A dual-mode communication unit fault early warning method based on timing characteristics

Abstract

This invention relates to the field of communication equipment fault early warning technology, specifically to a fault early warning method for dual-mode communication units based on time-series features. The method includes: acquiring historical operating parameters and communication mode identifiers of the dual-mode communication unit to construct a multi-channel input tensor; introducing the controller's hidden state as a cross-modal context at mode switching moments; inputting the tensor into a memory-enhanced neural network, comparing it with a normal state prototype set to generate a state deviation vector; using this vector to weight the information to be written at the feature level and generating a write address through mapping offset superposition; updating the memory by selecting overwrite or linear fusion according to the norm; iteratively searching, selecting candidate memory items for attention fusion in each round, forming an ordered search path sequence; aggregating the path sequence into a path embedding vector and concatenating it with the hidden state, inputting it into a fault prediction network to output the fault probability. The solution of this invention can improve the reliability and accuracy of fault early warning for dual-mode communication units.

Description

Technical Field

[0001] This invention relates to the field of communication equipment fault early warning technology, and specifically to a fault early warning method for dual-mode communication units based on timing characteristics. Background Technology

[0002] As a core node for data acquisition and interaction, the dual-mode communication unit's operational status directly impacts the stability of the entire communication network and the normal operation of business systems. In actual deployment environments, dual-mode communication units typically face severe electromagnetic interference and variable channel conditions. Aging internal components or sudden changes in the external environment can easily lead to communication interruptions or performance degradation. By monitoring the timing parameters of the dual-mode communication unit in real time and implementing proactive fault warnings, timely maintenance interventions can be taken before substantial equipment downtime. This has significant engineering application value in reducing system maintenance costs and ensuring the continuity of critical business data.

[0003] Dual-mode communication units frequently switch communication modes based on channel quality during actual operation. This cross-mode switching causes drastic fluctuations in operating parameters, accompanied by contextual coupling. Traditional time-series analysis methods cannot identify and utilize the implicit correlation features before and after mode switching. The changing operating state of dual-mode communication equipment is a long-term process, and the early warning features of faults contained in large amounts of operational data are often extremely weak and easily masked by normal fluctuations in noise. Existing early warning models, when processing long-term operational data, cannot establish a deep time-series mapping between the current operating state and historical normal or abnormal states. These problems prevent existing technologies from deeply decomposing potential fault change patterns when facing the complex operating conditions of dual-mode communication units, and thus fail to meet the requirements of modern high-reliability communication networks for intelligent operation and maintenance.

[0004] Chinese patent document CN116156588B discloses a communication method and apparatus for a dual-mode communication network. Based on monitoring the communication information transmission process, it constructs an early warning model to analyze the communication status and combines this with communication mode switching to provide early warnings of potential faults, thereby improving the operational reliability of the dual-mode communication network. However, the early warning modeling in this patent document mainly relies on conventional analysis of communication characteristics. It struggles to maintain a consistent representation of the operational status during the transition between communication modes, and lacks the ability to continuously track and adaptively perceive early, subtle, and gradual fault precursors. In scenarios with frequent changes in operating conditions or slow drift in operational behavior, the reliability and timeliness of the early warning remain insufficient.

[0005] Chinese patent document CN115144699B discloses a method for locating fault sections in distribution networks based on memory-enhanced neural networks. This method utilizes memory-enhanced neural networks to learn and memorize operational data of the distribution network, achieving adaptive location of fault sections with the help of external memory. However, this patent focuses on section location after a distribution network fault occurs, constituting a post-event discriminative application. It struggles to predictively warn of gradual changes in operational status in the early stages before a fault manifests. Furthermore, it is geared towards relatively stable distribution network operation scenarios, lacking adaptability to the discontinuities introduced by dynamic switching of communication modes, and its predictive stability remains insufficient under cross-modal operating conditions. Summary of the Invention

[0006] To achieve accurate monitoring and early warning of faults in the operation of dual-mode communication units, this invention provides a fault warning method for dual-mode communication units based on time-series characteristics, comprising:

[0007] The process involves acquiring the timing parameters and communication mode identifiers of the dual-mode communication unit within a preset historical time period to construct a multi-channel input tensor. At the communication mode switching moment, the controller hidden state at the moment before the switch is extracted as a cross-modal context channel, and this channel is merged with the timing parameters within the preset historical time period. The multi-channel input tensor is input to the controller module of the memory-enhanced neural network, which generates unnormalized score vectors of the information to be written and the basic write address, and calculates a state deviation vector. The state deviation vector is generated by comparing the current multi-channel input tensor with the normal state prototype set in the external memory matrix, and the normal state prototype set is maintained during real-time operation. The state deviation vector is used to perform feature-level weighting on the information to be written, and the state deviation vector is transformed into an address offset vector through a mapping network, which is then superimposed with the unnormalized score vector of the basic write address to generate the write address. A memory update strategy of overwriting or linear fusion is selected based on the norm of the state deviation vector. The controller module generates an initial read query vector, iteratively searches the external memory matrix, and extracts data based on similarity in each round. Each candidate memory item is weighted and fused using an attention mechanism to update the read query vector. The next iteration forms an ordered retrieval path sequence; the ordered retrieval path sequence is input into the sequence aggregation module to generate a path embedding vector, and the vector is concatenated with the current hidden state of the controller and input into the fault prediction network to output the future fault probability.

[0008] By acquiring the temporal operating parameters and communication mode identifiers of the dual-mode communication unit within a preset historical time period, a multi-channel input tensor is constructed. At the moment of communication mode switching, the controller hidden state at the moment before the switch is extracted as a cross-modal context channel and merged with the temporal operating parameters, avoiding the loss of continuous information of the operating state during mode switching. This allows the input features to simultaneously carry normal operating attributes and cross-modal temporal context. By comparing the state deviation vector with the normal state prototype set, the information to be written is weighted at the feature level and superimposed with the address offset to generate the write address, so that the information written to the external memory matrix focuses on fault precursor factors. An ordered retrieval path sequence is formed through iterative retrieval, and the path embedding vector is generated by sequence aggregation and then concatenated with the controller hidden state and input into the fault prediction network. This makes the output of the future fault probability integrate the state continuity features and multi-hop association reasoning results, improving the reliability of fault early warning of the dual-mode communication unit.

[0009] Furthermore, a multi-channel input tensor is constructed, including: at the moment of communication mode switching, truncating the current sampling period, obtaining a historical feature matrix composed of communication mode identifier, signal amplitude, system transmission delay and packet error rate within a preset historical time period with a fixed step size; synchronously extracting fixed-dimensional state features from the controller feedback layer, copying and expanding the fixed-dimensional state features by the fixed step size into a state feature matrix matching the time dimension of the historical feature matrix; and concatenating the expanded state feature matrix with the historical feature matrix in the feature dimension to generate a multi-channel input tensor that takes into account both normal operation attributes and state continuity attributes.

[0010] By truncating the current sampling period at the moment of communication mode switching and obtaining a historical feature matrix composed of communication mode identifier, signal amplitude, system transmission delay and packet error rate with a fixed step size, the fixed-dimensional state features of the controller feedback layer are simultaneously copied and expanded into a time-dimensional matching state feature matrix with a fixed step size and then concatenated in the feature dimension. This avoids the problem that the historical running attributes and the internal state continuity attributes of the neural network are inconsistent in the time dimension and cannot be integrated. The generated multi-channel input tensor takes into account both the fluctuation law of device operation and cross-modal temporal context, providing a complete and aligned feature basis for subsequent memory writing and retrieval.

[0011] Furthermore, the timing operation parameters include power supply voltage, operating temperature, and transmission power. The communication mode identifier is obtained by collecting high and low level data from the RF switching pins of the control chip, mapping the low level to power line carrier communication mode and the high level to microwave communication mode. After obtaining the timing operation parameters, a linear interpolation function is used to fill in the missing timing operation parameters, and a standard deviation normalization algorithm is used for processing. An edge detection algorithm is used to monitor the level transition edge of the communication mode identifier to determine the communication mode switching time.

[0012] Furthermore, maintaining the normal state prototype set includes: extracting time-series features from the equipment's historical fault-free operation cycles and inputting them into an unsupervised clustering algorithm; calculating cluster centers as initial normal state prototypes and storing them in an external memory matrix; during the real-time operation phase, calculating the Euclidean distance between the real-time operation data and each initial normal state prototype; when the Euclidean distance is less than a preset safety threshold and the Euclidean distance between the real-time operation data and the corresponding initial normal state prototype is less than a preset absolute benchmark threshold, updating the feature vector of the target prototype using an exponential moving average algorithm at a preset update rate to adapt to the behavioral changes of the dual-mode communication unit.

[0013] By extracting time-series features from the equipment's historical fault-free operation cycles and performing unsupervised clustering to obtain cluster centers as initial normal state prototypes, which are stored in an external memory matrix, exponential moving average updates are triggered under the dual conditions of Euclidean distance less than a preset safety threshold and Euclidean distance from the corresponding initial normal state prototype less than a preset absolute benchmark threshold. This avoids the mistaken inclusion of abnormal data in the prototype update, which would cause prototype drift. The normal state prototype can suppress transient noise interference while smoothly absorbing the gradual behavior caused by changes in ambient temperature and natural aging of hardware, thus improving the ability to perceive the evolution of early minor faults.

[0014] Further, calculating the state deviation vector includes: obtaining the current input feature vector, which is mapped to the normal state prototype dimension after feature reduction from the current multi-channel input tensor; calculating the Euclidean distance between the current input feature vector and each normal state prototype in the external memory matrix; selecting the normal state prototype with the smallest Euclidean distance as the reference state vector; performing element-wise subtraction between the current input feature vector and the reference state vector; retaining the positive and negative evolution directions of the difference vector obtained by subtraction to extract the deviation components; and outputting the state deviation vector.

[0015] By mapping the current multi-channel input tensor to a current input feature vector with the same dimension as the normal state prototype through feature dimensionality reduction, and selecting the prototype with the smallest Euclidean distance as the reference state vector for element-wise subtraction, the deviation component is extracted by preserving the positive and negative evolution directions of the difference vector. This avoids the problem of absolute value transformation or squaring processing masking the deviation direction information, so that the state deviation vector can characterize the microscopic direction and degree of deviation of the current communication state from the nearest healthy prototype, providing a directional discrimination basis for subsequent feature-level weighting and addressing offset.

[0016] Furthermore, the state deviation vector is used to perform feature-level weighting on the information to be written, including: inputting the state deviation vector into a multilayer perceptron composed of fully connected layers and activation functions to generate a feature weight vector that retains positive and negative adjustment information; performing a Hadamard product operation between the feature weight vector and the information vector to be written, and amplifying or attenuating the value of the feature dimension related to the state deviation according to the evolution direction, so as to adjust the representation weight of abnormal states during the memory writing process.

[0017] By inputting the state deviation vector into a multilayer perceptron consisting of fully connected layers and activation functions, a feature weight vector that retains positive and negative adjustment information is generated. This weight vector is then subjected to a Hadamard product operation with the information vector to be written. Based on the evolution direction, the feature dimension values ​​related to the state deviation are amplified, while the feature values ​​in normal convergence fluctuations are attenuated. This avoids the problem of background noise features and fault precursor features being treated equally during memory writing. As a result, the content information stored in the external memory matrix focuses on the precursor factors that cause dual-mode communication failures, thus improving the purity and focus of the written information.

[0018] Further, the state deviation vector is transformed into an address offset vector through a mapping network, and then superimposed with the unnormalized score vector of the base write address to generate a write address. This includes: inputting the state deviation vector into a feedforward neural network with a single hidden layer, and outputting an intermediate vector with the same address dimension as the external memory matrix; processing the intermediate vector through a Tanh function to generate an address offset vector containing positive and negative offset features; performing a vector addition operation between the address offset vector and the unnormalized score vector of the base write address, and normalizing the sum using a Softmax function to generate a write address whose sum of elements is 1.

[0019] By inputting the state deviation vector into a feedforward neural network with a single hidden layer and processing it with the Tanh function to generate an address offset vector containing positive and negative offset features, and then adding it to the unnormalized score vector of the base write address, followed by Softmax normalization to generate a write address whose sum of elements is 1, the problem of determining the write position solely based on the input itself while ignoring the direction of state deviation is avoided. This allows the addressing weights to be adjusted according to the positive or negative offset direction and magnitude of each slot, resulting in a differentiable continuous write address, thereby enabling more targeted control of soft overwrite operations on memory blocks.

[0020] Further, the memory update strategy of overwriting or linear fusion is selected based on the norm of the state deviation vector, including: calculating the L2 norm of the state deviation vector; when the L2 norm is greater than a preset hyperparameter threshold, triggering the overwriting memory update strategy, extracting the maximum value index in the write address probability distribution, and replacing the storage slot data corresponding to the maximum value index in the external memory matrix with the feature-weighted information to be written; when the L2 norm is less than or equal to the preset hyperparameter threshold, triggering the linear fusion memory update strategy, and updating and writing the original memory data corresponding to the write address and the information to be written to the external memory matrix by linearly weighting and summing them according to a fusion coefficient proportional to the L2 norm.

[0021] By calculating the L2 norm of the state deviation vector and selecting a memory update strategy based on its relationship with a preset hyperparameter threshold, a write-over strategy is triggered when the norm is large, replacing the storage slot corresponding to the highest probability index with the feature-weighted information to be written. When the norm is small, a linear fusion strategy is triggered, which sums the original memory data and the information to be written in a weighted manner according to a fusion coefficient proportional to the L2 norm. This avoids the problem of diluting obvious abnormal information or overwriting small fluctuations by using a single update method for states with different degrees of deviation. This makes the memory update amplitude match the degree of state deviation, improving the stability and adaptability of the external memory matrix content.

[0022] Furthermore, in each round, extraction is performed based on similarity. The candidate memory items are weighted and fused using an attention mechanism to update the read query vector, including: calculating the current read query vector and... The dot product similarity of the content vectors of each candidate memory item is calculated, and the attention weight of each candidate memory item is obtained by normalization using the Softmax function; the attention weight is then used to... The content vectors and address vectors of each candidate memory item are weighted and summed to obtain a fused content vector and a fused address vector. The current read query vector, the fused content vector, and the fused address vector are concatenated along the feature dimension and used as the input of the gated loop unit at the current time step for state update, outputting the read query vector for the next iteration.

[0023] By calculating the dot product similarity between the current read query vector and the candidate memory item content vector and normalizing it using Softmax, attention weights are obtained. The content vectors and address vectors of the candidate memory items are then weighted and summed to obtain the fused content vector and fused address vector. The read query vector, fused content vector, and fused address vector are then concatenated and input into the gated recurrent unit for state updates. This avoids the problem that a single round of retrieval cannot cover deep-level related information. Each round of retrieval uses both content information and addressing information to guide the correction of the query direction, outputting a more accurate read query vector for multi-stage progressive related reasoning.

[0024] Further, the ordered retrieval path sequence is input into the sequence aggregation module to generate a path embedding vector, including: The fused address vector and fused content vector generated in the next iteration are concatenated along the feature dimension to form a structure of length [length missing]. The time sequence is input into a recurrent neural network module to extract the temporal dependencies and logical deduction features on the iterative path, and the output of the last time step of the recurrent neural network module is used as the path embedding vector.

[0025] The beneficial effects of this invention are as follows: By introducing a hidden state of the controller during communication mode switching to construct a multi-channel input tensor that takes into account both operational fluctuation patterns and cross-modal temporal context, the state continuity information during the switching process is preserved; relying on the dual-condition triggering and exponential moving average update of the normal state prototype set, transient noise is suppressed while the gradual behavior caused by environmental changes and hardware aging is smoothly absorbed, enhancing the perception of early minor fault evolution; retaining the state deviation vectors of positive and negative evolution directions guides the feature-level weighting and addressing offset of the information to be written, so that memory writing focuses on fault precursors and matches overwrite or linear fusion strategies according to the degree of deviation, improving the purity and stability of the external memory matrix content; by using the ordered path sequence formed by multi-hop retrieval to extract temporal dependence and logical deduction features through sequence aggregation, the fault probability prediction integrates state continuity and multi-hop correlation reasoning results, improving the reliability and accuracy of fault early warning for dual-mode communication units. Attached Figure Description

[0026] Figure 1 This is a schematic flowchart illustrating a fault early warning method for a dual-mode communication unit based on timing characteristics according to an embodiment of the present invention; Figure 2 This is a weighted schematic diagram of abnormal state features in an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the predictive performance evaluation in an embodiment of the present invention. Detailed Implementation

[0027] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] An embodiment of a fault early warning method for dual-mode communication units based on time-series characteristics: like Figure 1 As shown, the present invention provides a fault early warning method for a dual-mode communication unit based on timing characteristics, comprising: S1. Obtain the timing operation parameters and communication mode identifier of the dual-mode communication unit within a preset historical time period, and construct a multi-channel input tensor. At the moment of communication mode switching, extract the controller hidden state at the moment before the switch as a cross-modal context channel, and merge the channel with the timing operation parameters within the preset historical time period.

[0029] The system reads the power supply voltage, operating temperature, and transmission power timing parameters of the dual-mode communication unit within a preset historical time period. It also collects high and low level data from the RF switching pins of the control chip to obtain the communication mode identifier. Low levels are mapped to power line carrier communication mode, and high levels to microwave communication mode. Missing timing parameters are filled using a linear interpolation function and processed using a standard deviation normalization algorithm. The `torch.stack` function of the PyTorch framework combines the timing parameters and communication mode identifier along the feature dimension to construct a multi-channel input two-dimensional tensor multiplied by the feature dimension. An edge detection algorithm is used to monitor the level transition edges of the communication mode identifier in real time to determine the communication mode switching time. At this time, the hidden state vector of the Long Short-Term Memory network controller module at the moment before the switching is extracted using PyTorch's hidden state call interface. The `torch.unsqueeze` function expands the dimension of this hidden state vector to generate a cross-modal context channel. The `torch.cat` function is then called to concatenate the cross-modal context channel with the timing parameters along the channel dimension, completing the merging of the multi-channel input tensor.

[0030] As an optional implementation, a multi-channel input tensor is constructed, including: At the moment of communication mode switching, the current sampling period is truncated, and a historical feature matrix composed of communication mode identifier, signal amplitude, system transmission delay and packet error rate with a fixed step size is obtained within a preset historical time period. Synchronously extract fixed-dimensional state features from the controller feedback layer, and copy and expand the features according to the time step to form a state feature matrix that matches the time dimension of the historical feature matrix; The expanded state feature matrix is ​​concatenated with the historical feature matrix along the feature dimension to generate a multi-channel input tensor that takes into account both normal operation attributes and state continuity attributes.

[0031] At the moment of switching between dual-mode communication such as power line carrier and low-power wireless, for example in The mode switching command is triggered at any time, truncating the current sampling period and retrieving past data. Time series data at time steps, where the preset historical time period is... The sampling frequency is Step size is These data specifically include communication mode identification data and normalized signal amplitude, such as signal strength mapped to the range of 0 to 1, and system transmission delay, for example... to Compared with packet error rate, for example to The above four types of feature data are combined along the time dimension to form a structure of size [size missing]. The historical feature matrix. From the feedback layer of the controller of the memory-enhancing neural network, such as the Long Short-Term Memory neural network, extract the fixed-dimensional state features at the moment before the communication switch, that is, the hidden layer state vector with a dimension of 64.

[0032] The flat expansion operation was used to expand the area. The state feature vector is copied 50 times along the time axis and expanded to the same size. The state feature matrix. In terms of feature dimensions... Historical feature matrix and The state feature matrices are concatenated and spliced ​​together to output a size of [size missing]. The system uses multi-channel input tensors to identify fluctuations in device operation and coherent cross-modal temporal context features within the neural network. The Long Short-Term Memory (LSTM) neural network takes a multi-channel input tensor as input and outputs the hidden layer state vector at the current time step. The LSM's structure comprises four core components: a forget gate, an input gate, an output gate, and cell states. Gating mechanisms control the retention and discarding of information.

[0033] S2, the multi-channel input tensor is input to the controller module of the memory-enhancing neural network. The controller module generates unnormalized score vectors of the information to be written and the basic write address, and calculates the state deviation vector. The state deviation vector is generated by comparing the current multi-channel input tensor with the normal state prototype set in the external memory matrix, and the normal state prototype set is maintained during real-time operation. The state deviation vector is used to perform feature-level weighting on the information to be written, and the state deviation vector is transformed into an address offset vector through a mapping network. This offset vector is then superimposed with the unnormalized score vector of the basic write address to generate the write address. The memory update strategy of overwriting or linear fusion is selected based on the norm of the state deviation vector.

[0034] The multi-channel input tensor is fed into a controller module based on a Long Short-Term Memory (LSTM) network. The hidden layer features output from this module are then fed into two fully connected layers constructed using `torch.nn.Linear`. Forward computation generates the feature vector of the information to be written and the unnormalized score vector of the base write address. The `torch.cdist` function is used to calculate the feature vector of the multi-channel input tensor and the feature vector of the external memory matrix. The mean clustering algorithm pre-extracts the Euclidean distances between prototype vectors within the normal state prototype set, selects the prototype vector with the smallest Euclidean distance, and uses vector subtraction to calculate the difference between the input feature vector and this smallest prototype vector to generate a state deviation vector. During real-time operation, the exponential moving average algorithm is used to maintain the normal state prototype set, with a momentum parameter set to 0.9. The torch.lerp function is called to fuse and update the current normal state features into the corresponding prototype vectors.

[0035] The state bias vector is input into the Sigmoid activation function for nonlinear mapping to obtain feature-level weights between 0 and 1. Element-wise multiplication is used to multiply these weights with the information vector to be written, completing feature-level weighting. The state bias vector is then input into a mapping network constructed from a multilayer perceptron for feature dimensionality reduction, outputting an address offset vector. A tensor addition operator is used to add the address offset vector to the unnormalized score vector of the base write address. The superposition result is normalized to generate the write address vector. The L2 norm of the state bias vector is calculated. When the L2 norm is greater than a preset hyperparameter threshold, a write memory update strategy is triggered. The maximum index in the write address probability distribution is extracted, and the storage slot data corresponding to this maximum probability index in the external memory matrix is ​​replaced with the feature-weighted information to be written. When the L2 norm is less than or equal to the preset hyperparameter threshold, a linear fusion memory update strategy is triggered. A scalar multiplication operator is used to linearly weight and sum the original memory data corresponding to the write address and the information to be written according to a fusion coefficient proportional to the L2 norm, and then update and write this sum to the external memory matrix.

[0036] As an optional implementation, the state deviation vector is generated by comparing the current multi-channel input tensor with the normal state prototype set in the external memory matrix, and the normal state prototype set is maintained during real-time operation, including: Extract time-series features from the equipment’s historical fault-free operation cycles and input them into an unsupervised clustering algorithm. Calculate the cluster centers as the initial normal state prototypes and store them in an external memory matrix. During the real-time operation phase, the Euclidean distance between the real-time operation data and each prototype is calculated. When the distance is less than a preset safety threshold and the Euclidean distance between the operation data and the corresponding initial normal state prototype is less than a preset absolute benchmark threshold, the feature vector of the target prototype is updated using the exponential moving average algorithm at a preset update rate to adapt to the behavior changes of the communication unit.

[0037] During the offline initialization phase of the model, time-series data was collected from communication equipment that had been running continuously for 720 hours and confirmed to be fault-free. Deep features were then extracted using a one-dimensional convolutional neural network and applied... The mean-based unsupervised clustering algorithm divides the features into N clusters, for example, N=20. The arithmetic mean of the feature vectors of all samples within each of these 20 clusters is calculated, serving as the initial 20 normal state prototypes. Each prototype is a 128-dimensional floating-point vector, and these prototypes are persistently stored as reference keys in the dedicated address space of the external memory matrix. During the real-time prediction phase, the model extracts the 128-dimensional mapping features of the current multi-channel input tensor at fixed time intervals and calculates the Euclidean distance between this real-time feature vector and the 20 normal prototypes in memory in parallel. A system safety distance threshold D=2.5 is set, and an absolute reference distance threshold of 3.0 is set. If the calculated minimum Euclidean distance, for example, 1.8, is less than 2.5, and the Euclidean distance between the current feature vector and the j-th initial normal state prototype (i.e., the fixed, non-updated anchor key in the external memory matrix) is less than the absolute reference distance threshold, then it is determined that the device data at the current moment has not failed, but is in a newly added normal operating evolution state. At this point, the j-th target prototype with the smallest distance is locked, and the weight is updated using the exponential moving average algorithm, as shown in the formula: Among them, the update rate parameter The preferred setting is 0.02. This is the 128-dimensional vector of the target prototype at the previous moment. This is the current input of 128-dimensional features. This mechanism can smoothly absorb the gradual behavioral characteristics of dual-mode communication units caused by environmental temperature or natural hardware aging while suppressing transient noise interference. The input of the one-dimensional convolutional neural network is time-series data of the device running continuously for 720 hours and confirmed to be fault-free, and the output is the dimensionality-reduced depth feature vector. The structure of this one-dimensional convolutional neural network includes one-dimensional convolutional layers for extracting temporal local features, nonlinear activation layers for feature mapping, and pooling layers for feature dimensionality reduction and preservation of salient features.

[0038] As an optional implementation, the state deviation vector is calculated, including: Obtain the current input feature vector that is consistent with the dimension of the normal state prototype after feature dimensionality reduction mapping of the current multi-channel input tensor, and calculate the Euclidean distance between the vector and each normal state prototype in the external memory matrix; The prototype with the smallest Euclidean distance is selected as the reference state vector, and the current input feature vector is subtracted from the reference state vector element by element. The deviation components are extracted by retaining the positive and negative evolution directions of the difference vector obtained by subtraction, and the state deviation vector is output.

[0039] The 50×68 multichannel input tensor is flattened and then fed into a mapping module consisting of two fully connected layers with 256 and 128 neurons respectively. After nonlinear activation by a linear rectified function, a 128-dimensional current input feature vector is output. Parallel computation is achieved by utilizing the broadcast mechanism of the tensor computation framework. The Euclidean distances to the N existing 128-dimensional normal state prototypes in the external memory matrix are calculated. The prototype vector with the smallest distance (e.g., 1.25) is selected using a minimum pooling operator and extracted as a high-confidence reference state vector. .implement and The element-wise subtraction operation is This generates a difference vector that is also 128-dimensional.

[0040] To represent the actual direction of communication feature evolution, such as whether communication latency is gradually increasing or decreasing, absolute value conversion or squaring is skipped, and the difference vector containing positive, negative, and zero values ​​is retained as the output state bias vector. For example, the element of the 5th dimension in this vector might be positive 0.45, and the element of the 8th dimension might be negative 0.32. This bias component represents the microscopic direction and degree of deviation of the current dual-mode communication state from the most recent healthy prototype. The input of the fully connected network is the flattened multi-channel input tensor, and the output is the current input feature vector. The structure of this fully connected network includes an input layer, a first fully connected layer, an activation function layer, and a second fully connected layer. Each node is connected to all nodes in the previous layer.

[0041] As an optional implementation, feature-level weighting is performed on the information to be written using the state deviation vector, including: The state bias vector is input into a multilayer perceptron consisting of fully connected layers and activation functions to generate a feature weight vector that retains positive and negative adjustment information. The feature weight vector is subjected to a Hadamard product with the original information vector to be written. The value of the feature dimension related to the state deviation is amplified or attenuated according to the evolution direction in order to adjust the representation weight of abnormal states during the memory writing process.

[0042] The extracted 128-dimensional state bias vector is input into a multilayer perceptron consisting of a single fully connected layer and a customized activation function with a scaling range of 0.5 to 1.5. This calculation maps the feature space containing positive and negative bias values ​​to a 128-dimensional feature weight vector. Dimensions representing positive bias expansion are mapped to weights greater than 1, such as 1.35, while dimensions representing negative bias or convergence are mapped to weights less than 1, such as 0.85. The original 128-dimensional information vector to be written, generated internally by the controller through logical computation, is extracted in parallel, and then the vector is multiplied element-wise by a Hadamard product with the newly generated feature weight vector. For example, when the original value to be written for a feature dimension representing an abnormal increase in packet loss rate is 0.6, and the weight calculated from the corresponding state deviation is 1.35, the multiplied feature value is amplified to 0.81. Conversely, if the evolution of an irrelevant environmental noise feature is in a normal convergence fluctuation, with a calculated weight of 0.85 and an original value of 0.5, the corresponding feature value is attenuated to 0.425. This operational mechanism enhances abnormal features based on the fault evolution direction and actively suppresses background interference, ensuring that the information subsequently stored in the external memory matrix focuses on the precursor factors that cause dual-mode communication failures. The input of the multilayer perceptron is the state deviation vector, and the output is the feature weight vector that retains positive and negative adjustment information. The structure of the multilayer perceptron includes a single fully connected layer and a customized activation function layer, with fully connected nodes handling the mapping between feature dimensions. The calculation process of the multilayer perceptron is as follows: In the formula, The input is the state bias vector of the multilayer perceptron. This is the weight matrix of the single fully connected layer. The bias vector of the single fully connected layer. The intermediate feature vector is obtained by linearly mapping the state deviation vector through the single fully connected layer. The feature weight vector is obtained by mapping using the customized activation function. It is the hyperbolic tangent function. For example... Figure 2 As shown, this figure is a weighted curve of abnormal state features. The horizontal axis corresponds to the 128-dimensional feature index, and the baseline write value is fixed at 0.6. The curve increases the weight difference as the dimension increases, and the gray area distinguishes the variance decay and amplification intervals.

[0043] As an optional implementation, the state deviation vector is transformed into an address offset vector via a mapping network, and then superimposed with the unnormalized score vector of the base write address to generate the write address, including: The state bias vector is input into a feedforward neural network with a single hidden layer, and the output is an intermediate vector with the same dimension as the address of the external memory matrix. The intermediate vector is processed by the Tanh function to generate an address offset vector containing positive and negative offset characteristics; The address offset vector is added to the unnormalized score vector of the base write address output by the controller module, and the result is normalized by the Softmax function to generate a write address whose sum of elements is 1.

[0044] The 128-dimensional state deviation vector obtained from the preprocessing is fed into a feedforward neural network containing a single hidden layer with 64 neurons. This vector is then projected through a linear output layer to generate a 256-dimensional intermediate vector with dimensions equal to the total number of slots in the external memory matrix. The external matrix contains 256 memory slots. This intermediate vector is mapped using a hyperbolic tangent nonlinear activation function, smoothing and truncating the element values ​​of these 256 dimensions within a continuous interval from -1 to 1. This yields an address offset vector representing the direction and magnitude of the positive or negative offset for each slot addressing. Based on this, a 256-dimensional unnormalized score vector of the basic write address, generated by the controller from the current input data (i.e., the unactivated logical data), is extracted. The element values ​​are typically distributed between -5.0 and +5.0. At the tensor operation level, element-wise vector addition is performed between the basic score vector and the address offset vector. For example, if the basic score of a slot is 2.5 and the corresponding address offset vector value is -0.8, the summed score is 1.7.

[0045] The superimposed and adjusted 256-dimensional tensor is input to a normalized exponential function to perform global probability normalization, transforming the addressing weights of all 256 slots into a probability distribution, for example, a maximum weight of 0.42, a second highest of 0.15, and a minimum of 0.0001, with the mathematical sum of all 256 elements being 1. This process generates a differentiable, continuous write address vector, thereby controlling which memory blocks are soft overwrites performed on. The feedforward neural network's input is a state bias vector, and its output is an intermediate vector with the same address dimension as the external memory matrix. The feedforward neural network structure includes an input layer, a single hidden layer, and a linear output layer, with information propagating unidirectionally from input to output.

[0046] S3, the controller module generates an initial read query vector and iteratively searches the external memory matrix, extracting data based on similarity in each round. Each candidate memory item is weighted and fused using an attention mechanism to update the read query vector. The next iteration forms an ordered retrieval path sequence; the ordered retrieval path sequence is input into the sequence aggregation module to generate a path embedding vector, and the vector is concatenated with the current hidden state of the controller and input into the fault prediction network to output the future fault probability.

[0047] The initial read query vector is calculated by inputting the hidden state vector of the Long Short-Term Memory (LSTM) network at the current moment of the controller module into a nonlinear mapping layer composed of a single fully connected layer and a Tanh activation function. The retrieval iteration count is set to five times, and a loop retrieval is performed. In each round of retrieval, the cosine similarity score between the current read query vector and all memory entries in the external memory matrix is ​​calculated. The top three memory entries with the highest similarity scores are extracted as candidate memory entries in descending order. The attention weights between the read query vector and each candidate memory entry are calculated using a dot product attention mechanism. After normalization using the Softmax function, the candidate memory entries are weighted and summed using matrix multiplication operators to generate a fusion context vector. A gated loop unit single-step state update formula is used, with the current read query vector as the historical hidden state and the fusion context vector as the current input, to calculate and generate an updated read query vector for the next round of retrieval. After the loop ends, the read query vectors generated by the five iterations are stacked along the time step dimension to form an ordered retrieval path sequence tensor. The ordered retrieval path sequence tensor is input into a sequence aggregation module containing a torch.nn.Conv1d one-dimensional convolution operator and a global max pooling layer to extract the spatiotemporal correlation features of the retrieval path sequence and compress it to generate a one-dimensional path embedding vector. The torch.cat function is called to concatenate and fuse the path embedding vector with the hidden state vector of the controller module at the current time step in the feature dimension. The fused comprehensive feature vector is then fed into a fault prediction network consisting of a multi-layer fully connected network and a ReLU activation function cascaded for forward inference. At the output layer of the network, a Sigmoid activation function is used to map the inference result to a floating-point value between 0 and 1, outputting the future fault probability of the dual-mode communication unit.

[0048] As an optional implementation, extraction is performed in each round based on similarity. The candidate memory items are weighted and fused using an attention mechanism to update the read query vector, including: Calculate the current read query vector and The dot product similarity of the content vectors of each candidate memory item is calculated, and the attention weight of each candidate memory item is obtained by normalization using the Softmax function. Using attention weights The content vectors and address vectors of each candidate memory item are weighted and summed to obtain the fused content vector and fused address vector. The current read query vector, the fused content vector, and the fused address vector are concatenated along the feature dimension and used as the input to the current time step of the gated loop unit for state updates, outputting the read query vector for the next iteration.

[0049] In the multi-hop retrieval iteration of the external memory, the model uses the approximate nearest neighbor algorithm to initially extract from the full memory in each round. For example, let's define four candidate memory entries as the most matching. Each candidate memory entry contains a 128-dimensional content vector and a 256-dimensional continuous soft-addressable address vector. Soft addressing means that when the network needs to read memory, it weights and sums the data in all memory slots according to the above probability. When the network needs to overwrite memory, it also updates the data in all slots according to the probability distribution. The current 128-dimensional read query vector is extracted and its dot product with the four candidate content vectors is performed to calculate a similarity scalar score, for example, an array consisting of scores of 18.5, 14.2, 8.7, and 5.1. These four scalars are then subjected to exponential scaling and normalization using a normalized exponential function to generate an attention weight distribution array with a sum of 1, for example, an array consisting of 0.65, 0.25, 0.08, and 0.02. Based on the aforementioned attention weight coefficients, the four 128-dimensional content vectors are weighted and summed to generate a single 128-dimensional fused content vector. Simultaneously, the four 256-dimensional address vectors are weighted and summed to generate a single 256-dimensional fused address vector. The model concatenates the original 128-dimensional read query vector, the generated 128-dimensional fused content vector, and the 256-dimensional fused address vector along the last feature dimension, the negative 1 axis, forming a 512-dimensional hybrid tensor. This hybrid tensor serves as the input to the gated recurrent unit at the current time step for state updates, outputting a new 128-dimensional feature vector. This vector serves as a more precise read query vector for the next retrieval iteration, guiding the model to complete multi-stage associative reasoning and progressive deep information retrieval.

[0050] As an optional implementation, the ordered retrieval path sequence is input into the sequence aggregation module to generate a path embedding vector, including: Will The fused address vector and fused content vector generated in the next iteration are concatenated along the feature dimension to form a structure of length [length missing]. The time series; The time series is input into the recurrent neural network module to extract the temporal dependencies and logical deduction features on the iterative path, and the output of the last time step of the recurrent neural network is used as the path embedding vector.

[0051] In the continuous query and retrieval phase of the memory-enhanced network, the total number of global iterations is set. At the completion of each independent iteration, the model concatenates the 256-dimensional continuous retrieval address vector locked by the query with the corresponding extracted and fused 128-dimensional content vector along the feature axis to form a single-step path representation vector of length 384. After three complete iterations, these three independent 384-dimensional path vectors are stacked sequentially along the time dimension to construct a 3×384-dimensional tensor in memory with a time step of 3 and a feature dimension of 384, which serves as the ordered temporal sequence representing the entire multi-hop retrieval trajectory. This 3×384 tensor is then fed into a gated recurrent unit network (GRU) as the sequence aggregation module, according to the iteration time sequence. This GRU has 128 hidden units specifically responsible for extracting the temporal dependency features and multi-hop association evolution relationships along the complete link from the first macroscopic communication state localization to the third microscopic fault mechanism tracing. After the network completes data processing for the third time step, which is the last time step of the retrieval process, the 128-dimensional hidden state tensor output by the gated recurrent unit network at this point is extracted. This tensor encapsulates the deep association reasoning results of the entire cross-modal retrieval path. The tensor is output as a path embedding vector to the fault prediction classifier module to calculate the probability of communication failures occurring within a specific future window. The input to the gated recurrent unit network is an ordered time sequence of the retrieval path, and the output is the path embedding vector of the last time step. The structure of this gated recurrent unit network includes reset gates and update gates, updating and merging past states with the current input through a gating mechanism.

[0052] The experiment used multidimensional time-series data collected over six months of continuous operation from a real dual-mode communication device as the experimental dataset, which was randomly divided into training, validation, and test sets in a 7:1:2 ratio. The experimental hardware platform was configured to have... The high-performance graphics processing unit (GPU) with dedicated video memory uniformly employs a deep learning tensor framework for computation. Experimental settings included a baseline learning rate of 0.001, a batch size of 128, a maximum training epoch of 200, and an early stopping mechanism. The control model group comprised a baseline standard memory-enhanced network model, variant one (without the normal-state prototype maintenance mechanism), variant two (without bias-guided feature-level weighting and address offset writing), and variant three (without the multi-hop retrieval path sequence aggregation module). These were then ablation-compared with the complete model of this invention, which incorporates all innovative mechanisms.

[0053] The baseline model's fault prediction accuracy is The recall rate was The comprehensive evaluation index score is Variant 1's prediction accuracy improved to The recall rate was The comprehensive evaluation index score is Variant 2 achieved a prediction accuracy of 100%. The recall rate was The comprehensive evaluation index score is The prediction accuracy of variant 3 is... The recall rate was The comprehensive evaluation index score is The complete model of this invention performs optimally across all metrics, achieving a prediction accuracy of up to [percentage missing]. The recall rate was The comprehensive evaluation index score reached ,like Figure 3 As shown in the figure, this is a bar chart of the predictive performance of five ablation models. The groups include the baseline, three variants, and the complete model. Each group is simultaneously labeled with accuracy, recall, and overall score. The bar height increases progressively for each group, with the complete model showing higher scores in all three metrics than the other comparison groups. The performance gap between Variant 1 and the complete model demonstrates that the normal prototype set maintenance mechanism and exponential moving average prototype update enhance the model's sensitivity to gradual behaviors such as natural aging and the evolution of early minor faults. The comparison between Variant 2 and the complete model simultaneously verifies that retaining the calculation of positive and negative state biases and address offset adjustments can amplify abnormal correlation features and suppress environmental noise, improving the purity and focus of information written to the external memory matrix. The performance of Variant 3 confirms that concatenating multi-hop retrieval addresses and content into an ordered temporal sequence and using gated recurrent units for feature extraction can identify deep statistical correlations and temporal logical links in the iterative addressing process, improving the reliability and prediction accuracy of fault inference.

[0054] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0055] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fault early warning method for a dual-mode communication unit based on time-series characteristics, characterized in that, Includes the following steps: The timing operation parameters and communication mode identifier of the dual-mode communication unit within a preset historical time period are obtained, and a multi-channel input tensor is constructed. At the time of communication mode switching, the controller hidden state at the moment before the switching is extracted as a cross-modal context channel, and the channel is merged with the timing operation parameters within the preset historical time period. The multi-channel input tensor is input to the controller module of the memory-enhancing neural network. The controller module generates an unnormalized score vector of the information to be written and the base write address, and calculates a state deviation vector. The state deviation vector is generated by comparing the current multi-channel input tensor with the normal state prototype set in the external memory matrix, and the normal state prototype set is maintained during real-time operation. The state deviation vector is used to perform feature-level weighting on the information to be written, and the state deviation vector is transformed into an address offset vector through a mapping network. This offset vector is then superimposed with the unnormalized score vector of the base write address to generate the write address. The memory update strategy of overwriting or linear fusion is selected based on the norm of the state deviation vector. The controller module generates an initial read query vector, iteratively searches the external memory matrix, and extracts data based on similarity in each round. Each candidate memory item is weighted and fused using an attention mechanism to update the read query vector. The next iteration forms an ordered retrieval path sequence; the ordered retrieval path sequence is input into the sequence aggregation module to generate a path embedding vector, and the vector is concatenated with the current hidden state of the controller and input into the fault prediction network to output the future fault probability.

2. The method for early warning of faults in a dual-mode communication unit based on time-series characteristics according to claim 1, characterized in that, Constructing a multi-channel input tensor includes: at the moment of communication mode switching, truncating the current sampling period and obtaining a historical feature matrix composed of communication mode identifier, signal amplitude, system transmission delay, and packet error rate within a preset historical time period with a fixed step size; synchronously extracting fixed-dimensional state features from the controller feedback layer, and copying and expanding the fixed-dimensional state features by the fixed step size into a state feature matrix that matches the time dimension of the historical feature matrix; concatenating the expanded state feature matrix with the historical feature matrix in the feature dimension to generate a multi-channel input tensor that takes into account both normal operation attributes and state continuity attributes.

3. The method for early warning of faults in a dual-mode communication unit based on time-series characteristics according to claim 1, characterized in that, The timing operation parameters include power supply voltage, operating temperature, and transmission power. The communication mode identifier is obtained by collecting high and low level data of the RF switching pin of the control chip, mapping the low level to power line carrier communication mode and the high level to microwave communication mode. After obtaining the timing operation parameters, a linear interpolation function is used to fill in the missing timing operation parameters and a standard deviation normalization algorithm is used for processing. An edge detection algorithm is used to monitor the level transition edge of the communication mode identifier to determine the communication mode switching time.

4. The method for early warning of faults in a dual-mode communication unit based on time-series characteristics according to claim 1, characterized in that, Maintaining the normal state prototype set includes: extracting time-series features from the equipment's historical fault-free operation cycles and inputting them into an unsupervised clustering algorithm; calculating cluster centers as initial normal state prototypes and storing them in an external memory matrix; during the real-time operation phase, calculating the Euclidean distance between the real-time operation data and each initial normal state prototype; when the Euclidean distance is less than a preset safety threshold and the Euclidean distance between the real-time operation data and the corresponding initial normal state prototype is less than a preset absolute benchmark threshold, updating the feature vector of the target prototype using an exponential moving average algorithm at a preset update rate to adapt to the behavioral changes of the dual-mode communication unit.

5. The method for early warning of faults in a dual-mode communication unit based on time-series characteristics according to claim 4, characterized in that, The calculation of the state deviation vector includes: obtaining the current input feature vector, which is mapped to the normal state prototype dimension after feature reduction from the current multi-channel input tensor; calculating the Euclidean distance between the current input feature vector and each normal state prototype in the external memory matrix; selecting the normal state prototype with the smallest Euclidean distance as the reference state vector; performing element-wise subtraction between the current input feature vector and the reference state vector; retaining the positive and negative evolution directions of the difference vector obtained by subtraction to extract the deviation components; and outputting the state deviation vector.

6. The method for early warning of faults in a dual-mode communication unit based on time-series characteristics according to claim 5, characterized in that, The state deviation vector is used to perform feature-level weighting on the information to be written, including: inputting the state deviation vector into a multilayer perceptron consisting of a fully connected layer and an activation function to generate a feature weight vector that retains positive and negative adjustment information; performing a Hadamard product operation between the feature weight vector and the information vector to be written, and amplifying or attenuating the value of the feature dimension related to the state deviation according to the evolution direction, so as to adjust the representation weight of abnormal states during the memory writing process.

7. The method for early warning of faults in a dual-mode communication unit based on time-series characteristics according to claim 1, characterized in that, The process of transforming the state deviation vector into an address offset vector via a mapping network and superimposing it with the unnormalized score vector of the base write address to generate a write address includes: inputting the state deviation vector into a feedforward neural network with a single hidden layer, outputting an intermediate vector with the same address dimension as the external memory matrix; processing the intermediate vector with a Tanh function to generate an address offset vector containing positive and negative offset features; performing a vector addition operation between the address offset vector and the unnormalized score vector of the base write address, and normalizing the sum using a Softmax function to generate a write address whose sum of elements is 1.

8. The method for early warning of faults in a dual-mode communication unit based on time-series characteristics according to claim 7, characterized in that, The memory update strategy of overwriting or linear fusion is selected based on the norm of the state deviation vector, including: calculating the L2 norm of the state deviation vector; when the L2 norm is greater than a preset hyperparameter threshold, triggering the overwriting memory update strategy, extracting the maximum value index in the write address probability distribution, and replacing the storage slot data corresponding to the maximum value index in the external memory matrix with the feature-weighted information to be written; when the L2 norm is less than or equal to the preset hyperparameter threshold, triggering the linear fusion memory update strategy, updating and writing the original memory data corresponding to the write address and the information to be written to the external memory matrix by linearly weighting and summing them according to a fusion coefficient proportional to the L2 norm.

9. The method for early warning of faults in a dual-mode communication unit based on time-series characteristics according to claim 1, characterized in that, Extracting based on similarity in each round The candidate memory items are weighted and fused using an attention mechanism to update the read query vector, including: calculating the current read query vector and... The dot product similarity of the content vectors of each candidate memory item is calculated, and the attention weight of each candidate memory item is obtained by normalization using the Softmax function; the attention weight is then used to... The content vectors and address vectors of each candidate memory item are weighted and summed to obtain a fused content vector and a fused address vector. The current read query vector, the fused content vector, and the fused address vector are concatenated along the feature dimension and used as the input of the gated loop unit at the current time step for state update, outputting the read query vector for the next iteration.

10. A fault early warning method for a dual-mode communication unit based on time-series characteristics according to claim 9, characterized in that, The ordered retrieval path sequence is input into the sequence aggregation module to generate a path embedding vector, including: The fused address vector and fused content vector generated in the next iteration are concatenated along the feature dimension to form a structure of length [length missing]. The time sequence is input into a recurrent neural network module to extract the temporal dependencies and logical deduction features on the iterative path, and the output of the last time step of the recurrent neural network module is used as the path embedding vector.

Images