of the transmission system, wear of guide rollers, unstable tension, and electrical component failures under high loads and long-term operation. Traditional equipment maintenance mainly relies on manual inspections or threshold alarms, which are insufficient to detect potential faults in a timely manner, easily leading to serious consequences such as downtime, cable breakage, or quality fluctuations. With the rise of machine learning, machine learning and intelligent diagnostic technologies are gradually being applied to the field of condition monitoring of manufacturing equipment. By extracting features and recognizing patterns from multi-source time-series data, a certain degree of intelligent fault identification can be achieved. However, existing technologies generally rely on static features or shallow models, making it difficult to effectively characterize the dynamic evolution characteristics and long-term correlation patterns of equipment signals, resulting in insufficient prediction accuracy under complex operating conditions and degradation stages.

[0003] To address the operational characteristics of cable production equipment, existing technologies have attempted to introduce models such as recurrent neural networks, long short-term memory networks, or nonlinear autoregressive networks to handle the nonlinear relationships and short-term dynamic dependencies in time series data. These methods, through joint modeling of the input sequence and historical outputs, can improve the predictive ability of fault trends to some extent. However, because equipment operating signals often exhibit significant long-term memory and multi-scale nonlinearity, traditional models have limitations in capturing long-range temporal correlations and dynamic characteristics of degradation processes. Furthermore, model outputs are susceptible to noise and fluctuations in operating conditions, resulting in insufficient predictive stability.

[0004] Therefore, how to provide a machine learning-based fault diagnosis method for cable production equipment is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a machine learning-based fault diagnosis method for cable production equipment. This invention fully utilizes a combination of fractional calculus algorithms and an improved NARX neural network model, detailing the entire process of determining equipment fault status and degradation trend levels under multi-source time-series data conditions. By processing multi-source signals such as temperature, pressure, current, vibration, and tension collected from the cable production equipment using fractional calculus, fractional derivative sequences with long-term memory characteristics and memory kernel convolution features are extracted. These, along with historical output lag sequences, are input into the improved NARX neural network model to establish a nonlinear mapping relationship of the complex dynamic behavior of the cable production equipment, enabling prediction of equipment operating status and degradation trend analysis. This method can simultaneously characterize the long-term dependence and short-term time-series variation of signals, significantly improving the sensitivity of fault diagnosis and the stability of prediction. It possesses advantages such as strong early warning capabilities, good model generalization, and high adaptability to complex operating conditions, and can be widely applied in the fields of health management and predictive maintenance of intelligent manufacturing equipment.

[0006] A fault diagnosis method for cable production equipment based on machine learning according to an embodiment of the present invention includes:

[0007] Collect multi-source time-series data of cable production equipment and record the corresponding fault status of cable production equipment as the actual observation output. Preprocess the multi-source time-series data to generate a standardized dataset.

[0008] Fractional calculus is performed on the standardized dataset. The fractional derivative operator with an adaptive order adjustment factor and a weighted memory kernel function is introduced into the fractional derivative operator defined by Caputo to obtain the fractional response sequence and dynamic memory weight distribution.

[0009] Based on the fractional-order response sequence and dynamic memory weight distribution, fractional-order derivative sequence, memory kernel convolution sequence and fractional-order energy index are extracted to construct long-term memory features and form an exogenous input vector sequence.

[0010] The exogenous input vector sequence and the corresponding real observation output are input into the improved NARX neural network model. The improved NARX neural network model is trained and the parameters are optimized to obtain a converged improved NARX neural network model.

[0011] An improved NARX neural network model that converges by inputting the exogenous input vector sequence and the historical output lag sequence together outputs the state prediction value of the cable production equipment, and the output layer is remapped based on fractional-order feature weights to generate the fault risk index of the cable production equipment.

[0012] The residual sequence between the predicted state value and the corresponding measured value of the cable production equipment is calculated. The residual sequence is then weighted and integrated using a fractional energy function to obtain a fractional residual energy index. This index is then weighted and fused with the fault risk index to generate an anomaly scoring sequence.

[0013] Based on the adaptive calculation of the quantile adaptive threshold using the abnormal scoring sequence, the abnormal scoring sequence is compared with the quantile adaptive threshold to determine and output the fault status and degradation trend level of the cable production equipment.

[0014] Optionally, the multi-source time-series data specifically includes temperature data, pressure data, current data, vibration data, and tension data of the cable production equipment.

[0015] Optionally, the preprocessing of multi-source time-series data specifically includes noise filtering, time synchronization, missing value imputation, normalization, and sampling frequency alignment.

[0016] Optionally, obtaining the fractional-order response sequence and the dynamic memory weight distribution includes:

[0017] The signals in the standardized dataset are divided into channels, and time series of temperature, pressure, current, vibration and tension channels are established respectively. The maximum historical memory length and sliding time window required for fractional-order operation are set to obtain the standardized time series data of each channel.

[0018] An adaptive order adjustment factor is constructed based on the amplitude changes, fluctuation intensity, and frequency band energy of each channel within the sliding time window.

[0019] Based on the adaptive order adjustment factor, the value of the fractional order is automatically adjusted according to the dynamic characteristics of different channel signals within the sliding time window. The initial value of the fractional order is smoothly updated and upper and lower bounds are constrained during the time series progression, forming a fractional order that adapts to time.

[0020] Based on fractional order, a weighted memory kernel function that monotonically decays with time lag is constructed for each channel. The influence weight of historical samples on the current moment within the sliding time window is calculated, and a dynamic memory weight distribution describing the long-term memory characteristics of each channel signal is generated. The historical samples refer to the historical standardized time series data of the cable production equipment within the sliding time window.

[0021] Based on the fractional order and dynamic memory weight distribution of each channel, a fractional operator is constructed using the fractional derivative defined by Caputo. Fractional differential operations are performed on the standardized time series data of each channel to obtain the fractional response sequence of each channel that reflects the long-term dependence and time series variation of the signal.

[0022] Optionally, the construction of long-term memory features and the formation of an exogenous input vector sequence includes:

[0023] The fractional response signals of each channel are normalized and time-synchronized according to the dynamic memory weights to generate a fractional feature matrix, and boundary consistency correction is performed at the time boundaries of each channel.

[0024] Based on the fractional characteristic matrix and the corresponding fractional order of each channel, the fractional derivative sequence of each channel, which characterizes the nonlinear variation law of the operating signal of the cable production equipment in the time dimension, is calculated.

[0025] Based on the fractional-order feature matrix and dynamic memory weight distribution of each channel, a weighted convolution operation is performed on the signal of each channel to generate a memory kernel convolution sequence;

[0026] Based on the fractional derivative sequences and memory kernel convolution sequences of each channel, a fractional energy index representing the cumulative energy intensity of the signal's fractional-order response is calculated within a set sliding time window. Specifically, the calculation of the fractional-order energy index involves:

[0027] The fractional derivative sequences and memory kernel convolution sequences within the sliding time window of each channel are normalized and detrended. The fractional response intensity at each time point within the window is weighted and accumulated according to the dynamic memory weight. Robust truncation or piecewise penalty is introduced to limit the amplitude of abnormal values. After completing the weighting and robust processing, the cumulative fractional response intensity within the corresponding sliding time window is calculated as the fractional energy index.

[0028] The fractional derivative sequence, memory kernel convolution sequence, and fractional energy index are fused to form a comprehensive long-term memory feature set through time step alignment and feature concatenation.

[0029] The comprehensive long-term memory feature set is unfolded into an exogenous input vector sequence in chronological order.

[0030] Optionally, the improved NARX neural network model that achieves convergence includes:

[0031] An improved NARX neural network model is constructed, which consists of a fractional-order memory-gated layer, a structured exogenous fusion layer, a delay-shared basis hysteresis mapping layer, a residual feedforward backbone layer, and a parallel output branch, wherein:

[0032] Fractional memory gating layers characterize the long-term dependency characteristics of operating signals of cable production equipment through fractional memory factors;

[0033] The structured exogenous fusion layer fuses exogenous input vector sequences to generate fused high-dimensional temporal features and extracts the dynamic coupling relationship between multiple physical quantities;

[0034] The delayed shared base lag mapping layer uniformly encodes the historical output lag sequence and establishes short-term time dependencies. The historical output lag sequence refers to the sequence formed by automatically caching the predicted outputs of the improved NARX neural network model at several previous time steps and arranging them in reverse chronological order.

[0035] The residual feedforward backbone layer performs nonlinear mapping and error correction on the fused high-dimensional temporal features;

[0036] The parallel output branches include a state prediction branch and an uncertainty output branch. The state prediction branch outputs the fault state prediction result, and the uncertainty output branch outputs the prediction variance information corresponding to the result.

[0037] Training samples are constructed based on standardized datasets and corresponding real observation outputs, and the real observation outputs at corresponding times are used as supervision signals to form a training sample sequence arranged in chronological order.

[0038] The training objectives and constraints are set, and the prediction error is measured by heteroscedastic regression. The error term is dynamically weighted based on the uncertainty output of the improved NARX neural network model. At the same time, the prediction sequence is subject to smoothing constraints, structural constraints, and process boundary constraints.

[0039] A gradient-based optimization method is used to iteratively update the parameters of the improved NARX neural network model. At the same time, the channel order parameters of the fractional memory gate layer are jointly updated. The values ​​of the channel order parameters are limited to zero and one. During the iteration process, the loss value is gradually reduced according to the joint objective of training and constraints until convergence on the training set.

[0040] The improved NARX neural network model is evaluated on an independent validation set. Based on the comprehensive index of the validation set and the early stopping criterion, the convergent network weight parameters and channel order parameters are determined. The model structure and parameters are fixed to obtain a convergent improved NARX neural network model.

[0041] Optionally, the failure risk index of the generated cable production equipment includes:

[0042] The exogenous input vector sequence and the historical output lag sequence are used as the input at the current time. Fractional-order memory gating layer processing is performed on the exogenous input vector sequence. The historical segments are weighted and converged according to the variable-order parameters of each channel, the fractional-order memory factor and the dynamic memory weight to obtain a memory-enhanced representation containing long-term memory information.

[0043] The memory-enhanced representation is input into the structured exogenous fusion layer, and multi-level fusion is performed in the channel dimension and time dimension according to the preset structured weights to generate fusion features that represent the dynamic coupling relationship of multiple physical quantities.

[0044] The historical output lag sequence is input into the delayed shared basis lag mapping layer, and the output of different lag steps is uniformly encoded according to the shared delay weight to obtain the lag mapping representation that represents the short-term time dependence.

[0045] The fused features and the hysteresis mapping representation are cascaded and then input into the residual feedforward backbone layer. Nonlinear mapping and error correction are completed through multiple residual units to obtain a high-order implicit representation.

[0046] The parallel output branch maps the higher-order implicit representation to obtain the state prediction value and the corresponding prediction variance information of the cable production equipment. The state prediction value is calculated by the state prediction branch, and the prediction variance is calculated by the uncertainty output branch.

[0047] Based on fractional-order feature weights, the output layer remapping is performed on the state prediction values ​​to weight and integrate the fractional-order long-term memory information with the current prediction results to generate the fault risk index of cable production equipment.

[0048] Output the predicted state value, prediction variance information and fault risk index, and write the current state prediction value into the prediction cache to update the historical output lag sequence. At the same time, record the time identifier for easy retrieval at the next time step.

[0049] Optionally, generating the abnormal scoring sequence includes:

[0050] The difference between the predicted state value and the corresponding measured value is calculated to generate the residual data at the current moment. The residual data at multiple consecutive moments are collected within a preset sliding time window to form a residual sequence. The corresponding measured value refers to the actual operating data collected synchronously at the prediction moment of the converged improved NARX neural network model, specifically including temperature data, pressure data, current data, vibration data and tension data of the cable production equipment.

[0051] Based on the dynamic memory weight distribution, a memory weight is set for each lag step. The memory weight is determined by the order parameter of the fractional derivative and the weighted memory kernel function, which describes the degree of influence of the residual at different time steps on the current energy calculation.

[0052] The residual sequence is weighted and squared to obtain the fractional residual energy index at the current time step by performing weighted summation based on memory weights, which represents the weighted cumulative strength of the residual in the time dimension.

[0053] Based on the prediction variance information, the fractional residual energy index is standardized. The standardized fractional residual energy index and the fault risk index are aligned by time and then risk-weightedly fused to obtain the standardized anomaly score.

[0054] The standardized anomaly scores obtained at each time point are arranged in chronological order to form an anomaly score sequence and then output.

[0055] Optionally, determining and outputting the fault status and degradation trend level of the cable production equipment includes:

[0056] Within the sliding time window, the distribution characteristics of abnormal scoring sequences are statistically analyzed to form a benchmark sample set, and the benchmark sample set is segmented and managed according to different process stages based on the production cycle.

[0057] The quantile values ​​are calculated for the benchmark sample set according to the preset quantile ratio to obtain the initial quantile threshold. The initial quantile threshold is then smoothly updated based on the change range and stability of abnormal scores within the sliding time window to form the quantile adaptive threshold.

[0058] The current abnormal score is compared with the quantile adaptive threshold. If the abnormal score exceeds the quantile adaptive threshold, a threshold overrun event is recorded, and the number of consecutive overruns and the duration are accumulated.

[0059] Based on the number of consecutive boundary crossings, duration, and prediction variance information at the corresponding time, a comprehensive judgment rule is executed to output the fault status judgment result, and the degradation trend is classified according to the boundary crossing magnitude and persistence.

[0060] Write the fault status determination result and degradation trend level into the diagnostic log, update the reference sample set and statistics of the quantile adaptive threshold, and output the fault status determination result, degradation trend and time identifier to the upper system.

[0061] The beneficial effects of this invention are:

[0062] This invention constructs an intelligent fault diagnosis system that combines long-term memory modeling and nonlinear dynamic mapping capabilities by integrating fractional-order calculus algorithms with an improved NARX neural network model. This effectively enhances the fault diagnosis and degradation trend prediction capabilities of cable production equipment under complex operating conditions. Compared with existing diagnostic methods that rely on single signal thresholds or static feature modeling, this invention utilizes multi-source time-series data fusion and fractional-order feature extraction mechanisms to fully explore the multidimensional time correlation and potential dynamic characteristics during the operation of cable production equipment, achieving simultaneous characterization of long-term degradation patterns and short-term fluctuation anomalies. By introducing an adaptive order adjustment factor and a weighted memory kernel function into the fractional-order derivative operator defined by Caputo, the improved NARX neural network model can dynamically adjust the memory depth and response speed, thereby accurately characterizing the historical dependence and nonlinear time-varying characteristics of cable production equipment signals.

[0063] The improved NARX neural network model introduces a structured exogenous fusion layer, a delayed shared basis hysteresis mapping layer, and a residual feedforward backbone layer to establish a nonlinear causal relationship model across multiple time scales. This structure, combined with heteroscedastic regression and uncertainty output branches, adaptively balances the confidence level of prediction errors under different operating conditions, achieving robust prediction and dynamic judgment of abnormal intervals for fault diagnosis. Through fractional residual energy analysis and quantile adaptive thresholding, this invention achieves sensitive detection of early, minor faults and graded determination of degradation trends. This method significantly improves the accuracy and stability of fault identification in cable production equipment, possessing strong robustness, good adaptability, and real-time online updating capabilities. It overcomes the technical bottlenecks of traditional methods, such as low identification accuracy, prediction lag, and weak model generalization ability in noisy environments, providing effective technical support for intelligent operation and maintenance and predictive maintenance of cable production lines. Attached Figure Description

[0064] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0065] Figure 1 This is a flowchart of a fault diagnosis method for cable production equipment based on machine learning proposed in this invention;

[0066] Figure 2 This is a schematic diagram of the improved NARX neural network model structure of a machine learning-based fault diagnosis method for cable production equipment proposed in this invention. Detailed Implementation

[0067] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0068] refer to Figure 1 and Figure 2 A machine learning-based fault diagnosis method for cable production equipment includes:

[0069] Collect multi-source time-series data of cable production equipment and record the corresponding fault status of cable production equipment as the actual observation output. Preprocess the multi-source time-series data to generate a standardized dataset.

[0070] Fractional calculus is performed on the standardized dataset. The fractional derivative operator with an adaptive order adjustment factor and a weighted memory kernel function is introduced into the fractional derivative operator defined by Caputo to obtain the fractional response sequence and dynamic memory weight distribution.

[0071] Based on the fractional-order response sequence and dynamic memory weight distribution, fractional-order derivative sequence, memory kernel convolution sequence and fractional-order energy index are extracted to construct long-term memory features and form an exogenous input vector sequence.

[0072] The exogenous input vector sequence and the corresponding real observation output are input into the improved NARX neural network model. The improved NARX neural network model is trained and the parameters are optimized to obtain a converged improved NARX neural network model.

[0073] An improved NARX neural network model that converges by inputting the exogenous input vector sequence and the historical output lag sequence together outputs the state prediction value of the cable production equipment, and the output layer is remapped based on fractional-order feature weights to generate the fault risk index of the cable production equipment.

[0074] The residual sequence between the predicted state value and the corresponding measured value of the cable production equipment is calculated. The residual sequence is then weighted and integrated using a fractional energy function to obtain a fractional residual energy index. This index is then weighted and fused with the fault risk index to generate an anomaly scoring sequence.

[0075] Based on the adaptive calculation of the quantile adaptive threshold using the abnormal scoring sequence, the abnormal scoring sequence is compared with the quantile adaptive threshold to determine and output the fault status and degradation trend level of the cable production equipment.

[0076] In this embodiment, the multi-source time-series data specifically includes temperature data, pressure data, current data, vibration data, and tension data of the cable production equipment.

[0077] In this embodiment, the preprocessing of multi-source time-series data specifically includes noise filtering, time synchronization, missing value imputation, normalization, and sampling frequency alignment.

[0078] In this embodiment, obtaining the fractional-order response sequence and the dynamic memory weight distribution includes:

[0079] The signals in the standardized dataset are divided into channels, and time series of temperature, pressure, current, vibration and tension channels are established respectively. The maximum historical memory length and sliding time window required for fractional-order operation are set to obtain the standardized time series data of each channel.

[0080] An adaptive order adjustment factor is constructed based on the amplitude changes, fluctuation intensity, and frequency band energy of each channel within the sliding time window.

[0081] Based on the adaptive order adjustment factor, the value of the fractional order is automatically adjusted according to the dynamic characteristics of different channel signals within the sliding time window. The initial value of the fractional order is smoothly updated and upper and lower bounds are constrained during the time series progression, forming a fractional order that adapts to time.

[0082] Based on fractional order, a weighted memory kernel function that monotonically decays with time lag is constructed for each channel. The influence weight of historical samples within the sliding time window on the current moment is calculated, and a dynamic memory weight distribution describing the long-term memory characteristics of each channel signal is generated. The seven historical samples refer to the historical standardized time-series data of the cable production equipment within the sliding time window. The generation of the dynamic memory weight distribution is specifically as follows:

[0083] Historical standardized time series data of each channel are selected within a sliding time window. The fractional order of the channel is determined based on the fluctuation intensity and energy distribution of the historical standardized time series data. A weighted memory kernel with a larger time lag and a smaller weight is constructed based on the fractional order. Finally, all weights are normalized to form a dynamic memory weight distribution that describes the long-term memory characteristics of each channel signal.

[0084] Based on the fractional order and dynamic memory weight distribution of each channel, a fractional operator is constructed using the fractional derivative defined by Caputo. Fractional differentiation operations are then performed on the standardized time-series data of each channel to obtain the fractional response sequences of each channel that reflect the long-term dependence and temporal variation patterns of the signal. Specifically, the fractional response sequences of each channel are obtained as follows:

[0085] Based on dynamic memory weight distribution, a fractional order matching the memory strength is configured for each channel. Within the sliding time window, the standardized time series data of the corresponding channel are numerically discretized according to the fractional derivative defined by Caputo. Samples from closer time moments are given higher weights, while samples from farther time moments gradually decay according to their weights and participate in the accumulation. The calculated time series is the fractional response sequence of the channel.

[0086] In this embodiment, the construction of long-term memory features and the formation of an exogenous input vector sequence includes:

[0087] The fractional response signals of each channel are normalized and time-synchronized according to the dynamic memory weights to generate a fractional feature matrix, and boundary consistency correction is performed at the time boundaries of each channel.

[0088] Based on the fractional-order characteristic matrix and corresponding fractional-order order of each channel, the fractional-order derivative sequence representing the nonlinear variation law of the cable production equipment operating signal in the time dimension is calculated. Specifically, the calculation of the fractional-order derivative sequence is as follows:

[0089] Based on the fractional-order feature matrix and fractional-order order of the channel, the fractional-order feature matrix is ​​accumulated by fractional-order difference in time order within the sliding time window, and dynamic memory weights are used to control the participation intensity of near and far samples; the channel-level fractional-order derivative value is output at each time step, and time alignment and amplitude normalization are performed to form a continuous fractional-order derivative sequence.

[0090] Based on the fractional-order feature matrix and dynamic memory weight distribution of each channel, a weighted convolution operation is performed on the signal of each channel to generate a memory kernel convolution sequence;

[0091] Based on the fractional derivative sequence and memory kernel convolution sequence of each channel, the fractional energy index representing the cumulative energy intensity of the signal fractional response is calculated within a set sliding time window.

[0092] The fractional derivative sequence, memory kernel convolution sequence, and fractional energy index are fused to form a comprehensive long-term memory feature set through time step alignment and feature concatenation.

[0093] The comprehensive long-term memory feature set is unfolded into an exogenous input vector sequence in chronological order.

[0094] In this embodiment, obtaining the convergent improved NARX neural network model includes:

[0095] An improved NARX neural network model is constructed, which consists of a fractional-order memory-gated layer, a structured exogenous fusion layer, a delay-shared basis hysteresis mapping layer, a residual feedforward backbone layer, and a parallel output branch, wherein:

[0096] Fractional memory gating layers characterize the long-term dependency characteristics of operating signals of cable production equipment through fractional memory factors;

[0097] The structured exogenous fusion layer fuses exogenous input vector sequences to generate fused high-dimensional temporal features and extracts the dynamic coupling relationship between multiple physical quantities;

[0098] The delayed shared base lag mapping layer uniformly encodes the historical output lag sequence and establishes short-term time dependencies. The historical output lag sequence refers to the sequence formed by automatically caching the predicted outputs of the improved NARX neural network model at several previous time steps and arranging them in reverse chronological order.

[0099] The residual feedforward backbone layer performs nonlinear mapping and error correction on the fused high-dimensional temporal features;

[0100] The parallel output branches include a state prediction branch and an uncertainty output branch. The state prediction branch outputs the fault state prediction result, and the uncertainty output branch outputs the prediction variance information corresponding to the result.

[0101] The construction of the improved NARX neural network model specifically involves:

[0102] A fractional-order memory-gated layer is added before the input layer of the NARX neural network model to enhance long-term dependence on the exogenous input vector sequence by introducing a fractional-order memory factor. The input layer of the NARX neural network model is replaced with a structured exogenous fusion layer, which constructs a structured weight matrix based on the mutual information strength between channels and performs weighted fusion in the channel and time dimensions to output fused high-dimensional temporal features. Shared delay weights are introduced into the lag input path of the NARX neural network model to form a delayed shared basis lag mapping layer, which uniformly encodes the historical output lag sequence to generate a lag mapping representation representing short-term time dependence. In the hidden layer of the NARX neural network model, a single fully connected mapping is replaced by several residual units in series to form a residual feedforward backbone layer, which completes nonlinear mapping and error correction through multiple residual units. A parallel output branch is added to the output layer of the NARX neural network model, one branch outputs the fault state prediction result, and the other branch outputs the prediction variance information corresponding to the prediction result.

[0103] Training samples are constructed based on standardized datasets and corresponding real observation outputs, and the real observation outputs at corresponding times are used as supervision signals to form a training sample sequence arranged in chronological order.

[0104] Training objectives and constraints are set, and heteroscedastic regression is used to measure prediction error. The error term is dynamically weighted based on the uncertainty output of the improved NARX neural network model. Simultaneously, smoothing constraints, structural constraints, and process boundary constraints are applied to the predicted sequence, where:

[0105] Smoothing constraints limit the magnitude of changes in predicted values ​​between adjacent time points; structural constraints make the weights of the delayed shared basis hysteresis mapping layer approach a blocky Toplitz low-rank structure; process boundary constraints impose a secondary penalty on predicted values ​​that exceed the upper and lower limits of the temperature.

[0106] A gradient-based optimization method is used to iteratively update the parameters of the improved NARX neural network model. At the same time, the channel order parameters of the fractional memory gate layer are jointly updated. The values ​​of the channel order parameters are limited to zero and one. During the iteration process, the loss value is gradually reduced according to the joint objective of training and constraints until convergence on the training set.

[0107] The improved NARX neural network model is evaluated on an independent validation set. Based on the comprehensive index of the validation set and the early stopping criterion, the convergent network weight parameters and channel order parameters are determined. The model structure and parameters are fixed to obtain a convergent improved NARX neural network model.

[0108] In this embodiment, the generation of the fault risk index for the cable production equipment includes:

[0109] The exogenous input vector sequence and the historical output lag sequence are used together as the input at the current time. A fractional-order memory-gated layer is applied to the exogenous input vector sequence. Historical segments are weighted and converged based on the variable-order parameters of each channel, the fractional-order memory factor, and the dynamic memory weight, resulting in a memory-enhanced representation that includes long-term memory information.

[0110] The variable order parameter is a specific order value that is dynamically changed at different time steps and channels, which is adaptively determined by the adaptive order adjustment factor in fractional derivative operations.

[0111] The fractional memory factor is embedded in the calculation formula of the dynamic memory weight, which controls the degree of influence of historical fragments on the current input. When performing weighted aggregation, each sample in time lag step is multiplied by a memory weight that decays monotonically with the lag time, and the decay rate and weight distribution are determined by the fractional memory factor.

[0112] Historical segments refer to standardized data samples formed by weighted aggregation of exogenous input vector sequences within a sliding time window prior to the current moment and historical output lag sequences;

[0113] The memory-enhanced representation is input into the structured exogenous fusion layer, and multi-level fusion is performed in the channel and time dimensions according to preset structured weights to generate fusion features representing the dynamic coupling relationship of multiple physical quantities. Specifically, the generation of fusion features is as follows:

[0114] Channel feature encoding is performed on the input memory augmentation representations of each channel to extract the main change trend and energy features of each channel in the current time period. The structured exogenous fusion layer performs weighted superposition of the memory augmentation representations of each channel in the channel dimension according to the preset structured weight matrix. The weights are determined by the mutual information strength between channels.

[0115] In the time dimension, the memory-enhanced representations of adjacent time steps are smoothly weighted to form a temporal continuity constraint, generating a fused multi-layer weighted feature.

[0116] The fused multi-layer weighted features are then normalized and nonlinearly transformed to obtain a stable and robust fused feature representation.

[0117] The historical output lag sequence is input into the delayed shared base lag mapping layer. Based on the shared delay weights, the outputs of different lag steps are uniformly encoded to obtain a lag mapping representation characterizing short-term time dependence. Specifically, the uniform encoding of the outputs of different lag steps involves:

[0118] The historical output lag sequence is standardized and its amplitude is normalized. The delay shared base lag mapping layer configures the same set of shared delay weights for all lag step outputs and performs weighted convergence in the time axis direction. The historical information of different lag steps is uniformly encoded and projected in the same weight space to form a lag mapping feature representation with temporal consistency.

[0119] The cascaded features and hysteresis mapping representations are then input into the residual feedforward backbone layer. Nonlinear mapping and error correction are performed through multiple residual units to obtain a higher-order implicit representation. Specifically, obtaining the higher-order implicit representation involves:

[0120] The fused features and hysteresis mapping representations are first processed by a nonlinear transformation unit composed of linear projection and activation function to extract high-order combined features, and then the stability of high-order combined features is enhanced by layer normalization and small-scale noise suppression.

[0121] The higher-order combined features are formed into progressively refined higher-order representations through a series of multi-layer residual units; taking the heteroscedasticity weighted loss as the target, combined with the predictive sequence smoothing constraints, structural constraints and process boundary constraints, the sample error is compressed into the residual increment by backpropagation to obtain the higher-order implicit representation;

[0122] The parallel output branch maps the higher-order implicit representation to obtain the state prediction value and the corresponding prediction variance information of the cable production equipment. The state prediction value is calculated by the state prediction branch, and the prediction variance is calculated by the uncertainty output branch.

[0123] Based on fractional-order feature weights, the output layer remapping is performed on the state prediction values ​​to weight and integrate the fractional-order long-term memory information with the current prediction results to generate the fault risk index of cable production equipment.

[0124] Output the predicted state value, prediction variance information and fault risk index, and write the current state prediction value into the prediction cache to update the historical output lag sequence. At the same time, record the time identifier for easy retrieval at the next time step.

[0125] In this embodiment, generating the abnormal scoring sequence includes:

[0126] The difference between the predicted state value and the corresponding measured value is calculated to generate the residual data at the current moment. The residual data at multiple consecutive moments are collected within a preset sliding time window to form a residual sequence. The corresponding measured value refers to the actual operating data collected synchronously at the prediction moment of the converged improved NARX neural network model, specifically including temperature data, pressure data, current data, vibration data and tension data of the cable production equipment.

[0127] Based on the dynamic memory weight distribution, a memory weight is set for each lag step. The memory weight is determined by the order parameter of the fractional derivative and the weighted memory kernel function, which describes the degree of influence of the residual at different time steps on the current energy calculation.

[0128] The residual sequence is weighted and squared to obtain the fractional residual energy index at the current time step by performing weighted summation based on memory weights, which represents the weighted cumulative strength of the residual in the time dimension.

[0129] Based on the prediction variance information, the fractional residual energy index is standardized. The standardized fractional residual energy index is then aligned with the fault risk index over time and fused using risk weighting to obtain a standardized anomaly score. Specifically, obtaining the standardized anomaly score involves:

[0130] The standardized fractional residual energy index and the fault risk index are normalized to maintain the consistency of dimensions. The weighting coefficients are dynamically adjusted according to the magnitude of the risk index. When the risk index is large, the weight of the standardized fractional residual energy index increases, and the weight of the fault risk index increases. During the fusion process, the fault risk index also acts as an amplification factor to monotonically amplify the anomaly score. Finally, the comprehensive anomaly score is obtained by weighted summation.

[0131] The standardized anomaly scores obtained at each time point are arranged in chronological order to form an anomaly score sequence and then output.

[0132] In this embodiment, determining and outputting the fault status and degradation trend level of the cable production equipment includes:

[0133] Within the sliding time window, the distribution characteristics of abnormal scoring sequences are statistically analyzed to form a benchmark sample set, and the benchmark sample set is segmented and managed according to different process stages based on the production cycle.

[0134] The quantile values ​​of the benchmark sample set are calculated according to a preset quantile ratio to obtain an initial quantile threshold. The initial quantile threshold is then smoothly updated based on the variation and stability of abnormal scores within a sliding time window to form an adaptive quantile threshold. Specifically, the calculation of quantile values ​​of the benchmark sample set according to the preset quantile ratio involves:

[0135] A certain number of abnormal rating samples are collected within the sliding time window and formed into a baseline sample set in chronological order. According to the preset quantile ratio, the abnormal ratings of the samples are sorted from smallest to largest, and the values ​​at the corresponding positions are selected as the quantile values ​​of the corresponding sliding time window.

[0136] The current abnormal score is compared with the quantile adaptive threshold. If the abnormal score exceeds the quantile adaptive threshold, a threshold overrun event is recorded, and the number of consecutive overruns and the duration are accumulated.

[0137] Based on the number of consecutive boundary crossings, duration, and prediction variance information at the corresponding time, a comprehensive judgment rule is executed to output the fault status judgment result, and the degradation trend is classified according to the boundary crossing magnitude and persistence.

[0138] Write the fault status determination result and degradation trend level into the diagnostic log, update the reference sample set and statistics of the quantile adaptive threshold, and output the fault status determination result, degradation trend and time identifier to the upper system.

[0139] Example 1:

[0140] To verify the feasibility of this invention in practice, it was applied to the entire process of fault diagnosis and health management of a 110kV cross-linked polyethylene cable production line at a special cable manufacturing plant in a certain city. This production line includes multiple continuous processes such as conductor stranding, extrusion, cooling, and winding. Twelve sets of multi-source sensor monitoring terminals were installed on-site to collect real-time multi-dimensional time-series data on the cable production equipment, including temperature, pressure, current, vibration, and tension, combined with operating status information and maintenance logs recorded by the process control system. All collected data was accessed to a centralized data control platform via industrial Ethernet, and after standardized preprocessing, it entered a joint analysis process using fractional calculus and an improved NARX neural network model.

[0141] During the pilot phase, from late June to early July, production operated at high loads continuously with ambient temperatures reaching 36°C, causing slight tension fluctuations in multiple winding machines. Traditional threshold monitoring systems could not accurately distinguish between normal fluctuations and potential bearing degradation signals. This invention's system performs real-time analysis of 1.8TB of multi-source time-series data over 10 consecutive days, automatically extracting fractional derivatives and energy characteristics, and dynamically judging abnormal intervals using an adaptive quantile threshold model. At 10:22 AM on July 3rd, the system first identified an increase in the fractional residual energy of the cooling section vibration signal and issued a potential imbalance warning for the cooling water pump 12 minutes later. Manual inspection confirmed that the pump impeller had an eccentricity of approximately 2.5 mm. After replacement and adjustment, the equipment vibration RMS value decreased from 1.12 mm / s to 0.37 mm / s, and the production line returned to stable operation.

[0142] Comparative analysis shows that after adopting the method of this invention, the average early warning time for equipment failure reaches 42 hours, the accuracy rate of anomaly detection increases to 96.3%, and the false alarm rate decreases to 5.4%, which is about 20 percentage points higher than the traditional experience-based judgment method. During a three-month operation period, unplanned production line downtime decreased by 23%, the average equipment utilization rate increased by 8%, and the product defect rate in the extrusion section decreased from 4.1% to 2.9%. Experimental results demonstrate that this invention can accurately identify potential equipment failures and degradation trends in complex, multi-condition cable production environments, achieving intelligent fault diagnosis with high precision and low false alarm rate.

[0143] Table 1. Data Table of Key Anomaly Identification and Handling Process in Pilot Production Line

[0144] Date and Time Process / Equipment Core monitoring quantity (unit) Real-time / Forecast Anomaly score S(t) Quantile adaptive threshold Predictive variance Judgment result Disposal and Confirmation Early warning lead time 2025-07-0310:22 Cooling section / water pump Vibration RMS (mm / s): 1.12 / 0.78 3.21 2.35 0.09 Warning Impeller eccentricity 2.5 mm, cleaning and correction required. 12 min 2025-07-0514:08 take-up machine / bearing Tension fluctuation (%): ±7.9 / ±5.1 2.84 2.10 0.07 Warning The clearance increased by 12%, so the bearing was replaced. 3 h 2025-07-0809:45 Extruder / Heating Chamber Temperature (°C): 102.6 / 98.9 2.42 2.05 0.06 abnormal Thermocouple aging, calibration and replacement required. 5 h 2025-07-1209:34 take-up machine / bearing Peak vibration (g): 0.78 / 0.52 3.46 2.40 0.11 abnormal Confirmation of bearing inner ring degradation 43 min 2025-07-1516:20 conductor stranding Motor current (A): 86.4 / 80.2 2.15 1.95 0.05 Warning Belt loosening, tension adjustment 2 h 2025-07-1811:10 Cooling section / water circuit Cooling water pressure (MPa): 0.18 / 0.23 2.01 1.88 0.05 Warning Clogged filter, clean 1 h 2025-07-2107:52 Take-up tension control Tension (kg): 32.8 / 30.6 2.63 2.06 0.06 abnormal Tension sensor drift 2.1% 1.5 h 2025-07-2413:05 Extruder / Screw Frame vibration (mm / s): 1.35 / 0.96 3.08 2.32 0.08 abnormal Screw end face wear 4 h 2025-07-2710:41 Cooling section / spray Wire diameter deviation (mm): +0.12 / +0.05 1.92 1.70 0.04 Warning Uneven spray distribution, adjust 50 min 2025-07-3018:26 Guide wheel system Shaft end vibration (g): 0.66 / 0.43 2.77 2.18 0.07 abnormal The guide wheel is eccentric by 0.6 mm. 3 h 2025-08-0308:18 Electrical cabinet Power factor: 0.82 / 0.87 1.74 1.60 0.04 Warning The contactor's contact resistance is too high. 2 h 2025-08-0815:12 take-up machine / servo Speed ​​fluctuation (%): ±4.5 / ±2.8 2.29 1.98 0.05 Warning Servo gain resetting 1 h 2025-08-1209:02 Cooling section / water pump Vibration RMS (mm / s): 1.05 / 0.76 3.02 2.28 0.08 abnormal Shaft seal worn, replace 35 min 2025-08-1611:37 Extrusion die Exit temperature difference (°C): 6.1 / 3.9 2.36 2.00 0.06 Warning Uneven heat distribution, air distribution optimization 1.2 h

[0145] As shown in Table 1, a total of 14 valid alarm events were recorded during the pilot period, of which 8 were judged as warnings and 6 as anomalies, covering key links such as the take-up machine, extruder, and cooling section. Summarized by process, the cooling section had the most related events (4), followed by the take-up machine (3) and the extruder (2), consistent with the characteristics of high summer temperatures and fluctuating cooling loads. Quantitatively, the anomaly score S(t) ranged from 1.74 to 3.46, all higher than the quantile adaptive threshold (1.60–2.40), with an average score of 2.56, indicating that the fractional residual energy-driven discrimination has a stable margin for exceeding limits. The average fractional residual energy was 0.093 (0.056–0.133), significantly higher in events involving rotating equipment such as pumps and bearings, reflecting sensitivity to mechanical degradation. Regarding uncertainty, the prediction variance is concentrated between 0.04 and 0.11, with a mean of approximately 0.065, indicating that the confidence intervals given by the parallel uncertainty branches are relatively convergent, which is beneficial for the heteroscedasticity weighting and robustness determination of S43. The warning lead time ranges from 12 minutes to 300 minutes, with an average of approximately 112 minutes; short lead times are more common in sudden changes in operating conditions (such as eccentric cleaning of water pumps), while long lead times are more common in slow drift problems (such as heating chamber temperature and screw end face wear). In terms of handling and confirmation items, bearing-related problems account for a high proportion (one case each of increased clearance and inner ring degradation), two cases each of eccentricity and wear, and other types of anomalies such as sensor drift, filter blockage, uneven heat distribution, uneven spray distribution, excessive contact resistance, and servo gain misalignment, indicating that the method can maintain consistent detection capability under multiple physical quantities and multiple fault mechanisms.

[0146] In summary, the coupling of fractional calculus operations and the improved NARX neural network model enables the scoring threshold link to still have discriminative power in high-noise scenarios. The cooling and line-collecting stages are the main risk points. The prediction variance of the online output provides a quantitative basis for the credibility of the alarm. The overall diagnostic link has interpretable and implementable engineering characteristics.

[0147] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fault diagnosis method for cable production equipment based on machine learning, characterized in that, include: Collect multi-source time-series data of cable production equipment and record the corresponding fault status of cable production equipment as the actual observation output. Preprocess the multi-source time-series data to generate a standardized dataset. Fractional calculus is performed on the standardized dataset. The fractional derivative operator with an adaptive order adjustment factor and a weighted memory kernel function is introduced into the fractional derivative operator defined by Caputo to obtain the fractional response sequence and dynamic memory weight distribution. Based on the fractional-order response sequence and dynamic memory weight distribution, fractional-order derivative sequence, memory kernel convolution sequence and fractional-order energy index are extracted to construct long-term memory features and form an exogenous input vector sequence. The exogenous input vector sequence and the corresponding real observation output are input into the improved NARX neural network model. The improved NARX neural network model is trained and the parameters are optimized to obtain a converged improved NARX neural network model. An improved NARX neural network model that converges by inputting the exogenous input vector sequence and the historical output lag sequence together outputs the state prediction value of the cable production equipment, and the output layer is remapped based on fractional-order feature weights to generate the fault risk index of the cable production equipment. The residual sequence between the predicted state value and the corresponding measured value of the cable production equipment is calculated. The residual sequence is then weighted and integrated using a fractional energy function to obtain a fractional residual energy index. This index is then weighted and fused with the fault risk index to generate an anomaly scoring sequence. Based on the adaptive calculation of the quantile adaptive threshold using the abnormal scoring sequence, the abnormal scoring sequence is compared with the quantile adaptive threshold to determine and output the fault status and degradation trend level of the cable production equipment.

2. The fault diagnosis method for cable production equipment based on machine learning according to claim 1, characterized in that, The multi-source time-series data specifically includes temperature data, pressure data, current data, vibration data, and tension data of the cable production equipment.

3. The fault diagnosis method for cable production equipment based on machine learning according to claim 1, characterized in that, The preprocessing of multi-source time-series data specifically includes noise filtering, time synchronization, missing value imputation, normalization, and sampling frequency alignment.

4. The fault diagnosis method for cable production equipment based on machine learning according to claim 1, characterized in that, The process of obtaining the fractional-order response sequence and the dynamic memory weight distribution includes: The signals in the standardized dataset are divided into channels, and time series of temperature, pressure, current, vibration and tension channels are established respectively. The maximum historical memory length and sliding time window required for fractional-order operation are set to obtain the standardized time series data of each channel. An adaptive order adjustment factor is constructed based on the amplitude changes, fluctuation intensity, and frequency band energy of each channel within the sliding time window. Based on the adaptive order adjustment factor, the value of the fractional order is automatically adjusted according to the dynamic characteristics of different channel signals within the sliding time window. The initial value of the fractional order is smoothly updated and upper and lower bounds are constrained during the time series progression, forming a fractional order that adapts to time. Based on fractional order, a weighted memory kernel function that monotonically decays with time lag is constructed for each channel. The influence weight of historical samples on the current moment within the sliding time window is calculated, and a dynamic memory weight distribution describing the long-term memory characteristics of each channel signal is generated. The historical samples refer to the historical standardized time series data of the cable production equipment within the sliding time window. Based on the fractional order and dynamic memory weight distribution of each channel, a fractional operator is constructed using the fractional derivative defined by Caputo. Fractional differential operations are performed on the standardized time series data of each channel to obtain the fractional response sequence of each channel that reflects the long-term dependence and time series variation of the signal.

5. The fault diagnosis method for cable production equipment based on machine learning according to claim 1, characterized in that, The construction of long-term memory features and the formation of an exogenous input vector sequence include: The fractional response signals of each channel are normalized and time-synchronized according to the dynamic memory weights to generate a fractional feature matrix, and boundary consistency correction is performed at the time boundaries of each channel. Based on the fractional characteristic matrix and the corresponding fractional order of each channel, the fractional derivative sequence of each channel, which characterizes the nonlinear variation law of the operating signal of the cable production equipment in the time dimension, is calculated. Based on the fractional-order feature matrix and dynamic memory weight distribution of each channel, a weighted convolution operation is performed on the signal of each channel to generate a memory kernel convolution sequence; Based on the fractional derivative sequence and memory kernel convolution sequence of each channel, the fractional energy index representing the cumulative energy intensity of the signal fractional response is calculated within a set sliding time window. The fractional derivative sequence, memory kernel convolution sequence, and fractional energy index are fused to form a comprehensive long-term memory feature set through time step alignment and feature concatenation. The comprehensive long-term memory feature set is unfolded into an exogenous input vector sequence in chronological order.

6. The fault diagnosis method for cable production equipment based on machine learning according to claim 1, characterized in that, The improved NARX neural network model that achieves convergence includes: An improved NARX neural network model is constructed, which consists of a fractional-order memory-gated layer, a structured exogenous fusion layer, a delay-shared basis hysteresis mapping layer, a residual feedforward backbone layer, and a parallel output branch, wherein: Fractional memory gating layers characterize the long-term dependency characteristics of operating signals of cable production equipment through fractional memory factors; The structured exogenous fusion layer fuses exogenous input vector sequences to generate fused high-dimensional temporal features and extracts the dynamic coupling relationship between multiple physical quantities; The delayed shared base lag mapping layer uniformly encodes the historical output lag sequence and establishes short-term time dependencies. The historical output lag sequence refers to the sequence formed by automatically caching the predicted outputs of the improved NARX neural network model at several previous time steps and arranging them in reverse chronological order. The residual feedforward backbone layer performs nonlinear mapping and error correction on the fused high-dimensional temporal features; The parallel output branches include a state prediction branch and an uncertainty output branch. The state prediction branch outputs the fault state prediction result, and the uncertainty output branch outputs the prediction variance information corresponding to the result. Training samples are constructed based on standardized datasets and corresponding real observation outputs, and the real observation outputs at corresponding times are used as supervision signals to form a training sample sequence arranged in chronological order. The training objectives and constraints are set, and the prediction error is measured by heteroscedastic regression. The error term is dynamically weighted based on the uncertainty output of the improved NARX neural network model. At the same time, the prediction sequence is subject to smoothing constraints, structural constraints, and process boundary constraints. A gradient-based optimization method is used to iteratively update the parameters of the improved NARX neural network model. At the same time, the channel order parameters of the fractional memory gate layer are jointly updated. The values ​​of the channel order parameters are limited to zero and one. During the iteration process, the loss value is gradually reduced according to the joint objective of training and constraints until convergence on the training set. The improved NARX neural network model is evaluated on an independent validation set. Based on the comprehensive index of the validation set and the early stopping criterion, the convergent network weight parameters and channel order parameters are determined. The model structure and parameters are fixed to obtain a convergent improved NARX neural network model.

7. The fault diagnosis method for cable production equipment based on machine learning according to claim 1, characterized in that, The failure risk index of the generated cable production equipment includes: The exogenous input vector sequence and the historical output lag sequence are used as the input at the current time. Fractional-order memory gating layer processing is performed on the exogenous input vector sequence. The historical segments are weighted and converged according to the variable-order parameters of each channel, the fractional-order memory factor and the dynamic memory weight to obtain a memory-enhanced representation containing long-term memory information. The memory-enhanced representation is input into the structured exogenous fusion layer, and multi-level fusion is performed in the channel dimension and time dimension according to the preset structured weights to generate fusion features that represent the dynamic coupling relationship of multiple physical quantities. The historical output lag sequence is input into the delayed shared basis lag mapping layer, and the output of different lag steps is uniformly encoded according to the shared delay weight to obtain the lag mapping representation that represents the short-term time dependence. The fused features and the hysteresis mapping representation are cascaded and then input into the residual feedforward backbone layer. Nonlinear mapping and error correction are completed through multiple residual units to obtain a high-order implicit representation. The parallel output branch maps the higher-order implicit representation to obtain the state prediction value and the corresponding prediction variance information of the cable production equipment. The state prediction value is calculated by the state prediction branch, and the prediction variance is calculated by the uncertainty output branch. Based on fractional-order feature weights, the output layer remapping is performed on the state prediction values ​​to weight and integrate the fractional-order long-term memory information with the current prediction results to generate the fault risk index of cable production equipment. Output the predicted state value, prediction variance information and fault risk index, and write the current state prediction value into the prediction cache to update the historical output lag sequence. At the same time, record the time identifier for easy retrieval at the next time step.

8. The fault diagnosis method for cable production equipment based on machine learning according to claim 1, characterized in that, The generation of the abnormal scoring sequence includes: The difference between the predicted state value and the corresponding measured value is calculated to generate the residual data at the current moment. The residual data at multiple consecutive moments are collected within a preset sliding time window to form a residual sequence. The corresponding measured value refers to the actual operating data collected synchronously at the prediction moment of the converged improved NARX neural network model, specifically including temperature data, pressure data, current data, vibration data and tension data of the cable production equipment. Based on the dynamic memory weight distribution, a memory weight is set for each lag step. The memory weight is determined by the order parameter of the fractional derivative and the weighted memory kernel function, which describes the degree of influence of the residual at different time steps on the current energy calculation. The residual sequence is weighted and squared to obtain the fractional residual energy index at the current time step by performing weighted summation based on memory weights, which represents the weighted cumulative strength of the residual in the time dimension. Based on the prediction variance information, the fractional residual energy index is standardized. The standardized fractional residual energy index and the fault risk index are aligned by time and then risk-weightedly fused to obtain the standardized anomaly score. The standardized anomaly scores obtained at each time point are arranged in chronological order to form an anomaly score sequence and then output.

9. The fault diagnosis method for cable production equipment based on machine learning according to claim 1, characterized in that, The determination and output of the fault status and degradation trend level of the cable production equipment includes: Within the sliding time window, the distribution characteristics of abnormal scoring sequences are statistically analyzed to form a benchmark sample set, and the benchmark sample set is segmented and managed according to different process stages based on the production cycle. The quantile values ​​are calculated for the benchmark sample set according to the preset quantile ratio to obtain the initial quantile threshold. The initial quantile threshold is then smoothly updated based on the change range and stability of abnormal scores within the sliding time window to form the quantile adaptive threshold. The current abnormal score is compared with the quantile adaptive threshold. If the abnormal score exceeds the quantile adaptive threshold, a threshold overrun event is recorded, and the number of consecutive overruns and the duration are accumulated. Based on the number of consecutive boundary crossings, duration, and prediction variance information at the corresponding time, a comprehensive judgment rule is executed to output the fault status judgment result, and the degradation trend is classified according to the boundary crossing magnitude and persistence. Write the fault status determination result and degradation trend level into the diagnostic log, update the reference sample set and statistics of the quantile adaptive threshold, and output the fault status determination result, degradation trend and time identifier to the upper system.

Images